Simulation of farmer decision on land use conversions using decision tree method in Jiangsu Province , China

Understanding farmer decisions on land use conversions (LUC) in rural areas has significant importance to understand and predict the patterns of land use changes in China. Many methods have been developed to search for the influencing factors on land use changes at farm household level. However, these methods have difficulty in evaluating the intertwined influences between factors and achieving farmer decision rules on LUC. Taking three regions located in Jiangsu Province as the study areas, the present work proposed a data mining method, classification and regression tree (CART), to simulate farmer decisions on LUC at farm household level. The accuracy of the simulated LUC by CART was above 85.00%, indicating that the proposed method could be used to simulate farmer decisions. The simulation results also showed that farmer decision rules on LUC presented regional characteristics. In Jiangsu Province, 20 rules were inferred for LUC using 10 factors which were related to the household resources, land market, and so on. These factors were ranked in the decreased importance on LUC as labor transfer, land market, location, resources of household, and characteristics of household. In Rudong County, the which were related factors with LUC included land market, labor transfer, and household resource, while in Changshu County, the additional factor of location was involved. Additional key words: classification and regression tree, determinants of rural land conversions, farm household level, regional differences.


Introduction
Since the economic reforms and an open door policy in 1978, China has been experienced rapid urbanization and industrialization, especially in the coastal regions, such as Yangtze River Delta region (Xu, 2004) and Pearl River Delta region (Weng, 2002;Li and Yeh, 2004).The accelerated industrialization and urbanization have greatly affected land use changes through increasing private enterprises and industry plants (Wu et al., 2004).Although there have been numerous studies analyzing land use conversions (LUC) in eastern and coastal China, most of which focused on urban fringe areas (Weng, 2002;Tan et al., 2005).In fact, many industrial plants have been boosted in rural areas (Xu, 2004), which leaded to nonfarm jobs in rural and small town enterprises growing rapidly.The transferred farm labor forces would influence the development of enterprises, and then influence LUC.Therefore, it is important to monitor land use changes in rural areas.
Jiangsu Province, located in the Yangtze River Delta, is one of the most developed provinces in China.In the recent three decades, heavily land use changes have been occurred, which focused many studies to investigate the land use changes and their driving forces in this area.For example, Ho and Lin (2004) believed that the cause to convert farm land to nonagricultural use were rural-urban migration, rapid economic growth, and increased investments in roads; Streets et al. (1995) used remote sensing data and social and economic data revealed the rapid growth of urban centers, commensurate declines in water surface area, and changing patterns of agriculture from 1976 to 1984 in southern Jiangsu Province; Xiao et al. (2007) studied the land use change patterns and their influence on population resource environment development system, and they concluded that the land use changed most heavily in the southern areas, and central and northern areas were followed in Jiangsu Province.These studies gave the insight to LUC and their driving forces.However most of them did at the macro level, not at farm household level.Farmers are the operators of LUC, thus it is significant important to explore driving forces of land use and land cover changes at farm household levels (Alisson et al., 2005).
A wide range of factors affects farmer decisions on LUC.These factors include social, biophysical, economic and institutional components (Moran, 1981;Walker and Homma, 1996).To perform such complex processes, the remotely sensed data and household social surveys have often been used.Remotely sensed imageries can provide a continuous spatial coverage of land cover changes, but the data processing and field validating requirements necessary to obtain a higher order land cover classification are considerable (Evans et al., 2001).While it is not easy to identify more concrete land use types from remotely sensed imageries because of the limited spectral and spatial resolution (Lillesand and Kiefer, 2000).Moreover, remote sensing techniques leave completely unaddressed the forces underlying farmer decision behind these trends (Pichón, 1997).Household social survey requires large teams of researchers due to travel time requirements and the complexity of the survey instruments, but it can provide indepth information at the level where land-use decisions are made.In this study, the analysis on driver forces for LUC was conducted based on household social survey data.
In the past decades, substantial advances have been made by developing a wide range of analytic tools to identify and simulate the driver forces for land use changes.The descriptive statistic methods (Odihi, 2003), linear regression techniques (Coxhead et al., 2002), Probit model (Alix-Garcia et al., 2005;Fisher and Shively, 2005), Logit model (Thapa Keshari et al., 1996;Adesina et al., 2000;Herath and Takeya, 2003;Tasser et al., 2007), Tobit model (Godoy et al., 1997), and so on, were often used in land use changes.These studies gain insight in the factors that influence land use decisions, and provide information about decision-making processes and human behaviors.However, in general, they have difficulty in evaluating the intertwined influences between these factors, especially intertwined effects caused by more than tow factors.They also have difficulty in explaining the distinguishing reaction of decision makers with different features defined by index value, because these analysis techniques are based on marginal analysis principle (Zhong, 2007).For example, young farmers and old farmers have weak motivation to change land use, and middle-aged farmers maybe have strong propulsion to convert land uses, however these methods mentioned above have difficulty to handle such situation.
The influencing factors could be used to simulate LUC, based on the method of decision tree analysis (DTA).The basic concept of a decision tree is to split a complex decision into several simpler decisions, which may lead to a solution that is easier to interpret.This method is especially to resolve complicated things involved of many intertwined influencing factors.DTA is increasingly being used in thematic mapping from remotely sensed data and habitat modeling in ecology (Zhou et al., 2003;Zhang, 2006;Zhang et al., 2008).However, to our knowledge, the potential of DTA to understand the mechanism of LUC has received little attention.
Taking Jiangsu Province as the study area, the broad goals of this paper are to explore the intertwined influences of socio-economic, demographic, biophysical, and geographical variables on land use conversion decision (LUCD) by farmers, and to find LUCD rules for predicating the LUC situations using decision tree analysis.

