J. Agric. Engng Res.

(1999) 73, 13}18

Article No. jaer.1998.0385, available online at http://www.idealibrary.com on

Fractal Geometry and Geostatistics for describing the Field Variability

of Soil Aggregation
A. Castrignanò; M. Stelluti
Istituto Sperimentale Agronomico, via Celso Ulpiani 5, 70125 Bari, Italy; e-mail: [email protected]

(Received 24 November 1997; accepted in revised form 24 October 1998)

Fractal geometry has become a widely accepted descriptive tool for speci"c physical properties of natural soils
and fractal scaling has recently been proposed as a model for soil particle size distribution. In this work, the
cumulative mass distribution of dry soil aggregates, M(r(R), was estimated and shown to be proportional to
R-", where r is the aggregate size, R is a speci"c measuring scale and D is the fractal dimension, which is
a measure of soil fragmentation: the larger its value, the greater the fragmentation.
Aggregates were collected from a depth of up to 40 cm of a clay soil located at Metaponto (South Italy). The
"eld of 23;25 m was sampled in 81 sites at the nodes of a semi-regular grid and aggregate size distribution was
obtained by dry sieving the soil through a nest of sieves (sized 10, 5, 2, 1 and 0)5 mm). The estimated D values
were found to vary from 2 to 3 and were then interpolated by using the geostatistical procedure called ordinary
kriging. The results were shown in the form of grey-coloured maps, which could be used as a useful tool for
describing "eld variability in soil aggregation.
 1999 Silsoe Research Institute

1. Introduction many dry aggregate size distributions are lognormal.

More recently, Castrignanò et al. 4 used geometric mean
Soil spatial variability within "elds has been widely diameter (GMD) and geometric standard deviation
demonstrated by soil testing results and crop yield di!er- (GSD) to characterize the in#uence of cropping treat-
ences. There are numerous causes of this variation in soil ments on aggregate fragmentation. Baldock and Kay 5
characteristics and potentials, including soil forming fac- described the cumulative percentage of aggregates by
tors (parent material, topography, vegetation, climate and a weight less than a characteristic linear dimension using
time), farming practice (tillage, crop rotations and fertiliza- a power function. Other authors applied the probability
tion) and erosion. Each cause of variation may operate of aggregate failure to show how wet aggregate stability
independently or in combination with other factors and changed due to cropping history.6
over a wide range of spatial and temporal scales. These Fractal scaling has recently been proposed as a model
factors are re#ected in variations in tillage draught energy, for soil aggregate size distribution.7 Fractal geometry
root penetration and transport of air, water and solutes. was initially formulated by Mandelbrot8 and expanded
The problems that face scientists involved in soil upon by Feder.9 According to Mandelbrot, fractals are
research are twofold. The "rst is to characterize soil characterized by a power law relation between the num-
structure with a single parameter that retains most in- ber and size of objects, whose exponent D is called the
formation. Traditionally, the most direct approaches &&fractal dimension''. Fractal geometry o!ers a powerful
have been pore size distribution and aggregate size distri- descriptive tool for soil scientists, because it provides
bution. As regards the latter, in particular, several empiri- a quantitative framework for integrating soil biological,
cal indices have been proposed for describing the entire chemical and physical phenomena over di!erent spatial
distribution with a single value. Van Bavel1 used the and temporal scales. Moreover, Perfect and Kay 7 com-
mean weight diameter, Mazurak 2 suggested the geomet- pared the performance of D as a statistical descriptor of
ric mean diameter (GMD) may be more appropriate. fragmentation to that of the other indices. They showed
That was supported also by Gardner's3 observation that that D was more sensitive to cropping treatments.
0021-8634/99/050013#06 $30.00/0 13  1999 Silsoe Research Institute
14 A. C AST RIG N AN OE ; M. S TEL L U TI

