Spatial prediction of flood susceptible areas using machine learning methods in the Siahkhor Watershed of Kermanshah province

This study assesses flood modeling sensitivity in the Siahkhor Watershed in the northeastern part of Islam Abad Gharb city within Kermanshah Province. In this regard, the data-driven methods of machine learning (ML) algorithms, namely Random Forest (RF), Support Vector Regression (SVR), Bayesian Ridge Regression (BRR) and Gradient Boosting Regressor (GBR) were employed to zone the flood-susceptible areas. A flood distribution map was constructed to predict future flood occurrences based on historical flood events. The dataset comprised fifty-three recorded flood events, with forty-nine regions considered non-flood susceptibility maps. Spatial locations were randomly divided into 70% and 30% for modeling and validation, respectively. Eight influential factors for flood zoning; slope, drainage density, river proximity, land use, geology, precipitation, slope aspect, and elevation were selected and prepared based on the interpretation of aerial photographs, historical data, Google Earth, and field surveys. Evaluation results of the model outputs indicated that, in the RF, SVR, BRR, and GBR methods, the coefficient of determination during the training period stood at 0.98, 0.97, 0.77, and 0.99, respectively. These values were 0.67, 0.70, 0.59, and 0.64 during the validation period. To evaluate model accuracy, ROC was used in the R software environment. The RF, SVR, BRR, and GBR validation results indicated that the Area under the ROC Curve scored 0.95, 0.99, 1, and 0.94, respectively. The flood susceptibility map (FSM) was categorized into five classes: very low, low, moderate, high, and very high sensitivity. According to the RF, SVR, BRR, and GBR methods, approximately 52%, 57%, 12%, and 36% of the study area fell into the very low-risk category, while approximately 35%, 43%, 61%, and 54% were classified as very high to high-risk areas. The most significant variables for predicting flood susceptibility were land use (0.35), slope (0.25), and elevation (0.22). According to the statistical criteria and ROC, all the methods provided favorable results. Still, in general, the SVR and GBR algorithms were more accurate based on the statistical results and are more compatible with the reality of the region. Flood risk zoning maps, coupled with necessary preparedness measures and formulation of flood prevention and optimal management strategies, can substantially mitigate flood-induced damages within the country.