ارزیابی و مقایسه روش‌های یادگیری ماشین در پهنه‌بندی خطر سیل استان ایلام با تلفیق داده‌های توپوگرافی، هواشناسی و سنجش‌ازدور

Flooding is one of the major natural hazards that can cause substantial human, economic, and environmental losses. Identifying flood-prone areas and producing flood hazard maps are among the effective non-structural measures for flood risk management and mitigation. This study aimed to evaluate and compare machine learning methods for flood hazard zoning in Ilam Province, Iran, through the integration of topographic, meteorological, and remote sensing data. Twelve flood-conditioning variables were considered, including digital elevation, slope, curvature, aspect, flow accumulation, river proximity, soil texture, land use/land cover, maximum precipitation, Stream Power Index (SPI), Topographic Wetness Index (TWI), and Normalized Difference Vegetation Index (NDVI). The maximum precipitation layer was generated using the WRF model, while flooded areas were extracted through the processing of Sentinel-1 radar images acquired before and after the October 2015 flood event. Subsequently, 1,000 randomly distributed sample points were selected, and four models, namely Multiple Linear Regression, Support Vector Machine (SVM), Decision Tree, and Bagged Ensemble, were evaluated using 10-fold cross-validation. The RMSE values obtained for these models were 0.471, 0.322, 0.466, and 0.427, respectively, indicating the best fitting performance for the SVM model. However, the correct classification rates of flooded points were 52%, 62%, 54%, and 75%, respectively, while the corresponding accuracies for non-flooded points were 78%, 91%, 79%, and 87%. Although the SVM model achieved the lowest prediction error, the Bagged Ensemble model showed the highest capability for identifying actual flooded locations while maintaining a suitable balance between flooded and non-flooded classification. Therefore, the Bagged Ensemble model was considered the most reliable method for producing the final flood hazard map. Analysis of high-hazard areas also indicated that these zones were mainly characterized by low elevation, gentle slopes, and sparse vegetation cover. Overall, the results demonstrate the applicability of integrating geospatial, meteorological, and remote sensing information with machine learning algorithms for flood hazard zoning at the provincial scale.