Study area
The study area, Jiangsu Province ranges from 116°18É to 121°57´E, and from 30°45´to 35°20´N, with an area of 106,700 km 2 (Fig. 1).Jiangsu is the lowest province in China, and most of the area is below 50 m in altitude.Low hilly region area accounts for about 14.30%, and low plain and water surface account for 85.71% of the study area.The northern subtropical monsoon climate dominates Jiangsu province year-round, with mean daily temperature of 13-16ºC, mean annual rainfall of 1060 mm, and mean annual non-frost period of 299 days, all of which are beneficial for agricultural production.
Jiangsu is one of the most developed provinces in China with great disparities among local regions within the province.According to the Statistical Yearbook of Jiangsu Province in 2007, the highest gross domestic product (GDP) per capita at county level reached to 141,064 yuan 1 (the base unit of the renminbi, people's currency in the mainland of the People's Republic of China), which gained 26.8 times of the lowest GDP per capita.The arable land accounted 45.10% of its total area.Orchard, grass, forest and the other farm lands accounted 2.90%, 0.02%, 3.10 and 12.40% respectively.The construction land and unused land accounted 16.90% and 17.60% respectively.From 1978 to 2006, farm land, particularly arable land, has been shrinking due to both urban sprawl and land requirements of villages, rural industries, and infrastructure.The arable land decreased by 16.9%, while urban settlements increased by 45.3%, rural settlements by 2.7%, industrial land by 73.8%, and communications and transport by 13.2%.
Jiangsu Province has been divided into three-fold divisions, namely developed southern Jiangsu (Sunan), moderate developed mid-Jiangsu (Suzhong) and poor northern Jiangsu (Subei) respectively.In our study, Changshu, Rudong and Tongshan County were selected to respectively represent the south, middle and north area of Jiangsu Province.According to the statistic yearbook of Jiangsu Province in 2007 (Table 1), Changshu County had the lowest total population, with the highest population density among the three counties.Changshu, with only 7.44% of the total population engaged in farming, provides most of the labor forces to produce commercial gains.While, in Rudong and Tongshan County, the percentages of population engaged in farming was respectively 21.99% and 47.12%.All of the GDP, GDP per capita, average wage of fully employed staff and worker and annual net income per capita of rural households in Changshu were the highest among the three counties, and they were listed at the lowest in Tongshan except annual net income per capita of rural households.While, Rudong had the highest percent of GDP from agriculture, Tongshan was ranked in the medium, and Changshu was listed at the last.
The crop land accounting for its agricultural land in Rudong, Tongshan, and Changshu County was 75%, 56% and 53% respectively, while its multiple crop index was 1.58, 1.56 and 1.38 respectively in 2006.The amount of fertilizer put into arable land was 5800, 4400 and 9000 Mg ha -1 respectively in the above three counties, while its crop production was 58900, 63900, and 55800 Mg ha -1 respectively.This indicated that the cultivated land use efficiency in Tongshan was lowest among the three counties.

