Modeling of Segzi plain farmer's decision based on cultivation type using the multinomial logistic regression model

Complex interactions between human decision-makers and their biophysical environment can be observed in land-use systems. These complexities are due to the differences between biophysical and socio-economic variables. For Modeling of human decision-making, we need to know interactions between landscape, community, or ecosystems. In reality, humans make decisions using variety of strategies. We need to simplify the complex interaction between all individual agents and their environment by formulating an agent typology. In this paper, given the complexity of the decision-making in agent based models, the agricultural land use changes are simulated by the multinomial logistic regression model to determine the socio-economic and environmental factors. Also the proportional random rules use to implement the bounded rationality law for unique individual decision making. Considering the particular environmental conditions of the region and the severity of the risk of desertification, the economic, social, physical factors and the chemical parameters of soil along with other environmental factors in the decision-making processes were investigated.
Diverse data (including GIS and household data) were used for initializing the coupled human–landscape system and farmer household decision making simulations. GIS data consist of landscape agents (grid cell or patch): Land use/cover (based on Landsat 8), soil Physico-chemical properties (EC, SAR, PH (pH), texture, moisture), institutional variables (i.e. ownership, village territory), and topography. Household data consist of socioeconomic attributes: labour force, educational status, income structure, and land properties. They were derived from an intensive household survey conducted in Segzi plain in Isfahan province (central Iran) during the spring 2013. The agent-based decision-making method has been presented by Le (2005). For determining decision-making approach, a mechanism of livelihood groups dynamics were considered as follows. At first, principal component analysis (PCA) was used to identify key factors differentiating household characteristics. These factors were then employed to classify the population into certain household groups using K-Means Cluster Analyses. The identified agent groups were interpreted and their types specified. Regression logistic multinomial model (M-logit) was employed for land-use choices modeling in each typological household agent group. The dependent variable of the model is land-use choice by farming households (Puse). The independent variables of the M-logit model include two groups of spatial variable and socioeconomic characteristics of farming households. Environmental features of lands were defined including Pwet ( soil moisture), Pslope (derived from Digital Elevation Model), Pelev (elevation), PEC (Electrical conductivity as salt factor), Pgroundwater (measuring the reduction of ground water), PPH (PH), PSAR (Sodium Absorption Ratio). Socioeconomic characteristics of farming households that influence farmer's decisions, include Hage (the age of the household head), Hedu (educated farmers), Hincome/pers (annual gross income per capita), Hholding/pers (land holding per capita), Hcultivation/pers (land cultivating per capita), Hlabor (number of workers of the household), and Hdepend(family members of the workers).
We reduced the dimensionality of 14 potential criteria by using PCA. The six (6) principle components were extracted with total Eigenvalues greater than 1.0, explaining 77.4 % of the total variance of original independent variables. The PC1 was strongly correlated to land variables: Hholding=0. 911, Hcultivation=0. 925. The principle components 2 (PC2), 3 (PC3), 4 (PC4), 5(PC5) were most weighted by percentage income from other off-farm activity factors (Hinother=0.843), household size (Hsize=0.833), percentage income from grain (HinGrain=0.773), percentage income from wheat (HinWheat=0.898), respectively.
The K-means run extracted three groups. The group I consists of households which are rich regarding both land resources and income. The group II includes households with average livelihood standard and the group III comprises the poorest households having the lowest amount of land and income. After determining the typological livelihood group, the variables affecting the decision-making was identified using the M-logit model. The effect coefficients were estimated with respect to the fallow land, i.e., the base case. The chi-square test shows that the empirical M-logit model is highly significant (p