Spectral Response Function-Guided Deep Optimization-Driven Network for Spectral Super-Resolution.

Area of Science:

• Remote Sensing
• Image Processing
• Computer Vision

Background:

• Hyperspectral images (HSIs) are vital for diverse research applications.
• Spectral super-resolution (SSR) aims to generate high-spatial-resolution (HR) HSIs from HR multispectral images.
• Existing SSR methods rely on traditional model-driven algorithms or fully data-driven deep learning approaches.

Purpose of the Study:

• To propose an optimization-driven convolutional neural network (CNN) for spectral super-resolution (SSR).
• To develop a physically interpretable network by unfolding a variational method with a deep spatial-spectral prior.
• To enhance the reconstruction of hyperspectral images (HSIs) by integrating auxiliary spectral response functions (SRFs).

Main Methods:

• An optimization-driven CNN was developed by unfolding a variational method, incorporating a deep spatial-spectral prior.
• Auxiliary spectral response functions (SRFs) were utilized to guide the CNN in grouping spectrally relevant bands.
• Channel attention module (CAM) and a reformulated spectral angle mapper loss function were employed for effective reconstruction.

Main Results:

• The proposed method demonstrated significant spectral enhancement on both natural and remote sensing datasets.
• Experiments confirmed the physically interpretable nature of the developed network.
• Classification results on a remote sensing dataset validated the enhanced information content.

Conclusions:

• The developed optimization-driven CNN effectively performs spectral super-resolution (SSR).
• The integration of SRFs and attention mechanisms improves the reconstruction of hyperspectral images (HSIs).
• The method provides a valid approach for enhancing spectral information in remote sensing applications.