The second problem lies in describing the observed lower boundary de"ned by the sieve diameters. The
variations in soil structure. Researchers have tradition- number-size distribution can also be predicted from the
ally attempted to remove spatial and temporal variability mass-size distribution, if the shape and bulk density of
by blocking or statistical averaging procedures. The price aggregates in each size fraction is known. However, some
to pay has often been a failure to understand processes errors may be introduced by assuming invariant scale
acting in the soil. values for the shape and bulk density. 14,15 Therefore, it
In contrast geostatistics, developed by Matheron,10 was considered more appropriate to investigate the ag-
provides a body of statistical techniques aimed at detect- gregate size distributions expressed in terms of mass,
ing and modelling the patterns of spatial dependence of which is an easily measured quantity. In this case, strict
attributes in space, rather than evaluating linear spatial self-similarity is preserved.14
average values. 11 Tyler and Wheatcraft14 developed a mass-size based
The object of this paper was to explore the applicabil- model for the estimation of fractal dimension:
ity of fractal theory to study spatial variability in soil
M(r(R)/M "(R/R )-" (2)
aggregation. The fractal dimension D together with an- 2 *
other model parameter which estimates the size of the where M is the cumulative mass of aggregates of size
largest aggregate R were used as statistical descriptors r less than R (e.g. sieve aperture), M the total mass,
* 2
of fragmentation. Their values, estimated at di!erent loca- R a parameter which estimates the size of the largest
tions of the "eld, were interpolated using the geostatistical *
aggregate and D the mass-fractal dimension.
technique of ordinary kriging12 and then mapped. The model of Eqn (2) sets an upper limit to the value of
D at 3. In fact, for D'3 the cumulative mass exceeds the
total mass with decreasing observation size, which is
2. Materials and methods a physically impossible situation. At the limiting value of 3,
the distribution is independent of scale. Equation (1)
2.1. Fractal theory suggests that D may take any positive value, because for
D less than zero, the cumulative number of aggregates
Fractals are spatial and temporal model systems that greater than R decreases as R is decreased; and that is
exhibit scaling symmetry, i.e. they are constructed by another non-physical situation. Moreover, at D"0 the
repeatedly copying a pattern or &&generator'' on a starting distribution is independent of observation size. There-
object known as the &&initiator''.8 fore, the range of variability of D is strictly limited to
Fractal theory can be applied for characterizing ag-
gregate size distributions in soil. Turcotte13 and Mandel- 0(D(3 (3)
brot8 suggested the fractal relationship for aggregate size In the model of Eqn (2), aggregate density and shape
distributions of the form are still assumed to be scale invariant.14 In the past, D in
N (r'R)"KR\" (1) Eqn (2) was obtained from the slope of the log}log plot
by a least squares "t performed over the entire range of
where N (r'R) is the cumulative number of objects of scales for which data were available.13,16 However, asym-
size r larger than a speci"c measuring scale R as deter- metries originating during the early stages of fragmenta-
mined by the sieve diameter, K is a constant equal to N at tion cause the distributions to deviate from linearity, as
R"1, relating to the shape of the objects and D is the logP0. In this paper, therefore, the variation of the
fractal dimension. The value of D depends on both the D and R estimates obtained by non-linear regression
shape of individual objects and the extent of fragmenta- *
was investigated. As proved elsewhere,17+19 this approach
tion across all the spatial scales considered. In the case of can be employed as an unbiased estimator of the fractal
regular objects, such as cubes, D is a measure only of dimension.
fragmentation. Perfect and Kay 7 showed D to be inde-
pendent of the Euclidean geometry used: when aggreg-
ates were represented by spheres instead of cubes, the 2.2. Geostatistical procedures
estimated values of D remained unchanged. The larger
the value of D, the greater is the fragmentation. At D"0, Semivariograms12 were used to examine the spatial
the size distribution is dominated by few in"nitely large dependence between measurements at pairs of locations
objects. As D increases, the number of small objects as a function of separation distance h, called lag.
increases at the expense of the larger ones.7 Semivariance c(h) was computed using the expression
In most soil analyses, Eqn (1) cannot be applied dir-
ectly, since it is not practical to count the number of 1 KF
c(h)" [z(x )!z(x #h)] (4)
aggregates, whose size ranges between an upper and 2n(h) G G
G
 FRA CT AL G EO M ET RY O F S OI L AG G RE G AT IO N 15

