USwin: a U-shaped Swin transformer for landslide extraction from remote sensing images

Accurate landslide extraction is critical for timely disaster response and emergency management. However, existing methods often struggle to generalize across diverse terrains due to variations in morphological, spectral, and textural features. To address challenges in transferability and boundary precision, we propose USwin, a U-shaped semantic segmentation model featuring a dual-branch decoder architecture. The first branch utilizes convolutional kernels of multiple sizes to capture multiscale spatial patterns, whereas the second incorporates a global attention mechanism to integrate fine-grained details with contextual information. A dynamic feature fusion module is further introduced to effectively combine encoder and decoder features, enhancing the model’s capacity to extract complex landslide signatures and improve generalization across heterogeneous landscapes. We evaluate USwin against nine state-of-the-art semantic segmentation models using four landslide datasets. Trained on the Iburi and Lushan datasets and tested on the Bijie and Jiuzhaigou datasets, USwin achieves a minimum improvement of 2.46% in mean intersection over union on Jiuzhaigou and 0.99% on Bijie. These results demonstrate the robustness and effectiveness of USwin in extracting landslides under varied real-world conditions.