A Spatio-Spectral Fusion Method for Hyperspectral Images Using Residual Hyper-Dense Network

Area of Science:

• Remote Sensing
• Computer Vision
• Image Processing

Background:

• Hyperspectral (HS) and panchromatic (PAN) image fusion is crucial for enhancing spatial resolution in remote sensing.
• DenseNets have shown promise in image super-resolution tasks due to efficient gradient flow.

Purpose of the Study:

• To propose a novel Residual Hyper-Dense Network (RHDN) for spatio-spectral fusion of PAN and HS images.
• To improve the spatial resolution of HS images by effectively combining information from PAN images.

Main Methods:

• A two-branch network architecture to separately process features within and outside the visible spectrum.
• A two-stream strategy within each branch for individual PAN and HS image feature extraction.
• Convolutional Neural Network (CNN) with cascade residual hyper-dense blocks (RHDBs) for complex feature learning and residual learning for efficiency.

Main Results:

• The RHDN method effectively fuses PAN and HS images, leading to significant improvements in spatial resolution.
• Benchmark evaluations demonstrate superior performance compared to existing state-of-the-art fusion methods.
• The network architecture facilitates direct connections across layers for enhanced feature combination.

Conclusions:

• The proposed RHDN is an effective deep learning approach for spatio-spectral fusion of PAN and HS images.
• RHDN offers significant advancements in improving image spatial resolution for remote sensing applications.
• The method demonstrates the potential of extended DenseNet architectures for complex image fusion tasks.