Process-based seasonal crop yield forecasting for early warning in sub-Saharan Africa
Abstract
Seasonal crop yield forecasts can support early warning and short-term adaptation in sub-Saharan Africa (SSA), where rainfed agriculture dominates, but systematic multi-crop, multi-region evaluations remain scarce. Here, we develop and evaluate a forecasting system in which bias-corrected, downscaled SEAS5 forecasts drive the LPJmL crop model. We generated 30 year hindcasts (1994–2023) for all major crops in Ghana, Kenya, and Zimbabwe, three contrasting SSA agro-climatic regimes, and assessed probabilistic accuracy and event discrimination from four months before to six months after sowing, against a climatological reference from an ERA5-Land-driven simulation. Probabilistic skill (continuous ranked probability skill score) was weak or negative around sowing and improved for later initialisations, most consistently in Zimbabwe, where the country mean rose from about 0.08 at sowing to about 0.7 four months later. Kenya improved similarly but with greater spatial heterogeneity, partly reflecting limited climate-forecast skill for its long rains, while Ghana showed predominantly negative early-season skill. All crop–country pairs reached positive skill in at least one metric by one month before harvest. Event discrimination was more informative than probabilistic accuracy at long lead times, separating above- from below-normal yield years for most crop–country pairs even before sowing. For post-sowing initialisations, area under the relative operating characteristic curve exceeded 0.7 for about 80% of grid cells on average, improving on recent climatology by up to about 0.5. The forecasts can therefore signal the direction of simulated yield anomalies well before their magnitude becomes predictable. This skill reflects the predictability of climate-driven yield variability as simulated by LPJmL. However, agreement between simulated and reported yield anomalies varied by crop and country, and was highest for maize in Zimbabwe and lowest in Ghana.