Robust Ensemble Machine Learning for Flash Flood Susceptibility Mapping Across Semiarid Regions

Abstract

Flash floods cause severe environmental and socio-economic impacts in arid and semi-arid regions. This study aims to improve the accuracy of flash flood susceptibility mapping in southwestern Morocco’s Assaka watershed by using an ensemble of machine learning models. Four models, Logistic Regression (LR), Multivariate Discriminant Analysis (MDA), Naïve Bayes (NB), and Multilayer Perceptron (MLP), were trained on a flood inventory of over 1.5 million data points and 14 environmental factors (e.g., altitude, slope, land surface temperature, soil moisture index). Each model produced a susceptibility map, and a voting ensemble of the top-performing models (all above 80% accuracy) further improved reliability. The MLP achieved the highest predictive performance, followed closely by LR and MDA. Sensitivity analysis identified altitude, topographic position index, land surface temperature, and soil moisture index as the most influential factors. The ensemble susceptibility map highlights densely populated areas near the city of Guelmim and infrastructure along major rivers as most prone to flash flooding. These findings enable practical mitigation measures such as improved drainage, early warning systems, and better land-use planning in high-risk zones. Integrating multiple models in an ensemble is a novel approach that reduces uncertainty and provides a more robust tool for flash flood risk prediction.