where n(h) is the number of samples separated by a dis- as Typic Epiaquerts according to USDA (Soil Survey
tance h and z represents the value calculated from Eqn (2) Sta! 23 ) and located at Metaponto-Matera (Southern
for D or R . Italy, 40324 latitude N; 16348 longitude E; 16 m above
*
A spherical model was "tted to the semivariance data sea level). The experimental 21;25 m "eld was sampled
of Eqn (4) using the non-linear least-squares method. The at the nodes of a grid at intervals of 2)5 m eastwards and
chosen model is given by: 3 m northwards for a total of 81 samples. Owing to the
small size of the agronomic plots, the grid could not


C #C [1)5(h/a)!0)5(h/a)], 0)h)a
c(h)"   be combined with a nested scheme, as proposed by
C #C , h'a McBratney et al. 24 and Webster et al. 25
 
(5) Approximately 1000 g of the material oven dried at
403C was placed on a nest of sieves (sized 10, 5, 2, 1 and
where h is the separation distance between observations; 0)5 mm) in a dry-sieving apparatus. The apparatus was
and a is a model coe$cient known as the &&range'', which operated for 2 min with a speed of 150 cycles min\ for
is a measure of the maximum distance over which the a total of 96 vertical strokes. The weight of aggregates
property z is spatially correlated. At separation distance retained on each sieve was recorded after oven drying at
greater than the range, the semivariance remains con- 1053C for 24 h. The values were corrected by subtracting
stant at a value known as the &&sill''. The model coe$cient the percentage of sand in the aggregate fractions of size
C is known as the &&nugget'' and C is another coe$cient less than 2 mm; thus the resulting size distributions did
 
(structural coe$cient), which equals the sill minus the not include detached primary particles. The mass-size
nugget. Ideally, the experimental variance should pass distribution data were used to "t Eqn (2), i.e. for the
through the origin when the distance of sample separ- estimation of R and D, using non-linear least-squares
ation is zero. However, many soil properties have non- *
methods.26
zero semivariances as h tends to zero. This non-zero
variance is called 20 the &&nugget variance'' or &&nugget
e!ect'' (C ) and represents unexplained or &&random vari-
 3. Results and discussion
ance'', often caused by measurement errors or variability
of the measured property at a spatial scale smaller than The results from the dry-sieving procedures (Fig. 1 )
the one of sampling. The structural coe$cient C then
 show that more than 50% of the aggregates had sizes
represents the component of total variance originating greater than 10 mm. The remaining 50% was distributed
from spatial patterns in the soil.
Ordinary punctual kriging20 was used to interpolate
the fractal parameters D and R in the "eld area. Lag
*
value was chosen equal to the minimum sampling dis-
tance of 2)5 m. Values for each parameter were estimated
on a regular grid at spacing of 1)0;1)0 m using 10 neigh-
bourhood points for punctual interpolation. Semivario-
grams were optimized by cross-validation.21
Ordinary kriging is a method for making optimal,
unbiased estimates of regionalized variables at unsam-
pled locations using the structural properties of the
semivariogram and the initial set of measured data.20
A useful advantage of kriging compared with the other
traditional linear interpolators is that an error term,
expressing the estimation variance or uncertainty in es-
timation, is calculated for each interpolated value. More-
over, kriging has the property of exactitude i.e. it returns
the datum value for the estimate if the location to be
estimated coincides with a sampled location.22

