Zero-Shot Hyperspectral Sharpening.

Area of Science:

• Remote Sensing
• Image Processing
• Computer Vision

Background:

• Hyperspectral image (HSI) sharpening via fusion with multispectral images (MSIs) is crucial for enhanced spatial detail.
• Deep convolutional neural networks (CNNs) show promise but struggle with limited training data and poor generalization.

Purpose of the Study:

• To develop a zero-shot learning (ZSL) method for HSI sharpening that overcomes data limitations and improves generalization.
• To enhance the accuracy and efficiency of HSI sharpening through novel fusion techniques.

Main Methods:

• Proposed a quantitative method to estimate sensor spectral and spatial responses.
• Implemented a ZSL approach using spatially subsampled HSI and MSI for training CNNs.
• Applied dimensionality reduction to HSIs and incorporated an imaging model-based loss function for CNNs.

Main Results:

• The ZSL method demonstrated high efficiency and accuracy in HSI sharpening.
• The approach effectively exploits inherent information from both HSI and MSI.
• The trained CNN exhibited strong generalization capabilities on test data.

Conclusions:

• The proposed ZSL method significantly improves HSI sharpening accuracy and efficiency.
• The technique addresses the challenges of limited training data and generalization in deep learning-based fusion.
• This method offers a robust solution for enhancing spatial resolution in hyperspectral imaging.