Field surveys
One of the reasons for choosing Changshu, Rudong and Tongshan County to study is that we enjoyed particularly good access to villages in these counties.In Changshu, Tongshan, and Rudong County, three, three, and four towns were selected based on the distance from their capitals to these towns.Considering the village numbers in every town and the town areas, one or two villages were randomly selected to conduct farm household surveys.Totally, 12 villages were selected.The detailed information about the selected villages is listed in Table 2 and the locations of the villages are described in Figure 1.
In every village, about 20 to 40 farm households were randomly selected and interviewed in July 2006.
The data relating to farm households were obtained through structured interviews, informal discussions with village elders and local government leaders.In total, we obtained 343 questionnaires.However, some of them were incomplete or one questionnaire answered by different respondents, which were thought invalid.Finally, 329 copies were determined as valid.There were 89, 104, and 136 valid questionnaires in Changshu, Tongshan, and Rudong County respectively.
The following information was included in the questionnaires: agricultural production, resource of household, land uses, LUC in the recent 10 years, labor transfer, credit and saving, land market, locations, and so on.LUC in this study referred to the conversions determined by farmers themselves, not by the government.It should be noted here that some LUC, such as arable land to construction land, was illegal, but it in fact occurred.In total, 50 households converted their land uses in the three counties.Among them, 16 households converted arable lands to vegetable plots and orchard lands, 17 converted arable lands to fishery and aquaculture lands, and 17 converted arable lands to construction lands.Fisheries are very popular here because Jiangsu province is one of the famous "a land of fish and rice" due to its special climatic and land characteristics.

Theoretical framework and independent variables
The decision on land use at a given parcel is made by a profit-maximizing "operator" of the land, and the "operator" may be "a single person, household, or group of people in the case of common property ownership" (Nelson and Geoghegan, 2002).The underlying motivation for operator to covert one use to another use is assumed to be maximization of expected return over an infinite time horizon (Irwin and Geoghegan, 2001).The return of land uses is determined by the costs and benefits, so the operator makes the decision on LUC through comparing costs and benefits of alternative land uses (Gellrich et al., 2007).Based on this theoretical supposition and given that there are no cost of converting from one land use to another land use, a model describing the operator's decision of land use can be developed.
The parcel k will be converted from state i to state j in time Where v jt is defined as the net present value of the future stream of net returns to parcel k in state j at time t, v it is the net present value of the future stream of returns to parcel k in state j at time t, w jt is the future stream of benefits to parcel k in state j at time t, c jt is the future stream of cost under operator control all for land use j at parcel k, w it is the future stream of benefits to parcel k in state i at time t, and c it is the future stream of cost under operator control all for land use i at parcel k.
The underlying response function is expressed through: Where Y* is a dummy variable, X is a vector of covariates including benefit-related variables and costrelated variables.
Land use will be converted on parcel k, if Y* takes the value 1; otherwise, it will not be converted.According to the theoretical analysis of farmer decisions on land use choice, and the former studies in LUC (Moran, 1981;Walker and Homma, 1996;Zhong, 2007), factors influencing farmer decisions were selected.They represented the characteristics of households, resources of households, labor transfer, land market, credit and saving, and location respectively.The demographic structure indicative of the characteristic of households exerts significant effects on the prominence of land uses (Perz, 2001), and demographic structure changes have the most significant effects on land use and land cover change (Evans et al., 2001;Stephen et al., 2006;VanWey et al., 2007).Resources of household included the area of arable land, the area of total land including arable land, fishery, forest land and orchard land, and the number of labor force aged between 16 to 65 years old, which impose strong influence on land uses (Dogliotti et al., 2005).The labor transfer characteristics involved five factors: number of wage employees, number of selfemployees, number of off-farm employment (the sum of the number of wage and self employees), income from off-farm activities, and per capita income from off-farm activities.The opportunity for off-farm employment and the income from off-farm activities have been proven to strongly influence land use changes (Godoy et al., 1997;Perz, 2001;Shriar, 2002;Shively and Pagiola, 2004).The land market is also believed an important factor to influence LUC because market will determine what is economically optimal (Drozd and Johnson, 2004).However, farm land sale market is forbidden in the mainland of China, where only agricultural land leasing market is permitted by law.Therefore, the area for rent-in and rent-out arable land, and the rent-in fishery land are selected here.The relation between the credit and saving, and land conversions is often discussed, on which there exists contrary views (Angelsen, 1999).Location is considered to be one of the most important determinants of land use and land cover change (Verburg et al., 2004).In this study, the distance from the household to its county city and the local counties were selected to represent the location characteristics.