2.3. Data set

One composite soil sample was collected from a depth

of 2}40 cm using a hand auger. The soil is clay, classi"ed Fig. 1. Proportion of aggregate size distribution
16 A. C AST RIG N AN OE ; M. S TEL L U TI

among the ranges 1}2, 2}5 mm (&12%), 5}10 mm Table 2

(&20%). The "ner particles ((0)5 and 0)5}1 mm) were Summary of coe7cients of spherical semivariogram models for
not very frequent (&2%). the fractal parameters RL and D
At "rst, calculated data for fractal parameters R and
* Variable Value Standard error
D were analysed using classical statistical techniques27 to
obtain values for the mean, standard deviation, min- Mass-fractal dimension (D)
imum, maximum, coe$cient of variation and correlation Range, m 7)233 1)864
coe$cient. Descriptive statistics of the two parameters Sill}Nugget 0)006 0)002
Nugget 0)007 0)002
(Table 1) showed that the coe$cients of variation range Residual 0)001
from 5)2% for D to 34)2% for R , indicating a low
*
variation for the degree of fragmentation D and a me- ¸argest aggregates size (R )
*
dium variation for the presence of coarser material in the Range, m 5)478 3)576
"eld R , according to the classi"cation by Warrick and Sill}Nugget, mm 26)618 34)706
* Nugget, mm 41)552 34)923
Nielsen.28 The foregoing results then depict a soil, Residual, mm 5)224
homogenous enough in "ner particles but quite change-
able in coarser materials.
The correlation coe$cient between R and D was quite
*
low (!0)12), though signi"cant at a 1% level of prob- pairwise sample variation depends upon the distance of
ability. The above result suggests that R and D supply separation. Moreover, D has the structural component
*
di!erent information types: R may be linked mainly to (sill}nugget) which is signi"cantly larger than nugget,
*
the presence in the "eld of coarser materials; whereas which indicates that spatial variability is not completely
D may be considered a fragmentation indicator, a!ected random. Figure 2 shows the experimental variogram of
by both soil texture and past management. Moreover, D and the spherical model "tted to data.
the negative correlation between the two parameters As regards R , the range was smaller than that for
*
indicates that the probability of occurrence of coarse D and of 5)5 m, indicating variability of the short-range
material was higher where the degree of fragmentation type. In fact, R exhibited essentially random variability,
*
was lower. the structural component of the variance being less than
Geostatistical methods are more suitable for analysis nugget variance. That can be inferred also from the
of properties that show spatially correlated behaviour. examination of its semivariogram (Fig. 3 ), which appears
Direct semivariograms were computed for each soil para- not very well structured with a very high nugget e!ect
meter and the optimized coe$cients for the best "tting compared with the total sill.
spherical models were estimated (Table 2). The cross- The classical and geostatistical results presented sug-
variogram was not signi"cantly di!erent from 0, which gest that spatial patterns in soil fragmentation may be
con"rms the result of the classical analysis that the two detected in the study "eld. Two maps of spatial patterns
parameters are not signi"cantly spatially correlated. As
regards D, the range of in#uence was 7)2 m, so at distan-
ces shorter than the range, variability is non-random and

Table 1
Descriptive statistics for the fractal parameters RL and D

Statistics Largest aggre- Mass-fractal

gates size dimension
(R ), mm (D)
*
Mean value 26)186 2)199
Median value 24)151 2)209
Standard deviation 8)258 0)114
Variance 67)361 0)013
Skewness 1)238 !0)401
Kurtosis 4)790 2)642
Minimum data value 12)390 1)938
Maximum data value 57)495 2)398
Coe$cient of variation 0)342 0)052 Fig. 2. Experimental variogram (point) with xtted spherical
model (line) for the mass-fractal dimension D
 FRA CT AL G EO M ET RY O F S OI L AG G RE G AT IO N 17

