Stormwater management modeling and machine learning for flash flood susceptibility prediction in Wadi Qows, Saudi Arabia

Abstract

Predicting flash flood-prone areas is essential for proactive disaster management. ‎However, such predictions are challenging to obtain accurately with physical hydrological models owing to the scarcity of flood observation stations and the lack of monitoring systems. This study aims ‎to compare machine learning (ML) models (Random Forest, ‎Light, and CatBoost) and the Personal ‎Computer Storm Water Management Model ‎‎(PCSWMM) hydrological ‎model to predict flash flood susceptibility ‎‎maps (FFSMs) in an arid region (Wadi Qows in ‎Saudi Arabia). Nine independent factors ‎that influence FFSMs in the study area were ‎assessed. ‎Approximately 300 flash flood sites were identified through a post-flood survey after the ‎‎extreme flash floods of 2009 in Jeddah city. The dataset was randomly split into 70 percent for training and 30 percent for testing. The results ‎show that the area ‎under the receiver operating curve (ROC) values were ‎above 95% for all tested ‎models, indicating evident accuracy. The FFSMs developed by the ML ‎‎methods show acceptable agreement with the flood inundation map created using the ‎PCSWMM in terms of flood extension. Planners and officials can use the outcomes of this study to improve the mitigation measures for flood-prone regions in Saudi Arabia.