County
In total, 19 independent variables were selected in this study.The detailed information is described in 3. The correlation coefficients between independent variables above 0.5 will bring multicollinearity problem (Wang and Guo, 2001), and the independent variables should remove redundancy before performing the simulation processes (Zhong et al., 2008a,b).However, considering the ability of the decision tree model can process nonlinear problems, the correlation analysis is not conducted in this study.

Decision tree
This study attempts to simulate farmer decisions to pursuit the maximization of expected return through LUC at farm household level, based on a data mining approach, decision tree model.The basic concept of a decision tree is to split the complex decisions using most independent variables into several simpler decisions using conditional methods, which may lead to a solution that is easier to interpret.In a decision tree approach, the social-economic variables are the predictor variables whereas the LUC are referred to as the target variable.Several decision tree models have been developed to solve the environmental problems, such as ID3, CART, C4.5, See 5.0, Orange (Tooke et al., 2009).
Among them, the model of CART (classification and regression tree), proposed by Breiman et al. (1984), has been widely used because of their non-parametric nature, simplicity, flexibility, and computational efficiency (Friedl and Brodley, 1997).
In CART, a tree-structured decision space is estimated by recursively splitting the data at each node on the basis of a statistical test that increases the homogeneity of the training data in the resulting descendant nodes (McLachlan, 1992).The decision tree grows by means of the successive subdivision until a stage is reached when there is no significant increase in homogeneity with further nodes.At this stage, the node will not subdivide further and automatically becomes a terminal node.Training accuracy is used to evaluate the simulation results, which equals to the simulated LUC of the samples inferred by simulation rules dividing by their actual land use conversations.In this study, the CART analysis was conducted with the software S-PLUS 2000.
To perform the accuracy assessment on the process of CART, the training accuracy is used.Firstly, the rules for simulating farmer decision on LUC are inferred based on the social and economic factors and farmer decisions on the sample questionnaires, and then farmer decisions are simulated using the social and economic factors on the sample questionnaires.The training accuracy is calculated by the simulated farmer decisions and the investigated farmer decisions.The sum of training accuracy and misclassification error equals to 1.

Farmer decisions on land use conversions in Jiangsu Province
Using the 329 household questionnaires in Jiangsu Province, the CART model was used to simulate farmer decisions on LUC.The accuracy of the CART training process in rightly attributing LUC was 93.31%, indicating that the simulation results were acceptable.The simulated tree included 20 nodes and 10 independent variables, which were COU, ARL, AHW, AINA, AGE, TAL, DIS, TNE, INE, and NSE respectively (Figure 2).There were 20 branches to predict farmer decisions, among which 13 branches were for not convert land uses, and the other seven were for converting land uses.
CART provides the relative importance of independent variables on LUCD.As in our study, the other nine variables did not appear in the simulated decision tree.However, CART can not quantitatively assess the influence of these variables involved in the simulated tree on LUCD.To understand the quantitative importance of independent variables, the 239 training samples were used to evaluate the 10 variables of COU, ARL, AHW, AINA, AGE, TAL, DIS, TNE, INE, and NSE.The predication error and the number of terminal nodes of the constructed tree model when using different variable groups are listed in Table 4.
The numbers of terminal nodes for the 14 tree models ranged from 17 to 21, indicating these tree models have similar complicated structures.Variables related to labor transfers including AINA, TNE, INE, and NSE were most important on LUCD (omitting these variables increased misclassification error to 10.64%), followed by the land market factors (misclassification error increased to 9.73%), location characteristics (misclassification error increased to 9.42%), household resources (misclassification error increased to 8.51%) and the characteristics of households (misclassification error increased to 7.90%).If comparing the importance The mis-predication errors and the number of terminal nodes of the constructed tree model when using different variable groups in Changshu are listed in Table 5.The number of terminal nodes for the eight tree models ranged from eight to 13, indicating these tree models had similar complicated structures.Variables related to labor transfer including AINA and NSE, and household resource involved of NLA, TAL and ARA were most important on LUCD (omitting these variables increased misclassification error to 14.71%), followed by labor transfer factors (misclassification error increased to 12.24%) and location (misclassification error increased to 12.50%).If comparing the importance of single dependent variable on LUCD, it was concluded that NSE, DIS, TAL, and AINA had the strongest influence on LUCD, followed by APA and NLA and NLA.
of single dependent variables on LUCD, it was concluded that ARL, AHW and NSE had the strongest influence on LUCD, followed by TAL, TNE and INE, the factors of AGE, AINA, COU and DIS had the relative lower effect on LUCD.
Since there were social and economic differences among Rudong, Changshu, and Tongshan, the following analyses were conducted respectively in the three counties.In this study, only two households converted their land uses during the past 10 years in Tongshan County; and the other 48 households were located in Changshu and Rudong County.Therefore, the following analyses were conducted in Rudong and Changshu.

