Automatic identification of active landslides over wide areas from time-series InSAR measurements using Faster RCNN

With the combined effects of climate change and anthropogenic disturbance, landslide hazards have progressively increased and emerged as one of the most significant natural threats to socio-economic safety and human life. Synthetic aperture radar interferometry (InSAR) can measure subtle ground displacement and thus has immense potential for detecting active landslides. However, the operational application of InSAR for landslide detection and inventory update in wide-area is still hindered by the high labor and time costs for visual interpretation and manual editing of InSAR results. Aiming at this problem, we developed a novel method using InSAR and convolutional neural network (CNN) for automated identification of active landslides over wide areas. It first performs InSAR analysis to produce a surface displacement velocity map of the target region and then employs an improved Faster RCNN based on attended ResNet-34 and Feature Pyramid Networks (FPN) to detect active landslides from the velocity map. Taking the Guizhou province in southwest China as a case study, we processed 1168 scenes of Sentinel-1 images and 473 scenes of PALSAR-2 images to derive the surface displacement and identified 1627 active landslides, including 326 manually labeled landslides and 1301 landslides automatically detected by Faster RCNN. The improved Faster RCNN achieved good recall, precision, F1 score, and average precision (AP) at 91.49%, 91.33%, 0.914, and 0.940, respectively. Further experiments indicated that the trained Faster RCNN showed satisfactory applicability and result accuracy for different test areas and various InSAR techniques. Therefore, the proposed approach has great potential applications for building inventories of active landslides over wide areas and regularly updating the records, which is crucial for preventing landslide disasters and mitigating losses.