Research on the acquisition of topographic interpretation capabilities using deep learning and the generation of regional landslide susceptibility maps.

The purpose of this study is to improve the prediction and risk assessment of landslides that cause slope hazards. We propose a method for the automatic detection of landslide topography over a regional area by using a deep learning algorithm to learn and replicate expert topographic interpretation techniques. Landslide topography and geological conditions have a significant impact on landslide prediction. In Japan, hazard maps have been created in the past based on expert topographic interpretation [1]. Deciphering requires a high degree of expertise and a considerable amount of time. Since Japanese landslide topography hazard maps have not been updated with high-resolution topography data since the 2000s, there is a need to develop more efficient and precise prediction techniques.

In this study, a deep learning model was used to acquire the topographic information characterizing landslide topography. This process involves acquiring expert topographic interpretation skills through deep learning. Approximately 10,000 landslide-related images were used as training data. These images were selected based on features that experts could recognize as landslide topography. An area of 18,000 km² in southwestern Japan was analyzed for topographic information, with half of this area used for training data. Model performance was verified in an unused area of 100 km².

The results showed that the detection rate reached approximately 80%, confirming that the automatic detection of landslides is feasible to some extent. The analysis was completed in about one hour, whereas it would take an expert several weeks. After deep learning inference, it took several hours to create a regional susceptibility map via GIS.

The acquisition of topographic interpretation techniques through deep learning is feasible and can be a method to accelerate and objectify the assessment of the likelihood of landslide topography over large areas in the future. When combined with remote sensing technology, dynamic hazard assessment will be possible. This is expected to be a next-generation tool for landslide hazard assessment. However, there are some points to keep in mind when introducing this method. It is necessary to have experts prepare training data and check inference results, and it is important to maintain an accurate disaster inventory. Additionally, it is crucial to continue fact-checking the results of deep learning inference. By fulfilling these requirements, deep learning can be used as a reliable analysis method. As a result, disaster preparedness planning will become more efficient, and society’s resilience to disasters will be improved.

[1] National Research Institute for Earth Science and Disaster Resilience. 2008. landslide topography map series.