Glacier segmentation algorithm that combines self-cross attention and lightweight context aggregation

We propose HFBFormer, a lightweight glacier segmentation network for complex optical remote-sensing scenes affected by bright-background interference, shadow/cloud contamination, complex terrain, and blurred boundaries. Built on an MiT-B2 encoder, HFBFormer introduces a hybrid self/cross-attention mechanism (HSC-AM) at the 1/16 and 1/32 scales to enhance cross-scale semantic interaction and long-range dependency modeling. A feature enhancement module (FEM) further improves discriminative representation by combining pyramid pooling with variance-guided channel recalibration, while a lightweight boundary sharpen module (BSM) refines glacier contours and narrow tongues. Experiments on the Animaqing glacier dataset show that HFBFormer achieves 95.11% accuracy and 89.36% mIoU with only 16.37 M parameters, demonstrating strong robustness and efficiency for glacier extraction in challenging mountain environments.