Comparison of Machine Learning Models in Flood Susceptibility Zoning in KarajDam Basin

The present research aims to determine areas with flood susceptibility using CART, RF and BRT models. Twelve factors affecting flood susceptibility including altitude )DEM), slope, aspect, distance from stream, lithology, rainfall, land use, SPI, TPI, TWI, curvature and RSP were selected. Out of 82 flood points, 70 percent to 30 percent were randomly classified as training and validation data. Also, random forest method was used to determine the most important parameters. The ROC curve was also used to validation the model. According to the random forest model, DEM, distance from stream, rainfall, land use and RSP were the most important factors affecting the susceptibility and probability of floods, respectively. According to the ROC chart, the accuracy of the RF model as a superior model has been very good in both training )0.884) and validation )0.856). According to the final flood susceptibility map, 32.7 percent of the study area has a medium to high flood susceptibility. The results showed due to the high accuracy of the spatial distribution map of flood susceptibility can be promising for decision makers, local managers and policymakers to reduce flood damages.