Elucidating observed flood occurrences in the Yangtze River Delta with interpretable machine learning

Understanding the relationships between flooding and potential factors is fundamental to effective risk management; however, progress is often impeded by the scarcity of spatially explicit flood data and the limited capacity of conventional models to decipher complex nonlinear relationships. To address these limitations, this study develops a regional-scale explanatory framework using remote sensing-derived flood records, multi-source environmental variables, and an interpretable machine learning model. Taking China’s Yangtze River Delta (YRD) as a case study, the results demonstrate that the XGBoost-based SHAP model outperforms random forests and logistic regression in regional flood prediction, achieving an accuracy of 0.970. The model characterizes nonlinear relationships between flood occurrences and environmental factors, revealing opposite associations of specific drivers across identified thresholds. These relationships vary across terrains and land-use types. Floods in mountainous areas are primarily associated with topographic wetness and surface texture indices, whereas those on plains are jointly influenced by river network density and surface texture. Notably, the flood effects of the largest patch index shift from negative in forests to positive in urban areas. These findings advance the understanding of regional flood occurrences and support site-specific flood mitigation strategies.