Upscaling sediment source prediction for watershed management

Abstract Identifying sources of erosion is critical for effective watershed management, but existing fingerprinting methods remain resource-intensive and difficult to standardize or scale. To address this, we compiled a global synthesis of 142 sediment tracing studies across 267 watersheds and developed an explainable machine learning framework that predicts sediment provenance using remotely sensed attributes. Four sediment categories capture global erosion sources—subsurface, cultivated, non-cultivated, and infrastructure—and our model predicts their contributions with robust accuracy (R 2 = 0.40–0.53). Model interpretation confirms that predictions align with expected hydroclimatic and land cover relationships, indicating consistency with the physical processes governing sediment sourcing. When applied to six watersheds in the United States and United Kingdom, the model reveals contrasting erosional regimes: subsurface erosion dominates the Upper Mississippi and Chesapeake Bay, whereas non-cultivated surface erosion prevails across the UK watersheds. This predictive framework provides a transferable, interpretable, and data-driven tool for mapping sediment sources at large scales, enabling targeted watershed management.