C Modeling and Simulation
C Modeling and Simulation
MARCELDEKKER,
INC.
D E K K E R
ISBN: 0-8247-0041-4
The publisher offers discounts on this book when orderedin bulk quantities.
For more information, write to Special Sales~rofessionalMarketing at the
address below.
This book is printed on acid-free paper.
ht~:jjww~
dekker.corn
Current printing (last digit):
1 0 9 8 7 6 5 4 3 2 1
$TATE$ OF
CA
iv
reface
reface
Preface
tems, or decision support systems-is presented here as a method of finetuning the parameter values of a crop model which includes the process
of moisture movement in the
soil (Chapter 6). Funda~entalmathematical
methods applied to biological processes, such as moisture and gas movement
throughthestomataofaleaf,arecovered(Chapter
3 ) to roundoutthe
mathematical and computational methods.
Modeling of livestocksystems probably does not have as long a history
as that of crop systems, because modeling livestock
on pasture involves
modeling the pasture as well. Four very different approaches are presented.
First, a fundamental method of modeling, using differential equations and
classical solutions for those equations, is well documented in Chapter 11.
Then, a dairy cow is simulated with the usual time steps and physiological
processes, but the feed supply does not include growth modeling of a crop
to grazing (Chapter
Chapter 13 shows how the detailed process of the
animal chewing on and consuming theforage can bemodeled and simulated
as part of the animal-pasture interaction. Chapter 10 uses the physiological
or process approach, with numerical integration, and models the growth of
the pasture and harvesting of the pasture by the animal (or
stopping this
grazing for a time) as interacting processes.
As the population of the world continues to grow, and with current
grainsurplusesbecomingalmostnonexistent,agriculturalproduction
has
become very important. Modeling and simulation of
food production systems will be increasingly used in the future, to help meet the global need
for food. We hope this bookwill help to spread the knowledge of agricultural
systems modeling and simulation to meet this task.
Robert M. Peart
R. Bruce Curry
Preface
Contributors
iii
ix
113
157
Steven J. Thomson
vii
viii
7. Evaporation Models
John L. Monteith
8. Simulation in Crop Management:
G O S S ~ M / C O ~
Harry E: Hodges, Frank D. FVhisler, Susan M. Bridges,
K. Raja Reddy, and James M. Mc~inion
9.
235
283
Neil L. Lecler
10. G W E : A Beef-Forage Model of Selective Grazing
Otto J. Loewer, Jr.
301
419
543
John C. Siemens
hole-Farm Simulation of Field Operations:
Object-Oriented Approach
Harbans Lal
16. Funda~entalsof Neural Networks
567
597
~ i l l i D.
a ~Batchelor
17. Object-Oriented Programming for Decision Systems
629
John Bolte
18. Simulation of Crop Growth: CROPGRO Model
Kenneth J. Boote, James W Jones,
and Cerrit Hoogenboom
651
Index
693
arrett Agricultural and Biological Engineering Department, Agricultural Research Service, U.S. Department of Agriculture, Purdue University, West Lafayette, Indiana.
atchelor AgriculturalandBiosystemsEngineering,Iowa
ioresource Engineering Department, OregonState University,
Corvallis, Oregon
State
e cur^ AgriculturalandBiologicalEngineeringDepartment,
University of Florida, Gainesville, Florida
ent Institute of Ecology and Resource Management, University
Edinburgh, Edinburgh, Scotland
of
ix
, Edinburgh, Scotland
arry F. Hodges Department of Plant and Soil Sciences, Mississippi State
niversity, Mississippi State, Mississippi
GerritHoogenboom
Department of BiologicalandAgriculturalEngineering, University of Georgia, Griffin, Georgia
James W. Jones AgriculturalandBiologicalEngineeringDepartment,
University of Florida, Gainesville, Florida
arbansLal
etteville, Arkansas
. McKinion
search
Service,
Mississippi
of
Contributors
xi
.C. Ward
AgResearch
Ruakura,
Hamilton,
New Zealand
Agricultural researchers have tried from the turn of the 20th century
to develop data and information bases that when interpreted can be
used to improve the management of agricultural systems. Early on,
researchers recorded and published descriptions that, for the most
part, were observations of situations involving abnormal and unusual
occurrences. With the invention of adding machines and calculators,
scientists began to define and determine relationships. With the advent of computers, engineers began to consider continuous and discrete happenings with respect to time. Beginning in the late 1950s,
descriptive and mathematical modeling of processes evolved. These
mimics were called simulations. Figure 1 is suggested, somewhat
with tongue in cheek.
In the rnid-l960s, the concept of systems dynamics emerged,
facilitating time-related representations of process flows and interactions. During the
systems dynamics became formalized. Rel
~nementscontinued through the 1980s, primarily in computer programming techniques, in verification, validation, and evaluation,
Simulation of dynamic agricultural and industrial systems has become an integral part of agricultural science. Increasedunderstanding
of ecosystem interactions,as influenced by the environment and management practices, has greatly expanded the potential for decision
support systems. In the mid-1990s we experienced the emergence of
the era of information technologies.
Bridging the gap between simulations that have been developed
to mimic and describe agricultural processes and procedures and the
use of these simulations to supply informa~onthat supports decisions being made by managers is difficult and challenging. Supplying
information to managers involves interfacing between computer output and the people who are the information users (see Fig. 2).
II.
A first point is that all decisions are made in a person's mind. Some
human makes each, any.and all decisions, either directly or indirectly.
Further, decisions are always based on whether to continue doing
things as they are being done or to change to a different procedure.
And, people who control the resources have the authority to make
the decisions. The icon of Fig. 3 emphasizes all this.
Knowledge
Engineering
Domain
Relationships
The mind, i n c l u ~ n gthe brain and sensory mechanisms for perception, functions to process and store symbolic expressions. An individual feels, perceives, thinks, remembers,and reasons in anadaptive
conscious and nonconscious manner. One way to think of this is that
the conscious mind is the interface between the source of symbols
that have been sensed and no~consciousmemory where they are
stored. Contained in a persons memory is everything that has been
observed and learned from all experiences, right or wrong, good or
bad. Certainly some things have become buried so deeply that they
are obscure. Stored in memory may be data, information, and/or
knowledge. Words are stored symbolically.
In resolving a problem, the contents of memory are reviewed,
perhaps intuitively and ordered according to importance, and alternatives are identified and automatically ranked. Risks are considered.
Short- and long-term consequences are explored, and choices are
made. Several matters may be thought through interactively and
simultaneousl~
When an individual is confronted with a problem to be solved,
several things happen. Some considerations occur almost instantly
and some are more protracted. What cognitive actions occur, the order inwhich they occur, and whether some of these actions recur are
a function of experience and knowledge of the domain where a decision is to be made. Most important, they are not a function of how
others see the problem, but of how the individual with the problem
sees it.
Mostly decisions are choices to continue a course of action orto
take a divergent course. Many preparatory dilemmas are resolved
over time before a decision is made to take action. Tpically when
an individual is confronted with a new situation or one where he or
she wants the action taken to be more effective, an attempt will be
made to review the situation. Thus, a fresh look may be taken to
reevaluate and reduce risks.
The blackboarding process that has been developed for computer information searching and problem solving is a mimic of what
in complex fashion goes on in a humans mind. That is, several problem areas may be addressed simultaneousl~The processing goes on
night and dayasleep or awake, consciouslyand nonconsciously until
the equations are complete or the situation is resolved to the point
that risk is adequately reduced and it is felt that action can be safely
taken. Then the resolution of the dilemma becomes evident. This may
umani~ationof Decision S u ~ ~ o ~
Benefits
FIGURE4 Decisionmaker.
So what does all this mean to decision support? Foremost, the job is
easier than may have been thought. Recognizing that decisions are
always made in a human's mind, support of the decision-making
process can.best be accomplished by supplying missing or incomplete
data or correcting inaccurate information about interactive critical areas in logic and achieving wording that is understandable and acceptable to the decision maker. This should point out the methodology that potentially can be used in any decision support process.
Simple? Not necessarily, because we are instilled through training andexperience to view problems from scientific perspectives,not
"from the applications-oriented viewpoints of people who manage
businesses. Scientists use the scientific method, which gives verified
and
Barrett
Nearing
and validated information with defined error, while others seek the
interpretation of information and data in changing situations. Scientists think in terms of probability, and decision-making managers look
to possibility.
A.
Excellent farmers who produce grain said that the following sources
are most used for needed information (Barrett and Jacobson, 1995):
Networking with other experts.
Suppliers of chemicals, equipment, seed, etc.
Extension and university specialists, trusted, with no profit ties.
Consultants, although rarely used.
Manufacturers technical departments.
Publications, trade and other.
Specification statements and sheets.
Computers for grain price and weather information.
Note that this list does not include social contacts.
B.
ManagerialTime
umanization
Supportof Decision
Computer Programs
~ n ~ o r ~ a t i o Information
n.
must be continuously searched for
with respect to prices, yield, policies, inte~ationalaffairs, tillage, variety performance, new technologies, etc. Networking
is useful here.
Cropeffects. Status of growth and development, local and regional; performance considering moisture and soilstatus;
vigor after emergence; stress factors; stability at harvest; shatter; size; stamina; etc.
Weedsuppression.
Existing and potential pressures related to
tillage; effectiveness of chemicals, spray s c h e d u ~ g etc.
, The
producers were observed to think of conventional tillage, reduced tillage, and no-till as variations of seedbed preparation
directly affecting potential weed pressures.
~ l a n t i n ~Varieties;
.
scheduled over
week period; when to
start with respect to harvest; tillage; labor; use of planter or
drill; etc.
~ e r ~ i l i t y Maintenance,
.
soil testing, scheduling, etc.
Harvest. Schedule over l + month period; preparation; storage;
labor; timing relative to planting schedule; etc.
Erosion. Environmental stewardship, sustainable agriculture,
sustainable production, conservation, compliance, etc.
Diseases,insects.
Scouting; chemical selection; spraying; etc.
O t h e re n t e ~ r i s e s .
All observed expert farmers maintain an
associated non-production-agriculture business enterprise
where resources are used interactively.
Labor, personal preferences, other crops, other factors.
Any information from simulations about interactions of the above
areas can potentially improve decision making.
r o ~ r a mto Supply l ~ f o r m a t i o ~
Humanization
Supportof Decision
11
information user, who is usually not the keyboard user. Relating the
outputs of these programs with theinformational needs of a manager
of production is not a straightforward task. The information being
sought relative to the critical areas must be extracted and phrased in
words and logic appropriate to the decision maker.
Early erosion simulations were empirical equations. Later, processbased models were developed. Information on rates of erosion related
to land management on a specified site continues to be objectively
sought.
A.
The first widely used erosion simulation was the Universal Soil Loss
Equation (USLE) (Wischmeier and Smith, 1978). This statistically derived regression model was developed to estimate average annual
soil lossfrom a hillslope.USLE does not address sediment deposition
on the toe slope of the hill, the amount of sediment leaving a field,
or the composition of the sediment eroded.
USLE has been the basis for the development and implementation of conservation practices throughout the world. It has been
applied to nearly every acre of cropland in the United States. The
primary reason for its success is its simplicity. The mathematics can
be done on a calculator, on the back of an envelope, or in the head.
The answer is also simple, being a single average annual soil loss
value for a hillslope. USLE is easily implemented and easily integrated with large and complex simulations and GIS applications.
evised Universal Soil loss Equation
wmanizati~nof Decision S w ~ ~ o r t
13
and
14
Barrett
Nearing
will vary among farms, with the technology used,and with the levels
of management abilities. The underlying ideas of precision farming
are solid and consistent with the concepts of good management. One
farmer of the grain production example said that he had been practicing prescription farming all his life. His only questions were: How
precise should he be, what size areas should he consider,and whether
and when he should change! He said that practicing precise, sitespecific farming was what made excellent management.
Weed Suppression Example
F e ~ i l i z a t i o nExample
15
DairyManagement
In similar thinking, dairy production can be improved by using simulations to evaluate the effect of changes in herds, both numbers and
quality of animals. This combination of cow and herd genetics, resource management, economics including sales methods, government
programs, and quality control is basic to good business management.
Report No. 10. West Lafayette, IN: USDA-ARS, National Soil Erosion Research Laboratory.
Jacobson, B. M., J. W. Jones, and S. M. Welch. 1995. Decisionsupport system
to assess agronomic, economic,
and environmental impacts of soybean and
corn management. ASAE Paper No. 95-2696. St. Joseph, MI: ASAE.
Nearing, M. A., G. R. Foster, L. J. Lane, and S. C. Finkner. 1989. A processbased soil erosionmodel for USDA-Water Erosion Prediction Project technology. Trans, ASAE 32:1587--1593.
Peart, R. M., and J. R. Barrett. 1979. The role of simulation. In: ~ o d ~ c a t i o n
of the Aerial E n v i ~ o n ~ ofe nPlants.
~
Monograph No. 2. St. Joseph,MI: ASAE,
pp. 467-480.
Renard, K. G., G. R. Foster, G. A. Weesies, D. K. McCool, and D. C. Yoder.
1996. P r e ~ i c ~ i nSoil
g Erosion by Water: A Guide to Conservation ~ l a n n i n gwith
the Revised ~niversalSoil Loss E ~ ~ a t i o n ( R ~ SUSDA
L E ) . Handbook No. 703.
Washin~on,DC: Government Printing Office.
Rockart, J. R. 1979. Chief executives define their own data needs. ~ a r v a ~ ~
~ u s i n e s Review.
s
March/ April, p. 81.
Wischmeier, W. H., and D. D. Smith. 1978. Predicting ~ i n ~ aErosion
ll
Losses:
A Guide to Conservation Planning. USDA Handbook No. 537. W a s h ~ ~ o n ,
DC: Government Printing Office.
ulation has proved to be a powerful tool in basic and applied biological sciences. The incredible growth and acceptance of personal computers during this same time period has made it possiblefor
simulation to become an integral part of biological researchin many
basic and applied fields. The purpose of this chapter is to present an
introduction to the concepts and techniques used in the simulation
of agricultural and biological systems with a few examplesthat demonstrate the approach.
Because of our own interest and experience in the simulation of
agricultural production systems (particularly crop systems), we focus
on the simulation of crop systems in this chapter and include examples related to plants, soil,and insects. Additional material on crop
s~mulationcan be found ina number of references, including Thornley and Johnson (1990), Leffelaar (1993),
Penning de Vries et al.(1989),
and Goudriaan and van Laar (1994). The conceptsand me tho do lo^
could have been presented equally well with examples on animal,
disease, or other biological systems (e.g.,see Curry and Feldman,
1987; Keen and Spain, 1990; France and Thornley, 1984; Odum, 1973,
and Dent and Blackie, 1979).
t
N
G
E
E
N
T
I
22
Jonesand ~ ~ y t e n
~ o r n ~ u t Sirnulation
er
Computer simulation includes the processes necessary foroperationalizing or solving a model to mimic real system behavior. Developing
computer logic and flow diagrams, writing the computer code, and
implementing the code on a computer to produce desired outputs are
necessary tasks in the simulation process. In practice, modeling and
computer simulation are usually interrelated. One should have in
mind techniques for i m ~ l e m e n t i nthe
~ mathematical model as it is
being conceptualized and developed. Some authors define simulation
to include the modeling process (Pritsker,1995).In this chapter, "simulation of biological systems" refers to the overall procedures outlined below, including mathematical model development and computer simulation to produce numerical model results for analysis.
iological Processes
23
State Varia~les
The interrelationships between components in a system, and therefore between state variables in the system, exist because of various
processes. We sometimes use the term "process-oriented, to describe
models that describe the flow and storage of mass, energy, or other
substances. For example, the crop biomass state variable changes as
a result of photosynthesis and respiration processes, and the soil water state variable changes as a result of rainfall, runoff, percolation,
and evapotranspiration processes. A crop model is the set of mathematical relations~psthat describe the changes in state variables as a
result of the various processes that occur.
C o ~ ~ i ~ u omodels
us
are characterized by state variables that can
change smoothly over small time intervals and are not restricted to
integer values. Discrete models, on the other hand, are those in which
the variables describing the system states take on integer values. Crop
models are usually classified as continuous models. An example of a
discrete biological model is one that predicts the birth and death of
individual insects. The number of insects alive at any point in time
is a whole number. Discrete models usually require information on
the times required to complete activities, such as the time required
for an insect egg to hatch. In contrast, continuous models require
~fo~matio
onn processes such as the flow rate of material or energy
between components and between components and theenvironment.
Continuous models are usually represented by a set of differential or difference equations derived from the structure of the system
and the interrelationships among components. Some systems can be
modeled as either continuous or discrete, depending on the purpose
of the model. For example, the fate ~ i r t hdeath)
,
of individual insects
may be viewed as a series of discrete events. The population number
at any time would be a count of the total number of individuals, each
considered separately. Obviously, if the population is large, a lot of
computer time is required to keep up with each insect. One may also
view the insect population density ( n u m ~ eof
r insects per unit area)
as a continuous state variable for large numbers and when population
rocesses (births, deaths) occur smoothly overtime, and thus use
differential equations to model their dynamics.
Validation is the process of comparing simulated results to real system data not previously used in any calibration or parameter estimation process.The purpose of validation is to determine if the
model is sufficiently accurate for its application as defined by objectives of the simulation study. Simulated state variables are compared
with measured values of state variables. Usually in crop simulation
studies only a few state variables out of many possibilities are measured, and thus a complete comparison is usually not possible. Validation involves subjective judgment. First, the areas for comparison
must be selected. Then a measure of "accuracy" or "closeness of fit"
must be established, such as the final crop yield. The choiceof criteria
and important state variables for comparison should be based on
model objectives. Validation effortsare essential in the application of
crop models.
.
Simulation involves the development of a model and its use in gaining insight about the system itself or its management. The Simulation
approach usually entails the use of a number of steps, which are
summarized below. Following this summary, techniques for biological model development and implementation are presented along with
an example of a simplified crop model.
i ~ ~ o g i cProcesses
a~
27
~ i t e r a t ~ rReview
e
and
ships. The mathematical representation of the system should be developed, including specific functions and relationships to be used in
the model. Experiments may be
needed to develop functions and
estimate parameters for the model. Themodel is translated into computer code for simulating the behavior of the real system. Computer
flowcharts are usually helpful in translating the model to computer
code and documenting the relationships between the model and the
code. Techniques formodel development are presented in Section IV.
mine whether the model behavior is differentfromrealsystem behavior to some stated level of significance. Theprecision of such tests should not overshadow the conceptual
problems in applying them. There is no mechanical procedure
that permits the conditions for acceptablevalidation by a
gle (or even a series of) statistical comparison(s)of the modelperformance against some recorded or measured data from
an existing system which the model represents. Any statistical
tests that may be carried out and which are positive in favor
of the model add to the model-builder's confidence that his
creation is functioning well.Theacceptablelevel
of confidence is achieved normally by a series of assessments and
subsequent modifications until such time as the model is to
be applied in support of decision making.
.
Continuous system modeling is a process-oriented approach for describing the behavior of a system. Processes fall into three categories:
transport (or flow), transformation, and storage (Smerage, 1977).
Jonesand L ~ ~ e n
iological Processes
31
32
from the system into a sink in the environment has no effect on the
environment. The rate symbol represents flow from onecompartment
to another, from a source to a compartment, or from a compartment
to a sink, Auxiliary variables are shown by circles and represent factors (inputs and parameters) that influence the rates. For example,
temperature may affect the rate of biomass accumulation in a crop
growth model and would be classified in this diagram as an auxiliary
variable. Solid lines represent pathways for material flow whereas
dashed lines represent information flow.
After the compartment model is constructed, one can write the
mathematical model representing the system using structural constraints and component descriptions. The Appendix lists variables
used in this chapter along with their ~efinitions and
units. From the
diagram itself the structural constraint or continuity can be applied
to each compartment level, or
33
where
For example, consider the simple water flow system in Fig. 3a. The
volumes of water in each tank are the integrated extensive variables
of interest in the system. Equations describing this system are
FIGURE3 (a) Schematic of a two-tank water flowsystem; (b) the corresponding compartment model using Forrester (1961) notation; and (c) example results.
parameters shown in the figure. When the input to the first tank
was set to 5 m3/h for the first 5 h of simulation, then set to 0.0 for
the remaining time, the volume of water in the first tank first increased to 16.9 m3, then droppedto 0.0 after 19.1h.Volume in
the second tank lagged behind that in the first tank, peaking at only
7.1 m3.
This example demonstrates the simplicity with which this approach can beused to obtain a first-order differential equation model
of a system, Agricultural examples of compartment modeling are insect population dynamics and growth of crops. Insome such systems,
the mechanisms causing flow are not well enough understood or are
so complex that research may be required to determine relationships
between flow and other system characteristics and inputs.
In general, a model developed in this way will have state variables defined by levels xl,x,, . . . , x,. The mathematical model will
be of the form
35
Sim~~ation
of Biological Processes
20
15
10
0
0
15
10
HOURS
FIGURE3 Continued
20
25
-=Ax+
dt
37
Layer n - l
Layer n
38
We can equate this expression with the right-hand side of Eq. (7) and
expand it into
iological Processes
0 and as Ax
0 results in
a8 a
-"
Daxax at
where 8 is a function of both x and t. Equation (13) is the partial
differential equation for water flow in one dimension through a soil
of unit area, the same equation presented by Gardner (1959). The firstorder differential equation system [Eq. (7)] is an approximate
"lumped" representation of the distributed (one-dimension)soil system more accurately represented by the partial differential equation,
Eq. (13). When approximate solutions are adequate, the lumped representation provides a relatively quick and simple method for simulating the behavior of distributed systems.
The same procedure has been used to lump age structures in
population dynamics models. Figure 5 shows the Forrester diagram
for an insect model with the population divided into age classes to
simulate age structure. The levels are population numbers for individuals in specific age classes. Anage class maybe defined as a range
of ages, for example 0-10, 10-20, 20-30, and >30 represent four age
classes of insects. Flowbetween age classes occursas insects age. The
partial differential equation describing this system is
aN(a, t ) aN(a, t )
+= -u(a, t)N(a, t )
at
aa
(14)
where
N(a, t ) = population of insects age a at time t, dimensionless
u(a, t ) = net flux of insects age a at time t across the system
boundary, S-*
I
l
l
I
l
1
I
I
l
I
I
I
l
I
I
l
I
I
I
I
I
l
l
I
I
and
&+At
xt
+ (axt + b)At
(16)
= xk
+ Rate, At
(17)
where Rate, is any expression for the rate of change of the state variable x at time t. In the example above, Rate, = axt + b.
Equation (16) is a finite-difference expression of the differential
equation. In general, Eq. (16) is also referred to as the rectangular or
Euler method of integration of a first-order differential equation. The
value of x is known at time t and possibly prior to time t. The model
expresses the instantaneous rate of change of the variable x. Although
we do not know the value of x at t + At, we can choose a small
and predict a value of x at time t + At based on the current rate of
change of x. This procedure is repeated as long as desired. For example, by continuing with x at t + At and projecting forward by one
more At we can estimate x for time t + M t .
The result will be a time series behavior of x defined by the
model. If the system has more than one first-order differential equation, the finite difference procedure can be applied to each. The rate
expression for any one differential equation may be functionally dependent on any of the state variables as well as on inputs to the
system. This method is dependent on the system model being described by first-order differential equations (or difference equations
as described earlier). Second-order differential equations can be simulated by transformation to two first-order equations. For example,
the second-order equation
and
dx2
dt
-=
ax2 - x1 + u(t)
is the formula for calculating yf+At. Note that Rate:+t is on the righthand side of Eq. (20). We have to "predict" Rate:+,, since we do not
h o w y or Rate at time t + At. The Euler method is used to predict
is used to compute
as a first approximation. Then that Y:+,~
Rate;',,, using the first-order differential equation model. We then
have all that is needed to compute Y f + A t above. The Euler and trapezoid methods are ex~Zicitmethods; all terms on the right-hand side
of the equation are h o w n or can be computed from what is known.
In general, a Taylor's series expansion about yt can be used to
express y;",,. This expression is
where
represents the nth derivative of y with respect to t and n!
represents factorial numbers (i.e., n! = 1 2
4 5
- n).
If we approximate yt by using the first three terms of Eq. (2l), then
where c t is the total truncation error, or the error caused by not including all the terms in the Taylor expansion.
Since
can be approximated by
- y',")/ At, Eq. (22) can
be written
Yt+At
= Yt
Et
Equation (23) is equivalent to the trapezoid predictor-corrector technique with the added error term. The Euler technique is equivalent
to the Taylor series expansion with only the first derivative term included explicitly.Thus, the trapezoid method is a more accurate
method for simulating first-order differential equations. However, the
Euler method is used for most models because of its simplicity and
ease of use.
iological Processes
43
Because that is a numerical approximation, it is subject to numerical errors. The local truncation error using the Euler method is
[(At)z/2][dzy(~)/dtz],
where 6 is some point in the interval [t, t + At].
The truncation error is the error with each integration step and
can thus accumulate. Truncation error for the trapezoid predictorwhich, for a small At, is less
corrector method is [(At)3/6][d3y(~)/dt3],
than that of Euler method (Ortega and Poole, 1981). Generally the
higher order methods are more stable than lower order methods.
Relative sensitivity a,(y I is often used to provide a normalked measure for comparing the sensitivity of a model to several variables. Relative sensitivity is defined by
.
There are many options for computer implementation of biological
models today. Becausethere are many reasons for developing biological models, some consideration should be given to which option to
choose for a particular model. Some biologicalmodels are developed
by an individual or small team of scientists to test a hypothesis, and
there is no attempt to make these models available to others except
These new languages have potential for helping improve the modularity of biological models.
There are a number of specialized computer simulation languages that allow model developers with little or on computer programming skills to implement models. Continuous simulation languages solve systems of continuous differential equations. For
example, the SLAM system (Pritsker, 1995) and FSE (van Kraalingen,
1995) packages provide modules that perform most of the required
computer simulation tasks. Users only have to "program" their specific model equations and provide parameters and other inputs in
order to obtain simulated behavior of a system. Probably the simplest
and easiest to use simulation software allows users to build conceptual models on computer screens using icons, such as Forrester symbols presented earlier in this chapter. Two widely used packages are
Stella (High Performance Systems, 1990) and VisSim(VisualSolutions,1990), both available for use on personal. computers. These
packages guide users in linking together different icons to depict a
system with its storage compartments and flow paths, then ask for
needed parameters and other inputs to run the
model, They willthen
simulate the system and graph or print results for users; no programming is needed. These packages have the distinct advantage of rapid
evaluation of variations in model structure to test different hypotheses about the system. Each computer language has its advantages
and disadvantages. The best choice for a particular model will depend on many factors such as model complexity, intended users of
the model, how it will be used, the need to integrate it with other
software and databases, and the need to maintain it.
Dynamic growth and yield models have been developed for a wide
range of crops, including cotton, grain sorghum, wheat, corn, soybeans, alfalfa, and others (Jones and Ritchie, 1990). Researchershave
used these models for a number of purposes, such as irrigation management (Boggess et al., 1981; Swaney et al., 1983), nutrient management (Keating et al.,1993; Singh et al. 1993), pest management
(Pinnschmidt et al., 1990; Batchelor et al., 1993), land use planning
(Beinroth et al., 1997),crop sequencing (Thornton et al., 1995),climate
change assessment (Curry et al., 1995; Rosenzweig et al., 1995; Penning de Vries, 1993), and yield forecasting (Swaney et al., 1983). These
applications have demonstrated conclusively the value of simulation
applied to cropping systems. However, these models have limitations
S i m ~ ~ a t of
i o Biological
~
Processes
47
because they do not include all factors that occur in reality and
contain empiricism that may require calibration and testing for sitespecific applications. Research directed to improving these shortcomings and limitations is needed to enhance the applicability of the crop
models.
A useful characterization of crop models for application at the
field scalewas attributed to C. T. de Wit (Bouman et al., 1996;Penning
de Vries and van Laar, 1982). Initially, four levels were defined, but
these were later simplified to three levels (Lovenstein et al., 1993).
First is the potential yield level; crop growth is dependent onweather
factors, such as temperature, solar radiation, CO, concentration, and
day length, and on genetic factors of the crop. This level of model
includes basic crop growth processes such as photosynthesis, respiration, tissue growth, and development. Its main use is to gain an
understanding of how these factors affect potential production of a
crop, assuming that water and fertilizers are adequate and that no
pest damage occurs. A second level of detail is characterized as attainable yield; at this level, the model has components that describe
water and nitrogen fertilizer limitations to production. Attainable
yield thus depends on water and nitrogen supplied by management
as well as by rainfall and soil organic matter. Many existing crop
models have the capability to describe crop growth and yield with
these assumptions (Jones, 1993).The third level of detail is actual
yield level, which attempts to explain all the factors that influence
growth in a field, including pest damage and mismanagement. This
characterization of levels of detail has been useful in providing both
a practical pathway for model development and a conceptual framework for model applications. It has also served to determine data
requirements for model development, taking into consideration the
need for more comprehensive data sets as the level of model detail
increases.
There are many choices of state variables for crop models. Simple models may have only one or two state variables, such as leaf
area and dry weight, whereas others may have hundreds of state
variables describing leaf area, leaf weight, number of leaves, stem
weight, number of fruit, and weight of fruit for various age classes
(Jones et al., 1991). Crop models with many state variables provide
more details about crop behavior, including lags in fruit growth and
size distributions, but require much more time to develop and test
and more time to solve in a computer. For some problems, models
with two to four state variables provide adequate predictions of crop
responses. In this chapter, a simple crop model at the potential yield
~ l l o t ~ ~ t h ~ l s
param~e~
h h t (I.)
temperature (7"
.""""""":""""""""""-"-
""""
IGURE 6 Forrester diagram for the crop model with three state variables
at the potential productionlevel of detail.
- = E(P,
dt
R,W)
where
51
where
I, = Imsin{Zn[( t , -
/ 241)
(31)
52
where I,nis the maximum light flux density (at noon) and th is the
solar time in hours,
Equation (29)contains environmental variables (T, C, Io),various
parameters a,K, m, D), and thecanopy leaf area, L, which depends
on time. Leaf area ratio (square meters of leaf area per gram of plant)
and specific leaf area, (m2leaf)/ (g leaf), vary considerably with env~onmentalconditions. Under cool temperatures, leaves appear more
slowly [Eq. (26)], accumulate more dry weight during the extended
growth period, but expand to about the same final area over a practical temperature range, as shown by Jones et al. (1991) for tomato.
Under low light, leaves will tend to be thinner because of lower rates
of photosynt~esisper unit of development. The assumption that L is
a function of N provides a model that mimics observed leaf area ratio
responses to light, temperature, and CO,. The expolinear equation
(~oudriaanand ~ o n t e i t h ,1990) was used to fit leaf area vs. node
number using data reported by Jones et al. (1991) for tomatoes grown
at two CO, levels and three temperatures, The equation is
L =
where
P ) W + exp[P(N - %)l>
(32)
dN
-= rn,r(T)
dt
(33)
p = 0.38
= 0.074
+h
$1
=I
30.0
= 5.0
= 0.0664
= 4.0
k , = 0.0006
m = 0.10
n b = 13.3
T
r, = 0.021
C = 350
D = 0.108
E = 0.70
fc(N) = 0.85
1, = 11200
K = 0.58
There are limitations in the use of this model. Itdoes not account
for senescence or picking of plant material, nor does it partition dry
matter into different plant components, such as fruit. The model
could be extended to describe the growth of fruit and other organs,
thereby creating more state variables, equations, and a need for more
coefficients and relationships.
Figures 7a and
show the effects of temperature, light, and CO,
concentration on the rates of photosynthesis predicted by Eq. (29)and
the effects of temperature on respiration for a canopy with a leaf area
index ( L ) of 4.0. These results demonstrate the importance of weather
on the rate of dry matter accumula~on.Figure 7c shows the effect of
temperature on leaf area expansion over time. Low temperature reduces the rate of node formation [Eq. (26)], which results in delays
in leaf area development. When plants are young, rapid leaf area
expansion is important to create a h11 canopy for capturin~light for
photosynthesis.
The three-state variable model, Eq. (33), was simulated for 80
days under assumed constant environmental conditions to demonstrate the integrated and cumulative effects of growth and de
ment rates on yield. For each set of ~omputations,all variab
one were held at reference values and one variable was chan
successive runs. Reference conditions were 30/20"C day/ night temperatures, 12 h days, 1200 (pmol photons)/ (m2 S) maximum light
flux density, and 350 (pmol CO,) / (mol air).
Figure 7d de~onstratesa major effect of temperature on
weight over the 80 day simulations. Temperature affects both development and photos~thesis,but its major effect is on development.
For the l8 / 8C temperature case, 59days were required to reach node
l 0 in contrast at 23 days for the 30/20"C te~peraturecase. ~hotosynthesis for the 18/8"C case was reduced by only about 30%. For this
500
l000
1500
2000
LIGHT FLUX DENSITY(~mol(photon)/mA2-s)
F I ~ U R E7 Simulated results from the three state variable crop model. (a)
Photosynthesis VS. maximum light flux (I,) at four CO, levels at temperature
ln =: 30C; (b) photosynthesis vs temperature at four maximum light flux
levels for a canopy with leaf area index L = 4.0; (c) leaf area development
vs time for four temperatures; (d) total dry weight vs. time for four day/
night temperatures; and (e) total dryweight vs. time for fourlight levels at
3O/2O0C day /night temperatures.
coldest case, dry matter at $0 days was lower by a factor of 15. Note
that growth was lower at 38/28"C thanat 30/20"C day/night
temperatures.
As light was varied from 400 to 1600 (pm01 photons)/ (m' S),
dry weight yield increased by a factor of2.5 at 80 days (Fig. 7e).
Numbers leaves and leaf area were not affected by light, Increasing
GO, levels from an ambient level 350 (pmol CO,) / (mol air) to 700
resulted in a 28% increase in dry weight yield at 80 days (data not
shown).
LES
Ai
Ci
D
55
-l
IGURE 7
10
10
15
20
25
TE~PE~TURE
(c)
,
20
Continued
30
40
DAYS
50
30
60
35
70
40
80
10
20
30
40
DAYS
50
60
l
LIGHT FLUX
10
IGURE
20
7 Continued
30
40
DAYS
50
60
70
80
57
volume in tank i, m3
volume of water stored in layer i, m3
level of ith state variable, various units
volumetric water content, m3/ m3
absolute sensitivity of output y to input k, dimensionless
relative sensitivity of output y to input k, dimensionless
assimilates, (g CH20)/ m
CO, concentration of the air, (pm01 CO,)! (mol air)
photosynthesis conversion
coefficient,
(g C
CO,)
conversion efficiency at CH20 to plant tissue, (g tissue)/(g
CHZO)
fraction of total crop growth partitioned to canopy dimensionless
actual light flux density at top of canopy (pm01 photons)/
(mZ S)
maximum light flux density at top of canopy (pm01 photons)/ (m2 S)
respiration rate at 25"C, (g CH20)/ [(g tissue) h]
canopy light extinction coefficient, dimensionless
canopy leaf area index, (mZleaf) / (m2ground)
light transmission coefficient of leaves, dimensionless
empirical coefficient for expolinear equation, ~imensionless
leaf number or node number, dimensionless
photosynthesis reduction factor, dependent on T, dimensionless
canopy gross photosynthesis rate, (g C
rate of development, h-'
maximum rate of leaf appearance, h-'
+
Qdum, H. T. 1973. An energy circuit language for ecological and social systems. In Systems Analysis and Simulation in Ecology. Vol.11. B. C. Patten
(Ed.). New York Academic Press, pp. 140-211.
Qrtega, J. M., and W. G. Poole, Jr. 1981.An Introduction to ~umericalet hods
for Di~erential E~uations.
Marshfield, MA: Pitman.
Patten, B. C. 1973. Systems Analysis and Si~ulation inEcology. Vol.11. New
York Academic Press.
Penning de Vries, F. W.T. 1993.Rice production and climate change. In:
Systems Approaches for Agricultural Developme~t.F. Penning de Vries, P.
Teng, and K. Metselaar (Eds.). Dordrecht, The Netherlands: Kluwer Academic, pp. 175-189.
Penning de Vries, F. W.T., and H. H. van Laar (Eds.). 1982. Simulation of
Plant Growth and Crop Production. Wageningen, The Netherlands: PUDQC.
Penning de Vkies, F. W. T., D. M. Jansen, H. F. M. ten Berge, and A. Bakema.
1989. Simulation ofEcophysiologicalProcesses
of Growth in Several Annual
Crops. Simulation Monographs 29. Wageningen, TheNetherlands: PUDQC.
Pinnschmidt, H., P. S. Teng, and J. E. Yuen. 1990. Pest effectson crop growth
and yield. InProceedings of the ~ o r ~ s h on
o pMode~ingPest-Crop Interactions.
P. Teng and H. Yuen (Eds.). Honolulu: University of Hawaii.
Pritsker, A. A. B. 1995. Introduction to Simulation and S L A M II. 4th ed, New
York Wiley.
Rabbinge, R., P. A. Leffelaar, and H. C. van Latesteijn.1994.Therole
of
systems analysis as an instrument in policy making and resource management. In Opportunities, Use, and Transfer of Systems Research Met~ods in Agriculture to Developing Countries. P. R. Goldsworthy and F. W. T. Penning
de Vries (Eds.). Dordrecht, The Netherlands: Kluwer Academic, pp. 6779.
Ritchie, J. T. 1995. International Consortium for Agricultural Systems Applications (ICASA): Establishment and purpose, Agric. Syst. 49:329-335.
Rosenzweig, C., J. W. Jones, G. Y. Tsuji, and P. Hildebrand (Eds.). Climate
Change and A ~ i c u l t u r eAnalysis
:
of Potential International Impacts. ASA Spec.
Publ. No. 59. Madison, WI: American Society of Agronomy.
Sinclair, T.R. 1990. Nitrogen influence on the physiology of crop yield. In:
heo ore tical Production Ecology: Reflections andProspects.
R. Rabbinge, J.
Goudriaan, H. van Keulen, F. W.T. Penning de Vries, and H. 3. van Laar
(Eds.).Simulation Monographs 34.Wageningen,The
Netherlands: PUDOC, pp. 41-55.
Singh, U,, P. K. Thornton, A. R. Saka, and J. B. Dent. 1993. Maize modeling
in Malawi: A tool for soil fertility research and development. In: Systems
Approaches for Agricultural Development. F. Penning de Vries, P. Teng, and
K. Metselaar (Eds.). Dordrecht, The Netherlands: Kluwer Academic, pp,
253-273.
Smerage, G. H. 1977. On the extension ofcomplementary variables modeling
theory to biology and ecology. In:Problem Analysis in Science and Engineering. F. H. Branin, Jr. and K. Huseyin (Eds.). New York Academic Press,
pp. 427-462.
62
Jonesand Luyten
.
The process of creating new knowledge involves the exploration of a
focus question or puzzlement. Figure l depicts this process (Novak
and Gowan, 1984). Guided by the focus question phenomena of interest are examined. Beginning at the base of Gowins Knowledge
Vee certain events or objects are observed. eaning is attached to
these events or objects by an intimate interplay of conceptual (or
t h i n ~ g and
) methodological (or doing) considerations, as depicted
by the two sides of the Vee.
Consider first the methodol~gicalconsiderations (on the right
side). Records of the events or objects are made. The facts are then
ordered or otherwise organized (transformations). The results are
then presented as data using tables, graphs, etc. This is followed by
interpretations, explanations and genera~ations. These lead to
knowledge claims about what was found. The final step consists of
value claims about the meaning or significance of the knowledge
claims.
On the conceptual side of the process, ~ e g i n n i n g
at the top-left
of the Tee); going from the most general, i.e., World View
(that
nature is orderly and knowable) to Philosophies, to Theories (logically related concepts) to Principles (derived from prior knowledge
claims) to Constructs and Conceptual Structures to Concepts (signifying regularities). During the creation of new knowledge there is an
active interplay between the conceptual and methodological sides of
nowledge Claims:
New
generalizations, in answer to the
telling questions, produced in the
context of inquiry according to
Toulmin)
Inte~retations,Explanations,
~ e n e ~ i i z a t i o n sProduct
:
of
nnciples:
methodologyprior
and
knowledge used
warrent of claims.
esults:
ransformations:
or objects.
theory
acts:judgment,
The
the inquiry
Events/Objects:
Phenomena of interest apprehended
through concepts and record-marking:
occurrences, objects
IGURE
Thevee heuristic with descriptions of the interacting elements
involved in the construction or analysis of knowledge in any discipline. Although all elements are involved in any coherent research program, the major sources of difficulty in individual inquiryusually begin at the bottom of
the vee, where concepts, events/objects, and records must be scrutinized.
(From Novak and Gowin, 1984, Reprinted with permission of Cambridge
University Press, New York, New York.)
the Vee. Novak and Gowin (1984) provide numerous specific illustrative examples from a range of subjects.
In a nutshell,this is the process we use to create new knowledge.
The process is so familiar that we often take it for granted.
being conscious of these steps makes the modeling process clearer; it
is fundamentally an artistic process in which there is no single right
CQQke
Scientific inquiry progresses more rapidly when there is a rich interplay between the conceptual and experimental at each step in the
process of creating knowledge rather than relying upon only one or
the other of these legs of the vee. However, the literature is replete
with disciplines in which the theoretical and experimental aspects
have matured at d~amaticallyd ~ e r e nrates.
t
Often, however, we as individual scientists rely more heavily
upon either one or the other, depending upon our training. For example, Fry (1941) contrasted the habits of thought of mathematicians
and engineers:
[l] Greaterconfidence in: Thetypical mathematician
feels greater confidence in a conclusion reached by careful
reasoning and is not convinced to the same degree by experimental evidence.
An engineer has less interest in pure forms which
have no apparent connection to the world of physical reality;
but may have great interest in such useful information as a
table of hardness which may be totally unrelated to any theory and which the typical mathematician would find quite
boring.
[Z] Severe criticismof details: A mathematician is highly
critical towards the details of a demonstration. For almostany
other class of people, an argument may be good enough, even
though some minor question remains open. For a mathematician an a r ~ m e nist either perfect in every detail, in form as
well as substance, or elseit is wrong. The mathematician calls
this rigorous thinking and it is necessary for work to have
permanent value; the engineer calls it hair-splitting and says
if indulged in would never get anything done.
[3] Idealization: Themathematician tends toidealize any
situation-gases are ideal, conductors, gases perfect, surfaces smooth. What the mathematician calls getting down
to essentials, the engineer is likely to dub somewhat contemptuously as ignoring the facts.
141 Generality: When confronted with solving
x3-1=0
Using ~ a t ~ e m a t i cas
s a roblem-Solving Tool
0.
ELS
At the interface region between engineering and biology there persists the lingering and mistaken belief that biologically based problems are inherently so complex that they simply are not amenable to
mathematical treatment. One might imagine that this same attitude
could have been applied to quantum mechanics, were it not for the
astonishing success achieved using mathematics as a central tool.
Experimentalists, especiallythose who ordinarily work withbiological problems, are sometimes inclined to be suspicious of mathematical work. Although a healthy skepticism is highly desirable, the
role of mathematical modeling has many times been unfairly criticized (but certainly not in all cases). A particularly common difficulty
is the failure to note the necessity and desirability of making assump-
tions-in both mathematical and experimental studies. In both instances attention is directed to a limited problem and theconclusions
reached are valid only within the constraints of "the experimental
assumptions." ath he ma tical models (Noble, 1967) can be used to
cause attention to be directed to the most hportant questions in an
experimental study. If data agree with a theory, the level of confidence
in the predictive power of the model is enhanced. In other instances,
mathematics can be used to reduce or eliminate certain experimentation.
Nahikian (1964) graphically depicted the successive approximations associated with a mathematical model (see Fig.2). ~ e ~ n n ~
with a clear statement of a "problem," the most important properties
of a physical or biological problem to be studied are abstracted and
isolated by making judicious use of assumptions (left side of Pig. 2).
A mathematical problem, using basic physicalor chemical principles,
is developed to embody the "essence" of the problem. The "mathematical crank" (bottom) is turned and checked. The implications of
the model are articulated. These implications are then related to the
original problem through the observed phenomena. Discrepancies are
noted, and the cycle begins again. However, notice that the arrows
are bidirectional throughout, indicating that the logical progression
may reverse at any point.
The advent of the computer is having a great liberating effect
on the use of mathematics. There is no longer the overriding need to
formulate problems to conform to the limited tools of classical anal-
r o ~ ~ e ~ - S o lTool
vin~
ysis. Onthe other hand, when using a computer we still must abstract
and idealize a given engineering problem. We still must break the
original problem into its elements, decide what is significant, specify
the data we must start from, what the steps of the analysis are, and
what end results are required. These steps require the habit of
thought of the mathematician.
Cooke
Using ~ a t ~ e m a t i c s a ro~lem-Solving
71
the temperature varies with distance into the soil. Recognizing analogies can greatly strengthen ones confidence in a solution.
The greater the insight into the underlying processes you can
muster at any stage of development, the more likely an acceptable
balance between simplicity and complexity canbe achieved. The crucial point to remember is that you are using a mathematical model
to gain insight and that no single representation or model serves all
purposes. A novice modeler is tempted to beg for a universal representation that fits all situations, but alas, if that is your goal, you are
likely to trade an obscuring complexity for the clarity that often can
be achieved more readily through simplicity.
Bowen (1966) described the basic issues to be considered in formulating a problem (Fig. 4). Before a problem can be properly defined, the decision maker must be identified. Usually the decision
maker is both the person or group whose objectives are not being
met and the one who controls the resources needed to effect a solution. Next the decision makers objective and resources must be clarified and weighted. Usually objectives conflict and cannot be simultaneously satisfied. Furthermore, the environment with its associated
constraints must be specifiedand understood. A solution t ~ignores
~ t
Resources
Selection of Decision
Maker
I
1
ProblemDefine
problem.
these const~aintsis not a solution, Only after all three of these issues
73
After your solution has been implemented, there is a responsibility to determine whether your solution works. This usually involves some form of experimentation and an evaluation of whether
you have met the objectives previously articulated.
Although numerous texts on mathematical modeling have been published, a very old text by Ver Planck and Teare (1954) serves well as
the focal point for this discussion. Table l summarizes the approach
presented in their book, but Ive also made some adaptation. These
are the time-honored steps that engineers routinely use in problem
solving. Because the steps are so familiar, their significance is often
overlooked.
Explicitly identifying these steps makes the process more comprehensible and more than just an art form. The steps are (A) define
the problem, (B) plan its treatment, (C) execute the steps, (D) check
your work, and (E) learn and generalize from your analysis. Lets
consider each of the steps in this process.
r o b l e ~ - S o l v Tool
i~~
75
77
Cooke
and to help you gain insight), so lets consider some specific techniques, including some that can beapplied rather easily as your work
progresses.
Kinds
Checking
apples = 11 apples
+ 1 cow =
apples
79
A (baskets)
The units for each term are treated algebraically. Contemporary symbolic mathematical software can be used to automate
this task, as well as handle the mundane but error-prone task
of conversion of units.
There are a few special situations to be considered. Numerical constants in an equation may represent pure numbers
(e.g.,
but may also be dependent upon a particular set of
units. For this reason, symbolic parameters are usually preferred because their units canbechecked
more easily if
treated on a basis symmetrical with the variables. For example, if the variable in the quadratic equation has the units
of meters, then clearly the parameters b, and c must have
units too.
The arguments of certain special functions, such as the
trigonometric functions, exponential functions, and Bessel
functions, are pure numbers and must be dimensionless. Beware that using units of measure such as degrees rather than
radians can createan apparent discrepancy. Similarly because
the logarithm of a product can be separated, an apparent discrepancy may exist. Sometimesa physical law must be introduced to complete the conversion process to demonstrate dimensional homogeneity.
This is similar to units check, but each factor in the equaiton is replaced by its dimensional formula in terms of primary quantities such
as mass (M), length (L), and time (T). Unlike the units check, this
does not verify consistency of system units. Ver Planck and Teare (p.
illustrate the procedure:
Ast40,
Symmetry as a Check
Symmetry also provides a powerful means for checking your derivations. Checking a result for even and odd functions is easy and
often helpful. For example, if a variable has a positive direction, what
result is expected if the sign of that variable is reversed? Will the sign
of the expression be the same or reversed?
Ver Planck and Teare (p. 242) provide two examples of using symmetry to identify potential errors.
T =hA
L [~1 - exp
(-~)]
Ast-0,
T-0
Ast-+a,
T+- 4
hA
IGURE
Circuit diagram.
82
Cooke
where
P = load (lb)
= deflection at the load (in.)
E = Young's modulus for the leaf material (lb in.-')
b = width of leaves (in.)
n = number of leaves
h = thickness of each leaf (in.)
Zl, E,E, = lengths (in.) as shown
I
I
11
Using ~ a t h e ~ ~ tas
i cas~ro~lem-Solviflg
Tool
Let f = / and 1 - f = /
would read
An analysis should always be concluded with a summary of important findings, with particular emphasis on the limitations imposed by
the assumptions. Of course, all important assumptions should be
made explicit rather than implicit. Mathematical equations are not
useful to the engineer until they have been interpreted in terms of
the original problem. You must verbally state the meaning of the
appropriate mathematical equations. If you cannot state their meaning, quite lilcely you do not yet understand them.
The analysis should usually but not always, involve numerical
testing of the results. This will clarify the meaning and the feasibility
of the particular problem. Thisprovides a sense of scale, alerting you
to gross blunders or possibly to your lack of success in finding a
solution. Finally the ident~ableuncertainties that still exist should
be clearly stated. Mistakes follow if you carelessly exceed the conditions on which your analysis is based. The analysis should not be
concluded until the possibility of simplifymg and/or generalizing the
analysis has been considered, explicitly and individually.
Consider now a specific analysis in order to make the foregoing
discussion more concrete. This alsoillustrates how a simple analysis
can make an expensive experiment unnecessary.
disc on ti nu it^
at~ematicsas
~ r o ~ l e ~ - S Q lTool
vi~g
FIGURE7 Seed-soilschematic.
Since the soil temperature at large distances from the center of the
seed will not be affected appreciably by the heat generated by the
seed within the soil (which would be at uniform temperature if the
seed were not present), that temperature may betaken as a reference.
In other words, all other temperatures may be calculated with
lim vu2= 0
r-m
where
V
= scaled
= diffusivity, cm2/s
= heat generation rate, cal/ (S cm3)
K = thermal conductivity, cal/ [S cm2(OC/ cm)]
and
K
Cooke
[1?1a (P
"> -1
d2v
d e sin2@
Due to the symmetry in 0 and
Assumptions 6 and 7, Eq. (1)may
be applied to the seed and soil separately:
~v
= --p
sin 0 80 (sin 0
+,
[?,-(..~)]
"p
l a
0,
Y >
The conditions required for solution of Eqs. (4) and (5) are as
follows. The temperature at great distances from the seed must approach the reference value of zero.
lim v2(r) = 0
(6)
where v, is not a temperature but a temperature difference-Celsius
degrees from the reference temperature.
In order for the boundary between the seed and soil not to increase in temperature without bound, the continuity relation must be
imposed. The flowof heat per unit area per unit time in the direction
of increasing variable is given by
= "v
(7)
where
dv a, dv
a+ d v
-+-"l""
ar Y 80 Y sin 0 a+
By symmetry, the temperature does not change in the 0 and directions for a surface Y = so
av
"- a+ = O
At the interface (Y =
so
The minus sign appears in Eqs. (7) and (11)because the flow
occurs in the direction of decreasing temperature, i.e., d v / a r will contain an inherent negative sign td compensate for the minus sign in
Eq.
XZE-1
H
This H should not be confused with the one in the expression
H
h=K
Equations (4) and (5) may be solved by integration and using
the boundary conditions (6), (ll), and (12).
Carslaw and Jaeger (1959, p. 232) give (without proof) the following solution:
+ 2XaK0 + 2a2(K0/K)]/6K0,
= Ao[a2 = A0a3/3Kr,
a(15)
>a
(16)
Equations (15) and (16) satisfy the differential equations (4) and
(5) and the associated boundary conditions.
Since solutions to Poisson's equation are known to be unique,
we may be confident that the actual solution has been found.
Now that the temperature has been found at all points, a check
should be made.
~~~~~
~ Suppose
e
c heat
~ generation were zero. Then
lim
=0
OK
lim
=0
OK
Ao-0
=0
OK for
l i m vO, K
=O
forr>a
If R
v"a-
OK
rea+
1im v3 =
all r c a
K-0
Other limiting checks may be performed, but a sense of underand confidence has begun to develop.
S embed this problem in a more general context to clarify
what has been learned.
In order to obtain the results in terms of the rate of heat prouction per seed, the flux over the surface Y = a may be obtained
by integration.
= (4/3)n;a3A0
Using equation.
(29)
v, =
Q
~
4TKa
This is the same temperature that would have resulted from an isolated point source in a medium of conductivity K at a distance r = a.
~
~~~e~~ It~ will nowebe instructive
~
to check
~ the results
~
numerically to establish reasonablebounds on the temperatures to be
expected. If we now assume intimate seed-soil thermal communication ( R = 0), i.e., seed coat temperature equals adjacent soil temperature, then the tempera~reof the center of the seed is found to be
A o a 2 &a2
v, = -+ -,
seed center
4Ko
3K
The seed surface temperature is
A,a
v, = 3K
seed surface
(33)
Then,
v, = soil temperature at Y = C
Q = rate of heat production, cal/ (S. seed)
a = radius of seed, cm
K = soil thermal conductivity, cal/ [S cm2(C/ cm)]
~ ~ e r ~ a l C o n ~ ~ c Carslaw
t i v i ~ y and
:
Jaeger givefor average
soil
K = 0.0023 call [S cm2(oC/cm)]
Kernel Size:
A 12/44 in. round hole screen is used by the US.
Department of Agriculture (USDA) as a basis for determining corn
breakage. If we take twice that size, say radius of kernel
a =
(g (in>
in.)
2.54 cm = 0.475 cm
Cooke
90
given. Varieties of corn with peak respiration rates of 3.0,2.1, and 2.3
cal/ (h. g) with four seeds were given.
= [3.0 cal/ (h-g)]
(3ioi (-
0.000208 cal/(s.seed)
v, =
(=)
l .22 Btu.in*
(0,00413) = 0.00042
ft?* hr*"F 12 in.
Since
A,a
3K
v, = -
K
+ V, = 0.056"C
Zr;c,
This is the maximum temperature rise above the undisturbed soil
temperature. The temperature difference v, - v, = 0.041"C is also below the limits for a thermocouple. Under the most careful laboratory
conditions,a thermocouple can be used to detect differences no
smaller than (0.05OF) (5/9) = 0.03"C.
v, = -v, + v, =
In view of thermal conduction along a thermocouple and the difficulty of correctly placing the probe in the seed, this appears to be a
futile experimental approach to the problem. Thegradients that occur
Using ~ a t ~ ~ ~ aast ai c s
normally in the soil will be vastly greater than that produced by the
heat of respiration.
However, if the contact resistanceof the surface were great, then
the internal temperature might be affected enough to alter the biochemical reactions within the seed. Therefore, the problem may be
redefined, if desired, to explore this new problem suggested by the
analysis.
Seed and soil temperatures will frequently be below those of the
calorimeter tests with a consequently lower respiration rate. The water absorbed by the seed (imbibition) will,of course, alter the thermal
energy status of the seed. The thermal conductivity of soil changes
radically with moisture content. Also, the corn kernel is not spherical.
However, the above computations may be taken as bounds on the
actual values.
If the seed were very near the surface, the method of superposition of heat source and sink could be used. The other extreme case
of a solid spherical seed producing heat at the constant rate
per
unit time per unit volume in air at constant temperature can be
examined.
v = -[h(a2-
6hK
24,
25
<a
(37)
I wish to express my appreciation to Henry D. Bowen, who stimulated my interest in this subject and helped me appreciate the magic
inherent in modeling.
Panesar
The features that vary and/or are of concern for regional planning
on the micro or macro levels are defined as spatial features. Spatial
features are taken as attributes for database management of a region.
These features help in identification and classification of homogeneous subregions in a region of interest. The spatial features are the core
of spatial models, and their accurate identification by use of a set
me tho do lo^ becomes important to eliminate any errors. Both the
variability and the equality among subregions have been used in the
literature for the identification of spatial features (Bennett and Tan,
1981; Nelson, 1987; Negahban et al., 1993;Panesar et al., 1989; Pari&,
1985; Pari& and Kromer, 1985; Smith, 1992). Thetwo approaches are
dual, i.e., equality is the dual of variability and vice versa.
It is a known fact that spatial variability exists for almost all
levels of systems, for example, in a ~ i c u l t u ~production,
al
transport,
industrial production, and the environment at themacro level and in
nutrient transport in soils and nutrient uptake by plants and inbiogas
digesters at the micro level. Spatial variability is of concern to planners and modelers because they want equal performance from or
equal distribution of benefits to spatially variable subregions. This
leads to the concept of equality among subregions. Hence, spatial
variability and equality concepts can be used to delineate homogeneous subregions or areas however small these subregions might be.
Spatial variability will help in identifying some characteristics of the
region, and equality might help to identify some other spatially variable characteristics of the region. For example, in allocating inter-
and
l n t e ~ r a t i nSpatial
~
~emporal
governmental grant resources, the equality concept will help in identification of the tax base as a characteristic fora region that otherwise
might be skipped.
. T
Temporal models are dynamic and hence require that past and future
behavior of the system be considered in any analysis. The past is of
utmost significanceforbiologicalsystems,for
these systems have
memory. The future has great importance for man-made nonbiological systems such as industries, transport, roads, electricity supply,
sewerage, and water supply. Man-madesystems using biological systems as components require that analysts take into account both the
future and past behavior of the system. The future behavior would
determine how the system would meet the future requirements of the
society for which it was made. All these factors add dynamism to
both natural and man-made systems, which complicates the task of
system analysts and modelers.
The certainty in model output means that once the model output is
predicted for an event it (the output) does not change when the
event actually occurs. Thus model outputs shouldhave a minimum
of uncertainty. Uncertainty, the dual of certainty, is due mainly to
four factors: (1) the stochastic nature of model inputs, (2) the stochastic determination of model parameters that are otherwise deterministic, (3) the deterioration of system behavior over time, and
(4) uncertainty about model structure. The stochastic nature
model inputs is self-explanatory. The model parameters are deterministic, but the methods (experiments, instruments, etc.) used to
determine them make them stochastic, For example, the spring constant of a spring is a deterministic feature of the spring, but the
technique used to determine it makes it stochastic. The spring constant is ete ermined by measuring force and displacements with the
help of instruments (which have errors); plotting force vs, displacement, and finally determining slope with the use of regression analysis makes the spring constant a stochastic parameter.
