|
 |
In this study, our aim is to use image recognition techniques to identify slopes at high-risk of slope failure. The feature of this study is that a three-dimensional terrain model is generated using the surrounding elevation values of each grid obtained from aerial laser surveying and the high-risk areas of landslide disasters are assessed by a Convolutional Neural Network (CNN) model trained on three-dimensional color images in which the three terrain factors are converted into color index. First, a standardized three-dimensional terrain model is constructed by extracting the elevation values of the surrounding grid and subtracting its own elevation values using the coordinate values of the grid data. We added per-grid slope and Laplacian terrain quantities to this three-dimensional terrain model. We have developed a method to quantitatively represent these three terrain quantities as color images by reflecting them in a 16-bit RGB color index. With the developed method, the three-dimensional terrain model becomes a single-color image. More than 100,000 images could easily be generated for each of the three watersheds included in the study area. A high-risk area was defined as the case where the center of the three-dimensional terrain model was included in the slope failure areas had occurred in the past. These images were used as explanatory variables to train the CNN. As a result, the occurrence and non-occurrence rates were balanced and could be evaluated at approximately 80%. Despite the high resolution of the evaluation unit being a 1m grid, the proposed method resulted in high-risk grids forming clusters that matched the surrounding terrain, providing a terrain-consistent evaluation. Additionally, the results were consistent with the terrain characteristics of slope failure phenomena during heavy rainfall as previously indicated. This suggests that the proposed method can accurately capture the fine terrain irregularities compared to traditional methods.
Research papers (academic journals)
