Enhanced Deep Blind Hyperspectral Image Fusion.

Area of Science:

• Remote Sensing
• Computer Vision
• Image Processing

Background:

• Hyperspectral image fusion (HIF) aims to reconstruct high spatial resolution hyperspectral images (HR-HSI) from low spatial resolution hyperspectral images (LR-HSI) and high spatial resolution multispectral images (HR-MSI).
• Existing HIF methods often rely on known observation models, which are unrealistic in many practical applications, leading to the blind HIF problem.

Purpose of the Study:

• To develop a robust deep learning-based method for blind hyperspectral image fusion (HIF).
• To address the limitations of existing methods by optimizing the observation model and fusion process simultaneously.
• To enhance spatial and spectral accuracy in reconstructed HR-HSI.

Main Methods:

• A novel deep learning approach that iteratively optimizes the observation model and fusion process for bidirectional data consistency.
• Incorporation of an invertible deep neural network component using modified spectral normalization to prevent information loss.
• Introduction of a Content-Aware ReAssembly of FEatures module and an SE-ResBlock model to reduce spatial distortion and feature redundancy, boosting performance and model compactness.

Main Results:

• The proposed method effectively addresses the blind HIF problem by enforcing bidirectional data consistency.
• The invertible network component successfully mitigates information loss inherent in general deep neural networks.
• Experimental results show superior performance compared to existing methods in both nonblind and semiblind HIF scenarios.

Conclusions:

• The developed deep learning method significantly improves the accuracy of hyperspectral image fusion, particularly in challenging blind scenarios.
• The innovative architectural enhancements lead to more precise spatial and spectral information reconstruction.
• This work offers a more realistic and effective solution for hyperspectral image fusion applications.