Spatial Modeling of Rainfed Bean Production in the Semiarid Central Mexico: Climate Surface and Vegetation Remote Sensing Determinants

Abstract

The goal of this study is to develop a method to estimate dry bean crop production from non-irrigated farmlands in central Mexico to improve crop forecasts. The proposed method has two components, estimating the dry bean cropland area (ha) and predicting the annual bean yield per area (tons/ha). The method utilized readily available data, so it could be applied operationally without additional investment. Senescence ratios based on MODIS (Moderate Resolution Imaging Spectroradiometer) Normalized Difference Vegetation Index (NDVI) and Enhanced Vegetation Index (EVI) were used to estimate the dry bean cropland area. A senescence ratio is calculated by dividing the vegetation index value 32 days after peak date by the vegetation index value at peak date. Senescence ratios between 0.6 and 0.7 were found to a good indicator for bean croplands with a mean absolute percent error (MAPE) of 34%. The data that was used to develop the bean yield model included MODIS remote sensing imagery and climatic data (precipitation and temperature) from local weather stations. The study examined a suite of empirical modeling approaches to estimate annual bean yields, including non-spatial regression methods (ordinary least squares) and spatial regression methods (Spatial Lag and Spatial Error models and Geographically Weighted Regression). The independent variables that accounted for most of the bean yield variability were accumulated precipitation and EVI. Spatial principal components were also used to examine temporal variability of bean yields. The best model was found to be the Spatial Error model using spatial principal components with a coefficient of determination of 0.84 and an Akaike Information Criterion statistic with small sample size correction (AIC) of -87. Total bean production estimates (tons/year), using the best cropland area estimations * best bean yield estimations from this methodology, showed acceptable results (coefficient of determination of 0.70) with a 17% underestimation.