Abstract
Soil moisture depletion precedes declines in vegetation productivity by days to weeks,creating a natural prediction window for agricultural drought impacts. We test whethersatellite root-zone soil moisture (RZSM) from the Soil Moisture Active Passive (SMAP)Level-4 product can provide skillful sub-seasonal forecasts of crop productivityNear-Infrared Reflectance of Vegetation (NIRV ) anomalies across drought-prone croplandsof Eastern and Southern Africa. We test three input configurations: NIRV -only, NIRVwith concurrent RZSM, and NIRV with explicitly lag-shifted RZSM, to forecast NIRVanomalies at lead times of 8 to 40 days using a Convolutional Long Short-Term Memory(ConvLSTM) model chosen for its ability to jointly capture spatial teleconnections andtemporal lag structure that scalar or pixel-wise models cannot resolve. All configurationsoutperform the climatological N IRv and reduce the RMSE by 22–24%. The concurrentRZSM configuration achieves the highest domain-averaged anomaly correlation coefficient(ACC = 0.555), with the soil moisture contribution growing monotonically from negligibleat 8-day leads to +0.049 ACC at 40 days. When we stratify by aridity, the RZSMadvantage concentrates in water-limited croplands, and the ACC improves by up to +0.09at 40 days when sufficient temporal context (80+ days) is provided. In energy-limitedregions, soil moisture adds minimal predictive value. Our case study of the 2024 southernAfrican El Niño drought demonstrates that the RZSM-informed model resolves the spatialpattern of crop failure approximately 32 days in advance, improving spatial correlation by29% over the autoregressive. Our results establish that SMAP root-zone soil moisture is aviable predictor of sub-seasonal agricultural drought impacts in water-limited Africancroplands.