Computers, Environment and Urban Systems 49 (2015) 1–14

Contents lists available at ScienceDirect

Computers, Environment and Urban Systems

journal homepage: www.elsevier.com/locate/compenvurbsys

A land-use spatial optimization model based on genetic optimization

and game theory
Yaolin Liu a,b,⇑, Wei Tang a, Jianhua He a,b, Yanfang Liu a,b, Tinghua Ai a,b, Dianfeng Liu a,b
a
School of Resource & Environment Science, Wuhan University, 129 Luoyu Road, Wuhan 430079, China
b
Key Laboratory of Geographic Information System, Ministry of Education, 129 Luoyu Road, Wuhan 430079, China

a r t i c l e i n f o a b s t r a c t

Article history: Land-use patterns can be considered as a consequence of competitions between different land-use types.
Received 20 May 2013 How to coordinate the competitions is the key to land-use spatial optimization. In order to improve the
Received in revised form 5 August 2014 ability of existing land-use spatial optimization models for addressing local land-use competitions (the
Accepted 4 September 2014
competitions on land units), a loosely coupled model based on a genetic algorithm (GA) and game theory
is constructed. The GA is repeatedly executed to separately optimize the spatial layout of each land-use
type. The land-use status quo is overlaid with the optimization results to find local land-use competi-
Keywords:
tions. The concept of land-use competition zones is introduced in this study. Using the competition zones
Land-use spatial optimization
Genetic algorithm
as the basic units, the model utilizes multi-stakeholder games and the knowledge of land-use planning to
Local land-use competition coordinate the local land-use competitions. The final solution is obtained after the land-use coordination.
Game theory Gaoqiao Town, Zhejiang Province is selected as the study area to verify the validity of the model. The
experimental results confirm that the model is feasible to undertake land-use spatial optimization and
to coordinate the competitions between different land-use types.
Ó 2014 Elsevier Ltd. All rights reserved.

1. Introduction Cellular automata models are based on land-use conversion rules

for local areas and can generate different land-use patterns under
Land-use spatial optimization is a complicated spatial decision- different conditions using a bottom-up approach (Li & Yeh, 2000,
making problem with multiple conflicting objectives. How to coor- 2002). Irrational land use in local areas can be adjusted using
dinate land-use competitions is the key to land-use spatial optimi- well-designed land-use conversion rules. However, optimization
zation. The quantitative constraints of land resources and their objectives and other important macroeconomic factors cannot be
multi-suitability are the main reasons for these competitions easily incorporated into cellular automata models.
(Chen, 2007; Liu, Wang, & Long, 2008). From the spatial perspec- Land-use spatial optimization is a complicated combinational
tive, land-use competitions can be considered as a scramble for optimization problem. The computational intensity increases expo-
land units between multiple land-use types; however, in essence nentially as the scale of the problem increases when using an
these competitions are a conflict of interest between multiple exhaustive search to solve the problem (Xiao, 2008). For many
stakeholders (Zhang, Li, & Fung, 2012). real-world applications, numerous spatial variables are involved,
Many researchers have performed extensive research in the and many spatial and non-spatial constraints need to be handled.
field of land-use optimization. The existing optimization models Conventional mathematical models are not able to find the optimal
can be roughly divided into the following three categories: linear solution within an acceptable timeframe (Aerts, Eisinger, Heuvelink,
programming models, cellular automata models, and models based & Stewart, 2003). Therefore, as a compromise between the time-
on intelligent algorithms. Linear programming models can quickly frame and the optimality of the final solutions, various intelligent
determine the optimal land-use structure according to specific algorithms for land-use spatial optimization have been developed,
objectives and constraints (Arthur & Nalle, 1997; Chuvieco, 1993; such as simulated annealing algorithms (Aerts & Heuvelink, 2002;
Sadeghi, Jalili, & Nikkami, 2009); however, these models are unable Santé-Riveira, Boullón-Magán, Crecente-Maseda, & Miranda-
to change the land use of parcels and to undertake spatial optimi- Barrós, 2008), particle swarm algorithms (Liu, Li, Shi, Huang, & Liu,
zation (Santé-Riveira, Crecente-Maseda, & Miranda-Barrós, 2008). 2012; Liu, Liu, Liu, He, Jiao, et al., 2012; Liu, Wang, Ji, Liu, & Zhao,
2012), ant colony algorithms (Li, Shi, He, & Liu, 2011; Liu, Li, et al.,
⇑ Corresponding author at: School of Resource & Environment Science, Wuhan 2012), and genetic algorithms (GAs). While these algorithms may
University, 129 Luoyu Road, Wuhan 430079, China. not be able to obtain the optimal solution for every case, near-opti-

http://dx.doi.org/10.1016/j.compenvurbsys.2014.09.002
0198-9715/Ó 2014 Elsevier Ltd. All rights reserved.
2 Y. Liu et al. / Computers, Environment and Urban Systems 49 (2015) 1–14

