Extracting deforming landslides from time-series Sentinel-2 imagery

Detecting landslide precursors early is crucial for mitigating regional landslide hazards. Yet, it is challenging to recognize true regional surface deformation among abundant background noises. Previous work has been focused on the detection of already-known landslide precursors or retrospective deformation of landslides. We proposed a new method to extract deforming landslides from background noise by using time-series Sentinel-2 images. The method is tested along a section of the Jinsha River in southwest China. First, we selected 104 clear images from 412 Sentinel-2 images December 2018 to October 2021 to compose 141 image pairs. Second, we derived three clusters of cumulative deformations starting from December 2018 with approximately 1-year interval. Third, we used linear regression models to fit cumulative deformations and calculated the p-values for each pixel. Finally, we used the low p-value (p = 0.0005) as a threshold to suppress background noises. The proposed method is very efficient to eliminate almost all background noise and isolate deforming landslides. Application of this method enables extraction of landslide precursors over large mountain regions.

Automated summary

A new method is proposed in this study to extract deforming landslides from background noise using time-series Sentinel-2 images. The method was tested along a section of the Jinsha River in southwest China and was found to be effective in eliminating background noise and isolating deforming landslides.