Scaling a phenological crop model to produce sub-national maize statistics in a scarce data environment: Application at African scale

Reliable sub-national maize yield statistics are largely unavailable across Africa, limiting the development of index-based insurance, catastrophe models, and evidence-based food security policy. We address this data gap by applying a Light Use Efficiency (LUE) framework to produce a spatially explicit, multi-decadal yield dataset at the second administrative level across the African continent. The approach combines satellite-derived fPAR (MODIS) and reanalysis meteorology (ERA5-Land) within the Monteith production efficiency formulation, extended with a dynamic phenology module and phase-specific stress penalties for heat and drought around silking. Model parameters are calibrated country-by-country against FAOSTAT yield anomalies via Bayesian optimization. Applied to 1,672 admin-2 units across 29 countries, the framework produces a continuous annual record from 2001 to 2024. Validation against national statistics shows that the framework captures the timing and relative severity of climate-driven yield shocks, with Pearson correlations strongest across Southern Africa — reaching 0.82 in South Africa and 0.64 in Zimbabwe. The resulting dataset provides a harmonised, independent view of sub-national maize yield variability — one driven entirely by climate and satellite signals rather than administrative reporting — enabling direct comparison of yield patterns over time and across space. The spatial consistency of the dataset further supports continental-scale analysis, including the identification of risk zones and the characterization of differential responses to large-scale climate phenomena such as ENSO across regions. The dataset and underlying code are openly shared for replication, validation, and application in agricultural risk management and climate adaptation contexts.