2019 Volume 56 Issue 5 Pages 255-263
LiDAR (light detection and ranging) before and after collapse in 38 cases of collapse induced by Typhoon Talas in 2011 suggested the presence of gravitational deformation in the terrain before the collapse. In this study, applying a wavelet function to the LiDAR data, a micro topology map was prepared to conduct deep learning. The authors analyzed 9,206 pieces of50-pixel tiled images including those of non-failure sites using a convolutional neural network, and achieved 80.8％ accuracy rate in the failure sites and 91.1％ accuracy rate in the non-failure sites. As a result of detailed studies, analysis results of the failure sites did not include clear scarps, but included irregular unevenness and small cliffs, in which learning effect was recognized. On the other hand, among the analysis results of the non-failure sites, those tiles of images determined as collapsed lands were interpreted as errors, but had topography including features found in the learning outcome. From these results, we were able to understand that they were not misreading but they had terrain elements that would become unstable in future. We can expect that those results will be utilized for forecasting future slope failure.
