Creation of a Landslide Susceptibility Map Using Short‐Term Data From the July 2018 Heavy Rainfall in Southern Hiroshima Prefecture

ABSTRACT Landslides are a severe geohazard that can cause significant loss of life and property. Given budgetary constraints, landslide susceptibility maps are an effective tool for determining the priority of mitigation activities to minimise the impact of landslide‐related damage. The goal of this study was to improve the accuracy of susceptibility maps by using short‐term event‐triggered data from a single disastrous event. A landslide susceptibility map was created using event history data from subsequent landslides in southwestern Hiroshima Prefecture, Japan, in July 2018. First, 13 factors and triggers related to the likelihood of slope failure were collected from various agencies, including topography, geology and rainfall data. Then, six machine learning models were applied to investigate the relationships between causal factors and observed landslide events: Naive Bayes classification, logistic regression, support vector machine, random forest, gradient boosting and artificial neural networks. The results showed that the random forest model achieved the highest accuracy (ACC) with a value of 0.774, while gradient boosting had the highest area under the curve (AUC) with a value of 0.829. Consequently, ensemble learning using decision trees produced a highly accurate map. The importance evaluation (Permutation Importance) indicated that the watershed area had the most significant impact on accuracy. Short‐term triggers were found to have less importance. The study demonstrated the significance of using proper watershed delineation methods when creating maps with short‐term inventory and trigger data. In the future, developing methods that ensure accuracy regardless of watershed processing could result in more versatile susceptibility maps.