Advancing hydrological prediction in South Africa with differentiable multi-source meteorological data fusion

Study region South Africa, covering 188 medium- and small-sized basins with diverse climatic and physiographic settings and scarce ground-based observations. Study focus Runoff prediction in South Africa is severely constrained by sparse gauge networks and the inconsistencies among precipitation products. This study develops a differentiable multi-source meteorological data fusion framework that integrates three precipitation datasets (CHIRPS, ERA5-Land, TAMSAT), additional meteorological variables, and basin attributes in an end-to-end model. Unlike conventional bias correction or data assimilation methods, the framework jointly optimizes data fusion and runoff prediction, enabling task-driven weighting of precipitation sources without relying on ground observations. New hydrological insights for the region The proposed framework consistently outperforming non-fusion baselines, achieving a median Nash–Sutcliffe efficiency of 0.38, representing a greater than 52 % improvement the single-source input models. Even when compared to the direct splicing method, the differentiable fusion approach shows a 23 % increase in median NSE, highlighting its ability to effectively mitigate data inconsistency and redundancy. In arid regions, which are particularly difficult for hydrological modeling, the fusion mechanism automatically captures the complementary strengths of different precipitation products. The differentiable fusion model also significantly reduces physically unreasonable negative flow predictions, a common issue with simple data splicing in drought-prone basins. These findings underscore the methodological value of embedding differentiable fusion into hydrological prediction frameworks, offering a scalable and more stable pathway for data-poor regions worldwide.