Farmer decisions on land use conversions in Changshu County
The 136 questionnaires were used to simulate farmer LUCDs in Changshu County.Among these investigated households, 24 converted land uses.The training accuracy by CART was 89.71%, indicating the simulated results were acceptable.The simulation tree for LUCD included 12 nodes and seven independent variables (Figure 3).The factors of DIS, AINA, TAL, ARA, AGE, and NSE were involved in the simulation results.The branches for predicting LUCD in Changshu were 12, among which seven were for not a Abbreviations: see Table 3. a Abbreviations: see Table 3.
Table 5. Mis-predication errors on LUCD using different independent variables in Changshu County

Farmer decisions on land use conversions in Rudong County
Using the 89 questionnaires in Rudong County, CART model was used to simulate farmer decisions on LUC at the farm household level.Among these investigated households, 24 converted land uses, and 65 did not convert their land uses.The accuracy of the CART training process in rightly attributing land use changes was 85.39%.The simulated tree included seven nodes and four independent variables, which were ARL, AINA, ARA, and INE respectively (Figure 4).There were three rules for changing land use, and the other four for keeping the same land use format.
Among the four independent variables involved in the simulated tree in Rudong County, ARA was considered the most important influencing factor on LUCD (omitting this variable increased misclassification error to 21.35%, Table 6).This demonstrated that the arable land directly affected farmer decisions on LUCD in Rudong County.The labor transfer variables of AINA and INE were also included in the simulated decision tree, indicating that the number of employment and the income by employment had strong influence on LUCD.If INE and AINA were not involved in the simulated tree model, the misclassification error increased to 22.47%.The factor ARL also showed effect on farmer LUCD.

Comparison of the simulated results by CART, Logit and Tobit model in Jiangsu Province
Based on the same questionnaires used in this study, Zhong et al. (2008a,b) simulated the influencing factors on LUC in Jiangsu Province using a Binary Logit Model and Tobit Model respectively.The simulation results showed that GEN, TAL, NSE, AHW, DIS and COU had strong influence on farmer LUCDs by Logit model, while an additional factor SAV was involved in the simulation results by Tobit.In this study, the five factors of TAL, NSE, AHW, DIS and COU also showed strong influence on farmer decisions, which was almost agreed with the work by Zhong et al. (2008a,b).
The COU in the three studies presented strong influence on farmer decisions on LUC, indicating that farmer LUCDs had regional characteristics.The COU was the root node in the simulated tree, and most of the braches to predict farmer decisions were for Changshu and Rudong County, because only two households converted their land uses in Tongshan County.The simulation branches on the two household decisions were cut during the simulation process, considering the balance between the cost and the efficiency (McLachlan, 1992).The reason for regional difference might be the social and economic differences.In Changshu County, there are many industries which provide many opportunities for employment; in Rudong County, the home workshops for texture are popular there; and in Tongshan County, there are not many industries and home workshops.
In the two work by Zhong et al. (2008a,b), DIS showed negative influence on LUC.The households leave more far to its county city, the land would not be converted uses.This was mainly because the distance to the county city nearly represented the distance to the land market (Zhong et al., 2008b).In this study, DIS had effect on farmer LUCDs under some conditions of COU, ARL, AWH, AINA, AGE, TAL, and INE, and it did not show the always negative or positive influence on LUCDs (Figure 2).Under the conditions mentioned above, the households would not convert land uses when DIS greater than 22.5 km; while the conditions combing different conditions of TAL and NIAN when DIS less than 22.5, the households would have different  a Abbreviations: see Table 3.
Table 6.Mis-predication errors on LUCD using different independent variables in Rudong County decisions on LUC.This is the merit of the DTA method.It could fully consider the intertwined interactions among the factors in the simulating process.The area of the contracted land by households showed positive influence on farmer decisions in the work by Zhong et al. (2008a,b).The more area of the contracted land by household, the more possibility of the LUC occurs.In this study, the influence of TAL on LUC would also determined by the other conditions of COU, ARL, AWH, AINA, AGE, DIS, INE and TNE.The influence of land market on LUCDs showed the area of the rent-in fishpond had positive influences (Zhong et al., 2008a,b).In this study, under some conditions of COU and ARL, the households would change their land uses when AWH was greater than 0.15 ha; while the condition of AWH smaller than 0.15 ha and other conditions would determine farmer decisions on LUC.The number of self-employees of the households showed positive influence on farmer LUCDs in Zhong et al. (2008a,b).In this work, similar results to Zhong et al. (2008a,b) were obtained under some conditions of COU, ARL, and AHW.When NSE was < 1, land use would not be converted; otherwise, it would.