mal solutions are often good enough to support decision-making for which may lead to difficulty in implementing the coordination
regional land-use planning. Geographic information systems (GISs) results.
play an important role in the application of the intelligent algo- Game theory can simulate the decision behavior of various
rithms to land-use spatial optimization (Wu & Grubesic, 2010). stakeholders in a conflict of interest and assist them in making
These algorithms and GIS can be loosely coupled through the the most favorable decision (Rasmusen, 2001; Zhang, 2004). Rele-
exchange of data files. GIS are utilized to process and visualize spa- vant studies have applied game theory to monitor land-use
tial data for these algorithms. Moreover, for open GIS, part of GIS changes (Wu, Wu, & Shen, 2005), allocate water resources (Liu,
functions can be reused in the development of these intelligent Sun, Gu, & He, 2002), cope with multi-objective optimization
algorithms. (Lee, 2012), and solve land-use conflicts (Hui & Bao, 2013). How-
GAs are widely used among the aforementioned intelligent algo- ever, game theory methods are usually applied in isolation and
rithms (Goldberg, 1989; Zhou & Sun, 1999). This technique has been have been rarely coupled with land-use optimization models. In
applied to solve common spatial optimization problems, e.g., facility consideration of the difficulty in coupling the genetic operators
location (Brookes, 1997, 2001; Li & Yeh, 2005), forestry planning with game theory methods, a loosely coupled model based on a
(Ducheyne, De Wulf, & De Baets, 2006; Fotakis, Sidiropoulos, genetic algorithm and game theory is constructed in this study.
Myronidis, & Ioannou, 2012), water resource allocation (Fotakis & The model is intended to improve the ability of existing land-use
Sidiropoulos, 2011), and land-use optimization (Holzkämper & spatial optimization models for addressing local land-use competi-
Seppelt, 2007; Lautenbach, Volk, Strauch, Whittaker, & Seppelt, tions. The second part of this study provides the details of the
2013; Porta et al., 2012; Stewart, Janssen, & van Herwijnen, 2004). model. The third part introduces the study area and relevant data.
Pareto front based methods are often integrated with GAs to illus- The corresponding experimental results are described and ana-
trate the trade-off among conflicting optimization objectives, which lyzed in the fourth part, and the relevant conclusions are given in
assists planners in addressing regional-scale land-use competition the final part.
(Balling, Taber, Brown, & Day, 1999; Bennett, Xiao, & Armstrong,
2004; Cao et al., 2011; Morio, Schädler, & Finkel, 2013). However, 2. The land-use spatial optimization model
these GA-based optimization models do not consider local land-
use competitions (the competitions on land units). The genetic oper- The land-use spatial optimization model in this study consists
ators have been improved from the random search techniques to of two parts (Fig. 1). In the first part, an improved GA is repeatedly
being combined with the knowledge of land-use planning in previ- executed to separately optimize the spatial layout of each land-use
ous studies (Cao, Huang, Wang, & Lin, 2012); however, the combina- type. In the second part, the land-use status quo is overlaid with
tion is simple. These operators lack an effective game-based the optimization results from the first part to find local land-use
coordination mechanism for local land-use competitions. The competitions. The concept of land-use competition zones is intro-
important role of interest factors is ignored in the coordination, duced in this study (see Section 2.2). Using the competition zones

Fig. 1. Model framework.

 Y. Liu et al. / Computers, Environment and Urban Systems 49 (2015) 1–14 3

as the basic units, the model utilizes multi-stakeholder games and land resources. The two objectives are integrated together using
the knowledge of land-use planning to coordinate the local land- a weighted sum. The landscape shape index is selected as the met-
use competitions. Interest factors are introduced into the coordina- ric for determining land-use compactness. The relevant formulas
tion process through multi-stakeholder games (see Section 2.3.1). are as follows:
Different game models can be constructed according to the charac- PM PN
teristics of different competition types. The knowledge of land-use i¼1 j¼1 sij  uij
S¼ PM PN ; ð1Þ
planning ensures the rationality of the coordination results. In this i¼1 j¼1 uij
model, the local land-use competitions are roughly divided into
two categories: agricultural land competition and competition XH
P
between agricultural land and development land. The former cate- C¼ phffiffiffiffiffi ; ð2Þ
gory is the competition between agricultural lands, e.g., farmland h¼1 4 Ah
versus garden and farmland versus forest. The latter category is
S C
further subdivided into two types: agriculture (status quo)-devel- Maximize : F ¼ w1  f norm ðSÞ þ w2  f norm ðCÞ: ð3Þ
opment land competition and development (status quo)-agricul-
In Formula (1), sij is the suitability of the focused land-use type in
ture land competition.
the cell indexed by i and j. Moreover, uij is a binary-state variable
The land-use coordination is a multiple stakeholder game. Dif-
that is 1 if the cell indexed by i and j is allocated to the focused
ferent stakeholders and interest demands are involved in different
land-use type; otherwise, the value is 0. In Formula (2), H is the
local land-use competition types. In this study, solving the agricul-
number of patches in the chromosome, Ph is the perimeter of a
ture-development land competition problem is used as an example;
patch, and Ah is the area of a patch. A patch is a set of cells that
a game model that simulates the negotiations between farmers and
are contiguous with each other and are allocated to the focused
the government is constructed to coordinate this competition type.
land-use type. Smaller values of Formula (2) correspond to better
The other local land-use competition types are solved using guid-
compactness. Formula (2) favors the chromosomes with compact
ance from land-use planning knowledge. S C
and fewer patches. In Formula (3), f norm and f norm are utilized to nor-
malize the objective values (see Section 4.1 for details). Moreover,
2.1. Genetic optimization w1 and w2 are the weights of objective S and objective C. These
weights satisfy the following constraints: (1) w1 + w2 = 1, (2)
2.1.1. Chromosome structure and initialization 0 6 w1 6 1, and (3) 0 6 w2 6 1.
The vector chromosome and the grid chromosome are the most
commonly used structures in land-use spatial optimization 2.1.3. Land-use constraints
(Balling et al., 1999; Cao et al., 2011). The grid chromosome is Land-use conversion is not random and is subject to constraints
selected in this study because of the effective representation of from multiple aspects, such as the economy, society, and ecology.
the study area and the convenient manipulation of land units. Each Violations of the constraints often lead to irrational allocation of
execution of the GA focuses on optimizing that spatial layout of land resources. We divide the constraints into two types: (1) area
only one land-use type. Through repeated executions of the GA, constraints and (2) spatial constraints. The area constraints macro-
the model separately optimizes the spatial layout of each land- scopically control the use of regional land resources (Cao et al.,
use type (Fig. 1). Therefore, the grid cells (namely genes) in a chro- 2012). These constraints are utilized to maintain a rational land-
mosome are all binary variables. When a cell is allocated to the use structure in land-use spatial optimization, such as the mini-
focused land-use type, its value is 1; otherwise, the cell has a value mum farmland area and the maximum development land area.
of 0. The land-use patterns in the chromosomes are too fragmented The spatial constraints restrict land-use conversion within specific
when the random strategy for initialization is utilized. Therefore, grid cells (the land-use data are provided in a raster format). For
each chromosome in the population is initialized to the land-use example, land use in a nature reserve should remain unchanged.
status quo (Fig. 2). The two initialization strategies are compared To satisfy the area constraints, a penalty method is introduced into
in Section 4.1. the fitness evaluation of chromosomes (see Section 4.1). The spa-
tial constraints are considered in the execution of the mutation
2.1.2. Objectives operator (see Section 2.1.4).
Land-use suitability S and compactness C are the objectives con-
sidered in the spatial layout optimization of each land-use type. 2.1.4. Genetic operators
Improving the suitability is conducive to the rational use of land The random search adopted in the conventional crossover and
resources. Compact land use is important for achieving sustainable mutation operations leads to arbitrary land-use conversion and
economic development, especially considering the deficiency of generates irrational land-use patterns. The conventional crossover

Fig. 2. The initialization strategy of chromosomes.