Deterioration in a systems performance, i.e., changes in the
systems behavior, with time is not uncommon. For example, the
field capacity of a soil at a given location is a constant parameter.
Rut the soil structure may change due to such activities as puddling
and inundation of water, which are part of the rice production system. Thus, the field capacity of soil changes with multiple use of the
rice production system, which therefore results in uncertainty in
model output.
The model structure of a system is based on knowledge available at the time the system was modeled. Advances in scientific
knowledge are being made at a rapid
pace. These scientificadvances
might change the model structure, thereby creating uncertainty in
model outputs at later
a date. The certainty feature of temporal models is of utmost importance to planners and forecasters.
.
Spatial and temporal features can be derived from time-space
functions. The following discussion indicates how time-space functions can leadtospatial/temporal
models and how missing
temporal/ spatial characteristics can be added to spatial! temporal
models.
y for two-
dimensions)
= set of model parameters
t = time
Let t = T where T is a specific time. Then, from Eq.
(l),the spatial
anesar
parameters, and outputs or in the form of a di~erentialequation involving time and/or space derivatives of the model outputs along
with the model inputs, parameters, and outputs. Either the model
input or model parameters or both are functions of time or space or
of both time and space. Thus, in temporal models, some or all of the
model inputs are functions of time, and in some cases timemay also
appear as a model input, but model parameters are not a function of
time. However, both Parameters and inputs of temporal models
might be functions of space. In spatial models also, similar to temporal models, some or all model inputs are functions of space variables, but model parameters are not functions of space. However,
both parameters and inputs of spatial models might be functions of
time.
If we compare Eqs. (2) and (3) in light of the above discussion,
it can be inferred that P and P consist of two distinct subsets of
model inputs and parameters. The members of the subset of P or
that are functions of space and time, respective
model inputs; and the members of the subset of
functions of space and time, respectively are referred to as model
parameters in more general terms.
~
~ The ~modified
~ expression
~
for
Z crop water
e
requirements
without water stress (CWm)to include spatial character is (Panesar,
1993):
K ET0 + I
where ET, is the reference crop eva~otranspiration, m/cropseason;
K is the crop factor; and I is the total in~ltrationloss, m/crop season.
CW,, is the model output.
The first factor in the estimation of CW;, is the crop factor K,
which depends on the crop, season of planting or sowing, growth
stage, and weather conditions such as wind speed and relative humidity. Doorenbosand Pruitt (1977) gave values of K for several crops
during four stages of crop development, i.e., initial stage (approximately 10% ground cover), crop development stage (ground cover
increases from 10% to about 70--80%), midseason stage (crop cover
remains nearly complete), and late season stage (crop cover decreases
from h11 to harvest or maturity crop cover). They outlined a procedure to determine crop factor K for each day after planting. Smith
(1992) considered water demand during land preparation and nursery raising but only for rice. Waterdemand during the two
operations
is important for other crops also. Thus, factorK has spatial and temporal variations. Hence, K will form part of the model inputs.
ET,
oraland Spatial
Integrating
99
900
T + 273
lOOO[A + y(l + 0.34U)I
+
anesar
The model for crop water requirements without water stress has
output
Inputs
Parameters
cwxn
K ; R,, G, Tt U,
900
ed,
4 t, Io, ki
The inputs are functions of time and space as discussed earlier. Therefore output (CWm)is also a function of time and space in addition to
the parameters listed above. Thus, the model output ( C ~has
~ a)
form similar to Eq. (l). The exact mathematical functional form for
inputs with time and space is not known. However, only an implicit
functional relationship in theform of time-space tables for each
input
exists, Thereforethe model has to be implemented in the form given
by Eqs.
Panesar (1993) has developed the model as a temporal
model by defining CW,, ETo, I, and Io in m/ day and removing the
time derivative from Eq. (5). Hence, the Panesar model does not provide spatial crop water requirements with water stress.
lntegrat~n~
Spatia~and Temporal Models
MODEL INPUTS
PARAMETERS
REGION-1
REGION- 2
SPATIAL
AND
TEMPORAL
DATA
MODULE OUTPUT
MODEL OUTPUTS
FOR
REGION. 1
REGION 2
REGION - 3
REGION 4
~
SPATIAL
AND
TEMPORAL
OUTPUT
1.
2.
3.
4.
5.
anesar
structured program for the task and testing and debugging of the
structured program modules are needed to eliminate programming
errors. The spatial input module adapts spatial and temporal data for
the spatialltemporal model such that the model processes all data.
The adapted data of the system will be in the form of several files or
subdirectories, each representing either a specific period or a specific
homogeneous region. The fileor subdirectory with model inputs and
parameters is processed by the spatial/temporal model. Hence, the
input module carries out two tasks: (1)data adaptation of inputs and
parameters into files or subdirectories and (2) processing of adapted
data from the file or subdirectory.
anesar
I
1
anesar
Electricity
(GWh)
Labor
of
man-hr
Water
(ha m)
Energy
(GJ)
46,809.68
34,738.44
49,619.08
58,852.33
31,102.42
17,866.85
r 35,997.36
20,632.91
48,323.08
41,174.13
12,458.31
49,350.39
446,925.10
186,508.60
27,325.54
42,922.50
114,302.70
109,266.20
96,769.38
181,792.30
107,014.30
248,312.10
190,691.10
40,538.72
202,080.50
1,547,524.00
396,871.2
248,950.8
409,184.4
467,543.5
317,750.7
184,199.6
417,121.5
140,911.0
439,159.9
372,796.1
152,361.5
210,476.7
3,757,328.0
638,287
733,073
471,335
714,324
266,147
215,237
361,110
235,944
513,610
384,488
106,591
575,977
5,216,123
11,373,274
7,435,317
11,019,501
13,000,204
6,368,922
4,442,695
9,485,957
4,652,270
13,220,959
12,339,237
2,545,513
14,459,249
110,343,072
District
Amitsar
Bathinda
Faridkot
Ferozpore
Gurdaspore
Hoshiarpore
~a~a~da
Kapurthala
Ludhiana
Patiala
Rupnagar
Sanpr
State totals
Models
Temporal
~ n t e and
g r a t Spatial
i~~
107
1980
1985
1991
1980
1985
1991
1980
1985
1991
1980
1985
1991
1980
1985
1991
1980
1985
1991
1980
1985
1991
1980
1985
1992
1980
1985
1991
1980
1985
1991
1980
1985
1991
1980
1985
1991
1980
1985
1991
h)(man.
1.ooo
1.331
1.445
1.000
1.259
1.339
1.ooo
1.266
1.350
l .ooo
1.312
1.472
1.ooo
1.530
1.521
1.ooo
0.848
1.028
1.ooo
1.380
1.540
1.ooo
1.220
1.408
1.ooo
1.283
1.398
1.ooo
1.329
1.415
1.ooo
1.289
1.347
1.ooo
1.406
1.641
1.ooo
1.300
l .424
1.000
1.370
1.617
1.ooo
1.659
1.849
1.ooo
1.678
1.693
1.ooo
1.355
1.531
1.278
1.392
1.000
1.068
1.547
1.ooo
1.551
1.911
1.237
1.461
1.ooo
1.756
2.093
1.ooo
1.345
1.517
1.ooo
1.333
l .544
1.ooo
1.953
2.556
1.000
1.460
1.740
1.ooo
1.299
1.408
1.182
1.195
1.ooo
1.287
1.374
1.ooo
1.387
1.617
1.ooo
1.462
l.424
1.ooo
0.850
1.021
1.ooo
l .200
1
1.ooo
1.l45
1.302
1.ooo
1.086
l .068
l .ooo
1.224
1.223
1.ooo
1.l42
1.111
1.ooo
0.990
0.882
1.ooo
1.199
1.263
(ha-m)
(GJ)
1.ooo
1.370
1.617
1.ooo
1.660
1.850
1.ooo
1.677
1.692
1.ooo
1.355
1.531
1.000
1.279
1.392
1.ooo
1.068
1.547
1.ooo
1.551
1.912
1.ooo
1.237
1.461
1.ooo
l .755
2.092
1.ooo
1.345
1.518
1.000
1.334
1.546
1.ooo
1.955
2.559
1.000
1.876
2.070
1.ooo
1.696
2.002
1.ooo
1.608
1.824
1.ooo
1.649
1.922
1.ooo
1.951
1.915
1.ooo
1.035
1.579
1.ooo
l .722
1.922
1.ooo
1.241
1.463
1.ooo
1.476
1.583
1.000
1.622
1.762
l .ooo
1.677
1.606
1.ooo
1.689
2.070
l .ooo
1.619
1.833
1.491
1.733
anesar
added to the model output. The results show relative diesel, electricity, water, labor,and total energy requirements in relation to 1980, yet
the table effectively displays the spatial and temporal behavior of the
agricultural production system.
The spatial output from the S E N ~ A Pmodel is in the form of
either maps or tables, Thus the outputmodule could also be designed
and implemented in Fortran, dBASE, and ARC-INFO such that the
spatial changes over differenttime periods could be displayed in the
form of maps (results of one period overlaid over the results of another period to pinpoint differences) or in the form of tables with
columns covering different periods and rows representing different
regions. TheFortran language would be very useful for complex calculations, such as finding totals, computing results for presentation
in tabular form, and getting hard copy of result tables. The dBASE
language/ environment helps in transferring model output to attributes of various homogeneous regions. ARC-INFO helps in developing maps for the model output, overlaying maps of different time
periods for visual display of changes, and getting hard copy of these
maps. Hence, implementation of the output module in a combination
of several small modules formulated in a suitable computer language
(which may differ among modules) may be helpful and easy for better presentation of modeled results.
The integrating methodology is a sequential step method comprising revisiting the system that has been modeled; studying the
modeling methodology used; understanding the assumptions and
model structure; noting model inputs, outputs, and parameters;
studying units and dimensions of model inputs, outputs, and parameters; finding values of the model parameters used for computer implementation of the model; studying the language used for the implementation of the model on the computer; finding computer data
formats for inputs, outputs, and parameters of the model used for its
implementation; designing input and outputmodules; implementing
input and output modules on the computer in conjunction with the
model; and presenting model results to facilitate interpretation. The
input module is designed and implemented for two tasks: data adaptation of inputs and parameters into files or subdirectories for use
by the model and processing of adapted data from the file or subdirectory. Similarly the output module is designed and implemented
for two tasks: storing the model output of a specific period or homogeneous region into a file or subdirectory and integrating the output files or subdirectories to predict spatial as well as temporal behavior of the system in a desired fashion.
The integration methodology has been demonstrated by taking
an example from an agricultural production system that has both
spatial and temporal character.
11
Panesar
Spatial Integrating
111
c.
Sowell and ~ a r d
11
3. Constraints
The decision variables relate to the decision that must be made and
are expressed with algebraic symbols. In the above example, the decision variables (xj,j = i, 2, 3) can be defined as
+ c2x2+ c,x3
(1)
or
3
j=
15
Subject to:
x1
+ xz + x3 S 200
+ 3xz + 4x,
300
(land constraint)
(labor constraint)
and
EXA
The diet, or feed mix, problem deals with determining the amounts
of alternative feeds to provide for livestock. The objective is to provide feeds that meet certain nutritional requirements at minimum
cost. The decisionvariables are often the amounts of the various feeds
to be included in the diet. The objective function coefficients define
the cost per unit of each feed so that in minimizing the objective
~ n ~ t i one
o n is minim~ingthe cost of feeding the livestock. The constraints express the need to meet minimum ~utritionalrequirements.
Each nutrient has its own constraint.
Consider the situation where a farmer, among his several enterprises, raises hogs for market. He has three types of feed available
for his hog operation and is mainly concerned about nutrients. He
has had each feed analyzed to determine the amounts of the three
nutrients in a pound of each feed. He has also determined the
imum daily requirements of each nutrient for each hog and the cost
per pound of each feed. He now wants to know how much of each
feed should be fed to each hog if he is to meet the mini mu^ nutritional requirements at a minimum cost. The data he has obtained are
given in Table 1. Each entry under feed has units of nutritional
value per pound of feed.
The decision variables for this problem can be defined as
+ 6x2 + 8x3
17
Nutrient
Daily nutritional
requirement
Feed
Corn Ib"
Tankage lb-l
Alfalfa Ib-l
___
1
2
150
50
cost
60
55
5
12
18
40
8
6
40
75
12
Each nutrient will have a constraint that expresses the fact that
its daily requirements must be met. Thus, the constraints are
Sowell and
l1
Field
(ha)
m3/ha
150255
152
80
175
210
1
2
3
4
1.7
1.9
298l .7
2.1
9
Hauling distance
Total
(m3)
Storage 1
Storage 2
10
29
M1
14
6
12
18
+
+
x43
5 725
x42
5 520
As before, the xs 0.
This formulation will allocate the beans from each of the four
fields to the two storage sites in such a way as to minimize the total
transpo~ationcost.The problem ensures that all the beans are
shipped (first constraint set) and that the amount shipped to each
storage location does not exceed its storage capacity (second constraint set).
z = clxl
+ c2x2 +
+ c,,x,
(3)
Subject to:
ailxl
+ ai2x2+
+ airZxn
i = 1, . . . , m
=, 2)b,
(4)
and
x1, x2,
. ,x ,22
(5)
* *
+ c2x2 + - + c,x,
*
Subject to:
ailxl
+ ai2x2+
+ ailzx,= b,
i =
. . .,m
and
x2,
* /
x, 2 0
+ a12x2 + al3x3 S
x1
+ a12x2+
+ x4 = b,
For Eq. (10) to hold, x4, like the other variables, must be nonnegative.
From the standpointof meaning relative to theinitial problem, x4 can
be interpreted, for example, to be the resources (represented by the
constraint) not used in the problem, i.e., the slack resources.
For the constraint, consider the following:
%lXl
+ a12 x2 + 43x3 2 bl
+ a12x2 +
- x4 = bl
(12)
Again, for Eq. (12) to hold, the surplus variable, must be nonnegative. To see the physical meaning, recallthe diet example. In that
situation the surplus variable represents the nutrients supplied over
the m ~ i m required.
u ~
Slack and surplus variables, while added to convert in inequalities to equations, do have meanings with respect to the original problem and are therefore a part of the problem. They carry valuable
.
In trying to understand how linear programming optimizes a linear
function (objective function) in which the variables must satisfy a set
of simultaneous linear equations, it is useful to graphically view what
is happening. Since no combination of variable values that does not
satisfy the constraints can beconsidered, it is of interest to graphically
see this set of solutions and how it related to the objective function.
.
Consider the problem
Maximize z = 80x, + 90x2
Subject to:
+ x,
2x1 + X ,
3x1 + 5x2
x1
40
x,, x2 2 0
x, + x2 5 10
If the equal sign holds, i.e.,
x1
+ x, = 10
we have the equation for a straight line, and anypoint on that straight
122
line satisfies the equation. Notice that any point lying in the halfplane below the line satisfies the constraint
3-
22
owever, only those below the line and in the first quadrant satisfy
both the constraint and the nonnegati~tyrestrictions. Consider the
other constraints in a similar manner, and wefind a set of points that
satisfy all the restrictions of the problem. This set of points is represented by the cross-hatched area of Fig. 1. All points in this shaded
area are feasible solutions since they satisfy all the restrictions. This
set of points is a convex set,* and it can be shown that the set of
*A set is convex if all points on a line connecting any two points in the set is also in
the
e.g.,
set,
is a convex
set.
isconvex
not a
set.
123
is a straight line. Any point on this straight line will give a value of
= 900. For each different value of z we obtain a different straight
line, but the lines are all parallel. We wish to find the straight line
with the largest value of z that has at least one point in the region of
feasible solutions. Notice (Fig. 1)that the line
has no points in the region of feasible solutions and thus is not the
optimal feasible solution to the problem. Let z2 = 720, and note that
the line
has points in the region of feasible solutions but that z can have a
larger value and still be in the region of feasible solutions. By drawing
lines parallel to and z2 we find that the maximum value that z can
attain and remain in the region of feasible solutions is found where
the lines x1 + x2 = l 0 and 3x1 + 5x2 = 40 intersect, or at x1 = S, x2 =
5. The value of z at the point is 80(5) + 90(5) = 850. Note that the
maximum value of z occurs at an extreme point or corner of the
region of feasible solutions. It can beshown thatthe optimal solution
of a linear programming problem always occurs at an extrerhe point
of the convex set. In this example the optimal solution is unique, i.e.,
only one (xl,.x2) pair lies in the region of feasible solutionsand optimizes the objective function. The following example is one for which
there are an infinite number of optimum feasible solutions.
+ 4x2
Sowell an
Subject to:
x, = 4
xz = 0
and
2x1 + 3x2 = 13
and that the minimum value of z occurs along this line. Although
this is an extreme point of the convex set, the objective function has
the same minimum value of -16 anywhere along the line
25
x1
x, =
xz = 0
and
2x, + 3xz = 13
Maximize z = Zx, - x,
Subject to:
we see that it lies in the region of feasible solutions. We also see that
for any other value, zzr of the objective function such that z, > z1 the
line
2x1 + x, = z,
also lies in the region of feasible solutions. This means that the objective function does not attain a maximum value on the set of points
satisfying the constraints. This set is unbounded, and the problem
has no solution. The situation is also referred to as a problem with
an unbounded solution.
126
~ o ~ e ~ with
i n glinear Programming
27
*2
x1
bered pairs (x,/ in the coordinate plane that satisfy all the restrictions. Therefore/ the problem has no solution. (See Fig. 4.)
Summarizing, we have looked at examples of linear programming problems for which (1)there is a unique optimal solution, ( 2 )
there are an infinite number of optimal solutions, (3) the solution is
unbounded, or (4) there is no solution.
Suppose a farm tool manufacturer has two primary resources, machine time and labor hours. In a certain period of time he has 200
machine-hours and 300 labor-hours to devote to three products, xl,
x,, and x3. Product xl takes 15 machine-hours and l 0 labor-hours per
unit. Product x, takes 10 machine-hours and 25 labor-hours per unit,
and product x3 takes 10 machine-hours and 20 labor-hours per unit.
The manufacturer wishes to determine the mix of products that will
maximize profits while not exceeding the machine-hours available.
He wishes to provide full employment for his workforce and therefore requires that all available man-hours be used. This is a competing
x,
x3
The profits to the manufacturer will be $5 per unit of xl,$10 per unit
of x,, and $12 per unit of x,. Thus, the optimizing to be performed is
Maximize 5x,
+ lox2 + 12x,
Since the first constraint is a S inequality we must add a slack variable and rewrite the set of constraints as
x1
+ lOx, = 200
10x1 + 25x2 300
15x1
=I
X1
7.2727
5
30
0
0
x2
x3
9.0909
0
0
12.5
0
0
- 20
40
12 120 0
0 180 15
x4
0
0
-250
0
80
50
Objective function
(5x1 + 10x2 + 12x3)
127.27
175
*See text.
cause the manufacturer is primarily trying to determine what products to make to maximize profits with the resources available.
~ o ~ e l i with
n g Linear ~rogramming
31
To further explain the mechanics of the simplex procedure, the tableau format is now described and applied to a simple maximization
132
+ c2x2+
+ c,x,,
Subject to:
The ~~~~~a~
tableau, using the above notation, is shown in Table
4. In this tableau, xBirepresents the value of the variables in solution
(initially xBi = hi); zj = cBiyij,where summation on i over variables in
solution; and zj - ci represents the relative measure of effectiveness
coefficient (sometimes referred to as rj).
The initial tableau, after a few iterations, can be generalized as
shown in Table 5. In the general tableau, the are the variables in
solution at a given iteration and yij are the values of ay at an iteration
after the initial tableau.
CB1
cl
Yll
y12
CB2
c2
y21
yz
y1
l*
Y2n
y1,1,+1
0
y2,w+2
...
* *
0
0
The specific mechanics of performing the four previously discussed steps of the simplex procedure are as follows, where the terminology is that of the above tableaus.
Step
The new tableau is computed by pivoting. Pivoting is
nothing more than establishing the identify matrix with the variables
in solution. As noted before, this simply means that in each column
associated with a variable in solution, there must appear one unit
coefficient and the rest zeros. Thispivoting can beexpressed through
the equation
bl
b2
13
[~],
+ r;i =
. . ., n
+ m; i =
. ., m
2x1 +x2+ x3
4x12x22x3
6
5
x11 Xzr
0
First, the inequalities must be converted to equalities by adding slack
variables. Thus, the constraints become
2x1 +
+ + =6
4x1 + 2x2 + 2x3 + x5 = 5
The initial tableau, using the above data, is shown in Table 6.
Here x4 and x5 are the variables in solution as they are readily
and x3 are set to zero. So we have an initial
evaluated when xl, xz,
basic feasible solution to begin the simplex procedure. The zj and zj
- cj were computed using the data in the tableau from the initial
statement of the problem.
Proceeding to step 1, a review of the zj - row reveals that
optimality has not been reached. All nonbasic variables (xl,x,, and
x3)have negative values, indicting that any one could be brought into
solution in step 2 and result in animprovement in the objective func-
Variable
in soln
cZ3i
x1
x2
x5
x4
x5
0
-2
zj
zj - cj
xBi
0
-5
6
5
3"
5
- = 2.5
2
y22
y12
g12= 2 - 2
(:)
Q22
y12
(E)
CBi
Variable
in soln
x5
0
5
7/2
512
zj
5.
0
zj - cj
5/2
0
0
25/2
5/2
= 0;
x, =
5
-*
= 0;
x* =
7
2
-*
x5
=0
farmer has the equipment to produce three different crops and wants
to h o w which crops, and how many hectares of each, to produce in
order to maximize profits. The farmer has the crop data shown in
Table 8.
The decision variables can be defined as follows.
+ x2 +
3x1 3- 4x2 + 5x3
300x, + 260x2 + 230x3
x1
(land)
x3
400
(labor)
25,000
(capital)
0,
x2
0,
x3
Crop 1
Crop 2
Crop 3
Capital needed
per hectare
Profit per
hectare
4
5
260
60
55
The initial simplex tableau can now be established using as the initial
basic feasible solution the values of x4, x,, and x6 when x1 = x, =
x3 = 0. These variables are placed in the solution initially and the
basic solution is immediately available.
x4 = 100 ha of land
x, = 400 h of labor
x6 = 25,000 dollars of capital
Thus, the initial tableau is as given in Table 9. From this tableau, it
can be seen that optimality has not been reached. Since x3 has the
smallest zj - cj value, it is chosen to enter solution. Since x5 has the
minimum ratio, it is selected to leave solution.
The second tableau, after pivoting, is shown in Table 10. Optimality has not been reached as - cl and zz - c, are still negative.
Since x1 has the smallest zj - cj, it enters solution. In determining the
variable to leave, cannot be considered as it has a negative under
the xi column. Since row 3 has the smallest ratio, x6 leaves solution.
The third tableau, after pivoting, is shown in Table 11. At this
tableau, optimality has been obtained. The optimal solution is
Maximize z = $6222 profit
cBi
Variable
in soln
x3
x5
x2
x1
x4
x5
x4
100
0
25,000
x6
0
zj - cj
Table
Cj
Variable
in soln
CB;
x4
0
68
0
60
55
68
x1
x2
x3
x4
x5
x6
-0.20
0
0
1
20
80
6600
0
0
5400
0.40
0.60
162
x3
x6
68z; 54.4
z; - cj
40.8
- 19.2
0.20
0.80
76
0
1
0
1
0
0
-0.6
0
0
0.20
-46
19.66
13.6
xl3i
with
ha of crop 1 grown
ha of crop 3 grown
0 ha of crop 2 grown
40.74
55.56
and with 3.70 ha of land not utilized. These hectares could not be
used because all the labor and capital were exhausted (x,and x6 have
zero values).
Table
cBi
0
68
60
Variable
in soln
x4
x2
x3
0
0
1
0.01
0.52
0.47
0
1
0
68 zj 63.42 60
z; - cj
0
8.42
x3
x1
x4
x5
x6
1
0
0
-0.09
0.37
-0.28
-0.002
-0.004
0.006
6222
0 0.09 8.36
0 0.098.36
3.70
55.56
40.74
slack variables to all constraints. What happens if there are greaterthan-or-equal-to constraints or equations as constraints in the initial
formula~on?Forexample, suppose that in the previous problem,
land was an equality in the initial formulation. This would have
meant that x4 would not be needed. Thus, the initial basic feasible
solution would have been difficult to determine.
To easily find the first basic feasible solution needed to initiate
the simplex procedure, a ~ ~ ~ ~c ai~ ~i albare
l ~ used.
s
These are variables
that are added to the problem after all inequalities have been converted to equalities. They are therefore added to equations! They are
added simply to get a basic feasible solution and to get the simplex
under way with a minimum of effort. Since they are added to equations, however,they cannot be in the final solution at a positive level.
They have no meaning to the problem (as slack and surplusvariables
do)-they are simply added to facilitate solution of the problem.
Since artificial variables are added for solution purposes only
they must be removed from the problem. One technique used to
"force" artificial variables out of the solution to a linear programming
problem is called the Big M method.
The Big M method "forces" artificial variables out of solution
by giving them a large (Big M) positive or negative objective function
coefficient d e p e n d ~ gupon whether the problem is a minimization
or a maximization problem, respectively. In this manner, it is highly
desirable, from the standpoint of optimizing the problem, to remove
the artificial variable and not let it return to the solution.
To illustrate the procedure, consider the problem
Maximize z = 3x1 + 2x2
Subject to:
+ x2 2s 4
x1 + x2 =
2x1
2x1
+ x, + x, = 4
x1 + x, =
x1,
x, 22
o ~ e l i with
n ~ Linear
2x1 + x2
x1 + x,
+ x3 = 4
+ x4 = l
Artificial variable x,, however, must not appear in the final solution
at a value other than zero. If it does, the solution is not feasible because the second constraint is violated!
In setting up the first tableau for solving this problem by the
simplex method, x, will be assigned a large negative objective function coefficient, --M. In this manner the simplex procedure will find
that once x, is removed from solution, it will be highly undesirable
to bring it back into solution. Thus, in the initial tableau, the objective
function coefficients will be (3, 2, - M ) for the respective variables.
The first tableau is shown in Table 12.
x3 and x4are in the initial basic feasible solution,and their values
in the solution are the right-hand-side values from the two constraint
equations, as before.The only difference is that x4 is an artificial
variable.
Since x1 has the smallest value of zj - cj, it is to enter solution
on the next iteration. The minimum ratio rule results in x4 leaving
solution. Pivoting results in the second tableau, as shown in Table 13.
Since all zj - cj are positive, optimality has been reached. With x4 out
of solution, the solution is feasible. Thus, x4 permitted the simplex
procedure to be easilyinitiated, and the large negative cost coefficient
helped ensure that, once removed, the variable would not reenter
solution. Again, if x4 had been in the optimal solution at a zero value,
CBi
Variable
in s o h
x2
-M
zi
2;
- cj
-M
-M - 3
-M
-M - 2
0
0
0
0
-M
142
CBi
Variable
in soln
X1
x3
X1
Zj
0
0
zj - cj
x2
x3
x4
-2
2
1
3
3+M
~ Q d e l i n with
g
Linear ~rQgramming
143
In using a computer routine to solve a linear programming problem, the student should carefully read the routines documentation.
The specifics of how the routine operates, how data are input, and
what is output are described. The understanding of linear programming obtained to this point will permit students touse most routines
to solve a linear programming problem.
URAL A P P L I C A T I ~ ~ S
~ a c ~ i n e r y a n a ~ e ~ and
e n t Crop ~ r o d ~ c t i oPro
n
xl = acres of corn
x, = acres of soybeans
x3 = acres of grain sorghum
x4 = acres of wheat
x5 = dummy variable associated with fixed cost
In addition to limitations on acres of cropland available, it is
assumed that there are also limitations on time available in the spring
for row crop planting and in the fall for row crop harvesting. These
times are 90 h and 160h, respectively both at an 85% probability
level. Time available for planting and harvesting wheat is assumed
to be unlimited. There is also a minimum wheat acreage of 40 acres.
This is the minimum required to justify special wheat equipment.
In order to develop the coefficients for the constraints involving row crop planting and harvesting time it is necessary to relate
machine parameters to the operations (i.e., a six-row planter
equipped for 30 in. rows, traveling at 4.2 mi/h, with a field efficiency
of 0.6 is assumed). The coefficients must have units of hours per
acre, or the inverse of field capacity. In the case of the six-row corn
planter, the coefficient is
8.25
-speed width field efficiency 4.2
= 0.22 h/acre
mi/h
15 ft
0.6
x1 + xz + x3 + x, 5 680
0.22~1+ 0.19~2+ 0.20~3 90
0.60~1+ 0.57Xz
x1,
xzr
(land constraint)
(row crop planting time
constraint)
+ 0 . 5 5 ~=S ~l60
x31
x,, x5
22
(nonnegativity restrictions
constraint)
The coefficients for all decision variables except x5, fixed cost,
represent the net profit per acre. For each unit of x5 the net profit will
be reduced by one unit; consequently its cost coefficient is -1. The
objective function for this problem is
Maximize z = 8 4 . 5 0 ~+~9 3 . 9 0 ~+~9 3 . 0 0 ~+~4 5 . 1 0 ~-~x5
Solving this problem we find the maximum net profit to be
$6102.54. The optimal crop mix is 291 acres of grain sorghum and
389 acres of wheat.
+ x31
+ x21
x3l
x12
x13
+ x32
+ x22
+ x23
400
I
600
x33
I
300
+ x12 + x13
I
700
~ 3 % x32
I
800
+ x33 S
To model the policy of a uniform set-aside acreage, the total percentage of planted acreage per farm must be equal for each pair of the
three farms. Therefore, the equations
Water available
(acre ft)
Usable acreage
1500
2
900
-
Maximum
Water
(acre
acreage
Crop
consumption
ft / acre)
Expected
profit
($/ acre)
5
4
3
x12 + x22 +
600
x11
x32
x13
400
300
100
+ x 2 3 + x33
300
300
400
must be satisfied. Since the first two equations imply the third, the
third equation can be omitted from the model. Furthermore, these
equations are not yet in a convenient form for a linear programming
model, as all of the variables are not on the left-hand side. The final
forms of the uniform workload restrictions are
3(x1,
= 225 acres
Maximize z = $342,500
~ Q d e l i n with
g
linear ~rQgramming
Max
Min
'Expected benefit
($ benefit/$ spent)
2.00
0.2
0.3
0.7
1.75
0.1
0.4
2.25
0.2
0.8
2.50
...,
j=1,2,
c2x2c3x3
+
225x6 + 2.5OX7
+ 2.25X5
2.5ox8
+ + x3 + x5x4
+ xx76
x2
+ x8
+ + x5 + x7 I10
x1 + x3
2x,
+ 2x7
or
x1 - x3 + 2x5 + 2x7
x3
and
"3x5
+ x7
Sowell and
15
x1
+ x2
2 0.2
Similar constraints that exist for the other three conservation measures are written below.
For managing existing rangelands,
0 . 7 ~- ~0 . 3 ~
2 ~0
0.3~~
-0.7
0~~
X3, X41
= 5.00;
x5 = 2.19;
x51x6x,7fx8
x2 = 5.00;
x6 = 19.71;
x3 = 1.67;
x7 = 1.14;
x4
I=
x8
0.72
4-57
odeling with ~ i n ~ a r
l51
Practice
Conservation cropping
2. Range management
3. Sediment reservoirs
On-farm structures
Total
Total expenditure,
optimal
($/million)
share,Federal
optimal
23.9
219.0
50%
70%
10%
57.1
20%
From a practical viewpoint, one might ask, If the above problems were developed from an actual government conservation program, would the administrator allocate the money as noted in the
solution? The answer to this question depends upon how well the
LP formulation, with all its assumptions, describes the real-world
problem. The administrator, without any further work on the LP solution, must try to evaluate (somewhat subjectively) the validity of
the assumptions, the probability that the linear programmin~coefficients will remain the same over time, and other factors not included
in the model (e.g., the likelihood that all structural measures will be
used in appropriate situations and adequately maintained and the
likelihood that nonstructural measures will be adequately maintained). Such an evaluation might cause the administrator to modify
the linear p r o ~ a m m i n results
g
somewhat in deciding the appropriate
funding levels between alternatives and the fraction of the cost of
each measure to be funded by the government.
It is possible, via sensitivity analysis, to examine the effects of
changes in the assumptions or benefit estimates on the solution. If
the model results are greatly affected by small changes in benefit
estimates or by removal or modifications of assumptions, then the
administrator would need to have confidence in the validity of his
assumptions in order to have confidence in the results.
.
This example is taken from Sowell et al. (1980).
A field with initial elevations as shown in Table 17 is to be landformed for improved water ~anagement.Based on soil characteristics and crops to be grown, it has been determined that in its final
52
9.76
ft apart in both
design the row slope of the field should be in the range of 0.1-0.5%
and the cross-row slope in the range of 0-1% for optimum water
management. The shrinkage characteristics of the soil are such that
the volume of cuts in the field should exceed the volume of fills by
a factor of 1.25-1; a tolerance of k0.05 is permitted. It is desired to
form the field to meet the above requirements while minimizing the
amount of earth moved and consequently the cost of land forming.
There are several types of land forming designs depending upon
whether constant or varying slopes are desired and where cross-row
drainage is desired (Sowell et al., 1980). For this example we choose
the design that assumes constant slopes in both the row and crossrow directions.
Formulation of the landforming design problem as a linear program is based on the assumption that the volume of cut or fill around
a field station is a linear function of the depth of cut or fill at that
station. The objective of a land form in^ design is to form the field to
meet the specified range of slopes with a minimum of earth movement, which results in minimum cost. Correspondingly, the linear
p r o ~ a m m i n gobjective function is expressed in terms of minimizing
the sumof cuts in the field. Beforegoing further with the fo~mulation
of the linear program it is necessary to define the notation that will
be used in the model.
xij = cut at station j ] ,
i = 2, j =
yii = fill at station j ] ,
i = 2, j =
U = slope in the row direction, ft/ 100 ft
v = slope in the cross-row direction, ft/
2, 3
2, 3
ft
53
It was previously stated that fills and cuts are directly proportional;
therefore, fills are not included in the objective function for they will
be minimized when cuts are minimized.
There are five groups of constraints in the linear programming
formulation of the land forming design problem. The first two are
similar, one for the cross-row direction and one for the row direction,
and represent the majority of the constraints. They are based on the
arithmetic of the differences between the elevations of two adjacent
field stations after the land has been formed. (Recall that the stations
are spaced 100 ft apart in both the row and cross-row directions.)
Consider the first two stations in the first cross row of the example
(Table 17). The difference between elevations will be equal to the
slope between the two stations, i.e.,
(10.01 - x11
0.1;
0.5;
1.0
"
"
$ 4
"
P J ?
ct'?
"
$ $
"
I
7
"
+
"
. c $
l?
a,
S
bi
P
i?
'2
~ o ~ e ~ Xi it^
n gLinear ~rogramming
155
Table
Design Cuts and Fills for Land Forming Design
Example"
Cross-row direction
Row
direction
-0.06
3-0.12
--0.17
3-0.18
0.00
2
-0.08
0.00
i=l, 3
j=l,3
i=l, 3
j=l,3
j=1,3
j=l,3
The solution to the land forming design problem, involving original field elevations as shown in Table 17, is given in Tables 18 and
19. The design slopes in the row and cross-row directions are 0.105
ftfl00 and 0.109 ft/lOO ft, respectively. The cuts or fills at the individual field stations are given in Table18, and the final design
elevations are given in Table
Table
____~
Cross-row direction
Row
1
2,
10.19
10.08
9.98
10.08
9.97
9.87
9.97
9.86
9.76
Department of Agriculture,
T~omso~
appropriate feedback. Of course, sensors cannot accomplish these adjustments by themselves. A bit of intelligence derived from field experience and keen practical knowledge of the processes involved
must be programmed into the system.
The following questions could be raised. Why might it be necessary to combine sensor- and model-based approaches? Why not
dispense with themodel and use sensors with appropriate algorithms
as the primary i n s t ~ m e n t sof decision making? The latter approach
is appropriate in many cases. For example, high volume instruments
( W I ) for cotton grading use sophisticated sensing systems that are
continually being improved upon (Thomasson, 1993). Although the
output could be used in a larger simulation of machinery sequences,
nothing needs to be modeled to determine grade because sensing
systems by themselves can do the job and do it well.
Soil water sensors, used for years to help the farm manager
schedule irrigations, are inexact in their characterization of crop water
stress. In their simplest mode, these sensors placed in the soil can
help the farmer determine when to irrigate on the basis of a preset
"trigger point," which usually indicates soil water potential or the
amount of energy the plant exerts to extract water from the soil. Soil
water sensors have been around for years, but farmers are still reluctant to use them for a variety of reasons. Some types of sensors respond well to temporal changes in water content but require servicing. Other types do not require servicing but respond more slowlyto
changes in actual soil water content. For proper irrigation scheduling,
a sensor must "track" the soil water conditions in situations requiring
frequent replenishment of water as in the dripirrigation of tomatoes
(Smajstrla, 1985). Low frequency irrigation systems such as those of
the center pivot type can, however,be scheduled with sensor readings
that cannot respond immediately to soil drying (Thomson and
Thread@, 1987). This is because sensors can be read once a day (in
the morning) for successful control.
Realizing limitations of soil water sensors,researchers have
spent much time refining crop water use models for actual irrigation
scheduling. These models need to be calibrated for soil type and crop
characteristics in a specific en~ronment.Sensed weather data are input to the models in many cases. Accuracy of critical input parameters becomes more critical as the required interval to read sensors
becomes shorter,as with crops requiring frequent replenishment. Another problem is that many of these models are site-specific or need
to be recalibrated in new environments. Periodic soil sampling or soil
rocess S i ~ ~ ~ a t i o ~
15
water sensing is wise to ensure that the model stays on track, but this
can be labor-intensive.
In situations like the above where there may be imperfections
in both sensor- and model-based management methods, a new
method of combining the two has shown promise. Sensor responses
may give clues as to which model input parameters require adjustment-a sort of dynamic calibration. Algorithms can be in place to
regulate the amount of adjustment, interpret responses to the adjustment, and take appropriate action. Ideally sensors would not be
needed as the system "learns" and provides more accurate control of
the process.
The purpose of this chapter is to summarize the development
of the method and the results from a study that used sensors and a
water use model together. Other potential applications for the method
are discussed as well as potential problems in developing the method
for a particular application.
Irrigation management can be accomplished using properly interpreted readings from sensors placed in the soil, water balance models,
or both. As has been mentioned, both methods have drawbacks. Realizing this, an objective of combining the two to provide better control was implemented.
The model chosen for the study was PNUTGRO 1.02 (Boote et al.,
1989a). PNUTGRO is a process-oriented peanut crop growth model
that is an adaptation of a soybean model,SOYGRO (Wilkerson etal.,
1983; Jones et al., 1988). The model predicts dry matter growth, leaf
area index (LAI), crop development, and final yield depending on
daily weather data and specific soils. Primarily developed as a research tool, PNUTGRO allows scientists to test the effect of different
soil, weather, cultivar, and management factors on crop yield. The
model can be runon a time step of 1 day covering the entire growing
season. Water balance components of this model were implemented
in the system described herein.
The knowledge system described here could be used with any
crop simulation model that uses the IBSNAT (1986) format for input
and output files. In addition to PNUTGRO, models that follow the
IBSNAT format include CERES-MAIZE (Jones and Kiniry 1986),
Two weighing lysimeters (300 220 cm) were inst~mentedto determine the soil water balance. Both lysimeters were installed at the
Irrigation Research and Education Park (IREP) by Butts (1988). For
this study each lysimeter was instrumented with four load cells, a
net radiometer, and thermocouples. Peanuts (Flomnner)were planted
by hand on May 20, 1988 in a moist soil cm (1.5 in.) deep and
cm (1.5 in,) apart in each row. The plants were thinned to cm (3
in.)apart in thenorth lysimeter when a suitable stand was observed
15 days after planting (PAPL), June 3. Plants in the south lysimeter
lagged somewhat behind the ones in the north lysimeter in the early
stages and were thus thinned 121PAPL, June 9.
The e~uivalentof560 kg/ha of a 5-10-15 mix of fertilizer was
incorporated into the soil for each lysimeter. A decision was made
not to incorporate a preplant herbicide because the lysimeters were
small enough to be weeded manually, Watermark* soildater sensors
(Irrometer, 1989)were placed at several depths (Fig. l) on June21 in
one location of each lysimeter that showed a good stand population.
These sensors were read manually at least once a day using the Watermark 30KTC meter. To obtain soil water tension from meter readings, a curve was derived for the Watermark 30KTC meter (Thomson,
1990). This curve was related to the sensors ac resistance, whichis a
function of soil water tension and temperature (Thomson and Armstrong, 1987). To provide temperature input or compensation of the
soil water sensors, thermocouples were placed
20, 40, 60, and 90
cm below the surface, the same depths as the sensors. Since some
sensors were placed under the growing canopy and some under exposed soil early in the season, two banks of thermocouples were
placed in each lysimeter to represent both covered and exposed conditions. Thermocouples and load cells were read every half hour.
A manually activated sump pump and dual tank system were
used to drain the lysimeters after large rain events. The quantity of
water removed was measured and subtracted from the total water
removal indicated by the load cell measurements. In this way water
Use of trade names in this publication does not imply endorsementby the authors
of the products namedor criticism of ones not mentioned.
IGURE
use by the crop during the period of sump pump operation could be
estimated. Details of calibration procedures for the load cells and a
description of the data acquisition system are given by Thomson et
al. (1989) and Thomson (1990).
An expert system to interpret soil water sensor readings was developed from previous knowledge about soil water sensor responses to
l6
Thomson
rocess Simu~ation
Julian Day
163
1
"""""""..
l
Sensor
Parameter Input
S E N R E ~ . W DATRUL-l/
~.
written inthe rule development language of VP-Expert, evaluated sensor data taken from the spreadsheet and placed flags
directly
into
the
spreadsheet file,
~ E ~ R E ~ next
. W to~ readings
/
that were questionable. Readings
that were out of the sensor's range of operation received a 1 in the
flag location. Sensors that were judged to be giving false readings
based on criteria 4 and 5 above were given a 2. Different flags were
placed forthese two situations because SENPOS, the parameter modification programf still used out-of-range readings that were increasing to modify a root weighting function WR(I), which is discussed
later. Absolutevalues of sensor readings that were out of the sensor's
range of operation could not be used for other adjustments. Spread-
To construct the knowledge base for PNUTGRO parameter modification, routines were developed to modify the PNUTGRO weather
file, accept soil water tension data from the user, extract soil water
contents and root length densities from PNUTGRO output files, convert soil water contents to tensions, and display the results.
The user is asked to enter sensor data or run the system with
default sensor files. If the default option is chosen, the sensor data
acquisition routines are bypassed. If the user has sensor readings, the
program asks how many sensor stations are in the field and requests
values of the previous days total solar radiation, maximum and minimum air temperatures, rainfall, and photosynthe~callyactive radiation (PAR).These are necessary inputs to the PNUTGRO crop
model. If the user does not have a reading for PAR, the program
calculates an approximate value from solar radiation. (PAR is needed
in photosynthesis routines and influences specific leaf area. PAR is
not used directly for water balance calculations but does influence
root growth and root water uptake.)
If the program is being run for the first time, the user is asked
for the number of sensors being read at each station and for sensor
depths and lateral spacings in sequence. The program presents the
user with a graphical representation of the input sensor layout for
approval. Thisallowscorrection of mistakes the user might have
made in entering the layout. The user is then asked for the Julian
day time of day the sensors were read, and tension values for each
sensor in sequence. The program stores each stations readings in a
separate data file. Each data file keeps the latest 2 days of sensor
readings.
M e n all sensor data have been entered, the program sends
weather data to the input weather file for PNUTGRO at Julian day
1. The program uses weather data for a typical high stress day for
several days inthe future to forecast the earliest date irrigation should
be required (Ross, 1984). Irrigation trigger level is based on a userentered soil water tension value that is converted from water content
using a programmed soil water characteristic curve.
The number of future days that use weather data for irrigation
forecasting is variable. After the last forecast day data for a typical
T~omso~
season are entered into the datafile, If data without rain were placed
in the PNUTGRO weather file for the entire season, the crop would
show si@ficant stress during each run of PNUTGRO. Although the
crop stressing later in the season would not affect the short-term irrigation forecast, other useful results (such as yields) could not be
obtained. The program is thus designed to allow other PNUTGRO
variables such as yield to be examined as desired using a prior
weather file and daily updating.
To make later comparisons between soil water contents based
on sensor data andon the crop model, the program weights soil water
content and root length per unit area in the defined PNUTGRO soil
horizons to zones around thefield sensors. The program must do this
because PNUTGRO soil horizons will not be at the same depths as
zones that sensors represent. Even if the user consistently placed sensors at the midpoint of defined PNUTGRO horizons, those horizons
would not be wide enough to accommodate the field sensor's zone
of influence. Each fieldsensor has an assumed zone of influence that
is automatically expanded or contracted depending on itsproximity
to other sensors in the zone. The initial zone of influence was set to
14 cm based on experience and can be changed by the user.
A procedure in the program also performs error checking if the
user enters questionable weather data. If total radiation or maximin
temperatures are outside a reasonable range, the user is alerted to
take corrective action. This may involve checking the sensor or instrumentation. Under these conditions, the weather file is not modified until the problem is fixed. If the value for PAR is questionable,
the user is alerted, but the weather file is modified with a value calculated from the total radiation. If the user enters zero rainfall for
that day but observed an appreciable amount of rain, he or she is
told to check the data collection system or strume en tat ion. In this
case, the weather file is still mod3ied if rainfall was the only questionable input value.
eter A ~ j u s t m e ~
Routi~es
t
The parameter adjustment algorithms were developed for soil characteristic and root distribution functions based on my knowledge and
the experience of two other specialists. PNUTGRO was modified to
accept ET data from the field test so differences in sensor- and modelbased results could be attributed to different soil-water and rooting
characteristics only. Complete program listings and further details on
all PWTGRO Fortran code modifications conducted for this study
can be found in Thomson (1990).
rocess Sim~lation
167
where
0.9
.-.-
0.8
0.7
0.6
-4
0.5
0.4
0.3
0.2
0.1
0
0
20
80 60 40
100
120 160
140
(cm)
FIGURE4 Root weighting function WR(1) as a function of depth.
rocess Simulatio~
171
equations derived for this study (Thomson and Armstrong, 1987) are
potential sources of error.
If the sensor-determined rooting depthwas unequal to the
model-determined rooting depth, the procedure WRModif was activated. The procedure modified root weighting factors so that future
model-based representations of water use distributions in sensor
zones could converge on sensor-based representations. The equation
to modify the root weighting factors was
Wr,,,
= Wr ir F(Wr/Wr~*~)
where
Wr,,, = new root weighting factor
Wr = previous root weighting factor
F = adjustment factor (or "gain" term) = 0.075
Wrtot= sum of root weighting factors in root zone
As with modification of the DUL, the gain term was chosen to provide a conservative adjustment. This could be done because the root
weighting function was reasonablyclose to the "sensor-inferred"
function initially.
172
T~omso~
Drying occurs after a saturating rain where field soil moisture conditions are assumed to be uniform. Nonuniform soil moisture conditions occur after the first irrigation cycle, so sensor readings cannot
be averaged if the sensors were widely spaced.
For the present strategy, a primary requirement is that the system arrive at the last part of the field on time so the crop does not
stress. If calculations show that the center pivot cannot apply the
required amount of water in time, a lesser amount is applied proportional to the system capacity and the time required to reach the
end of the field. Theuser is told how much water should be applied,
how much will actually be applied, and the percentage timer setting
required to run the system.
.
To test the parameter adjustment methods of
SENPOS,
daily
PNUTGRO runs using the IREP weather data were made. The purpose of the test was to determine if model-based representations of
soil water potential converged on sensor readings as SENPOS adjusted the three selected model input parameters simultaneously. The
testalsoverified
proper operation of each parameter adjustment
routine.
l.
Drying periods July 26-30 and August 19-23 were selected to illustrate convergence of soil water potentials. To facilitate the test, shortcuts were taken to enter data quickly using a screen editor instead of
running the system in the user mode (illustrated in Fig. 3).
Data from a field test weather file were imported by PNUTGRO.
Soil water sensors began to show drying on July 27, Julian day 207
after a rain event. This was consistent with soil water drying trends
noted in the PNUTGRO output file. PNUTGROsown calculations of
soil water redistributio~after a rain event set the initial conditions.
Root weighting factors were set to values derived from examination
of soil water sensor data for the entire season. This is represented by
the extrapolated curve of Fig. 4 and served as the initial best guess
rooting function.
Soil water sensor data from the north lysimeter for July 27 and
28 were manually entered into a data file input to SENPOS. A weak
drying response was noticed from two shallow sensors placed in the
17
T~omso~
75
e l f - A ~ j u s t Process
i~~
I
0.13 I
0.12
0.11
0.10
0.09
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0
213212211210
237 236"
235 234
FIGURE 5 Drained upper limit 0)UL) value of the top fivesoil zones as
modified by parameter adjustment routine.
600
500
Sensors
.-
300
[CT
200
L.
-.-0
100
0
208
209
210
21 1
21 2
IGURE 6 Composite soil water tension results from sensors; model using
unmodified DUL, LL, and VVR(1); and model using daily modified DUL, LL,
and WR(1) for (a) July 26-30 and (b) August 19-23 in 24 cm zone of
regulation.
0.9
0.8
0.7
--
0.6
0.4
0.3
0.2
0.1
0
210
21 1
212
Julian Day
2;3
214
0.9
0.8
0.7
0.6
0.5
0.3
0.2
0.1
0
235
237
236
238
Julian Day
0-5cm
5-1 5 cm
15-30cm
30-45cm
45-60cm
60-90cm
IGURE 7 Root weighting factors in several soil zones modified by the parameter adjustment routine for
July and @) August.
208
209
2io
2i1
2i2
Julian Day
FIGURE8 Comparison of modeled soil water tension using three root functions to sensor tension at
8 cm depth and (b) 20 cm depth.
Test of l r r i ~ a t i o ~
Sc~e~~ler
130
100
90
70
.g
60
50
40
30
20
0
Julian Day
l80
2 0
290 270 250 230 210 ld0
220
240
260
280
300
Julian Day
ICURE 9 Soil water tension data at 8 cm depth for Watermark sensors in
(a) north lysimeter and (b) south lysimeter.
S e ~ f - A ~ j u s tProcess
in~
Si~ulation
(a)
Julian Day
T~omson
34
32 30 28 26 24 "Tii 22
f?i. 2018
16 -c
-:
E ;;r
10 8642-
I 8 0 I 200
230210190
(a)
220
240
1 260 I 280
250 290270
300
300
Julian Day
28 I
26 24 22 20 18 a, 1
25 6'
.-g 14
12108m
::
6 -
42 ' l80
210
190
200
220 I 240
250
230
260
290
270
2bO
Julian Day
11
water tension data at 40 cm depth for Watermark sensors
north lysimeter and (b) south lysimeter.
IGURE
in
Self-A~jutin~
Process Simu~ation
183
180 I
170
1
130
120
E
c
L
1 1 0 ~
100 t"
g
2
$j
90
80
70
60
50
40
30
20
10
0
(a)
Julian Day
12
11 10
9-
8 76 -
5 4 -
3211
27 0
280
Juiian Day
JGURE 12 Water applications (a) from rainfall and (b) from supplemental
irrigation.
adjustments have to be made under conditions like these when evidence is not very strong that a particular parameter needs adjustment. ~ u n n i n gthe system in a drier region could yield many more
drying cycles, allowing full corrections to occur early. This would
allow even more conservative adjustments than were already effected. A system that would work better under drier conditions is
desirable because d v regions are, of course, more critical with regard
to irrigation. For this study; it is fortunate that adjustments were
made because model-based predictions left uncorrected would have
aled many unnecessary irrigations.
Environmental conditions and characteristics of the soil at the
lysimeter site provided a stringent test of the method as evidenced
by the next irrigation event at day268. On this day; 0.931 cm of water
was applied to the lysimeters based on sensor readings and a composite trigger level of 60 kPa (Fig. 112b). There was some question as
to whether irrigation should be delayed because of forecast rainfall.
For this reason, it was decided that the amount of water applied
should be a little less than the amount needed to replenish a full 40
cm water regulation zone. (The sensors indicated significant water
uptake at the 40 cm depth as seen in Fig, 11.)Therefore, the amount
of water applied was calculated to be the amount needed to refill a
cm zone to field capacity(or the drained upper limit). This amount
could allow for additional recharge by rainfall. Irrigation was decided
on because sensor tension levels were climbing very rapidly due to
the low water retention characteristics of the illh hopper sand. Indeed, enough rain to replenish the zone fell within a couple of days
after irrigation commenced.
A run of the model also indicated that irrigation was warranted
on day268. This late in the season, parameters had long since settled
on their "static" values, and model-based soil water tension was apking the sensors quite well. However, the amount of wawas calculated to be 2.5 cm, an amount required to replenish a zone of 40 cm. The model does not presently account for
prevailing weather conditions and thus does not allow a deficit irrigation strategy. Theuser would need to override the calculated water
amount if less water were desired.
Although present parameter adjustment routines adjust only soil water parameters and the rooting function, evapotranspiration (ET)
should be considered because inaccuracies in modeling crop water
use can also influenceresults. Many trends can be observed by comparing model-based ET with measured ET.
Figure 13a compares simulated and measured ET rates over the
season. By examining trends over a shorter time scale, more detailed
observations can be made. Figure 13b illustrates an ET comparison
over the period August 12-29 (days 225-242). Data for days 231237 represent a drying cycle immediately after a high intensity rain,
and ET measured during this period should be close to potential ET
rates. During this period, measured ET was greater than simulated
ET by an average of 28%. This might be partially explained by sensible heat advection occurring over exposed lysimeters. The quantity
LE/ Rn (latent heat of vaporization of water times ET rate divided by
net radiation) can be used to ascertain sensible heat advection (Verma
and Rosenberg, 1977). If LE/ h > 1.0, sensible heat consumption is
occurring at the soil surface. Over a sunny period of days 230-237,
LE/Rn averaged 1.47. It showed a low of 1.09 on day 230 and a peak
of 1.85 on day 236. Measured ET was significantly higher than simulated ET during periods of sensible heat consumption.
All days where simulated ET was greater than measured ET
were characterized by low radiation loads (with or without rain). On
days 226 and 238, for example, weather was cloudy with no precipitation. PNUTGRO used a form of the Priestley and Taylor (1972) ET
model, which included an advection modification based on air temperature (Ritchie, 1985; Jones and Kiniry, 1986). Ritchie (1985) used a
multiplier of 1.1 on the ET equation to account for the effects of unsaturated air and increased the multiplier to allowfor advection
when the maximum temperature was greater than 24C. In PNUT6110,however, the multiplier was increased when the maximum
temperature exceeded 34C. Air temperatures on days 226 and 238
were well below this threshold, so potential ET on these days was
simulated by the unmodified Priestley-Taylor equation. A study by
Steiner et al. (1991) indicated that the unmodified Priestley-Taylor
model under predicted ET at low ET rates and overpredicted ET during periods of high evaporative demand. The authors suggested that
a more responsive advection function could improve predictions at
both ends of the ET spectrum.
If, in the future, rules are developed to modify the ET model,
methods should be applied separately to evaporation and transpiration components. This would be very important in the early stages
of crop development because the evaporation component is a large
part of the total ET. Although sensor-based feedback would be difficult to implement in the earliest parts of the season, information
T~ornso~
Measured
Modeled
"
l
8
a,
U:
.-0
a
v)
--_.
Measured
Modeled
225
227
229
231
233
235
237
239
241
could soon be obtained as roots begin to proliferate. Soil water sensors cannot provide reliable information on soil evaporation in the
shallowest soil horizons, but they can provide information on the
combined evaporation and transpiration in horizons where roots
proliferate.
Thornso
1
170
263
261
259
257
Julian Day
IGURE
Additional rules could be developed to elaborate on the characteristics of different sensor types. For example, my experience with
Watermark sensors indicates that the sensors themselves usually do
not "fail'' in the wet range of operation unless wires break or fray. If
one of these sensors does not show water use from drying and other
sensors at the same depth do indicate water use, for example, it is
much more likely that the sensor needs to be relocated to the zone
of active roots. Thesesensors show a very gradual degradation in the
wet range of operation characterized by readings that do not "return
to zero" when saturated (Pogue,1991). Additional decision rules
could be written to accommodate these cases.
It was also mentioned that tensiometers can "break tension" in
the dry ranges of operation (approximately 80 Wa). As coded, there
are no rules that presently accommodate this situation. Breaking tension would be an infrequent problem when scheduling irrigation in
a field with sandy soil. Usually; a crop is never allowed to approach
this stress level because a sandy soil holds very little water at this
tension level. Itis thus common practiceto irrigate much sooner than
80 Wa. By contrast, data from a field study undercorn in Shenandoah
County; Virginia (Fig. 15) showed deep rooting and tension levels
frequently exceeding 80 Wa in some horizons before irrigation was
called for. Figure 15aindicates when tensiometers broke tension and
rocess Si~ulation
80
Julian Day
I
Julian Day
URE
rocess S i m ~ ~ a t i o ~
reason, a farm manager may opt to leave them in the soil and install
new ones the next year.
PNUTGRO was modified to accept daily ET as input from the
field study so differences in sensor- and model-based results could
be attributed to different soil water and rooting characteristics only.
For actual application, ET cannot be an input, and analysis presented
in Thomson (1990) and herein indicates that the model tracked measured ET very well. Forthis study, modeled ET could have been used
with little influence on the results.
Care must be taken when adjusting parameters that influence
the same variable. The developer should be keenly familiar with the
processes involved and how variables interact. Many model runs and
sensitivity studies can increase the developer's knowledge and thus
allow intelligent examination of variable interactions. For this study
conservative adjustments to parameters were made to achieve the
desired goal. By analogy with control system theory, the output reflected somewhat of an "overdamped" response. This type of response was desired because, as indicated for the DUL, evidence was
not strong enough to warrant large adjustments. In general, a conservative approach is probably the best course to follow until the
knowledge base grows and more evidence (extracted and interpreted
from field data) is presented. I believe that the system was quite stable with the adjustment factors used.
