Sahel storm forecasts: AI model beats baselines up to six hours ahead

Flash flooding from intense rainfall causes major damage and loss of life across Africa, and weather models struggle in data-sparse regions to predict when storms start, evolve and organise. Short-range forecasts from satellite data could help fill that gap, though the work reports benchmark skill rather than flood outcomes.

Explainable AI analysis on a single-location version indicated that predictions rely on physically meaningful factors: how close, how large and how intense nearby cores are. Ablation tests showed self-attention contributed most to skill, with temporal information mattering more at longer lead times.

The Sahel's rainfall is dominated by mesoscale convective systems, whose cores change quickly. That limits short-term predictability, particularly where observations are sparse and numerical weather prediction models carry substantial uncertainty about storm initiation, evolution and organisation.

The study took an object-based approach: instead of predicting raw pixels, it detects convective cores as objects and models how their populations evolve and interact. It first tested the idea at one location, then scaled it to a gridded probabilistic forecast for the western Sahel.

Explainable AI analysis suggested the single-location predictions were driven by core proximity, size and intensity. Ablation experiments showed self-attention was the largest contributor to skill, and the authors conclude that object-based learning has potential for reliable short-range nowcasts in data-sparse environments.