Application and Challenges of a Machine Learning-Based Zero-Order Basin Extraction Method for Preventive Maintenance Against Large-Scale Disasters During Havey Rainfall: A Case Study the 2024 Noto Peninsula Heavy Rain

This study focuses on the applicability of AI-based zero-order basin extraction method to large-scale landslides, using the analogue analysis techniques of terrain expert analysts as a reference. Based on the results of the large-scale landslides that occurred in the Noto Peninsula of Japan during the heavy rains in September 2024, this study verifies the reliability of the extraction method. The extraction of zero-order basin, which are presumed to be the sources of slope failures and debris flows, has traditionally been identified by the analysts based on criteria such as the numerical topographic characteristics of zero-order basins and other micro-topographic classifications effective for zero-order basin extraction. Using the analogue extraction results obtained by the analysts as a reference, we developed a model that extracts zero-order basins from digital terrain maps and digital elevation models as potential hazardous areas prone to landslides using a fully convolutional network (FCN), a machine learning technique (References). Regarding the AI extraction accuracy, we evaluated the AI model's recall rate for missed detections and the precision rate for false detections based on extraction data from the analysts in the verification area, confirming 88% and 64%, respectively (References). We applied the AI method to the 2024 Noto Peninsula heavy rain disaster. The results are as shown in Figure 1 and Table 1. The heavy rainfall exceeding 500 mm resulted in landslides and surface collapses in an area spanning 40 km east-west in the northern part of the Noto Peninsula. Using aerial photographs, we identified 7,468 sites of landslides and debris flows. As a result of extracting zero-order basins using the AI method, the total number of extracted zero-order basins was 12,111 sites. To assess the accuracy of the AI method, we calculated the precision, capture rate, and miss rate based on actual disaster sites. The results were 31%, 50%, and 50%, respectively. The current capture rate of the AI method is 50%, which is not so low but there is room for improvement. However, we demonstrated that the AI method could extract thousands of potential landslide sites within one hour, which would otherwise take ten days and over using the workforce, thereby improving the efficiency of the extraction process.