A key factor in acceptance of this method is how universally
applicable it may be. Soil and rooting characteristics, for example,
vary widely with the soil and crop. Running this system on fields
with a few more soil-crop combinations would further strengthen
the knowledge base. Other soil characteristics that were not thought
to be significanthere might require consideration. A major advantage
is that once a robust knowledge base is constructed it can be selfcalibrating. That is,many exhaustive field trials should not be needed
because the rules should "know" what trends to look for. Machine
learning could be used to ensure that the proper parameter was adjusted. If, for example, an inappropriate adjustment was made, later
trends in data might give a clue that the adjustment was wrong. If
an adjustment to the correct parameter were then made and subsequent water status data showed more consistent tracking, the system
could recall this and apply the correct adjustment later if similar environmental conditions were encountered. Self-learning systems that
use neural networks might be used for the parameter adjustment
portion of this work and have shown promise in estimating model
outputs with limited input data (Altendorf et al., 1992).
19
Thornson
ocess
Self-Adjusting
195
Although the evaporation of water from land surfaces and from vegetation is a process of profound importance for agriculture, ecology,
and geophysics, mechanismsgoverning the process remained obscure
until relatively late in the history of science. In 1867, G. J. Symons
described evaporation as the most desperate art of the desperate
science of meteorology, and it was an art that made little progress
until well into the 20th century.
Reviewing the partitioning of available energy at the earths surface, Brunt (1939) wrote: A fraction, whose ~ a ~ i t u isd not
e ~eadil~
e s t i ~ a t e d[italics mine],is used in evaporating water from the surface
of grass, leaves, etc.and inhigh latitudes a portion is used in melting
ice and snow. A few years later, however, the first measurements of
fluxes of water vapor over uniform vegetation by Thornthwaite and
Holzman (1939) signaled the start of major advances in both experimental and theoretical aspects of evaporation science. Progress was
then interrupted by World War 11, and it was not until 1948 that the
publication of three seminal papers initiated a field of research that
is still expanding 50 years later.
197
onteit
At Rothamsted, England, Penman (1948) used a set of small lysimeters, installed many years previously, to measure evaporation
from open water, bare soil, and grass. By combining a thermodynamic equation for surface heat balance and anaerodynamic equation
for vapor transfer, he derived a general formula for the rate of evaporation from open water as a function climatic elements (temperature, vapor pressure, wind, and radiation) measured at screen
height. The key to this analysis was the elimination of surface temperature (a rarely measured and often unmeasurable quantity) which
was achieved by assuming that saturation vapor pressure was a hear function of temperature over small ranges. If Penmans equation
had been published a few decades later it would almost certainly
have been known today as Penmans model!
Working independently in the USSR, Budyko (1948) used the
same primary equations as Penman but estimated surface temperature by a process of iteration, thereby deriving estimates of evaporation more accurately than Penman but more laboriously. (This1948
monograph was an English translation of the original Russian text
that predated Penmans classic paper.)
Thornthwaite (1948) proposed a completely differentsolution to
the problem of estimating evaporation from watersheds in the United
States, given minimal climate records. Like Penman, he argued that
evaporation was driven primarily by energy available fromradiation,
but, for the practical reason that very few measurements of radiation
were available, he used air temperature as a surrogate variable. This
argument is still valid in some circumstances, and a recent prediction
of global evaporation by Mintz and Walker(1993) used Thornthwaites method in preference to Penmans.
This review of evaporation model opens with a brief description
of Thornthwaites formula but deals mainly with developments from
the theoretical base established by Budyko and Penman. One major
concept emerging from their work was the distinction between potential rates of evaporation from vegetation or soil obtained when
the supply of water was abundantand actual rates prevailing
when the supply was restricted.
After examining a large number of hydrological records for climatically contrasting sites in the United States, Thornthwaite concluded
that the potential evaporation in any month could be correlated with
199
~va~oration
Models
E = 16(10T/I)
(2)
where
a = (0.67513 - 77.11
This set of equations contains eight numerical constants, all determined empirically and some quoted with unwarranted precision. As
needed, the constants were modified to allow for the dependence of
daylength on latitude and for monthly temperatures outside the
range 0-26.5C.
In a well-balanced critique of Thornthwaites formula, Pelton et
al. (1960) concluded that it was most reliable for periods of a month
or longer because, over a year and at many sites, monthly mean temperature and potential transpiration were strongly correlated. However,
Because of the severe limitations of mean temperature methods, they will yield correct s ~ o r t - ~ e r estimates
io~
of evapotranspiration only under fortuitous circumstances and cannot be
relied on for general use . . . To expend effort improving mean
temperature methods for short-period potential evapotranspiration estimates and adjusting the estimates to a given
locale and crop when mean temperature has little or no physical foundation for this use appears to be kicking a dead
horse.
Despite this unequivocal warning, agronomists and engineers,
especially in the United States, continued to refine and use empirical
models for estimating potential evaporation from mean temperature
(Blaney and Criddle, 1962) or diurnal temperature range (Hargreaves
and Samani, 1982). When locally calibrated, they perform well considering how little input they need as shown in a recent comparison
by Allen et al. (1994). However, becausethey are products of current
climate they are prone to error when used to predict changes of evaporation rate as a consequence of global warming, particularly in dry
climates as shown by McKenny and Rosenberg(1993).Empirical
o~teit
Starting from the basic physics of heat and vapor transfer, Penman
(1948) developed a mechanistic model for evaporation from free water surfaces and extended it empirically to bare soil and grass. This
model was a practical application of the First Law of Thermodynamics which requires that the net rate at which radiant energy is absorbed at any point on the earths surface, often written R, (W/m2),
must be balanced by the rate at which energy is (l)transformed to
latent heat of evaporation, XE (W/ m), where h (J/ g) is the latent heat
of evaporation for water and E [g/ (m. S)] is a mass flux of water
vapor, or (2) dispersed to the atmosphere by convection, or (3) stored
in a substrate (soil, rock, water, etc.). (Energystored by photosynthesis, essential for growth, is energetically negligible.)
Because instruments for measuring R, were not available in the
1940s (and are still very rare at climatological stations 50 years later),
Penman calculated this quantity from its components. Incoming
short-wave (solar) radiation was estimated from hours of sunshine
using a relation derived by Angstrom [cited by Brunt (1939)] and net
long-wave radiation was estimated from air temperature and vapor
pressure using an empirical formula derived by Brunt.
Further empiricism was needed to obtain (1) a vapor transfer
coefficient (the flux of water vapor per unit difference in vapor pressure between a water surface and the air passing over it at some
arbitrary height) and (2) a corresponding heat transfer coefficient. For
both heat and vapor, the transfer coefficient used by Penman was
obtained from measurements of evaporation from small (76 cm diameter) evaporation tanks and had the general formf(u) = a ( l + bu),
where a and b were empirical constants and u was wind speed measured ata standard height of 2 m. This function was used to estimate
fluxes of both sensible heat and latent heat from a surface, given air
temperature and vapor pressure (as measured at the height of a climatological screen) and the egective temperature and vapor pressure
of a wet surface.
By eliminating surface temperature and vapor pressure from this
set of equations, Penman was able to express the latent heat of evaporation from a wet surface as
XE = (AH + YE,)/ (A
where A is the rate of change of saturation vapor pressure with temperature (which increases rapidly with temperature from 45 to 145
Pa/K between 0 and 2OOC) and y is the psychrometer constant in the
same units of pressure per unit temperature difference. (The factthat
this constant is a weak function of both temperature and pressure
is usually ignored. At a standard pressure of 101.3 Wa, it increases
from 65 to
Pa/K between 0 and 20C.
In Penmans 1948 paper, H was the amount of energy per unit
of ground surface as received from radiation, less a fraction stored in
the soil. (A few workers, including the writer, still observe this tradition, but most use the symbol H for sensible heat flux.) Penman
assumed that over periods of a week or longer, soil heat
storage could
be neglected compared with the net amount of heat received from
radiation.
Using contemporary units, the composite variable E, can be
defined as
E a
PADftU)
(5)
Penmans model combined robust thermodynamic and aerodynamic principles and provided a sound basis for theoretical developments in evaporation science. Itcould not be used in practice, however (to estimate irrigation need, for example), without recourse to a
substantial amount of empiricism.WhereasThornthwaites model
contained eight empirically determined constants, Penmans, in its
original form contained six-four for estimating net radiation from
The water vapor content of air is conveniently expressedas dimensionless specific
humidity q (kg of water vapor per kg of moist air), related to vapor pressure by
=
[ ( p - e)
==
where = 0.622 is the ratio of molecular weights for
water vapor and air, e is vapor pressure, and p is air pressure in the same units.
The approximation is valid because e is invariably two orders
of magnitude less
than p . The parameters A and y then assume units of K- and y = c,/X (K-), where
cp[J/ (kg-K)] is the specific heat of air at constant pressure.
202
~onteit
sunshine hours and vapor pressure and two for the wind function.
A seventh constant was needed to estimate the "actual" rate of evaporation from soil or vegetation as a fraction of the "potential" rate
for open water in the same environment. Subsequent theoretical and
instrumental developments, now reviewed, made much of this empiricism unnecessary.
A major generalization of Penman's original equation was made possible by introducing "resistances" -simple electrical analogs for the
potential difference needed to drive unit flux in systems that involve
the transport of momentum, heat, and water vapor (Monteith and
Unsworth, 1990). (Such resistances have dimensions of time per unit
length and are usually quoted in units of seconds per meter or per
centimeter.)
What subsequently became known as the en man-Monteith
equation (PME hereafter), incorporating both an aerodynamic resistance Y , and a surface or stomatal resistance Y,, was first published in
a somewhat obscure contract report from a group led by F. A. Brooks
at Davis, California (~onteith,1963), and a more extended treatment
followed (Monteith, 1965). An almost identical contemporary model
was developed wholly independently by Rijtema (1965).
In the PME, the empirical wind function f(u) used in Penman's
equation was replaced by the reciprocal of an aerodynamic resistance
so that the second term in the numerator became
E: =
(7)
+ 7")
(9)
The appearance of a wind-dependent function in the denominator as well as in the numerator of Eq. (9) implies that the rate of
evaporation calculated from the PME is always less dependent on
wind speed than the rate from the corresponding Penman equation
when other elements of climate are unchanged. In general, estimated
rates are usually insensitive to the magnitude of r,, and the error
203
= pc,(A-'
+ y")D/H
(10)
X/H = A/(A
gives
+ 7)
It follows that when the surface resistance is close to the value set by
Eq. (lo), the evaporation will be insensitive to wind speed and close
to AH/(A + y) (Fig. l; see also Section IV).
90
IGURE
Dependence of latent heat of evaporation hE on aerodynamic
resistance r, and surface resistance r, when total available heat H = 105
W/m2, saturation vapor pressure deficit D = 0.2 @a, and A = 0.15 W a / K
In mostcircumstances, evaporation rate decreases as Y, increases (top
curve). However, there is critical value of Y, (90 s/m in this case) at which
h E / H = A/(A +
[see Eq. (15)] so that E is independent of
When r, is
smaller than this value, the evaporation rate increases as ra increases.
Because the net heat term AH is usually several times larger than AE,,
it is the main source of error in the numerator of the PME. This error
has three components: the net flux of radiation received by vegetation
over the period of application; the net amount of heat stored in the
vegetation; and the net amount of heat stored in the soil. Forfractions
of a day or for forests with a large heat capacity, the magnitude of
the storage term must be evaluated and retained if it is comparable
with the net radiative flux over the same period [see Thom (1975),p.
93 et seq.].
Net radiation can either be measured directly or estimated from
its components. Unfortunately, net radiometers are notoriously prone
to several types of error, mainly associated with changes in the transmission properties of polythene domes during exposure (Field et al.,
1992; Duchon and Wilk, 1994). Climatological measurements of net
radiation are therefore both rare and unreliable. Records of solar radiation are more trustworthy and more readily available but must be
complemented by estimates of long-wave incoming radiation (usually
obtained from empirical formulas) and from outgoing radiation assumed equal to blackbody radiation at air temperature because the
radiative surface temperature is rarely available. However,the difference between air and surface temperatures can be accounted for by
defining an effective radiative resistance as rR = pcP/4aT':, where a is
Stefan's constant and T',is the air temperature in Kelvin (Monteith
and Unsworth, 1990). Thevalue of this resistance decreases from245
s / m at 10C to 186 s / m at 30C.If the effective resistance for heat
transfer by convection and radiation, treated as parallel processes, is
taken as (ri' + ri')-', it is legitimate to estimate net radiation as the
flux density Xni (W/mz)for a surface at air temperature Ta ("C) (sometimes referred to as the "isothermal" net radiation). For clear skies
an appropriate algorithm for the long-wave component of net radiation .Rnl (Monteith and Unsworth, 1990) is
Rd
(12)
= 107 - 0.3Ta
c.
Within comprehensive models of crop growth and yield, the PME
can readily be exploited as a submodel for transpiration either at the
scale of a single leaf or at canopy level provided the relevant physical
and physiological variables are specified a priori or are available as
output from other parts of the model. Following an exhaustive comparison of lysimeter measurements of evaporation and predictions
from a range of formulas, Allen et al. (1994) concluded that the PME
was more reliable than other models over a wide range of climates.
The PME has therefore been adopted by FAO as a substitute for the
Penman equation modified by Roorenbos and Pruitt (1977) in an irrigation manual usually referred to as FAO-24. Several other evaporation models, including the original Penman equation, fitted Allen's
measurements data almost as well as the PME and can therefore be
confidently used for practicalapplications when the surface resistance
is known.
Doorenbos and Pruitt altered Penman's original formula by introducing an alternative wind function that predicts a smaller value
of the aerodynamic resistance at a given wind speed. However, Thom
and Oliver (197'7) demonstrated that an error in the original wind
function (which was appropriate for a smooth water surface rather
than for rough vegetation) fortuitously compensated for the assumption that the surface resistancewas effectively zero. Eventually, Allen
et al. (1994), using measurements from high-class lysimeters, demonstrated conclusively that when this compensation was removed,
the "improved' formula overestimated evaporation by 10-30% (the
precise figure depending on the relative magnitude of radiation and
humidity deficit terms).
Since 1981, the British Meteorological Office has used the PME
operationally over the whole of Britain to provide weekly and
monthly estimates of evaporation and soil water deficit (Thompson
et al., 1981). The correct type of wind function was used along with
empirical relations between surface resistance, leaf area (Grant,1975),
and soil water potential (Russell, 1980).
206
Single Leaves
For the simplest case of a single transpiring leaf, the total resistance
to the diffusion of water vapor from substomatal cavities to the external airstream is analogous to an electric current passing through
two resistances in series, usually referred to as "stomatal" and "aerodynamic" resistances.
The magnitude of the stomatal resistance can be estimated in
principle from the number of stomata per unit leaf area and from the
mean diameter and length of pores. These are the main variables of
diffusion models for stomata (e.g., Penman and Milthorpe, 1967), but
because they are rarely known and are difficult to determine with
precision, stomatal resistances are usually calculated from measured
rates of transpiration and estimated gradients of vapor concentration.
When stomata are closed,leaves of most species continue to lose
water very slowly by diffusion through waxy cuticles, but their resistance is so large compared with stomatal resistance that it is usually assumed to be infinite,
There is an aerodynamic resistance r, to diffusion in the boundary layer surrounding the leaf within which the transfer of heat, water
vapor, etc. proceeds at a rate governed by molecular diffusion. Provided the wind speed is great enough and the temperature difference
between leaf and air is small enough to ensure that transfer processes
are not affected by gradients of air density, the boundary layer resistance depends on air velocity and on the size, shape, and attitude of
the leaf with respect to the airstream. In very light wind, however,
rates of transfer are determined mainly by gradients of temperature
and therefore of density, so that the aerodynamic resistance depends
more on the mean leaf-air temperature difference than on wind
speed. Empirical formulas for estimating leaf boundary later resistance as a function of wind speed and characteristic dimension can
be obtained from the engineering literature as presented by Monteith
and Unsworth (1990) and by Jones (1992), for example.
To obtain the correct total (stomatal plus aerodynamic) resistance to vapor diffusion, rst, it is necessary to distinguish between
"amphistomatous" leaves with stomata on both abaxial (ab) and adaxial (ad) surfaces and "hypostomatous" leaves with stomata on adaxial surfaces only. Forthe amphistomatous class, the total resistance
is found by combining the resistance for the two surfaces in parallel
to give
For the h~ostomatousclass, the resistance is simply Y ,
+ r,.
Models
207
+ Xo.25)
(15)
and
or
zo = 0.3h(l
d/h)
when
0.2 S X
1.5
(16b)
(1990), for example], whereas zo may either decrease or increase depending on changes in foliage density. Businger(1956) suggested that
Tanner and Pelton (1960) confirmed that the reciprocal of raMis equivalent to the wind function f(u) in Eq. (5).
When foliageis warmer or cooler than the air above it, mechanically generated turbulence is respectively enhanced or inhibited by
the action of buoyancy. For these so-called non-neutral conditions,
Eq. (17a) can, in principle, be modified by introducing a stability parameter such as the Richardson number or the M o n i n - ~ b u ~ o v
length (Thorn, 1975; Monteith and Unsworth, 1990), but the former
requires measurements of temperature and wind speed gradients that
are not generally available and the latter requires a sensible heat flux
that cannot be obtained without iteration.
An aerodynamic resistance for the transfer of scalars such as
heat or water vapor can be obtained by adding to the momentum
resistance raMan additional component rb to account for the absence
of a process equivalent to the blufiody forces that enhance momentum transfer ( ~ o n t e i t hand Unsworth, 1990). For arable crops and
with U* in the range 0.1-0.5 m/s, Thorn (1975) suggested the empirical relation rb = 6 . 2 ~ , . ~Alternatively,
~.
heat and vapor resistances
can be obtained directly from q. (17a) by replacing zo with scalarecific roughness lengths SUC as zoE3for heat or zov for vapor in
place of the omentum roughness usually written zo. Equation (17a)
can then be rewritten for heat or water vapor as
Garratt (1992) suggested that ln(zo/zoII) ln(zo/zov) 2 was appropriate for diverse types of vegetation, but from the analysis of many
measurements over short grass, Duynkerke (1992) reported values of
ln(z,/zoE.3)
r a n ~ n gfrom about -2 to 13 and increasing both with leaf
area index and with U*. Precise modeling of evaporation rates therefore requires a careful assessment of differences in the flux- adient
relation for heat, for vapor, and for momentum.
Because it is possible to identify an aerodynamic resistance for
extensive uniform vegetation [using Eq. (17a)], it is also possible to
identify a surface resistance imposed by stomata, assuming that a
complete canopy is endowed with the properties of a single big
leaf. As for individual leaves, the surface resistance of a canopy
is usually estimated indirectly from fluxesand gradients using a procedure discussed in the next section.
n v e ~ s i oof
onteith
~ the
E ~ ~ ~ t i o n
Whereas the original Penman model has been used mainly to estimate the irrigation need of crops or seasonal changes in the water
balance of catchments, a common use of the PME is to estimate the
surface resistance of vegetation when its rate of ~anspirationhas been
measured independently The magnitude of this resistance can then
be used as anindex of ground cover, water stress, nutrient deficiency
etc. The procedure is to invert Eq. (9) to obtain
Evaporation Models
where D, is the specific humidity deficit at the leaf surface and Dsois
a scaling factor. As the transpiration rate E is given by gD, (assuming
21
o~teit
negligible temperature gradient across the cuticle), the stomatal component of resistance is given by
For the common condition that g is much larger than go, rearrangement of Eq. (22) gives
0.5
""".
c
onteith
I
~ - - r
The model used above to explore the exchange of heat and water
between vegetation and the atmosphere was essentially an electrical
analog containing a network
resistances but neglecting capacitances-elements of the system capable of storing heat or water for
periods usually longer than an hour but less than a day. Their properties are now considered briefly.
1 . Heat Storage i n
and Vegetation
Diurnal changes in the storage heat in the soil are needed for precise hourly estimates of evaporation using Penman-type equations.
21 5
Water Storage
Passioura (1983) suggested that when the roots of a uniform crop start
to extract water at a specified depth, the volumetric water content 0
can often be expressed as a negative exponential function of time,
e.g.,
0 = 0, exp(--t/.r)
(24)
Models of extraction of water by plant roots are still at a primitive stage of development compared with models for the evaporation
of extracted water. When the rate of evaporation is determined by
weather, this disparity is usually tolerable, but when growth is limited by the availability of water to roots, most models resort to somewhat crude empiricism and predictions are therefore unre~able.
3.
hen rain starts to fall on vegetation, leaves and other surfaces intercept a fraction that depends on factors such as area, angular distribution, and roughness. Eventually the rate of interce~tionis balanced by the rate of loss by dripping from wet surfaces, and the
a ~ o ~ ofn water
t
retained after saturation is known as the canopy
storage capacity (S), a quantity usually lying in the range 0.5-2.0 mm.
Rutter et al. (1975), Calder (1977, 1986), and others explored the relation between interception and evaporation rates by writing the net
~ )a balance between (l)
rate of change of intercepted water ( ~ I / das
the rate at which water is intercepted, assumed to be a fraction of
the pre~ipitationrate P, and (2) the rate at which it is subsequently
lost either by evaporation ( E ) or by dripping [assumed to be a function f(I) of I ] .It follows that
and is therefore independent of wind speed [see Eq. (ll)].More generally provided the saturation deficit of air is constant with height
so that air in contact with foliage has the same value ( D ) as air at
screen height, the latent heat of evaporation is
When this equation is used to eliminate D from Eqs. (8) and (IO),
rearrangement of terms again gives Eq. (26).
In practice, saturation vapor pressure deficit is not constant with
height. In the absence of cloud and for reasons outlined below, it
increases with height within a convective boundary layer (CBL) that
is often several hundred meters deep at dawn, expanding to several
thousand meters in depth as the day progresses.
The CBL is capped by an inversion through which relatively
warm dry air is mixed into the boundary layer from above. At the
ground, however, air passing over vegetation receives water vapor
by transpiration and may be heated or cooled depending on the difference between surface and air temperature. Commonly the net flux
of water vapor into the CBL is small, so vapor pressure changes little
during the day. In contrast, temperature invariably increases in response to radiative heating and to an input of sensible heat from
above the CBL that is modulated but never reversed by sensible heat
exchange at the ground. An important consequence of these intimately linked processes is that saturation vapor pressure deficit increases with height in such a way that rates of transpiration exceed
predictions from Eq. (26) by 20-30%. To deal with this excess empirically Priestley and Taylor (1972) suggested that the equation should
be modified by introducing a multiplier (a)such that
21
o~teit
S1 = [l
+A+y
(")l
r,
-1
ary Layer
+ my,
a- =
(33)
N (Eq. 29) is expected to increase as surface resistance increases(during a spell of dry weather, for example) because of an increase in
a!
100
surface r e s i s ~ (SI
~c~
Reciprocal of the Priestley-Taylor coefficient, a, estimated as a
function of surface resistance by McNaughton and Spriggs (1989), who coupled a boundary layer model to meteorological measurements for 9 days
over grassland at Cabauw, Netherlands (Driedonks, 1981). Nine lines join
sets of points at seven arbitrary values of surface resistance assumed constant
throughout the day so that each set represents daily mean estimates (a)-1
as
a function of resistance for daylight hours on 1 day. Sets of points for each
day are displaced vertically to avoid confusion. Values of for zero resistance ranged from 1.30 to 1.57 with mean of 1.40 0.1.
1 + aornrs
(35)
1t-N-X
l + f - a&
where X = f /mr, can be regarded as a climatological variable depending on mean air temperature and wind speed and on the struc-
The rate at which water evaporates from wet soil depends partly on
the state of the atmosphere as specified by radiation, wind speed,
humidity, etc. and partly on the state of the soil as established by the
distribution of water held in pores and by vertical gradients of temperature and vapor pressure. Fully comprehensive models of evaporation from soil, as described by ten Berge (19901, for example, account rigorously for the way mass and heat transfer are coupled
within soil and at the soillair interface.
illel (1976) described a representative model of
soil evaporation in which a loam soil with a depth of 1.13 m was
treated as 14 layers of increasing depth starting with a l cm layer at
the surface. Soil water potential and hydraulic conductivity were
specified functions of water content, and fluxes of water and heat
were assumed to be independent rather than coupled as in the more
rigorous model of Philip and de Vries (1957). The short-wave reflectivity and long-wave emissivity of the soil surface were functions of
onteit
222
o Bingharnton
Phoenix
0
0
0
3
**
**
0.4
0.8
FIGURE5 Rate of evaporation from bare soil estimated from the model of
Van Bavel and Hillel and plotted (as fractions of maximum rate) against the
inverse square root of time fromthe onset of drying. Sites chosen for climatic
contrast were Binghamton, New York, and Phoenix, Arizona. (From Van
Bavel and Hillel, 1976.)
the water content of the top layer, and the aerodynamic resistance
between the surface and a reference height of 2 m was found by using
conventional equations incorporating wind speed, roughness, and a
stability correction.Climate records for seven contrasting stations
within the United States were used to explore the behavior of the
model when run with an hourly time step in a CSMP program. During the so-called "first stage" of drying when the soil surface was
continuously wet, daily values of potential evaporation estimated by
the model were very close to values obtained from an equation of the
Penman type.
Lee and Pielke (1992) and Mihailovic et al. (1995) reviewed a
range of empirical models in which the evaporation rate was estimated from the aerodynamic resistance r, between a uniform surface
and the air at a reference height z;and from the difference between
the specific humidity of air at z
and atthe soil surface,assumed
to be thesaturation value
at mean surface temperature. In these
models, the reduction of evaporation rate during the second stage of
E v a ~ o r a t i oModels
~
23
drying is taken into account by introducing an empirical factor depending on water content at the surface. Some workers suggest that
the surface humidity should be reduced by a fraction so that
Others place outside the brackets implying that the surface water
content determines the evaporation rate directly rather than indirectly
through the surface humidity In this second category; the model developed by Deardorff (1978) performed well in a comprehensive field
test of 9 models by Mihailovic et al.
Less sophisticated methods can be used to estimate daily rather
than hourly rates of evaporation from soil. For example,
a simple
linear relation between evaporation rate and the square root of time
(in days) is characteristic of a system in which an upper layer of soil
is treated as completely dry anda lower layer as saturated(Monteith,
1981). Predictions from such a model are very similar to those obtained from the complete Van Bavel-Hillel model (Fig. S). In particvc3 rate of actual evaporation from a hot dry station
(Phoenix, Arizona) was less than the rate at a humid station (Binghamton, New York) whereas the potential rate of evaporation was
greater. This is because the soil resistance increases more rapidly at
the dry site.
"""""-""""""""H
"
"
"
"
"
"
"
"
"
"
"
"
"
25
226
onteit~
from the corresponding profile. The essence of his theory is a distinction between the systematic movement of small but discrete elements water vapor close to the point of evaporation ("near-field"
behavior) and the subsequent random movement of elements under
the influence of turbulence ("far-field' behavior).
Van den Hurk and ~ c ~ a u g h t o(1995)
n
have shown that nearfield behavior can be simulated by introducing an additional resistance in a lumped canopy model of the type described by Shuttleis defined asthe increase in
worthand Wallace.Theresistance
concentration per unit
attributable to displacement within the
near field. In Fig. 7, such a resistance would therefore be added in
series with the resistance. Van den Hurkand ~ c ~ a u g h t tabulated
on
plausible values of the additional resistance that lie mainly between
0,3/u" and 0.41U", where U" is the traditional friction velocity. They
then showed that predicted rates of evaporation from the canopy and
from wet soil beneath are not very sensitive to the introduction of the
near-field resistance becauseit is an order of magnitude less than the
resistances already included in the Shuttleworth-s all ace model.
FIGURE7 Resistance network for transfer processes in vegetation incorporating a "near-field'resistance r, in series with a bulk boundarylayer resistance rb (equivalent to r: in Fig. 6) and a physiological resistance r, (After
van den Hurk and ~cNaughton,1995.) (C is the concentration of an entity
at screen height, vegetation height, or the soil surface, subscripts R, v, and
S, respectively. F is a
from the soil, from vegetation, or from the total
soil-vegetation system, subscripts S , h, and t, respectively.)
~ v a ~ o r a t i oModels
n
227
In earlier sections of this review, it was assumed implicitly that measurements of climate obtained at some standard height and subsequently used to estimate evaporation rates were representative of the
surrounding area. This assumption is invalid when the physical properties of the surface are discontinuous as a result of major changes in
vegetation height, for example, orin the availability of water following irrigation. Several models have been developed to explore
changes in evaporation rate and temperature when air in equilibrium
with a relatively hot dry surface such as bare soil encounters the cool
wet surface of irrigated vegetation. In early work (e.g., Philip, 1959,
1987; McNaughton, 1976), surface resistance was assumed to be independent of the imposed climate, but Kroon and de Bruin (1993)
described numerical model in which resistance was a function of
vapor pressure deficit and temperature and in which changes of surface roughness were accounted for.
Itier and Perrier (197'6) and Brunet et al. (1994) derived a function for the change in concentration of a scalar downwind of a step
change in the scalar flux. Theyapplied this to the case of an irrigated
field downwind of dry land with step changes in temperature and
humidity both specified and with the net flux of heat assumed constant throughout the system. They also argued that the evaporation
rate downwind of the discontinuity could be assumed constant because many laboratory measurements appear to show that stomatal
resistance increases with saturation vapor pressure deficit. However,
it is more likely that stomata respond to the flux of water vapor
through them than to a vapor pressure deficit [Eq. (23)].
Crop canopies emit radiative signals that can be used to derive information about their rates of transpiration and water status. Atten-
o~teit
=L
E = [H - Ta)/r,]/h + W
(39)
where H is total heat flux and W is an error term that includes the
soil heat flux. According to Choudhury and de Bruin (1995), typical
values of these parameters estimated by remote sensing over periods
on the order of a day are W = -0.1 mm/day (range -0.1 to +0.1
mm/day) and pc,/ra = 0.37 mm/(day .K) [range 0.25-0.64 mm/
(day. K)]. This technique is valid only when the surface is unobscured
by cloud, and the authors warn that "correction of infrared temperature observations for the surface emissivity and atmospheric effects
remains a challenging problem." Another unresolved problem is the
existence of systematic differences between radiative surface
ature and aerodynamic surface temperature as observed by
and ~ o n t e i t h(1986), for example.
By eliminating E from Eqs. (9)and (39), the aerodynamic surface
temperature of uniform vegetation can be expressed as
(41)
where Tr0is a maximum surface temperature estimated by extrapolation for the (unattainable) condition D = 0. Once the parameters of
this relation have been established, it can be used to detect anomalously large values of surface temperature that are an indication of
water stress.
Idso (1987) later suggested that because this simple linear model,
was valid within narrow limits for a wide range of crop types, the
canopy resistance must be effectively constant during much of the
day, in conflict with extensive evidence from porometers. However,
it can be shown that the closure of stomata in response to increasing
demand for water [Eq. (23)] is compatible with a linear relation between Te and D.
This type of analysis is valid only when ground cover is complete so that radiative signals from foliage are not contaminated by
emission from the soil.Moregenerally, the spatial distribution of
evaporation rate and of surface resistance can be estimated from satellite images that provide estimates of ground cover, surface temperature, and reflectivity, supplementing surface-based measurements of
air temperature and aerodynamic resistance (e.g., Bastiaasen et al.,
1994).
Allen, R. G., M. Smith, A. Perrier, and L. S. Pereira. 1994. An update for the
definition of reference evapotranspiration. International Commission for
Irrigation and Drainage Bulletin 43, No. 2.
Ball, J. T., I. E. Woodrow, and J. A. Berry. 1987. A model predicting stomatal
conductance. In: Progress in P h o ~ o s ~ n ~ h eResearch.
sis
Vol.IV. (J. Biggins
(Ed.). Dordrecht: Martinus Nijhof, pp. 221-234.
230
onteit
Bastiaasen, W. G. M., D. H. Hoekrnan, and R. A. Roebeling. 1994. A methodology for the assessment of surface resistance and soil water storage
variability at mesoscale based on remote sensing measurements. IAHS
SpecialPublication No. 2. Wageningen:Wageningen Agricultural University.
Blaney H. F., and W.D. Criddle. 1962, Determining Consumptive Use and
Irrigation Water Requirements. Tech.Bull.1275.Washington,DC:Agricultural Research Service, USDA.
Brunet, Y., B. Itier, K. J. McAnaney and J. P. Lagouarde. 1994. Downwind
evolution of scalar fluxes and surface resistance under conditions of local
advection. I. A reappraisal of boundary conditions. Agric. Forest Meteorol.
71:211-225.
Brunt, D. 1939. ~hysicaland Dynamical Meteorolo~.London: Cambridge Univ.
Press.
Budyko, NI. I. 1948. vapora at ion Under Natural Conditio~s.Leningrad: GIMIZ.
English Transl.: Jerusalem: Israel Program Sci. Transl., 1963.
Bunce, J. A. 1997. Doestranspiration control stomatal response to vapor pressure deficit? Plant, Cell and Env. 20:131-135.
Businger, J. 1956. Some remarks on Penmans equation for evapotranspiration. Neth. J. Agric. Sci. 4:77-80.
Calder, I. R. 1977. A model of transpiration and interception loss from a
spruce forest: in P l y ~ m o n Central
,
Wales. J. Hydrol, 33:247-265.
Calder, I. R. 1986. The influence of land use on water yield in upland areas
of the U.K. J. H y ~ r o 88:201-211.
~.
Choudhury, B. J., and H. A. R. de Bruin. 1995. First order approach for estimating unstressed transpiration from meteorological satellite data. Adv.
Space Res. 16:10167-10176.
Choudhury, B. J., and J. L. Monteith. 1988. A four-layer model for the heat
budget of homogeneous land surfaces. Quart. J. Roy. Meteorol. Soc. 124:
373-398.
Deardorff, J. W. 1978. Efficientprediction of ground surface temperature and
moisture with inclusion of a layer of vegetation. J. Geophys. Res. 83:18891903.
Dolman, A. J. 1993. A multiple-source surface energy balance model for use
in general circulation models. Agric. and Forest Meteoroz. 65:21-45.
Doorenbos, J., and W. 0. Pruitt. 1977. Crop water requirements. Irrigation
and Drainage paper, No. 24 (revised). FAO, Rome.
Dougherty R. L., J. A. Bradford, P. I. Coyne, and P. L. Sims. 1994. Applying
an empirical model of stomatal conductance to three C-4 grasses. Agric.
Forest Meteorol. 67269-290.
Driedonks, A. C. M. 1981. Dynamics of the Well-Mixed Boundary Layer. Sci.
Rep. WR 81-2, ISNMI, de Bilt, Netherlands.
Duchon, C. E., and E. Wilk. 1994. Fieldcomparisons of direct and component
measurements of net radiation under clearskies. I; Appl. Meteoroz. 33:
245-251.
Duynkerke, P. G. 1992. The roughness length for heat and other vegetation
parameters for a surface of short grass. I; Appl. Meteorol. 32:579-586.
Field, R. T., L. J. Fritschen, E.T. Kanemasu, E. A. Smith, J. B. Stewart, S. B.
Verma, and W. P. Kustas. 1992. Calibration, comparison and correction of
net radiation instruments used during FIFE. 1. Geophys.Res. 97017):
18681 -18695.
Frenzen, P. and C. A. Vogel.1995. A further note on the magnitude and
apparent range of vanation of the von Karman constant. B o ~ n ~ a ~ - L a y e r
Meteorology 75315-317.
Garrat, J. R. 1992. The A ~ ~ o s p h e rBi c o ~ n Layer.
~ a ~London: Cambridge Univ.
Press.
Gash, J. H. C., W. J. Shuttleworth, C. ]R. Lloyd, J. C. Andre,and J, P, Goutorbe.
1989.Micrometeorological measurements in Les Landes forest during
Hapex-Mobilhy Agric, Forest. Meteorol. 46:131-147.
Grant, D. 1975. Comparison of evaporation from barley with Penman estimates. Agric. Forest Meteorol. 25:49-60.
Hargreaves, G. H., and Z. A. Samani. 1982. Estimating potential evapotranspiration. Tech. Note. I; Irrig. Drainage, ASCE 108:225-230.
Huband, N. D. S., and J. L. Monteith. 1986. Radiative surface temperature
and energy balance of a wheat canopy. I. Comparison of radiative and
aerodynamic canopy temperature. B o ~ n ~ a ~ - Meteorol.
~ a y e ~36:l-17.
Hunt, E. R., S. W. Running, and C. A. Federer.1991. Extrapolating plant
water flow resistances and capacitances to regional scales. Agric. Forest
Meteorol. 54:169-195.
Idso, S. B. 1982. Non-water-stressed baselines: A key to measuring and interpreting plant water stress. Agric. Meteorol. 2759-70.
Idso, S. B. 1987. An apparent discrepancy between porometry and infrared
thermometry relative to the dependence of stomatal conductance on air
vapour pressure deficit. Agric. Forest Meteorol40:106--107.
Itier, B., and A. Perxier. 1976. Presentation d u n e etude analytique de l'advection. I. Advection liee aux variations horizontales de concentration et
de temperature. Ann. Agron. 27111-140.
Jarvis, P G. 1976. The interpretation of the variations in leaf water potential
and stomatal conductance found in canopies in the field. Phil. Trans, Roy.
Soc. Lond. B 273:593-602.
Jensen, M.E., and H. R. Haise. 1963. Estimating evapotranspiration from
solar radiation. I; Irrig. Drainage, ASCE 89:15-41.
Jones, H. G. 1992. Plants and Microclimate (2nd Ed.). Cambridge University,
Cambridge.
Jury, W. A., and C. B. Tanner. 1975. Advection modification of the Priestley
and Taylor evapotransp~ationformula. Agron. J. 678404342.
Kelliher, F. NI., R. Leuning, M. R. Raupach, and E. D. Schulze. 1995. Maximum conductances for evaporation from global vegetation types. Agric.
Forest Meteorol. 73:l-16.
Kroon, L. J. M., and H. A. R. de Bruin. 1993. Atmosphere-vegetation interaction in local advection conditions. Agric. Forest Meteorol. 64:1--28.
23
Lee, T. J., and R. A. Pielke. 1992.Estimating the soil surface specific humidity.
]. Appl. Meteorol. 32:480-484.
Leuning, R. 1995. A critical appraisal of a combined stomatal-photos~thesis
model for C, plants. Plant Cell ~nviron.18:339-355.
Lohammar, T., S. Larsen, S. Linder, and S. 0. Falk. 1980. FAST simulation
model of gaseous exchange in Scots pine. col. Bull. 32:505-523.
McKenny,M. S., and N. J. Rosenberg.1993. Sensitivity of some potential
evapotranspirationestimation methods to climate change.Apic. Forest Meteorol. 64:81-110.
McNaughton/ K. G. 1976. Evaporation and advection. 11. Evaporation downwind of a boundary separating
regions having different surface resistances
and available energies. Quart. ]. Roy. Meteorol. Soc. 202:193-202.
McNaughton/ K. G., and P. G. Jarvis. 1983. Predicting effects of vegetation
changes on transpirationand evaporation. In: Water DeficitsandPlant
Growth. Vol. 7. T. Kozlowski (Ed.). New York Academic Press.
McNaughton, K. G., and T. W. Spriggs. 1989. A mixed-layer model for regional evaporation. ~ o u n ~ ~ - ~Meteorol.
a y e r 34:243-262.
Makkink, G. F. 1957. Testingthe Penman formula by means of lysimeters.J.
Inst. Water. Eng. 11:277-288.
Mihailovic, D. T., R. Rajkovic, L. Dekic, R. A. Pielke, T. J. Lee, and U. Zhopia
1995. The validation of various schemes for parameterising evaporation
from bare soil. ~ o u n ~ a ~ - L a y e r ~ e t76:259-289.
eo~ol.
Mintz, Y., and G. K. Walker. 1993. Global fields of soil moisture and land
surface evapotranspiration derived from observed precipitation and surface air temperature. ]. Appl. Meteorol. 32:1305-1333.
Monteith, J. L. 1963. Calculating evaporation from diffusive resistances. In:
Investigation of Energy and Mass TransfersNear the Ground. Final Report
under Contract DA-36-039-SC-80334. US. Army
Electronic
Proving
Ground's Technical Program.
Monteith, J. L. 1965. Evaporation and the environment. Symp. Soc. Exper. Biol.
29:205-234.
Monteith, J. L. 1981, Evaporation and surface temperature. Quart 1. Roy. Meteorol. Soc. 207:1-27.
Monteith, J. L. 1986. Manipulation of water by crops. Phil. Trans, Roy. Soc.
Lond. A 326:245-259.
Monteith, J. L, 1995. A reinterpretation of stomatal responses to humidity.
~ l a n tCell
/ Environ. 28:357-364.
Monteith, J. L., and M. H. Unsworth. 1990.Principles of ~ n v i r o n ~ e n tPhysics.
al
London: Edward Arnold.
Passioura, J. R. 1983. Roots and drought resistance. Agric. Water Manage. 7:
265-280.
Pelton, W.L., K. M. King, and C. B. Tanner.1960. An evaluation of the
Thornthwaite and mean temperature methods for determining potential
evapotranspiration.Agron. ]. 52:387-395.
Penman, H. L. 1948. Natural evaporation from open water, bare soil, and
grass. Proc. Roy. Soc. Lond. A293:120-145.
33
23
Thompson, N., I. A. Barrie, and A. Ayles. 1981. The Meteorological Office
rainfall and evaporation calculation system: MORECS (July 1981). Meteorological Office Bracknell,UK, Hydrological Memorandum No. 45.
Thornthwaite, C.W.1948.An approach towards a rational classification of
climate. Geogr. Rev. 38:55-94.
Tuzet,A.,A. Perrier, and A. K. Oulid Aissa. 1993. A prediction model for
field drying of hay using a heat balance method. Agric. Forest Meteoroz. 65:
63-89.
Van Bavel, C. H. M., and D. I. Hillel. 1976. Calculating potential and actual
evaporation from a bare soil surface by simulation of concurrent flow of
water and heat. Agric. ~eteoroZ.17:453-476.
Van den Hurk, B. J. J. M., and K. G. ~cNaughton.1995. Implementation of
near-field dispersal in a simple two-layer resistance model. J. Hydrol. 166:
293-311.
Van den Hurk, B. J. J. M., A. Verhoef, A. R. van den Berg, and H. A. R. de
Bruin. 1995.An intercomparison of three vegetation/soil models fora
sparse vineyard canopy. Quart J. Roy. Meteoroz. Soc. 121:1867-1889.
Viterbo, P. and A. C.M. Beljaars.1995.An Improved Land SurfaceParamete~ationScheme in the ECMWF Model and Its Validation. Tech. Report TR-75. Reading, LJK ECMWF.
Vogt, R., and L. Jaeger. 1990.Evaporation from a pine forest. Agric. and Forest
Meteoroz. 50:39-54.
GQSSYM was intended to simulate crop responses to physical conditions and provide producers with information that would aid in
making management decisions. To do that, a comprehensive understanding of the cotton plant itself is needed, inclu~ingmathematical
relationships among weather and soil and the responses of various
plant processes to those physical conditions. The infor~ationwas
organized into relatively small closely related packages sometimes
called subroutines or modules. COSSYMs program flow and model
structure are presented in Fig. 1.
CQSSYM is the main program from which all of the subroutines
vertically belowit in the diagram are called. ~LYMATreads the daily
weather information and calls DATES, which keeps track of both Jul-
ian day number and the calendar date being simulated. CLYMAT
then calls TM~SOL,which calculates the soil temperatures by soil
layer. SOIL is a mini main program, which calls the soil subprograms
(Boone et al., 1995). The soil routines provide the plant model with
estimates of soil water potential in the rooted portion of the soil profile, an estimate of the nitrogen entrained in the transpiration stream
available for growth, and an estimate of metabolic sink strength in
the root system.
The below-ground processes are treated in a two-dimensional
grid. The material balances of water and nitrate, ammonia, and organic matter are maintained and updated several times eachday.
FERTLI~distributes ammonium, nitrate, and urea fertilizers in the
soil profiles. GRAFLO simulates the movement of both rain and irrigation water into the soil profileby gravitational flow. ET estimates
the rate of evaporation from the soil surface and t r a n s ~ ~ a t i o
from
n
the plant. UPTAKE calculates the amount of the nitrogen and water
taken from the soil region where roots are present. CAPFLO estimates
the rewetting of dry soil from wetter soil by capillary flow. NI
calculates the conversio
mmonium to nitrates by bacterial a
in the soil medium. C
is also a mini main program, one that
238
Hodges et al.
239
term weather records are available, the producer may run the model
with many weather and management scenarios to increase the probability of selecting the management practice that most nearly fits his
or her objectives.
Such a model, based on important physiological processes and
their responses to physical conditions, should be able to correctly
simulate unique combinations of conditions that were not explicitly
represented in the model development process. Such model capability provides the characteristics that are needed in combination with
precision agriculture. Precision agriculture will cause producers to
question why certain areas yield more than other areas. Process-level
mechanistic models will help provide that information.
Y
henology is defined as the length of time required for a articular
developmental event to occur or the period between events. For example, the length of time from emergence to first flower bud (commonly called
in cotton) formation is a phenolo~calevent. Also,
the time between unfolding of leaves on the mainstem or f ~ i t i n g
branches are phenolo~callymeaningful periods, as is the duration of
a leaf or internode expansion.
1.
Rays to
enced by temperature. In GOSSYM, daily progress to floral bud emergence (the reciprocal of time is defined as daily developmental rate)
is calculated by summing the values for average daily temperature
(Fig. 2). Themodel predicts that a square will be produced when the
summed values equal 1or greater. About40 days are required at 20C
and only 23 days at 30C. Temperatures above 28C do not speed the
floral initiation process, nor does additional photosynthates (K. R,
Reddy et al., 1993a). Progresstoward square formation actually slows
at temperatures above 28"C, suggesting that injury occurs at higher
temperatures. No progress toward floral development occurs below
15C.
When Baker et al. (1983) developed GOSSYM, the days to first
square value was based on dataof ~ o r a g h a n
et al. (1968). That model
predicted 20% more time to produce first square than was observed
in field conditions. They calibrated the model by introducing a multiplier into the temperature response function for this phenological
event. This adjustment improved the predictions at near-optimum
temperature, but predicted rates were too fast at below- and aboveoptimum temperatures. We have since learned that modern cultivars
respond to temperature differently than those grown several years
0.06
y = - 0.1265 +
r =
0.00
15
25
ago. Modern cultivars initiate squares much earlier than older cultivars and are able to produce squares and fruiting structures at higher
temperatures (K. R. Reddy et al., 1993a).
Others have used growing degree-days as a way of summarizing temperature data to predict time of first square (Jackson, 1991).
This procedure is appropriate for predicting phenological events if
the temperature response of the ~ h e n o l o ~ crate
a l is linear over the
range of temperatures experienced but is i n a ~ p r o p ~ a when
te
the
event being simulated is nonlinear with temperature.
2.
= - 0.11482641 + 0.00Q67002*
- 0.~0014~18* X*, 6
0.94
except at low temperature where they had only limited data. Apparently changes in varieties during the past few decades have not influenced the sensitivity of cotton to temperature during the square
maturation period. About 46 days were required at 20C to develop
flowers after squares were visible, but only 21 days at 30C.
Nodes of ~ a i n s t ~ ~ , F r uand
i t Vegetative
i~~,
Branches
IGURE 6 The number of days required for cotton plants to produce a node
at different positions on the mainstem when grown at near-optimum temperature. The arrows are the positions at which the first fruiting branch and
the node of the unfolding leaf at first flower were produced. (From K. R.
Reddy et al., 1995b)
odges et al.
ExpansionDuration
Leaf andinternode
The duration of mainstem leaf expansion and of each mainstem internode elongation decreased gradually as temperature increased to
about 30C. The daily progress of leaf expansion and internode elon-
0.1
0.0~
Therefore, they reasoned that the ratio of plant height to the distance
between rows is a better index of ground cover and a better indicator
of light interception in cotton. However, in a mature, defoliating canopy this relation does not remain true. They also empiricallyadjusted
the light capture relationship from field data so that during maturation, when the LAI became less than 3.1, the light interception decreased linearly with LAI. Photosynthesis i s estimated in GOSSYM
on a canopy basis. The rationale for using a canopy model, which is
relatively empirical in nature, includes several considerations for the
alternatives. The basic physiology of photosynthesis is controlled at
the leaf level; however, there is considerable difficulty in scaling up
to a canopy level. The reasons for such difficulties include the problem of defining leaf environment. How much radiation strikes a leaf
varies with position in the canopy and the angle to the sun. That
angle is also modified by the heliotropic nature of cotton leaves. In
addition, the age and nutritional status of individual leaves need to
be taken into account, and, perhaps most important, the difficulty of
validating a leaf element model requires data that are not readily
available. For these reasons, Baker et al. (1983) chose to model photosynthesis on a canopy basis. Their aim was to provide the model
each day of the growing season with anappropriate increment of dry
matter produced that was a function of the plant age, size, physiological status, and environmental conditions and to distribute it to
the growing points in the plant.
Canopy gross photos~thesis(Pg) responses to radiation are
plotted and shown in Fig.8.Gross photosynthesis is equal to net
~ h o t o s ~ t h e splus
i s respiration. Respiration is a function of biomass
and temperature. Canopy photosynthesis was measured under a
range of temperature, vapor pressure deficits, and fertility conditions
but with good soil moisture conditions. These individual factors sig~ficantly in~uenced ~hotosynthesis,
but none changed the estimate
of photosynthesis by 5% over the conditions tested. Therefore, Baker
et al. (1983) used the relationship in Fig. 8 to calculate hourly photo~ynthateyield and daily totals. Those values are used from daily
total radiation to simulate daily potential .Pg in GOSSYM. They did
not adjust P, for nitrogen status because in their tests the nitrogen
range was not sufficient to cause differences.
Data more recently acquired but not incorporated into a commercially used version ofGOSSYM
show that nitrogen-deficient
plants are also carbon-deficientplants because of the effect of nitrogen
deficits on photosynthesis (K. R. Reddy et al., 1997a). A 25% decrease
in photos~thesisoccurred when leaf nitrogen declined to 20% of
FIGURE8 Paily canopy gross photosynthesis (net photosynthesis plus respiration) as a function of solar radiation. (From Baker et al., 1983.)
optimum nitrogen levels,and photosynthesis decreasedto 50% of maximum when leaf nitrogen was reduced to 60% of its maximum value.
The effectsof water deficits on photosynthesis are shown in Fig.
9. Cotton canopy photosynthesis declined as midday leaf water potentials decreased. It appears that little change occurs in canopy photosynthesis at leaf water potential values above about -1.5 to -1.7
MPa, although the trend line causes one to infer that the maximum
values are nearer zero.
The size and age of each organ on the plant is considered, and the
potential growth rate is calculated as a function of temperature, assuming no shortage of photosynthesis or nitrogen. Potential growth
is calculated for day and night periods separately in COSSYM, using
temperature and water deficits appropriate to those respective time
periods.
Leaves
odges
al.
Potential plant height growth rates are based on the estimated expansion of the top three nodes. Controlled en~ronmentaldata were
not available for stem expansion rates at the time ~ ~ S was
S first
Y ~
developed, so field data were used (Bruce and Romkens, 1965). In
the model, stem expansion is based on the age of the topthree nodes.
Thus, simulated stem expansion is related to developmental rate, for
which there were no reliable data from controlled environmental
studies in the literature. The expansion rate is written as a function
of temperature and of metabolite supply, which may be reduced by
water deficits and PIX. A total carbohydrate demand is calculated for
day andnight periods separately, using appropriate temperature and
FIGURE10 Relative leaf expansion rate and specific leaf weight of mature
cotton leaves as functions of the temperature at which the leaves were
= Relative Leaf Expansion Rate; o = Specific Leaf Weight. (From
Baker et al., 1983.)
water deficit for those time periods. Since data are now available/
plant height can be calculated using the prevailing conditions present
at the time each internode is elongating (K. R. Reddy et al., 1997a).
Root expansion is simulated as a function of soil temperature and
biomass of roots in each of three age categories: (l)5 5 days, (2) 515 days/ and (3) >l5 days. The rootenvironment is estimated by keeping track of the physical and chemical properties of a hypothetical
grid that is the width of the row spacing and 2 m deep. There are 20
40 cells in the grid in which the properties and conditions are
initialized at the beginning of the season and updated daily. Further
discussion of this is presented in Section 111. A materials balance of
nitrogen, water, and root mass is maintained for each cell, The concept of age dependency on water uptake and root growth is discussed
by Graham et al. (1973). Direction of root growth from a parent cell
to adjacent cells in the matrix is controlled by relative amounts of
mechanical impedance in the various cells where growth may occur.
Potential root growth is reduced in each cellof the matrix where
physical conditions or nitrogen supply is limiting. If growth is limited
and
matrix with a shortage of nitrogen and/or higher penetration resistance. This makes additional carbohydrate available for root growth
in grid positions with good nitrogen supplies and/or low penetration
resistance. Third, based on the level of water stress in the plant and
the ratio of root to stem plus leaf biomass, potential root growth may
be increased. Thiseffectivelyincreasesroot
growth in periods of
drought if there is a small or negligible fruit load. Finally the directions of root growth from a parent soil cell to adjacent cells in the
root matrix is controlled by relative amounts of mechanical impedance in the various cells in which growth may occur and weighted
for downward growth (geotropism). Actualroot growth is then
calculated as the product of potential growth multiplied by the carbohydrate supply/demand ratio, GSTRES. This partitions photosynthate to each organ on the plant purely on the basis of the
contribution of that organ to the total carbohydrate demand.
Potential fruiting structure growth is calculated from temperatureand adjusted downwardunder water-deficitconditions.The
temperature responses of squares and boll growth are presented in
Figs. l1 and 12.
.
The model calculates the carbohydrate requirements for the various
organs, then the su~routineNITRO is called. This subroutine esti-
Y = -0.010168 + 0.001253 *
-0.~0195
25
Hodges et al.
20
35
mates the nitrogen required for the assimilation of the carbon just
estimated for all the organs. The nitrogen requirements are summed
for the vegetative and fruiting structures. These sums are used to
estimate the total nitrogen requirement (nitrogen demand) for each
day.
The nitrogen supply is the nitrogen taken up that day plus any
that may have been taken up but not used on previous days. Such a
nitrogen reserve is assumed to be stored in the leaf, The daily supply/demand ratio is calculated to estimate the maximum fraction of
the potential carbohydrate that can be assimilated in above-ground
vegetation, roots, and fruiting organs considering the nitrogen limitations. These calculationsallow the model to estimate organ growth
taking into account the amount of nitrogen available. A conceptual
diagram of the nitrogen supplydemand rationale is shown in Fig.
13. A more mechanistic nitrogen supply and stress model is needed.
A carbohydrate supply/demand ratio is calculated in a similar
manner to the nitrogen supply/demand ratio (Fig. 14). In this case
the carbon supply is estimated from photosynthesis, which is a function of the light intercepted by the canopy and leaf turgor. The C
demand is estimated as a function of the total growth potential of all
plant organs. That potential is controlled by turgor, age, and temperature of all the organs. Carbon stress is assumed to be the ratio of C
supply to C demand. When C stress and N stress are equal to 1, the
various organs are not limited by carbon or nitrogen. As the ratio of
255
FIGURE
Factors that control nitrogen supply and nitrogen requirements
of plants. When the supply/demand ratio is less than 1, plant growth is
nitrogen-limited.
FIGURE14 Factors that control available carbon. When the carbon supply/
demand ratio is less than 1, plant growth is carbon-limited.
Nitrogen
Stress
Carbohydrate
Stress
Delay
Abscission
and plant
The soil is treated as a slab 1cm thick, 2 m deep, and extending from
the center of one planted row to the next (Fig. 16). This mathematically convenient grid is superimposed over the natural soil horizons.
The soil physical and hydrological properties are given by horizons.
A complete description of the processes and how they are formulated
and coded is given in Boone et al. (1995) and follows the program
flowchart of Fig. l. The commercially available GOSSYM/ COMAX
considers the plants to be at the upper corners of the soil slab, but a
revised version (Theseira, 1994) places the plants in the middle of the
slab in order to consider alternate furrow irrigation and alternate side
of the row banding of fertilizer.
In the commercially available version of the simulator, rainfall
and irrigation are treated as daily amounts. These are broken up into
five increments during the daylight hours and one increment during
the dark period. Water is simulated to move into each layer, starting
at the surface and moving downward, as calculated by the Darcey
fluxes (Hillel, 1980). Within each layer, water is moved from side to
side by the same type of flux equations based on the soil water gradient and the soil water diffusivity. A revised version of the model
uses the soil matrix potential gradient and the hydraulic conductivity
to calculate these fluxes. Both versions have advantages and disad-
15
20
Column
FIGURE16 A hypothetical grid between plant rows that allows a description
of soil physical conditions and growth and distributionof plant roots. COSS W assumes that each cell is 5 cm 5 cm.
vantages. Data needed for these calculations are, for each horizon,
the
depth of the horizon, the saturated hydraulic conductivity, the soil
moisture release curve (from 0.01 to 15 bar), the bulk densitypercent
sand, and percent clay. These latter three properties are used in other
calculations or model modules. Any soluble chemicals, primarily ni-
25
odges et al.
trate nitrogen, are assumed to move with the water fluxes. This mass
movement is assumed to be much greater than diffusion.
The runoff from rainfall or irrigation is based upon the Soil Conservation Service "curve number method" [USDA (1972); for details see
Boone et al. (1995)l. The runoff is based on the soil water content at
the soil surface for the preceding 5 days andthe surface horizon sand
and clay content. There is a variable in the soil hydrology file that
can be changed if the user does not want runoff to be considered. A
grower might select this option if the field has been landleveled, such
as for rice. The runoff values are subtracted from the rainfall or irrigation amounts for each day, thus giving the "effective" rainfall value
to be infiltrated into the soil.
In the commercially available version of GOSSYM/COMAX, an empirical model of soil temperature is used that is based on a regression
of air temperature and measured soil temperature at four depths
(Boone et al., 1995). The equations used in this module are based on
the work of McWhorter and Brooks (1965). Daytime and nighttime
Evapotranspiration is calculated using a modification of Ritchies approach (Ritchie, 1972). The modifications
involve using light interception instead of leaf area index and m i ~ m u mdaily temperature instead of relative humidity or dew point. These latter two properties
are very difficult to measure in a weather station located near agricultural fields owing to dust. For complete details, see Whisler et al.
(1986) and Boone et al. (1995). Theevaporation amount is taken from
the surface layer of soil cells and by soil water gradients from the
lower cells on an iterative basis each day. The transpiration amount
is taken from the rooted cells and weighted for root amount (mass)
and root age in each cell containing roots. This causes moisture gradients among the cells with the lowest water content in the cells with
the most active roots. Water is simulated to move from regions of
high water content to drier regions by Darcey fluxes as described
earlier (Section 1II.A).
.
