Trans-CAUNet: a swin transformer-based network for automated mapping mountain glaciers from multisource remote sensing imagery

Accurate and automated extraction of glaciers is pivotal for water resources and ecological systems in the cryospheric. Given the expanding access to remote sensing imagery, there is an urgent requirement to develop efficient methods to identify glaciers in High Mountain Asia (HMA). Proposed here is a novel UNet-based model embedding the Swin Transformer and channel attention mechanism (CAM) parallel modules. This design effectively captures global-local and contextual information from multisource imagery, further enhancing spectral features. Trained on a dataset from three regions, it achieved impressive results relative to other popular deep learning models and distinguished glaciers from similar terrain elements, with an F1 score of 0.93 and a mean intersection over union (MIoU) of 0.87. Besides, it demonstrated robust transferability when applied to more challenging glacier environments in Southeastern Tibet. The proposed model presents a viable solution for the rapid and accurate monitoring of glaciers dynamics.