Water pollution incidents in large river basins are often documented as isolated emergencies, yet their recurrence can reveal broader risk regimes shaped by hydrological connectivity, environmental assimilation capacity, and human activity. From a One Health perspective, these incidents are not only water-quality disturbances, but also interconnected threats to ecosystem integrity, drinking-water security, food-system safety, and human health. Here, we constructed a standardized event-based database of 2,606 water pollution incidents in the Yangtze and Yellow River basins from 1985 to 2024. An NLP-assisted standardization framework was used to extract, classify, and georeference heterogeneous historical records, achieving an accuracy of 86.0% and an F1-score of 78.0%. Building on this database, a hybrid 2D-CNN-CapsNet-WOA framework reconstructed continuous basin-scale risk surfaces from discrete incident locations and multi-period environmental covariates, achieving an accuracy of 91.8% and an F1-score of 91.5%. High-severity events were concentrated during 2006–2015, whereas lower- to moderate-grade events remained frequent thereafter. Reconstructed risk surfaces revealed two contrasting spatial regimes: the Yangtze River basin evolved from localized hotspots into connected urban-industrial risk corridors, whereas the Yellow River basin developed fragmented, polycentric clusters around resource-development and midstream industrial zones. Event-type analysis showed that illegal discharges contributed most strongly to continuous risk corridors, production accidents clustered around industrial nodes, transportation-related incidents followed infrastructure networks, and safety accidents formed dispersed operational risk points. SHAP-based interpretation further revealed stage-dependent driver reorganization, highlighting stronger socio-economic and corridor-related controls in the Yangtze River basin and stronger sensitivity to water scarcity, terrain constraints, and resource-development intensity in the Yellow River basin. This study provides a generalizable event-based framework for converting fragmented pollution records into spatially explicit risk evidence, supporting anticipatory basin governance and One Health-oriented prevention across environmental, ecological, and public-health systems.

Water pollution incidents,Event-based database,Basin-scale risk reconstruction,Interpretable deep learning,One health