One can readily see that processes involved in plant growth and development are complex. GOSSYM simulates only a few of the seemingly more important processes. However,there are many other processes that are critical to crop growth. These involve both plant and
soil processes that appear important to simulating crop responses to
real-world situations. As additional information becomes available,it
should be incorporated into the model to allow crop simulation in a
wider array of environmental conditions.
Unfortunatel~model complexity requires that additional information about the crop or conditions be provided by the user, Although users typically want all the information possible about their
crop, during busy times they may also be unwilling to provide that
information or examine and interpret the details of the model simulations. We attempted to hide some of the complexity and simplify
the operational aspects of running the model by providing an expert
system that could aid the decision-making process.
of COMAX
The management decisions currently addressed by a COMAX analysis are irrigation, nitrogen fertilization, plant growth regulation, and
harvest timing. Figure 18 illustrates how COMAX divides the growing season into three sections: (1)the portion of the growing season
that has already elapsed and for which "actual" weather data are
available, (2) a short period of time (at most 1 week) following the
current date for which predicted weather data are available, and (3)
the remainder of the growing season. Historical weather data are typically used as input for this third portion of the simulated season.
Each historical weather file is called a weather scenario or future
weather file. COMAX can run the simulation with up to three future
weather files. When COMAX runs the GOSSYM simulation, it runs
the simulation through the last day of actual weather, saves the Val-
odges et al.
262
ues of all the GOSSYM state variables at the end of the simulated last
day of actual weather, and then runs the simulation from this point
to the end of the season with up to three different weather scenarios.
COMAX uses the stored state variables to reinitialize the simulation
to the state at thelast day of actual weather before running each new
weather scenario.
Management decisions that a COMAX user wants to consider
are selected using the GUI. Prior to a COMAX analysis, all weather,
soil, and cultural i n f o ~ a t i o nare specified as for GOSSYM. Users are
also required to provide additional information about crop management practices that are relevant to the task, For example, a user who
wants COMAX to determine an irrigation schedule for the crop must
provide information about how irrigation is delivered (drip, sprinkler,
etc.), the m a ~ m u mamount of each irrigation, the minimum time
between irrigations, and the point in the growing season when the
system should start to consider applying irrigations. COMAX assumes that any cultural inputs specified prior to the last day of actual
weather have already been applied to the crop and cannot be
changed. It develops a strategy for using the GOSSYM model to analyze the set of cultural management decisions specifiedby the user.
This task includes planning a schedule of simulation runs, determining what the inputs should be for each simulation run, determining
what data need to be collected from each simulation run, and determining what data analysis is needed at the endof each run. COMAX
then runs the model one or more times using the strategy, modifies
the strategy after each run as necessary, and evaluates the output of
these GOSSYM runs for the user.
Future
f
Emer~~nce
End
Today
of the season
2.
New
COMAX is a toolbox of advisors that can be used together or independently. The COMAX tools have two purposes: (1)to improve the
users efficiency when using the GOSSYM simulation and (2) to automate the use of the GOSSYM model for making recommendations
to manage the cotton crop. Two categories of tools have been developed to fulfill these purposes. Scenario tools are used primarily
to automate running the simulation model and tosummarize results.
Scenario toolsdo not make recommendations but rather run the simulation under conditions specified by the user and summarize the
results of the Simulation. Advisor tools, on the other hand, use
rules to determine how to plan a set of simulation runs and make
recommendations for management decisions based on the simulation
runs. Figure 19 shows the COMAX toolbox that is currently implemented. Any one of these toolscan be selectedfor a particular
COMAX analysis. In addition, one irrigation advisor (long-term or
water conservation), the fertilization advisor, and the PIX scenario
tool can be used in any combination.
Several factors were taken into account in the design of the
mechanism for using multiple tools:
1. The number of simulation runs required for the analysis
should be minimized.
2. Changing or enhancing the knowledge content of one tool
in future versions should have minimal impact on the operation of the other tools.
3. It should be straightforward to add new tools to the toolbox
and to integrate the activities of the new tools with activities
of existing tools.
1.
User-Specified
Many users had requested changes to the original riga at ion and
fertilization tools that would allow the recommendations of the advisors to be adapted to cultural practices of individual growers. For
example, the rules used by the original irrigation advisor were illsuited for the high plains ofTexas, where water resources are ex-
o ~ g e set al.
tremely limited. Two alternative solutions to this problem were considered. The first was the development of sets of region-specific
rules based on the expertise of cotton specialists and growers from
major cotton-growing regions. The second was the implementation
of user-specified rules that would allow individual growers to customize the tools to their specific crop management practices. Although the use of regional rules has the advantage of allowing the
system to make very specific recommendations, this alternative was
rejected for the following reasons:
1. The number of regions is very large. Texas alone has five
distinct cotton-growing regions.
2. There are large variations in crop mana~ementpractices
within each regionthat are due to a variety of factors-other
crops being grown on the same farm, field history equipment available to the grower, the economic status of the
grower, and many more.
3, The use of a specific set of regional rules would limit the
users ability to use COMAX to experiment with the use of
different cultural inputs and practices.
User-customized rules, on the other hand, allow the recommendations to be tailored to each growers crop manage~entpractices.
Since mostgrowers have adapted extension servicerecommen~ations
to their particular situation, it was considered preferable to give them
a mechanism for customizing COMAX rules. Some aspects of the
customization allow users to adaptthe rules to their particular delivery capabilities, while others reflect the types of rules of thumb
developed by growers and cotton specialists through years of experience. Several examples of the type of custo~izationenabled by
us~r-specifiedrules are illustrated on the Irrigation Advisor screen
from the GUI in Fig. 20. The check boxes on the left require the user
to select one of the three irrigation tools (short-term, long-term, or
water conservation). Each of these tools is described in more detail
in the examples of the following section. To customize the irrigation
rules, the user specifies the maximum amount of irrigation that
should be delivered for a single application, the minimum number
of days between applications, and the application method (furrow,
sprinkler, or drip). These allow the user to tailor the rules to his irrigation delivery capability. They also allow the user to play what-if
games with different types and amounts of irrigation delivery. The
s t ~ ~and
i nstopping
~
criteria for irrigation allow the grower to specify heuristics for when irrigation should be initiated and terminated
267
Field Profile
Profile: SAMPLE91
COMAX lnfo~ation
Clear
IRRI~ATION
ADVISOR
App
COPY
(in)
1061
Days
Between
Short
Term
Min
App
at er Conservation
I.l.00)
Meth~
lete
Edit
New
0
Rename
Run
Select
ormal
irrigation; that date is usually based on the average first frost date.
In other regions, where boll rot is a particular problem, growers almost never irrigate after the first open boll. Thecustomization options
available forother advisors are discussed briefly in the following sections describing each advisor.
2.
tiun Three different irrigation advisors have been developed to accom~odatethe way different growers use the model for
irrigation timing. These three irrigation advisors and their input requirements are shown in Fig. 20. All three advisor systems interact
with GOSSYM on a daily basis. Each of these advisors examines the
average soil water potential each day and if it falls below -0.5 bar,
an irrigation is scheduled.
The short-term irrigation advisor is meant to be used in those
situations when the onlydecision the user is trying to make is
whether the crop will need to be irrigated within the next 2 weeks.
It cannot be used in conjunction with any other advisor. The shortterm irrigation advisor runs the simulation for 2 weeks past the last
day of actual weather and then informs the user if the GOSSUM:
model is predicting the need for an irrigation application within that
time period. Users often use a weather scenario with all rainfall removed from the future weather file so they can determine when they
will need to irrigate if there is no rainfall.
The long-term i r r i ~ ~ t i advisor
on
develops an irrigation schedule
for the remainder of the growing season using each weather scenario
specified by the user. Thisadvisor should be used in situations where
the grower wants to h o w what type of irrigation schedule will be
needed under different weather scenarios for the entire growing season. This is also the advisor that will most often be used in conjunction with the fertilization advisor and PIX scenario tool for an irrigated field.
The water conservation advisor is a modification of the longterm advisor that modifies the amount of irrigation r e c o ~ ~ e n at
de~
each irrigation using the predicted evaporative demand of the crop.
One s i ~ ~ l a t i run
o n is made in which irrigation is applied in the same
manner as for the long-term advisor. A subsequent run is made that
uses the evaporative demand from the first run to determine the
r ~ Z ~ zThe
e ~fertilizer advisor has been modified to accommodate more detailed information about the users fertilizer delivery
practices. A wide variety of fertilizer application amounts, materials,
and schedules are used by cotton growers. Lemmons (1986) original
fertilizer advisor was mainly concerned with determining the amount
of fertilizer to apply. The system had little knowledge of how to
schedule fertilizer applications to not only meet the nitrogen requirements of the plant but also accommodate the application practices
and capabilities of the grower. The new fertilizer advisor divides the
growing season into four segments for purposes of scheduling nitrogen applications. The general markers for these segments are emergence, first square, first bloom,and late season, respectively. For each
of these markers, the user is asked to specify the form of nitrogen
that will be applied, the method of application, and the percentage
of total nitrogen requirement that the grower will try to supply. A
percentage of zero for a particular segment indicates that the user
does not want to consider an application at this marker. COMAX
makes a series of simulation runs to determine how much nitrogen
should be applied to the crop and develops an application schedule
that meets the percentages specified by the user as closely as possible.
Earlier versions of GOMAX used two different strategies to determine how much nitrogen should be applied to the crop. The first
version developed by Lemmon (1986) ran the simulation with no
added nitrogen until a nitrogen stress was detected, applied a fixed
amount of nitrogen a few days before the simulated stress was encountered, and then repeated the process until the crop did not encounter nitrogen stress. This strategy worked quite well when the
amount of nitrogen that needed to be applied was relatively small,
but it required a large number of simulation runs when a lot of nitrogen was needed to meet the demand. James Siefker (personal communication) modified COMAX to use a demand strategy rather
than the previous stress strategy. In Siefkers strategy,an excessive
amount of nitrogen was applied to the crop on the first simulation
run, the actual demand of the crop was calculated from the amount
of nitrogen actually used by the crop, and then this modified amount
was applied on the second simulation run. Additional simulation
27
Hodges et al.
runs were often necessaryto refine the nitrogen amount. This strategy
worked very well when a large amount of nitrogen was needed but
required too many simulation runs when only a small amount of
nitrogen was needed, such as late in the growing season. The new
version ofCOMAX implements both of these strategies and uses a
set of heuristics to select the most appropriate strategy. The user is
asked to enter the typical yield for the crop, and this is used to compute a rough estimate of the total nitrogen requirement of the crop.
The heuristics use this estimate, along with the amount of nitrogen
already available forthe crop (residual nitrogen and fertilizer that has
already been applied), the point in the growing season, the availability of irrigation, and the cotton variety to select either the "demand" strategy or the "stress" strategy. These heuristics were developed on the basis of discussions with many experienced GOSSYM/
COMAX users, both cotton growers and scientists. For both strategies, the goal of the advisor is to find an amount of nitrogen to apply
that will allow only a slight nitrogen stress late in the growing season
so there will be sufficient nitrogen to support plant growth and boll
development but there will not be nitrogen available to support late
season vegetative growth that will interfere with boll maturation and
harvest. The time window used by COMAX is the
days prior to
the date the crop reaches 80% of the maximum yield.
~ l a ~ t
~ r o ~ tThe
~ most
~ common
e ~ l areasons
t o r for
s using
plant growth regulators with cotton are to reduce the size of the plant
structure while increasing yield. GOSSYM models the effect ofPIX,
the most widely used growth regulator. The PIX scenarios tool was
developed to automate the process of making simulation runs that
users typically make in order to decide if they should apply PIX. With
each scenario,the tool makes onesimulation run with no PIX applied
and one run with PIX applied as specified by the user. A summary
comparing the results of these runs is then presented. This tool does
not currently determine when and how much PIX should be applied,
but it does allow the user to experiment with the effects of different
PIX tools. An experimental PIX advisor has been developed that uses
an analytical model to analyze the sigmoidal growth curve of the
plant and examines the growth rate during the early part of the growing season to determine how much PIX should be applied. It was
very straightfo~ardto integrate this new experimental advisor into
the existing toolbox.
retest ~
i The ~harvesti timing
~ scenarios
~
tool was implemented to automate the analysis that many users have been doing
271
~ e ~ t S ~c ee~ ~r r i oWhen
s
the weather scenarios option is selected, COMAX will run GOSSYM with the specified weather scenarios and summarize the results of each run. This tool is useful if
the user wishes to run the GOSSYM model with more than one
weather scenario and a specified set of inputs. Using the weather
scenarios option has several advantages over making a single GOSSYM run with each scenario. The user needs only specify one set of
inputs for all three files; the use of a state file by COMAX results in
a significant time savings in running the simulation, and a summary
of all simulation runs is provided, thus facilitating analysis of the
simulation results.
The heart of the new COMAX design is the simulation run planning
component. After the user selects a set of COMAX tools for a particular analytical session and initiates a COMAX analysis, the COMAX
simulation planner must schedule a set of simulation runs to achieve
the goals of each advisor. This includes specifying how to modify
input values, deciding which data values should be recorded during
the simulation run, and determining how the results need to be analyzed. The heuristics used by the planner are based on the expertise
of many GOSSYM users. The work of each of the advisors is called
a task, and there are a set of available actions that can be used to
achieve each of the tasks. The planner has a set of rules that it'uses
to develop an initial schedule of simulation runs and the actions that
are to beexecutedbefore, during, and after each simulation run.
These rules are very straightforward in cases where only one advisor
has been selected by the user. When several advisors are selected for
one analysis, the planner must try to achieve the goals of each advisor
while using a minimum number of simulation runs, When planning
a set of simulation runs with multiple tasks, the planner constructs a
.
ICOMAX has many applications for on-farm use (MCal., 1989). A grower might be considering leasing or buying
a new field or farm. That person can determine from the county soil
survey what soils are there, but he would like to h o w what he might
expect from his normal production practices on this new area. If the
necessary soil files are in the ~ ~ S S ~ / file
C ~
(we~have
A over
X
350 soils, mainly from the Cotton Belt, in that file), then he can make
several simulations using his normal production methods and selecting different years of weather data. The results would indicate
whether or not he might want to consider irrigation for that field if
it was not irrigated, how much nitrogen he should apply preplant
and/or as side-dresing, what varieties might bebest for that specific
soils, what row spacings to use, etc. We have found that on some
soils cotton generally yields more than on other soils, some varieties
yield more than others, some planting dates are more optimum than
others for a particular climatic location,and some row spacings seem
better most of the time under those same climaticconditions and soils
(Wang and Whisler, 1994). Another exercise a grower might do is to
iming and rate of chemical applications such as TE
(a pesticide, plant growth regulator, and harvest
ical, respectively) by simulating the crop with these different management practices, with different weather and soil scenarios. These
types of exercises, which can be done before the planting decisions
are finalized, are often called strategic exercises or strategic decision
making.
During the growing season, there are several other types of decisions to be made, and these are called tactical decisions. They
might
I irrigate tomorrow, or should I wait? Should I side-dre
en or wait and apply foliar nitrogen after a
rain? Should I apply PIX today or wait? Should I apply
defoliant tomorrow or wait? All of these decisions are made on a
field-by-field basis and might be made with precision a ~ c u l t u r e ,
using variable rate applications on a soil-by-soil, variety-by-v
subset of a field. This assumes that the grower has access to
date weather data, either from his own weather station or from one
within close proximity to his field and has rainfall and/or irrigation
for the specificfield.In the ids south, that pro xi mi^ is
miles. In the more arid regions it might be greater.
e simulator-assisted decisions result is greater economic rence of better information
AX. Scientists from Texas
ension Service, stud
duction decisions and prac
ig and Thomas (1992) found that the av
user attributed $55/ acre net return to the cro
model when it was used in his cotton management. F~st-yearusers
217
gained less, and more experienced users found it more helpful and
required less time. More experienced users found more ways, sometimes unexpected by the developers, in which the information provided by the simulator could be useful.
er
i ~ ~ t i o ~ s
relatemodel-calculatedphysiologicalprocesses
withthe
threedimensional plant representation. Through colorization, such effects
as variations in plant temperature can be displayed to show the effect
of water deficits on stomatal closure, with the resulting slowing of
transpiration rate and diminishing evaporative cooling. Another example is a three-dimensional soil representation with roots, soil water,
chemicalfertilizer, and/or pesticidemovement in both time and
space.
With the addition of expert systems technology that can beused
to hide system complexity from the user, the combined scientific visualization system will alsoprovide a unique tool for academictraining of both undergraduate and graduate students for departments
that teach agronomy, soil science, botany, agricultural engineering,
and environmental science. Students, through the use of an expert
system that facilitates com~unicationbetween the model and the
user, will be able to easily accessand exercise the visualization system
and the generic plant modeling system. The expert system-SVS combination wiU provide the student with analytical capability so that
soil, weather, and cultural practice effects areapparent in the context
of the physical and biological processes involved. The student will
be able to quickly gain insights into plant physiology, micrometeorology, soil physics, crop management, and many other associated
areas of expertise.
Another area of rapid technological development that has a possible role in crop m~nagementis a combination of geographic information systems (GIs), global positioning systems (GPS), and intelligent implements (11) technologiesforprecision
agriculture. These
techniques may be integrated with the simulation model and expert
system technologies to deliver dynamic, precision agriculture decision s ~ p p o rsystems
t
that can respond to within-season problems and
provide real advantages to productivity. The primary product of this
research will be the maintenance and improvement of food and fiber
production systems while minimizing erosion and the use of agrochemicals, thus lowering the negative environmental impacts of agricultural production.
Crop models are the only automated source of dynamic h o w l edge within computer technology, or any other known t e c ~ n o l o ~ ,
that can respond to crop status, current meteorology, and management practices. Crop models can provide the essential linkages between the crop manager (decision maker) and the formation made
available with GIs, GPS, and The crop model is the only technique
with which we are familiar that has a possibility of accepting, pro-
277
cessing, and responding to the vast quantities of information provided by the GIS and GPS to deliver variable rates of input through
intelligent implements. Appropriate responses to precision agriculture information will include recommendations on a soil type-by-soil
type or situation basis within single management units using intelligent implements to deliver variable seeding and agrochemical rates.
Crop models are the ideal source of dynamic information that can
optimize crop production and minimize resource utilization while
protecting the environment. Crop models also have the advantage
that they allow producers to simulate the application of new technology, test it on their soils with historic weather data for each site,
and develop the optimum management practices to use with the new
technology.
27
o ~ ~ete al.
s
Baker, D. N. 1980. Simulation for research and crop management. In: World
Soybean Research Conference 11, Procee~ings.F.T. Corbin (Ed.). Boulder, CO:
Westview, pp. 533-546.
Baker, D. N., and J. A. Landivar. 1991. Thesimulation of plant development
in GOSSYM. In: Predicting Crop P k e n o l ~T.~ Hodges
,
(Ed.). Boca Raton, FL:
CRC Press, pp. 153-170.
Baker, D. N., J. D. Hesketh, and R. E. C, Weaver. 1978. Crop architecture in
relation to yield. In: Crop P h ~ s i o Z o U.
~ , S. Gupta (Ed,). New Delhi, India:
Oxford and IBH, pp. 110-136.
Baker, D. N., J. R. Lambert, and J. M. McKinion. 1983.GOSSMM: A Simulator
of Cotton Crop Growth and Yield. South Carolina Exp. Sta. Tech.Bull.
1089.
Booch, G. 1991. Object Oriented Design. New York Benjamin/Cum~ngs.
Boone,M. Y. L., D, 0. Porter, and J. M. McKinion. 1995. RHIZOS 1991: A
Simulator of Row Crop Rhizospheres. IJSDA, ARS Bull. 113. Washington,
DC: Govt. Printing Office.
Bruce, R. R., and M. J. M. R o d e n s . 1965.Fruiting and growth characteristics
of cotton in relation to soil moisture tension. Agron. J. 57:135-139.
Campbell, R. B., D. C. Reicosky and C. W. Doty. 1974. Physical properties
and tillage of paledults in thesoutheastern coastal plans. 1. Soil Wuter Conserv. 29~220-224.
Graham, J., D. T. Clarkson, and J. Sanderson. 1973. Water uptake by roots of
narrow and barley plants. Ann. Rep. L e f c o ~ b Lab,
e
Wuntuge, p, 9-12.
Hesketh, J. D., and A. Low. 1968. Effect of temperature on components of
yield and fibre quality of cotton varieties of diverse origin. Cofton Grower
Rev. 45243-257.
Hesketh, J. D., D. N. Baker, and W. G. Duncan. 1972. Simulation of growth
and yield in cotton. 11. Environmental control of morphogenesis. Crop Sci.
22:436-439.
I
I
Lecler
In the design and planning of irrigation projects, most "water requirement" studies have focused on climatic factors that influence
water use for maximum crop production. This approach may be losing its relevance. In many parts of the world the demandson limited
water resources are of major concern and the available water resources are subject to increasing pollution. The net result is that water
users, especially the larger users like irrigated agriculture, will face
shortages and in order to survive will need to make more efficient
use of diminishing and more costly water supplies. For example, in
South Africa, a country that is about 90% of the way toward developing water resources infrastructure to accommodate the annual storage potential, more than halfof the 70 state irrigation schemes had
to cut farmers' water quotas by 50-100% in 1983 when the country
nominally required only 40% of the annual storage potential (Green,
1984).
:.AP~LICATION LOSSES
1.A
IGURE
diminish. The net result is that if a plot of crop yield versus actual
transpiration and applied irrigation water is drawn, the yield versus
applied irrigation water line will curve away from the yield versus
t r a n s ~ ~ a t i oline
n as shown in Fig. l. The more generous the irrigation, the more modest the proportional contribution to crop water
requirements (transpiration) is likely to be. This is one of the reasons
the a~plicationof the large amounts of irrigation water need to maximize yield can be inef~cientand is not always likely to result in
~ a x i m u meconomic return.
S
Lecler
286
--
REVENUE PUNCTlON
SYSTEM
COSTFUNCTION
a2
bl
b2
- DESXQN CAPACITY
1
- DXSIQN CAPACITY SYSTEM 2
- MAXIMUM NET RETURN FOR SYSTEM
IGURE 2
MAXIMUMNBTRBTURN
FOR SYSTSM 2
SYSTEM 1
SYSTSM 2
To develop a viable deficit irrigation strategy, information on the response of a crop to different irrigation watering r e m e s and environmental constraints is needed. This information can be determined
from experimental trials or by using crop growth/ yield simulation
models. Some of the problems with experimental trials are that
1. Experimental trials are expensive and time-consuming, and
therefore it is difficult to test a wide range of crops and
water application and management strategies.
ecler
290
Lecler
ter (PAW) is often chosen when irrigating to avoid crop water stress.
If this method of irrigation scheduling is used for deficit irrigation,
the crop is allowed to deplete soil water during certain growth stages
to a level that results in some stress and some reduction in yield.
The determination of the timing and magnitude of the
depletion-to-stress-level events is not a simple exercise. In the planning stages of an irrigation project, multiple simulations of different
depletion levels at different crop growth stages could provide a database for the selection of a suitable strategy. It is important when developing this database to also consider the actual physical characteristics of the irrigation systems that would be required to implement
the simulated strategies. It is difficult to optimize a depletion level
schedule and hardware requirements concurrently. Thisis an important consideration, as the level of flexibility built into an irrigation
system during the design phase can have a major influenceon capital
equipment costs and therefore on the potential returns (see Fig. 2).
To implement a given depletion level schedule, an irrigation system
with a large capacity is often required even though the potential capacity is only used for relatively short but possibly crucial periods
during a crops growing season.
In practice, the logistics of moving irrigation hardware to maintain an optimum depletion level schedule over all sectors of a field
is extremely difficultunless a solid set irrigation system is used. These
difficulties are compounded when the water budget of different sectors of a field are unequally influenced by rainfall events. Many farmers find the management of a simple depletion level schedule overtaxing and would be doubtful participants in a more complex and
varied approach. In addition, untimely irrigation equipment breakdowns and unforeseen delays that may occur in practice may result
in severe crop water stress if irrigation applications are routinely
triggered when more than 50% ofPAW is depleted and there is no
safety buffer of soil water.
When scheduling according to soil water depletion levels, the
soil profile is normally recharged to the drained upper limit (Dm).
An additional option for which depletion level scheduling to a
planned deficit is implemented is worth consideration. The root zone
soil profile is then deliberately recharged by irrigating to below the
DUL when the soil water threshold is reached. Theirrigation amount
is therefore planned to leave a portion of the potential soil water store
to be filled by the expected rainfall. One assumption in this mode of
scheduling is that irrigation is supplementary to rainfall in areas
where there is a high probability of rain falling between irrigation
Deficit
Planning
291
In this mode of irrigation either a preselected or otherwise predetermined amount of irrigation water is applied in a fixed cycle. The
selected cycle length is assumed to continue throughout the growing
season. Simple improvements to this strategy can considerably enhance the efficiency of irrigation applications. For example, the irrigation cycle could be stopped for a period of time after a rainfall
event, and different cycle lengths could be used for different parts of
the growing season. The length of time that the cycle is halted can
be related to the magnitude of the rainfall event and the effective
daily irrigation application according to the equation.
t = -P
ilc
where
hardware specifications that coincide with irrigation management considerations and yield consequences by using
computer simulation models.
294
Lecler
Irrigation
system
1
2
3
4
Seasonal
Maximum
irrigation
capacity
application
(mm/7 days)
(mm)
20
70
35056
45065
500
504
Crop
yield"
(t/ha)
7.6
8.5
9.0
9.0
Net
returns
to
land
($/ha)
(a)
60
57
Irrigable
area
relative to
system 4
(ha)
(b)
1.440
1.120
1.008
1.ooo
Relative
net
returns if
water is
limited
(a)
(b)
80.64
7'2.80
60.48
57.00
eficit ~rrigation~lanning
V.
CONVEYANCE
of various op-
cle er
29
Lecler
ning
Deficit Irrigation
299
d ~ o ~ o ~ cModeZZ~ng
aZ
System. R.E. Schulze (Ed). Report TT69/95. Pretoria:
Water Research Commission, pp. AT17-1-AT17-16.
Schulze, R.E., F. B. Domleo, P.W. Furniss, and N. L. Lecler. 1995a. Crop
yield estimation. In: ~ y d ~ o and
Z o A~~ o h y ~ ~ o Z oA# :Text to Accompany the
ACRU: 3.00 A ~ o h y d ~ o Z o ~ ' cModeZZ~ng
a2
System. R. E. Schulze (Ed). Report
TT691 95.Pretoria: Water Research Commission,
pp. AT19-1 -AT19-14.
Schulze, R. E., J. C. Smithers, N. L. Lecler, K. C. Tarboton, and E. J. Schmidt.
199513. Reservoir yield analysis. In: ~ y d ~ o Z o #and A ~ o h y ~ ~ o AZ Text
o ~ :to
Accompany the ACRU: 3.00 A ~ o h y d ~ o Z o ~ModeZZ~ng
'c~2
System. R. E. Schulze
(Ed). Report TT69/95. Pretoria: Water Research Commission, pp. AT14-1AT19-17.
Tsuji, G. Y., G. Uehara, and S. Balas.1994. DSSAT v3. Honolulu, Hawaii:
University of Hawaii.
.
The quantitative description of either a plant or animal system is
indeed complex. Describing the plant/ animal grazing interface presents an even greater challenge. The beef-forage grazing model deeffort by
scribed in this chapter is the result of nearly 20 years
many ~dividuals.Model developments are dynamic; that is, models
change as information and research objectives change. With this in
mind, the objectives of this material are to (1)present a brief history
of the submodels contained in the GRAZE model and (I;) conceptually describe the logic and mathematics
the model as related to
q u a n t i ~ biological
~g
processes.
302
Loewer
ode1 of Selective ~ r a ~ i n g
303
For the above reasons, the initial objective of the Southern Regional Research Project "Simulation of Forage-Beef Production in. the
Southern Region" (S-156), initiated in 1981, was to develop an improved beef-forage model that could more accurately test the effects
of nutrition and environment on animal maintenance, growth, and
production. This larger model was called GRAZE and was the emphasis of another Southern Regional Research Project S-221 "Development of Profitable Beef-Forage Production Systems for the Southern Region.'' At the end of this project in 1994, a user's guide and a
collection of "validation" case studies for GRAZE version 2.3 were
prepared (Loewer and Parsch, 1995; Parsch and Loewer, 1995). The
GRAZE software package includes both a DOS version (Version 2.3,
available since mid-1995)and a Windowsversion (Version 3.02,available since June 1996). Both versions are available for downloading
free of charge from the Internet at the home page of the University
of Florida Agricultural & Biological Engineeringat the following Universal Resource Locator (URL) address:
http: / / www.agen.ufl.edu
vervie
EVALUATED
LEAST
EVERY
15 ~ I N U
~IGESTI~
DRY
L ~~
EACH 24 HOURS
A I ~ T~A K E RR A E
simulated on a daily basis, and the growth and development of animals are computed at least every 15 min. The model is composed of
three submodels: plants, beef animal, and plant/animal interface. The
plant/ a ~ m a interface
l
simulates selective grazing and is based on
t that an unbound grazing area may be conceptually disubareas or "partial fields" that are created during the
ing process (Fig. 2). These subareas may differ in forage quality
density. The animal then selects among these as desired based
on the premise that grazing priority is established by an attempt to
ze the digestible dry matter intake rate. Subareas may be creconsolidated on the basis of the selective grazing process.
Subar~asare modeled as separate entities with regard to plant
AZE represents a significant improvement over the BEEF
that it more fully describes plant and animal physiology
e grazing process. However, GRAZ may be
used to simulate forage growth without grazing, or beef animal growth in a feedlot situation. The plant submodel is based on the premise that all
forages may bedescribed collectively by a number of parameters. The
animal submodel is somewhat unique in that it uses body composition as the key to defining physiolo~caldevelopment of the animal.
A study conducted in the early stages of this model's development
el of Selective ~ r a z i ~ ~
indicated that the body composition approach closely followed National Research Council (NRC) predictions for average energy and
animal inputsbut provided deeper insights into distribution of
growth under exceptionally high or low energy feed supplies (Loewer
et al., 1982,198330). These
results were further enhanced by comparing
the protein deposition predictions ofGRAZE to those of other researchers (Hintz et al., 1982).
3-
IGURE
Industrial dynamicssymbols.
Loewer
3~~
2.
3.
4.
5.
AZE:
307
(1)
.
The biological processes associated with an animal may reflect far
ifferent levels of aggregation. For example, the following modeling
efforts, which originated in the 1960s and 1 9 7 0 ~
represent
~
different
levels of me tho do lo^ and aggregation in their approach to describing beef animal hysiology. These models are all discussed in chapters of a book edited by Spreen and ~aughlin(1986) and have been
used as the basis for other modeling efforts.
California Net Energy Model
The California net energy system was mentioned earlier (Lofgreen
and Garret, 1968). ~ s s e n t i a l this
l ~ is a very simple static regression
model that relates intake energy to beef animal gain. Its equations
can be used as part of a dynamic simulation.
2.
Texas
Model
model (Sanders and Cartwright, 1979a, 1979b) describes beef animals in terms of a herd composed of various catego-
el of Selective ~ r a z i n ~
The Kentucky BEEF model is like the Texas A&M model in that it
considers a herd of animals composed of various categories within
the total. However, the emphasis of this model is the daily prediction
and associated interaction of forage and animal performance. This
model formed the basis for the GRAZE model, which came out of
the same working groups. GRAZE emphasizes the internal functioning of an individual beef animal (but not to the extent of the UC
Davis model) and its interaction with forage growth and availability.
NonspecificModels
Genetics
Loewer
31
....
FIGURE4 Factors affecting plant growth are determined for each species
growing on each subarea once each 24 h.
311
Daily Growth
Energy flow is expressed in the plant portion of GRAZE by converting sunlight, as represented by day length and temperature, into
plant growth. The ability of the plant to convert energy into dry matter may be limited by a number of other factors, including leaf area,
photoperiod, water, and available nutrients. Also, stored sugars in
the various physiological states of the plant contain stored energy
that may be used for plant growth.
2.
Water
GRAZE contains a water balance model that considers rainfall, runoff, percolation losses, and plant utilization. The availability of water
has a direct impact on plant growth potential.
3.
Nutrients
Nutrients (N, P, and K) are extracted from the soil by the plant. The
availability of soil nutrients may limit plant growth. Nutrients may
be replenished by a fertilizer application, and nitrogen may be
leached from the soil. Currently, GRAZE does not consider recycling
of nutrients from deposition of fecal material by grazing animals.
States
CRAZE places dry matter into one of three physiological states: new,
old, or dead material. Dry matter is also divided into three cellular
states (cell contents, potentially digestible cell wall, and indigestible
cell wall), all of which are subject to change based on their physiological states. In general, physiological states flow from new to old
to dead material with an accompanying change in cellular makeup.
*
+ B*TEMP + C*TEMP**2
(2)
where
MAX'
-!-
INI~U
AXIMU
diurnal period. Penrod (1960) and Smith et al. (1968) did considerable
research on the diurnal variation in air and soil temperatures. In the
time period from sunup to solar noon to sundown, daily air temperature follows a function of the form
TEMP = A
+ B*H + C*H**2
(3)
where
A,B, C = constants
H = time in hours
TEMP = temperature ("C)
From sundown to sunupl the function is of the form
TEMP = D*&+*(-E*H)
(4)
Loewer
(5)
and 240:
315
MAX
MIN
SUNRISE
MIDNIGHT
SOLAR
NOON
MIDNIGHT
SUNSET
15.0
31
XL = 0.1333*ARCQSE(- (SIN(O.Ol745~XLAT)
~ S ~ ~ ( O . O l ~ 4 5 ~ ~/E( ~
LO
TA
S ()O
) .Ol745~X~~~)
~CQS(O.01745~~ELTA)))~57.31
(10)
(11)
and
SUNSET = 12.0
(12)+ XL/2.0
1.0
0.0
IGURE 8
el of Selective ~
LAF = A
+ B*ADM + C*ADM"*2
LAF = 1.0
and
LAF = D
+ E*ADM + F*ADM**2
(15)
where
LAF = a leaf area factor that is the proportion of the daily
growth rate that will occur, 0 LAF 1.0
M = accumulated dry matter, which is the sum of daily
31
Loewer
l.O)/(B - A)}*(DL
A)]
where
1.o
0.0
319
C = maximum effect of photoperiod as a proportion of the optimum growth rate (0.0 C S 1.0)
DL = day length in hours ( B DL A)
As day length continues to decrease,
PF = C
(17)
The water stress factor calculatedin GRAZE ranges in value from 0.0
to 1.0 and is a direct multiplier of potential plant growth (Fig, 10).
For the most part, GRAZE does not "borrow" logic from other mod-
1.o
IL
FIGURE10 Plant growth per day as influenced by soil water.
oe~er
els, the exception being that GRAZE uses water balance logic from
CERES-Maize (Jones and Kniry; 1986) and specifically subroutines
SOILRI and a modification of WATBAL (called WATBL2) (Costello et
al., 1988). A new subroutine, NEWAT1, replaces the original GRAZE
Version 1 subroutine PRECIP.
In GRAZE, the water balance logic in WATBL2 calculates a daily
water stress factor as a function of daily rainfall, average daily radiation, and various soil and plant parameters. The following explanation of WATBL2 (WATBAL) is a condensation from Jones and Kniry
(1986).
Conceptually, the subroutine WATBAL receives values of daily rainfall and calculates runoffaccording to the US. Soil Conservation Service (SCS) curve number method. If precipitation is greater than a
computed minimum, runoff occurs. Potential in~ltrationequals the
difference between precipitation and runoff. Potential and actual soil
evaporation and potential and actual evapotranspiration are calculated on the basis of specific soil, plant,and weather conditions. Water
flow between soil layers is determined. Root growth and associated
potential root water uptake from the soil is calculated. The actual
water uptake is computed, and the soil water content of each layer
is updated.
Two soil water deficit factors are calculated. The less sensitive
factor is used to affect photosynthesis and is the ratio of the potential
root water uptake and transpiration. The more sensitive factor affects
cell expansion and is a weighted ratio of the same two factors of
potential root water uptake and transpiration. In CERES-Maize, plant
growth and leaf senescence associated with water stress are reduced
according to the water stress indices calculated in the o ~ g i n a subl
routine WATBAL. The less sensitive of the two water deficit factors
is used in GRAZE as the water stress factor. The~ ~ c t i of
o the
n water
stress factor is to reduce optimum plant growth in proportion to the
available water.
el of Selective ~ r a z i ~ ~
322
Loewer
transpiration and anaverage root length per unit of land surface area.
The initial root length per unit of soil volume is expressed as an
exponential function of soil layer depth. The root length density rate
of decline is an estimated, plant-dependent input variable.
Model
The WATBL2 water balance routine in GRAZE requires input parameters describing soil, plant, and water conditions. Many of the factors
used by subroutine WATBL2 must be estimated because of a lack of
measurement or scientific description. Input variables are listed and
described by Parsch and Loewer (1995). Soil water input parameters
for each paddock include
Number of days from planting to emergency
Root length per gram root weight (cm/@
Maximum increasein root depth per degree-day accumulated if
there is no water stress (cm/degree-day)
Weight of roots @/m2)
Maximum daily root water uptake per unit root length (cm3/cm
root)
Leaf area index factor for summer growth
Leaf area index factor for spring growth
Soil Conservation Service curve number for calculating runoff
Number of soil layers (maximum of 10)
Bare soil albedo
Soil water constant for calculating drainage rate
First,second,
and third coefficientsforroot
water uptake
equation
Upper h i t of stage l soil evaporation, mm
Depth of each layer, cm
Drained upper limit of soil water, decimal fraction
Drained lower limit of soil water, decimal fraction
Field saturated soil water content, decimal fraction
Initial soil moisture content, decimal fraction
Weightingfactorfor
soil depthto
determine new growth
distribution
Water inputs include daily values of precipitation, extraterrestrial radiation, and maximum and minimum air temperatures.
lant ~ o m ~ o s i t i oand
n ~~ysioio~ica
eef-Forage Mode!
Selective
of
Crazing
323
FIGURE11 Dry matter types and physiological states for each species on
each subarea created by the GRAZE simulation model.
All plant dry matter is further described as falling into one of the
following cellular states.
1. Cell content material
2. Potentially digestible cell wall material
3. Indigestiblecell wall material
The following plant parameters describe cellular nutrient and carbohydrate composition for each plant species in GRAZE (Fig. 12).
1.
2.
3.
4.
2.
There are three physiological states of the dry matter (Fig. 11) in
GRAZE, referred to as(1)new material, (2) old material, and (3) dead
material. Material in each of these physiological states ages according
to a physiological time scale, with new plant material eventually becoming old plant material, old material eventually becoming dead
material, and dead material eventually disappearing. The following
plant parameters foreach plant species in GRAZE describes the
32
el of Selective Grazing
325
OLD
Ma~mum
physiological ageof the initial quantity of dry matter
in the new partition, which is analogous to the life span in
days of a stated day length (this may be viewed as the time
required for a crop leaf to reach its mature size).
Maximum physiologicalage of the initial quantity of dry matter
in the old partition, which is analogous to the age when leaf
senescence is essentially complete.
Changes in physiological state should be viewed as a c o n t i n u u ~
from new material to old material, with each physiological state also
changing the ratio of cellular composition. In GRAZE, a range of
functions are used to "connect" the defined plant parameters, including linear, parabolic, negative exponential, and, under extreme conditions, step function relationships. Figure depicts the relationships
between cellular composition and physiological age.
Physiological aging in GRAZE relates physiological age to chronological age. Physiological age advances at the same rate as chronological age when (1) day length is exactly the same as that defined
as a plant parameter for the context in which physiological age was
specified or (2) the temperature remains above freezing.
In GRAZE, physiological age advances in direct proportion to
the ratio of day length to the reference day length provided as a plant
parameter (Fig. 14). Each of the physiological states (new, old, and
326
Loewer
L
GRAZE:
327
1.O
0.0
tion of that nutrient, except when soil nutrient levels are relatively
low. For this situation, the potential rate of removal is reduced until
a specified removal rate is reached. Nutrient extraction from the soil
is modeled using the following soil parameters:
1. Fraction of nitrogen in the soil that is unavailable for immediate crop growth
2. Fraction of phosphorus in the soil that is unavailable for
immediate crop growth
3. Fraction of potassium in the soil that is unavailable for immediate crop growth
32
~~1
"Applicatio
Application
~~
"l
Plant
Growth
Status
FIGURE16 Soil nutrients may be either transferred to the plant or, as with
nitrogen, lost by leaching through the soil.
=c:
XN*(l.O - CWALLI)*C
W C = XP*(l.O - CWALLl)*C
RKC = XK*(l.O - CWALLl)*C
el of Se~ectiveGrazing
32
where
RNG, RPG, RKC = amounts of N, ]p, K, respectively required for
growth, kg/ ha
XN,XP, XK = fraction of the cell content that is N, ]p, or K,
respectively
CWALLl = fraction of the plant growth that is cell wall
material
C = projected plant growth, kg/ha
GRAZE assumes that there exist minimum amounts of N, ]p, and
K in the soil that are always available to the plant. The soil input
parameters that define these values are
1. Minimum amount of nitrogen always in the soil (kg/ ha)
2. Minimum amount of phosphorus always in the soil (kg/ha)
3. Minimum amount of potassium always in the soil (kg/ha)
Plant species competition is modeled in CRAZE version 3.02 (Windows version) but not in Version 2.3 (DOS version), which was presented by Parsch and Loewer (1995) and Loewer and Parsch (1995).
Loewer
330
NEW
~ ~ y s i o l o ~Age
i c aof
~ ater rial, (days)
FIGURE17 Conceptual changes in potential digestibility of cell wall as impacted by physiological age and state of the plant.
331
Loewer
33
.
The beef animal portion of the GRAZE model (Fig. l)uses engineering principles to describe biological processes.Although each animal
genotype is somewhat unique, there are overriding principles that
govern them all.
The biological processesof the beef animal in the GRAZE model
may be quantified using
1.
2.
3.
4.
Body composition
Heat transfer and thermodynamics
Intake and digestion
Energy and nutrient priorities
"""""""""".
The model contains a priority for distributio~of chemical energy Chemical energy may bestored in the form of tissue or "excess"
fat (the term "excess fat" is discussed in detail later). These stores
may be converted back into chemical or thermal energy forms if there
is the need to increase body temperature or maintain. the fetus ~ r o ~ t h
Loewer
33
rate, respectively. "Excess" fat can also be converted into the chemical
energy needed to synthesize tissue or produce milk. The chemical
energy level maytrigger the initiation or termination of eating. A high
body temperature also may terminate eating. Nutrient distribution is
handled similarly.
werwiew of the Types of Flows
Laws
Conservation
~
~ The elaw of ~conservation
~
.of energy states that energy
can be neither created nor destroyed, However, it can be changed
from one form to another.
~~e~ Matter is the material of which substances are made
and on which energy acts. For purposes of the discussion of animal
physiology, matter is not destroyed but may be converted into different forms.
2.
Dry Matter
Dry matter refers to material that does not contain any water. Interms
of the animal, water may be viewed as anessential component of the
body. However, water contains no energy that can be used to convert
animal feed dry matter into another form nor is any energy derived
directly from water when it is lost from the tissue of the animal body.
In GRAZE, it is assumed that water is always available to the animal.
Nutrients
Nutrients are categories of feed dry matter that are consumed by the
animal and then converted into another form as part of the animal
body tissue.
Energy
Energy is used to convert the feed dry matter that is consumed into
matter that makes up the body of the animal. In terms of the animal,
energy is generally viewed as being either chemical or thermal.
Chemical energy is contained within the material
consumed or in the body of the animal.
Beef-Forage
GRAZE:
A
Model
Selective
Grazing
of
335
Feed DryMatterComposition
The intake of "feed' dry matter provides the material and energy
required to "build' (grow), maintain, and operate the animal body.
An animal is literally what it eats. While the GRAZE plant submodel
simulates the presence of nitrogen, phosphorus, potassium, and
stored sugars (carbohydrates), the animal submodel also considers
several other categories. In GRAZE Version 2.3, some of the values
were assumed to be constant percentages of grazed (as opposed to
fed) forages as noted below. GRAZE Version 3.02 (Windows version)
considers these components to be part of the plant parameter inputs.
~ l a ~ t In GRAZE, plant material is divided into cell content material and cell wall material (Fig. 11).
Cell Content: Cell content is the material contained within a
cell. It is generally considered totally digestible in a relatively short
period of time, less than 24 h. The cell content contains the soluble
cell minerals (not silica), soluble protein, nonprotein nitrogen, organic
acids, sugars, soluble carbohydrates, starch, and pectin.
Cell Wall: Cell wall material may be categorized as
1. Potentially digestible cell wall: material (cellulose, hemicellulose) that can be digested in the animal body given sufficient time. However, some
of this material would be expected to exit the body before being totally digested.
2. Indigestible cell wall: material that will not be digested by
the animal regardless of the time that the material spends
in the digestive tract. Indigestible material is composed of
lignin, cutin, silica, tannins, essential oils, and polyphenols.
C Q ~ ~ Q ~ eThe
~ t feed
s . dry matter is composed of
minerals, protein and nonprotein nitrogen, fat, and carbohydrates.
~ i ~ e r a l sAlthough
:
at some level all minerals are important,
calcium and phosphorus are the key minerals considered by the
GRAZE animal submodel. Minerals do not contain energy for purposes of animal growth or maintenance. GRAZE Version 2.3 (DOS
version) assumes that the cell content of grazed material is
calcium and 27.13% miscellaneous minerals not including calcium and
phosphorus. In GRAZE for Windows (Version 3.02),these parameters
are assigned as part of the input data file. The potenti~llydigestible
~ t ~ o ~
i io^. The rate
at which nutrients and energy are made available for animal growth
and maintenance is a function of the rate of digestibility. Oneway of
estimating digestibility is through a series of laboratory procedures,
referred to as thedetergent system, that result in quantifying the neutral detergent fiber and acid detergent factor(Fig. 20)(VanSoest,
1982).
F Z ~ e n tdetergent fiber ( N D ~ ) . In simplest terms, N D F is the dry
matter that is left over after the plant material is exposed to
% = 100 - NDF, %
(22)
after the neutral detergent fiber material has been further exposed to an acid solution. ADF material constitutes the indigestible cellwall material. Thus, by process of e~mination,the
337
1.O NDF
I
DETERGENT
FIBER
(ADF)
IGURE 20
Divisions
NRF, %) - ADF, %
(23)
6.
33
Loewer
GROSS FEED
ENERGY INTAKE
FECAL
ENERGY
l
FIGURE21 A model of animal utilization of energy supplied by feed.
~ ~ s e ~or o~~su c t s
Gaseous products of digestion
include the combustible gases produced in the digestive tract^ by the
fermentation of the ration. GRAZE computes the production of methane, which is by far the dominant gas produced and represents a
s i ~ i f i c a nloss
t of energy only in ruminants such as beef cattle.
Urinary energy is the gross energy of the
urine, it includes the energy content of the nonoxidized portion of
the absorbed nutrients and the energy contained in the endogenous
33
(body) fraction of the urine. GRAZE computes an energy loss associated with conversion of the various types of body stores including
maintenance.
~ e t a b ~ l i z a b l e E n e Metabolizable
r~.
energy is the energy that
remains after subtracting from the gross energy intake the fecal energy, energy in the gaseous products of digestion, and urinary energy.
Essentially GRAZE computes metabolizable energy, from which it
then proceeds to determine the extent to which other biological processes occur.
ea^ ~ n c ~ e ~ eHeat
n ~ .increment is the increase in heatproduction following consumption of food when the animal is in a thermally neutral environment. It consists of increased heats of fermentation and of nutrient metabolism. GRAZE considers heat increment
in that theinefficiencies of digestion are converted to thermal energy
and added tothe body heat energy pool.
Many of the concepts just discussed are mental models that form
the basis for mathematical models that are needed to quantify biological processes. The key concept to remember in GRAZE is that
there is conservation of energy and mass as the feed supply moves
through the animal. At the same time, energy and mass are being
converted to different forms by biological and chemical processes.
The GRAZE model attempts to relate these processes in a mechanistic
manner reflective of the actual biological processes.
1.
Laewer
PhysiologicalAgeandWeight
el of Selective ~ r a z i n ~
FIGURE212 Relationshipbetweenchronologicalage
weight.
342
Loewer
Relationship between emptybody live (full) body and physiological body weight.
A s s ~ ~ ~ ~ ~ Q ~~ ~s ~ Given
Q
s the~above
~ categories,
Q
~
.
the following assumptions are made:
There is an independent physiological weight-age relationship for eachtissue component that may be expressed mathematically as a segmented sigmoid curve extending from
conception to maturity (Fig. 25).
2. Each unit of physiological growth will contain each of these
components in a predictable ratio that changes simultaneously with physiological age (Fig. 26).
oewer
6 Physiologicalweight-agerelationships.
3. Water cannot be stored as part of tissue in excess of its physiological maximum for a given physiological age,
thus making empty body water content the most accurate measure of
physiological age.
4. Nitrogen and essential minerals may be stored in excess of
their physiological maxima at a given physiologicalage but
not in quantities greater than constant percentages of their
hysiological maxima (Fig.27).
at storage is not limited by a physiological m a ~ m u m .
ese assumptions is addressed ~dependently.
where
RGAIN = potential change in weight in a particular component (kg/ day)
RGAINM = maximum potential rate of gain of a particular component (kg/ day)
XNOW = physiological age (days)
If the animal has not yetreached the inflection point
of the sigmoid curve,i.e., where maxi mu^ daily
rate of gain occurs,
RGAINC = the growth rate of that component at conception
(kg/day)
TCON = number of physiological days from conception to
the inflection point
If the animal is past the physiological age where
maximum daily rate of gain is obtained
Loewer
Each component is assumed to follow its own physiological weightage relationship independently of the others. However, all components have physiological age as a common measure of comparison
(Fig. 26). Thus, when physiological age is known, the instantaneous
rate of physiological change of any component can be computed. As
physio~o~cal
age changes, the relative ratios among components
change, but always in a mathematically predictable manner for a
TCON
IGURE 28
TCON
-7
Eoewer
IGURE 29
Nutrient pathways.
ii
'
I
'
""_
" "
35
Loewer
PRODUCTION
NUTRIENTS OR ENERGY
FOR FETAL DEVELOPMENT
NUTRIENTS OR ENERGY
FOR BODY ~ A I N ~ N A N ~ E
IGURE
"L
"""""-
I
I
I
'
35
I
I
PRESENTAGE
352
oewer
PR~SENT
el of Selective ~ r a z i n ~
35
FIGURE34 Utilization process when there is insufficient energy for maximum gain. Step l a for fat or protein only. Steps 2 and 3 for excessnonenergy
Component.
35
Loewer
PRESENT AGE
355
35
Eoewer
PRESENT
In Situation 4 (Fig. 363, the ratio of amount of component available above maintenance (from either digestion or"excess") to the
maximum daily physiological growth potential Wp' is the fraction of
physiolo~calgrowth that will be obtained. The situation in which
excess material is used is shown in step 1. Otherwise, only step 2
occurs. In either case, the changes in physiological weight gain and
age are depicted by Wg' and Wa', respectively, and physiological age
will advance no further than that of the limiting component.
S i t ~ ~ t5.i ~There
~
are sufficient quantities of both energy and
component to exceed both maintenance and physiological growth
potential.
In Situation 5 (Fig. 37), in which neither energy nor components
are limiting, maximum physiological growth potential is obtained
(step I). Material that remains after physiological growth is satisfied
is converted to excess up to themaximum limit allowed, Wsmax(step
2).
357
The situations just described apply to all components. However, components differ in their levels of excess storage and in the energy required for maintaining and synthesizing tissue. Giventhe component
categories of water, fat,nitrogen (in the form of protein), and essential
minerals (grouped into a single category), consider how these factors
interact.
Let Em be the bodys daily demand for maintenance energy for
all components, and Ed the energy available each day from digestion.
The only sources of body energy are fat and nitrogen (in the form of
protein). The energy above maintenance required for maximum physiological growth (Emaxpg) is
(25)
Emaxpg = Wpn*Epn + Wf*Ef
where
Wpn, Wf = maximum physiological weight change potential
for nitrogen (as protein) and fat, respectively
Epn, Ef = energy content per unit of weight for nitrogen (as
protein) and fat, respectively
The total energy available to meet maintenance and growth
needs in the living animal (Eta) is
Eta = Ed + Rens*Wen*Epn + Wpn*Epn
+ Refs*Wef*Ef + Wpf*Ef
where
Ed = energy available from digestion
Rens, Refs = maximum daily decimal percentages of excess nitrogen (as protein) and fat, respectively, that may
be removed form excess material
Wen,Wef = excess weights of nitrogen (as protein) and fat,
respectively
Wpn, Wpf = weights of physiological nitrogen (as protein) and
fat, respectively
Let Eag be the total energy available for physiological growth
and additions of excess material. Then
Eag = Ed + Rens*Wen*Epn + Refs*MTef*Ef
(27)
Let Eexbe the total energy available from excess protein nitrogen and
fat for physiological growth. Then
Eex = Rens*Wen*Epn + Refs*Wef*Ef
(28)
358
Loewer
If Eag is less than Em, physiological age will decrease with accompanying losses in physiological weight all components to the extent
necessary to maintain an energy balance (Em - Eag) (see Situation 1
and Fig. 33). If Eag is greater than Em but less than Emaxpg, then
PA' = (Eag - Em)/Emaxpg
(29)
(30)
(see Situation 5 and Fig. 37). By definition, PA' cannot exceed 1; that
is, changes in physiological age cannot be greater than changes in
chronological age. Energy supply in excess of Emaxpg is converted
to excess nitrogen protein until (1) nitrogen becomeslimiting,(2)
maximum nitrogen storage is obtained, or (3) energy becomes limiting. Inthe first two cases, the remaining energy is converted to excess
fat, there being no limit as to excess fat storage.
In the case of nitrogen or essential minerals, the model assumes
that a constant percentage of the physiological weight is excreted, this
quantity being referred to as the maintenance level. The energy required for maintenance may be specified as a function of body composition or total body metabolic weight, the latter being used in our
current modeling efforts. Situations 3,4, and 5 apply as shown
Figs. 35-37.
In summary this conceptual approach to modeling limits physiological age and physiological growth either to the most limiting
body component or to maximum genetic potential. "Excess" growth
may occur in some components even when there is negative physiological growth.
7.
Laboratory data must be used to mathematically describe the concepts discussed above. We have drawn heavily on bodycomposition
studies conducted at the University of Missouri as reported by Trowbridge et al. (1918), Moulton et al. (1922,1922a, 1922b, 1923),Moulton
(1923), and Ellenberger et al. (1950). Although the studies are old,
they are complete with regard to data and represent major efforts in
determining body composition. The overall goal of these researchers
35
Loewer
8. Observations
el of Sekxtive Grazing
361
to gain at the maximum physiological rate while adding excess stores of all components.
When there is a severe shortage of energy or nutrient(s) (Situations 1 and 3), the animal will becomephysiologically
younger, always maintaining an energy and nutrient balance. Backward integration is used when the animal becomes physiologically younger. Currently,
the model does not
limit the animals ability to regress along the physiological
weight-age curve. Realistically, at some point the process becomes irreversible and death occurs.
ConcludingComments
(31)
363
AIN
FIGURE
Heat
in an animal.
The"ExternalDrivers"-Weather
input
36
Loewer
l
"
IGURE
date (day of the year), (2) latitude, (3) minimum daily temperature,
(4) maximum daily temperature, (5) time lag for minimum temperature after sunrise, (6) time lag for maximum temperature after solar
noon, and (7') coefficient that controls temperature decrease at night
(Fig. 6). Note that factors 5 and 6 were not mentioned in the discussion on plant growth but may be used to "shift" the temperature
curve over the day if desired.
The model computes sunrise, sunset, and day length, using latitude and the Julian data. Temperatures over the day are computed,
using relationships developed by Parton and Logan (1981).
~~~~~.
of Selective ~ r a z i ~ ~
H U ~ I DRATIO
I ~
W
366
Loewer
MAX
MIN
MID NIGH^
SUNRISE
SOLAR
NOON
SUNSET
F
FIGURE 41 Radiation distribution based on
projections
maximum.
of the daily
often the case. Usingthe above assumption, GRAZE assumes that the
absolute humidity can then be determined by entering its psychrometric chart subroutine with the existing temperature and the absolute humidity value (Fig. 40). A maximum daily relative humidity
other than 100% also maybe specified by the model user; the default
value currently used is 90%.
~ 0 ~ 2 ~~ i~ Although
~
~ solar
o radiation
~
. is not used directly
in the plant submodel, it can be a significant consideration in the
animal submodel. Solar radiation is entered as the maximum received
per hour,which is assumed to occur at solar noon (Fig. 41). The
amount of radiation received from sunrise to sunset is assumed to
follow a parabolic function of the form
R = Rmax - ( 4 , 0 ~ R m a x / ~ A Y ~ ~ ~ 2
-)TIME
~ ( 1 4+
4 TIME~~2)
(32)
where
CRAZE:
367
The following sections examine how the model uses these environmental factors.
3.
ResistancetoHeat
~ n t e ~ n a 2The
. internal resistance to heat flow from the body, RI,
was cited by Webster (1974) as
RI = (115.0 + 0.387*WT + 2.07*T)/1000
(33)
where
RI = internal resistance, C m/ W
WT = live animal weight, kg
T = ambient temperature, C
Webster states that Eq. (33)is valid for temperatures below 0C. Note
that the resistance to heat flow decreases with a decrease in temperature, reflecting an attempt by the animal to maintain the temperature
of its outer extremities. However, one also would expect the resistance to heat flow to decrease at higher temperatures when the animal is experiencing heat stress and is trying to void itself of as much
heat as possible. Themodel reflects this situation by assuming a linear
decrease in internal resistance to heat flow from the value computed
at 0C to a zero value at body temperature, but never to be less than
a stated minimum value. In effect, the assumed relationship between
internal resistance to heat flow and ambient temperature is as shown
in Fig. 42.
(34)
\
Internal resistance to heat flow.
flow,
winter coat begins 12 weeksafter the longest day of the year (Yeates,
1955), or Julian day 256. Linear growth and shedding rates are assumed (Fig.44). Hair lengths were 1.12-2.02cm in one group of
exposed cattle (Webster et al., 1970), with another report implying a
1.0-2.0 cm range (Webster, 1974).
The combination of Eqs.(33) and (34) is used to
compute heat loss potential per hour per square meter of surface area
U = (1/RE
+ 1/R1)*0.86kcal/W-h
(35)
SA = 0 . 0 9 ~ ~ * * 0 . 4 7
(34)
where SA = surface areal m2, and W" = live weight of the animal,
kg*
75
165
256
346
365
Loewer
5. SensibleLossesfrom
Surface
AMBT)
(37)
where
HTLSUR = heat losses from the surface, kcal/h
RECTLT = core temperature of the animal, "C
AMBT = ambient temperature, "C
6.
