Yosuke KATAHIRA

Development of Wide-Area Hazard Assessment for Slope by Combining Machine Learning and Image Recognition

Satoru OHTSUKA

Conventional slope hazard assessments have a problem which does not suitably consider the topographical features of slope because it is based on individual mesh units. This study, therefore, developed an analytical method to identify hazardous slopes and to evaluate the hazard level of slopes by incorporating geomorphic factors of slopes at various scales into the hazard assessment. In the first stage, the slope hazard assessment was conducted by considering macroscope topographic information of 5m and 10m squares around the mesh to the conventional mesh-based prediction. In the second stage, using the results of the slope hazard assessment by machine learning, the hazardous slope blocks were extracted through the artificial intelligence technology with image recognition. As the results, it was found that in the first stage of machine learning, the accuracy of mesh-by-mesh prediction was improved by inputting the macroscopic information of various ranges on the surrounding area. Among the prediction methods in machine learning, ensemble learning, such as bagging trees and boosting trees, was found to be the most effective for evaluating hazardous slope. In the second stage of image recognition, it was shown to extract collapsed blocks from the results of hazard assessment obtained by machine learning. YOLO and pix2pix were able to extract collapsed blocks, but the collapsed blocks were smaller and appeared thinner than the actual extent of collapse. On the other hand, Semantic Segmentation was able to extract collapsed blocks by compensating for misclassified areas in the machine learning. This confirmed that Semantic Segmentation is currently the most effective method for image recognition. In conclusion, the input of topographic information at different scales is highly effective in machine learning for assessing the wide-area hazard of slopes at the mesh level. Besides, It is made clear that image recognition using machine learning results can extract collapsed blocks. However, this research is at the beginning of slope hazard assessment, it is strongly expected to further developed by settling many subjects.