There has been an increasing demand in recent years for detailed and accurate landslide maps and inventories in disaster-prone areas of subtropical and temperate zones. Standard Landslide susceptibility mapping methods require detailed fieldwork to be conducted by knowledgeable, skilled professionals. When predicting landslides, it is important to understand past landslide cases and prepare for situations in which the same phenomena occur. Developing automatic analysis methods using deep learning can contribute to the sophistication and cost of screening.
In this study, models using the automatically construct high-performing convolutional neural network (CNN) and Visual Transformation　(ViT) architectures for landslide detection, were applied and their outcomes were compared for landslide susceptibility mapping at Kii-peninsular, Japan. As a first step, a total of 101 landslide and non-landslide points were identified and divided into 70% training and 30% validation datasets. Eight landslide influence factors were used, including slope angle, eigenvalue ratio, curvature, underground-openness, overground-openness, topographic witness index (TWI), wavelet and elevation. Experimental results of model evaluation using receiver operating characteristics and area under the curve (ROC, AUC), and accuracy showed that the CNN models were more accurate than ViT model in predicting landslides spatially. Furthermore, the landslide susceptibility map is consistent with the trends in the distribution of gentle slopes and knick lines unique to the study area and can be used as a powerful method for future landslide prediction.