Evaluating the Efficiency of Machine Learning Models in Preparing Flood Probability Mapping

Flood is one of the most devastating natural disasters that annually causes financial and life losses. Therefore, developing a susceptibility map for flood management and reducing its harmful effects is essential. The present study was conducted to prepare a flood susceptibility map using data mining models including Random Forest (RF) and Gradient Boosting Machine (GBM). At first, 275 flooding locations flood and 275 non-flood locations were identified in the Komijan watershed of Markazi province. Spatial locations were randomly divided to 70% (190 location) and 30% (82 location) for modeling and validation, respectively. Then, 12 factors affecting the occurrence of flood including slope, aspect, altitude, rainfall, land use, distance from river, drainage density, plan curvature, profile curvature, lithology, soil and stream power index were determined. The ROC curve was used to evaluate the models used. The results showed that in the validation stage, the under curve for RF and GBM models was 0.83 and 0.75%, respectively, which indicates that the RF model is more accurate in producing a flood susceptibility map. The most important factors affecting the flood are rainfall, distance from river and altitude.