LatentLossesfrom
Surface
Latent heat exchange is associated with a change of state; that is, for
animals, it is the heat exchange associatedwith evaporation of water.
NET HEATEXCHAN~E
a. Reradiation
b. Surface sensible
c. Sky temperature
d. Ambient temperatu~
HAIR T E ~ P E ~ A T ~ R E
SKIN T E ~ P ~ R A T ~ R E
CORE T E ~ P E R A T ~ ~ E
in an animal.
Beef-Forage
GRAZE:
A
~o el of Selective Grazing
371
Animals may be classified as "sweaters" or "nonsweaters." Beef animals are classified as "nonsweaters" in that they do not lose significant amounts of moisture through their skin surface. Thus, GRAZE
considers the latent heat loss from the surface to bezero, whereas the
heat loss would need to be considered for the "sweater" category
7.
RespiratoryHeatLosses(SensibleandLatent)
Heat loss through respiration includes both sensible and latent losses.
When the animal takes in air at a given temperature and relative
humidity; heat from the body warmsthe air in thelungs to near body
temperature, resulting in sensible heat losses for the animal. At the
same time, the warmed air in the lungs approaches saturation following anadiabatic process. Thisremoval of heat from the body through
the removal of moisture is referred to as "latent" heat loss. (By definition, latent heat loss is the heat transfer associated with the removal of moisture.)
The animal can increase its respiration rate above some minimum level up to some maximum value. Respiration is highly correlated with surface area. Minimum latent heat loss is given by Baxter
and Wainman (1964) as 17 W*m2,which is about equivalent to the
latent losses resulting from a volumetric respiratory airflow rate of
0.084 m3/min per square meter of surface area. The maximum volumetric respiratory airflow rate was given by Guyton (1947) as 0.114
m3/min per square meter of surface area.
The GRAZE model allows the model user to specify other respiratory airflow rates if so desired. Likewise, the degree to which
ambient air will be warmed and saturated with moisture also may
be altered for both the lower and upper respiratory rates. The ambient
air conditions are specified in terms of temperature and relative humidity.Thesensible and latent heat losses can then be computed
usingstandardpsychrometric
re~ationships(Figs.
and 4'7). The
computations are carried out in the following order.
~ T S M N
= RESMIN~PSYS~N~SA~C
where
RHT~MN= respiratory sensible heat losses associate
minimum respiratory rate, kcal/ h
RESMIN = m ~ i m u mrespiratory area, m3 (min m2)
(38)
Tdp
IGURE
Twb
Tdb
46 Psychrometricchart.
where
TSMX = respiratory sensible heat losses associated with the
maximum respiratory rate, kcal/ h
~ E ~ M A= X
maximum respiratory rate, m3/(min m)
PSYS~X= sensible enthalpy difference between incomin
existing air as calculated through psychrometri~relationships, kcal/ m3 air
37
l00 % RELATIVE ~ U ~ I D I N
DIN
IN EQUILIBRIU~IT^
BODY TE~FERATURE
TLMN = RESM~N~PSYLMN~SA~C
where
TLMX = RESMAX~PSYLMX~SA~C
lloewer
37
where
= respiratory latent heat losses associated with the
~ T L M N
minimum respiratory rate, kcal/ h
PSYLMX = latent enthalpy difference between incoming and
exiting air as calculated through psychrometric relationships, kcal/ m3 air
+ RHTSMN + RHTLMN
(42)
where HTLSMN = the minimum possible heat losses that the animal
can experience,kcal/ h.
x i m Total
~ ~ Heat Loss
HTLSMX = HTLSUR
+ RHTSMX + WTLMX
(43)
Zone
ThermalNeutrality
HeatProduction
e ~ ~ ~ o l i sTo
m survive,
.
a warm-b~oodedanimal must
to maintain its body temperature within a relatively
narrow range. The functionings of the body, such as circulating blood,
breathing, and digestion, require the expenditure of energy, with heat
being a by-product. The minimum heat production level associated
with maintaining life is referred to as basal metabolism, and the reis
lationship, given by Crampton and Harris
ASL = 7 0 ~ A C T ~ ~ ~ 0 . 7 5
(44)
rk. ~ e t ~ ~ o l i Basis:
c ~ Ife the
i ~animal
~ t
is active, the total heat produced includes some allowance for the
el of Selective Grazing
SURFACE LOSSES
375
SURFACE LOSSES
4-
~ I N I ~ U ~
RESPIRATORY
LOSSES
RESPIRATORY
RESPIRATORY
HEAT E X C H A ~ G ~
a. Latent
b. Sensible
e. Respiration rate
d. Ambient temperature
e. Ambient relative humidity
f. Core temperature
g. Exchange efficiency
FIGURE
work associated with this activity; and the relationship used (Lofgreen and Garrett, 1968) is
MAINT = 77*ACT~T**O.75
(45)
+ Ewm*VV
Ewm are energy required for maintaining a unit of protein, fat, minerals, and water, respectively. At present, however, Eq. (45) is being
used. APL maintenance energy is added to the body heat energy pool.
~ ~ ~ tof~~ ~i ss s i ~~s ~x ~ ~ s When
s
tissue (tissue is defined as ~h~siolo~caPLy
essential material) or excess fat is formed,
there is an inefficiency that results in part of the pool of body chemical energy being converted to heat energy. Themodel adds this heat
energy source to the total body heat energy pool. The process also
occurs if tissue or excess fat is converted to chemical energy.
S.
In GRAZE, the growing fetus and its associated body tissue are treated the same as other tissue insofar as efficiency of gain is concerned. The differenceis that the body will not
pull fetal energy back into the chemical energy pool.
~
~ As
0 with
~ tissue
.
and fat, thermal energy is produced
when chemical energy is converted to milk (Moe and Flatt, 1969). All
milk that is produced is assumed to exit the body, and therefore the
model does not allow milk to be converted back into chemical energy
for reconversion to body tissue or to meet other energy demands.
Heat Gain: In daytime, the animal may receive direct energy from the sun in the form of radiation. However,
only a fraction of the animal's surface will be exposed. It would be
logical forthe animal to increase exposure during colder weather and
ecrease it in the summer, but no attempt has been made to model
this occurrence. Instead, the model user specifies the fraction of the
animal's surface area that is exposed to radiation. The quantity of
heat received is expressed by
*(144.0 - 2~.0~TIME
+ TIME*"2)]
*O.O1433*6O~O*SA*~CA~EX/lOO.O
(47)
377
where
ENRDHR = radiation received, kcal/ h
RMAX = radiation at solar noon, kcal/h
D A K = day length, h
TIME = time of day, h
SA = surface area, m2
PCANEX = percent of animal exposed to radiation
Heat Loss: The animal may reradiate heat energy to the sky at
night. Equations (48) and (49) are used to simulate this situation for
the skin and hair coat temperatures.
(48)
Tskin = (Tair + RE/RI*Tcore)/(1.0 + RE/RI)
Thair = (Tskin - Tair)/4.0
+ Tair + 273.0
(49)
where
Tair = temperature of the air, "C
Tcore = body core temperature, "C
RE = external resistance to heat flow [see Eq. (34)]
RI = internal resistance to heat flow [see Eq. (33)
The effective sky temperature (K) is then calculated as
Tsky = 0.0552*(Tair + 273.0)**1.5
(50)
Emittance of the hair coat is assumed to be 1.0, in which case the
reradiation of the animal to the sky becomes
ENRDHII = l.O*SA*SF*SBC*(Tsky**4
- Tair**4)*0.1443*60.0(51)
where
ENRDHR = radiation losses from the animal, kcal/ h
SA = surface area, m2
SF = fraction of surface area exposed
ST3C = Stefan-Boltzmann constant, 5.6697E-8
The major weaknesses in this logic is not the basic heat transfer
mechanisms but rather the computation of the fraction of surface area
that is exposed. Future logic may consider amaximum and minimum
possible exposure and allow the animal to use solar radiation or reradiation to its best advantage insofar as body temperature is concerned. Presently the model user has the option to bypass the solar
radiation logic.
37
oewer
~~~~~.
~~~~~~.
Environmental factors (i.e., temperature, relative humidity, and solar radiation) establish the short-term heat losses from
the animal. Wind, rain, and snow effects may be added to the model
at a later date if desired. In GRAZE, the animal may alter its chemical
and physical systems using such options as regulating blood flow,
respiration, shivering, and eating to maintain a constant body
temperature.
379
" " " " " " " " " " " "
'
"""-"""_""""-
terial that is actually digested is dependent upon the time that the
material resides in the animal's body.
In GRAZE, the user may specify a feed source to simulate a
feedlot situation or the animal may obtain its feed supply through
grazing. In addition, pastures may be supplemented with external
feed sources, the assumption being that the external sources will be
consumed first, regardless of quality. If an external feed source is
used, the following inputs must be specified. These inputs are computed internally in a grazing situation.
Loewer
(534
PKG, CCNPKG,
CCFTKG,
CCCAKG,
CC
are the portions of a kilogram of cell content
tein, nonprotein nitrogen, fat, calcium, phosphorus, and other minerals, respectively.
The remaining portion of the cell content is designated as
carbohydrates:
wh
cc
tract), kcal/ kg
PAFIBE = energy content of fiber (fiber extract), kcal/kg
+ CCCR~G~PANFEE
(56)
where GEFGCC is the gross energy per kilogram of cell contents, kcal,
GEKGCW = CWPKG*PAPRE
+ CWFTKG*PAEEE
+ CWCWG~(PANFEE+ PAFI~E)/2.0
(57)
Loewer
1.o
POINT AT WHICHDRY
1 MAYTERINTAKERATE
BEGINS TO BECOME
LIMITEDBYDRYMATTER
!
0.5
0.0
IGURE 50
centage of the animals body weight. In beef cattle, this percentage generally ranges from 2 to 3% per day. Depending
upon the degree of accuracy required, it may be necessary to
make some allowances for physiological maturity compared
to actual weight if the animal is overweight for its age.
Counting ~ u ~ and/or
~ e Size
r of Bites. Another approach is to
assume that the animal will take a given number of bites of a
certain size over a given time interval.
~ s t i ~ a t i nDigestibility.
g
Estimates of digestion can be obtained
for different feed sources. If the amount of fecal material deposited per day is known, daily intake can then be estimated.
~ e c h a n i s t i c A p p r o ~ c h e s . In CRAZE, the intake feedsource is
placed into three "quality" categories based on cell content
and physiological age of the material. Feeds in each category
may be consumed at a rate that depends on the maturity of
the animal.The intake rate is determined based on the
weighted average of the three categories of material that is
consumed after accounting for forage density, total availability, and whether the animal may eat as influenced by physical
fill, chemostaticlimits, thermostatic limits, or rest requirements.
4.
EAT IN^
ENDS
IGIJRE
intake.
el of Sele~tive~ r a ~ i n ~
PTRMAX = maximum rumen pass-through rate, percent of mmen capacity per hour
As dry matter enters, it is placed into (1)cell content (assumed
to be 100% digestible), (2) potentially digestible cell wall, or (3) indigestible cell wall. The quantities of cell content and potentially digestible cell wall material are either digested in the rumen (each following a modifiedhalf-life digestive rate) or passed through the
rumen. Cell content that passes through the rumen is digested later
in the body.Potentiallydigestiblecell
wall material that passes
through the rumen before being digested exits the body. Indigestible
cell wall can only pass through the rumen and be excreted.
The pass-through rate function is a symmetric sigmoid curve
based on the integral of a parabola (Fig. 52). The flow rate is zero
when the rumen is empty and equals
(PTMAX/ lOO~RCFAC*physiolo~cal
weight
when the rumen is totally full. The pass-through material is in proI.o
0.0
EMPTY
PASS-THROUGH
(kglhr)
FULL
MULTIPLEB
FACTOR
MAXI~UM
PASS-THROUGH
RATE
(kgkr)
386
loewer
CELL CONTENT
POTENTIALLY
DIGESTIBLE
CELL WALL
INDIGESTIBLE
CELL WALL
TY
ULL
IT
Beef-Forage Model
Selective
Grazing
of
387
ELDPCW,
ELDFCW,
ELDCCW,
ELDMCW,
ELDCAW,
ELDPHW, and ELDNPW = the thermal energy losses associated with the digestion of l kg of cell wall protein, fat, carbohydrates, other minerals, calcium, phosphorus, and nonprotein nitrogen, respectively.
The total digestive losses per kg of dry matter component become
DLCCJSG = CCPKG*ELDPCC + CCFTKG*ELDFCC + CCC
*ELDCCC + CCMNKG*ELDMCC + CCCAKG
*ELDCAC + CCPHKG*ELDPHC + CCNPKG*ELDNPC
(59)
where DLCCKG is the digestive energy losses per kilogram of cell
contents digested, and
DLCWKG = CWPKG*ELDPCW + CWFTKG*ELDG~~
CWCRKG*ELDCCW + C ~ N K G * E L D M C W
+ CWCAKGeELDCAW + CWPHKG*ELDPH~
+ CWNPKG*ELDNPW
where DLCWKG is the digestive energy losses per kilogram of cell
wall digested. The digestive losses may be used to calculate the heat
increment losses of the animal once the digestive rates are determined.
It is difficult to determine experimentally the portion of digested
energy that is lost as fermentation heat. Therefore, theoretical analyses were used to obtain the fermentation-associated heat losses for
protein, carbohydrate, and fat using, in part, the theoretical heats of
combustion (Lange and Forkew, 1952).
Hershberger and Hartsook (1969) reported experimental values
for fermentation heat loss of 5.5 and 5.1% of the heat of combustion
of alfalfa hay fermented. Also cited is work by Marston (1948) in
which 6% of the combustible energy of the fermentation products was
lost as heat where cellulose was the substrate. A theoretical analysis
for carbohydrate by Wolin (1960) was used to calculate a heat loss
value of 6.4% of substrate heat content.
Reaction equations were formulated, and efficiency calculations
were made. The equation used by Wolin (1960) for carbohydrate is
57.5(C6H,,O,) = 65 acetate +
+ 60 CO,CH,
+ 35
+ 25 H20
propionate
+ 15 butyrate
(61)
(62)
(63)
An assumption is made that the heat loss values are the same
whether the material is from cell content or cell wall.
Energy losses also are associated with the addition of tissue
components. Moe and Flatt (1969) reported an efficiency of 75% for
the addition. of body fat from metabolizable energy.Pullar and Webster (1977) reported an energy requirement of 53 kJ/g (12,661 kcal/
kg) of fat or protein tissue added. Therefore, the efficiency should be
el of Selective ~ r a z i ~ ~
(65)
(6%)
Energy losses associatedwith the removal of tissue must also be considered. The energy associated with the removal of nitrogen is 7.9
kcal/gN removed (Crampton and Harris, 1969). This nitrogen compound then contains 1.264 as much energy as is normally associated
with protein. The energy loss is estimated to be
(1.264 - 1.0)*5686kcal/ kg of protein
= 1501.1 kcal/kg of protein removed
The efficiency of fat removal is estimated to be 99% because very
little heat loss is associated with fat removal. The energy loss associated with removal of fat is
0.01*9367 kcal/kg of fat = 93.67 kcal/kg
7 . Regulation
(69)
Intake
Z
~ GRAZE
~
~integrates
~
the
. sum of all rates
into and out of the rumen (Fig. 49). Oncethe animal initiates eating,
the process continues until the animal is filled (Fig. 54), unless other
limits (discussed later) are imposed first. All this time, digestion and
pass-~hroughempty the rumen. Emptying the rumen will not trigger
390
Loewer
391
was compared for both groups with the essential fat curve used
the model that was developed from Moulton data on starvation. The
animals in Group I1 followed the model's essential fat curve fairly
well. Theanimals fed for maximum possible gain showed much more
fat addition. It was hypothesized that the maximum rate of fat addition was proportional to the amount of essential fat present in the
body. This allows the rate to increase as the animal increases in physiological age until maturity. At maturity, the maximum possible rate
should be constant. The actual maximum rate may be much larger
because of environment or other factors, such as increased maintenance requirements or restricted feed availability, and because excess
fat may be deposited during only part of the day due to ~uctuations
in the chemical energy level of the blood.
The data show that from approximately 34 to 48 months of age
the rate is a constant, which is assumed to be the least maximum
daily rate of fat addition. This rate is calculated as
((285 kg - 169 kg)*12mo*l yr)/((48 mo - 34 mo)
*lyr*365 days) = 0.272 kg/day or = 0.60 lb/day
(70)
392
Loewer
393
where
BW = blood weight, kg
~TFTFR
= fat-free empty body weight, kg
FTWT = weight of total fat, kg
The R2 value for the equation based on these data is 0.996, The
values predicted agree fairly well with data from Marcilese et al.
(1966) for Herefords of 5.12 L/lOO kg of live body weight assuming
a specific gravity of blood of 1.05. Using Eq. (73), an animal with
20.10 kg of Eat would have 4.69 kg blood/kg empty body weight.
Converting to live weight using a factor of 1.14 gives 5.35 kg blood/
100 kg live weight, which is near the previously mentioned value.
Paine (1971) used a value of 7.7 kg/lOO kg live weight, with provision
far some adjustments. Haxton et al. (1974) showed that calves with
a mean age of 9 weeks had blood/weight ratios of 11.6 kg/lOO kg
live weight. The regression equation, therefore, provides a reliable
estimate of blood weight, except perhaps in the case of very young
animals.
At present, the energy concentration that initiates eating is set
at the base of zero energy level. The energy concentration above that
level is computed by dividing the excess energy in the blood by the
total kilograms of blood as calculated from Eq. (73). At some energy
concentration above the base, the animal will stop eating. Eating will
resume when the energy concentration in the blood returns to the
base level. Paine (1971) used a single value to initiate and stop eating
and based his values on the acetate energy concentration only. A total
energy concentration value comparable to Paine's acetate value of 5
kcal/kg would be 8.77 kcal/kg. However, the GRAZE model value
is for energy concentration above a base level, which is therefore not
directly comparable, Currently, GRAZE uses a single value near the
base for both initiating and terminating eating. The use of a single
value sometimes contributes to relatively rapid cycles of on-off eating actiuity, which merits further review in later versions of GRAZE.
~
~ The
~ body
~ heat
0 pool
~ determines
.
body
temperature. Once the maximum rate of heat removal has been
reached [Eq. (43)j, the body temperature may begin to rise. If body
temperat~rebecomes sufficiently high (presently 39"C), the model assumes that the animal will stop eating to prevent a d ~ t i o ~heat
a l load
(Fig. 56). In the model, low temperatures may trigger eating indirectly
by increasing maintenance and thereby reducing the body chemical
energy level more rapidly.
394
BODY ~ ~ P E R A T ~ R E
BODY TEMPERATU~E
WHERE EATING
WHERE EATING
~~
g ~ of ~ ~ ~ ~ There
~~~is a atime
' u during
~ ~ ~the
~
evening when no eating occurs (Castle et al., 1950). The model uses
two inputs to regulate nighttime eating:
TMEATE = miitary time that intake ends
TMEATB = military time that eating may begin
When TMEATE is reached, the animal ceases to eat irrespective of
fill, chemical, or thermal conditions (Fig. 57). Eating may begin only
when TMEATB is reached. However, fill, chemical, and thermal regulation may override TMEATB,
intake
involve
a .number
~
~ Dry matter
~
~ and digestion
u
~
of feedback mechanisms and governing regulations, In GRAZE, eating may be initiated only by a sufficiently low blood energy level.
However, eating will cease if body temperature rises sufficiently
above normal, if blood energy levels are sufficiently high, if sleeping
occurs, or if physical fill is reached. These concepts are illustrated
simultaneously in Fig. 58.
.~ o
Selective
~
e Grazing
~
FIGURE5'7 N i g ~ t t i ~
rest
e regulation.
The fourth condition helps to explain why two similar females of the
same weight may not both conceive if one is gaining and the other
is losing weight. Effectively; the stored levels of energy and/or nutrients may be significantly different in the two animals, reflecting a
body composition difference. This concept is still in the verification
phase.
396
TIM
IGURE 58
2.
Fetal Development
Loewer
LIMITED BY
GENETIC
POT~NTI~
ENERGY. OR
NUTRIENTS
ENERGY OR N U ~ R I E ~ S
'
PRIM~ILY
BY
NUTRIENTOR
! ORENERGY
N
I TAKE!
demand with
399
DIGESTED
FEED
l
lactation begins. These two points are connected by a parabolic function similar to that shown in phase 1 of the lactation curve. Thus,
GRAZE effectively assumes that reaching "genetic potential" may be
limited by factors other than milk components until m a ~ m u mproduction is reached. Mills production cannot exceed the genetic potential and may be limited to lesser milk output by demand for milk or
the availability of milk components.
GRAZE determines the daily milk output by first"looping"
through each simulated day every DT time step tomaintain a current
value of nutrient stores and blood concentrations of nitrogen and
energy. DT will typically be 15 min. The milk logic begins by establishing a proposed mills composition based on minimum percentage
levels of calcium, phosphorus, other minerals, fat, and protein (nitro-
A brief overview is presented to tie together the three major submodels ofGRAZE (Figs. 1 and 2).
Loewer
2.
3,
4.
S.
Physical fill of the rumen. If intake ceases because of physical fill, a portion of the material must be emptied from the
rumen before the animal can resume eating regardless of
other control factors.
Upper blood energy chemostaticlimits.Relatively
lower
chemostatic limits must be reached beforeeating can resume
regardless of other control factors.
Internal body temperature. Before the animal can resume
grazing, its temperature must fall below a defined upper
limit.
Time of day. Animals do not eat during certain designated
hours during the evening, regardless of the status of the
other control mechanisms.
The computational interval for the animal ,model. The
grazing status of the animal remains constant over the computational interval. This interval is an input parameter that
may be varied (typically, a
min interval is used).
03
The GRAZE model contains selective grazing logic that interfaces the
plant and animal submodels. The major divisions of the GRAZE interface logic relate to plant quality and the conceptual division of a
pasture into partial grazing areas. The grazing animal selects from
among the partial grazing areas based on differences in forage quality
and density (kg/ ha).
I.
to Plant Quality
more, the faster dry matter is digested, the greater the potential for
intake as long as chemostatic, thermostatic, and restricted gra
time controls are not in effect, For example, Holloway et al. (1
observed that the daily dry matter intake of lactating beef cows increased with increases in forage quality, The higher the ratio of cell
content to cell wall material, the greater the dry matter digestion rate
ven the same percentage of indigestible dry matter. Inthe partitions
of dry matter, new material generally contains a higher percentage of
cell content material. Accordingly, in GRAZE it is assumed that for a
given forage the dry matter intake rate for a diet of entirely new
material would be higher than for a diet composed entirely of old
aterial. Similarly the animal would be expected to ingest old maerial at a faster rate than dead material. In essence, "palatability"
and "quality" are defined in GRAZE to be functions of the new, old,
and dead plant partitions. An input to the model is the rate (kg/h)
at which the animal would independently consume each of these
ant categories. By definition, "new," "old," and "dea
fferent physical components of the plant bu
nce, GRAZE assumes that each bite the animal takes is in
direct proportion to the availability of new, old, and dead material.
The rate at which material is ingested is a weighted average of the
portion of the plant category and its potential intake rate. A somewhat similar approach was used in a modeling effort by Johnson and
Parsons (1985) in that the independent weights of four structural
plant components were computed. The contributio~of each of the
structural components to the leaf area index was used to develop
their algorithm for modeling selective grazing.
*
2.
SelectiveGrazingLogicasRelated
to ForageAvailability
The m a ~ potential
m ~ intake
~
rate may be reduced by forage availity (kg/ha). The premise is that at some point dry matter availity will begin to limit intake rate &/h) independent of forage
nctional form used in GRAZE is shown in Fig. 50. This
similar to that reported by the N
and discussed by Minson (1983) and
The following equations are used to determine maximum potential
intake rate:
aximum potential intake rate (kg/h)
= (maximum intake associated with quality for a
,kg/ h)*(forage availability factor)
(74)
where
Maximum intake associated with quality for a given
forage (kg/ h)
= (% of available dry matter that is new/ 100)
where
FAF = forage availability factor that serves as the fractional multiplier of the potential dry matter intake rate as in~uenced
by dry matter availability less than D(0 S FAF 1.0).
D = available dry matter mass level below which the dry matter intake rate will begin to become limiting, kg/ha
A = available dry matter mass for grazing, kg/ha
If A 2 D, then FAF will equal 1.0. FAF will equal 0.0 when A = 0.
owever, A is defined as the available dry matter on the field. Some
minimum dry matter level is present even after the animals have
removed as much dry matter as is physically possible.
urposes of example, suppose the proportions of new, old,
material are 40%,
and 2576, respectively.Similarly,
assume that the associated independent intake rates for ne
ead materials are 4.0, 2.5, and 1.0 kg/h, respectively, a
forage availability factor(FAF) limits intake to 50% of potential. Using
the above equations,
Loewer
+ (25/ lOO)*l.O
+ (35/100)*2.5kg/h
kg/ hp0.5
= 1.36 kg/h
3.
orage ~ o d eof
l Selective Grazing
FIGURE
Loewer
Partial
NO.
3
la
2
3
4
5
-2
(1) l b
2a
6
(2)
(3)
lb
2b
3a
-+
(4)
(5)
lb +
2b +
3b3b (6)
4a
(7)
--
lb
2b
+
+
lb
2b
+
3b
4b "* (8) 4b
5a
(9) 5b
6a
+
+
+
(10)
(11)
"Grazing priority designates the steps of selective grazing in two parts ("a" and "b'
except for partial field 1). The first number in each cell indicates the partial field to
be grazed in terms of maximizing digestible dry matter intake. For each partial field,
"a" and "W' refer to the portion of dry matter above and below, respectively the dry
matter level that begins to limit intake. The numbers in parentheses indicate the order
of grazing. The "+" connects those partial fields in the same column that are being
grazed simultaneously.
availability begins to become limiting. Likewise, assume that the animal grazes to point 5". Given these assumptions, GRAZE assumes
that step 5" did, in fact, reach the intake-limiting point but on partial
field 3a rather than 3b. The areas of partial fields 3a and 3b are adjusted with regard to relative area so as to appear as shown in Fig.
62. At midnight, partial field 3 (3a and 3b) becomes 3 and 4 with
separate calls to the crop growth model.
An input to GRAZE is the number of allowable partial fields
within the major field that are available for grazing, This number is
3
IGURE 62
Loewer
divided into the grazing area per animal to obtain a minimum partial
field size. If a new projected partial field is not sufficiently large,the
last partial field to be grazed retains its identity.
In summav, the selective grazing logic in GRAZE serves the following functions:
It calls the plant submodel at midnight of each simulated day
foreach partial field and updates plant growth and composition,
It computes the digestible dry matter intake rate (kg/h) associated with the grazing animal for each partial field.
It determines the order in which each partial field is subject to
grazing based on its forage quality and availability.
It callsthe animal submodel as required over the day andmoves
the animal systematically fromthe highest ranked partial field
to the lowest ranked as necessary so as to maximize the animal's digestible dry matter intake rate.
It adjusts the size and number of partial fields based on the dry
matter removed by the animal during the day.
It controls the reading and writing of input and output information, respectively and the scheduling of management practices such as fertilizer applications and moving of animals to
other fields.
GRAZE is a Fortran program that does not use any simulation language. It contains approximately 17,000 lines of computer code and
can be executed on a microcomputer (minimum configuration requirements are 386 machine, math coprocessor, NIB of RAM, and
MB of disk storage). The DOS release of GRAZE is Version 2.3,which
is described fully by a user guide (Parsch and Loewer, 1995) and a
case study publication (Loewer and Parsch, 1995). GRAZE for Windows, Version3.02, was first released to the public in June1996. Major
enhancements of the Windows version allow GRAZE to simulate
competition among plant species (Loewer et al., 1989), gives greater
flexibility in examining fertilizer applications, provide direct access
to the definition and default values of input parameters, and allow
for direct creation of output graphs that can be used to compare dif-
11
ferent scenarios.GRAZE for Windows is the version that will be supported in the future.
Input to GRAZE is accomplished through the use of two files,
one related to weather and the other containing the scenario that
defines an individual system to be evaluated. GRAZE may generate
several output files as specified by the model user. Maximum frequency of the output is l day. Each output file (except for the one
tracing file) can be imported into a spreadsheet, where it can be
manipulated at the discretion of the model user, In this situation,
comparisons among systems, graphic displays, statistical analysis,
etc. are the responsibility of the user, and these kinds of operations
can usually be done within the spreadsheet. In GRAZE for Windows,
Version 3.02, graphical displays may be generated, printed, and copied within the program. Again, details concerning executionand utilization of GRAZE may be found in the GRAZE User
(Parsch,
1995).
~eef-~ora
Model
~Selective
e of
15
National Research Council. 1987. Beef Cattle. Predicting Feed ~ n t a ~of eFoodProcessing AnirnaZs. Washington, DC: National Academy Press, Chapter 6.
Nelsen, T. C., C. R. Long, and T. C. Cartwright. 1982. Post-inflectiongrowth
in straightbred and crossbred cattle. I. Heterosis for weight, height and
maturing rate. 11. Relationships among weight, height and pubertal characters. J. Anirn, Sei. 55(2):280-292, 293-304.
Oltjen, J. W., A. C. Bywater, and R. L. Baldwin. 1986a. Development of a
dynamic model ofbeef cattle growth and composition. J. Anirn, Sei. 62:
86-97.
Oltjen, J. W., A. C. Bywater, and R. L. Baldwin. 1986b.Evaluation of a model
of beef cattle growth. J. Anirn, Sei. 62:98-108.
Paine, M.D.1971. Mathematical modeling of energy metabolism in beef
animals. PhD. Dissertation. Stillwater, OK Oklahoma State University.
Parsch, L. D., and 0. Loewer (Eds.). 1995.GRAZE Beef-Forage Simulation
Model: Users Guide. Southern Cooperative Ser. Bull. 381A. Fayetteville,
A R University of Arkansas.
Parton, W. J., and J. A. Logan. 1981. A model for diurnal variation in soil
and air temperature. Agric. Meteurul. 23:205-216.
Penrod, E. B. 1960. Variation of Soil Temperature at Lexington, Kentucky.
Univ.KentuckyEng.Exper.Sta.Bull.57.Lexington,
Icy: University of
Kentucky.
Preston, R. L. 1966. Protein requirements for growing-finishing cattle and
lambs. 1. Nutr. 90:157.
Pullar, J. D., and A. J. F. Webster. 1977. The energy cost of fat and protein
deposition in the rat. Br. J. Nutr. 37355-363.
Salisbury, F. B., and C. Ross. 1969.P ~ a Physiology.
n~
Belmont, CA: Wadsworth.
Sanders, J. O., and T. C. Cartwright. 1979a. A general cattle production systems model. I. Structure of the model. Agric. Syst. 4:217.
Sanders, J. O., and C. Cartwright. 197933. A general cattle production systems model. 11. Procedures used for simulating animal performance.Agric.
Syst. 4:289.
Simlcins, K. L., Jr., J. W. Suttle, and B. R. Baumgardt. 1965a. Regulation of
food intake in ruminants. 3. Variation in blood and rumen metabolites in
relation to food intake. J. Dairy Sei. 48:1629-1634.
Simlcins, K. L., Jr., J. W. Suttle, and B. R. Baumgardt. 1965b. Regulation of
food intake in ruminants. 4.Effectof
acetate, propionate, butyrate and
glucose on voluntary food intake in dairy cattle. J. Dairy Sei. 48:1635-1642.
Smith, E. M., and 0.J. Loewer. 1981.
A nonspecific crop growth model. ASAE
Paper No. 81-4013. St. Joseph, MI: ASAE.
Smith, E. M., and 0. J. Loewer.1983. Mathematical logic to simulate the
growth of two perennial grasses. Trans. ASAE 26(3):878--883.
Smith, E. M.,T. H. Taylor, and L. Brown.1968. Interpretation of diurnal
variation in soil temperatures. Trans. ASAE i!i!(2):195-197,
Smith, E. M., L. S. Ewen, and 0. J. Loewer. 1981. Growth of fescue and red
clover as influenced by environment and interspecific competition. ASAE
Paper No. 81-4020. St. Joseph, MI: ASAE.
Loewe
el of elective ~ r a z i ~ ~
17
Simo~
. I
One approach to modeling farm systems is to make measurements of the variables at intervals through time. A mathematical
curve can then be fitted to these time series data, Provided the other
variables can be provided, such a statistical regression model can be
used to predict future trends of the system based on observa
past trends.
Unlike traditional regression modelsof farm system behavior,
dynamical systems are constructed from relationships based on component experiments. The relationships are described
mathematica~y and
may be nonlinear,and the model allows the prediction of the evolution
of the interactingsystem vari
S forward through time,similar to
computersimulationmodelinowever,dynamicalsystemsmodels
are amenable to a variety of useful mathematical analyses and o * mization techniquesso that the long-term stability
of the modeled a
cultural ecosystems may be completely understood and so that the
model may explicitly focus onoptimizing farm management.
For optimization and management applications, simpler models
are more effective. Not only
does additional complexity make paramete~zationdifficult and reduce the chance that simple management principles may be identified,but a complex model may also be
prone to instability (May, 1971) or chaotic behavior (Sol6 and Bascompte, 1995). Nevertheless, it is perhaps unlikely that chaos will be
odwar
22
animal, the pasture, and the grazing interface (Fig. 2). Indeed, many
models have focused on the pasture component alone, assuming that
changes in grazing pressure per hectare are small.
Ecologists have been studying natural plant-animal ecosystem
dynamics for many years (e.g., Caughley and Lawton, 1981; Crawley,
1983; Stocker and Walters, 1984; May, 1971) where the animal populations may change in response to forage availability. However,in an
intensively managed farming application, animal numbers are tightly
controlled. Nevertheless, dynamical systems models are still useful.
Consider the situation of growing cattle grazing on a seasonally
growing pasture. The objective of beef finishing operations, for example, is to maximize the animals liveweight gain over a limited
growing season and thus their market value. Optimal management
includes choosing the right stocking rate so that animals are well fed
while pasture supply is maintained throughout the grazing period.
A valuable skill in dynamical modeling is the ability to rapidly
construct a prototype model of the system that contains all the basic
structures even if some of the finer details have not been firmly established, The process of analyzing a prototype model sheds significant light on what kind of model will ultimately be needed to provide
answers to the management issues and what biological knowledge
will be essential for the realism of those answers.
The situation of growing cattle requires a model of pasture availability and animal weight gain that includes the effects of changing
stocking rate. We will work through the construction of a prototype
model for this system.
90
4-
80
23
Growth
Rate \
70
60
50
40
30
20
10
0
0
1000
2000
3000
Yield, Y (kg ha")
4000
5000
FIGURE3 A sketch of the logistic growth curve for spring and winter pastures.
-=
dt
a(t)Y - bYZ
where Y (kg/ha) is the pasture biomass at any point in time. The first
term on the right represents the net balance between the rate of new
growth and the rate of senescence and depends linearly on the pasture mass according to the seasonal parameter a(t). The second term
represents a damping of growth due to self-shading within the sward
canopy, Thisterm does not vary seasonally but is nonlinear. The first
and second terms taken together comprise a logistic growth curve
with ceiling yield
war
Therefore,
a(t) = 4
gmax
-,
gmax
b=4yy-
ymax
1 max
rY
c+Y
1000
2000
3000 5000
4000
6000
dt
Pasture
accumulation
= [a(t)Y- b y z ] -
n(t)
Pasture
growth
Stocking
rate
rY
~
c+Y
Intake
26
war
2.5
-1.5
IGURE 5 The dependence of animal liveweight gain on feed intake. The
solid curve is the usual empirical relationship. The dotted line represents a
linear approximation.
the standard liveweight gain response curve (Fig. 5) (e.g., Sheath and
Rattray, 1985). This may be expressed mathematically as
Feed
intake
Body
growth
Body
maintenance
dt
- -L"-. rY v ci-Y
Liveweight
Body
gain
Feed
intake
dL
-
-L
V
(8)
maintenpce
ynamic
~ o d e l sand ~ r a z i n ~
27
U,&/ha
I, kgDM / (animal * day)
n, animal/ ha
L, kg/ animal
U,, = 5000 kg/ha
gmaX
=
kg/(ha day)
a = 0.06 kg/ (kg - day)
b = 1.2
ha/(kg.day)
Y = 0.035 kg/(kg. day)
c = 1100 kg/ha
V = 4 kg/kg
m = 0.02 kg/(kg .day)
Description
Pasture yield
Daily dry matter intake per animal
Stocking density
Animal liveweight
Maximum yield
Maximum pasture growth rate
Seasonal relative pasture growth rate
Damping of pasture growth
Potential intake per kilogram liveweight
Yield at which intake is half-maxi~um
Feed conversion efficiency
Feed maintenance requirement
.
Even very simple models can be used to give useful grazing management information. There have been a number of grazing models
war
IGURE
500
1500
2000
P = kY(t) = kY (0)e-knt
(10)
43
o~war~
t=
-
h-"---
"
kn
kY(0)
1
4
= 6 days
0.005
30 In 0.005 2000
"
This is a much more satisfactorymethod than calculations based simply upon "herbage allowance," as it takes into account the change in
intake rate as pasture mass declines. The simple model of McGall et
al. (1986) provides simple management prescriptions appropriate to
its intent.
where Ymaxis the ceiling yield, as before. This function is the same
"logistic growth equation" that we met in Fig. 3 and is based on field
data of ryegrass pasture growth gathered by Brougham (1956).
If a paddock of a rotational grazing system has a postgrazing
mass of Y (0) kg/ ha when the animals are removed at time t = 0, then
solving Eq. (12) (using separation of variables, see Appendix) gives
the accumulation of regrowth at some later time t as
Y(t) =
YmaXY(0)
Y (0)
+ [YmaX- Y
This is the typical sigmoid pasture growth curve (Fig. 8). Assuming
negligible pasture growth during the grazing phase for the paddock,
Morley wished to calculate the time at which the average rate of
pasture regrowth was maximized after the animals were removed
from the paddock (Fig. 8). That is,
~ a t h e m a t i c to
a ~maximize
~
somequantity we must set the derivative equal to zero, as follows:
31
6ooo
,
0
20
40
t*
60
Time (days)
FIGURE8 Tpical S-shaped growth curve for a pastureregrowing after grazing in a rotational grazing system. The dotted line is the steepest line that
touches the growth curve. The point where it touches is the point of maximum average regrowth rate. [Redrawn from Parsons and Penning
d Y(t) - Y(0)
dt
t
-(
>
YmaxY(0)
Y (0) + [U,, - Y (o)]e-"
With a bit of work we can deduce from this that the average pasture
growth rate is maximized if the animals are introduced into the paddock at time t satisfying the relation
This is not soluble in an explicit form but can easily be solved numerically Taking an initial pasture (postgrazing mass) of Y(0) = 600
kg/ ha and the parameters in Table 1, the optimal resting time is t" =
48 days.
This was an elegant way of using a simple logistic model of
pasture growth to make recommendations about the length of resting
in a rotational grazing system, Once again, a simple model was appropriate to answering the management problem.
The work of ecologist Noy-Meir (1975,1976,1978a,197813) in modeling grazing systems is well known. His seminal paper in 1975 analyzed grazing models of the form
dY
- = g(Y) - nI(Y)
where g ( Y ) is the net instantaneous rate of growth of new herbage
(rrtinus losses due to senescence) depending on the ~hotosynthe~c
biomass Y, and I(Y) is the rate of intake of a single animal when
offered herbage Y. The stocking density is n.
The relative merits of a number of explicit functional forms for
g (U)and I(Y) are discussed in Noy-Meir (1978b), but the 1975 paper
carries out a simple graphical stability analysis (see Fig. 9) for a generic model [Eq.
Noy-Meir shows that for someparticular forms
of the growth and intake functions, there might be two stable steadystate values of pasture mass Y for a given stocking density-one at
high pasture mass and one at low pasture mass.
A steady state is a point (in this case a value of pasture yield
Y) at which all of the system fluxes are in balance, hence the system
0
0
2000
4000
5000
or
6.3 c n c 10.6 animals/ha
(23)
35
F
2000
)r
=W
1000
3j
0
l
+
15
war
-=.aydt
where ni is the stocking density in paddock i, that is, the total number
of animals divided by the area
the paddock being grazed. This is
a reasonable model over a short time period at low to ~ e d i biou ~
mass. If the animals sequentially graze two paddocks areas h, and
h2 and initial pasture masses Y,(O) and Y2(0),respectively, then the
total intake of the animals over T days and the total pasture remaining in the two
paddocks at time T were calculated to be, respectively,
and
U,(")
+ U,(?') = hlY~(0)e(a-r*lk)~l
e
+
1)
(26)
h2y2(~)eut1e(~-~~2k)(T-ll)
37
applied. The intake over ?' days, C(?'), was maximized with respect
to tl when
Description
Stocking density in paddocks 1 and 2
Relative pasture growth rate
Areas of paddocks 1 and 2
Initial pasture mass in paddocks 1 and 2
Relative intake rate per animal
Total time in the two fields
Optimal swapover time
Optimal 10-day intake per animal
Woodvvar
dY
- =
dt
aY
Growth
- n(t)kY - aY(t
Grazing
I)exp
Senescence
Pasture
(kg h a )
(L'=O)
that we wish to feed the animals more than their maintenance requirement, we notice from the phase plane that this is not possible
to maintain in the long term. Initially, when animals are small, it is
possible to sustain above-maintenance feeding from pasture. However, as animals increase in weight, their intake rate begins to outstrip
pasture production, until at some point the pasture barely provides
maintenance for them. In order to maintain animal growth, the stocking density must then be reduced; this has the effect of stretching the
parabolic isocline in Fig. 12 out to the right so that the current state
(L, Y) is in the region of the graph where pasture mass and animal
liveweight are both increasing.
Since the maintenance isocline is a horizontal line, a second prediction is that animals of all sizes require the same pasture mass for
maintenance. Naturally this depends on the exact functions we have
chosen, in particular the linear liveweight model [Eq. (")l. Nevertheless, it is probably not far from the truth. Although larger animals
eat more feed,their requirements are also greater, so the pasture mass
on which they achieve maintenance may not be much different from
that providing maintenance to a smaller animal.
odwar
dY
- = aY - bY2 - nL dt
c+Y
dL - 1
"--.L""
dt z1
rY
c+Y
mL
z1
(33)
The isoclines are curves that show where the trajectories (curved arrows in Fig. 12)turn around, that is, where they are either vertical or
horizontal. This occurs when either Y or L is not changing, that is,
when dY/dt = 0 or dL/dt = 0.
First we consider all the points where d L /dt = 0, that is, the L'
= 0 isoclines. From Eq. (33), we see that d L / d t = 0 when
Y m=
mc
r-m
(35)
(Um is the pasture yield level at which the animals' intake is equal to
their maintenance requirement.) The value of Y mcalcdated with the
parameters in Table 1 would be, for example,
1.2
(37)
b y 2 - nL
rY
c+Y
-c
(39)
nr
r-m
r-m
L, = -(0.06 - 1.2
0.035n
3109
1467)(1100+ 1467) = n
(41)
So, for example,if n = 5 cows/ ha, the liveweight L, that wouldmaintain the pasture at Y = Ymwould be 3109/5 = 621 kg. This high value
reflects the high value of the parameter a for spring growth and in-
dicates that at this time of year even very large animals can gain
weight at a high stocking rate.
.
The directions of the small arrows (the "flow") on each isocline are
found by looking at the sign (+ / -) of the other equation. Equation
(34) shows us that d L / d t > 0 when L > 0 and Y > U,. Everywhere
else, d L / d t 5 0. This means that the direction of the trajectories (or
the "direction of flow") is to the right in this region and to the left
everywhere else. Similarly, from Eq. (38) we see that dY/ d t 0 when
> 0 and L is to the left of the parabolic isocline. To the right of the
parabola, d Y / d t c 0. We can now draw in the direction of flow across
or along the isoclines.
This information is summarized in Fig. 13. In fact, the directions
of many of the arrows can be deduced from common sense. For example, it is obvious that when liveweight is zero, pasture mass will
increase toward maximum yield and that when there is no pasture,
animal liveweight will decline to zero.
It is a fundamental property of dynamical systems that the trajectories of the system (the curves that join up the arrows) can never
cross except at a steady state. This is a very important property to
in mind when sketching trajectories on phase planes. Along
nc
- (a -
Lr
bYm)(c+ Ym)
If a steady state has at least one trajectory leading away from it, it is
unstable. So although system states near an unstable steady state
may initially move toward that equilibrium, eventually they will
move away. From Fig. 12, we can see that the two steady states on
the Y axis have at least one trajectory leading away from them; both
are unstable.
The stability of the remaining steady state at (Ls, Y,) is more
difficult to determine. From the sketch, it seems that the trajectories
spiral clockwise around the steady state, perhaps spiralling inward.
46
odward
Dynamic Systems
and Models
Grazing
the Y state equation [Eq. (32)] as fy and the right-hand side of the L
equation [Eq. (33)] as fL, then the eigenvalues e,,e, are the solutions
of the matrix determinant equation
det
22
51
bmc(r
+ m)
r(r
m)
(discriminant)
(44)
where
m2)a2+ [2mvbc(m2- 1)
mvb2C2(r + m)2 + 4bmcr(r - m),
Discriminant = mv(?
- 4r(r - m)]a
(45)
Dynamical systems theory states that for the steady state to be stable,
the real parts of both eigenvalues must be negative. In addition, when
the discriminant of Eq. (44) is negative, both of the eigenvalues are
complex and their real parts are equal. Then the steady state is the
center of a spiral; this occurs when a is in the range (Bl, B,), where
Bl, B2
bc(r
m)
Y - mmv
2r(r - m)
2r
it - [(Y- m)
mv
+ mvbc]12
a c 0.093
(47)
war
500
1500 2000
Time (days)
2500
500
200
300
400
500
IGURE 315 (a) Out-of-phase oscillationsin yield (solid curve) and liveweight
(dotted curve) when the steady state is a spiral point (in this case an unstable
spiral point). (b) The same oscillations plotted on the phase plane.
damped, this indicates a stable spiral; if they are increasing in magnitude, an unstable spiral.
When a is in (B3, B2), the real part of the eigenvalues is given by
the first two terms only in Eq. (44) because the value of the square
root is imaginary. The real part is zero when
bmc(r + m
(Y(Y - m ) ) =
That is,
bc(r + m)
r-m
This value of a (0.048 for the parameters in Table 1)is the watershed
between a stable spiral point and an unstable spiral point. The real
parts of the eigenvalues are plotted in Fig. 16, and we see that the
steady state is unstable when a 0.048. In this case the trajectories
spiral outward toward a limit cycle (Fig. 14). At lower values of
they spiral inward (as shown in Fig.. 12).
When the eigenvalues are outside the range specified in Eq.
the eigenvalues are real, the spiralling behavior disappears, and the
steady state is a node. The critical value for stability of this node
occurs when
bcm
r-m
a=-"--
at which point
= 0
and
8.
6 Theeigenvalues e,,
of the yield-liveweightsystemat
the
steady-state point (Ls, U,) plotted against the pasture growth rate parameter
Dotted lines indicate changes in stability - Re(el), --- Re(e2).
odward
450
\
FIGURE17 Phase plane diagram for the yield-liveweight model in the case
when the steady state at (Ls, Y
), is a stable mode.
Agricultural systems are biological systems under intensive management. The farmer exerts a significant degree of control over the system. In a grazing system, the farmer is able to change the stocking
and
~ yodels
~ a Systems
m i ~
Grazing
51
rate, apply fertilizer or water, mow the pasture, sow seed, feed out
supplements, and so on. All of these are managed controls.
Many agricultural problems involve identifyrng better control
strategies. How should a farmer change the stocking density or grazing management to maximize animal growth? How much fertilizer
should be applied, and when, to maximize pasture response? How
much excess pasture should be conserved, and when should it be fed
out? Optimal control theory allows us to take a dynamic systems
model of a system, define some quantity that we wish to maximize,
and then calculate the optimal management strategy to maximize this
quantity. In the language of control, the quantity to be maximized is
called the objective function, the management variable is called the
control variable, and the system dynamics equations are called the
state equations.
Control theory is well established in physics and economics and
is equally useful for biologicalapplications. An excellentintroductory
text to the application of optimal control methods in bioeconomic
manage~entis ~ a t ~ e ~ a t i c a l ~ i ~ e c by
o n ~Colin
~ i c sClark (Clark,
1990), which gives simple methods and examples for calculating the
optimal harvesting strategies for fisheries, which are biological systems not unlike agricultural systems. Here,we work through a simple
problem in grazing management as an example of how the methodology works.
We take the example of calculating the optimal stocking rate
(adjusted daily) to maximize production per hectare (measured as
animal intake) over a year.
~orm~la~io~
(52)
The method of optimal control is as follows. We form a special function H called the Hamiltonian of the problem. This is done by appending the state equation to the objective function using a Lagrange
m~ltiplier,
is called the adjoint variable.
= n(t) Icy
~
c+Y
3.
")
KY
(53)
The optimal stocking rate n ( t )is found when either the partial derivative of H with respect to n(t)is zero or n(t)itself is zero. Inaddition,
the partial derivative of H with respect to Y ( t ) ,the state variable, must
also be zero. These partial derivatives are
KY
an c +c Y+ Y
dH
dH
-= n(t)Kc +
dY
(c +
(54)
a ( t ) -2bY - n(t)---"-
(c +
Kc
=o
(55)
= l. ~ u b s t i t u t i n ~= 1
and
odels
razing
53
dY0,t
- a(t)Yo,t - by:,, - n(t)dt
c + yo,,
(57)
We rearrange this to find the stocking rate, n(t), that holds pasture
mass at the optimal level for maximum production:
n(t) =
a ( q 2 / 4 b - a(t)l2b
[h(f)/2b]l[c + a(~)/2bI
n(t) > 0
Notice that the optimal stocking rate depends on a(t), which is the
rate of change of pasture growth rate with time. This is because we
need to anticipate changes in pasture growth in order to adjust stocking rate in advance to allow the pasture to remain at the optimal
level.
Expressing a ( t ) as a sine function facilitates a numerical example. Figure 18 shows a typical course of Yo,, and the optimal n ( t )
calculated from Eq. (58)
It is possible that Eq. (58) might suggest that the stocking rate
n(t) should be negative. This occurs when pasture growth rate is increasing rapidly. In this case we take n(t) = 0, so that Eq. (58) is not
needed and we do not need to find the optimal stocking rate-it is
zero. It is then more difficult, however, to calculate the optimal pasture level. The model recommends removing all of the animals to
allow pasture mass to grow to .the optimal level, which itself is increasing.The animals are reintroduced when the pasture level
achieves the optimal level recommended by Eq. (56). So in practice
we cannot achieve the optimal pasture mass at all points in time,
because we cannot have n(t) 0. We are forced to accept a suboptimal
pasture level during those periods when growth rate is increasing
rapidly.
4000
3500
40
35
.(ET 3000
30
2500
2000
1500
15
10 tj
500
0
5
autumn
winter
spring
summer
FIGURE18 Optimal pasture mass for maximum growth, and optimal stocking density to hold pasture as near to this level as possible. In spring, the
actual pasture may be suboptimal for a time.
The recommendation is that when pasture growth rates are increasing, for instance in the spring flush, one should withdraw animals for a short time so that the pasture can get to its maximum
growth rate level. At this point, animals are reintroduced. Farming
practice in early spring often follows this line-farmers hold back
from increasing grazing pressure in the early spring inorder to allow
the pasture mass to accumulate.
However, this recommendation alerts us to some deficiencies in
the model. First, there is no consideration of pasture quality. In the
late spring we need to graze hard in order to maintain high green/
dead ratios in the sward. Second, the model has no limitation to the
rate at which animals may be added to or removed from the system.
In a real farm this will be limited. Further constraints could be introduced to reflect the number of animals available to the farmer.
.
Up to this point we have considered pasture models that express
pasture using a single variable, pasture biomass
However, milk
production in New Zealand dairy systems is adversely affected by
summer-autumn buildup of dead leaf in the pasture sward. A bio-
Dynamic Systems
and Models
Grazing
55
56
war
OO~!O
90%
80%
70%
50%
40%
30%
20%
10%
0%
~ ~ aLeaf
r dArea (m2
located in .the lower sward strata at most times of the year, its effect
in intercepting light may be assumed to be negligible, and the rate
of new growth of green leaf may be directly related to green leaf mass
by a ~itscherlichcurve. A more sophisticated model might attempt
to include the effect of light interception by dead matter in the sward.
We approximate this relationship by a Michaelis-Menten curve
(for ease of analysis):
BmaxG
Rate of new growth = kgL>M/(ha. day)
q+G
(59)
where gmax
is the maximum rate of new growth, which may vary
seasonally with changes in available light,temperature, and moisture,
and q is the herbage mass at which 50% of incident light is captured
by the sward. The net rate of green leaf accumulation d G / d t in the
absence of grazing is then modeled as
dG
-
dt
accumulation
Net Growth
-" gmaxG
-
GG
9t-G
Senescence
Net change in D
Senescence
57
Decay
Equations (60) and (61) taken together form a two-dimensional dynamical system modeling the fluxes of green and dead leaf mass in
regrowing pasture in the absence of grazing animals.
Keep ~ e ~ r c h i n g
(62)
I=NS=
tp
1440 - ti
S
+ S ( t m + tr)
and the total number of bites per day [constrainedby Eq. (62)]would
be modified accordingly. Then we would have the daily intake of
green leaf mass, I@ as
59
and a similar expression for the daily intake of dead leaf mass, Id
where t, and td are the bracketed expressions from Eq. (64), that is,
the total processing time (mastication and rumination) per kilogram
of green or dead leaf eaten, respectively.
C.
Se~ectio~
of
The challenge is to predict bite size and composition from the sward
variables. Thisis complicated owing to one significant problem:herbage is not uniformly distributed in a pasture, and the mechanisms
animals use to select which patch they will graze are still poorly
understood. Animals use the patchy distribution of pasture mass and
composition to their advantage by selecting a diet of higher quality
than the average quality of the pasture and in doing so they create
further patchiness. It has been suggested that within the constraints
of their speed and size, animals forage to obtain an optimal diet from
their environment (Newman et al, 1994). It may be that this theory
of optimal foraging will provide the way forward. However, there
are still many obstacles, not the least of which is the lack of suitable
statistics for characterizing the spatial distribution of plant material
in a pasture. The question of how one predicts the size, rate, and
composition of an animals bites when it is selectively grazing in an
environment with a nonuniform spatial distribution of pasture mass
and components is still very much open.
Animals prefer green mass, and the more green mass present,
the more they will eat. However, increasing levelsof dead mass will
depress green intake, as more time is spent in discrimination. Similarly, the more dead leaf in the sward there is, the more will be eaten
(assuming a principally green, vegetative sward). But increasing levels of green mass result in less dead matter being eaten as animals
can more easily selectthe green leaf in the sward surface layers.
A prototypical approach might be to consider that the average
bite size of each component is proportional to the availability (herbage mass) of that component, i.e.,
(x,
where
and (xd are constants (ha/bite). Such h e a r relationships
may be derived from encounter and rejection arguments of selective
foraging as described in Parsons et al. (1994). If Y = G + D is the
total pasture yield and P, = D I Y is the proportion of dead matter in
war
the sward, then the total bite size S and proportion of dead mass in
the diet Pd are predicted to be
S = S,
+ S,
=I
a,G
+ adD = [a,(l
P,)
+ aJ?,]Y
(674
and
This simple model gives predictions of total bite size and composition
consistent with the experimental observations, of Laca et al. (1992)
that total bite size increases linearly with herbage mass (or height)
and of Rattray and Clark (1984) and Parsons and Penning (1993) that
animals select a diet of higher quality than the average quality of the
pasture (Fig. 22).
If for convenience we define t, = 1440 - ti as the total nonidling
time in Eq. (65), then substituting the component bite sizes from Eq.
(66) into Eq. (65) gives predictions for the daily intake of green mass
1, and dead mass 1, respectively
=L
t,
+ t,a,G + tdadD
and
These are simple formulae for the daily intake of green and dead leaf
mass by a selectively grazing animal and may be appended to the
green-dead dynamics equations [(60) and (61)] in order to analyze
the e q u i l i ~ ~ ugreenldead
m
ratio in a grazed sward.
.
~e now have a simple model of green and dead pasture fluxes in a
pasture grazed at a stocking density of n animals per hectare:
0.5
0.45
2000
50%
O0/0
50%
Predictions of (a) total bite size and (b) composition when the
bite size of each component is proportional to the availability of that
component.
gmax
=
Description
kg/(ha. day)
=
kg/ha
CT = 0.03 kg/(kg*day)
t, = 0.02 midbite
growth. The green/dead ratio at this equilibrium may well be influenced by the stocking density chosen, so a farmer may be able to
choose a stocking density that gives an improvement in pasture
quality.
The greenldead ratio at equilibrium can be studied by using
phase plane analysis (recall Section IV). However, the isoclines for
this system are complicated. There is one isocline along the straight
line G = 0, where there is no green leaf and the amount of dead leaf
decays away to zero. The other isoclines are hyperbolas. Nevertheless,
we can plot these curves numerically for the parameters given in
Table 3. Figure 23 shows the isoclines of the GD phase plane for
several stocking densities. The D = 0 isocline, although a hyperbola,
is approximately straight in the feasible portion of the phase plane
(where G
0 and D
0). It intersects the D axis at (0,0), where
there is a steady state. The slope of the D = 0 isocline is approximately G = ( S / C F ) D
and is largely unaffected by stocking density.
The hyperbolic G = 0 isocline is a curve whose position changes
markedly as stocking density increases. When n is small, the isocline
Dynamic Systems
and Models
Grazing
463
all
/
S
0
IGURE 24 Bifurcation diagram showing the steady states of the greendead pasture model for a range of stocking densities. The curve shows
the position of the steady states, which are labeled S (stable, heavy line)
or U (unstable, dotted line). The horizontal axis is also a branch of steady
states,
.
There is no oneuniversal methodology that is suited toall problems.
The art of applied modeling is the ability to synthesize the empirical
knowledge and insight of experimental scientists into an appropriate model so that the applied problem can be answered. ~ynamical
systems models may be oversimplified for the purposes of coordinating diverse research efforts studying the detailed components of
a complex system; however, when simple insights or robust management recommendations are required, simplicity is a virtue, and
dynamical systems models come into their own-they allow a thorough analysis of the systems stability and are ideal for the purpose
of calculating optimal management strategies.
