Flood susceptibility assessment using three machine learning techniques and comparison of their performance

Abstract One of the most common natural disasters is flooding, which has the potential to seriously harm environments and infrastructure. Flood susceptibility mapping (FSM) is the main way to manage flood risk. It measures how likely a region is to flood in a quantitative way. The purpose of this study was to develop state-of-the-art ensemble machine learning (ML) models for flood prediction and to identify the most suitable approach for accurate flood susceptibility mapping. This study leverages diverse datasets, including elevation, slope, aspect, plan curvature, topographic wetness index, stream power index, distance from rivers, soil, rainfall, land use/land cover, and drainage density, which were used as conditioning factors to evaluate flood susceptibility in the Choke Watershed. Three machine learning (ML) algorithms were employed: Random Forest (RF), Gradient Boosting (GB), and Extreme Gradient Boosting (XGBoost). Model performance was assessed using confusion matrix metrics and the area under the receiver operating characteristic curve (AUROC). The Gradient Boosting (GB) and Extreme Gradient Boosting (XGBoost) models scored the highest in terms of test accuracy (0.97), followed by RF (0.96). This study is the first application of these models in the Choke Watershed for flood susceptibility mapping, with potential for broader applications to other natural disasters, including earthquakes and landslides. The results help strengthen global efforts aimed at mitigating natural disaster risks, particularly in Ethiopia, and advancing environmental sustainability.