Takero AIKAWA

Research on AI Wide-Area Hazard Assessment of Slopes Combining Machine Learning and Image Recognition Techniques

Satoru OHTSUKA

Conventional slope risk assessments are based on individual mesh-based assessments and do not take into account the topographical characteristics of slopes. In this study, we attempted two-stage prediction by machine learning with wide-area information as input and image recognition with the results as input, referring to the findings of previous studies.In the first stage, point-wise risk assessment by machine learning, we used 5m and 10m moving averages of topographic information, and additionally input the above-ground openings, underground openings, and ridge and valley openings used in red 3D maps to improve the accuracy of the existing model with wide-area information for the July 2011 heavy rainfall event. In this case, the wide-area information was used.
The effectiveness of the model proposed in this study was demonstrated by examining a previous model with wide-area information and a 1-meter mesh model without wide-area information.The applicability of the model to other regions with different triggers and predispositions for slope failure was examined, and it was found that the failure characteristics are highly regional. However, the proposed method is considered to be highly applicable to regions where the triggers and predisposing factors for slope failure can be treated as equivalent.In the second stage, extraction of collapse blocks using pix2pix, we first examined the optimal model for extracting collapse blocks in pix2pix, and the size and type of training images. The results showed that a model trained by dividing the image into small pieces with a grayscale hazard rating was optimal.The applicability of the model created by image recognition to cases where the accuracy of machine learning is low was examined, and it became clear that the accuracy of the model depends on the accuracy of the machine learning. Furthermore, the applicability to other regions was examined, and it was confirmed that the generative model has little influence on regional characteristics and can be applied to other regions.Finally, we discussed the issue of improving the accuracy of the generative model, and showed the need to improve the quality of risk assessment rather than simply increasing the number of training images.