Machine learning reveals disruptive nutrient pollution shifts in Chinese rivers to 2100

Abstract Anticipating nutrient pollution under changing conditions is urgent for water security. Nonetheless, high-resolution predictive frameworks capturing nonlinear driver responses remain limited. Here, we present a nationwide assessment of China’s water-quality evolution from 2023 to 2100, integrating over 3 million daily records with 41 climatic, landscape, and socioeconomic drivers via regionally tailored Random Forest models ( R 2 of 0.88–0.92). Our results reveal a disruptive spatiotemporal shift, with projected NPI ranges from –50.9% to +218.1% under SSP5-8.5. Seasonal patterns restructure toward unimodal peaks, with pollution increases in spring/autumn (up to 28.3%) but decreases in summer (up to 27.0%). Spatial homogenization emerges via a westward/southward shift of pollution centers, with localized increases exceeding 200% from coldspots with low baselines (<0.4 vs >1.0 in hotspots). Landscape configuration dominates (64.5% feature importance) over climatic forcing (7.2%–35.5%), reinforced by minimal climatic projection uncertainty. Strategic land-use planning could be a cornerstone of future water security.