Semantic-Aware Co-Parallel Network for Cross-Scene Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Cross-scene classification of hyperspectral images is challenging due to varying data distributions and limited prior knowledge.
• Previous methods had restricted use of cross-modal knowledge, hindering performance.
• Recent large-scale language-vision models show promise for cross-modal assisted learning.

Purpose of the Study:

• To propose a novel network, Semantic-aware Collaborative Parallel Network (SCPNet), for cross-scene hyperspectral image classification.
• To mitigate data distribution discrepancies by integrating linguistic modalities for learning cross-domain invariant representations.
• To enhance feature clustering and separation using supervised contrastive learning within an optimized semantic space.

Main Methods:

• SCPNet employs a parallel architecture with spatial-spectral and multiscale feature extraction modules.
• Features are mapped to a semantic space for improved supervised contrastive learning.
• Linguistic modalities are utilized to mine cross-domain invariant representations, bridging visual and linguistic gaps.

Main Results:

• SCPNet significantly outperforms existing methods on three public hyperspectral image datasets.
• The proposed method demonstrates effectiveness in cross-scene classification tasks.
• Integration of linguistic modalities enhances the learning of domain-invariant features.

Conclusions:

• SCPNet offers a robust solution for cross-scene hyperspectral image classification.
• The semantic-aware approach effectively leverages cross-modal information.
• This work advances the field of hyperspectral image analysis through innovative deep learning techniques.