Fig. 3. Experimental variogram (point) with xtted spherical

model (line) for the largest aggregates size R
*

in R and D (Fig. 4 ) were produced by interpolating

*
from calculated values of the two fractal parameters
using the semivariogram models in Table 2 and the
method of punctual kriging. The high variability in
soil fragmentation is quite clear from the observation
of the two maps, especially for R . However, it is
*
still possible to detect some spatial patterns: aggreg-
ates of the smallest sizes were present more often in
the lower half of the "eld (low R values) and a wide
*
and fairly homogenous spatial pattern was located paral-
lel to the principal diagonal of the "eld (medium
D values). The highest degree of fragmentation was ob-
served especially in the left lower corner of the "eld. As
suggested by the small negative correlation between the
two parameters, the most fragmented zones (high D)
corresponded to the ones with coarse materials of the
smallest sizes (low R ). The foregoing result shows that,
*
in spite of the very small value of the linear correlation
coe$cient, there was a signi"cant spatial correlation be- Fig. 4. Maps of kriged values for the largest aggregates size R
*
tween the two parameters. A probable explanation of this and the mass-fractal dimension D
di!erence between the two approaches, traditional and
geostatistical, might be that classical statistical methods
do not account for sample locations and cannot detect 4. Conclusions
and model the spatial patterns of the soil attributes.
Spatial distributions of aggregates of di!erent sizes rarely Spatial variability of soil aggregation of an agricultural
are the consequence of chance, but rather they are the "eld of 23;25 m was studied using two known method-
result of natural processes and management practices ologies combined together: fractal geometry and geos-
producing special patterns of spatial correlation in the tatistics. First, a new fractal mass-based model was
soil. applied to describe dry soil particle size distribution.
However, owing to the small size of the study "eld, The two model parameters, the fractal dimension D and
it was quite di$cult to disclose well-de"ned spatial the size of the largest aggregates R , were assumed as
*
patterns. Soil fragmentation was essentially character- a descriptive tool for soil aggregation. The values of
ized by variability at short range, mostly induced by farm D and R , estimated at each sampling point, were then
*
management and cropping sequence. interpolated and mapped using geostatistics.
18 A. C AST RIG N AN OE ; M. S TEL L U TI

