Predicting Regional Food Production Under Climate Uncertainty with Probabilistic Deep Networks

Abstract

Regional food production forecasting under climate uncertainty represents one of the most pressing challenges at the intersection of environmental science and computational intelligence. Traditional deterministic models often fail to capture the compounding uncertainties arising from volatile precipitation regimes, temperature anomalies, and shifting growing seasons. This study proposes a probabilistic deep network (PDN) framework integrating Monte Carlo dropout with a hybrid recurrent encoder and a temporal attention mechanism to generate calibrated uncertainty bounds for multi-crop yield prediction across heterogeneous agricultural regions. The model is trained on a 35-year retrospective dataset covering six major crop types across 480 administrative units in East Asia and Sub-Saharan Africa, augmented with ERA5 reanalysis climate variables and MODIS-derived vegetation indices. Experimental results demonstrate that the proposed PDN achieves a root mean square error of 0.31 tonnes per hectare and a continuous ranked probability score of 0.18, outperforming baseline recurrent networks and ensemble regression methods by margins of 14% and 22%, respectively. Uncertainty quantification calibration evaluated through prediction interval coverage probability confirms that 90% prediction intervals contain the true yield in 88.4% of test instances, validating the practical utility of the probabilistic framework for climate adaptation planning and food security governance.