Early Detection of Disaster Areas Using a Landslide Occurrence Probability Model Considering Predisposing Factors and Triggers Combined with Deep Learning

Abstract

Early identification of damage after a landslide is important. However, image deciphering by specialists is time-consuming and costly. In recent years, there has been discussion on delegating the image interpretation part to AI. Yet, a comprehensive model capable of detecting widespread and simultaneous sediment-related disasters has not been established. In addition, these models often rely solely on image data and do not consider the specific characteristics of the affected areas. Therefore, in this study, we developed a damage detection model that enables early detection of landslides using YOLO. Additionally, we developed a model that categorizes the factors causing landslides in the July 2018 torrential rainfall into predisposing factors and triggers, and quantitatively evaluates them using binomial logistic regression analysis. The combination of these approaches was shown to complement the results obtained from using only image decipherment.