Hyperspectral Image Super-Resolution via Deep Spatiospectral Attention Convolutional Neural Networks

Area of Science:

• Remote Sensing
• Computer Vision
• Image Processing

Background:

• Hyperspectral images (HSIs) offer rich spectral information but often suffer from limited spatial resolution due to imaging constraints.
• Enhancing the spatial resolution of HSIs is crucial for detailed feature analysis and various applications.

Purpose of the Study:

• To propose a novel deep convolutional neural network architecture for fusing low-resolution HSI (LR-HSI) and high-resolution multispectral images (HR-MSI) to generate high-resolution HSI (HR-HSI).
• To preserve both spatial and spectral information during the super-resolution process.

Main Methods:

• A deep convolutional neural network architecture utilizing LR-HSI at the HR-MSI scale for spectral preservation.
• Integration of attention and pixelShuffle modules to extract spatial details and enhance image quality.
• Training the network using a mean squared error loss function.

Main Results:

• The proposed network achieved superior performance, both qualitatively and quantitatively, compared to existing state-of-the-art HSI super-resolution methods.
• Demonstrated excellent network generalization ability and robustness to the number of training samples.
• Showcased a limited computational burden.

Conclusions:

• The developed deep learning approach effectively addresses the spatial resolution limitations of HSIs.
• The fusion of LR-HSI and HR-MSI using the proposed network yields high-quality HR-HSI with preserved spatial and spectral characteristics.
• The method offers practical advantages including efficiency, generalization, and robustness for HSI super-resolution tasks.