Determination of Flood potential Using CART, GLM and GAM Machine learning Models

Flood is a phenomenon that causes a lot of environmental and socio-economic damage. The purpose of this study is to evaluate the efficiency of CART, GLM and GAM machine learning models in identifying flood risk areas in the Kashkan basin. Lorestan province and especially Kashkan basin, including: Selseleh, Delfan, Doreh, Khorramabad, Poldakhtar and Kuhdasht, is flooded and has suffered flood damage many times and in April 2019, experienced the largest flood of the last 200 years. In this regard, various factors including: height, slope direction, land curvature, slope percentage, distance from the river, drainage density, soil, lithology, land use and topographic moisture index were used. The digital map of all the mentioned factors was prepared in ArcGIS10.5 software and in the form of a database. The location of 123 flood events recorded in recent years in this basin was collected and randomly used in two categories of model training (86 cases) and model validation (37 cases) in modeling. Using machine learning models and environmental factors, flood potential prediction maps were prepared and then validated using AUC characteristic curve methods and TSS index. The results of model validation showed that CART machine learning model with AUC = 0.91 and TTS = 0.88 index was the most accurate model in predicting flood risk potential, followed by GAM model with AUC = 0.87 and TSS index = 0.84 and GLM model with AUC = 0.83 and TSS index = 0.88. Accuracy 0.91 CART model indicates the excellent accuracy of this model for the Kashkan basin. This model shows a larger area of ​​the basin under high potential and moderate flood risk conditions, which include most of the western areas as well as the central areas of the basin (Kuhdasht, Khorramabad and Poldakhtar), which are exactly parts The same areas were flooded in the great flood of 2019 and it is necessary to be in the first priority of flood risk planning and management in this basin.