Key points are not available for this paper at this time.
Landslides are among the most frequent and destructive geological hazards in mountainous regions, posing serious threats to infrastructure and human safety. With the proliferation of high-resolution remote sensing imagery, there is increasing potential for automated landslide extraction. However, most existing methods struggle with limited cross-region generalization and the under-detection of fragmented landslides, due to variations in terrain, scale, and sensor resolution. To overcome these challenges, we propose FMEformer, a frequency-domain–enhanced segmentation framework designed specifically for landslide extraction. The model integrates a Global–Local Frequency Modulation (GLFM) module, which combines Fourier-based global reconstruction with wavelet-based local enhancement guided by local variance, and a Multi-Directional Strip Boundary Enhancement (MD-SBE) module that captures directional boundary features along horizontal, vertical, and diagonal orientations. We evaluated FMEformer against eleven representative deep learning segmentation models, including UNext, UNetFormer, PSPNet, CSUNet, UHRNet, SwinUNet, SegFormer, TransUNet(2021), TransUNet (2024), SKENet, and LSRFormer, covering both CNN-based and Transformer-based architectures. Experiments show that FMEformer improves Intersection over Union (IoU) by at least 6.5%, substantially enhancing landslide extraction accuracy, boundary delineation, and robustness under cross-scenes. This framework provides a scalable solution for precise landslide inventory generation, with strong applicability to disaster risk assessment and emergency response in complex mountainous terrains.
Wang et al. (Wed,) studied this question.