Flood susceptibility prediction using four machine learning techniques and comparison of their performance at Wadi Qena Basin, Egypt

Abstract

Floods represent catastrophic environmental hazards that have a significant impact on the environment and human life and their activities. Environmental and water management in many countries require modeling of flood susceptibility to help in reducing the damages and impact of floods. The objective of the current work is to employ four data mining/machine learning models to generate flood susceptibility maps, namely boosted regression tree (BRT), functional data analysis (FDA), general linear model (GLM), and multivariate discriminant analysis (MDA). This study was done in Wadi Qena Basin in Egypt. Flood inundated locations were determined and extracted from the interpretation of different datasets, including high-resolution satellite images (sentinel-2 and Astro digital) (after flood events), historical records, and intensive field works. In total, 342 flood inundated locations were mapped using ArcGIS 10.5, which separated into two groups; training (has 239 flood locations represents 70%) and validating (has 103 flood locations represents 30%), respectively. Nine themes of flood-influencing factors were prepared, including slope-angle, slope length, altitude, distance from main wadis, landuse/landcover, lithological units, curvature, slope-aspect, and topographic wetness index. The relationships between the flood-influencing factors and the flood inventory map were evaluated using the mentioned models (BRT, FDA, GLM, and MDA). The results were compared with flood inundating locations (validating flood sites), which were not used in constructing the models. The accuracy of the models was calculated through the success (training data) and prediction (validation data) rate curves according to the receiver operating characteristics (ROC) and the area under the curve (AUC). The results showed that the AUC for success and prediction rates are 0.783, 0.958, 0.816, 0.821 and 0.812, 0.856, 0.862, 0.769 for BRT, FDA, GLM, and MDA models, respectively. Subsequently, flood susceptibility maps were divided into five classes, including very low, low, moderate, high, and very high susceptibility. The results revealed that the BRT, FDA, GLM, and MDA models provide reasonable accuracy in flood susceptibility mapping. The produced susceptibility maps might be vitally important for future development activities in the area, especially in choosing new urban areas, infrastructural activities, and flood mitigation areas.