Application of Multi-Input Deep Learning to the Setting of Steep Slope Warning Areas

Abstract

In Japan, the Ministry of Land, Infrastructure, Transport and Tourism (MLIT) designates “Steep Slope Warning Areas” as zones where there is a risk of slope failure on inclines of 30 degrees or more, particularly in locations where such events could threaten the lives and property of residents. These areas also include zones where certain land use restrictions are necessary to prevent the promotion or triggering of slope collapse. However, the field surveys required for such designations involve significant labor and cost, leading to growing interest in more efficient and automated methods. Previous studies have attempted to extract steep slope warning areas using image segmentation techniques based on deep learning models such as U-Net, achieving an accuracy of approximately 80%. Building upon these results of previous research, this study aims to further improve accuracy and efficiency by constructing a novel deep learning model utilizing multi-input data—including digital elevation models (DEMs), shaded relief maps, and aerial photographs—and evaluates the utility of each data type.