Holistic risk assessment using a hybrid approach in a flash flood disaster-prone area of the poyang lake basin

Flash floods, as prevalent and destructive natural disasters, pose significant threats to societal and economic systems. This study developed a holistic framework for flash flood risk assessment that integrates machine learning with multi-criteria decision analysis, while incorporating novel explainable artificial intelligence techniques, and applied it to assess flash flood risk in the Rao River Basin, a representative flash flood-prone area within the Poyang Lake system. Three ensemble learning models were employed for the evaluation of flash flood hazard, with their performance assessed via classification evaluation metrics and the results interpreted through explainability analysis. Multi-criteria decision analysis was used to assign weights to vulnerability-related factors, supporting the analysis of vulnerability. By synthesizing hazard and vulnerability assessments, a spatially explicit flash flood risk map was developed. The results show that the boosting model outperforms the other models in terms of hazard prediction, achieving superior performance in key metrics such as the F1-score(0.935) and AUC (0.977). Factors such as drainage density andruggedness number were found to significantly influence hazard levels. The proportions of cropland and impervious land areas, along with the urbanization level, were identified as significant contributors to vulnerability. The moderate to high-risk areas in the basin are predominantly found in settlement clusters located within hilly terrain. The analysis of historical flash flood events revealed that 77.7% of disaster events occurred in high-risk and very high-risk zones. Additionally, field surveys corroborated the reliability of the model outcomes. These findings offer insights into disaster prevention and management efforts in such regions.