Landslide detection by deep learning based semantic segmentation using Sentinel-1 intensity images

Abstract

Abstract Fast-moving landslides associated with earthquakes and/or heavy rains can cause significant human and socioeconomic damage. Satellite synthetic aperture radar (SAR) can observe wide areas regardless of the presence of clouds and sunlight, and thus is a promising tool for detecting landslides immediately after a disaster occurrence. Recently, the application of deep learning-based semantic segmentation to satellite optical imagery is active for the landslide detection, but its application to SAR imagery is still limited. Here we investigated landslide detection capability of deep learning-based semantic segmentation using pre- and post-disaster Sentinel-1 SAR intensity images. We used U-Net as the deep learning model and validated the method on four disaster cases composed of two earthquake and two heavy rain events. In the cases of the 2018 Hokkaido Eastern Iburi earthquake and the July 2017 Northern Kyushu heavy rainfall, detection models trained in the same single case or on multiple cases could identify large landslides and spatial concentrations of damage areas. The model trained in the 2018 Hokkaido Eastern Iburi earthquake provided the highest Kappa coefficient (0.624) when we applied it to the same case. The proposed method could detect landslides that had not been detected by emergency aerial observation, demonstrating the effectiveness of the combination of satellite SAR and deep learning for the landslide detection. However, the detection performance of trained models was degraded when they were applied to cases with different characteristics in terms of topography, vegetation and landslide occurrence mechanisms from training cases. Therefore, our results indicated that detection models should be trained in various cases for improving versatility. For future practical use, further validations by other disaster cases using various kinds of satellite SARs are needed.