Modeling tropical deforestation in the southern Yucatán peninsular region: Comparing survey and satellite data

This paper presents some initial modeling results from a large, interdisciplinary research project underway in the southern Yucatán peninsular region. The aims of the project are: to understand, through individual household survey work, the behavioral and structural dynamics that influence land managers' decisions to deforest and intensify land use; model these dynamics and link their outcomes directly to satellite imagery; model from the imagery itself; and, determine the robustness of modeling to and from the satellite imagery. Two complementary datasets, one from household survey data on agricultural practices including information on socio-economic factors and the second from satellite imagery linked with aggregate government census data, are used in two econometric modeling approaches. Both models test hypotheses concerning deforestation during different time periods in the recent past in the region. The first uses the satellite data, other spatial environmental variables, and aggregate socio-economic data (e.g., census data) in a discrete-choice (logit) model to estimate the probability that any particular pixel in the landscape will be deforested, as a function of explanatory variables. The second model uses the survey data in a cross-sectional regression (OLS) model to ask questions about the amount of deforestation associated with each individual farmer and to explain these choices as a function of individual socio-demographic, market, environmental, and geographic variables. In both cases, however, the choices of explanatory variables are informed by social science theory as to what are hypothesized to affect the deforestation decision (e.g., in a von Thünen model, accessibility is hypothesized to affect choice; in a Ricardian model, land quality; in a Chayanovian model, consumer-labor ratio). The models ask different questions using different data, but several broad comparisons seem useful. While most variables are statistically significant in the discrete choice model, none of the location variables are statistically significant in the continuous model. Therefore, while location affects the overall probability of deforestation, it does not appear to explain the total amount of deforestation on a given location by an individual.