A Lightweight State Space Model With Multiscale Morphology and Low-Rank Head for Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• State space models (SSMs) are effective for hyperspectral image (HSI) classification.
• Existing methods struggle with single-scale feature extraction and training instability on complex HSI data.

Purpose of the Study:

• To propose a novel lightweight multiscale morphology-enhanced low-rank head residual state space network (MMLH-RSSN).
• To address limitations in spatial feature modeling and training stability in current HSI classification approaches.

Main Methods:

• Developed a multiscale morphological module for hierarchical spatial feature extraction.
• Introduced an enhanced Residual SSM with residual connections and layer normalization for improved stability.
• Utilized a parameter-efficient low-rank decomposition head for a lightweight design.

Main Results:

• Achieved state-of-the-art performance on four benchmark HSI datasets.
• Obtained high overall accuracies (e.g., 98.51% on Pavia University, 99.69% on Botswana).
• Demonstrated efficiency with only 0.063 million parameters.

Conclusions:

• The synergistic combination of multiscale priors and a stabilized SSM backbone provides a highly accurate and efficient HSI classification solution.
• MMLH-RSSN is particularly suitable for resource-constrained scenarios.
• The proposed network effectively handles complex HSI data and varying spatial geometries.