LSTNet: A Reference-Based Learning Spectral Transformer Network for Spectral Super-Resolution.

Area of Science:

• Earth observation
• Remote sensing
• Image processing

Background:

• Hyperspectral images (HSIs) offer rich spectral data crucial for Earth observation.
• Current imaging limitations restrict large-area HSI acquisition, necessitating methods like spectral super-resolution (SSR).
• Existing SSR methods often lack spectral fidelity due to insufficient constraints.

Purpose of the Study:

• To propose a novel learning spectral transformer network (LSTNet) for high-fidelity hyperspectral image reconstruction.
• To address the limitations of existing SSR methods by incorporating reference-based learning for spectral fidelity.
• To improve the coverage of hyperspectral imagery over large areas.

Main Methods:

• Developed a novel learning spectral transformer network (LSTNet).
• Designed a spectral transformer module (STM) to exploit prior spectral information.
• Incorporated a spectral reconstruction module (SRM) utilizing reference-based learning for spectral structure knowledge transfer.

Main Results:

• The LSTNet effectively reconstructs hyperspectral images with improved spectral fidelity.
• Experimental results validate the method's ability to produce high-fidelity reconstructed spectra.
• The proposed approach leverages multispectral image (MSI) swath width for enhanced HSI coverage.

Conclusions:

• The LSTNet offers a significant advancement in spectral super-resolution.
• The reference-based learning strategy ensures high spectral fidelity in reconstructed HSIs.
• This method enhances the utility of hyperspectral imagery for large-scale Earth observation missions.