4 Y. Liu et al. / Computers, Environment and Urban Systems 49 (2015) 1–14

Fig. 3. Schematic representation of the mutation operation.

and mutation operations have been improved in some studies.

Knowledge of land-use planning is incorporated into the new oper-
ators. When changing the value of a gene, both land-use con-
straints and the land use in the neighborhood of the gene are
considered (Cao et al., 2012; Fotakis et al., 2012).
In this study, the crossover operator randomly selects one
CN  CN window in the study area and subsequently swaps the land
use in the windows in the two chromosomes. The number of the
‘‘select and swap’’ operations executed on the two chromosomes
in one iteration of the GA is N1, which is a model parameter. The
mutation operator randomly selects two 3  3 windows in the
study area. The average suitabilities of the focused land-use type
in the two windows are individually calculated. The window with
a better suitability is labeled ‘‘H’’. The other window is labeled ‘‘L’’. Fig. 4. The structure of a competition zone and its neighborhood.
The mutation operator ‘‘moves’’ the land use in window L to win-
dow H (Fig. 3). The cell values in window L are changed from 1 to
0, while the cell values in window H are changed from 0 to 1. How- development-agriculture land competition are solved using knowl-
ever, the number of value-changed cells in window L should be edge of land-use planning. For an AA competition zone (see Section
identical to the number of value-changed cells in window H. There- 2.2 for the definition of AA), the priority degree PD for each agricul-
fore, the moving process terminates if either of the two following tural land type is separately calculated. The formula is as follows:
conditions is reached. First, the cell values in window L are all 0 PDi ¼ C i þ Di ; ð4Þ
(Fig. 3(a)). Second, the cell values in window H are all 1 (Fig. 3(b)).
The spatial constraints are incorporated into the mutation operator. where Ci is the area of land-use type i in the competition zone and
If the land-use conversion on a cell (the cell value changes from 1 to its neighborhood on the land-use status quo map. Moreover, Di is
0 if it is in window L and from 0 to 1 if it is in window H) violates any the area of land-use type i in the competition zone on its optimized
of the spatial constraints, the conversion is withdrawn. This cell is spatial layout map. The land-use type with the highest PD is
not considered in the moving process. The number of the ‘‘select selected as the new land-use type of the competition zone. All cells
and move’’ operations executed on one chromosome in one itera- in the competition zone are allocated to the new land-use type,
tion of the GA is N2, which is also a model parameter. Roulette wheel which is helpful for compacting land use in local areas.
selection is utilized as the selection operator. The appearance of the development-agriculture land competi-
tion is primarily due to the sporadic emergence of development
2.2. Definition of land-use competition zones land in large tracts of agricultural land. The sporadic emergence
of development land is unfavorable to the scale management of
The same land-use competition type typically occurs within agricultural land. The strategy for solving this competition type is
several contiguous grid cells. Therefore, we introduce the concept to convert development land to agricultural land. When the entire
of competition zones into the model to promote land-use coordi- neighborhood of a DA competition zone is agricultural land (see
nation. Based on an 8-neighbor structure, a competition zone is Section 2.2 for the definition of DA), the area of each agricultural
composed of cells that are contiguous with each other and have land type in the neighborhood on the land-use status quo map is
the same land-use competition type. According to the aforemen- calculated. The new land-use type of the competition zone is the
tioned competition classification, competition zones can also be one with the largest area. All cells in the competition zone are allo-
divided into three types: agricultural land (AA), development-agri- cated to the new land-use type.
culture (DA), and agriculture-development (AD) competition The agriculture-development land competition is solved using
zones. The structure of a competition zone and its corresponding negotiations between farmers and the government. On one hand,
neighborhood is shown in Fig. 4. it is necessary for the government to expropriate some agricultural
land for economic development. The government must fully con-
2.3. Mechanism for land-use coordination sider the location and the developmental value of the agricultural
land for expropriation and the compensation to the land-lost farm-
Different coordination strategies are adopted for different land- ers. On the other hand, agricultural land is the primary source of
use competition types. The agricultural land competition and the farmers’ income. Their life will be largely affected if they lose their
 Y. Liu et al. / Computers, Environment and Urban Systems 49 (2015) 1–14 5

 u
land. In the negotiation process, farmers safeguard legitimate SD
V ¼ kþ  Q  S; ð7Þ
rights and personal interests. Snei

