Detecting slow-moving landslides using InSAR phase-gradient stacking and deep-learning network

Abstract

Landslides are a major geohazard that endangers human lives and properties. Recently, efforts have been made to use Synthetic Aperture Radar Interferometry (InSAR) for landslide monitoring. However, it is still difficult to effectively and automatically identify slow-moving landslides distributed over a large area due to phase unwrapping errors, decorrelation, troposphere turbulence and computational requirements. In this study, we develop a new approach combining phase-gradient stacking and a deep-learning network based on YOLOv3 to automatically detect slow-moving landslides from large-scale interferograms. Using Sentinel-1 SAR images acquired from 2014 to 2020, we developed a burst-based, phase-gradient stacking algorithm to sum up phase gradients in short-temporal-baseline interferograms along the azimuth and range directions. The stacked phase gradients clearly reveal the characteristics of localized surface deformation that is mainly caused by slow-moving landslides and avoids the errors due to phase unwrapping in partially decorrelated areas and atmospheric effects. Then, we trained the improved Attention-YOLOv3 network with stacked phase-gradient maps of manually labeled landslides to achieve quick and automatic detection. We applied our method in an ∼180,000 km2 area of southwestern China and identified 3,366 slow-moving landslides. By comparing the results with optical imagery and previously published landslides in this region, the proposed method can achieve automatic detection over a large area precisely and efficiently. From the derived landslide density map, we determined that most landslides are distributed along the three large rivers and their branches. In addition to some counties with known high-density landslides, approximately 10 more counties with high landslide density were exposed, which should attract more attention to their risks for geohazards. This application demonstrates the potential value of our newly developed method for slow-moving landslide detection over a nation-wide area, which can be employed before applying more time-consuming time-series InSAR analysis.