MB-Net: A network for accurately identifying creeping landslides from wrapped interferograms

Abstract

The efficient and automated identification of landslide hazards is essential for socio-economic development and human safety. Integrating the feature extraction capabilities of deep learning with the millimeter-level precision of Interferometric Synthetic Aperture Radar (InSAR) technology establishes a foundation for this task. However, current methods require unwrapping interferograms, and even converting them into deformation products before identifying landslide hazards. This process is susceptible to unwrapping errors, resulting in inefficient data utilization, and demands considerable time and labor. To overcome these challenges, wrapped interferograms are directly utilized for identifying creeping landslides. In this study, trigonometric functions are applied to improve the representation of interferograms and to further enhance the data through rendering. Secondly, a multi-branch semantic segmentation network (MB-Net) was designed, with parallel branch encoding and progressive feature fusion to optimize the model’s ability to learn interferometric phases. Experimental results indicate a good performance, with the F1-score of 80.91 %, the Intersection over Union (IoU) of 67.94 %, and the Matthews correlation coefficient (MCC) of 80.16 % on the ISSLIDE dataset. To further validate the generalization capability of MB-Net, the public COMET-LiCS Sentinel-1 InSAR portal data was utilized, focusing on the middle reaches of the Jinsha River in China. The results highlight MB-Net’s efficacy in spatial transferability analysis. These findings emphasize the potential of our approach for large-scale landslide hazard identification, providing a crucial foundation for the utilization of interferograms in creeping landslide detection.

Keywords

• InSAR
• Creeping Landslides
• Deep Learning
• Semantic Segmentation