LGSTNet: Efficient Local–Global Collaboration for Heterogeneous Remote-Sensing Landslide Segmentation

Purpose: Landslide segmentation from remote sensing imagery is challenged by scale variation, irregular morphology, blurred boundaries, and confusion with spectrally similar backgrounds. This study aims to develop an efficient framework that coordinates heterogeneous input regulation, local structural preservation, and hierarchical semantic modeling. Methods: LGSTNet is organized around two coupled design principles. An input-configuration-oriented local–global encoder combines input recalibration, a shallow Ghost branch, and a compact three-stage Swin Transformer. An accuracy–efficiency-oriented compressed context-fusion strategy enhances the deepest Transformer feature through compressed multi-scale, directional, and global context branches and integrates multi-level representations using a lightweight projection–fusion decoder. Results: Experiments were conducted on the Jiuzhai Valley, HR-GLDD, and Landslide4Sense datasets. On Landslide4Sense, LGSTNet achieved the best F1-score, foreground intersection over union (IoU), mean IoU (mIoU), and Matthews correlation coefficient (MCC) among the seven evaluated models, reaching 64.01%, 47.07%, 72.81%, and 63.39%, respectively, with 3.79 M parameters, 2.39 G FLOPs, and 63.93 frames per second. On HR-GLDD and Jiuzhai Valley, it provided competitive segmentation performance with favorable inference efficiency, although it did not outperform every baseline. A within-dataset input-configuration experiment on Landslide4Sense further showed that the F1-score of LGSTNet increased from 42.87% with RGB input to 64.01% with the full fourteen-channel input; SegFormer exhibited the same overall trend, increasing from 42.50% to 61.56%. Conclusions: LGSTNet provides a favorable accuracy–efficiency trade-off across the three evaluated remote-sensing configurations. The within-dataset input-configuration results provide additional evidence that the complete system can benefit from the information available in the full multi-channel input, while the cross-dataset comparison alone is not interpreted as causal evidence that channel number or modality richness by itself explains the observed performance differences.