基于可解释机器学习的山洪致灾因子研究OA

To faciliate scientific and precise prevention and mitigation of flash flood disasters,this study investigates the predic-tion of flash flood occurrence probability and the analysis of hazard-causing factors by integrating remote sensing data and ex-plainable artificial intelligence(XAI)techniques.First,18 hazard-causing factors across five dimensions-geographical loca-tion,topography,meteorology,human activities,and land cover-were extracted from remote sensing retrieval data.Next,combined this with flash flood disaster survey data,an XGBoost machine learning model was employed to construct a prediction model for flash flood occurrence probability.Finally,the SHAP(SHapley Additive exPlanations)method was applied at different scales(the entire Zhejiang Province and its sub-regions)to quantify the single-factor importance,impact effects,and multi-factor interactions among hazard-causing factors,thereby revealing regional heterogeneity in factor impacts and interactions.The results indicate that the model achieves a prediction accuracy of 0.90 and a precision of 0.87,demonstrating its capability to effectively determine the probability of flash flood occurrences.At the provincial scale,the predicted flash flood probabilities are highly con-sistent with the spatial distribution patterns of historical flash flood locations,further verifying the model's high extrapolative ro-bustness.The importance ranking of the hazard-causing factors reveals that the impervious surface proportion is the most critical factor affecting flash flood probability,while the prominent contribution of longitude and latitude coordinates highlights the signifi-cance of considering geospatial relationships in flash flood prediction.At the regional scale,the multi-factor interaction analysis successfully identifies the differentiated impact patterns of hazard-causing factors across various regions.These research findings provide a scientific reference for related studies,such as the understanding of flash flood disaster-causing mechanisms and pre-cisely quantifying their triggering conditions.