The modeling of problems in grazing management is still very
much a new frontier for the application of dynamical systems, and
there is vast scope for further progress in this field.
history the agriculture industry faces as many problems and opportunities today as ever in the past, and mathematical modeling is
another tool in the scientists workshop for addressing these issues.
Formulating models that are both realistic enough to give meaningful answers to our questions and simple and robust enough togive
reliable answers requires partnership between agricultural scientists
and mathematicians. Likewise, interpretation of results is a biological and economic issue as much as it is a mathematical one.
The intention of this review has been to survey the variety of
existing work in dynamical systems modeling of agricultural systems and to suggest some directions for the next step forward. Two
of these have been the coupling of simple animal growth functions
with pasturemodels and the expansion of pasture models to address
pasture quality issues. An additional lack is a simple model
dict the development of animal body composition during gr
In addition to a range of grazing applications, I have
introduce thebasic techniques of dynamical systems analysis. I have
also tried to provide references to useful books that explain the
mathematical techniques. It is hoped that the models and analyses
discussed in this chapter have proved thought-provo~
the relative strengths of mathematical modeling comp
466
~oo~ward
dt
where a ( t ) is the time-varying potential growth rate. Note that it is
often not necessary to fix the form of the ~ n c ~ ao( t n
) in advance.
The method is to rearrange the equation so that all of the
appear on the left and all of the t's on theright. For this example, we
divide by f ( Y ) and multiply by d t as follows:
67
50
40
20
3000
4000
5000
-Ymax In 1 - )xa:
);:
-Yrnax In 1 - -
+c=
actors
dY
- = a(t)Y + g ( t )
dt
where a and g are known parameters (that may vary with time if
desired). For example, consider Bircham and Sheaths (1986) model
for disappearance of pasture under grazing:
dY
- = go - knY
dt
where go is a constant pasture growth rate. The method is first to
move the term involving Y to the left:
dY
- + knY = go
dt
(All)
IF
Note: The general formula for the integrating factor for Eq. (A$)
is
!dt
E
IF 3- knY
IF = go
IF
d
- (Y IF) = go IF
dt
d
- (U exp(knt)) = go exp(knt)
dt
(Al4a)
(Al4b)
(A14c)
Y exp(knt) + C =
go exp(kns)ds = 8 0 [exp(knt) - 1
1
kn
(A15)
(A16)
C is then eliminated from Eq. (A15), and both sides are divided by
the integrating factor to give
Y(t) = Y(O)exp(-knt)
+ Bo
[ l - exp(-knt)]
kn
dt
= a(t)Y
+ b(t)Y2
= a(t)
"
z2 d t
dZ
dt
_z1 I-
==3
-=
1
b(t) Z2
-a(t)Z - b(t)
470
odwar
Systems
ynamic
and
Models
Grazing
71
Penning, P.D., A. J. Parsons, R. J. Om, and T.T. Treacher. 1991. Intake and
behavior responses by sheep to changes in sward characteristics under
continuous stocking. Grass Forage Sci. 46:15-28.
Press, W. H., B. P. Flannery, S. A. Teukolsky, and W. T. Vetterling. 1989. Numericul Recipes in Pascal: The Art of Scientific Computing. New York Cambridge Univ. Press.
Rattray, I? V,, and D. A. Clark. 1984. Factors affecting the intake of pasture.
N . Z. Aguic. Sci. 18(3):141-146.
Rook, A. J., C. A. Wuckle, and P. D. Penning. 1994. Effects of sward height
and concentrate supplementation on the ingestive behavior of springcalving dairy cows grazing grass-clover swards. Appl. Atlnim. Bekuv. Sci.
101-3112.
Sheath,G. W., and I? V. Rattray.1985.Influence of pasture quantity and
quality on intake and production of sheep. Proceedings of the XV International Grassland Congress, Kyoto, pp. 1131 -1133.
Snaydon, R. W. 1981. The ecology of grazed pastures. In: Grazing Animals. F.
H. W. Morley (Ed.). New York Elsevier, pp. 13-31.
Sol& R, V., and J. Bascompte. 1995. Measuring chaos from spatial information. J Theor. Biol. 275139-147.
Spa~nger,D. E., and N. T. Hobbs. 1992. Mechanisms of foraging in mammalian herbivores: New models of functional responses. Am. Nat. 140(2):
325-348.
Stocker, M., and C. J. Walters. 1984.Dynamics of a vegetation-ungulate system and its optimal exploitation. Ecol. Model. 25:151--165.
Thornley J. H. M. 1990. A new formulation of the logistic growth equation
and its application to leaf area growth. Ann. Bot. 66:309-312.
Thornley, J. H. M,, A. J, Parsons, J. Newman, and P. D. Penning. 1994. A
cost-benefit model of grazing intake and diet selection in a two-species
temperate grassland sward. Funct. Ecol. 8:5-16.
Ungar, E. D., and I. Noy-Meir. 1988.Herbage intake in relation to availability
and swardstructure: Grazing processes and optimal foraging. I; Appl. Ecol,
25:1045-1062.
Woodward, S. J. R. 1993. A dynamical systems model for optimizing rotational grazing. 1Ph.D. Thesis. Massey University, Palmerston North, New
Zealand.
Woodward, S. J. R, (in press) Formulae for predicting animals daily intake
of pasture and grazing time from bite weight and composition. ~ ~ v e s t o c ~
Production Science, accepted.
Wo~dwar~
S., J. R., and G.C.Wake.1994.
A differential-delay model of
pasture accumulation and loss in controlled grazing systems. Math. Biosci,
121 :37-60.
Woodward, S. J. R., G. C. Wake, A. B. Pleasants, and D. G. McCall. 1993a.
A simple model for optimizing rotational grazing. Agric. Sysf. 41:123155.
Woodward, S. J. R., G. C. Wake, D. G. McCall, and A. B. Pleasants. 1993b.
Use of a simple model of continuous and rotational grazing to com-
73
pare herbage consumption. Proceedings of the XVII International Grassland Congress, 8-21Feb.1993,
Palmerston North, New Zealand, pp.
788-789.
Woodward, S. J. R., G. C. Wake, and D. G. McCall. 1995. Optimal grazing of
a multi-paddock system using a discrete time model. Agric. Syst. 48:119139.
Yates, J. 1982. A technique for estimating rate of disappearance of dead
herbage in pasture. Grass Forage Sci. 37:249-252.
Zill, D. G. 1989. A First Course i n ~ i ~ e r e n t i a Z ~ ~ u a twith
i o n sAppZications. 4th
ed. Boston: PWS-Kent.
Systems-oriented researchapproaches have been given increased emphasis in publicly funded agricultural research. Theunderlying message seems to be that research should consider the detrimental implications (side effects and/or delayed effects) of research innovations
that were formerly often overlooked.
Systems-oriented research is a challenge for the applied livestock production scientist. The animal scientist was trained to do experimental research and has typically not studied systems theory and
model development. Although there is a history of modeling biological and physiological systems in animal science (Mertens,1977; Baldwin and Hanigan, 1990), the technique has had limited use at higher
levels of aggregation such as applied livestock production science,
The purpose of this chapter is to give an overview of modeling
as a systems research method and to describe how computer simulation models can be used in livestock production research and development.
75
Systems theory was introduced by the biologist Ludvig von Bertalanffy in the1940s (Bertalanffy, 1951).
He found that different systems
(plants, animals, societies,and ecosystems) could be described in general terms as general systems theories (~ertalanffy,1973).Being
holistic and general at the same time presents a paradox. As one of
the fathers of general systems theory (Boulding, 1956) said, All
that we can say about practically everything is almost nothing. The
relevance of systems theory is closely linked to the word model.
Systems theory may become clearer if we say as Le Moigne (1989)
that systems theory is the theory of modelling.
A system is a part of reality (physical or organizational) that is
separable from the environment. What is inside a conceptual border
is part of the system, and what is outside the border is the environment. Most systems experience exchanges (for example, energy or
information) with the environment; these exchanges are called inputs
and outputs. When we want to describe systems or simply to comm u ~ c a t our
e thoughts concerning systems, we create a hodel of the
system. A model is a simplification of the system and expresses the
system from a certain perspective and with a certain purpose. A more
detailed discussion of model definition in livestock production is
given by Spedding (1988).
An important property of a system is that it can be defined
within a hierarchy of systems (Rountree, 1977).A livestock farm as a
hierarchy of systems is described in Fig.
The farm may include
di~ferentflocks of animals (dairy cows, goats, swine, etc.)and different crops. Eachof these subsystems consists of its own subsystems.
The dairy herd, for example, consists of individual animals. In each
of the animals, organs can be identified, each composed
of tissue,
cells, and organelles. Each level in the hierarchy has its own characteristic rules of input / output transformations.
l
herd
n
The
dairy
farm
The
n- 1
The
n-2
An organ
11-3
A tissue
77
To understand a system it is often desirable to studythe behavior of its subsystems and the relationships among them, The number
of possible relationships or interactions between subsystems increases
exponentially as you move down in the hierarchy (Weinberg, 1975).
Although this means that a full description of a system by subsystems
necessitates reductions, a certain system cannot typically be reduced
to a simple function of subsystems. Some characteristics of a system
may disappear when it is divided into subsystems. These systems
characteristics were pointed out by the ph~osopherMario
described as emergent properties (Agassi and Cohen, 1982). An emergent property differs from an interaction between subsystems. An
emergent property cannot be linked to specific subsystems and their
relations, Theoretically a system with emergent properties cannot be
described by subsystems without reductions,
The word analyze originated from the Greek for to S
Traditional science is based on the analytical approach of ga
knowledge by splitting a problem into parts. A problem is reduced
to hypotheses suitable for testing in a designed experiment. The wor
systems is also of Greek origin and derives from the op
analyze, namely to aggregate.In systems-oriented rese
knowledge is obtained by
gation and not by splitting. Systemsoriented researchcan be c
ted in various ways but will in
era1 be based on the following steps:
/I
1. A problem is perceived.
2. The system of concern is identified conceptually.
3. A model of the system is developed.
It is necessary for the scientist to h o w how an animal is managed in order to understand the production of the animal. Since livestock management is conducted at herd level, the system of concern
is often the livestock herd. Herd management may be difficult to
separate from land management. If this is the case, the system of
concern might be the livestock farm. Research results may suggest a
certain general change in. management. This change should be evaluated at the level of the system of concern. Conceptual modeling of
a livestock farm is discussed by Sorensen and Gistensen (1992). The
following description focuses on the livestock herd as the system of
concern.
conceptual model of the livestock herd needs to describe the
relationship between production transformation of input to output)
and management (control of the transformation process). relevant
relationship between production and m a n a g ~ ~ e for
n t a livestock
farm is found in a cybernetic system (Sorensen and Kristensen, 1992).
A cybernetic system consists of an effector (a controlled system) and
a sensor acting on the effector (a controlling system). model of a
dairy herd as a cybernetic system is shown in Fig. 2. The herd as a
system is defined by two subsystems: a production system and a
management system. The production system consists of all the animals in the herd. The production from the herd is its output, and the
input to the production system is characterized as consisting of controllable and uncontrollable factors. Uncontrollable factors
include the
climate, legislative regulations, and price changes. The management
ProductionSystem
79
LIVEST~CK
Simulation models can be classified by the following three dichotomies (France and Thornley 1984):
1. Static versus dynamic
eterministic versus stochastic
3. Empirical versus mechanistic
The difference between a static and a dynamic model is that a dynamic model includes time as a driving variable. In such cases the
state of a system at time t + 1 is a function of the state at time t. A
dynamic model is often represented by a loop. The feedback of information in Fig. 2 illustrates a dynamic loop. If the measurement
refers to time the adjustment refers to time t + 1. The fact that a
feedback loop describes a time change is often overlooked. A livestock herd model as illustrated here mimics a self-recruitinggroup of
female animals over time and will therefore by dynamic as indicated
in Fig. 2.
The difference between a deterministic model and a stochastic
model is that input variables in a stochastic model are not described
solely by their mean values but also by their probability distributions.
The probability distributions will typically affect the mathematical
transformation of input to output made in the model, and a deterministic model ignoring these distributions would not produce mathematically correct results. In a stochastic model a change in the variance of one or more input variables can change the mean of one or
more important output variables, whereas the result in a deterministic model remains the same.
A deterministic livestock herd model uses classes of animals as
the simulatio~unit to deal with discrete events such as conception,
sex of offspring, death, and culling. A class of animals contains, for
example, 10.7 animals at a certain time. A much-used deterministic
cattle herd simulation model is one formulated at Texas A&M University of Sanders and Cartwright (1979).
A certain type of stochastic models are called probabilistic or
arkov chain models. In a probabilistic livestock herd mo
classes of animals constitute a transition matrix expressing a probability distribution of animal movement from a specific class to all
other classes in the next step. Probabilistic models follow the so-called
Markovian property in which the outcomes of each step dependonly
on the last state of the model. Examples of probabilistic livestockherd
models are dairy and swine models described by Jalvingh etal.
(1992a, 1993).
Another type of stochastic herd models are Monte Carlo simulation models. A Monte Carlo simulation model typically uses the
animal as the simulation unit. Discrete events are controlled by pseudorandom number generators linked to suitable probability distributions. "'Whole animals" are maintained through a simulation by a
set of status variables referring to the individual animal. It is necessary to make repeated runs from a specific data input in order to
establish a mean herd value by a stochastic model. Examples of
Monte Carlo simulation models are the dairy herd model SIMHER~
(Slzrrensen et al., 1992) and the sow herd model Pig-Oracle (Marsh,
1986).
The distinction between empirical and mechanistic models relates to the conception of the world as a hierarchy of systems. In a
hierarchy of systems any system is composed of subsystems and at
the same time is part of a larger system as illustrated in Fig. 1. An
empirical model output relates to input within the same hierarchy
level. A mechanistic model describes a certain level indirectly from
the behavior of one or more sublevels. The advantage of a mechanistic model is that parameters estimated at one level canpredict results
at the next higher level. Thus,the model becomes a medium for communication between levels of a system. It is important, however, to
be aware that when a system is described by lower levels in the
hierarchy, the degree of reduction is higher. A mechanistic model implies, in principle, a higher degree of reduction than an empirical
model. Livestock herd simulation models are typically mechanisticin
the sense that they simulate the production of the herd indirectly by
simulation of subsystems (classes of animals or individual animals).
Ssrrensen
482
Livestock Applied
~roductionScience
83
Description
Individual characteristics
Identification
Inherent
production capacity Expected FCM/day 1-24 wpp in 3rd
lactation
State variables
Days
Age
Weeks postpartum
Lactation stage
Lactation number
Cycling or not and weeks to estrus
Estrus status
Pregnant or not and weeks to
Pregnancy status
calving
Decision made or not
Decision for culling
Maximum yield (FCM/day) in the
Milk production potential
present week
Yield (FCM/day) in thepresent
Milk production
week
Kilograms
Liveweight
(4)
(1st lactation)
(older cows)
(5)
(6)
g,,,
= potential daily gain, kg/day; ypot= potential milk yield, kg/day; A = mature
weight; 7' = weeks postpartum; W = actual liveweight; y = actual milk yield, kg/day;
g = actual growth rate, kg/day; c, = net energy requirement per kg milk; c, = net
energy requirement per kg gain; El, = available net energy for lactation and growth,
SFU/day; K = bulk capacity, units/day; E , = available net enerw, SFU/ day; Ei, = net
energy intake, SFU/day; b, c, n, ko, k, are constants.
Slclrrensen
lactation
1st
All cows
0.0049
1.3110
0.0995
Older cows
0.9356
0.0611
1.22 x lo+
2.56
0.90
10.06
1.04
0.92
0.40
2.65
5.00
8
k
-8
h
OONCK)
"v!c?T?
OFIOO
Reproduction and replacement in a livestock herd have many delayed and side effects. These complex problems can be assessed by
livestock herd model simulations evaluating the technical and economic impact of different reproduction and culling strategies (e.g.,
Marsh et al., 1987; Dijkhuizen and Stelwagen, 1988; Jalvingh et al.,
1992b). Clausen et al. (1995)
evaluated nine culling strategies with the
combination of different heat detection rates, initiation of insemination after calving, and heifer purchase policies using SIMHERD. The
results show that the best culling strategy varies between herds depending on the total management strategy of the farm.
Grassland management is a highly complex interaction between animals and pasture growth. It is therefore a classical objective for computer simulation development and application (Spreen and Laughlin,
1986; Maxwell,1990).Grass growth and quality follow a seasonal
pattern. It is therefore often decided to use a certain calving pattern
in dairy herds to synchronize feed requirement and grass growth.
Srarensen et al. (1992) showed that the effect of seasonal calving and
grass utilization at herd level very much depend on the production
constraints of the herd. If the production is constrained by herd size,
the positive effect of spring calving is very little compared to a constraint on milk production per year. The effect of different stocking
rate and supplemental feeding levels on pasture was found to be
highly dependent on the herd management system (Kristensen et al.,
1997).
Health issues at herd level are very complex because the cause-andeffect relationships between production and health are very complex
(Srarensen and Enevoldsen, 1992). The effectof a certain management
change will therefore be almostunique for a certain herd. It therefore
seems more relevant to develop simulation models that allow the
advisor to describe the unique herd and evaluate certain herd-specific
strategies than to generate general rules of thumb. SIMHERD and a
set of other PC tools were therefore developed into a program package available for herd health management (Enevoldsen et al., 1995).
Health management is further complex due the contagious pattern of some diseases. Suchproblems are often dealt with by Markov
chain models (Carpenter, 1988).A model based on SIMHERD describing the contagious cattle disease BVD has been developed (Srarensen
et al., 1995), illustrating the relevance of combining interactions between herd management and contagious disease pattern.
ination of Case Studies and Simulat~on Ex~eriments
91
Marsh, W. E., A.A. Dijkhuizen, and R. S. Morris, 1987. An economic comparison for four culling decision rules for reproductive failure in United
States dairy herds using Dairy ORACLE. J. Dairy Sci. 70:1274-1280.
Marsh, W. E., D. T. Galligan, and W. Chalupa. 1988. Economics of recombinant bovine somatotropin use in individual dairy herds. J Dairy Sci.
2944-2958.
Maxwell, T. J. 1990. Plant-animal interactions in northern temperature sown
grasslands and seminatural vegetation. In: Systems 27zeory Ap~liedto Agriculture and the Food Chain. J. G. W. Jones and P. R. Street (Eds.). Elsevier
Applied Science: London, pp. 23-60.
Mertens, D. R. 1977. Principles of modeling and simulation in teaching and
research. 1. Dairy Sei. 60:1176-1186.
Ostergaard, S., J. T. Sorensen, V. F. ISristensen, and T. Kristensen. 1994. Modelling of the production of a dairy cow in a new energy system. Presentation and documentation of the PC model SIMCOW. Rep. 24. National
Institute of AnimalScience,
Denmark. (WithEnglish summary and
subtitles.)
Ostergaard, S., J. T. Sorensen, J. Hindhede, and A. R. Kristensen. 1996. Technical and economic effectsof feeding dairy herds TMR in one group compared to multiple groups under different herd and management characteristics estimated by stochastic simulation. Livest. Prod. Sci. 4523-33.
Rountree, H. 1977. Systems thinking-Some fundamental aspects. Agric,
Syst. 2~247-254.
Sanders, J. O., and T.C. Cartwright. 1979. A general cattle production systems model. I. Structure of the model. Agric. Syst. 4:217-227.
Sarensen, J. T. 1990. Validationof livestock herd simulation models: A review.
Livest. Prod. Sci. 26:79-90.
Sarensen, J. T., and C. Enevoldsen. 1992. Modelling the dynamics of the
health-production complex in livestock herds: A review. Prev. Vet. Med.23:
287-297.
Sorensen, J.
and C. Enevoldsen. 1994. SIMHERD:A dynamic, mechanistic
and stochastic simulation model for dairy herd health management decision support. Kenya Vet. 28:431-434.
Sorensen, J. T., and E. S. Kristensen. 1992. Systemic modelling: A research
methodology in livestock farming. In: Global A p ~ r ~ i s aofZ Livestoc~F ~ r ~ i n g
Systems and Study on Their Organizational Levels: Concepts, M e t ~ o ~ o and
Zo~
Results. A. Gibon and G. Matheron (Eds.). EUR14479. The Commission of
the European Com~unities,pp. 45-57.
Sarensen, T.,E. S. &istensen, and I, Thysen. 1992. A stochastic model
simulating the dairy herd on a PC. Agric. Syst. 39:177-200.
Sorensen, J. T.,C. Enevoldsen, and T. &istensen. 1993. Effects of different
dry period lengths on production and economy in the dairy herd estimated
by stochastic simulation. Livest. Prod. Sci. 33:77-90.
Sarensen, J. T., S. Ostergaard, V. F. Kristensen, and T. Kristensen. 1994. A net
energy model describing the effect of feed management on live weight
changes along the lactational cycle in dairy cattle. Paper presented at the
49
S~rensen
ent, and
H. Fawcett
.
Grazing systems are one of the main types of agroecological systems
for food production in the world. These systems comprise about half
of the worlds land area (Stuth and Stafford-Smith,1993), and increased interest exists in improving their management and ensuring
sustainability. Thereis concern to prevent degradation of the resource
base and its consequent environmental, social, and economic effects.
A common approach in studying grazing systems has been by
way of mathematical modeling (Stuth and Stafford-Smith, 1993). Unfortunately, many scientists have looked at specific and detailed phenomena within a part of the system without taking into account interactions and effects at the whole-systemlevel(Demment et al.,
1995). The plant/animal interface is one of many subsystems of a
grazing system. It is a key element but is subject to a number of
forward and backward interactions with the other subsystems. Consequently, it cannot be fully understood in isolation nor can it be
e r r ~ r oet al.
.
.
A wide range of approaches to modeling the forage resource are
found in the literature. These vary widely in degree of complexity
number of species represented, variable and parameter definitions,
and simulated output.
Some of the simplest representations are based on the functional
form of plant growth curves [see Thornley and Johnson (1990) for a
review].Forexample,Brougham(1956),Morley(1968),Noy-Meir
(1975, 1976, 1978), Christian et al. (1978), and Woodward et al. (1993,
1995) used logistic growth curves to represent pasture growth in their
models:
where m = the maximum relative growth rate, W = initial plant biomass, Wmax= the asymptote plant biomass, and t = time. This description assumes that growth is proportional to plant biomass, rate
of growth is proportional to amount of substrate, and substrate is
finite (Thornley and Johnson, 1990).
E~uation(1)is easy to parameterize when smooth experimental
ata of the dynamics of plant growth are available. The appropriateness of the parameters will depend on thequality of the experimental
data andwill reflect onlythe particular conditions in which the plants
were growing for that particular data set. Nevertheless, the shape of
the curve can be explained physiolo~cally.Exponential growth occurs due to increased irradiance captured by increases in leaf area
index (LAI) during early stages of development, while growth progressively decreases to a plateau as respiration losses due to senes-
1991;
errero et ai,
lant/Anima~~ n t e ~ in
a c~ ~r a z i Models
n~
It is well recognizedthat apart from herbage availability, sward structure plays an important role in determining the intake of grazing
ruminants (Stobbs, 1973; Freer, 1981;Hodgson, 1985; Demment et al.,
1995) and that r u ~ n a n t select
s
~referentia~y
for the leaf component
of the sward (Laredo and Minson, 1973; Hendricksen and Minson,
1980; Cowan etal.,1986; Penning et al.,1994).In rangelands (i.e.,
Coughenour, 1984; Hanson et al., 1994), species composition plays a
major role in determining the diet selected by the animal reag again
and Schwartz, 1995), and, depending on animal numbers and environmental conditions or disturbances, this selection modifies t
sequent species composition of the sward (Humphreys, 1991;
~ c h m o n d1994).
,
Therefore, a basic requirement of pasture models, apart from their ability to predict total biomass, is that they
should be capable of differenti~tin~
between plant parts and their
density across the sward's vertical strata and/or species composition
*
the case of multispecies systems. Table 1 shows some of the main
lfferences in the representation of sward morphology and composition between models. It canbe observed from the table that several
els do not di~erentiatebetween plant parts.
The most common fractionation of morpholo~calcomposition
is between leaf, stem, and dead material, and this i s usually linked
to age c~aracteri~tics
of the sward. This fractionation occurs partly
because the models rely on photos~thesisand require an estimate
Total
No. of
biomass
species"
Model
Noy-Meir (1975,
1976,
1978)
Johnson and Thornley (1983,
1985)
Herrero (1995)
Parsons
al. et
(1991)
C ~ a ~ l e s - E ~ w aal.
et
r d s (1987)
91) al.
et
onal.et
(1988,
1994)
rlstian
al. et
(1978)
Cou~he~
al.oet
u r (1984)
3"
Lauenroth et al. 31.
(1993)
Lopez-Tirado and Jones (1991a)
~ u e ral.
r e r oet
(1984)
Rodriguez
al. et
(1990)
Woodward
al.
et
(1993,
1995)
al.
et Doyle
(1989)
e l ial.
~ a net
(1992)
Plant
partsh
Phenology
or age of
Species
plant
parts
comp.
1
1
1
2
1
1
3+
1
1
1
1
1
2
1
3"
1
+ f,Ws
+ Wsh, + Wsh, + f,Ws
+ WL, + Wsh, + Wsh, + f,Ws
(4)
where WL, and Wshj are the structural weights of leaves and stems
of different ages, respectively; where j = 1 is new material and j = 4
is senescent tissues. In this three-horizon example, note that the first
horizon (top of the sward) contains only leaf, which is commonly
observed in many pasture species (Hodgson, 1985; Humphreys,
1991). Ws is the storage compartment, and f;: the fraction of Ws associated with the plant material in dmi.Bulk density can be estimated
as
dmi
bdi = hi
and
dm
bd = h
where bdi = bulk density in the ith horizon and bd= total sward bulk
density.
The flexibility of the modeling structure presented in this chapter permits representation of a large number of sward structures,
since the number of horizons and thetype and quantity of plant parts
comprising them can be changed without altering the functionality
and robustness of the original plant models. At the same time, the
structure has useful attributes in modeling intake and diet selection
by different methods. The approach has been specifically developed
for tropical and temperate monospecific swards, but it can be extended to grass-legume swards.
Separation between green and dead material in other models is
mainly for the purpose of determining the swards nutritive value.
or example, to be able to represent diet quality; some models (Christian et al., 1978; Guerrero et al., 1984; Smith et al. 1985) subdivide the
al Interface in Grazing
method used to represent intake assumes that estimates of pasintake can be derived from the energy requirements of the animal
and the energy content of the pasture consumed. This last parameter
has been usually derived from in vivo or in vitro digest~bi~ty
estimates. The energy value of the forage as well as the animal's requireents have been expressed
t commonly as
, metabo~zableenergy ( ), or net energy
., 1995).
Two approa~hesare commonly used. The first is to estimate infrom the "inverse" of the nutrient requirements, and the second
use regression equations, which are often included in requirements s stems. These methods of intake estimation have been widely
ck models (Sanders and Cartwright, 1
ro et al., 1984; Gartner a
ever, although it is acce
requirements represent one of the most important driving forces of
the "reverse" calculation of intake from requirements is applied, intake is usually overestimated because animals are
not physicallyable to eatsufficient quantities of forage.
More recently Forbes (1993) suggested that, for cows, this
digestibility threshold may be higher depending on the level
of production. In view of these problems, several models
have incorporated static physical fill limitation constraints
on intake (Forbes, 1977; Kahn and Spedding, 1984; Mertens,
1987; Finlayson et al., 1995).
2. Maintenance energy ~equirements scale with metabolic
weight (Brody 1945), but rumen sizescales with body
weight (Demment and van Soest, 1985; Illius and Gordon,
1991), thus partly explaining why digestibility which is a
crucial parameter in requirements systems, is not a good
predictor of intake for low quality forages (Laredo and
son, 1973; Poppi et al., 1981; Kbon and rskov, 1993). In the
trial of Laredo and Minson (1973), sheep consu
leaf than stem with both plant fractions havin
digestibility suggesting that other factors, such
ical structure of plant parts, which influence parti
down and passage rates (Poppi et al., 1981, Ken
~ u r p h y1988; McLeod et al., 1990) play a
in the control of feed intake and also that
of digestion that consider these factors ma
timates of potential intake.
2.
EmpiricalRelationsBetweenHerbageMassandIntake
errero et al.
LA1 = LM*SLA
RI = 1 - exp(-O.OOl*DM)
RI = [imax*(DM/(~M
+ X))] /imax
RI = l
- exp(-O.O02503"DM)
RI = 1 - e~p(-2.4*10-~*Z)M~)
RI = 1 - exp(-0.000008*DM2)
RI = ~ M / ( D M+ 250)
RI = 1 - exp(-0.000002*D~2)
RI = [imax*(l - exp(-DMH/
imaX)l.23)(1/1.23)]
/imax
RI = [imax"(LAI/K)Q/(l + (LAI/
K)Q)]/ imax, where RI = imax/ 2
for LA1 = K; K = 1 and Q = 3 for
sheep grazing ryegrass
RI = 2"FA/B - FA2/B2,where B =
750
RI = 1 - ((1- O.l)/(HI LOW)2)"(HI- SH)2,where HI =
20 and LOW = 5
RI = exp[-1.016" exp(-1.038*A)],
where A = (~M/imax)*(area/
animals); 8 = 1 - 1.42"
exp(-0.00198"DM)
I
~
~The R
relationship
~
between herbage availability and intake at grazing in different models,
ot
IGURE 4 Effect of changes in the proportion of leaf on the functional response between herbage availability and intake at grazing.(-) pL
= 0.3;
("")
pL = 0.5; ("-) pL = 0.7.
errero et al.
51
graze shorter swards more efficiently and therefore swards can reach
lower herbage masses before intake is reduced (Illius and Gordon,
1987). Although all models modify potential intake largely on the
basis of body weight or a function of it, few models (Zemmelink,
1980; McCall, 1984)modify explicitly the functional response according to body size. Johnson and Parsons (1985) claim a value of K = 1
for ewes (i.e., 80 kg body weight) grazing ryegrass while a value of
K = 2 provides a suitable relation for mature dairy cattle (i.e., 600 kg
body weight) (Parsons, personal communication). However, they do
not provide a specific relation of this parameter with body weight.
We scaled parameter K to body weight using an allometric relationship derived from Illius and Gordon (1987), who claim that differences in the ability of animals of different sizes to graze are caused
by differences in incisor breadth, hence mouth size. The relationship
between parameter K and body weight then becomes
K = 0.229 BWo36
(9)
~ l a n t / A n i ~ lanl t e ~ a ~
ine~ r a z i ~ g
1
0.8
0.6
0.4
0.2
0
0
bite size and act as compensatory mechanisms when bite size is too
small to obtain the desired intake level (Hodgson, 1981). Chambers
et al.(1981) and Newman et al. (1994a) suggested that biting rate
declines at high bite sizes because of an increase in the ratio of manipulative to biting jaw movements and therefore it is also partly
dependent on sward characteristics. This subjecthas been clearly depicted by Laca et al. (1994), who found that time per bite (TB) was
linearly associated with the total number of jaw movements per bite
(JM) (Fig. 6a):
TB = 0.43
+ 0.682 JM,
r2 = 0.96
The proportion of total manipulative jaw movements that performed manipulation and mastication (MJM) increased asymptotically with bite size (Fig. 6b) according to the relation
MJM =
1.028 IB
0.234
0.246
+ IB
'
r2 = 0.69
(a) Relationship between time per bite and total number of jaw
movements per bite. (b) Relationship between bite size (IB) and the proportion of manipulative jaw movements relative to total jaw movements. [From
Laca et al.
12-13 h/day (Fig. 7) (Arnold, 1981). Thesevalues are similar for cattle and sheep (Hodgson, 1982,1985;Forbes,1988; Demment et al.,
1995).
ite size is positive related to herbage mass o
lack and Kenney 1984; odgson, 1985; Forbes, 1988
1991). Burlison et al. (1991),working in swards ranging from 5 to 55
cm in height, explained 78% of the variation in IB of sheep with the
relation
(13)
where H = sward height.
The slope of this relationship was similar to those reported by
odgson (1981) and Forbes (1982) when expressed on the basis of
bite size per kilogram body weight. Burlison et al.(1991) also argue
that due to the bias caused by changes in bulk density across graze
horizons, the responses often found were asymptotic, thus confirming
the results of other authors (Penning, 1986; Ungar and Noy-Meir,
aker et al., 1992).
better understanding of how changes in sward characteristics
affect bite size canbe achieved by describing this variable at a lower,
more detailed, level of aggregation. Burlison et al.(1991) describe the
components of bite size in Fig. 8.
ite depth is generally proportional to sward height (M
al.,1982;Wade et al.,1989;Laca et al.,1992; Ungar et al., 1992;
emment et al., 1995),but it may decreasedepending onthe relative
height of stem material in the grazed horizons (Barthram and Grant,
1984; Forbes, 1988; Floreset al., 1993). Burlison et al. (1991)
found that
the following relation for sheep explained 93% of the variation in bite
depth:
ite depth = -1.0(14)
+ 0.37H
Since sward height is a good predictor ofLA1 (Parsons et al.,
1994), which in turn reflects leaf mass [see Eq. (7)], these results are
in close agreement with the relation between leafiness and bite size
found
in
several pastures (St
Chacon and Stobbs,
1976;
hen, and Minson, 1980;
, 1986).
creases with swa
(Burlison
al.,
et1991;
Laca
wever, it also increases with decreasing bulk density
urlison, 1987), especially on higher swards (Laca et al., 1992),
ight be explained by the limitations posed by the shearing st
required to harvest a bite (Hodgson, 1985). Nevertheless,bite area is
less sensitiveto sward characteristics than bite depth (Hodgson, 1986)
51
errero et al.
DAIRY COWS
CAlTLE
la,
SHEEP
75
50
25
0
HOURS
515
and has been considered constant for a given body weight in some
models (Parsons et al., 1994).
Bite volume is also positively related to sward height (Ungar
and Noy-Meir,1988).Burlisonetal.(1991)
obtained the following
relationship for sheep, thus explaining 83% of the variation in bite
volume:
Bite volume = -32
+ 8.0H
(15)
why Dougherty et al. (1988) could not find i~erencesin grazing time
at various levels of supplementation. In temperate regions, the largest
proportion (~0-90%)of grazing time occursduring daylight (Penning
et al., 1991; Rook et al., 1994); however, in the tropics night grazing
quently observed due to high ambient temperatures during the
y (Humphreys, 1991). The largest proportion of rumination time
ccurs during daylight (Rook et al., 1994). Mastication and rumination increase with increased neutral detergent fiber
ntration in forages (Demment and Greenwood, 1988)
estibilit~(Arnold, 1981), in order to reduce particle
forage consumed for passage through the gastrointestinal tract. Fasttime and reduces min nation time (Greenwood
om par is on of predicted intakes from grazing behavior measurements with ~ m i t e dexperimental data has shown close a ~ e e m e n t
(Gordon, 1995). However, these models still require considerable effort to be widely validated; thus their range of application has been
limited mostly to research purposes. Nevertheless, they have provided a s i ~ i f i c a n t c o n t ~ ~to~ the
t i ounderstanding
n
of intake from
ds in the past two decades and have given valuable
design of appropriate management strategies in some
recowzed criticism of this approach is that large part of
grazing behavior is caused by the animals need for nutrient supply
(Ungar and Noy-Meir, 19 ). Most models do not integrate the two
rocesses, and those that
have not integrated mechanistic models
tion and metabolism at the same level of aggregatio
azing behavior. ~onsiderableresearch needs to be
e~reroet al.
51
.
From the viewpoint of whole grazing systems, it is now clear that the
plant/ animal interface is not completely represented if the consequences of grazing and other nutritional management practices (e.g.,
production) on the animal are not modeled. A series of papers and
books relate to the subject (e.g., Forbes and France, 1993; van Soest,
1994; Journet et al., 1995), but there appears to be no consensus on
the best approaches to modeling these processes. Nevertheless,those
discussed in the following subsections represent, broadly, the most
common approaches.
~~ant/Animal
Interface in
51
The energy requirements of ruminants have been estimated with reasonable accuracy, and differences between the systems used in different countries (i.e., Jarrige, 1989; NRC, 1989,1996; SCA, 1990; AFRC,
1993) seem to be small (McDonald et al., 1995). Traditional requirements systems were not designed to predict intake but to assess the
nutritio~al and
productive consequences of different feedstuffs for
the
animal once their intake was known. Therefore, a criticismthat often
arises is that the effective calculationof nutrient supply to the animal,
and hence the quality of the predictions of animal performance, are
largely dependent on the accuracy of the intake estimate used for the
calculations. Hence the importance of the representation of intake
prediction in grazing systems models.
Several models of grazing systems, whether designed for sheep
or beef or dairy cattle, rely on one form or another of an energy
requirements system to represent animal performance (Vera et al.,
1977; Christian et al., 1978; Sibbald et al., 1979; Konandreas and Anderson, 1982; White et al., 1983; Doyle et al., 1989; Walker et al., 1989;
Richardsonet al.,1991; Seman et al.,1991; Hanson et al.,1994;
Thornley et al., 1994; Freeret al., 1997). However, fromthe nutritional
management viewpoint, these systems per se present some inadequacies that need to be addressed by other mechanisms to improve
their flexibility.
1. These systems are static, and digestibility estimates are central to the calculation of energy in feedstuffs in the appropriate units
(e.g., DE, ME, NE). In requirements systems, these estimates are an
input and are fixed for a particular feedstuff. However, effective digestibility is a consequence of degradation and passage through the
gut, and therefore it dependent on plant and animal characteristics
(Demment and Greenwood, 1988; Illius and Gordon, 1991). Due to
the inherent selection by grazing animals on the basis of chemical
and physical characteristicsof different plants and/or plant parts (see
Section 111), it is necessary to model degradation and passage before
describing digestibility and consequent nutrient supply. This requires
dynamic models.
2. Even the most recent requirements systems do not take into
account explicit protein-energy interactions (Oldham, 1984; Preston
.
A wide range of dynamic models of digestion can be found in the
literature (e.g., Waldo et al., 1972; Mertensand Ely, 1979; Forbes, 1980;
lack et al., 1980; Bywater, 1984; Fisher et al., 1987; Hyer et al., 1991;
Illius and Gordon, 1991,1992; Sniffen et al., 1992; Fisher and Baumont, 1994). These types of models have been recently reviewed by
Illius and Allen (1994), and Baldwin (1995) reviewed research models
representing metabolism and the formation of end products of fermentation in ruminants" (e.g., Baldwinet al., 1970, 1977, 1987; Gill et
al., 1984; Murphy et al., 1986; Danfaer, 1990; Dijkstra et al., 1992, 1993;
Poppi et al., 1994). A range of approaches to model digestive processes can also be found in Forbes and France (1993).
The basic objectives of dynamic models of digestion are to predict potential intake, digestibility, and animal performance as a function of the nutritional quality of plants on offer and a range of animal
characteristics. There is evidence that such models are better at predicting nutrient supply and animal performance than requirements
systems (Fox et al., 1992, 1995; Ainslie
et al., 1993). However, there
are certain basic aspects that need to be considered that define the
accuracy and flexibility of the model. A discussion of these follows,
ons The basic fractionation of feedstuffs is represented in Fig. 9. The separation of dry matter into its
basic chemical entities is important because different feed fractions
of different forages have different degradation and passage rates (11lius and Gordon, 1991; Russell et al., 1992)and therefore have different digestibilities. Consequently they supply different amounts of nutrients to the animal (Murphy etal.,1982;Gilletal.,1990).These
fractionations are also important into predicting effects o
mentation on the rate of cell wall digestion (Argyle and
1988), modeling protein-energy interactions, and using recent standards of protein requirements (e.g., Fox et al., 1992; O'Connor et al.,
1993; AFRC, 1993). Nevertheless, other authors consider that the nutritional description of the potentially degradable fractions of feedstuffs requires yet further fractionations (Mertens and Ely 1979; Sniffen et al., 1992), although it is questionable that they provide better
predictions than simpler approaches (Illius and Allen, 1994). The fractionation presented here is robust, simple, and suitable for use in
whole grazing system models.
n e ~ ~The
c ~concentration and potential degracell wall of forages are among the important
determinants of intake and di~es~ibilj.ty
(Mertens, 1987; Illius an
"Readers are referred to Baldwin (1995) for further inforrnation, since metabolic models are not covered in this chapter.
52
errero et al.
undegradable(l
insoluble but
IGURE
nimal Crazing
Interface in
Models
523
BR = 0.144
ICW-0.144
BW-0.27,
r2 = 0.62
(17)
r2
(18)
= 0.76
SPR = 0.75PWC,
C.V.
15.5%
(19)
+ 0.00127~B)(3.96BW-0.27),
(20)
= 0.82
For nomenclature see Eqs. (16) and (17). These passage rates apply
to the B fraction and the indigestible fraction of cell wall of forages
and represent mostly small particles. Passage rates also depend on
the feeding level of the animal (AFRC, 1993). To account for effects
of feeding level (FL, in multiples of maintenance), these should be
multiplied by 0.25 FL, and a scaling rule similar to that claimed by
Sniffen et al. (1992) is obtained. Other relations can be found in Sauvant et al. (1995) but for total DM. Certainly more work is required
on this subject to understand the factors affecting passage rates (e.g.,
buoyancy and its relation to particle fer~entationand density). For
concentrated feeds,Sniffenetal.(1992)describe
the passage rate
(PRC) as
PRC(21)
= -0.424
PR)+ (1.45
r2 = 0.98
(22)
linking the animal with the soil fertility subsystems and their consequent effects on pasture growth in grazing models.
de~nitionof a
errero et al.
Models can be built solely to increase our understanding about systems under study and the nature of the functional relations between
parts of the system. For example, many authors have investigated the
effect of certain variables on the stability and steady states of grazing
systems (Noy-Meir,1975,1976,1978;
Johnson and Parsons, 1985;
Thornley and Veberne, 1989),and other scientists have looked at more
fundamental relationships between the animal and plant communities in terms of body weight effects (allometry) (Belovsky 1987; Demment and van Soest, 1985;Illius and Gordon, 31987; Taylor et al., 1987),
grazing behavior (Ungar and Noy-Meir, 1988, Ungar et al., 1991; Laca
et al., 1992; Parsons et al., 1994), diet selection (Belovsky, 1987; Parsons et al., 1994; Newman et al., 1994b), or digestive processes (Illius
and Allen,1994;Baldwin,1995).Increased
understanding of these
processes has led toimproved methods for modeling grazing systems
~ ~ a n t / A n Interface
i~a~
in Grazing
OUTPUT EXAMPLESOF
MULTIPLE RUNS
General
- farm sizes
- milk quotas
- labor inputs
Nutritiui\al andgrazing
management
concentrates offered
other supplements
continuous or rotational
grazing
..paddock rotation lengths
fertilizers used
herd structures
calving intervals
ages atfirst calving
mortality and culling rates
stocking rates
grazing
interface
grass growth
model
best optionsfo
sustainable use of
he farrn's resources
herbage mass
sward composition and
structure
nutritive value
FIGURE
A decision support system for grazing systems. [From Herrero
et al. (1996a).]
5 3 ~
errero et al.
lant/Ani~alInterface in Grazing
531
53
Edye, L. A., W.T. Williams, and W. H. Winter. 1978. Seasonal relations between animal gain, pasture production and stocking rate on two tropical
grass-legume pastures. Aust. I; Agric. Res, 29:103-113.
Finlayson, J. D., 0. J. Cacho, and A. C. Bywater. 1995. A simulation model
of grazing sheep. I. Animal growth and intake. Agric. Syst. 48:l-25.
Fisher, D. S., and R. Baumont. 1994. Modeling the rate and quantity of forage
intake by ruminants during meals. Agric. Syst. 4543-54.
Fisher, D. S., J. C. Burns, and K. R. Pond. 1987. Modeling ad libitum dry
matter intake by ruminants as regulated by distension and chemostatic
feedbacks. 3: Theor. BioE. 226:407-418.
Flores, E. R., E. A. Laca, T. C.Griggs, and M.W. Demment. 1993. Sward
height and vertical morphological differentiation determine cattle bite dimensions. Agron. I; 85:527-532.
Forbes, J. M. 1977.Interrelationships between physical and metabolic control
of voluntary feed intake in fattening, pregnant and lactating sheep: A
model. Anirn. Prod. 24:90-101.
Forbes, J. M. 1980. A model of the short term control of feeding in the ruminant: Effect of changing animal or feed characteristics.Appetite 2:21-41.
Forbes, J. M. 1993. Voluntary feedintake. In: ~uuntitutiveAspects o~Rurninunt
Digestion and ~eta~oEisrn.
J. M. Forbes and J. France (Eds,). Wallingford,
UK: CAB International, pp. 479-494.
Forbes, J. M., and J. France. 1993.Quantitative aspects of ruminant digestion
and metabolism. Wallingford, UK: CAB International.
Forbes, T, D. A. 1982. Ingestivebehaviour and diet selection in grazing cattle
and sheep. Ph.D. Thesis. University of Edinburgh, Scotland.
Forbes, T. D. A. 1988. Researching the plant-animal interface: The investigation of ingestive behavior in grazing animals. J. Anirn. Sci. 66:2369-2379.
Forbes, T. D. A., and J. Hodgson. 1985. Comparative studies of the influence
of sward conditions on the ingestive behaviour of cows and sheep. Grass
Forage Sci, 40:69-77.
Fox, D. G., C. J. Sniffen, J. D. OConnor, J. B. Russell, and P. J. van Soest.
1992. A net carbohydrate and protein system for evaluating cattle diets.
111. Cattle requirements and diet adequacy. 1. Anirn. Sci. 70:3578-3596.
Fox,D.G.,M.
C. Bamy, R. E. Pitt,D. K. Roseler, and W.C. Stone. 1995.
Application of the Cornel1 net carbohydrate and protein model for cattle
consuming forages. J. Anirn, Sci. 73:267-277.
Freer, M. 1981. The control of food intake by grazing animals. In: Gruzing
~ n i r n ~ E F.
s . H. W. Morley (Ed.). World Animal Science, B1, Amsterdam:
Elsevier, pp. 105-124.
Freer, M.,J. L. Davidson, J. S. Armstrong, and J. R. Donnelly. 1970.Simulation
of summer grazing. Proc. 11th. Int. Grassland Congress. Queensland:
Surfers Paradise.
Freer, M., A. D. Moore, and J. R. Donnelly. 1997. GRAZPLAN: Decision
support systems for Australian grazing enterprises. 11. The animal biology
model for feed intake, production and reproduction and the GrazFeed
DSS. Agric. Syst. (in press).
al ~nter~ace
in razing
533
53
errero et al.
~lant/Animali n t e ~ a ~ineCrazing
535
Kahn, H. E., and C. R. W. Spedding. 1984. A dynamic model for the simulation of cattle herd production systems. 2.An investigation of various
factors influencing the voluntary intake of dry matter and the use of the
model in their validation. Agric. Syst. 13:63-82.
Karnezos, T. R, N. M. Tainton, and D. I, Bransby 1988. A mathematical model
used to describe animal performance on kikuyu and coastcross I1 pastures.
J. Gruss~undSoc. S. Afiicu 5:38-41.
Kennedy, P.M., and M.R. Murphy 1988.The nutritional implications of
differential passage of particles through the ruminant alimentary tract.
Nutr. Res. Rev. 1:189-208.
Kibon, A., and E. R. Rlrskov. 1993. The use of degradation characteristics of
browse plants to predict intake and digestibility by goats. Anirn. Prod, 57:
247-251.
Kleyn, F. J., and R. M. Gous. 1988. Mathematical programming model for the
optimisation of nutritional strategies for a dairy cow. S. African J. Anirn.
Sci. 18:156-160.
Konandreas, P.A., and F.M. Anderson. 1982. Cattle herd dynamics: An integer and stochastic model for evaluating production alternatives. ILCA
Res. Rep. No. 2. Addis Ababa, Ethiopia:International Livestock Centre for
Africa.
Laca, E., and M. W. Demment. 1996. Foraging strategies of grazing animals.
In: The EcoEogy and nug gem en^ of Grazing Systems. J. Hodgson and A. W.
Illius (Eds.). Wallingford, U K CH13 International, pp. 137-158.
Laca, E. A., E. D. Ungar, N. Seligman, and M.W. Demment. 1992. Effects of
sward height and bulk density on bite dimensions of cattle grazing homogeneous swards. Grass Forage Sci, 47:91-102.
Laca, E. A., E. D. Ungar, and M. W. Demment. 1994. Mechanismsof handling
time and intake rate of a large mammalian herbivore. AppE. Anim. Behav.
SC~.39~3-19.
Laredo, M. A., and D. J. Minson. 1973. The voluntary intake, digestibility,
and retention time by sheep of leaf and stem fractions of five grasses.Aust.
I; Agric. Res. 242375-888.
Lauenroth, W.K., D. L. Urban, D. P. Coffin, W. J. Parton, H. H. Shugart,
B, Kirchner, and
M. Smith.1993. Modeling vegetation structureecosystem process interactions across sites and ecosystems. EcoE. ~ o d e ~ ~ i n g
67~49-80.
Leaver, J. D. 1986.Effects of supplements on herbage intake and performance. In:Grazing. J. Frame (Ed.). BritishGrassland Soc. Occasional Symp.
NO. 19, pp. 79-88.
Loewer, 0. J., K. L. Tad, L. W. Turner, N. Gay, and R. Muntifering. 1987.
GRAZE: A model of selective grazing by beef animals. A g i c . Syst. 25:
297-309.
Lopez-Tirado, Q., and J. G. W. Jones. 1991a. A simulation model to assess
primary production and use of BouteEo~agvacilis grasslands. Part I. Model
structure and validation. Agric. Syst. 35:189-208.
37
538
errero et al.
~lant/Anima~
lnte
53
et
errero
al.
Si~~ens
.
~ p t i m u mmachinery size is defined as that machinery set that results
in the minimum total cost, as shown in Fig. 1. In this chapter "optimum" and "least-cost" are used synonymously.
For corn or soybean farms and many other farms, machinery
size can bevaried by changing tractor sizes and matched implements
and/or combine sizesand matched attachments. This means that Fig.
can be visualized as three-dimensional, the dimensions being tractor size, combine size,and total cost, Total costis the sum of the costs
for machinery, labor, and timeliness. The least-cost machinery combination is determined by the low point on the three-dimensional
surface.
Y
IGURE
546
Siemens
and the fixed and variable costs can be estimated. Several sizes of
farm machinery sets could be used on this farm. As the size of the
machinery set increases, both productivity and initial price increase.
The annual machinery costs, fixed and variable, would increase with
increasing machinery size, as estimated by the lower line in Fig. 1.
It is necessary to h o w the relation between productivity of machinery and labor cost. For the owner-operator of a farm producing field
crops, the time used for operating machinery need not be considered
in selecting the optimum machinery size unless other enterprises
compete forthe time of the owner-operator and a value can be placed
on his or her time. Other enterprises might include a hog or beef
operation or a part-time job. Then the owner-operators time is important, and this cost should be considered.
If hired labor is available and used, the cost may be on anhourly
or annual basis. On an hourly basis, total labor cost is directly proportional to machine operating time and inversely proportional to
machine productivity. When labor is hired on an annual basis, total
labor cost is independent of machine operating time and productivity.
The question in either case is how to apportion labor costs between
field operations and other farm tasks. If the only reason for hiring
labor on a particular farm is to operate machinery, then the entire
annual cost of hired labor should be assessed against the field operations.
Figure 1 assumes that one man is paid on an hourly basis only
when he is operating field machinery. Labor cost
is constant until the
machinery is large enough to reduce total field time required below
the level of full-time employment for allthe hours the machinery can
be operated. When this occurs, labor costs will decrease becausetotal
field hours decrease.Field operations are completed when they
should be (decrease in timeliness cost) rather than when labor is
available.
Field ~ a c h i n e r ySelection
such as primary tillage, may have only an indirect influence on timeliness cost as they may affect the completion of subsequent operations,
It has been well established in the literature that yield decreases
if corn is not planted before the middle of May in the US. Corn Belt.
The value (or cost) of the yield decrease is the timeliness cost, The
date at which this cost begins has also been reasonably well established in the literature of the various state agricultural experiment
stations. There is little if any influence of factors other than the latitude of the location on this date. And there is general agreement t
the rate of yield loss is approximately 1 bushel per acre per day.
No such penalty has been established for soybean planting. Indeed, there seems to be general disagreement as to whether one exists
as long as time is available for the soybeans to reach maturity. A
realistic timeliness penalty for soybean planting and for planting of
other crops is needed.
It is logical that harvesting also has a timeliness cost associated
with it. Timeliness cost for harvest of both corn and soybeans depends on the variety, harvest losses, drying costs, and many other
factors. Thismeans that the many causes and their interrelationships
make the determination of a timeliness cost for harvesting complex.
For this reason, there is no general agreement on how to set a penalty
for late harvesting. For purposes of the program, a timeliness cost for
late harvesting is an option and is commonly used.
Productivity
(acres/ h)
hp combine
8 row corn head
ft grain platform
hp tractor
moldboard plow
ft chisel plow
ft disk
f t field cultivator
ft shredder
ft rotary hoe
Nine-knife applicator
ft fertilizer applicator
ft sprayer
ft drill
Subsoiler
Paraplow
Ro-till
ft co~binationtool
ft mower
l1 ft wheel rake
ft mower-conditioner
Baler (pto)wire
hp tractor
moldboard plow
ft chisel plow
ft disk
ft field cultivator
ft shredder
ft rotary hoe
11-knife applicator
ft fertilizer applicator
ft sprayer
24 ft drill
Subsoiler
Paraplow
Ro-till
f t combination tool
ft mower
l1 ft wheel rake
ft mower-conditioner
Baler (pto)wire
pto = Powered through power take-of3 of tractor.
List price
($)
Productivity (acres/h)
4.24
3.70
7.00
5.56
9.33
7.40
14.00
11.10
21
14.00
28.00
11.10
9,400
2,500
13,050
3,600
16,700
4,750
27,000
9,300
36,000
12,500
60,000
9,300
Siemens
55
12
12
4
4
4
4
4
4
4
4
4
4
4
100
120
140
200
220
12
12
24
12
12
12
12
12
12
12
12
12
12
12
12
12
12
12
12
24
24
24
24
24
24
24
The machinery cost factors file contains the constants for the equations used to estimate the fixed and variable machinery costs (Table
4). The constants and formulas were obtained from ASAE (1994).The
formulas for estimating remaining farm value are as follows:
Tractors
RFW
Combines
RFV
Balers, forage harvesters, and self-propelled RFV
sprayers
All other machines
RFV
= LP
= LP
= LP
= LP
where
(0.920)
ac~inerySelection
551
Table 4 Values of Repair Cost Constants RC1 and RC2 and Life
Life Machine
RC2
RC1
Tractor
Moldboard plow
Chisel plow
Disk stalks
Field cultivator
Stalk chopper
Plant corn
Row cultivator
Rotary hoe
Anhydrous applicator
Fertilizer applicator
Sprayer
Drill
Subsoiler
Paraplow
Forage chopper
Combination tool
Mower
Hay rake
Mower and conditioner
Baler
Self-prop. combine
Corn head
Grain platform
0.007
0.290
0.280
0.180
0.270
0.440
0.320
0.170
0.230
0.630
0.630
0.410
0.320
0.280
0.280
0.150
0.270
0.460
0.170
0.180
0.230
0.040
0.040
0.040
(h)
2.00
1.80
1.40
1.70
1.40
2.00
2.10
2.20
1.40
1.30
1.30
1.30
2.10
1.40
1.40
1h 0
1.40
1.70
1.40
1.60
1.80
2.10
12.10
2.10
12,000
2,000
2,000
2,000
2,000
1,200
1,500
2,000
2,000
1,200
1,200
1,500
1,500
2,000
2,000
2,500
2,000
2,000
2,500
2,500
2,000
3,000
3,000
3,000
Source: ASAE
RC1
(h/1000)Rc2
where
TAR = total accumulated repair cost, $
LP = list price of machine, $
h = total machine use, h
RC1, RC2 = constants (Table 4)
552
The life of each machine is used as a maximum limit for the time of
ownership.
Fuel consumption for tractors and combines is estimated by using the equation for fuel efficiency in ASAE (1994):
Diesel fuel, gal/(hp. h) = 0.52X
where X = the ratio of the equivalent pto power required to themaximum available from the pto. X is set in the program at 0.85.
Fuel and lubrication costs are estimated from the fuel consumption value, the price of fuel, and the assumption that lubrication cost
is 10% of the fuel cost.
The workday probability file lists the probabilities of a day being
suitable for field work based on historical data for weeks of the year.
Data are included for northern, central, and southern Illinois (Table
5). File space is provided for other locations. Theprogram user must
provide and store the workday roba ability data for his or her location
and when running the program must select the location to be used.
The probabilities fornorthern, central, and southern Illinois are from
Schwart (1981) and represent the fraction of time during the weeks
listed in which field work has been feasible at least 5 out of years
(83.3% of the time). The workday probability data may be changed
by the user.
User input files contain data that are likely to vary between users or
between farm situations. After the data are entered, these files may
be stored for a specific farm.
In addition to the list price of each machine and the field capacity of
each power unit-attachment combination, economicdata common to
each machine are needed to estimate the fixed and variable costs of
each machine. Thesedata are entered in the economic data file (Table
6). Data include the availability and cost of machinery operators;
price of diesel fuel; a housing, interest, and insurance charge as a
percent of remaining farm value; machinery purchase price as a percent of list price; and percent inflation. The maximum number of
machinery operators the program will utilize is six, and their cost
may be entered on an hourly basis. The program does not optimize
the number of machinery operators.
553
Northern
Central
Southern
0.00
0.12
0.36
0.44
0.39
0.59
0.46
0.61
0.61
0.64
0.64
0.66
0.47
0.47
0.48
0.63
0.46
0.56
0.56
0.42
0.00
0.00
0.14
0.26
0.36
0.28
0.46
0.46
0.51
0.51
0.60
0.70
0.70
0.51
0.56
0.56
0.71
0.51
0.58
0.50
0.26
0.00
0.07
0.17
0.30
0.30
0.43
0.50
0.57
0.57
0.62
0.62
0.68
0.59
0.54
0.54
0.58
0.54
0.40
0.40
0.11
0.00
1-13
14-15
16-17
18-19
20-21
22 -23
24-25
26-27
27-28
29-31
32-35
36-37
38-39
40
41
42-43
44-45
46
47
48 -49
50-52
0.00
Siemens
55
Timeliness factor
(7% yield loss per day of delay)
Crop
Planting
Harvesting
1.0
1.0
0.25
1.o
0.25
0.5
1.0
1.0
1.0
1.0
Corn
Soybeans
Wheat
Oats
Sorghum
7.
