Landslide susceptibility assessment of Kure City (Hiroshima Prefecture, Japan) with machine learning random forest and frequency ratio methods using XRAIN radar-acquired rainfall data

Abstract

Abstract In this paper, we perform an efficiency comparison between two methods of producing landslide susceptibility maps (LSM): the machine learning technique random forest (RF), which is evidenced to be the most efficient for landslide susceptibility assessment tasks (Yilmaz, 2009; Youssef & Pourghasemi, 2021) and the statistics-based frequency ratio (FR) method. The area considered for this study is southern part of Hiroshima Prefecture in Japan, an area known to suffer from rainfall-induced landslide disasters, the most recent one in July 2018. Both the above methods require a collection of landslide conditioning factors (LCFs), which in this study are: 1) geology, 2) altitude, 3) slope angle, 4) slope aspect, 5) drainage density, 6) land use, 7) distance from lineaments, and 8) mean annual precipitation. The rainfall LCF data comprise of XRAIN (eXtended RAdar Information Network) radar records, which are novel in this task of LSM production. The accuracy of the produced LSMs was assessed by the receiver operating characteristic’s (ROC) area under curve (AUC), which is 0.84 for the FR method and 0.92 for the RF method. It is also noteworthy that the RF method is substantially swifter and more practical than the FR method, and it allows for multiple and automatic experimentations with different parameters, providing fine and accurate results with the given data. The results also evidence that machine learning techniques such as the RF method are most advisable for dealing with hazard assessment problems such as the one exemplified in this study, and that XRAIN radar-acquired mean annual precipitation data are effective when used as a LCF in producing LSMs.