Equivariant High-Resolution Hyperspectral Imaging via Mosaiced and PAN Image Fusion

Area of Science:

• Remote Sensing
• Image Processing
• Optical Engineering

Background:

• Existing mosaic-based hyperspectral imaging systems lack high-resolution capabilities.
• Fusing low-resolution mosaiced images with high-resolution panchromatic images is a viable solution for HR HSI reconstruction.

Purpose of the Study:

• To propose a dual-sensor hyperspectral imaging system combining mosaiced and panchromatic sensors.
• To develop an unsupervised equivariant imaging (EI)-based training framework for HR HSI reconstruction.

Main Methods:

• A dual-sensor system with a 4x4 spectral filter array (SFA) mosaiced sensor and a co-aligned PAN sensor.
• An unsupervised EI training framework with a learnable degradation function, including mosaicing and downsampling.
• Modeling PAN image as a linear projection of HR HSI with learnable spectral response function (SRF) parameters.

Main Results:

• Successful reconstruction of HR HSI with high spatial consistency and spectral fidelity.
• Demonstrated effectiveness on a real-world dataset of 60 paired mosaiced/PAN images.
• The method maintains computational efficiency.

Conclusions:

• The proposed dual-sensor system and EI framework effectively overcome limitations of existing HSI systems.
• Unsupervised learning with learnable degradation functions enables accurate HR HSI reconstruction without ground truth.
• The system provides a practical solution for applications requiring high-resolution hyperspectral data.