A Novel Multi-Driver Explainable AI (XAI) Framework for Predicting African Precipitation Extremes at Grid-Cell Resolution: Insights from 15 Climate Drivers over Africa

Extreme precipitation is growing more frequent and severe in many regions worldwide, increasing risks to lives, infrastructure, and agricultural production. Nowhere is this challenge more pronounced than in Africa, where rainfall is substantially erratic, and populations are frequently subject to droughts, floods, and sudden changes in seasonal patterns. Although previous studies have explored the links between climate and rainfall in Africa, most use simplified methods that smooth over regional differences and cannot capture how specific drivers influence different types of extremes. To address the gap, an explainable AI (XAI) framework was designed to predict ten standard ETCCDI precipitation extreme indices at grid-cell resolution across Africa, informed by 15 major climate drivers from the Pacific, Indian, and Atlantic oceans. The approach is based on a broadcast-fusion XAI-CNN that incorporates scalar climatic indices with spatial precipitation patterns. The model learn how various large-scale factors influence local extreme behavior, expanding each climate driver into a spatial layer. The model is developed using CHIRPS data from 1981 to 2025, achieving a mean R² of 0.80 throughout all indices, with the highest performance for PRCPTOT (0.90) and CDD (0.88). The results exhibit that the large-scale drivers contain sufficient insight to predict wet and dry extremes at the continental scale. The findings challenge the long-standing view that ENSO is the most dominant influence on African rainfall. However, the Tropical Atlantic emerges as the strongest driver of extreme wet events, while the Indian Ocean Dipole, central Pacific ENSO, and Pacific Warm Pool exhibit regionally specific influences on East, Southern, and Central Africa, respectively. The study presents a transparent and scalable approach to characterizing hydroclimatic extremes by integrating deep learning with an explainability framework. The spatial explainability analysis improves interpretability and reveals physically consistent teleconnection patterns, opening avenues for regionalized climate prediction and disaster risk reduction.