Landslides are a prominent and frequently occurring geological hazard worldwide, posing an urgent demand for the advancement of accurate and timely landslide segmentation techniques. However, current deep learning methods based on remote sensing imagery still face challenges in precisely delineating landslide boundaries and comprehensively segmenting the landslide body. To address these challenges, this study proposes the dual-stream frequency-guided network (DFG-Net), a segmentation architecture based on frequency-domain identification. This network employs a parallel dual-stream backbone comprising a Swin Transformer and a ResNet to capture low-frequency contextual and high-frequency boundary information, respectively. A dedicated frequency-domain enhancement module, the Fourier multiscale low-pass module, and a boundary guidance module were also designed to refine these two types of features. Finally, a guided cross-attention fusion module is used to enable interactions and improve these two information streams. Experiments were conducted on this study’s self-built Mianyang landslide remote sensing dataset and the public GVLM landslide remote sensing dataset. The findings demonstrate that DFG-Net substantially outperforms various models on key metrics, achieving an intersection over union (IoU) of 72.13% and an F1-score of 83.81% on the former dataset, and an IoU of 82.13% and an F1-score of 90.19% on the latter. Furthermore, the model’s generalisation ability was further confirmed in Kupang, Indonesia.
Zhong et al. (Mon,) studied this question.