Enhancing Flash Flood Susceptibility Modeling in Arid Regions: Integrating Digital Soil Mapping and Machine Learning Algorithms

Flash floods in arid regions are among the most dangerous and destructive disasters worldwide, with their frequency increasing due to intensified climate change and anthropogenic activities. This study aims to identify susceptibility areas to flash floods in arid regions, characterized by high vulnerability, numerous complexities, and unknown mechanisms. 19-flash flood causative physiographic, climatic, geological, hydrological, and environmental parameters were considered. Using the Boruta wrapper-based feature selection algorithm, temperature, distance to the river, and elevation were identified as the most effective parameters. Four standalone and hybrid machine learning models (Random Forest (RF), Support Vector Regression (SVR), GLMnet, TreeBag, and Ensemble) were employed to model and determine flash flood susceptibility maps. Based on performance evaluation metrics (accuracy, precision, recall, and Areas Under Curve (AUC) indexes), the RF and Ensemble models exhibited the best performance with values of (0.94, 0.93), (0.97, 1), (0.92, 0.88), (0.94, 0.94), respectively. The findings highlighted the previously overlooked role of soil in flood susceptibility mapping studies, particularly in arid areas with high levels of silt and clay soils. This study introduced digital soil mapping for the first time in flood susceptibility studies, providing an effective approach for the spatial prediction of soil properties using easily accessible remote sensing data to generate soil maps in areas with limited available data. It emphasizes the importance of examining the role of soil in arid areas during flash flood modeling and recommends using Ensemble and RF models for their high flexibility in such studies.