Comparison of factors and rules for LUCD between local counties
It was obvious that farmer decisions on LUC were more complicated in Jiangsu Province than that in local counties (Figures 2, 3, and 4).Totally, 20 branches for simulating LUC using 10 independent variables were obtained in Jiangsu Province by CART.There were 12 rules with 7 variables in Chaungshu County, and 7 rules with 4 variables in Rudong County to simulate farmer LUCDs.The influencing factors on farmer decisions in Jiangsu Province included labor transfer, land market, location, resource and characteristics of household, while the characteristics of household were not involved in Changshu and Rudong County.
The simulated decision trees for farmer LUCDs in Changshu County had more complicated structure than that in Rudong County (Figures 3 and 4).In Rudong, the contracted arable land area by households, the total income from employment, the average income from off-farm activities, and the rentin land by households were the main factors to influence farmer decisions on LUC.In Changshu County, the three factors of resources of households, the number of self-employees, per capita income from offfarm activities, the area of rent-in land, and distance to its county city showed strong influence on farmer LUCDs.Comparison with the two simulation results in Changsu and Rudong County, the additional factor of location was involved in the simulated tree in Changshu.This demonstrated that farmer decisions on LUC were influenced by more complicated conditions in Changshu than that in Rudong County.

Further studies in simulating farmer decisions on land use conversions at farm household level
Although we tried to select enough influencing factors on LUC at farm household levels, there might be other factors to influence farmer decisions.For example, the government policy has strong influence on farmer decisions on LUC (Lichtenberg and Ding, 2008).This factor was not involved in our study because we have not enough data.In the further studies, the government policies would be involved in the simulation process.
Furthermore, it would be more useful to break down land-use changes to several categories surveys and identify the major changes from the surveys and investigate why these major changes happen.However, in our study area, there were only 16, 17, and 17 households converted their arable lands to vegetable plots and orchard lands, fishery and aquaculture lands, and construction lands.The samples were not enough when classifying the LUC.
Moreover, there are also some problems needed to further study.For instance, could the selection of villages be a potential problem for the findings about location which is important in one county but not in another?How the imbalance on the numbers of household that have changed and have not changed land use could affect the results?How many villages should be enough to study the problem of farmer decisions on LUC?

Conclusions
In this study, farmer decisions on LUC at farm household level were simulated by a decision tree model in Jiangsu Province.This model obtained reasonable accuracy for predicating farmer LUCDs of 93.31%, 89.71% and 85.39% in Jiangsu Province, Changshu County and Rudong County respectively.The simulation results also showed that farmer decisions presented regional characteristics.The decision tree structure in Jiangsu Province was more complicated than that in Changshu County, which was more complicated than that in Rudong County.There were 20 branches with 10 indepent variables, 12 branches with 7 variables and 7 branches with 4 variables in the simulated trees to predict farmer decisions on LUC in Jiangsu Province, Changshu County and Rudong County respectively.
Moreover, the decision tree model provides some insight into the social and economic variables that are most responsible for driving LUC.The factors influencing farmer LUCDs in Jiangsu Province were ranked as decreased importance as labor transfer, land market, location, resource of household, characteristics of household, according to the importance of the effect on LUCD.In Changshu County, the related factors on LUCD ranked as the decreased importance as NSE, DIS, TAL, AINA, ARA, NLA, and ARL; while in Rudong County, the 4 involved factors ranked ARA, AINA, ARL, and INE.

Figure 1 .
Figure 1.Spatial distribution of the investigated villages

Figure 2 .
Figure 2. Decision tree for simulating farmer decisions on land use conversions in Jiangsu Province

Figure 3 .
Figure 3. Decision tree for simulating farmer decisions on land use conversions in Changshu County

Figure 4 .
Figure 4. Decision tree for simulating farmer decision rules on land use conversions in Rudong County

Table 1 .
Social and economic conditions in Changshu, Rudong and Tongshan County a yuan is the base unit of the renminbi, people's currency in the mainland of the People's Republic of China; 1 yuan equals about €0.11.

Table 2 .
The surveyed villages and investigated samples

Table 4 .
Mis-predication errors on land use conversion decision using different independent variables in Jiangsu Province