From prediction to regionalization: Enhancing flash flood susceptibility mapping using machine learning and GeoDetector

Flash floods cause substantial economic losses and casualties worldwide. Susceptibility-based flash flood mapping supports the development of effective flood mitigation strategies. While machine learning (ML) models offer superior accuracy, converting their outputs into spatially coherent and actionable maps remains challenging. Existing susceptibility maps often rely on subjective discretization and exhibit fragmented spatial patterns, limiting their utility in practice. In this context, this study proposes a novel framework that achieves the effective transformation of susceptibility prediction results into a management-oriented regionalization map. The framework integrates supervised learning, unsupervised clustering, and spatial explanatory feedback to enable information fusion and spatial restructuring of multi-model outputs. Flash flood susceptibility was first modelled using two supervised algorithms: Random Forest and CatBoost. Their outputs, along with exposed elements, were integrated and discretized using a two-stage clustering approach based on Self-Organizing Maps (SOM) and Ward’s method. Finally, a GeoDetector-based iterative optimization process was implemented to refine the regionalization by maximizing alignment with historical flash flood distributions. Results show that all susceptibility models achieved excellent predictive performance (AUC > 0.95), with the CatBoost model trained on grid-based samples performing best (AUC = 0.997). The final regionalization map exhibits regional contiguity and effectively captures historical flood patterns, explaining 73% of their spatial variability. The integration of hybrid ML with explanatory feedback provides a novel perspective for generating susceptibility regionalization maps that are both expressive of flash flood risk and spatially coherent, in addition to providing support for exploring region-specific defense measures.