Flood risk prediction in data-scarce sub-Saharan African contexts demands models that reason about both spatial interdependencies and temporal change dynamics. Existing approaches treat the landscape as an independent collection of pixels, failing to capture how river networks, elevation gradients, and road connectivity propagate flood risk across connected regions. This paper proposes Graph-Transformer GeoAI, a novel deep learning framework that represents the Earth's surface as a spatial graph of 500 m × 500 m tiles and applies a three-stage pipeline: (1) a modified ResNet18 convolutional neural network (CNN) extracts 256-dimensional feature embeddings from 19-band multi-source satellite imagery; (2) a three-layer Graph Attention Network v2 (GATv2) learns spatially-aware node representations by propagating information across 196,566 spatial edges; and (3) a fourlayer Temporal Transformer captures seasonal change patterns from dry-season to flood-period signatures. Applied to the Niger-Benue Confluence in Lokoja, Kogi State, Nigeria, the full pipeline achieves Flood F1 = 0.9697, Accuracy = 99.87%, and AUC = 0.9998 on held-out test nodes. An ablation study across four configurations — all experimentally obtained — confirms monotonic improvement as each component is added. GAT attention weights independently rediscovered the Niger and Benue river flood plains without explicit river labels. A georeferenced GeoTIFF output was validated in QGIS confirming accurate spatial registration over Lokoja, Kogi State.
Abiodun Emmanuel Ofobutu (Fri,) studied this question.