SRViT: Self-Supervised Relation-Aware Vision Transformer for Hyperspectral Unmixing.

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Vision Transformers (ViT) offer scalability and strong representation but struggle with pixel-level spatial continuity in hyperspectral image (HSI) unmixing.
• Traditional ViTs partition images into non-overlapping patches, disrupting local structures and hindering fine-grained spatial dependency capture for dense prediction tasks.

Purpose of the Study:

• To develop a novel Self-Supervised Relation-aware Vision Transformer (SRViT) to address the limitations of traditional ViTs in HSI unmixing.
• To improve feature representation by preserving pixel-level spatial continuity and local structural relationships.

Main Methods:

• The proposed SRViT integrates a self-embedded module with encoders, a pixel-level position encoder (PLPE), and a self-supervised contrastive mechanism (SCM).
• The self-embedded module and PLPE maintain HSI local correlations across views for cross-view learning via SCM.
• A decoder utilizing Kronecker-factored approximate curvature (K-FAC) captures the spectral information's local geometric structure.

Main Results:

• SRViT effectively learns endmembers and fractional abundance, crucial components for HSI unmixing.
• Comparative experiments systematically validated the superior performance and competitiveness of SRViT against existing methods.

Conclusions:

• SRViT significantly enhances hyperspectral image unmixing by preserving spatial continuity and improving feature representation.
• The developed model demonstrates a promising advancement in deep learning approaches for dense prediction tasks in remote sensing.