Quick and automatic detection of co-seismic landslides with multi-feature deep learning model

Abstract

Co-seismic landslide detection is essential for post-disaster rescue and risk assessment after an earthquake event. However, a variety of ground objects, including roads and bare land, have spectral characteristics similar to those of co-seismic landslides, making it difficult to gather information and assess their impact rapidly and accurately. Therefore, an automatic detection method based on a deep learning model, named ENVINet5, with multiple features (ENVINet5_MF) was proposed to solve this problem and improve the detection accuracy of co-seismic landslides. The ENVINet5_MF method is advantageous for co-seismic landslide detection because it features a landslide gain index (LGI) that effectively eliminates the spectral interference of bare land and roads. We conducted two experiments using multi-temporal PlanetScope images acquired in Hokkaido, Japan, and Mainling, China. The accuracy evaluation and rationality analysis show that ENVINet5_MF performed better than comparative methods and that the co-seismic landslide areas detected by ENVINet5_MF were the most consistent with ground reference data. The findings of this study suggest that ENVINet5_MF can provide an efficient and accurate method for coseismic landslide detection to ensure a rapid response to co-seismic landslide disasters.