Daiki IGARASHI

Consideration on Sophistication of Slope Hazard Assessment by Machine Learning

Satoru OHTSUKA

In wide-area slope hazard analysis, statistical and AI evaluations are conducted for each hazard factor, but there is a problem of a large discrepancy with the actual phenomenon of slope failure, which is a mass movement phenomenon. This study aims to advance the assessment of wide-area risk assessment of slopes. We used a two-step analysis method consisting of machine learning that has been developed in our laboratory to simplify and increase the accuracy of the AI assessment model for expanding the range of its application.

In the first step of the machine learning of topographic information, we used moving averages to take into account the topography of the target location as well as nearby topographic features. Various analyses were performed by manipulating the amount and scale of terrain information, and the differences in prediction performance were verified. The analysis revealed an analytical method that is more accurate with information of high importance and with attention to correlations with other factors. 
In the second step of the analysis to extract collapse hazard blocks using image recognition AI, the hazard assessment maps obtained in the first step were combined with multiple terrain quantities and analyzed for different input images to evaluate the generated shapes and performance. Although the extraction of collapse hazard blocks was successful, the shape of the collapse blocks was affected by the choice of topographic information.

Using the above analysis method, we conducted a comprehensive analysis of its effectiveness and versatility in the study area and in areas with different triggers and predispositions to slope failure. Based on the analysis results, we proposed two methods for efficient slope hazard assessment. However, the proposed model is close to the limit of its predictive performance, and further improvements are needed to improve the performance of risk assessment.