An Embedding Swin Transformer Model for Automatic Slow-moving Landslides Detection based on InSAR Products

Interferometric Synthetic Aperture Radar (InSAR) technology is the most advanced and effective method for monitoring large-scale slow-moving landslides. However, automatic landslide detection based on InSAR products regarding landslide samples and deep learning models is still challenging. Different InSAR products are inconsistent for slow-moving landslide detection due to the fuzzy boundaries of potential landslides and complex deformation characteristics. In addition, the accuracy of existing landslide detection models is not high because multi-scale factors in feature abstract and feature fusion are rarely considered. This paper proposes a multi-scale Swin Transformer InSAR products detection network (MSIDNet) to detect slow-moving landslides automatically. Firstly, we adopt an advanced Swin Transformer as the backbone to extract features and multi-scale spatial-temporal attention blocks in the neck to improve the feature fusion capability. Meanwhile, to train the new model, we built a landslide dataset based on visual interpretation, including deformation rate, phase gradient, and C-index (GRCI). Experiments demonstrate that our proposed method outperforms the existing deep learning models like Faster R-CNN and Yolov3. The GRCI dataset is more efficient and less biased than the traditional deformation rate and phase gradient datasets. The precision of detected landslides in a given testing area is 0.89, and it has good generalization ability. This study provides a new dataset and method for slow-moving landslide detection based on InSAR products.