Examination of the efficacy of machine learning approaches in the generation of flood susceptibility maps

Abstract

AbstractFlash floods stand as a substantial peril linked to climate change, imposing a severe menace to both human existence and built structures. This study aims to assess and compare the effectiveness of four distinct machine learning (ML) methodologies in the production of flood susceptibility maps (FSMs) in Ibaraki prefecture, Japan. Additionally, the investigation aims to examine the influence of excluding plan and profile curvature factors on the accuracy of the resulting maps. The dataset comprised 224 spots, consisting of 112 flooded and 112 non-flooded locations, and 11 environmental factors. The models were trained using 70% of the dataset, while the remaining 30% was utilized for model evaluation using the ROC curve method. The results indicated that both the ANN-MLP and SVR models achieved notable accuracy, with area under curve values of 95.23% and 95.83% respectively. An intriguing observation was made when the plan and profile curvature factors were excluded, as it led to an improvement in the accuracy of the ANN-MLP model, resulting in an accuracy of 96.7%. Furthermore, the generated FSMs were classified into five distinct hazard levels. The northern region of the maps predominantly exhibited very low and low hazard levels, while areas located in the southern region, closer to main streams, demonstrated considerably higher hazard levels categorized as very high and high. Ultimately, this study marks novel endeavor to investigate the impact of the curvature factor on the precision of machine learning algorithms in the creation of FSMs, which serve as fundamental tools for subsequent investigations.