2.3.1. A game model simulating negotiations between farmers and the Rn ¼ ð1  eÞ  V  C n : ð8Þ
government
A game model simulating negotiations between farmers and the In Formulas (7) and (8), Snei is the area of the neighborhood, SD is the
government is constructed. Whether the agricultural land in an AD area of the development land in the neighborhood, k and u are the
competition zone is expropriated is determined by the result of the influence coefficients of the location conditions (0 < k < 1, u > 1),
game model (see Section 2.2 for the definition of AD). If the AD and e represents the cost the government should bear to make
competition zone is expropriated, all cells in the zone are con- the competition zone ready for development (0 < e < 1). The former
verted to development land. We execute the game model sepa- component of the tuples in Fig. 5 is the farmers’ income; the latter
rately for each AD competition zone. Details regarding the game component is the government’s net revenue.
model processes in an AD competition zone are provided below. The negotiations between the farmers and the government are
In the game model, farmers and the government take action in modeled as a dynamic game. The game model hypothesizes that
turn. When the government plans to expropriate the agricultural both the farmers and the government are rational. The farmers
land in an AD competition zone, the farmers must decide whether accept the land expropriation if and only if the compensation Cn
to agree to the expropriation after comparing their agricultural exceeds the income I from agricultural production. The govern-
income with the compensation. If the farmers reject the expropri- ment gives up the land expropriation if and only if the compensa-
ation, the government can choose to negotiate with the farmers or tion to the farmers becomes sufficiently large such that the profit
simply give up. When negotiations are adopted, the farmers must Rn becomes negative. In the game model, both the farmers and
decide whether to agree to the new compensation. If the farmers the government are clear that it is time for each side to make a
reject the expropriation, the government either renegotiates with decision; they each even know the exact node where the decision
the farmers or gives up. The process continues back and forth. is made. This property is known as complete and perfect informa-
The two sides can play the game for a long time; however, the tion (Rasmusen, 2001). With this property, backward induction can
game cannot go on indefinitely. More specifically, let N be the larg- be utilized to find the sub-game perfect Nash equilibrium of the
est number of rounds for the game. If the farmers reject the expro- game (Zhang, 2004). Each participant in dynamic games has his
priation in the Nth round of the game, the government will choose own strategy set, and a strategy is a complete action guide for
to give up the expropriation. The entire game process can be rep- the participant. The word ‘‘complete’’ means that a strategy of a
resented through a game tree (Hui & Bao, 2013). An example of a participant can provide advice to the participant on how to act at
game tree in which N equals 3 is shown in Fig. 5. all his decision nodes in the game tree. The sub-game perfect Nash
The following formula is applied to calculate the farmers’ total equilibrium is a strategy combination that indicates the strategies
income I from agricultural production in the coming M years: participants should adopt. At each decision node of the farmers, the
X strategy for the farmers indicated by the sub-game perfect Nash
I¼M S  Pi  F i ; ð5Þ
equilibrium chooses the action that brings a higher profit between
i
agreeing and disagreeing with the land expropriation. Similarly, at
where S is the area of the AD competition zone, Pi is the percentage each decision node of the government, the indicated strategy for
of agricultural land-use type i in the competition zone, and Fi is the the government chooses the action that brings a higher profit
annual income per unit area. between expropriating and not expropriating the land. According
Let T be the initial compensation. As the game continues, the to the indicated strategies, a path from the root node of the game
compensation to the farmers increases; however, the incremental tree to one of the leaf nodes can be found. The leaf node of the path
amount of compensation in one round gradually decreases. This gives the answer to the question of whether the competition zone
change is modeled using a decreasing geometric progression. The should be expropriated for development. The corresponding profits
first term is W, while the common ratio is q. The farmers’ compen- of the government and the farmers are also provided.
sation amount is S  T after the 1st round of the game, S  T + W Although farmers may have incomplete information about the
after the 2nd round, and S  T + W + W  q after the 3rd round. government’s profit in real-world negotiations, the game model
The farmers’ compensation after the nth round is calculated using constructed in the study is still a dynamic game of complete and per-
the following formula: fect information. The reason for this is that the farmers do not utilize
their information about the government’s profit to maximize their
W  ð1  qn1 Þ
Cn ¼ S  T þ : ð6Þ own profit in the dynamic game. According to the government’s
1q
profit, the government can be classified as two types: a high profit
In Formula (6), W = p  S  T and 0 < p, q < 1. government and a low profit government. For different types of
The government determines the lease price of the AD competi- the government, the farmers can put forward different compensa-
tion zone according to its future usage, e.g., industrial develop- tion demands. The government can also provide different compen-
ment, commercial and residential development, or development satory standards in consideration of its profit. If the farmers utilize
for public interests. The lease price differs widely among these their incomplete information about the government’s profit to max-
usages. In this study, we cannot foresee the future usage of the imize their own profit, the negotiations between the farmers and the
competition zone. Therefore, an average lease price for AD compe- government cannot be modeled as a dynamic game of complete
tition zones is introduced. We modify the average lease price to and perfect information. A dynamic game of incomplete informa-
obtain the final lease price of the competition zone based on loca- tion may be more suitable for the negotiations. The incompleteness
tion conditions. The percentage of the development land in the of the farmers’ information means that they only have the prior
neighborhood of the competition zone is utilized to evaluate the probability of the government’s type and do not know which type
location conditions. A high percentage indicates a high lease price. the government belongs to. After knowing which compensatory
If Q is assumed to be the average lease price and V is the govern- standard the government selects, the farmers can determine the
ment’s revenue from land leasing, the government’s net revenue posterior probability of the government’s type using Bayes’ theo-
Rn if the agricultural land is obtained in the nth round of the game rem. The dynamic game of incomplete information is more compli-
can be calculated using the following formulas: cated than the complete and perfect information. The sub-game
6 Y. Liu et al. / Computers, Environment and Urban Systems 49 (2015) 1–14

Fig. 5. A game tree representing negotiations between farmers and the government.

perfect Nash equilibrium and backward induction are not applicable The land-use data of the study area at a 1:10,000 scale were
to the dynamic game of incomplete information. Applying the derived from the second nationwide land survey. These data were
dynamic game of incomplete information to land-use coordination converted to a 546  516 grid with a resolution of 25  25 m. Rel-
requires future examination. evant yearbooks, suitability evaluation maps, and topographic
The real government in a Chinese town is not merely satisfied maps were also collected. Both land attributes and the land-use
by obtaining a positive profit in land expropriation. The profit from status quo were considered in the suitability evaluation. The
land expropriation is a considerable addition to local fiscal revenue results of the suitability evaluation were divided into four grades:
(Zhou, 2007). The government does their best to maximize the highly suitable (100), suitable (80), marginally suitable (60), and
profit to support local economic development. Occasionally, the unsuitable (20). This study primarily considers the spatial optimi-
government delays the compensation payment and even skimps zation of (1) farmland, (2) gardens, (3) forests, and (4) development
or embezzles the compensation to obtain more profit (Zhao, land. Paddy fields and upland fields are classified as farmland. Gar-
2009). Farmers will resist the infringement of their legitimate dens are comprised of orchards and tea plantations. Development
interests through various means. Negotiations with the farmers land includes rural settlements, the urban district, and land for
are an effective way for the government to handle the conflicts mining and industry. Transportation land, hydraulic construction
from land expropriation. Moreover, negotiations help the govern- land, water, and scenic areas remain unchanged in the
ment avoid the farmers’ violent struggle and appeal to higher optimization.
authorities for help, which may introduce additional losses to the
government (Zhao, 2009). In reality, farmers are typically in a weak 4. Results and discussion
position in the game between farmers and the government. There-
fore, the game model constructed in this study pays more attention 4.1. Spatial layout optimization
to the protection of the farmers’ interests.
The parameter values for the spatial layout optimization of each
3. Study area and data land-use type are listed in Table 1. A unary linear regression was
utilized to predict the population in the target year, i.e., 2020,
Gaoqiao Town is located in Zhejiang Province, eastern China according to the data from 1996 to 2005. The minimum farmland
(Fig. 6). The town covers an area of 104.03 km2 and has a popula- area per capita was calculated according to the following formula
tion of approximately 58,000 people. There are a variety of land- (Cai, Fu, & Dai, 2002):
forms within the town, e.g., low mountains, hills, basins, and Gr
plains. Forests make up approximately 64.2% of the town, 10.98% Sper capita ¼b ; ð9Þ
Pqk
is development land, and 16.3% is farmland. Gaoqiao Town has
been experiencing high-speed economic development, and simul- where b is the self-sufficiency rate of grain. The government plans
taneously some undesirable phenomena have appeared, e.g., to raise the rate to 80% in the target year. Moreover, Gr is the grain
excessively rapid expansion of development land and conflicts consumption per capita. From 1996 to 2005, the grain consumption
between farmland and development land. per capita declined continuously. For the target year, Gr was set to
 Y. Liu et al. / Computers, Environment and Urban Systems 49 (2015) 1–14 7

