Abstract
Flash flood susceptibility mapping is essential for identifying areas prone to flooding events and aiding decision-makers in formulating effective prevention measures. This study aims to evaluate the flash flood susceptibility in the Yarlung Tsangpo River Basin (YTRB) using multiple machine learning (ML) models facilitated by the H2O automated ML platform. The best-performing model was used to generate a flash flood susceptibility map, and its interpretability was analyzed using the Shapley Additive Explanations (SHAP) tree interpretation method. The results revealed that the top four models, including both single and ensemble models, demonstrated high accuracy in the tests. The flash flood susceptibility map generated by the best-performing eXtreme Randomized Trees (XRT) model showed that 8.92%, 12.95%, 15.42%, 31.34%, and 31.37% of the study area exhibited very high, high, moderate, low, and very low flash flood susceptibility, respectively, with approximately 74.9% of the historical flash floods occurring in areas classified as moderate to very high susceptibility. The SHAP plot identified topographic factors as the primary drivers of flash floods, with the importance analysis ranking the most influential factors in such descending order as DEM, topographic wetness index, topographic position index, normalized difference vegetation index, and average multi-year precipitation. This study demonstrates the benefits of interpretable machine learning, which can provide guidance for flash flood mitigation.