Progressive Spatial Information-Guided Deep Aggregation Convolutional Network for Hyperspectral Spectral...

Area of Science:

• Remote Sensing
• Computer Vision
• Signal Processing

Background:

• Hyperspectral image (HSI) spectral super-resolution (SSR) aims to enhance image resolution using deep learning.
• Current fusion-based methods struggle with effectively utilizing cross-modality information, limiting performance, especially at larger scales.

Purpose of the Study:

• To propose a novel deep aggregation convolutional neural network, SIGnet, for improved HSI spectral super-resolution.
• To enhance the fusion of high-resolution multispectral images (HR-MSI) and low-resolution hyperspectral images (LR-HSI).

Main Methods:

• Developed SIGnet, a network featuring dense residual channel affinity learning (DRCA) blocks and a spatial-guided propagation (SGP) module.
• DRCA blocks utilize a channel affinity propagation (CAP) module to model interdependencies between feature map channels.
• The SGP module refines HSI features progressively using a degradation simulation and deformable adaptive fusion.

Main Results:

• SIGnet demonstrates superior performance compared to existing state-of-the-art fusion-based methods.
• The proposed method achieves better reconstruction quality, particularly at larger upsampling scales.

Conclusions:

• SIGnet effectively addresses the limitations of current fusion-based HSI SSR methods.
• The network's architecture enhances the utilization of cross-modality information for superior spectral super-resolution.