Fig. 6. The study area.

Table 1
The parameter values for the spatial layout optimization of each land-use type (the unit of the numbers in the parentheses is cell).

Current area Minimum area Maximum area w1 w2 Smin Smax Cmin

Farmland 16.3% (27,067) 14.3% (23,780) – 0.5 0.5 20 100 260.28
Gardens 6.62% (11,020) 4.1% (6818) – 0.5 0.5 20 100 361.12
Forests 64.2% (10,6751) 60.7% (100,939) – 0.5 0.5 20 100 380.51
Development land 10.98% (18,265) 10.98% (18,265) 12.77% (21,235) 0.3 0.7 20 100 253.11

the value for 2005. Furthermore, P is the grain yield per hectare. Table 2
From 1996 to 2005, the grain yield per hectare exhibited a fluctuat- The parameter values of the GA.
ing yet increasing trend. The average grain yield per hectare for the
Population size Iteration Crossover rate CN N1 Mutation rate N2
ten-year period was 5499.42 kg/h m2; for the target year, P was set
100 10,000 0.9 15 100 0.9 300
to the 2005 value (7053.1 kg/h m2). Additionally, q is the ratio of the
sown area to the total area of farmland; k is the multiple cropping
index. We set q and k to their respective ten-year averages using the
assigned a higher weight. Details regarding the normalization of
period from 1996 to 2005. The minimum farmland area in Table 1 is
objective functions in the spatial layout optimization of farmland
the product of the population in 2020 and the minimum farmland
are provided below. The same normalization methods were also
area per capita. The minimum garden and forest areas were set to
applied for gardens, forests, and development land. The suitability
their respective minimum values between 1996 and 2005. The
value of farmland ranges from 20 to 100 (see Section 3). Therefore,
growth rate of the urban district of Gaoqiao Town from 1996 to
the maximum value of objective S (Smax) was set to 100 in the spa-
2000 was too fast. The urban district area in the target year (i.e.,
tial layout optimization of farmland; the minimum value of objec-
2020) was predicted according to the data from 2001 to 2005 using
tive S (Smin) was set to 20. The following formula was used to
unary linear regression. The area of rural settlements and the area
normalize the values of objective S:
of land for mining and industry were individually predicted accord-
ing to the data from 1996 to 2005. The maximum area of develop- S x  xmin
ment land in Table 1 was set to the sum of the predicted areas of the f norm ðxÞ ¼ ð10Þ
xmax  xmin
urban district, rural settlements, and land for mining and industry.
The minimum area was set to the area of the existing development We executed the GA only using objective C to optimize the spatial
land. A penalty method was adopted to handle the area constraints. layout of farmland and to estimate the minimum value of objective
The chromosomes which violated the area constraints were elimi- C (Cmin). However, it is challenging to estimate the maximum value
nated in the iterative process. Two spatial constraints were consid- of objective C. Therefore, the following formula was utilized to nor-
ered in the execution of the mutation operators. First, the land-use malize the values of objective C:
type of an area where the gradient is greater than 25° should not be xmin
C
farmland due to the conservation of soil and water. Second, the f norm ðxÞ ¼ : ð11Þ
x
urban district cannot be converted to farmland, gardens, or forests
due to the large expense of this conversion. The parameter values of the GA are shown in Table 2. The optimiza-
No a priori knowledge is available for determining the objective tion problem in this study involves hundreds of thousands of vari-
weights. Objective S and objective C were given equal weights in ables. The GA must perform several generations to determine a
the spatial layout optimization for farmland, gardens, and forests. better solution because of the extensive search space. It is impossi-
For development land, the objective C was emphasized and ble to greatly improve the quality of chromosomes in one step of
8 Y. Liu et al. / Computers, Environment and Urban Systems 49 (2015) 1–14

Fig. 7. The values of the objective functions before and after the genetic optimization.

combinations of the parameters CN and N1 were utilized to optimize

the spatial layout of each land-use type. There are fluctuations in
the fitness values of the optimization results of the same land-use
type. The optimization result of farmland and the optimization
result of development land were both satisfactory when CN was
set to 15 and N1 was set to 100.
The model was implemented in C++ on the Windows platform
(Microsoft Corporation). An open-source library (Geospatial Data
Abstraction Library, GDAL) was utilized to access spatial data.
The commercial GIS software (ArcGIS Desktop 9.3, Environmental
Systems Research Institute) preprocessed spatial data and visual-
ized the results of the model. The spatial layout optimization of
farmland required approximately 4.5 h on a Lenovo desktop com-
puter with a Pentium(R) Dual-Core CPU E5800 @3.20 GHz and
2 GB RAM; 4.1 h, 4.9 h, and 4.2 h were required for gardens, forests,
and development land, respectively. The land-use coordination
required only 40 s.
Currently, the land-use suitability of the study area is high;
Fig. 8. Convergence curves of the genetic optimization.
therefore, the goal of the genetic optimization is to improve the
land-use compactness while minimizing suitability decreases.
The values of the two objective functions before and after the opti-
the iteration. Therefore, the number of iterations was set to a large mization are shown in Fig. 7. The suitability of farmland and devel-
value. The mutation operation in this study plays an effective role in opment land decreases slightly; however, the compactness of the
the land-use spatial optimization. The mutation rate and N2 were two land-use types is largely improved. The compactness of gar-
also set to large values, which contributed to reducing the iteration dens is slightly improved. For forests, both the compactness and
time. CN and N1 were determined by trial and error. Several value the suitability exhibit nearly no changes. The spatial optimization

Fig. 9. Tradeoff curves of objective function values (normalized) for different weight distributions.
Y. Liu et al. / Computers, Environment and Urban Systems 49 (2015) 1–14 9

Fig. 10. The spatial distribution of various land-use competitions.

