Deep Posterior Distribution-Based Embedding for Hyperspectral Image Super-Resolution.

Area of Science:

• Computer Vision
• Deep Learning
• Remote Sensing

Background:

• Hyperspectral (HS) image spatial super-resolution is crucial for detailed analysis.
• Existing deep learning methods often use empirically designed modules for feature embedding.
• Efficiently embedding high-dimensional spatial-spectral information remains a challenge.

Purpose of the Study:

• To develop a novel deep learning framework for hyperspectral image spatial super-resolution.
• To propose a new method for embedding spatial-spectral information in HS images.
• To improve the performance and interpretability of HS super-resolution.

Main Methods:

• Formulating hyperspectral embedding as posterior distribution approximation.
• Developing layer-wise spatial-spectral feature extraction and network-level feature aggregation.
• Integrating the embedding scheme into a source-consistent, physically-interpretable super-resolution framework (PDE-Net).

Main Results:

• PDE-Net iteratively refines high-resolution (HR) HS images from low-resolution (LR) inputs.
• Achieved superior performance over state-of-the-art methods on three benchmark datasets.
• Demonstrated the capability of providing epistemic uncertainty in network outputs.

Conclusions:

• The proposed probability-inspired HS embedding scheme is effective for super-resolution.
• PDE-Net offers a physically-interpretable and high-performing solution for HS image super-resolution.
• The probabilistic nature of PDE-Net provides valuable uncertainty information for downstream applications.