Reliable Early Prediction of Agricultural Drought Under Geographic Shift: A Continental Low-Resource Benchmark for Data-Sparse African Basins

Agricultural drought represents a devastating threat to food security and agropastoral livelihoods across Africa. While recent machine learning (ML) literature demonstrates high continuous correlation in regional drought forecasting, the overwhelming majority of existing frameworks evaluate models within the same geographical catchment using random or chronologically shuffled cross-validation. Consequently, the operational behavior of drought forecasting models under severe geographic distribution shift remains critically under-characterized. In this study, we present a reproducible, continental-scale, low-resource benchmarking framework evaluating ten forecasting architectures across six major African agricultural basins: the Nile River Basin, Sahelian Agropastoral Belt, Zambezi-Limpopo Basin, Maghreb Zone, Congo Agricultural Transition, and Horn of Africa, over a 44-year observational record (1981-2024) using 100% real CHIRPS v2.0 and ERA5-Land satellite and reanalysis data. Our key findings: (1) spatial domain shift degrades forecasting accuracy by +6.4% to +17.2% MAE; (2) uncalibrated models exhibit catastrophic event-detection failures despite high continuous correlation; (3) a modest 25% local historical calibration budget recovers over 50% of the lost transfer skill; and (4) all improvements are statistically significant (p < 0.05) while consuming less than 15 MB RAM. These results establish parsimonious, probability-calibrated ensemble architectures as a scientifically rigorous, computationally democratic foundation for regional drought early warning in resource-constrained environments.