SGFNet: Redundancy-Reduced Spectral-Spatial Fusion Network for Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Signal Processing

Background:

• Hyperspectral image classification (HSIC) faces challenges due to high dimensionality, spectral redundancy, and spatial noise.
• Existing deep learning models often struggle with feature redundancy and insufficient spectral-spatial coupling.
• Accurate HSIC requires effective reduction of uncertainty and preservation of informative spectral-spatial interactions.

Purpose of the Study:

• To propose SGFNet, a novel spectral-guided fusion network for HSIC.
• To address feature redundancy and uncertainty from an information-theoretic perspective.
• To improve the efficiency and accuracy of hyperspectral image classification models.

Main Methods:

• Developed a Spectral-Aware Filtering Module (SAFM) to suppress noise and encode spectra.
• Introduced a Spectral-Spatial Adaptive Fusion (SSAF) module for enhanced feature interaction.
• Designed a Spectral Guidance Gated CNN (SGGC) for efficient spatial representation extraction.

Main Results:

• SGFNet demonstrated superior performance across multiple metrics on four benchmark datasets.
• The proposed network consistently outperformed eight state-of-the-art HSIC models.
• Extensive experiments validated the effectiveness of SGFNet's spectral-spatial fusion approach.

Conclusions:

• SGFNet offers an efficient and effective solution for HSIC by balancing redundancy reduction and information preservation.
• The information-theoretic design of SGFNet enhances feature representation for hyperspectral data.
• The proposed modules (SAFM, SSAF, SGGC) contribute to improved HSIC accuracy and model efficiency.