9
The results of this study have shown that, in spite of Feder J Fractals. New York: Plenum Press, 1988
10
the small "eld size, spatial patterns in soil aggregation Matheron G La TheH orie des Variables ReH gionaliseH es et ses
were non-random. The "eld could be divided roughly Applications. (The theory of regionalized variables and
their applications). Fascicule 5, les Cahiers du Centre de
into two halves, with the left half characterized by a high- Morphologie MatheH matique, Ecole des Mines de Paris,
er degree of fragmentation and smaller aggregates. Con- Fontainebleau, 1970
11
sidering the small size of the "eld, it was very likely that Castrignanò A; Mazzoncini M; Giugaliarini L Spatial char-
most of the soil variability resulted from past farming acterization of soil properties. Advances in GeoEcology,
management (tillage, crop rotation, fertilization and 1998, 31, 105}111
12
Isaaks E H; Srivastava R M Applied Geostatistics. New
irrigation). York: Oxford University Press, 1989
The two methodologies used are not new and have 13
Turcotte D L Fractals and fragmentation. Journal of Geo-
already been widely applied as a descriptive tool of phys- physical Research, 1986, 91(B2), 1921}1926
14
ical systems. However, the originality of the proposed Tyler S W; Wheatcraft S W Fractal scaling of soil particle-
approach consists of using fractal geometry combined size distributions: Analysis and limitations. Soil Science
Society American Journal, 1992, 56, 362}369
with geostatistical procedures, which may have signi"- 15
Perfect E; Kay B D Brittle fracture of fractal cubic aggreg-
cant implications in soil speci"c management. ates. Soil Science Society American Journal, 1995, 59,
Fractal geometry, in fact, o!ers a quantitative tool for 969}974
16
integrating soil information of di!erent types; interpola- Perfect E; Rasiah V; Kay B D Fractal dimension of soil
ting the two fractal parameters D and R by using geos- aggregate-size distributions calculated by number and
* mass. Soil Science Society American Journal, 1992, 56,
tatistics has resulted in delineating homogenous areas in 1407}1409
the "eld, that could then receive the same agricultural 17
Perfect E; Kay B D; Rasiah V Unbiased estimation of the
operations. fractal dimension of soil aggregate size distributions. Soil
Matching management with the soil-speci"c di!erences Tillage Research, 1994, 31, 187}198
18
in the "eld can be an important means of reducing water Perfect E Fragmentation of Euclidean and fractal cubic
aggregates. In: Fractal in Soil Science. Baveye P; Parlange
and environmental contamination with agricultural chem- J Y; Stewart B A (eds). Boca Raton, FL: Lewis Publishers,
icals: the proposed approach, combining fractal geometry 1995
19
and geostatistics, may help to face this future challenge. Castrignanò A; Stelluti M Applicazione dello &&Scaling'' Frat-
tale alla suddivisione a secco degli aggregati: Analisi
e Limitazioni. (Application of fractal scaling to dry aggreg-
ate size distribution: Analysis and limitations). Rivista di
References Agronomia, 1997, 3, 832}839
20
Journel A G; Huijbregts C H Mining Geostatistics: New
1
Van Bavel C H M Mean weight-diameter of soil aggregates York: Academic Press, 1978
21
as a statistical index of aggregation. Soil Science American Journel A G Geostatistics for the environmental sciences.
Proceedings, 1949, 14, 20}23 EPA project no. cr 811893. Technical Report, US EPA,
2
Mazurak A P E!ect of gaseous phase on water-stable syn- EMS Lab, Las Vegas, NV, 1987
22
thetic aggregates. Soil Science, 1950, 69, 135}148 Deutsch V D; Journel A G GSLIB. Geostatistical Software
3
Gardner W R Representation of soil aggregate-size distribu- Library and User's Guide. (2nd Ed). Oxford University
tion by a logarithmic-normal distribution. Soil Science Press, New York, 1989
23
American Proceedings, 1956, 20, 151}153 Soil Survey Sta4 Soil Survey Manual. USDA } SCS. U.S.
4
Castrignanò A; Di Bari V; Colucci R; Marrone G Variazione Gov. Print. O$ce, Washington, D.C., 1992
24
temporale della distribuzione di aggregati di terreno sot- McBratney A B; Webster R; Burgess T M The design of
toposto a di!erenti avvicendamenti colturali. (Temporal optimal sampling schemes for local estimation and map-
variation of aggregate size distribution of a soil submitted ping of regionalized variables, I: theory and methods.
to di!erent cropping sequences). Agricoltura Ricerca, Computer and Geosciences, 1981, 7, 331}334
25
1993, 151/152, 183}191 Webster R; Atteia O; Dubois J P Coregionalization of trace
5
Baldock J A; Kay B D In#uence of cropping history and metals in the soil in the Swiss Jura. European Journal of
chemical treatments on the water-stable aggregation of Soil Science, 1994, 45, 205}218
26
a silt loam soil. Canadian Journal Soil Science, 1987, 67, SAS/STAT SAS Institute Inc. SAS/STAT User's Guide, Re-
501}511 lease 6.03 edition. Cary, NC: SAS Institute Inc., 1988,
6
Kay B D; Angers D A; Groenevelt P H; Baldock J A Quan- 1028pp
27
tifying the in#uence of cropping history on soil structure. Steel R G D; Torrie J H Principles and procedures of
Canadian Journal of Soil Science, 1988, 68, 359}368 Statistics: a Biometrical Approach. New York, NY:
7
Perfect E; Kay B D Fractal theory applied to soil aggrega- McGraw Hill, 2nd Ed. 1980, 633 pp
28
tion. Soil Science Society American Journal, 1991, 55, Warrick A W; Nielsen D R Spatial variability of soil physical
1552}1558 properties in the "eld. In: Applications of Soil Physics.
8
Mandelbrot B B The Fractal Geometry of Nature. 2 Edn (Hillel D ed.) pp. 319}344. New York: Academic Press,
Londra: W. H. Freeman, 1982 1980