Rin KOBAYASHI

Application of Slope Hazard Assessment Model by Machine Learning to Non-training Area

Satoru OHTSUKA

With the enactment of the Landslide Disaster Prevention Act in 2001, non-structural measures against slope hazards have been promoted in Japan, and AI-based research on slope hazard assessment has been conducted in recent years. In Step 1, hazard assessment was conducted on a mesh-by-mesh basis using machine learning. Bagging was the preferred classification method, and the best performing model was the one that used multiple scales of terrain information for all areas and added the degree of openness and ridge and valley. In Step 2, the hazard ratings from Step 1 were converted to images, and pix2pix was used to extract landslide-prone blocks. The comparison of model losses did not show much difference among the areas, but the verification using test data showed that the results differed depending on the area. This may be due to the fact that the size, number, and shape of collapses differed by area, and that the test data included collapses of a scale that had not been trained. This study obtained that large to middle-scale landslides are suitable for hazard assessment and it is difficult to grasp small-scale landslides, and that reducing the image size ensures good quality training data. Furthermore, it showed the possibility of improving performance by processing the hazard evaluation values. The model created in Step 1 and Step 2 was validated in regions outside the training area. In Step 1, if the elevation distribution differs significantly between the training and validation areas, the prediction cannot be performed correctly. When elevation is excluded, the true positive rate was greatly improved, indicating the effectiveness of the model. The model that considered topographic information at multiple scales and added the degree of openness and ridge and valley was found to be superior in non-training regions as well. The validation of the model in non-training regions showed that the performance was generally lower, but the results for areas with similar topography were relatively good. In Step 2, it was shown that image generation models created in non-training regions can be applied only when accurate hazard evaluation values are used. Further studies are needed to achieve highly accurate hazard assessment in non-training regions.