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

Area of Science:

• Remote Sensing
• Glaciology
• Computer Vision

Background:

• Glacier segmentation is crucial for monitoring climate change impacts.
• Complex optical remote-sensing scenes present significant challenges due to interference like bright backgrounds, shadows, clouds, and intricate terrain.
• Existing methods often struggle with accuracy and efficiency in these demanding conditions.

Purpose of the Study:

• To develop a lightweight and robust glacier segmentation network, HFBFormer, capable of handling complex remote-sensing scenes.
• To enhance the network's ability to model cross-scale semantic interactions and long-range dependencies.
• To improve the refinement of glacier contours and narrow glacier tongues.

Main Methods:

• HFBFormer utilizes a hybrid self/cross-attention mechanism (HSC-AM) at multiple scales within an MiT-B2 encoder.
• A feature enhancement module (FEM) combines pyramid pooling with variance-guided channel recalibration for improved feature representation.
• A lightweight boundary sharpen module (BSM) is employed for precise contour refinement.

Main Results:

• HFBFormer achieved 95.11% accuracy and 89.36% mIoU on the Animaqing glacier dataset.
• The network demonstrates strong robustness and efficiency with only 16.37 million parameters.
• Effective segmentation was achieved despite bright-background interference, shadow/cloud contamination, complex terrain, and blurred boundaries.

Conclusions:

• HFBFormer offers a highly accurate and efficient solution for glacier segmentation in challenging optical remote-sensing environments.
• The proposed network architecture effectively addresses common difficulties in glacier extraction.
• The results highlight HFBFormer's potential for reliable glacier monitoring and analysis.