10 Y. Liu et al. / Computers, Environment and Urban Systems 49 (2015) 1–14

Fig. 10 (continued)

process can also be reflected by the convergence curves (Fig. 8).

Similarly, the optimization effect of farmland and development land-use competitions can be seen in Fig. 10. Relevant statistical
land is better than for gardens and forests. information is listed in Table 3. The conflict between agricultural
Random initialization was tried in the spatial layout optimi- land and development land is obvious in the study area. The spatial
zation of farmland in this study. In each initial chromosome, distributions of the development-agriculture land competition and
farmland distributes randomly in the study area and the farm- the agriculture-development land competition have many over-
land area is 27,067 (Table 1). Random initialization was also laps. Most of the overlaps are concentrated in the southeastern
tried for development land. The relevant convergence curves portion of the study area, especially in the border areas of farmland
are shown in Fig. 8. The random initial chromosomes are of poor and development land. However, the spatial distribution of the
quality, which has significant disadvantageous effects on the agricultural land competition is more dispersed.
quality of the optimization results. Random initialization may The agriculture-development land competition is a conse-
not be suitable for land-use spatial optimization. In this study, quence of the rapid economic development and population
the spatial layout of farmland and the spatial layout of develop- growth. Farmland and development land are two dominant land-
ment land were optimized using multiple weight distributions. use types in the area marked by the rectangles in Fig. 10(a) and
The tradeoffs between the suitability objective and the compact- (b). Currently, the spatial layout of the development land in the
ness objective are illustrated in Fig. 9. The compactness objective area is very fragmented. Therefore, the area will be a hotspot for
is more susceptive to the weight distribution than the suitability land-use conflicts; the experimental results confirm the inference.
objective. The optimized value of the suitability objective linearly
increases as its weight increases. Simultaneously, the optimized 4.3. Land-use coordination
value of the compactness objective exhibits a nonlinear decreas-
ing trend. The knowledge of land-use planning is utilized in the land-use
The results of the genetic optimization indicate that the current coordination. The game model that simulates the negotiations
spatial layout of farmland and development land is partially irra- between farmers and the government targets the agriculture-
tional. Forests represent the dominant land-use type in the study development land competition. The parameter values of this game
area. The spatial layout of forests is compact even before the opti- model are listed in Table 4. The area constraints and the spatial
mization (Fig. 6). Therefore, forests do not have much potential for constraints are verified in the land-use coordination. The land-
the genetic optimization. use conversion of a competition zone is withdrawn if the conver-
sion violates any of the constraints. The annual income per unit
4.2. Land-use competitions area from agricultural land is relevant to the farmers’ management
of agricultural land. However, detailed data for the management of
After the spatial layout optimization of each land-use type, agricultural land in each AD competition zone is lacking and diffi-
there are 7918 cells within which land-use competitions occur. cult to obtain. Therefore, the annual income per unit area from
According to the above definition (Section 2.2), 1940 competition farmland is assumed to be identical for all of the AD competition
zones are generated. The spatial distribution of various zones. The same assumption is applied to the annual incomes
 Y. Liu et al. / Computers, Environment and Urban Systems 49 (2015) 1–14 11

per unit area from garden and forest areas. We estimated the revenue from recent land leases and the total area of the leased
values of the three parameters based on the yearbooks. The aver- land. The coefficients k and u have a substantial effect on the loca-
age lease price was calculated according to the government’s total tions of new development land. When k decreases and u increases,
the agricultural land surrounded by development land can be more
Table 3 easily converted to non-agricultural uses. After several experi-
Statistical information regarding land-use competitions.
ments, k and u were set to 0.5 and 4, respectively. Numerous
Agriculture- Development- Agricultural rounds are advantageous for farmers to express their compensa-
development agriculture land tion demand. Therefore, N was set to 10. To protect the farmers’
land competition land competition competition
rights and interests, M was set to 30. The maximum compensation
Cell 3134 1876 2908 amount stipulated by the Land Administration Law is 30 times the
Competition 654 508 778
annual income from agricultural land. Typically, farmers cannot
zone
obtain this much compensation in reality. However, in this study,

Table 4
The parameter values of the game model.

Annual income per unit area from farmland (RMBa/mub) 2000 Cost e 0.2
Annual income per unit area from gardens (RMB/mu) 3000 The largest number of rounds N 10
Annual income per unit area from forests (RMB/mu) 2500 Initial compensation T (RMB/mu) 35,000
Land-use term M (year) 30 p 0.4
Average land lease price Q (RMB/mu) 160,000 q 0.6
Influence coefficients k/u 0.5/4
a
RMB is Chinese currency. 1 RMB = about 1/6 U.S. Dollars.
b
Mu is an area unit in China. One mu = about 666.7 m2.

Table 5
Statistical information regarding the land-use coordination.

New land-use type Agricultural land Development-agriculture land Agriculture-development land

competition (778 zones) competition (508 zones) competition (654 zones)
Farmland 331 49 337
Gardens 199 3
Forests 248 3
Development land 0 453 317

Fig. 11. The spatial distribution of the new development land.

12 Y. Liu et al. / Computers, Environment and Urban Systems 49 (2015) 1–14

Fig. 12. The relationship between the area of new development land and the initial compensation.

