Author Name : SUGITANI Rinsuke

Research Topic : Analysis of slope attributes and risk of wide-area slope through case analysis using AI

Supervisor Name : OHTSUKA Satoru

Summary : In recent years, landslides caused by rainfall have been increasing in Japan. AI has been used to predict slope failures, but most of them use independent meshes and fail to take surrounding topographic features into account. Among actual collapses, large-scale collapses are affected by the macro topography, while small-scale collapses are affected by the micro topography. Therefore, conventional methods that investigate the relationship between collapse phenomena at the mesh level do not reflect the collapse characteristics of the slope. The objective of this study is to extract collapse blocks for regions with different topographic characteristics by performing prediction using machine learning with wide-area information as input and image recognition with the results as input. At the same time, we identify the best topographic data for collapse prediction in each region. For prediction by machine learning, collapse prediction is performed when the size of the DEM and the size of the moving average are varied, and the differences in the analysis system depending on the topographical data used are discussed. Prediction by image recognition is aimed at extracting the areas with high collapsed land densities as collapsed blocks using an image transformation algorithm.
Machine learning analysis was performed for each area, and in all areas, higher accuracy was obtained when moving average values were used. In the case of 1mDEM, the accuracy tended to be higher when using only 10m moving average topography (10 items) than when using only 10 indicators (10 items) or increasing the number of training items by combining indicators and moving average values (30 items). The superiority of the moving average was also observed in the case of the 5mDEM topographic data. In the hazard assessment map, the number of ambiguous areas with hazard assessment values between 0.5 and 0.7 decreased when the number of study items was increased, suggesting that the model would become more clearly defined. However, when 5mDEM and 10mDEM are used, there is a significant decrease in prediction performance, suggesting that it is necessary to take into account detailed topographic undulations.
Next, based on the results of machine learning, we applied the image transformation algorithm to the case with the highest accuracy in each area under the same conditions. As a result, collapsed blocks were extracted in areas with a high concentration of collapsed areas predicted by machine learning. In terms of performance metrics, the correctness rate and accuracy improved in all cases. In contrast, the true positive rate decreased in all cases. The reason why image recognition improves the correctness rate and accuracy is thought to be that the image recognition tends to extract collapsed blocks where the concentration of collapsed areas is high, while small collapsed areas are overlooked. This is thought to be the cause. In fact, Area-B, which has many large collapsed areas of 500 m2 or more, showed a significant increase in the percentage of correct answers and accuracy, and a significant decrease in the percentage of true positives.