Seasonal bean yield forecast for non-irrigated croplands through climate and vegetation index data: Geospatial effects

Seasonal crop estimates at large scales are essential for national food security. Therefore, this study aimed to predict non-irrigated dry bean yields for 41 districts in the semi-arid region in central Mexico before estimations from crop census. 13-year period data included: bean yields from official statistics, climate data from local weather stations, and Remote Sensing Imagery. The study examined a suite of econometric modeling approaches to predict seasonal bean yields under different precipitation categorization scheme (normal-wet seasons and dry seasons), as well as predictand/predictor log-transformations. The method tested Ordinary least Squares (OLS) and OLS + Dummy (dummy variable with spatial regimes), and spatial regression methods: Spatial Lag (SAR) and Spatial Error (SER), and Geographically Weighted Regression (GWR). At first, exploratory OLS regressions were used to create a subset of specified models before testing the spatial regression models. The predictors that accounted for most of the bean yield variability were precipitation and Enhanced Vegetation Index. The models that incorporated explicit and implicit spatial effects (SAR and OLS + Dummy, respectively), with log-transformations of predictand and non-transformation of predictors, showed the best performances (r² between 0.81 and 0.84 with an AIC between −81 and −99). Likewise, a prognosis of a13-year yield simulations (hindcasts) indicated that the latter models are adequate for normal-wet and dry seasons (mean absolute error less than 0.190 ton ha⁻¹). Overall, spatial regression techniques have the potential to estimate bean yields as an early forecast for crop official statistics, eluding large amounts of monetary and human resources on gathering ground data in poor or developing countries, particularly.