Large-sample characterization of flooding events in India

Sai Kiran Kuntla , Manabendra Saharia

Effective flood management requires a robust understanding of past floods. In India, such understanding is largely limited to case studies due to the absence of a standardized observed flood dataset. We address this gap by presenting a national dataset of 7500 flooding events, developed by merging observed streamflow records with official flooding thresholds and augmenting it with multiple catchment-scale variables. Spatial analysis reveals high normalized flood magnitudes along the southwest coast—an area with intense rainfall and mountainous terrain. Temporally, 86% of floods occur during the southwest monsoon. Using a random forest model combined with the game-theoretic SHAP approach, we find that precipitation of the wettest month is the most influential predictor of flood magnitude. Grouped feature importance shows climatology contributes 61% to model performance, while geomorphology accounts for 39%. This comprehensive large-sample study surpasses conventional case studies, providing a more robust understanding of flooding patterns and drivers across India.