Fig. 13. The final allocation solution.

the farmers can obtain more compensation than the stipulated The sporadic emergence of development land was more often
maximum amount through the game model. The parameters p located in farmland, which may have an unfavorable effect on
and q were empirically determined. Although the same set of the scale management of farmland.
parameter values was used for all AD competition zones, the farm- Furthermore, 317 of the 654 AD competition zones were con-
ers’ behavior was different in each zone. The farmers in different verted to development land using the parameter values listed in
AD competition zones put forward different compensation Table 4. The agricultural land surrounded by development land
demands according to their incomes from agricultural production. has a high priority to be expropriated for development if the influ-
The number of rounds in the game varies with the competition ence coefficients of location conditions are set to appropriate values.
zones; higher compensation demands correspond to more rounds. Therefore, the extensive expansion of development land can be
Statistical information regarding the land-use coordination is avoided (Fig. 11). 337 AD competition zones remained for agricul-
shown in Table 5. There are 778 competition zones within which tural use. The compensation demand of the farmers was not met
agricultural land competition occurred; the new land-use type is in 281 of these zones. In the remaining 56 AD competition zones,
farmland in nearly half of these zones (42.54%). The develop- the government was not satisfied with its profit and gave up land
ment-agriculture land competition occurred in 508 competition expropriation. With the other parameters fixed, the relationship
zones; however, only 55 of these zones (10.83%) were converted between the area of new development land and the initial compen-
to agricultural land. Most of the DA competition zones did not sation is shown in Fig. 12. When the government adopted a low
meet the requirements for the conversion from development land initial compensation, no agricultural land was converted to develop-
to agricultural land. Moreover, 49 of the 55 competition zones ment land. As the initial compensation increased, the area of new
were converted to farmland, three to gardens, and three to forests. development land increased, then remained steady at a relatively
 Y. Liu et al. / Computers, Environment and Urban Systems 49 (2015) 1–14 13

Fig. 14. The adjustment of local land-use patterns in the final solution.

Fig. 15. The objective function values of the final solution.

high level and ultimately decreased. The reason for this pattern is as 6.5%, 63.9%, and 12.0%, respectively, which meet the area constraints
follows. As the initial compensation increases (less than in Table 1. The experimental results confirm that the model is feasi-
50,000 RMB/mu), the number of cases of farmers rejecting land ble to undertake land-use spatial optimization.
expropriation tends to decrease. The area of new development land
increases. However, the government’s profit from land expropria-
tion is largely reduced if the initial compensation continues rising 5. Conclusion
(more than 50,000 RMB/mu). The government prefers to give up
land expropriation in this case, which causes the area of new devel- A land-use spatial optimization model is constructed in this
opment land to decrease. When the initial compensation was set to study, through coupling a genetic algorithm and game theory.
45,000 RMB/mu, the number of cases of farmers rejecting land The GA is repeatedly executed to separately optimize the spatial
expropriation was 82; however, when the initial compensation layout of each land-use type. The land-use status quo is overlaid
was increased to 50,000 RMB/mu, the number was 0. Therefore, with the optimization results to find local land-use competitions.
the government must balance the area of the expropriated land The concept of land-use competition zones is introduced. Using
and the compensation to farmers. In this study, a reasonable range the competition zones as the basic units, the model utilizes
for the initial compensation was from 35,000 RMB/mu to multi-stakeholder games and the knowledge of land-use planning
50,000 RMB/mu. to coordinate the local land-use competitions. The competitions
are divided into three types to promote the land-use coordination.
4.4. Final solution A dynamic game model of complete and perfect information is
constructed to solve the agriculture-development land competi-
The final solution generated by the model is shown in Fig. 13. The tion. The other competition types are solved using guidance from
spatial layout of agricultural land and development land in the solu- land-use planning knowledge.
tion is more compact relative to the land-use status quo (Fig. 14A In order to verify the validity of the model, the model was uti-
and C). Moreover, the sporadic emergence of development land in lized to optimize the land resource allocation in Gaoqiao Town,
agricultural land is substantially decreased (Fig. 14 B). Overall, the Zhejiang Province. The sporadic emergence of development land
land use of the study area became more rational. The objective func- in agricultural land is substantially decreased in the final solution.
tion values of the final solution are shown in Fig. 15. The land-use The land-use pattern of the solution is more rational than that of
suitability is well maintained. Although the compactness was the land-use status quo. The dynamic game model of complete
damaged during the land-use coordination, the land-use pattern of and perfect information balanced the interests of farmers and the
the solution remains better than that of the land-use status quo, government in solving the agriculture-development land competi-
especially for the development land. The percentages of farmland, tion, and helped the government find suitable locations of the new
gardens, forests, and development land in the solution are 15.7%, development land. Through revealing the relationship between the
14 Y. Liu et al. / Computers, Environment and Urban Systems 49 (2015) 1–14