DesiredFieldOperationsFile
The most critical information input by the user for a specific farm is
the list of desired field operations to be performed and the data related to these operations. The main objective of the program is to
select the set of machinery required to complete the desired operations in a timely fashion at least cost. A list of desired field operations
for an example farm is presented in Table 7.
Code numbers are used to enter the field operations. The list of
available code numbers for the 30 different field operations included
in theprogram can be viewed on the screen. Other data to be entered
Price
($/bu)
Penalty dates ( M D )
(bu/ acre)
2.50
7.50
3
7.50
4.50
90.0
12.0
3.5
1.oo
40
Planting
Harvesting
5/15
5/31
10/30
5/31
5/31
l1/ 15
10/ 15
7/20
10/ 15
10/ 15
55
Start
Finish
date
Land
date Labor
Field operations
(M/D)
(M/D)
Combine soybeans
Combine corn
Chisel plow
Disk harrow
Field cultivate
Plant corn
Disk harrow
Field cultivate
Plant soybeans
Row cultivate
Row cultivate
9/15
10/ 15
15
4/1
4/25
4/25
4/1
5/1
5/1
Area
(acres) (h/day)
500
500
500
500
500
500
500
500
500
500
500
10.0
10.0
area No.
1
2
2
1
1
1
2
2
2
1
2
for each operation include the earliest start date; latest finish date;
acres to be covered; available labor hours per day or the hours per
day the operation can be performed, whichever is least; and a land
area number. Entry of a latest finish date is optional and is needed
only if an operation must be completed by the date entered. The land
area numbers are used to help ensure that the operations are scheduled correctly; they are explained in more detail below.
The first operation listed in Table 7 is Combine soybeanswith
an earliest start date of 9/15. This is interpreted to mean that September 15 is the earliest date soybeans are commonly ready for harvest.
A latest finish date is not specified for soybean harvest. The next
entries are the acres to be covered for the operation and the number
of hours per day the operation can be performed. The latter figure is
assumed to be the number of hours spent in the field operating machinery, No allowance is made for time used to adjust and service
equipment or for traveling to fields.
The second operation listed for the example farm is Combine
corn, with an earliest start date of October 15. The third operation,
Chisel plow, is to be performed on the same land from which corn
is harvested. Thus, the land area number is the same for both combine
corn and chisel plow operations. Land area numbers are used in
scheduling the field operations. For the example farm,the chisel
556
Sie~ens
operation will not get ahead of the combine soybean operation when
the operations are scheduled.
The desired spring operations for corn are Disk harrow, Field
cultivate, and Plant corn. All of these operations are to be performed in sequence on the land that was in soybeans the previous
year, and therefore the land area number is 1 for each of these operations. For soybeans, the spring operations are disk harrow, field
cultivate, and plant soybeans, and the landarea number is 2 for these
operations. Both crops are to be cultivated once.
A primary objective of the program is to determine the optimum
set of machinery and the machinery-related costs for the list of desired field operations (Table 7).
Quantity Sizea
Tractors (max. 6):
Large
Small
1
1
5
1
60 hp
80hp
hp
120 hp
140 hp
5 = 160 hp
6 = 180 hp
7 = 200 hp
8 = 220 hp
Combines (max. 3)
Corn heads
Grain platforms
1
1
1
(4-row head)
(6-row head)
(8-row head)
(12-row head)
Implement
Implement
Spray pesticide
Drill
Subsoiler
Paraplow
Chop forage
Combination tool
Mow hay
Rake hay
Mow and condition
Bale hay
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
Moldboard plow
Chisel plow
Disk
Field cultivate
Chop stalks
Plant
Row cultivate
Rotary hoe
Apply anhydrous
Apply fertilizer
1
1
1
0
0
0
0
0
1
1
1
2
Three options are available when running the program. The first option is an attempt to provide the user with an estimate of the maery costs, work schedule, etc., for a specified number and specisizes of tractors and combines. Themachinery set specified might
be the current set of machinery on a farm or a set in which th
st farmers have a set of machinery currently
interested in an estimate of the total cost inc
potential timeliness penalties and the work schedule. Or they ma
want to evaluate the effects of changing only one or two machines,
or they would like to h o w the effects of using the present machinery
on a lar er farm. The first option of the program may be used for
these
oses.
e second and third options are programmed to determine the
timum set of machinery for the desired field operations.
second option the user specifies the number of tractors, combines,
55
Sieme~s
and laborers, and the program optimizes the sizes of the tractors and
combines. The third option is an attempt to determine the optimum
set of machinery including number and sizes of tractors and combines and number of laborers. The search for the optimal machinery
set is explained here using the second option.
Assume that the program user has entered the list of field operations as shown in Table 7, specified the number of tractors and
combines as shown in Table 8, and matched tractors to implements
as shown in Table 9. Note that in Table 8 for the example farm, the
number of tractors has been set to one large tractor and one small
tractor and the number of combines to one. Now the goal of the
second option is to find the sizes of the tractors and combines that
result in the least cost.
The program first sets the size of the large tractor to 220 hp, the
largest available in the program; the small tractor to one tractor size
increment less than the largest tractor size; and the size of the combine to 225 hp, the largest in the program. With this set of machinery,
the input data, the stored data files, and the field operations are
scheduled, and the costs for machinery, labor,and timeliness are calculated. These costs are summed to get the total average annual machinery cost.
Then both the small tractor and combine sizes are decreased,
and again the work schedule and costs are computed. The sizeof the
small tractor and combine continue to be decreased until the total
annual cost increases due to increased timeliness costs. Then combinations of small tractor and combine sizes are tried until the leastcost machinery set has been found with the size of the large tractor
fixed at 220 hp. Table l 0 duplicates the screen that shows the costs
in dollars per acre for the different machinery sets evaluated for an
example farm with the large tractor fixed at 220 hp. The
on
the screen indicate a machinery set that is not able to complete the
operations within a calendar year, a tractor of insufficient sizeto pull
an implement, or an operation that could not be completed by the
latest finish date listed in the input data.
Next the size of the large tractor is reduced by one size increment, in this case to 200 hp, and the above procedure is repeated.
The size of the large tractor is reduced until the least cost increases.
During this process, no trials are made with both tractors the
same size, which in some situations would result in the least cost.
Therefore, the final set of trials is made with both tractors the same
size. Tractor and combine sizes are varied until the least-cost combination is found.
Thus, the total cost is determined for all possible size combinations of two tractor sizes and combine sizesthat could result in the
559
60
80
100
120
140
160
180
200
220
180
60.23
66.32
63.88
64.66
66.46
69.51
140
Combines, hp
Cost ($/acre)
Large
Small
Set"
tractor
tractor
Combine
Machinery
Labor
Timeliness
Total
160
180
200
160
l80
140
220
200
80
80
80
100
100
80
80
100
215
215
215
215
215
215
215
215
49.95
51.61
52.55
52.26
53.92
49.16
54.84
54.86
6.10
5.93
5.79
6.10
5.93
6.37
5.40
5.79
1.19
0.00
0.00
1.19
0.00
4.64
0.00
0.00
57.24
57.53
58.34
59.55
59.85
60.17
60.24
60.66
"Each set consists of one large tractor, one small tractor, and One combine each with
appropriate attachments.
set. The presentation includes for each set only the sizes of the large
and small tractors and the size of the combines; costs for machinery,
labor, and timeliness; and total cost.
The program user may obtain a presentation of the output information for any of the eight lowest cost machinery sets found during the
search. Inthis section we discuss the output information for the leastcost machinery set for the example farm.
.
The first output screen is a list of the machinery set including the
assumed purchase price and annual use of each machine (Table 12).
The annual number of hours each operator spends operating machinery i s also given.
The computed work schedule is the most critical portion of the program output (Table 13). Considerable efforthas been devoted to making the computed work schedule realistic in terms of utilizing operators and machine~y.
Purchase price
Machine
(h)
215 hp combine
8-Row corn head
22 ft grain platform
160 hp tractor
11 ft chisel plow
25 ft disk harrow
29 ft field cultivator
80 hp tractor
8-Row planter
8-Row cultivator
106,020
25,200
13,320
67,500
8,190
500
15,930
11,970
35,100
20,430
6,390
(acres)
173
91
82
231
76
89
66
236
107
129
500
500
1,000
1,000
1,000
1,000
Calendar
Finish
Start
date date
Combine soybeans
Combine corn
Chisel plow
Disk harrow
Field cultivate
Plant corn
Disk harrow
Field cultivate
Plant soybeans
Row cultivate
Row cultivate
9/ 15
10/15
11/1
4/1
4/25
4/25
4/24
5/22
5/22
6/5
6/21
10/8
10/31
11/ 16
4/24
5/6
5/ 14
5/22
5/30
6/5"
6/21
7/5
days
Work
days
Acres
24
17
16
24
12
20
29
9
15
17
15
10.3
9.1
7.6
4.5
3.3
5.4
4.5
3.3
5.4
6.4
6.4
500
500
500
500
500
500
500
500
500
500
500
-~
562
Siemens
The program schedules the field operations on a day-to-day basis, giving priority to the order in which the operations are listed in
the input data. For the desired field operations listed in Table 7, the
first operations scheduled are the harvesting operations, When only
one machinery operator is available for harvesting it is assumed that
the operator spends 50% of the time hauling and processing the grain.
Thus, the number of hours per day one operator actually spends operating the combine is reduced. When two operators are available,
the combine operates the full number of hours per day listed in the
input data and theoperation requires the use of both operators. It is
realized that large variations exist in the time spent in unloading,
hauling, and processing grain during harvest operations. Revisions
of the computer program may be needed to account for these variations.
Each field operation starts on the date listed in the input data
unless an operator and machine needed to carry out the operation
are not available. Other operations could be under way utilizing all
operators or machines required. In that case the start date for the
operation in question is the first date that both an operator and required machines are available. The acres completed on a given date
equal the productivity (acres per hour) for the machinery being used
times the hours per day labor is available timesthe probability of that
date being suitable for field work.
If an operation is not completed by the latest finish date listed
in the input data or if all operations are not completed within one
calendar year, the machinery set is regarded as being unacceptable.
For the example fieldoperations listed in Table 7, the operations
are scheduled beginning with the fall operations and then the spring
operations are scheduled. No timeliness penalties occur in conjunction with combine soybean, combine corn, or plant corn operations
as these are completed before their respective penalty dates. Soybean
planting is not completed until after the,specified penalty date (Table
4), and therefore a timeliness penalty is calculated.
As the field operations are scheduled, the annual hours of use are
accumulated for each machine. The fixedand variable costs for each
machine are calculated as previously explained using the annual use
input data and data
from the stored files. The fixedand variable costs
are summed, and the result is used to calculate the cost per acre for
each field operation (Table 14).
Field ~ a c ~ i n e Selection
ry
Table
Machine
215 hp combine
Years
of use
Annual
hours
173
6.75
91
82
231
76
89
66
Power and
machine
cost
($/ acre)
Machine
only
($/ acre)
25.60
236
3.90
107
129
6.40
1.10
If desired, the detailed costs for each machine are provided for each
year up to a maximum of 10 years of ownership. For the example
farm the costs of the combine and corn head are presented in Tables
Table
Year
1
2
4
5
6
7
8
9
173
346
59,771692
65,090 866
69,797
1,039
1,212
77,6491,385
80,912
1,558
39,298
46,971
73,963
HIIb
Repairs
Average
cost
($/ acre)
12,722
19,928
31,950
36,944
42,365
45,277
48,739
51,803
118
522
2,250
3,597
5,278
7,297
9,660
12,371
312
167
147
133
123
115
109
104
"List price = $117,800; purchase price = $106,020; fuel price = $l/gal; fuel use = 10.88
gal/hr; machine use = 173 h/yr.
bHII = Housing + interest + insurance.
etailed Costs
for
Eight-Row
costs
Accumulated
Total
use
Year
(h)
1
2
3
4
5
6
7
8
9
10
91
181
272
363
454
544
635
726
817
907
Corn Head"
Average
($)
Percent
Depreciation
of life
3
6
9
12
15
18
21
24
27
30
9,341
11,165
12,779
14,207
15,471
16,590
17,580
18,457
19,232
19,918
HIIb
Repairs
($/acre)
3,024
4,737
6,253
7,594
8,781
9,832
10,762
11,585
12,313
12,958
7
32
75
138
220
323
447
591
757
945
24.75
15.90
12.75
12.00
9.80
8.90
8.20
7.65
7.20
6.75
"List price = $28,000; purchase price = $25,200; machine covers 500 acres/yr; fuel cost
= $1.97/acre; productivity = 5.51 acres/h; annual use = 90.7 h/yr.
bHII = Housing + interest + insurance.
Average Annual
ost item
achinery fixed cost
Fuel cost
Repair cost
Total machine^ cost
Timeliness cost
Total cost
achine~-RelatedCosts
$/F
41,874
4,708
3,368
49,950
6,098
1,190
57,238
$ / acre
41.87
4.71
3.37
49.95
6.10
1.19
57.24
cost is calculated for the acres of any crop planted or harvested after
the penalty dates specified. From the work schedule explained previously for the example farm, the "Plant soybeans" operation was
completed late. Thus, a timeliness costis calculated forthe oper~tion.
The costs for machinery, labor,
and timeliness are summed to provi
an estimate of the total annual machinery-related costs for the machinery set.
Siemens
57
La!
~ole-FarmS i ~ ~ ~ a tof
i oField
n
571
[Tractor-l, Tractor-2],4.74,7,Avail)
The other object-classes formodeling a whole-farmsystem could
include farm, labor, tractor, irrigation equipment, crop, field, operation, scheduled work hours, and daily weather. A schematic of an
object data model for the whole-farm system consisting of these
object-classes is depicted in Fig. 1. The attribute set for each of these
classes is presented in Table 1.
According to this model the farm consists of crops, fields, operators, tractors, and implements. A crop can be grown on one or
more fields. Each crop has a list of operations that need to be carried
out in proper sequence for it to grow and yield. Each operation requires an implement for its execution and would need to meet certain
agronomic requirements to be effective for the crop production. Implements can be self-propelled orrequire a separate power source. A
Has
Attributes
~-
Farm
Labor
Tractor
Implement
Irrigation equipment
Crop
Field
Operations
Scheduled work hours
.
The object-oriented model just described is a static representation
a farm system. It depicts the structure of its objects and their relationships at a single moment in time. Attribute values and links held
by objects are called their state. Over time, objectschange their states
in response to external or internal stimuli. Simulation models examine these changes to the objects and their relationships over time. In
developing such a model, we need to define and understand the dynamic aspects of the system and how they influence different objects.
These aspects, referred to as controls by Rumbaugh et al. (1991), describe the sequences of operations that occur in the system in response to stimuli. An individual stimulus from one object to another
is an event. The response of an event depends on the state of the object
receiving it and can include a change of state or the sending of another event to an original sender or to a third object.
Controls and operations for a whole-farm system with multicrop production activities are summarized below. A simulation
model should capture these controls in as realistic a manner as
sible. These controls are performed on each field every day in our
model.
1. Search through the farm knowledge base, find pen
erations, and select the operation that can be perfor
the field.
2. Check the conditions that need to be satisfied to make the
operation effectiveforcrop
production, for e
simulation day must be within the agronomic
the operation, that is, the time window when t
is appropriate for this crop.
3. Identify the machinery system (implement and an operator
for self-propelle~machinerysystems; implement, power
source, and an operator for the machinery systems that require a separate power source) for the operation, and check
its availability.
4. Estimate the actual work hours for the day based on the
operation type, scheduled work hours, climate, and soil- and
crop-related factors.
5. Estimate the operational capacity for the day based on the
selected machinery system-width, speed, and efficiencyand the actual work hours.
6. Calculate the actual work done, and update the farm knowledge base to reflect the changes in field con~itions-completed and pending operations and availability status of the
machinery system (implement, power source, and operator)
for the day.
7. Prepare and store the simulation reports, including the o
erations completed and the machinery system use
PROLOG (PROgramming in LOGic) is one of the two artificial intelligence (AI) languages (LISP is the other) and provides an environment for developing computer models with an object-oriented data
structure. It.is defined as a declarative language, in contrast to procedural languages such as Basic, Fortran, Pascal, and C. Of course,
many other software development programs are available, including
the various Visual packages for developing Windows-based programs. PROLOG contains a basic scheme forrepresenting knowledge
and also has a built-in inference engine. A program in PROLOG is a
collection of facts and rules @nowledge base) about the system. These
facts and rules are assembled in two principal components of the
program code, namely predicates and clauses. In a broader sense,
predicates are analogous to subroutine definitions of procedural languages, and clauses are similar to calls for a subroutine with specific
argument values. From the collection of facts and rules, PROLOG
derives solutions to the questions posed interactively by the user or
listed as goals within the program code. The search for solutions
to the question or the goal is made through all the clauses. This
permits achieving multiple solutions to a single query. An expert system written in PROLOG will check all the rules for possiblesolutions,
whereas some expert system development programs stop and present
only the first solution found. The capabilities such as symbol processing, list manipulation, and recursion facilitate handling heuristic,
in addition to quantitative and procedural, computations necessary
for simulating dynamic behavior of a system.
PROLOG has been employed mostly for writing expert systems.
There have been some efforts to develop simulation programs, simulation languages, and object-oriented pro~amminglanguages using
PROLOG (Adelsberger, 1984; Futo and Gergely, 1986,1987; Kornecki,
1986;Yokoi,1986;Fan and Sackett, 1988). However, there is no reference in the literature to the use of this language for agricultural
simulation modeling.
Turbo PROLOG (Borland, 1984) was used in this project. This
version ofPROLOG provided a convenient programming environment. It permits writing the code in natural-language-like sentences,
which facilitates better understanding of program logic by nonprogrammers. The Turbo PROLOG code can be compiled into an executable program and can be distributed freely to different users.
~ole-FarrnS i r n ~ ~ a t of
io~
Field
575
57
imulateSeason(SDay,~ay):asserta(currentday(Sday)),
repeat,
currentday
(SimDay)
,
makeAvailAl1,
simulateADay(Simday,
"Irrigation")
,
simulateADay(Simday,
"WarvestingOps")
,
simulatemay (Simday,
"PlantProtectionOps")
,
simulateADay(Simday,
"PlantFertiOps")
,
simulatemay (Simday,
"LandPrepOps")
,
Sdayl=Simday+l,
),
changeDay
(Sdayl
Sdayl>Fday.
imul
fieldc
(FieldN,
Crop)
c h k C o n ~ d W o r (FieldN,
k
fail.
imul~t~ay.(-,-)
:-
Crop,
Sday,
OpsType)
~hk~on~~Work(Fi~l~,~rop,-,-):fieldo.(FieldN,-,-,-,Crop,OpsList,-,
"I - /" / ? ) /
OpsList=[]
!.
577
57
FIGURE3 Tasks performed by the Operations Simulator for scheduling irrigation and other field operations for a field on a daily basis.
have been processed. This ensures that all the fields listed in the
fieldc database are checked daily for every type of operation. Having
processed all entries of this database, the clause fails even prior to
calling the predicate Chk~ondAnd~ork.
At this stage PROLOG
searches for an alternative sinnulateADay clause or rule and finds
simulateADay(-,-):-!., which always succeedsfor any value of
SDay and OpsType. It thus successfully completes the task of checking all fields on the farm for carrying out the operation type instantiated (the assigned value of OpsType) in the simulateADay predicate. Thesame process is repeated for other operation types by calling
the simulateADay predicate with different values of OpsType for
the current day.
hole-Farm Sim~~ation
of Field ~ p e r a t i o ~ s
57
such as
10) Get details for the operation
etc.
agronomic window, implement list,
11) Check suitability of the "current day"
f o r the
operation with respect to its agronomic window.
The "current day" should be within the work
window of the operation.
12) Check the availability of the required machinery
set for the operation; the implement and the
operator in case of self-propelled implements, and
the implement, the tractor (power unit) and the
operator in the case
of other types of implements.
FIGURE3 Continued
c ~ ~ o n d A n d W ois
r kthe most demanding predicate. It checks
several conditions to perform its task. However, its clauses are structured to perform efficiently. They avoid deep level searches for the
operations that have been already completed and also for the operations for which the field isnot ready. Thetime required for the day's
simulation decreases as crops on different fields are harvested and
the operations lists associated with the crop become empty.
The simulator produces three reports: work report, no-work report, and summary report. The work report contains daily information about the work performed during the simulation, and the nowork report contains daily information about the operations that
were attempted but could not be done because of factors such as
nonavailability of the machinery or excessive rain during theday. The
summary report summarizes the work and no-work reports for each
operation for the entire cropping season. Table presents the contents
of these reports.
Summary report
Contents
Julian day, month,
field,
crop,
operation, total area,
accumulated done area, days work, implement,
tractor, operator, and number of hours worked
Julian day month,field,crop
and operation, total
area, done area, reason for no-work, implement
list and its availability report, operator list and its
availability report, tractor list and its availability
report
Field,crop, operation, scheduled startand finish
times, actual start and finish times, number of
days and hours worked, total done area, number
of nonworking days
La1
SY:
Yes
(ha)
Cropping system
102.2
76.8
6.0
Cotton-Cotton
Cotton-Wheat
Cotton-Soybeans
79.6
Soybeans-Cotton
10.0
Peanut-Cotton
26.4
NoCrop-NoCrop
This led to a total of l0 fields (Table 3) with the total cultivated area
of the farm unchanged.
Wheat covered approximately 25% of the cultivated area as a
second crop on the farm. It is planted in the second half (OctoberNovember) of a year and harvested in the following year. Therefore,
it became necessary to simulate for two consecutive years to cover
the entire cropping season. The second yearcrops were suffixed with
the character 2, and their agronomic windows for different operations for the second year were adjusted by adding 365 to their values
of the first year.
Weather (especiallydaily rainfall) and scheduled work hours are
required inputs for the simulation. The actual weather of Tallahassee,
Florida, for the years 1954 (dry), 1960 (normal), and 1964 (wet) were
used for the test runs. The runs were made by selecting Dry followed by another Dry year.
The firsttwo shulations were made with all the machinery and
labor available on the farm for 6 and h of daily scheduled work for
the entire simulation period. These simulations resulted in zero done
area for a number of operations on several fields. Thisindicated that
a fixed daily work schedule of 6 or h for all calendar months during
the cropping season would not be sufficient to attend to the operational r e q ~ ~ e m e nwith
t s the available machinery and labor on the
farm. A flexible schedule with more daily work hours during agro-
w o r k r ~ p o r t ( j u l i a n D a y , Month, F i e l d , Crop,
Operation, CultArea, Day DoneArea, TotDoneArea,
Implement I Operator, Tractor, ActWorkHrf
workr~port
f t A p r i l t f , v g F i e l d - l t t , ffCottonff
,
ffPlowinglg,1 0 2 . 2 , 1 4 . 2 , 1 4 . 2 , tgBottomPlowff
,
tfJerryfl
, I8Tractor-l" , 1 2 )
workr~port
f t A p r i l f t , " F i e l d 2 " , tfCottonsf
,
"Harrowing" , 76 8 , 30.2 , 30.2:
ffDiskHarrowff
,
"Keithff,fgTractor-21f, 1 2 )
.)
workreport
f f F e r t i l i z e t,l 4 1 . 6
, f t A p r i l f f , f t F i e l d - 5 f,f
noworkre~ort
7 9 . 6 , " S o y b e a n s f g ,f g F e r t i l i z e r f 0
,,
WIachinesNotAvailff,
[ t f F e r t i S p r e a d e r L P t t,] ftNotAvailtl,
[ [ t 8 F e r t i S p r e a d e r L P g f i, e l d 4 " ,
rtCottonlg
,
f g F e r t i l i z e f t] l,
[ ttTractor-4tt, ftTractor-5ff , "NotCheckedtl, El ,
["Keithtf, rrRaytf,
flJoeft
, ftJerryF1]ltNotCheckedf1 11
noworkreport
' t A p r i l t 8 ffField-8ss
,
1 1 . 8 , ffSoybeansff,f t F e r t i l i z e 8 5 0, ,
llMachinesNotAvailff,
[ttFertiSpreaderLPf8], flNotAvailff,
[ [ f f F e r t i S p r e a d e r L P gI tf ,f i e l d 4 " , "Cottonff
lfFertilizeft] ,
[ tfTractor-4f1,
KfTractor-5111
, ftNotCheckedff, [l
[ f f K e i t h vlfRaya*,
f,
8 1 J e r r y *,fftNotCheckedtl
]
, [l
FIGURE6 Examples of work and no-work reports of the FARhfSYS Operations Simulator.
La!
4th run
5th run
6th run
Operation
AStSfn Sst
Ah
ASt
Afn AFnASt
181
183
184
186
173
182
184
186
181
183
184
186
179
187
190
192
186
189
191
192
187
189
191
192
182
190
193
194
189
192
193
194
190
198
202
203
197
200
202
203
188
190
193
194
190
193
195
296
190
193
195
196
202
203
204
206
202
203
204
206
192
193
195
196
191
194
197
198
193
194
197
198
204
210
211
212
209
210
211
212
200
200
197
197
547
554
555
557
553
554
556
557
548
556
557
558
555
556
557
558
184
195
290
184
195
290
183
192
290
183
192
290
~ o ~ e - FS
a ir~~u ~ a t i oofn Field
e er at ions
Table 4 Continued
S. window
Operation
Sst
Sfn
4th run
ASt
5th m n
6th run
Afn
ASt
A h AFnASt
186
151
186
196
151
186
196
152
184
193
152
184
193
554
558
559
560
554
558
559
560
556
559
560
561
556
559
560
561
109
117
118
202
109
117
117
117
118
118
118
202
119
208
119
208
548
555
561
567
568
552
557
561
567
568
547
557
567
568
569
548
558
567
568
569
SSt = Earliest start time, SFn = latest finish time, ASt = actual start time, AFn = actual
finish time. The timings of other operations on different fields remain unchanged,
b4th run: With Ray and Tractor-5 available for work. 5th run: With Ray withdrawn
from the farm. d6th run: With Ray and Tractor-5 withdrawn from the farm.
as much as the most used item of the same class and there
were one or more complementary units to perform the same
operations.
2. Recommend increasing the machinery capacity to ixnprove
the timeliness for the delayed operations (l)by increasing
scheduled daily work hours, (2) by increasing the speed of
operation for the machinery, or (3) by acquiring a bigger
machine if neither of the first two options is possible.
La1 et al. (1990) present these rules in detail and describe how
they were used for searching through the simulation reports and
farm knowledge base for developing recommendations by the
Expert Analysis System. The subsequent simulation scenarios were
created by incorporating recommendations of the Expert Analysis
System for the previous run.
The fourth simulation was made after removing ChiselPlow,
~OTillPlanter,PDSprayer2, GeneralPlanter, and PDSprayer.These
implements were not used at all during the third simulation and
were recommended for removal by the Expert Analysis System. This
simulation produced exactly the same reports as were produced
during the previous run.
The fifth simulation was made after removing "Ray" (the least
utilized operator during the fourth simulation and recommended
for removal) in addition toother removals. All operations were completed within their agronomic windows during this simulation also.
owever, certain operations such as Sprayingl on Field3 previously performed by Ray were performed by Keith and got shifted
by 3 days (from day 197 to day 200) in their start and completion
(Table 4). This happened because the substituted operator, Keith,
during the period of these operations was busy elsewhere and was
not available to start the operations on the days carried out during
earlier simulations.
The removal of Ray during the simulation also affected the
performance of TractorS. This tractor was mostly operated by Ray
~ u r i n gearlier simulations. There was a slight change in the operations performed by Tractor-5. Its overall utilization decreased from
120 h in 13 days
to 108 h in 1 2 days. This reduction can be attributed
mainly to the nonavailability of an operator for Tractor-5 when it
became a possible choice for the work.
The sixth simulation was made after removing Tractor3 in addition to Ray and other removals. This was done based on the recommendatio~of Expert Analysis System after the fifth Simulation.
All operations were completed within their agronomic windows
during the sixth simulation also. There were some additional shifts
in the start and/or completion days of certain operations such as
first year ~ultivation2for the peanu crop and second year Fertireading for the cotton crop on Fie -7. It was further interesting
to note that certain other operations such as Sprayingl for the first
year soybean crop on Field3 were expedited with the combined
withdrawal of Ray and Tractor3(Table 4). At first glance, it seemed
590
La1
k
42
p!
d
ti
~ ~ o l e - F a rSimulation
m
Field Operations
m a
+-)
ai
PDSprayer
PDSprayer2
Spraycoupe
General PLanter
NoTilPlanter
GrainDril
Bedder
General Grain Combine
Peanut Combine
Cotton Picker
BottomPlow
FertiSpreaderLP
Subsoiler
Planter-77
BMower
FertiSpreaderPF
Operators available
Laborer-3
Laborer-4
Laborer-l
Laborer-2
Tractors (HP) available
JD4450 (148)
JD4430 (125)
JD2950 (90)
JD3020 (70)
JD2640 (70)
Identified for withdrawal.
Theobject-oriented
approach to simulation of field operations
showed considerable advantages over the conventional approach to
simulation. The new approach helped in organizing, representing,
and manipulating farm knowledge in different object-classes. Farmers preferencesand priorities for different fields,
operation types, and
assigning implements to an operation, tractors to an implement, and
operators to a tractor were captured and used in the simulation process. The approach resulted in a flexible simulation model that rescheduled operations on different fields, as a farm manager would
do in a real situation, based on the availability of machinery and
labor. The model can also be used for a variety of farming situations
ranging from a highly mechanized farm (such as the test farm) to a
labor-intensive farm with little or no mechanization. It is feasible be-
Fan, I. S., and P. J, Sackett. 1988.A PROLOG simulator for interactive flexible
manufacturing systems control. Simulation 50(6):239-247.
Flowers, E. R. 1988. Failing with grace. Turbo Technix ( ~ o ~ l Language
an~
J.)
I(5):76-85.
Freeman, S., A. D. Whittaker, and J. M. McGrann. 1989. Knowledge-based
machinery management aid. ASAE Paper 89-7576. St. Joseph, MI: ASAE.
Futo,I., and T. Gergely.1986. TS-PROLOG: A logic simulation language,
Trans. SCS 3(4):112-119.
Futo, I., and T. Gergely. 1987. Logic programming in simulation. Trans. Soc.
C~mput.Simu. 3(3):195-216.
Glunz, D. J. 1985. Comparative analysis of Florida citrus harvest systems.
MS Thesis. Gainesville,FL: University of Florida.
Helms, G. L., J. W. Richardson, M. E. Rister, N. D. Stone, and D. K. Loh.
1987. COTFLEX:A farm-level expert system to aid farmers in making farm
policy decisions.Annual Meeting of Am, A@c. Econ. Assoc. East Lansing,
MI.
Kornecki, A. 1988. Simulation and artificial intelligence as tools in aviation
education. In: AI Papers. R. J. Uttamsingh (Ed.). Simulation Series Vol. 20,
No. 1. San Diego, CA: Society for Computer Simulation, pp. 121-126.
Labiadh, S., and J. C. Frisby. 1987. A simulation model to select alfalfa harvesting machines. ASAE Paper 87-1047. St. Joseph, MI: ASAE.
Lal, H. 1989. Engineering farm knowledge for a seamless decision-support
system. Ph.D, Dissertation. Gainesville, FL: University of Florida.
Lal, H,, R. M. Peart, J. W. Jones, and W. D. Shoup. 1990. An intelligent information manager for knowledge-based systems. A p p l . Eng. A g i c . 6(4):
525-531.
Lal, H., R. M. Peart, W. D. Shoup, and J. W. Jones. 1991a. Expert result analyzer for a field operations simulator. Cornput, Electron. A g ~ i c 6:123-241.
.
Lal, H., R. M. Peart, J. W. Jones, and W. D. Shoup. 1991b. An object-oriented
field operations simulator in PROLOG. Trans. ASAE 34(3):1031-1039.
Lal, H., J. W. Jones, R. M. Peart, and W. D. Shoup. 1992. FAMSYS: A wholefarm machinery management decision support system. A g i c . Syst. 38:
257-273,
Miles, G. E., and Y. J. Tsai. 1987.Combine systems engineering by simulation.
Trans. A S A E 30(5):1277--1281.
Nute, D., G. Zhu, and M. Rauscher. 1996. DSSTOOLS: A toolkit for development of decision support systems in PROLOG. Proc. 6th Int. Conf. on
Computers in Agriculture, Cancun, Mexico. St Joseph, MI: ASAE.
OKeefe,R.M., and J. W. Roach. 1987. Artificial intelligence approaches to
simulation. J. Opt. Res. Soc. 38(8):713-722.
Rumbaugh, J., M. Blaha, W. Premerlanim, F. Eddy, and W. Lorensen. 1991.
O~~ec~-Oriented
Modeling and Design. Englewood Cliffs, NJ: Prentice-Hall.
Smith, R: D.1985.Biomassstorage,decision
support systems and expert
systems in crop production and processing. Ph.D. Thesis. West Lafayette,
IN: Purdue University.
~ ~ o l e - F a rSim~lation
m
of Field
595
Stone, N. D., R. E. Frisbie, J. W. Richardson, and R. N. Coulson. 1986. Integrated expert system applications for agriculture. Proc. Int. Conf. on Computers in Agric. Ext. Programs, February 1986. Orlando, FL, pp. 836-841.
Thai, C. N., and C. P. Wilson. 1988. Simulation of peach post harvest operations. ASAE Paper 88-6058. St. Joseph, MI: ASAE.
Tsai, Y. J., J. W. Jones, and J. W. Mishoe. 1987.Optimizing multiple cropping
systems: A systems approach. Trans. A S A E 30(6):1554-1561.
van Elderen, E. 1981. Scheduling of field operations: Research report. M A G
82(3):865-871.
van Elderen, E. 1987. Sc~edulingFarm O ~ e ~ a t i o A
n s S~ ~ m u l ~ t i oModel.
n
Wageningen, Netherlands: Pudoc.
Whittaker, D. A.,E. J. Monke, and F. R. Foster. 1987. ADAM: An ADaptive
Assembler of Models. ASAE Paper 87-5536. St. Joseph, MI: ASAE.
Yokoi, S. 1986. A PROLOG based object-oriented language and its compiler.
In: Lecture Notes in C o ~ ~ u tScience.
er
G. Goos and J. Hartmanis (Eds.). Logic
Programming '86 Proc. 5th Conf., Tokyo, Japan.
applications to predict processes such as soybean phenology (Elizondo et al., 1994), aflotoxinconcentrations in peanuts (Parmar et al.,
1994), optimum temperatures for greenhouses (Seginer and Sher,
1992),insectpest treatment thresholds (McClendon and Batchelor,
1995), and recognition of patterns from digital images (Deck et al.,
1991).
There are many commercial neural network packages that are
available on both PC and workstation platforms. The most basicand
inexpensive software packages are simply collections of neural network algorithms that can be incorporated into other programs (see
Rummelhart and McClelland, 1986). The user must supply code to
input data into the algorithms and to output data in useful forms.
The advantage of this approach is that the algorithms can easily be
incorporated as components of larger models. These algorithms can
also be rewritten in many programming languages and can be implemented on different computer platforms. The most sophisticated
PC-based software packages are highly sophisticated Windows-based
programs. These programs typically require little time to learn, and
users can very quickly begin to develop a neural network. The only
thing that the user must do is provide input data, select an architecture, and set up the number of hidden nodes and learning rates.
These packages typically allow data to be input directly, or data can
be imported from spreadsheets, ASCII datafiles, orvarious databases.
Often, these packages allow the user to select from among several
different architectures. Outputs vary from tabular to graphical, and
outputs can often be exported into spreadsheet, ASCII, or database
files. Many of these packages can compile a neural network into an
executable code that can be called from simulation models.
~ a t h e ~ a t i crepresentation
al
of a biological neuron.
other neurons and derives the appropriate weights for the interconnected dendrites required to memorize the information. When patterns of information are presented to the biological neural networ
information is converted to an electrochemical representation, This
representation is repeatedly presented to the network u
essary interconnections with other neurons are form
weighting along each of the interconnected dendrites is f
point, the neural network acts as a permanent storage
information it has learned, The memoryis simply a result of the structure of the network and the weights (permeabi~tyto potass
along each dendrite. When this new information is prese
network at a later date, the network can rapidly recall the
information.
ici
x = [x,
x2
* *
Xi]
for i inputs. Input neurons typically pass the value of the input directly to its output using a linear transfer function with a range of
[O,l]. Each interconnection between the input layer and the hidden
layer has a weight associated with it. The weights from input node i
to hidden node j follow the convention W,. The matrix describing
this information is
The input signal for each node in the hidden layer is computed
by summing the products of the output from each input layer and
the weight of the interconnections:
3
i
(5)
i=l
ere yk and are the actual and computed output values and k is
number of output nodes. If the sum square of error is greater than
some threshold value, the weights are adjusted and the process is
adjustment of each weight, 0, is computed by the backn algorithm according to
0 = 6E
+ G@-,
(7)
termine the se
~rmnesswas
~~~
Process monitoring
Trend analysis
Robotic control
Optimization
Applications
of patterns from large databases
Identification of trends in data
Capture of human decision making
Identification of patterns in images
Analysis of satellite imagery
Electrical signal analysis
Wear detection
Quality control
Food inspection
Stock, commodity, futures markets
Weather prediction
Weather patterns
Chemical reactions rates
Pest population levels
Electrical loads
Object identification
Learning motion
Harvest date
Pest treatment thresholds
Greenhouse environments
Neural N e t ~ o r ~ s
individual digital
single output node was defined, with output values of 0.9 representing a fertile egg and 0.1 an infertile egg. A threshold of 0.5 was applied to the network out ut to determine if an egg was infertile (0.1)
or fertile (0.9).
Murase et al. (1994) developed an NN to relate textural features
to the growth stage of lettuce grown in en~onmentallycontrolled
growth chambers for a space station. Their goal was to use an NN to
determine lettuce growth stage as a biofeedback mechanism for the
environmental control system. The three-layer NN used textural features from digital images as inputs including contrast, homo gene it^
and local homogeneity. The output was leaf size. They trained the
NN and found that it gave reasonably good estimates ofleaf size
compared to measured values.
Uhrig et al. (1992) developed an NN to predict corn yields in
the m i d w e s t e ~
United States. Theinputs were weekly m a ~ ~ and
u m
minimum temperature, soil moisture from 0-1 m and 1-2 m depths,
cumulative growing degree-days, and yield trend. The output was
corn yield (bushels per acre). The NN was trained using weekly inputs. Twenty-nine years (1960-19$9) of data from an Indiana cr
reporting district were used for training and validation. The auth
used 1990 data for validation. Their results were encouraging.
Ruan et al.(1994) developed an NN to predict rheological pro
erties of dough from the torque developed during mixing. The properties considered were farinograph, extensibility, and extensional flow
force. They trained and validated the NN using 62 different batches
of dough with variable flour/water ratio and percent protein. The
best NN gave relative errors less then 2.3%, Z.S%, and 7.8% for farinograph, extensibilit~ andextensional flow force, respectively.
Thai and Shewfelt (1990) U d a neural network to quantify the
color of tomatoes and peaches.
e, chroma, and value of lightness
were the inputs to the network. The network was trained using judgments from experts on the color quality of tomatoes and peaches. The
network was then tested against experts and statistical evaluation
method. The neural network gave good predictions of color quality
in both peaches and tomatoes.
atchelor
If inputs that highly influence the process being predicted are left
out, the NN has little chance of adequately predicting the process.
For instance, crop yield is highly influenced by weather information.
If a NN was developed to predict crop yield and weather information was not an input to the
NN, it would likely notperform well.
However, NN can discriminate between inputs andessentially ignore
inputs that have minimal influence on the process being predicted.
Neural network architecture must be determined, and development
usually results in a sensitivity analysis of error in prediction over a
range of architectural configurations. The steps involved in development are
1:
insect
McClendon and Batchelor (1995) developed a neural network to determine the economic treatment threshold when up to four different
insects are present in soybean fields. The insectsconsidered were velvetbean caterpillar (VBC), soybean looper (SL), corn earworm (CEW),
and southern green stinkbug (SCSB). Economictreatment thresholds
based on population levels and crop growth stage exist forindividual
insect populations; however, these thresholds cannot be used when
insect complexes are present in the field. Entomologists who have
extensive field experiencein relating damage from pest complexes to
yield reduction often must be called to make such complex decisions.
They use intuition about population numbers of insects present and
soybean development stage, age of insects present, and historic damage to make a treatment decision. McClendon and Batchelor pro-
~ e ~ r~aelt ~ o r k s
DefinetheProblem
The first step is to clearly define the problem to be solved. For this
problem, we want to develop an NN that can capture the
and decisions of the expert entomologist in making pesticide treatment recommendations for four primary insects. Staticthresholds exist for individual insects; however,when a combination of these pests
are present, an expert has to determine when a treatment should be
given. Thus, if only a single pest is present, the solution is easy to
determine, but underrealistic fieldconditions the standardguidelines
cannot be followed and an expert must be called to determine the
recommendation. The goal of this NN is to use the same information
that the expert considers in the decision process and map this information to a treatment decision.
2.
DeterminetheInputsandOutputs
An expert must consider many factors before making a recommendation. Interviews were conducted with the expert to determine what
information is typically used to make treatment recommendations.
The mosti ~ p o r t a nfactor
t
is the population levels of each pest. If the
populations are at a sufficient level, a treatment is always recommended, but if each pest is present at a low level, the combination
of estimated damage from each pest must be evaluated. Timing of
the infestation is also critical. If the pests are feeding on foliage, the
treatment threshold is higher than if they are feeding on pods and
seeds. Thus, plant age is a critical decision factor. Any historicdamage must also be considered in the decision. Thus,input nodes should
define the plant age, stage"of plant development, pest levels, and
historic damage. The following inputs were considered:
Input node
Input node 2:
Input node 3:
Input node 4:
Input node 5:
Input node 6:
Input node 7:
Input node 8:
Input node 9:
Plant age
Vegetative stage
Reproductive stage before seed expansion
Reproductive stage after seed expansion
VBC population level
L population level
CEW population level
SGSB population level
Historic defoliation
These input nodes reflect the inputs that are important to the expert.
owever, not all of the inputs may be i ~ p o r t a in
~ tpredicting the
ecisions of the expert. After the NN is eveloped, a sensiti~tyanalsis on the inputs can be performed by eliminat~gone or more inuts and retraining and validating the NN. If elimination of an input
does not change the NN results, then the input is not considere
ortant in dup~catingthe experts recommendations using the
The expert may make one of three recommendations: (1)treat
i m ~ e ~ i a t e l (2)
y , wait 3 days and check again, or (3) do not tr
Thus, we could have three output nodes that reflect these three
cisions.
Output node
Do not treat.
Output node 2: Wait 3 days and check the populations again.
Output node 3: Treatimmediately.
ose value
1.01
Collect
eural n e t w o r ~
requir
tterns of input and output information for
t ~ a i ~ and
n g va~dation
this example, we propose
isting of different insect population levels, crop
historic damage. The inputs for each scenario a
The expert then made a recommendation to either treat, wait 3 da s
d check the populations again, or not treat. These scenarios for
database the NN needed to determine how th
treatment reco~mendations.
4.
Select the
For this case study we focus on a three-layer feedforward architecture. One uncertainty related in this architecture is how to determine
the optimum number of hidden nodes required for a given set of
patterns. The optimum number of hidden nodes determines the number of interconnections in the network. The number of hidden nodes
is a function of the relationships to be learned. Too few hidden nodes
will not provide enough interconnections for the NN to learn the
r~lationships,while too many hidden nodes will often cause the NN
to memorize the training patterns and it will not be able to generalize
well for other scenarios. Since the memory of a network is stored in
the interconnections, it may seem to be an advantage to have as many
hidden nodes as possible. owever, it has been shown that too many
hidden nodes results in excellent recallof training scenarios but gives
poor predictions of validation scenarios because the NN has a suficient number of interconnections to memorize specific scenarios in
the training patterns. In contrast, too few hidden nodes can also decrease the accuracy of the network because there is not enough memory to completely define all the relationships between the input and
output patterns. Often, a sensitivity analysis on error in predicting
training and validation scenarios versus number of hidden nodes is
performed to determine the optimum number of hidden nodes for a
atc~elor
Low High
0
0
0
0
0
0
0
0
0
0
0
0
150
25
1
1
20
20
20
20
50
1
1
1
Comments
Days after planting
Number of nodes on mainstem
0 = no; 1 = yes
0 = no; 1 = yes
Actual population/ m
Actual population/m2
Actual population/ m
Actual population/m2
Percent defoliation
0 = no; 1 = yes
0 = no; 1 = yes
0 = no; 1 = yes
Selecting the proper learning rate and momentum is also a trial-anderror procedure, and these values can vary for different problems,
architectures, and training scenarios. There is no mathematical way
to determine the best values. Recall that both of these are dependent
on the shape of the error surface of the problem. For smooth error
surfaces, convergence is rapid and there is little error of converging
to a local minimum. Typically, it is recommended to first train the
network with a learning rate and momentum of 0.9 (fast). If the network appears to oscillate rather than converge, then both factors
should be reduced to 0.5 or lower.
7. Select Training and Validation Scenarios
Neural Networks
15
Train theNetwork
Once the network has been defined and the training and validation
scenarios have been selected, the network should be trained to learn
the relationships between inputs and outputs
of the training patterns.
Training time depends on the number of nodes, number of training
scenarios, learning rate and momentum, and the speed of the computer. Convergence criteria can often be specified
in terms of the
number of epochs or a minimum error between predicted and actual
output values. Training time can range from a few seconds to several
days, depending on the number of training patterns; input, hidden,
and output nodes; learning rate; and complexity of the relationships
being learned. Note that training a network is very time-intensive;
however, obtaining an output from a trained network requires only
a minimal amount of time.
9.
ValidatetheNetwork
17
allow the NN to begin training from several points on the error surface. It is possible that for some starting points the NN converges to
a local minimum rather than a global minimum. If the error in the
validation scenarios is different for differentinitial weight values, the
error surface is "bumpy," and the initial weights that give the lowest
error should be considered the optimum network.
Soybean rust
~~~~~~~~2~ Syd.) occurs in both the eastern
and western hemispheres and is a major disease of tropical and subtropical areas (Bromfield, 1984;Sinclair,1989;Tschantz,1984).The
disease can cause severepremature defoliation. It causes considerable
yield loss in many Asian countries, with losses as high as 40% reported in Japan (Bromfield, 1984) and up to 80% reported in Taiwan
(Yang et al., 1991b). Theimportance of the disease has increased owing to increased soybean production and expansion of the crop to
new regions, and the disease was recently found in Hawaii (Killgore
et al.,1994).Soybean rust has been considered a potential exotic
threat to soybean production of the United States (Kuchler et
al., 1984;
Sinclair, 1994). Epidemiolo~of the disease has been studied extensively over the past two decades (Yang et al., 1991a). Regression or
simulation models have been developed (Yang et al., 199la, 1991b)
to describe and predict the severity of soybean rust based on many
sources of data describing the environmental effects on population
dynamics and disease severity (Casey 1979; Marchetti etal.,1976;
Melching et al., 1979; Tschantz, 1984).Yang et al. found difficulties in
predicting disease development using traditional modeling techniques for some datasets for the purpose of risk assessment. In this
case study we discuss the development of an NN to predict the disease progress for soybean rust (Batchelor and Yang, 1995),
l.
through day 343 at weekly intervals with a total of46 planting dates.
In the 1981 experiment, planting started on December 21,1980 and
continued through day 264, for a total of 27 planting dates.
For each planting date, disease rating began as soon as rust was
first observed. Each plot was divided into four equal sections. Disease
severity, defined as percentage of diseased leaf area to total leaf area,
was recorded at weekly intervals, eight times per planting date experiment. A weather station was established in the field. Rainfall,
daily maximum and minimum temperature, relative humidity, and
average temperature were recorded at 2 h intervals. For detailed information, see Yang et al. (1990) and Tschanz (1984). This gavea total
of 577 observations of plant and disease conditions. For the purpose
of NN modeling, each observation serves as a pattern that can be
used for training, testing, or validation.
2.
The seven inputs shown in Table 3 were used for each of the neural
networks (Fig. 3). The output of each network was percent soybean
rust severity. The first fiveinputs were determined from the individual experiments and weather conditions. Degree-days of pathogen
(CDDR) and crop (GDDS) development were computed based on
5. CDDR
6. CDDS
7. CRH
Description
Planting date
Days to soybean maturity
First day that disease was observed
Crop age on sampling date
Cumulative degree days for rust development
Cumulative degree days for soybean development
Cumulative days that relative humidity exceeded 90%
Neural N e t ~ o r k s
19
PLD
DM
OSD
AGE
Disease
Severity
CDDR
CDDS
CRH
ValidatetheNeuralNetwork
severity was independent of other observations for the Same epidemic. Thus, the 577 scenarios were randomly divided into training
(327), testing (125), and validation (125) scenarios, without consideration of the epidemic or planting date that each scenario was associated with. The 327 training scenarios were then used to train the
NN, and the error between predicted and observed disease severity
was used to adjust the weights in the network. The testing scenarios,
a concept developed by Ward Systems Group (1993), were used to
determine when the network training was completed. Periodically
during the training process, the testing scenarios were presented to
the network and the errors between predicted and observed soybean
rust were computed for the testing scenarios. If the overall error was
lower than the previous error in the testing scenarios, the weights
were saved as the optimum values of the weights. This iterative process of presenting training and testing scenarios to the network continued until the network could no longer find new weights that minimized error in the testing scenarios. At this point it was assumed
that the network had converged and that the optimum weights that
minimized errors in the testing scenarios were found. At no time were
the testing scenarios used to adjust the weights; rather, they were
used to determine how well the weights computed during the training process predicted disease severity on an independent set of scenarios. This approach for training an NN is unique to the NeuroShell
software and is an excellent technique to avoid overtraining the network; also it gives a network that can generalize relationships better
than using training scenarios alone. Thetesting scenarios ensure that
the network is trained to give the best predictions on scenarios inendent of the scenarios used for weight adjustments.
Once the network was trained, the validation scenarios were
presented to the network and the error in prediction (R) was comted for each network. The number of hidden nodes was altered
er a range of 10-48 to determine the best a~chitect~re
for the network that minimized the error in the va~dationscenarios.
In the second validation effort, it was assumed that scenarios
* ( I
each of the 73 epidemics, or planting date S
ndent but that scenarios in differ
dent
with
respect to other pla
dies.
The
disemics or plantings were rando
ded so that 51 (70%)
for training, l1 (15%)were
testing, and the remaining 11 (15%)were used to validate the NN. The scenarios within
each epidemic were then classified as traini
testing,or validation
scenarios based on the classification of the
In the first ~alidationof the NN, the 577 scenarios were randomly
divided into 327 training, 125 testing, and 125 validation scenarios.
Errors in predicting soybean rust severity are shown in Table 4. The
best network had 14 hidden nodes and gave an R' value of 0.925 for
the validation scenarios. The average error in predicting disease severity was 6.9% for each of the validation scenarios, and the largest
0.971
0.971
0.982
0.966
0.967
0.968
0.975
0.960
0.978
0.974
0.970
0.973
0.980
0.976
0.966
0.980
0.975
0.977
0.975
0.976
0.859
0.917
0.912
0.919
0,913
0.908
0.916
0.903
0.911
0.923
0.897
0.910
0.920
0.921
0.898
0.915
0.902
0.912
0.916
0.923
0.884
0.921
0.925
0.913
0.915
0.917
0.912
0.912
0.905
0.922
0.915
0.911
0.896
0.912
0.920
0.910
0.913
0.909
0.911
0.895
152
97
98
113
111
108
115
116
124
102
111
117
136
110
104
139
114
119
116
137
8.4
7.9
6.9
7.5
7.4
7.3
7.4
7.4
7.4
7.1
7.5
7.5
7.7
7.3
7.0
7.8
7.3
7.5
7.3
7.7
43.4
37.0
33.8
31.2
34.3
37.0
34.5
34.9
38.2
37.2
32.6
34.0
38.0
37.
36.6
49.6
37.7
36.0
41.2
52.3
622
atchelor
E'i
2
2
60
50
40
10
0
20
40
60
80
100
eural ~ ~ t w o r k s
623
ical system. Thus,the NN was able to predict disease severity for the
training patterns very well because it was trained using other observations within the same set of planting date studies.
In the second validation effort, it was assumed that scenarios in
each planting date study were not independent but scenarios in different planting date studies were independent. The planting date
studies were divided into 70% training, 15% testing, and 15% validation scenarios.Thethree-layer
NN was trained using different
numbers of hidden nodes, and the results are shown in Table 5. The
overall R' for eachNN decreased compared to the previous validation
effort, with R2 values ranging from 0,625 to 0.754. The best architecture had30 hidden nodes, with X2 = 0.754. This NN gave a maximum
absolute error of '71.276, but the mean absolute error was 11.5% over
all the validation scenarios.
In the second validation effort, it was assumed that scenarios
within individual planting date studies were not independent of each
other but that scenarios in different planting date studies were independent. This assumption decreased overall NN accuracy in predicting disease severity in the validation scenarios. This assumption
is likely more realistic than assuming that all scenarios are independent. In view of disease forecasting, this validation is equivalent to
projecting disease progress curves without referring to previous disease occurrence within a planting.
To use this neural network to predict severity of rust on a soybean crop, the user would have to supply the seven inputs listed in
Table 3. Let's assume that the prediction is to be made when the crop
R2
14.3 0.668
14.5 0.689
0.754
13.2 0.698
14.1 0.689
15.4 0.625
Mean
square
error
421
420
332
408
421
507
Mean
absolute
error (%)
Max.
absolute
error
(%)
"
11.5
71.2
76.8
is 50 days old. The user must first make some assumptions about
future weather data, namely temperature and number of days that
relative humidity exceeds 90%. This information can begenerated for
the remainder of the season using many different techniques, for ex* g weather generators or long-term weather forecasts.
then compute the cumulative degree-days for soybean
st (GDDR) development at a desired future date as
well as the number of days that relative humidity exceeds
The
st date that disease was observed should also be known. This inrmation can then be input into the NN, and the predicted soybean
rust severity will be computed on the date that corresponds to the
crop age entered by the user.
~ e u r a ~l e t w ~ r ~ s
627
Melching, J. S., K. R. Bromfield, and C. H. Kingsolver. 1979. Infection, colonization, and uredospore production on Wayne soybean by four cultivars
of P ~ ~ o p s opachyrhizi,
r~
the cause of soybean rust. PhytopathoZo~
1262-1265.
Murase, H., Y. Nishiura, and N. Honami. 1994. Textural features/neural network for plant growth monitoring. ASAE Paper No. 94-4416. St. Joseph,
MI: ASAE.
Parmar, R. S., R.W. McClendon, G. Hoogenboom, I? D. Blankenship, R. J.
Cole, and J. W. Dorner. 1994. Prediction of aflatoxin contamination in preharvest peanuts. ASAE Paper No. 94-3562. St. Joseph, MI: ASAE.
Priest, J. B,, J. McLaughlin, and R. R. Price. 1994. Neural network model of
naturally ventilated broiler house. ASAE Paper No. 94-4014. St. Joseph,
MI:ASAE.
Rigney M.P., and G. A. Kranzler. 1989. Seedling classification performance
of a neural network. ASAE Paper No. 89-7523. St. Joseph, MI: ASAE.
Ruan, R., S. Almaer, and J. Zhang. 1994. Prediction of dough rheological
properties using neural networks. ASAE Paper No. 94-6523. St.Joseph, MI:
ASAE.
te~
Rumelhart, D. E., and J. L. McClelland. 1986. ParaZZeZ D i s f r i ~ ~Processing:
~~plorations inthe ~ i c r o s t r ~ c t u rofe Cognition. Vol.
Foun~ations.Cambridge, MA: MIT Press.
Seginer,I., and A. Sher.1992. Neural-nets for greenhouse climate control.
ASAE Paper No. 927013. St. Joseph, MI: ASAE.
Sinclair, J. B. 1989. Threats to soybean production in the tropics: Red leaf
blotch and leaf rust. Plant Dis. 73504-606.
Sinclair, J. B. 1994.
Thai, C. N., and R. L. Shewfelt. 1990. Modeling sensory color quality: Neural
networks vs. statistical regression. ASAE Paper No. 90-6038.St. Joseph,
MI:ASAE.
Thai, C. N., A. V. A. Resurreccion, G. G. Dull, and D. A. Smittle. 1990. Modeling consumer preference with neural networks. ASAE Paper No. 90-7550.
St. Joseph, MI: ASAE.
Tschanz, A. T. 1984. Soybean rust epidemiology: Final Report. Shanhua, Taiwan: Asian Vegetable Research and Development Center.
Uhrig, J. W.,B. A. Engel, and W. L. Baker, 1992. An application of neural
networks: Predicting corn yields. Proceedingsof the 4th International Conference on Computers in Agricultural Extension Programs, Lake Buena
Vista, FL.
WardSystems Group. 1993. ~e~rosheZZ
2. Frederick, MD: WardSystems
Group, Inc.
Wilkerson, G. G., J. W. Jones, K. J. Boote, and J. W. Mishoe. 1985. SOYGRO
V5.0: Soybean Crop Growth and Yield Model. University of Florida.
Yang, X. B.,M. H. Royer,A.T. Tschanz, and B. Y. Tsia. 1990. Analysis and
quantification of soybean rust epidemics from seventy-three sequential
planting experiments. PhytopathoZo~80:1421-1427.
Yang, X. B., W. M. Dowler, and M. H. Royer. 1991a. Assessing the risk and
potential impact of an exotic plant disease. Plant Dis. 75976-982,
Yang, X. B., W. M. Dowler, and A. T. Tschanz. 1991b.A simulation model for
assessing soybean rust epidemics. J. P ~ y t u ~ ~ t233~187-200.
~ul.
Zhuang, X., B.A. Engel, and M. Benady 1992. Locating melons using artificial neural networks. ASAE Paper No. 92-7014. St. Joseph, MI: ASAE.
630
Bolte
4. Implement the simulation model in a computer representation to execute and test the models predictive capabilities.
Simulation models,then, become a vehicle forrepresenting in abstract
terms a slice of reality consistent with our specific interests in enhancing our understanding of our world. Representation is the key
word here: Simulation models must provide us with an abstract representation of reality. How we represent the conceptual model we
derive from observation of reality thereforeshould be of considerable
interest to the modeler. Itis here that various programming languages
and simulation tools have a central role to play.
