ABSTRACT Landslides represent a significant hazard to human life and property in the Hengduan Mountains. Precise identification of landslide boundaries plays a critical role in disaster mitigation and response efforts. Therefore, this study constructed a landslide sample dataset by integrating Sentinel‐2 multispectral remote sensing imagery with elevation and slope data, and employed the SegFormer semantic segmentation algorithm to build a landslide identification model. Additionally, three comparative algorithms—DeepLabv3+, PSPNet, and U‐Net++—were introduced for performance benchmarking. The performance of each model in landslide identification under the complex topographic conditions of the Hengduan Mountains was evaluated using accuracy metrics such as MIoU, F 1 ‐Score, and overall accuracy (ACC). Furthermore, error visualization techniques were applied to analyze the causes of false positives (FP) and false negatives (FN) in the SegFormer model. The results show that the SegFormer model achieved the highest performance among the four models, with MIoU and F 1 ‐Score values of 71.18% and 64.39%, respectively. The SegFormer model achieved the highest performance among the four models, with an MIoU of 71.18% and an F 1 ‐Score of 64.39%, demonstrating strong boundary extraction capability and suitability for landslide identification in the Hengduan Mountains. The FP and FN of the model were primarily attributed to mixed‐pixel effects, spectral similarities among surface features, and differences between new and old landslides. The findings of this study provide methodological references and technical support for landslide identification in the Hengduan Mountains and other regions with complex terrain.
Yin et al. (Sun,) studied this question.