11:15 AM - 11:30 AM
[STT39-03] Landslides detection by machine learning and deep learning using Sentinel-1 intensity images
Keywords:Landslide, Synthetic Aperture Radar, Machine learning, Sentinel-1, Intensity image
We used multiple features such as backscatter intensity difference between the time-averaged images in the pre-disaster stage and a single intensity image that is available immediately after the disaster occurrence and terrain data for training. Similarly, we performed pixel-based machine learning using Gradient Boosting Decision Tree method and object-based deep learning using U-Net model, and then compared the results. U-Net is a semantic segmentation model, which can learn without losing the location information of objects compared to a simple FCN (fully convolution network). The result of landslide detection was performed for square areas with several hundred meters on a side, because there were a lot of small noises in the raw-resolution image and we thought the detailed shape and size of each landslides is not so important in practical use for administrations and municipalities. The training of the model and detection accuracy was evaluated on two cases, the 2018 Hokkaido Eastern Iburi Earthquake and the 2017 heavy rain event in Northern Kyushu.
In the results of machine learning using Gradient Boosting Decision Tree, the detection model could detect part of landslides, most of which were relatively large-size landslides in the case of 2018 Hokkaido Eastern Iburi Earthquake, achieving a Recall of 12.01%, a Precision of 58.87%, a Kappa coefficient of 0.20. On the other hand, there were many false landslide detections in 2017 heavy rain event in Northern Kyushu, possibly due to sudden surface change of scatter characteristics due to the strong rainfall. Similarly, we performed deep learning using the U-Net model and found that the accuracy was significantly improved compared to pixel-based machine learning, achieving a Kappa coefficient of 0.48. We confirmed an improvement in accuracy by deep learning in 2017 heavy rain event in Northern Kyushu, but there were still many incorrect detections compared to the earthquake landslide event in Iburi.