Computer methods for the solution of simulation models had
their genesis with the introduction of the Fortran programming language in the 1950s. At that time, Fortran provided a powerful new
tool to allow scientists and engineers to solve complex mathematical
relationships. Both hardware and software tools for conducting simulations have continued to evolve since that time. Eventually computing languages were designed specifically for simulation. As the
needs of the simulation community grew more complex, new simulation languages were developed for different types of models. These
languages provided direct support for higher level representation of
fundamental systems concepts and typically provided automated
methods for the solution of complex equations. The General Purpose
Systems Simulator (GPSS) was one of the first examples of such a
specialized language. Designed by G. Grudden for IBM, GPSS is a
very general language used primarily for queuing problems and inventory control. GPSS/ PG and GPSS/ H are two modern derivatives
ofGPSS (Gordon, 1969; IBM, 1970; Schriber, 1972). Simscript, developed by H. M. Markowitz at the RAND Corporation and then at
CACI (Kviat et al., 1983; Caci 1988) was the next major simulation
language. Simscript is a general-purpose simulation language used
primarily to program discrete-event simulations but also allowing for
a process-oriented approach.
Several contributions have been made to the advancement of
simulation by A. Alan B. Pritsker. The firstsimulation language written by Pritsker was General All-Purpose Simulation Program (GASP).
GASP had its origins at U.S. Steel, and Pritsker improved it in 1969.
Pritsker and P. J. Kviat documented GASP 11, and then Pritsker produced the latest version, GASP IV (Pritsker, 1974). Itis largely based
on Fortran and requires the user to write a series of subroutines to
describe the simulation. Pritskers next project was the development
of Simulation Language for Alternative Modeling (SLAM),which was
~ j e c t - ~ r i e n~
t e r~ o ~ r a ~ ~ i n ~
631
spond toentities in the real world. Because these objects mirror reality
at both the conceptual and implementation levels, object-based software tends to be easier to design and maintain than conventional
software. This high level support for sophisticated representation,
coupled with additional supporting language features, has moved
OOP into the forefront of current programming paradigms.
Simula, the first object-orientedp r o ~ a m m i n language,
g
was developed in 1967 explicitly to provide simulation facilities for modeling real-world objects (Dah1 and Nygaard, 1966,1967a, 196713;
keud, 1989). Simulawas not only the first simulation language, it was
also the first object-orientedlanguage. Simula introduced the concept
of a class as the fundamental representational paradigm. Although it
was never a commercial success, Simula
developed a strong academic
following in Europe and had great influence on programming language design. In addition to basic language features, Simulaprovides
simulation for support through two key classes, SimSet and Simulation (Franta, 1977). The SimSet classprovides list processing by implementing a double-linked list. The Simulation class provides sche
uling, synchroni~ation, anda simulated time flow mechanism.
There have since been many implementations of object-oriented
simulation languages. DEMOS (Birtwistle, 1979) consists of approximately 30,000 lines of Simula code designed to model queuing netISCO (Helsgaun, 1980) is a further extension of
provides classes to assist in combined simulation and demonstrated
the ease with which OOP languages could be extended. The Smalltalk
"bluebook" (Goldberg and Robson,1983) explains the simulatio~
constructs that were implemented in Smalltalk-80,a widely use
object-based language. The
design of Smalltalk was strongly influence
by Simula. M ~ D I MXI, based on Modula-2, is an object-oriented
pose programming language that provides direct S
d discrete-event simulation (Belanger et al., 1990).
classes was introduced in 1980 (Stroustrup, 1983), and C++ followed
ortly thereafter. C+ + was originally developed by Strous
ovide support for event-driven simulations similar to those
in Simula but with the efficiency of code generation in C. A s i m i ~ a ~
language, Objective-C, was d
has been less
successful in
the commercial marketplace.
ascal p r o ~ a m m i n g
number of general software development efforts.
Of all these languages, C+ + has
and is currently the most widely use
opularity has been attributed to the fac
su
~ n ~ u ~ s u l u tisi othe
n mechanism for in
ation hiding and is one
of the most important properties of objectyer,1988;
Budd,1991).
Each
object hides its
own informa
never that information
is needed it can only be obtain
eans of the methods of the
object. In this way, object data can be protected, and unwant
ification due to software errors is prevented. Encapsulation
the possibi~tyof corrupting an objects data. The use of encapsulation
therefore improves the modular it^ r e ~ a b i ~ tand
y , maintainabi~tyof
system code de et al., 1993).
Messuge ~ ~ s s i nisga necessary result of enca~sulation.omm munication between objects is ssible only t ~ o u g hmessages.
action in an object-oriented
gram is initiated by objectsse
messages to one another. The method executed by an object m reetermined by the class of the receiver. All
use the same method in response to a given
message (~umbaughet al., 1991), operating on each specific instances data. A message states what should be done by an object,
whereas a method expresses how it will be done. A message is
34
Bolte
"bound" to a method either statically (at compile time) or dynamically (at run time) (Ege, 1992).
~ y n a m i cbinding is the ability to determine what method to call
at run time rather than resolving the call at compile time. Frequently,
the specific receiver classof the message will not be h o w n until run
time, so which method will be invoked cannot be determined until
then (Ege, 1992). Thus we say there is a dynamic or late binding between the message and the method used to respond to the message,
in contrast to early binding in traditional procedural languages such
as Fortran or C. This feature allows referencesto generic objects during compilation, with the run-time behavior of the system determined
by the actual (subclassed) objects that are active in the system at the
time. Dynamic binding is an extremely powerful tool for defining at
a generic level the specific types of behavior a class of objects must
provide. Dynamic binding provides an excellent mechanism for defining system-level interfaces forsimulation objects, with specific behaviors appropriate to each subclass invoked during program execution.
~ ~ l y is m
the ability
~ ~ to ~have~ different
~ s kinds
~
of objects respond to the same message in different ways. It is based on dynamic
b i ~ d i n g(Budd, 1991). For example, animal and plant simulation objects both might have the method Grow( ), but each one responds to
the Grow( ) message differently. Polymorphism allows this shared
code to betailored to fit the specific circumstancesof each individual
data type (Ege, 1992).
~n~eritance
denotes the ability of an object to derive its data and
behavior automatically from another object(Budd,1991).When
a
new class is defined as a subclass of an existing one, data members
and methods are automatically inherited from the parent class (superclass). This is based on creating new classes of objects as descendants of existing ones, extending, reducing, or otherwise modifymg
their functionality* Classes become
organized as a hierarchy, with subclasses inheriting attributes from a superclass higher in the tree. Inheritance is a powerful concept that can greatly enhance and simplify
the design and implementation of a system by capturing commonalities between similar components (Rumbaugh et al.,1991; Rubin
and Goldberg, 1992; Budd, 1991). It is an extremely valuable tool for
constructing and reusing complex simulation components, since the
inheritance hierarchy defines various levels of informatio~ andbehavioral abstraction appropriate to the system being represented.
Coupled with dynamic binding, inheritance allows the development
of abstract frameworks for model construction defined at a high level
b j e c t - ~ r i e n t e~rogramming
~
35
Simulation is one of the natural applications of object-oriented programming. From an object-oriented viewpoint, simulation models are
readily envisioned as an assemblage of interacting objects representing specific entities in the real world (Zeigler, 1990). Although the
mechanisms of OOP can be implemented with a traditional programming language supportingdata abstraction, better results can be
achieved with an object-oriented programming language (Bischak,
1991) that directly incorporates object orientation into the language
specification. The object-basedapproach provides powerful tools for
the representation of complex domains and results in very efficient
programming environments (Rumbaugh et al., 1991). OOPS features
and benefits are well documented (Booch, 1991).
The limitations of procedural approaches in simulation are well
established. One of the most significant problems of traditional simulation modeling has been the lack of reusability of the simulation
models and/or components (Rumbaugh et al., 1991). This problem
stems from the fact that most simulation models are developed for a
specific simulation experiment objective. When anew objective is encountered, a new model is developed from scratch even though it
may include elements contained in earlier models (Rumbaugh et al.,
1991; Sierra and Pulido, 1991). Inheritance, encapsulation, and dy-
r (1992) reported a
ecarnemore appar
63
olte
system increased. This has been consistently noted in domains outside the realm of agriculture as well.
639
are typically updated at a constant interval, while stochastic simulation objects may be updated at intervals established by their statistical
properties. The priority of a simulation object establishes its ordering
sequence in the simulation event list when two objects are being upated at the same time.
It is sometim~suseful to have simulation objects store data. Because generalized data storage is a requirement of many systems, the
framework defines a Datastore class that provides data storage caatastore is a repository for data. By ~eneralizingthe
r data storage into a specific class, any object can use
that class to handle its datastorage requirements. Thus, the datastorage object definition becomes a standard way of sharing and manipulating data. A SimQbj can acquire data storage capabilities by subclassing from a Datastore. Further, by extending the definition of a
atastore, any object that derives from Datastore automatically gets
access to its extended capabilities. Thesemight include statistics generation, file I/ 0, or visualization.
statistics
on the event list, updates thecurrent time to the event time, and then
updates the event. When the system is updated in theevent routine,
new events can be scheduled and added tothe event list. The process
repeats by removing the next object at the top of the event list, resetting the clock, and then u p d a t ~ gthe object. The procedure of a
vancing the simulation clock continues until no more items are on
the event list or a simulation stopping condition ends the simulation.
'Variable time increment simulations therefore jump through time
from one event to the next, and the state of the system is U dat
each event. This approach facilitates the coexistence of fixe
Bolte
/ / constants
const float a=0.3, b=0.2, rH=O.l, rP=0.3;
/ / parent Population class- derived from SimObj
class Population : public SimObj {
protected:
doublepop;
/ / populationsize(statevariable)
public:
/ / public "set" method for population
void SetPopulation( double p) { pop = p;
l;
* timestep;
1;
/ / declare globalvardables
Predator *pred;
/ / pointer to an instance
of a Predator
*prey;
Prey
/ / pointer to an instance
of a Prey
/ / main simulation program
int main() {
pred = new Predator;
/ / create an instance of a Predator
prey = new Prey;
/ / create an instance of a Prey
);
/ / Set initial conditions
pred->SetPopulation(
prey->SetPopulation ( 50 ) ;
Si.mEnv imEnv;
/ / create a simulation environment
simEnv.RegisterSimObj( pred ) ; / / tell it what to simulate
simEnv.RegisterSimObj( prey ) ;
simEnv.SetStopTime(
simEnv.
I;
Run
() ;
);
riented ~ r o g r a ~ ~ i n g
variable-step simulation objects, allowing the simulator to use whichever approach is most appropriate for each object.
SimEnv and associated SimObj-derived classes
provide the basis
requirements for conducting a simulation. A simplifiedexample
based on a two-population Lotka-Volterra predator-prey simulation
is implemented in Fig. 2. This simulation shows the creation of two
objects (Predator and Prey) and a simulation environment (SimEnv).
The virtual Update( ) method, implemented for each simulation object and called as necessary by the simulation environment, steps each
object through a single time step. This basic simulation system can
be readily extended by defining additional classes and methods to
support more sophisticated simulation capabilities, as described next.
Communication among objects is a required component of any objectbased system. Individual simulation objects send messages to other
objects to implement some actionor get some information. In a robust
simulation environment, message passing should be implemented at
several levels to provide either direct or indirect communication capabilities. In Fig. 3, object 1 sends a message to object 2. Object 2
takes some action in response to the message being sent and replies
ent
to the sender of the message. This is a form of o ~ ~ e c t - d ~ e n dcommunication becauseobject l must have somereference to (direct
knowledge of) object 2 to send it a message.
During the development of the framework, it was found that
object-independent communication (i.e., a protocol available so that
a simulation object can send messages to other objects in the system
without requiring a direct reference to them) was needed. There are
situations where the object sender may not know who the object receivers are. In Fig. 4, the "ice cream man" simulation object broadcasts to anyone within hearing range that it has ice creamto sell, with
no direct knowledge of who might be listening or how they may
respond. Anyone interested will receive the message and respond
accordingly. What is needed in this situation is an o~ject-ind~endent
communication mechanism.
The implementation of such a scheme is relatively straightforward. First, individual objects tell the simulation environment what
messages they are interested in. An object accomplishes this by registering a n o t ~ c ~ t ifilter
o n with the simulation environment, indicating
that it wants to be notified whenever a particular message is broadcast from one or more simulation objects. Second, objectsmust broad-
Dependent object ~ o ~ ~ u ~ i c a t i o n .
delivery.Broadcastmessa
are sent "now" (in simu
are then discarded from
system, whereas blackboa
remain on the blackboar
definitely until explicitly
amount of time between when a essage is posted to th
and when it is read is unspecifie
~imulationobjects c o m m ~ ate with other simu
st / Notify and/ or blackboard
sting to the blackboard are
es that allow for globalor obje
unication without i n t r o d ~ ~ i nobject
g
dependencies.
~e~resen~ation
is a central concept for both decision and si~ulation
systems, and modelers are generally well aware of issues of re
sentation in simulation. The objectparadigm provides an intrinsically
roach to re~~esenting
conventional simulation objects.
nt, however, is the ability to support ad
resentation paradigms, partic~larlyin the area of
'stic systems stress a separat
representation capabilities
encaps~lationusing an object-orie
In this context, a potentially
ual component of a
decision system is an agent, an intelligent entity that might
the system with diagnostic, control, or planning capabilities,
source management context, agents inthe system might
knowledge about the resources in question, using inform
rived from models and datasets to make suggestions rega
propriate mana~ementtactics under different circumstances. Agents
may beuseful for guiding the trajectory of complex systems and may
know how to answer questions and solve problems for other simu-
olte
lation objects. Theymay assist other simulation objects during a simulation run.
Agent capabilities were added tothe framework by defining an
Agent class,descending from SimObj, that can store and execute d e s
via an inference engine. Since agents are just another simulation object, albeit with somewhat different capabilities,they are added to the
decision framework just like any other simulation object. How agents
participate in a simulation is dependent on how quickly they are
required to solve some problem. For example,consider the case of a
simulation of a salmon hatchery facility. A salmon hatchery simulation is running, and one of the simulation objects (e.g., fish lot) has
reached a minimum oxygen threshold value for sustaining growth.
We might like an agent to interact with the simulation to monitor
such problems and take corrective action. Several
approaches are possible, depending on the requirements of the simulation. If the agents
should immediately respond to any problems (rarely the case in
the real world), the fish lot would either talk directly to the agent
about the problem (object-dependent communication) or broadcast a
message that the agent had previously been asked to listen for (objectindependent communication). Alternatively, the intent of the simulation may be to explore the effects of scheduling and time management on the agents availability for fixing problems. Inthis case, the
agent would be assigned a time step proportional to its availability
for action. Upon receiving an update message from the simulation
environment, the agent would either directly observe each object it
was monitoring (object-dependentcommunication) or look for problems previously posted on the blackboard that have not been resolved
(object-independent communication). In either case, it would invoke
its knowledge base for the particular problem, presenting options to
the user or taking steps to correct the problem.
As an example of a decision system based on the preceding integrative framework, we have developed a decision system for aquaculture production facilities management, where a facility consists of
multiple ponds, fish lots, and various supporting resources. At the
lowest level, the system relies on simulation models to make predictions about the state of the facility through time. Individual components of the system are dynamically created (instantiated) from object
templates reflecting the item in question and include fish ponds, fish
lots, pumps, water conditioning units, and other entities important
in the operation of actual facilities.
The need for tools to synthesize and integrate knowledge sources for
agriculture and resource management continues to increase. Fortunately, technologies for dealing with the problems encountered are
becoming increasingly available. Manyaspects of representation and
analysis in these management systems can be abstracted to
system-level services and provided through integrative frameworks.
Similarly, interfacesbetween the components of these systems can be
defined on an abstract basis using object-oriented tools. These tools
are extremely valuable for constructing and reusing complex decision
support components, providing mechanisms to define various levels
of information and behavioral abstraction appropriate to the system
being represented. They allow the development of abstract frame-
Bolte
Whittaker, A. D., M. L. Wolfe, R. Godbole, and G. J. V. Alem. 1991. Objectoriented modeling of hydrologic process. AI AppE. 5(4):49-58.
Wilde, N., P. Matthews, and R. H. Hutt. 1993. Maintaining object-oriented
software. IEEE Software, January, pp. 75-79.
Zeigler, B. P. 1990. O~ject-Orie~ted Si~ulatio~s with ~ierarchical, Modula~ Mod
els. San Diego, CA: Academic.
Jones
ath he ma tical simulation of crop growth and yield processeswas initiated about 30 years ago (de Wit, 1965; de Wit et al., 1978; Duncan
et al., 1967; Keulen, 1975). Thesepioneering modelers provided a systems framework for modeling carbon balance, water balance, and
energy balance in a complex crop production system. Thisis basically
the approach that has been followed by most crop modelers since
that time, Early crop modeling efforts were initially limited by the
availability and capacity of computers, by inadequate understanding
of physiolo~calgrowth processes, and by lack of input/output standards andaccess to common file accessand graphics tools. Early crop
models also focusedprimarily on potential production, often not considering water, nutrient, and pest limitations. As a result, there was
not much realismin the simulated outcomes. Inthe intervening years,
computer technologies have advanced manyfold and considerable
651
oote et al.
65
For many crop plants, the partitioning of dry matter to growth in the
different plant parts depends on the state of development. To accurately predict the growth and yield of these crops, one must be able
to accurately predict the timing and duration of the various crop
growth phases and the total life-cycle duration.
Crop development in CROPGRO during various growth phases
is d~erentiallysensitive to temperature, photoperiod, water deficit,
and N stresses. There is a vegetative development rate that affects
the rate of leaf appearance. In addition there are 13 possible life-cycle
phases from sowing to maturity, each having its own unique developmental accumulator starting at a unique prior endpoint growth
stage and ending when the accumulator reaches a defined threshold
when an event such as first flower or onset of seed growth occurs
(Fig. l).The physiological development rate, expressed as physiological days per calendar day (PD/ day), during any phase is typically a
function of temperature (T), photoperiod (P), and water deficit:
655
Sowing
Emergence
End Juven.
~
V1 stage
End MS
End Leaf
Harvest
Maturity
TE~PE~TURE,
IGURE 2 Effect of temperature on rate of vegetative node appearance, reproductive development rate before beginning seed stage, and reproductive
development rate after beginning seed for soybeans.
the cultivar files. Figure 1 illustrates the time line for the 13 phases
used for the three grain legumes-dry beans, peanuts, and soybeans.
Note that several phases have a common starting point, such as time
from first flowerto first pod, first flower to first seed, and first flower
to end of leaf expansion.
CROPGRO's flexibility to use different temperature functions
before and after flowering was found to be very helpful for correctly
predicting the reproductive phenology of soybeans. Using simplex
optimization techniques with extensive data on time to flower, beginning seed, and physiological maturity, we discovered that soybean
cardinal temperatures for reproductive growth differ forpreflowering
versus postflowering phases (particularly after beginning seed) and
that both of these processes have cardinal temperatures different from
those for rate of vegetative node
stage) development (Fig. 2)
(Grimm et al., 1993, 1994).
stage development has a higher base
and higher optimum temperature than does rate of progress toward
flowering. Of particular significance, the rate of development after
beginning seed had a much lower derived base temperature
Of course, soybeans do not survive freezing temperature, but their
rate of progress from beginning seed (R5) to physiological maturity
(R7) is scarcely slowed by cool temperature. This change (use of different temperature functions before and after seedset) greatly improved our ability to predict maturity dates of soybean cultivars
grown farther north in the United States or in cool conditions. The
flexibility of using different temperature and photoperiod sensitivities
during various life-cycle phases has been helpful in modeling a number of species with the same basic model.
c,
where
S,, SRIand
are abscised parts of leaf (L), stem (S), root
(R), and shell (SH), respectively; and
=
E(P,
X,)
dI4.JS- x,?&+dt
"
M,
c, - P,
oote et al.
The CROPGRO model has two options for photosynthesis calculations: (l)daily canopy photosynthesis or (2) hourly leaf-level photosynthesis with hedgerow light interception. The daily canopy photosynthesis option is similar to, but modified from, the method used
in SOYGRO V5.42. The leaf-levelphotosynthesis option, with hedgerow light interception, allows more mechanistic responses to temper-
~ i ~ u l a tof
i oCrop
~ Growth
659
rates by their respective LAIs. ourly canopy assimilationvalues are summed to give the total daily gross photosynthesis.
Leaf photosynthesis rate of sunlit and shaded leaves is computed with the asymptotic exponential light response equation, where
quantum efficiency (QE)and light-saturated photos~thesisrate (Pm,,)
variables are dependent on CO, oxygen, and temperature. The Farquhar and von Caemmerer (1982) equations for the RuBP regeneration limited region are used to model the basic h e t i c s of the rubisco
~n~ym
and
e tocompute the efficiency of electron use for CO, fixation.
S includes temperature effects on specificity factor of mbisco an
pensation point in the absence of mitochondrial respirate and Pickering (1994) and Pickering et al. (1995) for a
on of simplified algorithms for these computations].
Single-leaf P,, is modeled as a linear function of specific leaf weight
(SLW) and as a quadratic function of leaf N concentration. The SLW
and leaf N concentration (and thus P,,) are modeled to vary with
LA1 depth in the canopy. Leaf QEis modeled to have a mild dependence on leaf N concentration. Leaf carbohydrates are excluded from
calculations ofSLW and N concentration used to influence photogross canopy photosynthesis response to solar irradiance,
e,leaf N concentration, and LA1 are illustrated in Figs.
3-6, using the leaf-level assimilation compared to the daily canopy
ion option. Responses are shown for soybeanparameters for
canopy at 65 days after planting. Response to daily solar
irradiance shows a gradual saturation of response beginning at '20
/(m2.day), particularly so for the hourly version (Fig.3).The
eater degree of saturation of the hourly version is associated with
for computing fraction diffuse, the approach used for
absorption, and the fact that thesolar constant re
the day of simulation even though the "input'
ance was varied (this does not happen in realit ). Irradiance
mmer months is typically between 14
Simulated daily canopy photosynthesis
rature optimum for both the ho
the hourly version, the predicted
~ a x l ~ uformdaytime mean temperat
The leaf-levelhourly version gives a
cause leaf ~hotosynthesisdepends o
e~atingin opposite directions, QEd
P,, increases with incre
s s i ~ l ~ t i response
on
to temperature for the ]DAILY
LEAF option, using soybean parameters and con
S i ~ u l a t e dwith 12C spread between. ma^^^
imum t e ~ ~ e r a t uand
r e plotted versus mean daytime
oote et al.
662
also because many leaves in the canopy operate at low light flux,
which increases the importance of the temperature effect on QE.The
daily version temperature parameters were initially taken from SOYGRO V5.42, based on relative growth rate data of Hofstra and Hesketh (1975),but we decreased the base temperature from 5 to 3C and
decreased the first (lowest) optimum temperature from 24 to 22"C,
based on experience with SOYGROV5.42 in the northern United
States and in France. The hourly photosynthesis version has higher
rates at high air temperature than does the daily version; however,
this is supported by soybean measurements of Pan et al. (1994)taken
in controlled environment studies. For leaf photosynthesis, the base
temperature of 8C and linear response up to 40C for rate of electron
transport were derived from Harley et al. (1985).
Canopy assimilation response to leaf N concentration shows initially rapid response from 18 to 30 mg N/g, with a gradually decreasing response at higher N concentration, reaching a saturating
response above 52 mg N / g (Fig. 5). Canopy assimilation response
to leaf N is expected to saturate more readily than single-leaf lightsaturated photosynthesis because quantum efficiency is less sensitive
to leaf N concentration, The leaf-level option has relatively higher
rates than the daily option, especially at low leaf N concentration, a
phenomenon that we believe may be caused by the modeled vertical
distribution of leaf N (and SLW) about the canopy mean.
g-I
63
20
5;
5!
10
zt:
LEAF
INDEX
FIGURE 6 Daily assimilation response to leaf area index for the DAILY option and the HOURLY-LEAF option at a closed canopy stage,using soybean
parameters and conditions described in Fig. 3.
Growth and maintenance respiration are handled similarly to SOYGRO V5.42 (Jones et al., 1989; Wilkerson et al., 1983) and PNUTGRO
Vl.01 (Boote et al., 1986). Maintenance respiration depends on temperature, gross photosynthesis rate, and current crop biomass (minus
oil and protein stored in the seed). The relative sensitivity of maintenance respiration is similar to the Qlo response (1.85) reported by
McCree (1974).
Growth respiration and efficiency of conversion of glucose to
plant tissue are computed using the approach of Penning de Vries
Boote et al.
664
and van Laar (1982,pp. 123-125) and Penning de Vries et al. (1974).
This approach requires estimates of tissue composition in six types
of compounds: protein, lipid, lignin, carbohydrate-cellulose, organic
acids, and minerals [summarized for peanuts by Boote et al. (1986)
and for soybeans by Willcerson et al. (1983)and Jones et al. (1989)].
The glucose equivalent needed for biosynthesis depends on
chemical composition of tissue and biochemical pathways of synthesis. The approach accounts for
1. Glucose respired to provide ATE NADH, and NADPH for
biosynthesis
2. Condensation and reduction and increased energy content
of tissue
Table 1 shows the glucose cost of energy for synthesis versus the
glucose equivalent that remains condensed in the product. Note that
lipid synthesis is very costly (3.11g glu/g product), in terms of both
energy required and energy stored in the lipid. Proteins are also
costly, particularly if N is assimilated via NO, or Nz fixation. Even
though the cost of protein from NH, appears to be less, we compute
all costs of inorganic N assimilation as if originating from NO, because nitrification rapidly converts most of the NH, to NO, in most
cost
energy
C in
product
Total
cost
Conversion
efficiency, E
(g prod/g gl4
0.36
1.22
0.11
1.34
1.34
1.13
1.70
2.56
1.24
0.56
0.39
0.81
1.17
0.62
0.00
0.05
1.94
1.55
0.929
0.00
3.11
2.17
0.929
0.050
0.32
0.46
1.08
665
soils and because NK-fed plants have an uncertain added cost associated with cation balance. Proteins constructed with N from N2
fixation cost 2.83 g glu/g protein, based on theoretical calculations
for a hydrogenase-negative Nz-fixing soybean (K. J. Boote, unpublished calculations).
With these glucose-equivalent costs to synthesize six classes of
compounds and knowledge of the approximate composition of the
tissue, one can compute the total glucose-equivalent cost to produce
1 g of tissue (Penning de Vries and van Laar, 1982, p. 123):
E.
The partitioning algorithms in CROPGRO are similar to those of SOYGRO V5.42, PNUTGRO V1.02, and BEANGRO 1.0, except that N deficit can limit vegetative and reproductive growth and alter partitioning between root and shoot (described in Section 1II.F). New dry
weight growth of each plant component (Wi)
depends on the computed partitioning factor to each tissue type (XJ,
conversion efficiency
(Ei),gross photosynthesis, and maintenance respiration. The partitioning factor varies with crop developmental stage and addition of
reproductive sites.
WT = XiEi(P, - R,)
1,
_..,
,,
,",.,.,.,,
'.,,._,,_..,
..
",...
..
. .
..
...........,
,.
"
.....
..
oote et al.
\"""".
0.0
10
NDLEAF
V E ~ ~ A T I VSTAGE
E
FIGURE7 Partitioning of dry matter among leaf, stem, and root components
as a function of vegetative growth stage of soybeans. Taken from soybean
species file. Whenbeginning bloom stage occurs (whether at low or high V
stage), partitioning begins a gradual transition to the values at NDLEAF
(date when the last branch or main axis leaves complete their expansion).
and root being dependent onV stage progression (Fig. and further
m o d ~ e by
d water deficit and N deficit. As reproductive development
pro~esses,new sinks (podwalls, seeds) are formed, and assimilate is
increasingly partitioned to reproductive rather than vegetative tissue
growth. Thereafter, the vegetative tissues share in the fraction of
growth remaining after supplying reproductive tissues.
2.
to Re~rodu~tive
Growth
At the ~ e ~ ~ npod
i n stage,
g
GRQPGRQ starts adding new classes or
cohorts of reproductive sinks on a daily basis. Fruits each cohort
increase in physiolo~calage and pass through a shell growth phase.
Partway through the shell growth phase, seeds in each fruit "set"
and begin their rapid growth phase. Thus, reproductive "sink" demand comp~sesmany i n d i ~ i d u reproducti~e
a~
tissues, all of ~ i ~ e r ~ n t
ages,
~
a a genetic
~ potentia~
i
~ a s s ~~i l a t eema an^ d e p e n ~ e ~ t
on temperature. The priority order for assimilate use is seeds, podwalls, addition new pods, and vegetative tissues. Nitrogen fixation
and nodule growth are actually given higher priority than all other
tissues because CROPGRO computes N demand (based on tissue
growth and "normal" N concentrations) and takes the required car-
S i ~ u ~ a tof
i oCrop
~
Growth
667
Soil nitrogen balance and root nitrogen uptake processes were included in CRQPGRQ using the code from CERES"Wheat (Godwin et
al., 1989). The soil N balance processes-N
uptake, mineralization,
immobilization,nitrification,denitrification,leaching,etc.-are
the
same as described for the CERES models (Godwin and Jones, 1991).
f NO, and NHg are Michaelis- ent ten functions of NO,
concentration, soil water availability, and the root length
density in each layer.If N supply exceeds N demand, the act
crop N uptake is constrained to be equal to or less t
mand. Total crop N demand is computed from the
crease in each organ type multiplied by the m a ~ m u mN concentration allowed for each tissue type. The upper N concentrations for
each tissue type are specified in the SPECIES file. This upper limit
(for vegetative tissues) is subject to downre~lationas the cro
gresses through its reproductive cycle from beginning seed to
rity.
An NZ fixation component has been incorporated into CRQPe is a thermal time requirement for initiating first-nodule
n N uptake is sufficient, nodule growth is slow, receiving
a minimum fraction (bypass flow) of the total assimilate that is alloted to roots. When N uptake is deficient (less than N demand) for
owth of new tissues, carbohydrates are used for N, fixation to the
extent of the nodule mass and the species-defined nodule specific
~ c t i v i If
t ~nodule mass is insu~cient,then assimilates are used for
nodule growth at a rate dependent on soil temperature and speciesule relative growth rate. The NZ fixation rate is further
by soil temperature, soil water deficit, soil aeration (floo
lant reproductive age.
.
Carbohydrates accumulate in leaf and stem tissues under several conditions: (1)when N deficiency limits growth, (2) when there is insu
ficient sink during early sink-lirnited growth or after a full seed loa
is set, and (3) during programme^' accumulation from f l o ~ e r i n ~
until rapid seed growth begins. During this programmed accumulation, up to 30% of the daily vegetative growth increment can be allocated to carbohydrate storage in stems (primarily) and leaves (secondarily), ending when a full seed load occurs and assimilate is no
longer used for vegetative growth. Car~ohydratem o ~ i ~ z a toccurs
io~
continuously but is allowed to accelerate during see
~ o b i ~ z a t i oofn protein (C and N) from old t
tissue occurs slowlyduring vegetative growth, but
vegetative tissue becomes up to twofold faster W
grow (after 75 days in Figs. 8a and 8b). The maxi
m o b i ~ a t i o ndepends on the rate of reproduct
advantage of using mobilized protein is that
produced per day than when all reduced N must come from current
N assimilation or N, fixation. A disadvantage of protein mobilization
is that as leaf N declines (Fig. sa), leaf photosynthesis is decreased
some leaf area is abscised (Fig. 8b).Loss of no
mass also occurswhen protein mass is mobili
leaf mass, leaf protein, stem mass, and stem protein all decline.
Senescence of leaves and petioles is dependent on protein mobilization and is additionally enhanced by drought stress. All three
legumes have gradual senescence of foliage and loss of leaf area an
Boote et al.
670
60
50
40-
8
2
g 20-
DAYSAFTERPLANTING
7.0
6.0
5.0
4 4.0
a:
2.0
'.J
1.o
20
40
60
120
140
DAYSAFTERPLANTING
FIGURE8 (a) Leaf N concentration and (b) LA1 and average mass per seed
for 'Bragg' soybeans (MG 7) planted June 12, 1984 at Gainesville, Florida.
Lines are CROPGRO simulations, and syrnbols are field-measured values,
leaf protein as seed fill progresses, thus reducing seed growth rate;
however, soybeans and common beans additionally have a grand senescence phase that starts at physiolo~calmaturity
stage) and
causes almost all remaining leaves to abscise (Fig. 8b). The grand
senescence of soybean and dry bean foliage causes all seed growth
to cease even if seeds have not filled the pods. Peanut plants, by
contrast, are indeterminate, and their foliage normally remains green
l
will cease
to maturity even though the growth of i n d i ~ d u a seeds
when they reach the limits of their individual pod cavities ( m a ~ m u m
shelling percentage).
S i ~ ~ ~ a t i oCrop
n
Growth
671
The soil water balance in CROPGRO is the same as that in the CERES
growth models and is described in detail by Ritchie (1985). The soil
is divided into a number of layers, up to 10 or more. Water content
in each layer varies between a lower limit ELL( J)] and a saturated
upper limit[SAT(J)]. If water content of a given layer is above a
drained upper limit [DUL(J)], then water is drained to thenext layer
with the tipping bucket approach, using a drainage coefficient
specified in thesoil file.Infiltration and runoff of rainfall and applied
irrigation water depend on the Soil Conservation Service (SCS) runoff
curve number. Vertical drainage may be limited by the saturated hydraulic conductivity (Ksat) for each layer. This feature allows the soil
to retain water above layers that have been compacted or have natural impedance to water flow. In such cases, soil layers may become
saturated for aperiod of time, causing root death, reduced root water
uptake, and decreased N, fixation.
Root mass and root length in each soil layer are state variables
that can change daily. New root length growth each day depends on
the daily assimilate partitioned to roots and a length-to-weight parameter. The distribution of the new root length density into respective soil layers depends on (1) progress of downward root depth
front, (2) a soil-rooting preference
function [WR(L)] describing the
probability distribution of roots growing in each layer of soil (definable late in the season at the end of the root growth period), and (3)
soil water content in each zone. If the fraction available soil water
content in a given zone is less than 0.25 or within 2% of the saturated
value, then root growth in that zone is decreased. Thus, roots tend to
grow into moist soillayers rather than drier layers or saturated layers.
Root senescence in a given zone is accelerated when available soil
water content is below a critical fraction (0.25) or near saturation.
l.
CROPGRO allows several optional methods for computing climatic
potential evapotranspiration (Eo): (1) ThePriestley-Taylor method
(Priestley and Taylor, 1972), also described by Ritchie (1985); (2) the
FAO version of the Penman ET equation as described by Jensen et al.
(1990); and (3) an hourly energy balance that also solves for foliage
and soil temperature (Pickering et al., 1990, 1993, 1995). Only temperature and solar radiation are required to compute the PriestleyTaylor equilibrium evapotranspiration. TheFAO-Penman method
additionally requires windspeed and humidity data. The climaticpo-
Processes sensitive to water deficit include photosynthesis, transpiration, N, fixation, leaf area increase (viadecreased specific leaf area,
SLA), V stage progress, internode elongation (height and width increase), and partitioning to roots. When root water uptake is unable
to meet the potential transpiration of the foliage, total crop photosynthesis and transpiration are reduced equally in proportion to the
decrease in water uptake. Decreases in V stage development, SLA,
and internode elongation and increase in partitioning to roots begin
when the ratio of root water uptake to potential evapotranspiration
first falls below 1.5. These processes
are considered more sensitive to
water supply than photosynthesis. Rate of root depth progression is
accelerated when the ratio drops below 1.0. The N, fixation process
is decreased as the fraction available of soil water in the nodule zone
(between 5 and 40 cm depth) falls belowa species-dependent fraction
available of water (0.35-0.40). Leaf senescence is accelerated (with a
delay feature) following days when plant transpirational demand
cannot be met.
.
With the CROPGRQ model, there is one common set of Fortran code,
and all species attributes associated with dry beans, peanuts, an
soybeans, are input from species files,as well as ecotype and cultivar
files.There are no hardcoded, crop-specific subroutines in GRQ
GRQ; rather, all species or cultivar differences are handled externally
through input parameters and relationships described in these three
input files. The species files describe the basic tissue compositions;
photosynthetic, respiratory; N assimilation, partitioning, senescence,
phenological, and growth processes; and the sensitivity of those processes to en~ronmentalfactors (e.g., temperature, solar irradiance,
plant N status, plant water deficit, or soil water deficit). Cultivar and
ecotype aspects are found in another set of files. For example, the
cultivar file describes, in one line, a minimum set of 15 important
genetic attributes that primarily determine life-cycle attributes b
also include traits such as leaf photosynthesis, seed size, and see
fill duration.
The species files of GRQPGRQ are used to create soybean, peanut, or dry bean attributes and create speciesthat differ in sensitivity
to temperature, solar irradiance, water deficit, etc. Basic differences
in dry matter accumulation ability of species are described by the
curvilinear responses of daily assimilation versus daily solar irradiance (described by two parameters) when all other factors such as
LAI, water, temperat~re, and leaf N are optimum. For the hourly
version of canopy assimilation, the dry matter accumulation ability
depends primarily on single-leaf photosynthesis rate
which is
both a species attribute and a cultivar attribute. Species files have
674
oote et al.
675
cn
T ~ ~ P ~ R A T U RCE ,
lower (Thomas and Raper, 1977, 1978) or even 18C and lower (Hesketh et al., 197'3; Hume and Jackson, 1981), although the response is
cultivar-dependent (Lawn and Hume, 1985). The pod addition rate
function for peanuts is also similar except that the pod addition rate
declines to zero as the temperature drops from 21 to 1'7Caccording
to data of Campbell (1980). The pod addition function for dry beans
is assumed to have its entire response curve shifted to cooler
temperatures.
At higher than optimum temperature, these three grain legumes
exhibit poor reproductive fruit formation, extended vegetative
growth, and poor partitioning. Specific causes may be associated with
increased production of vegetative primordia and the expression of
more vegetative sites, ~gh-temperature-induceddelay in onset of reproductive sites, and failure of successful fertilization of reproductive
sites. Thus, we have included a "partitioning limit" function (Fig. 9)
that decreases the maximum fraction partitioning to pods as temperature exceeds
and 28C in soybeans, peanuts, and dry beans,
respectively. We have since confirmedthis general phenomenon with
experimental data on soybeans (Boote et al., 1994; Pan et al., 1994).
Harvest index declines progressively as temperature increases above
the optimum for soybeans (Baker et al., 1989; Pan et al., 1994) and
peanuts (Hammer et al., 1995).
For additional information on other traits and relationships in
the individual species files, readers are referred to each species file,
.
The CROP^^^ model uses the standard input/ output files specified
project, used in common with other crop models
S Decision Support System for Agrotechnology Transet al., 199410). Required weather data include daily solar radiation, maximum air temperature, minimum air
temperature, and rainfall. To run the Penman-~onteith option, daily
wind run and dewpoint temperature (or minimum daily relative humidity) are also required.
Soils information can typically be obtained or extracted from
U.S.Soil Conservation Service(SCS) ~ublications.Using this SCS
ata, the soils file creation software in the DSSAT can be used to
erive most of the soils traits needed to run the soil water balance.
Caution is recommended, however, in relying on the LL, DUL, and
SAT values obtained with this program, particularly for high clay
content soils. Direct field measurements of LL, DUL, and SAT values
for soil layers are better.
Crop cultural management practices such as planting date, row
acing, plant population, cultivar, fertilization, irrigation, crop resiue, and other treatment information are entered into a File X used
for running the crop model. The File
is used to specify cultural
management conditions for actual site-specificfields and experiments. It can also be used to specify hypothetical "what-if'' simulations. This file has cross-references that link a specific field or treatment to a given soil, to a given weather-year file, and to a given
cultivar in the cultivar file. Cultivar characteristics, 15 in all, are described in the "read-in" cultivar file. Procedures are available for developing some of the more criticalrequired cultivar traits from variety
analysis information. For more inormation,readers
oogenboom et al. (199430).
Boote et al.
678
DAYS AFTER P ~ N T I N ~
'0
20
40
60 80
120 140
DAYS AFTER P ~ N T I N ~
IGURE
(a) Leaf area index and (b) total crop mass and seed mass of
'Bragg' soybeans and 'Florunner' peanuts planted June 12, 1984 at Gainesville, Florida under irrigated conditions. Lines are simulations, and symbols
indicate field-observed values.
Sim~~ation
of Crop Growth
2.500
N
2,Lvv
,500
500
0'
20
100
120
40
60
DAYSAFTER
P~NTING
DAYSAFTER
PLANTING
120
40
20
140
140
FIGURE11 (a) Seed number per squaremeter and (b) average mass per seed
of 'Bragg' soybeans and 'Florunner' peanuts planted June12,1984 at Gainesville, Florida under irrigated conditions. Lines are simulations, and symbols
indicate field-observed values.
cycles and are less day length sensitive can be grown at northerly
latitudes in the United States, Canada, and Europe, whereas cultivars
grown in the southern United States and in the tropics have longer
life cyclesand greater day length sensitivity (approaching MG 12 near
the equator). Table 2 shows critical short day length values (CSDL),
photoperiod sensitivity slopes (PPSEN), and physiological day requirements to complete their life cycles (under short days and warm
temperatures) for MC 00-9 soybean cultivars as solved by Grimm
et al. (1993, 1994) from experimental data. Table 2 also shows simulated days to maturity and yield performance for MC 00-9 cultivars
model
Florida
simulation
Iowa
simulation
CSDL
PPSEN
EM-FL
FL-SD
SD-PM
Maturity Yield Maturity Yield
(h)
(days/h)
(days) (days) (days) (days) (kg/ha) (days) (kg/ha)
00
0
1
2
3
4
5
6
7
8
9
14.35
14.10
13.84
13.59
13.40
13.09
12.83
12.58
12.33
12.07
11.88
0.148
0.171
0.203
0.249
0.285
0.294
0.303
0.311
0.320
0.330
0.340
16.0
16.8
17.0
17.4
19.0
19.4
19.8
20.2
20.8
21.5
23.0
12.0
13.0
13.0
13.5
14.0
15.0
15.5
16.0
16.0
16.0
16.0
30.0
31.0
32.0
33.0
34.0
34.5
35.0
35.5
36.0
36.0
36.5
70.5
73.7
77.2
84.3
94.2
107.9
120.0
132.7
145.0
159.3
170.2
842
916
984
1108
1398
1984
2261
2572
2655
2785
2660
94.0
101.4
109.5
122.1
135.3
152.2
166.8
1514
1838
2086
2399
2463
2280
1730
-d
-d
-d
-d
aCrops were planted on day 123 at 30 plants/m2 and grown under rain-fed conditions, using
10 years of historical weather at each site (1978-1987 in Florida and 10 of 1984-1995 in Iowa).
Critical short day length (CSDL), photoperiod sensitivity (PPSEN),and physiological days from
emergence to flowering(EM-FL), flowering to first seed (FL-SD), and first seed to physiological
maturity (SD-PM) are given for each MG, as used by the CROPGRO soybean model. Physiological days requirement to emergence is 3.6 days.
bFreeze damage in 2 of 10 years.
Freezedamage in 5 of 10 years.
dFreeze damage in 10 of 10 years and maturity not reached.
Soybean yield response to planting date was simulated with CROPGRO for three sites (Florida, North Carolina, and Iowa) using soybean cultivar characteristics described above (Fig. 12). In Florida,optimum planting date for highest yield of MG 7 under rain-fed
conditions was April and May (days 93-123), with good yield levels
for planting in early June. Yields were lower with earlier ~lantings
because plants were shorter and flowered earlier, and plantings on
days 63 and 78 suffered killing frost damage in three out of
and
one out of 10 years, respectively. Later than optimum plantings had
pro~essivelylower yields because of shorter life cycles and smaller
plants. The response to planting date for MG 6 in North Carolina was
4000
~000
P
2000
g
"
60
90
Soybean yield response to planting date simulated under rainfed conditions for 10 years of historical weather (ten from 1984 to 1995) at
Ames, Iowa; 5years of weather (1984-1988) at Clayton, North Carolina; and
10 years (1978-1987) at Gainesville, Florida.Maturity groups
and 7 were
used in Iowa, North Carolina, and Florida, respectively grown at 30 plants/
m' in 76 cm rows in Iowa and 91 cm rows at the other two sites. Vertical
bars represent plus or minus one standard deviation, shown for Florida
simulations.
682
oote et al.
Water deficit decreases canopy assimilation, reduces leaf area expansion, and enhances leaf area abscision. These features are illustrated
in a simulation of 'Cobb soybeans planted June 26, 1981 at Gainesville, Floridaand grown under full irrigation compared to a treatment
that endured two drought periods during vegetative growth (Fig. 13).
During the two periods of soil water deficit, there was a complete
cessation in LA1 iFcrease and minor leaf abscision (Fig.13a).Leaf area
growth was prolonged after the water deficit was relieved, resulting
in only slightly lower LA1 values during seed fill. Biomass accumulation was slowed during each drought period and continued to lag
the fully irrigated treatment (Fig. 13b). Onset of seed growth was
slightly delayed for the plants suffering soil water deficits during
vegetative growth, but final seed yield was reduced less than 10%
because the crop received full irrigation during seed fill and the LA1
was nearly adequate. Droughts during reproductive growth typically
decrease seed yield more, as we have seen with data from other years.
i5
6.0
5.0
2.0
8
g
a
cn
- -
(8)
8,000
irrigated
* Veg. Drought
Lz
Irrigated
3.0 -
3
&.
(A)
6,000
Biorn---
2,000
OO
20
40
60
80
100
DAYS AFTER PLANTING
120
140
FIGURE13 (a) Leaf area index and (b) total crop mass and seed mass of
'Cobb' soybeans (MG 8) planted June
1981 at Gainesville, Florida and
grown under full irrigation compared to atreatment exposed to two drought
periods during vegetative growth. Lines are simulations, and symbols indicate field-observed values.
trates that different seed yields are possible with the same seasonal
E, owing to variation in drought patterns.
atic
Variation in t e m ~ e r a t ~ rconsiderably
e
influences yield potential
across locations, and year-to-year variation influences yield within a
location. Furthermore, global warming of 2-4"C has been predicted
as a consequence of a doubling in "greenhouse" gases (primarily
carbon dioxide increase from fossilfuel burning). For this reason, we
are interested not onlyin overall temperature effects on yield but also
in the effects of temperature variation at a given location (Jones et
al., 1996). Figure 15 illustrates predicted seed yield response to decrease or increase in temperature for MG 3 soybeans grown at Ames,
Iowa and MG 6 soybeans grown at Gainesville, Florida under rainfed conditions using 6 years of historical weather at each site. The
response to temperature change differed across sites and depended
on the site mean seasonal temperature. In Iowa where seasonal mean
temperature was 21.9"C, seed yield was decreased by 2 or 4C tem-
16.0
20.0
6 Det Florida
24.0
SEASONAL ~ E A N
28.0
32.0
T E ~ P E ~ TCU ~ E ,
5 Predicted seed yield response to decrease or increase in temperature for maturity group 3 soybeans grown at Ames, Iowa or for MG 6
soybeans grown at Gainesville, Florida under rain-fed conditions in 91 cm
row spacing at 30 plants/m2, using 6 years of historical weather (1984-1989)
at each site. Mean seasonal temperature over the soybean growth cycle for
the 6 years was 21.9 and 27.0"C at Ames and Gainesville, respectively. [From
Boote et al. (1996),]
et
perature decreasebecause
temperatures became suboptimal for
growth, photosynthesis, and seed fill in addition to extending the life
cycle into freeze hazards. A 2C increase had no effect, and a 4C
increase caused only a small yield decrease in Iowa. By contrast, in
Florida where seasonal mean temperature was much warmer
(27.OoC),yield was increased by 2 or 4C decreases in temperature,
and yield was decreased if temperature was increased. One implication of Fig. 15is that a global temperature increase of 2-4C would
be expected to decrease soybean yields in the southern states but with
little effect in the upper midwestern states. From Fig. 15, it appears
that 22-24C is the optimum mean seasonal temperature for seed
yield. Thisoptimum for seed yield is the integrated result of all model
processes and temperature influences on processes of vegetative development, reproductive development, photosynthesis, respiration,
N, fixation, vegetative growth, and seed growth processes. CROPGROs predicted soybean yield response to temperature is fairly similar to that predicted by SOYGRO V5.42 (Roote et al., 1989a), in part
because many of the temperature relationships are similar. The decline in seed yield at higher than optimum temperature is consistent
with recent experimental data (Pan et al., 1994).
el ~pplicationsin Cro
The CROPGRO model and itsprecursor models have been used considerably in the research mode to synthesize understanding of growth
physiology to hypothesize genetic improvement, to devise production strategies to optimize yield relative to inputs of water and other
natural resources, and to develop policy recommendations relative to
global climatechange (Roote et al., 1996). Possiblythe greatest future
potential (and challenge) is to operationalize the model to be used as
a decision support tool (DSS) for production agriculture (Hoogenboom et al., 1994a). The goal is to help producers make preseason
and in-season decisions related to cultivar choice, planting date, row
spacing, fertilization, irrigation, crop rotation, land area devoted to
each crop enterprise, replant decisions, and minimization of water,
soil, and agrochemical losses. We are addressing this DSS use of the
CROPGRO model in a large interdisciplinary project in soybeanproducing states of the United States. This project
is also dealing with
problems of data availability (historical weather records, current
years weather, soils information, cultivar traits, and prior field history). Site-specific production recommendations with the model will
potentially use geographic information systems and global position-
S i m ~ ~ a t i oCrop
~
~ r o ~ h
~~7
oote et al.
690
Boote et al.
Agric.Exp.Sta.,
Journal No. N-00379.Gainesville,FL:
University of
Florida.
Hoogenboom, G., J.' W. Jones, and K. J. Boote. 1991. Predicting growth and
development of grain legumes with a generic model. ASAE Paper No. 914501. St. Joseph, MI: ASAE.
Hoogenboom, G., J. W, Jones, and K. J. Boote. 1992. Modeling growth, development, and yield of grain legumes using SOYGRO, PNLJTGRO, and
BEANGRO: A review. Trans. ASA 352043-2056.
Hoogenboom, G., J. W. Jones, K. J. Boote, W.T, Bowen, N. B. Pickering, and
W. D.Batchelor.1993.
Advancement in modeling grain legume crops.
ASAE Paper No. 93-4511. St. Joseph, MI: ASAE.
Hoogenboom, G., J. W, Jones, L. A. Hunt, P. K. Thornton, and G. Y. Tsuji.
1994a. An integrated decision support system for crop model applications.
ASAE Paper No. 94-3025. St. Joseph, MI: ASAE.
Hoogenboom, G., J. W. Jones, P.W. Wilkens, W. D. Batchelor, W. T. Bowen,
L. A, Hunt, N. B. Pickering, U. Singh, D. C. Godwin, B. Baer, K. J. Boote,
J. T. Ritchie, and J. W. White. 1994b. Crop models. In: DSSAT Version 3.
Vol. 2. G. Y. Tsuji, G. Uehara, and S. Balas (Eds.).Honolulu, HI: University
of Hawaii, pp. 95-244.
Hoogenboom, G., J. W. White, J. W. Jones, and K. J. Boote. 1994c. BEANGRO,
a process-oriented dry bean model with a versatile user interface. A g o n .
J. 86:182--190.
Hume, D. J., and A. K. H. Jackson. 1981. Pod formation in soybean at low
temperatures. Crop Sci. 22:933-937.
IBSNAT. 1989. Decision Support System for Agrotechnology Transfer Version
2.1(DSSAT V2.1). International Benchmark Sites Network for Agrotechnology Transfer Project. Honolulu, HI: Dept. of Agronomy and Soil Sci.,
College of Tropical Agriculture andHuman Resources, University of
Hawaii.
Jeffers, D. L., and R. M. Shibles. 1969. Some effectsof leaf area, solar radiation, air temperature, and variety on net photosynthesis in field-grown
soybeans. Crop Sci. 9:762-764.
Jensen, M. E., R. D. Burman, and R. G. Allen (Eds.). 1990. vapotranspiration
and ~rrigation~ a ~ e r ~ e ~ u i rAe Manual.
~ e n t s New
:
York: Am. Soc. of Civil
Engineers.
Jones, J. W., and K. J. Boote. 1987.Simulation Models for Soybeansand Other
Crops. Concepts of Crop Systems. Tech. Bull. 106. Taipei City, Taiwan: Food
and Fertilizer Technology Center, pp. 1-7.
Jones, J. W., K. J. Boote, S. S. Jagtap, G. Hoogenboom, and G. G. Wilkerson.
1987. SOYGRO V5.4, Soybean Crop Growth Model, User's Guide. Florida
Agric. Exp. Sta., Journal No. 8304. Gainesville, FL: University of Florida.
Jones, J. W., K. J. Boote, G. Hoogenboom, S. S. Jagtap, and G. G. Wilkerson.
1989.SOYGROV5.42, Soybean Crop Growth Simulation Model. User's
Guide. Florida Agric. Exp. Sta., Journal No. 8304. Gainesville, FL: University of Florida.
oote et al.
Scholberg, J. M. S., K. J. Boote, J. W. Jones, and B. L. McNeal. 1997. Adaptation of the CROPGRO model to simulate the growth of field-grown tomato. In: Systems A ~ r o a c h e sfor Sustaina~le A~icultural ~evelopment:
Applications of Systems Approaches at the Field Level. M. J. Kropff et al. (Eds.).
Dordrecht, The Netherlands: Kluwer, pp. 135-151
Singh, P., K. J. Boote, A. Yogeswara Rao, M. R. Iruthayaraj, A. M. Sheikh, S.
S. Hundal, R. S. Narang, and Phool Singh. 1994a. Evaluation
of the groundnut model PNUTGRO for crop response to water availability sowing
dates, and seasons. Field Crops Xes. 39:147-162.
Singh, P., K. J. Boote, and S. M. Virmani. 1994b. Evaluationof the groundnut
model PNUTGRO for crop response to plant population and row spacing.
Field Crops Xes. 39:163-170.
Spitters, C. J. T. 1986. Separating the diffuse and direct component of global
radiation and its implication for modeling canopy photosynthesis. Part 11.
Calculation of canopy photosynthesis. Agric. For. Meteorol. 38:231 -242.
Spitters, C. J. T., H. A. J. M. Toussaint, and J. Goudriaan. 1986. Separating
the diffuse and direct components of global radiation and its implications
for model canopy photosynthesis. I. Components of incoming radiation.
Agric. For. Meteorol. 38:217-229.
Thomas, J. F., and C. D. Raper, Jr. 1977. Morphologicalresponse of soybeans
as governed by photoperiod, temperature, and age at treatment. Bot. Gaz.
238:321-328.
Thomas, J. E, and D. D. Raper, Jr. 1978. Effectof day and night temperature
during floral induction on morphology of soybean. Agron. J. 70:893--898.
Uehara, G., G. Y, Tsuji, and S. Balas (Eds.). 1994. DSSAT Version 3. Vols. 1Honolulu, HI: University of Hawaii.
Wilkerson, G. C., J. W. Jones, K. J. Boote, K. T. Ingram, and J. W. Mishoe.
1983. Modeling soybean growth for crop management. Trans. A S A E 26:
63-73.
Wilkerson, C. G., J. W. Jones, K. J. Boote, and J. W. Mishoe. 1985. SOYCRO
V5.0: Soybean Crop Growth and Yield Model.Technical documentation. Gainesville, FL: Agricultural Engineering Department, University of
Florida.
Approximations, successive,68
Aquaculture, 650
~ r t i ~ c iintelligence,
al
7
C++, 633
Calibration, 24
Capacitance
in evaporation, 214
Carbon cycle, 498
Chaos
Chaotic behavior), 420
Chaotic behavior, 420
Checks
s ~ m e t80
r ~
dimensions, 77
limiting case check, 79
Continuous processes,
568
costs
irrigation, 285
labor, 546
machinery, 545
repair, 551
timeliness, 546
Cotton
phenology of, 240
Crop management, 235
DEMOS, 633
DISCO, 633, 654
DSSAT, 288
Decision support, 5
support systems, 526
Decision-making, 2
Degree-days
growing, 235
Depreciation, 551
Digestion, 520
694
FARMSYS, 580
Feed intake, 483, 503
Feeding management, 489
Forage, 568
Formulation
mathematical, 83
of models, 69, 73
Forrester diagram, 32
Index
[Grazing]
process of, 503
rotational system, 430
systems, 495
Growth, logistic, 423
Hamiltonian, 452
Harvest
of fruit, 72
Hatchery, salmon, 628, 649
Heat conduction, 69
Herd health, 489
Heuristic, 65
IBSNAT format, 159
Implements, 556
Irrigation
deficit, 283
GOSSYM/ COMAX, 268
knowledge-based, 176
scheduling, 171, 179, 289
water supply; 295
Lactation, 481
Leaf expansion, 249
Livestock production, 475
Lotka-Volterra, 644, 645
MODSIM 11,632
optimum set, 559
selection, 543
Maturity; crop, 655, 659
Memory, human, 4
Michaelis-Menten, 424, 433, 464
Milk production, 483
Models
conceptual, 477, 478
of insects, 39
semantic, 571
spatial/ temporal, 93
whole-farm, 567
Index
Modeling
crops, 67
physiological, 239
process-based, 239
Nitrogen cycle, 499
Object-classes, 570
Object-oriented, 567
Object-Pascal, 633
Objective-C, 633
Optimal control theory, 450
Paradigm programming, evolution, 629, 632, 633, 634
Parameter adjustment, 173
Partitioning, 665
Penman evaporation equation,
198,200
Penman-Monteith evaporation
equation, 202
Phase plane, 439
Photosynthesis, 247, 498, 658
Plant / animal interface, 495
PNUTGRO, 159,652
Precision Farming/ Agriculture, 13,
276
Predator, prey,644
Programming
dynamic, 479
languages, 630
linear, 479, 543
mathematical, 479
object-oriented, 275, 629, 637,
654
evolution of, 629, 632, 633,
634
PROLOG, 570,574
Remote sensing
of evaporation, 227
Representation, 648
695
SIMAN, 631
SIMCOW, 482
SIMHERD, 482
SOYGRO, 159, 652
Seed
germination of, 82
Senescence, 438
Sensitivity analysis, 29, 43
Sensors
for soil water, 158, 160, 190
Sirnula, 630, 631, 654, 655
Simulation,
of biological processes, 19
of field operations, 567
tools, 630
with PROLOG, 575
Simulation models
deterministic, 480
dynamic, 480
empirical, 480
mechanistic, 480
static, 480
stochastic, 480
of biological processes, 19
SMALLTALK, 633
Soil-Plant-Atmosphere Units,
239
Stability, 445
Stocking rate, 518
Stress
crop moisture, 282, 290
[Stress]
crop nutrients, 253, 668
temperature, 674
Sward, 499
Systems
dynamics, 418, 419, 427
hierarchy of, 476
theory of, 476
Systems-oriented research,
477
Validation, 25
Verification, 24
Weather, 583
WEPP,
Workday probability, 553