Predicting East Africa's rainfall extremes with calibrated, hybrid physical and AI systems
Download
Abstract: This study presents a low-cost, fine-tuning method that calibrates ensemble forecasts to probabilities of exceedances for rainfall extremes, applied to physical and hybrid artificial intelligence rainfall forecasts over East Africa. The research demonstrates that hybrid AI systems, which integrate global physical forecasts with machine learning approaches, enable low-cost ensemble generation and extend skill to medium-range lead times, offering scalable systems to inform early warning systems particularly under resource-constrained settings. Case-study results show encouraging skill improvement from hybrid AI systems at higher rainfall thresholds, with sub-regional analysis revealing that hybrid AI systems better reflect observed chances of extremes at longer lead times. The tailored evaluation across rainfall and probability thresholds helps identify the ranges of rainfall and risk tolerances where most actionable improvement arises and guides forecast adoption for early warning systems to protect vulnerable communities from extreme weather hazards.