Answers are grounded in this paper and cite their sources.
Accurate segmentation of landslides from high-resolution optical remote sensing imagery is pivotal to proactive disaster prevention, mitigation, and postdisaster emergency response. However, current deep learning-based approaches still face critical bottlenecks in robustly handling multiscale landslide bodies and precisely delineating complex boundaries, particularly in challenging terrains. To address these limitations, we introduce multiscale context fusion network for landslide segmentation (SlideMS-Net), a novel multibranch encoder–decoder network specifically designed for fine-grained landslide segmentation. In the encoder, we propose a high-fidelity multiscale context-aware encoding mechanism. By leveraging parallel dilated convolutions, our design effectively captures multiscale semantic information, while integrating local high-order feature enhancement with global channel attention recalibration, thereby strengthening the model's capability to focus on critical landslide features. In the decoder, we adopt a jointly optimized strategy that combines efficient orthogonal-branch upsampling and multibranch feature fusion. This strategy not only significantly enhances the recognition performance for both large-scale landslides and subtle boundary details, but also substantially reduces computational complexity. Extensive experiments conducted on three public datasets, namely the BJL, LMHLD, and Landslide4Sense datasets, demonstrate that the proposed SlideMS-Net outperforms prevailing segmentation models across all evaluation metrics. The overall segmentation predictions achieve high consistency with the actual geomorphic boundaries of landslides, effectively reducing commission and omission errors while yielding more precise delineation along irregular landslide edges. Balancing segmentation accuracy and computational efficiency, SlideMS-Net exhibits prominent lightweight advantages compared with large-scale back-bone-based segmentation models. Moreover, the proposed network possesses favorable generalization capability across diverse geographic conditions and various landslide categories, paving a novel technical solution for intelligent remote sensing monitoring of landslide hazards.
No takes yet. Share an insight, caveat, or question.
Ren et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: