How Far Can We Predict African Crop Yields? A Multimodal Benchmark Across 1,109 Districts and 42 Years

How accurately can we predict crop yields across sub-Saharan Africa using publicly available data? We address this question by establishing the first comprehensive machine learning benchmark on HarvestStat Africa, covering 1,109 administrative districts in 33 countries over 42 years (1981–2022). We engineer 246 features from six data sources (climate reanalysis, satellite vegetation indices, burned area, soil and topographic properties, yield history, and country-level encodings) and quantify the marginal contribution of each modality through systematic ablation. We propose a novel dual-edge spatial graph combining geographic adjacency and climate-correlation edges, and evaluate gradient boosting, graph neural networks (GAT, RGAT), and a hybrid XGBoost+GNN residual pipeline. Our best model achieves R² = 0.809 (RMSE = 0.553 t/ha) for maize yield prediction on held-out test years (2019–2022), rising to R² = 0.893 when cropping system duplicates are resolved, and generalizes to sorghum and millet. Our analysis reveals that yield history dominates prediction (R² = 0.778 from 13 features alone), that all modalities contribute but with diminishing returns, and that GNNs add nothing to a well-engineered gradient boosting model because explicit spatial features already capture the available spatial structure. We report 15+ experiments including negative results and provide actionable guidelines for practitioners.