area of new development land and the initial compensation, the Lee, C.-S. (2012). Multi-objective game-theory models for conflict analysis in
reservoir watershed management. Chemosphere, 87, 608–613.
dynamic game model provided a referable range for the initial
Li, X., Shi, X., He, J., & Liu, X. (2011). Coupling simulation and optimization to solve
compensation. The incorporation of the dynamic game model planning problems in a fast-developing area. Annals of the Association of
improves the ability of the land-use spatial optimization model American Geographers, 101, 1032–1048.
to support decision-making for regional land-use planning. Li, X., & Yeh, A. G.-O. (2000). Modelling sustainable urban development by the
integration of constrained cellular automata and GIS. International Journal of
Local land-use competitions are a complicated problem involv- Geographical Information Science, 14, 131–152.
ing interests of multiple stakeholders. A dynamic game model of Li, X., & Yeh, A. G.-O. (2002). Neural-network-based cellular automata for
incomplete information for the agriculture-development land simulating multiple land use changes using GIS. International Journal of
Geographical Information Science, 16, 323–343.
competition can be constructed in future research. Game models Li, X., & Yeh, A. G. O. (2005). Integration of genetic algorithms and GIS for optimal
for the other land-use competition types can also be considered. location search. International Journal of Geographical Information Science, 19,
581–601.
Liu, X., Li, X., Shi, X., Huang, K., & Liu, Y. (2012). A multi-type ant colony optimization
Acknowledgement (MACO) method for optimal land use allocation in large areas. International
Journal of Geographical Information Science, 26, 1325–1343.
This study was supported by the National Natural Science Foun- Liu, Y., Liu, D., Liu, Y., He, J., Jiao, L., Chen, Y., et al. (2012). Rural land use spatial
allocation in the semiarid loess hilly area in China: Using a particle swarm
dation of China (Grant No. 41371429). optimization model equipped with multi-objective optimization techniques.
Science China Earth Sciences, 55, 1166–1177.
References Liu, W., Sun, Y., Gu, S., & He, J. (2002). Game analysis for conflicts in water resource
allocation. Systems Engineering-Theory & Practice, 1, 16–25.
Liu, Y., Wang, H., Ji, Y., Liu, Z., & Zhao, X. (2012). Land use zoning at the county level
Aerts, J. C., Eisinger, E., Heuvelink, G., & Stewart, T. J. (2003). Using linear integer
based on a multi-objective particle swarm optimization algorithm: A case study
programming for multi-site land-use allocation. Geographical Analysis, 35,
from Yicheng, China. International Journal of Environmental Research and Public
148–169.
Health, 9, 2801–2826.
Arthur, J. L., & Nalle, D. J. (1997). Clarification on the use of linear programming and
Liu, Y., Wang, L., & Long, H. (2008). Spatio-temporal analysis of land-use conversion
GIS for land-use modelling. International Journal of Geographical Information
in the eastern coastal China during 1996–2005. Journal of Geographical Sciences,
Science, 11, 397–402.
18, 274–282.
Balling, R. J., Taber, J. T., Brown, M. R., & Day, K. (1999). Multiobjective urban
Morio, M., Schädler, S., & Finkel, M. (2013). Applying a multi-criteria genetic
planning using genetic algorithm. Journal of Urban Planning and Development,
algorithm framework for brownfield reuse optimization: Improving
125, 86–99.
redevelopment options based on stakeholder preferences. Journal of
Bennett, D. A., Xiao, N., & Armstrong, M. P. (2004). Exploring the geographic
Environmental Management, 130, 331–346.
consequences of public policies using evolutionary algorithms. Annals of the
Aerts, J. C., & Heuvelink, G. B. (2002). Using simulated annealing for resource
Association of American Geographers, 94, 827–847.
allocation. International Journal of Geographical Information Science, 16, 571–587.
Brookes, C. J. (1997). A genetic algorithm for locating optimal sites on raster
Porta, J., Parapar, J., Doallo, R., Rivera, F. F., Santé, I., & Crecente, R. (2012). High
suitability maps. Transactions in GIS, 2, 201–212.
performance genetic algorithm for land use planning. Computers, Environment
Brookes, C. J. (2001). A genetic algorithm for designing optimal patch configurations
and Urban Systems.
in GIS. International Journal of Geographical Information Science, 15, 539–559.
Rasmusen, E. (2001). Games and information: An introduction to game theory. Wiley-
Cai, Y.-L., Fu, Z.-Q., & Dai, E.-F. (2002). The minimum area per capita of cultivated
Blackwell.
land and its implication for the optimization of land resource allocation. ACTA
Sadeghi, S., Jalili, K., & Nikkami, D. (2009). Land use optimization in watershed scale.
GEOGRAPHICA SINICA-CHINESE EDITION, 57, 127–134.
Land Use Policy, 26, 186–193.
Cao, K., Batty, M., Huang, B., Liu, Y., Yu, L., & Chen, J. (2011). Spatial multi-objective
Santé-Riveira, I., Boullón-Magán, M., Crecente-Maseda, R., & Miranda-Barrós, D.
land use optimization: Extensions to the non-dominated sorting genetic
(2008). Algorithm based on simulated annealing for land-use allocation.
algorithm-II. International Journal of Geographical Information Science, 25,
Computers & Geosciences, 34, 259–268.
1949–1969.
Santé-Riveira, I., Crecente-Maseda, R., & Miranda-Barrós, D. (2008). GIS-based
Cao, K., Huang, B., Wang, S., & Lin, H. (2012). Sustainable land use optimization
planning support system for rural land-use allocation. Computers and Electronics
using Boundary-based Fast Genetic Algorithm. Computers, Environment and
in Agriculture, 63, 257–273.
Urban Systems, 36, 257–269.
Stewart, T. J., Janssen, R., & van Herwijnen, M. (2004). A genetic algorithm approach
Chen, J. (2007). Rapid urbanization in China: A real challenge to soil protection and
to multiobjective land use planning. Computers & Operations Research, 31,
food security. Catena, 69, 1–15.
2293–2313.
Chuvieco, E. (1993). Integration of linear programming and GIS for land-use
Wu, X., & Grubesic, T. H. (2010). Identifying irregularly shaped crime hot-spots
modelling. International Journal of Geographical Information Science, 7, 71–83.
using a multiobjective evolutionary algorithm. Journal of Geographical Systems,
Ducheyne, E., De Wulf, R., & De Baets, B. (2006). A spatial approach to forest-
12, 409–433.
management optimization: Linking GIS and multiple objective genetic
Wu, Y., Wu, C., & Shen, L. (2005). Modeling the decision-making using game theory
algorithms. International Journal of Geographical Information Science, 20,
in monitoring land-use practice in China. Systems Engineering – Theory &
917–928.
Practice, 9, 65–70.
Fotakis, D., & Sidiropoulos, E. (2011). Combined land-use and water allocation
Xiao, N. (2008). A unified conceptual framework for geographical optimization
planning. Annals of Operations Research, 1–17.
using evolutionary algorithms. Annals of the Association of American Geographers,
Fotakis, D. G., Sidiropoulos, E., Myronidis, D., & Ioannou, K. (2012). Spatial genetic
98, 795–817.
algorithm for multi-objective forest planning. Forest Policy and Economics.
Zhang, W. (2004). Game theory and information economics. Shanghai People’s
Goldberg, D. E. (1989). Genetic algorithms in search, optimization, and machine
Publishing House.
learning. Addison-Wesley.
Zhang, Y., Li, A., & Fung, T. (2012). Using GIS and multi-criteria decision analysis for
Holzkämper, A., & Seppelt, R. (2007). A generic tool for optimising land-use patterns
conflict resolution in land use planning. Procedia Environmental Sciences, 13,
and landscape structures. Environmental Modelling & Software, 22, 1801–1804.
2264–2273.
Hui, E., & Bao, H. (2013). The logic behind conflicts in land acquisitions in
Zhao, D. (2009). The interactive relationship between farmers, local governments
contemporary China: A framework based upon game theory. Land Use Policy, 30,
and the state during the course of land acquisition. Sociological Studies, 2, l.
373–380.
Zhou, F. (2007). The role of government and farmers in land development and
Lautenbach, S., Volk, M., Strauch, M., Whittaker, G., & Seppelt, R. (2013).
transfer. Sociological Studies, 1, l.
Optimization-based trade-off analysis of biodiesel crop production for
Zhou, M., & Sun, S. (1999). Genetic algorithms: Theory and applications. National
managing an agricultural catchment. Environmental Modelling & Software, 48,
Defense Industry Press, Beijing.
98–112.