Hyperspectral Super-Resolution of Locally Low Rank Images From Complementary Multisource Data

Area of Science:

• Remote Sensing
• Image Processing
• Geospatial Analysis

Background:

• Hyperspectral images (HSIs) are often low rank, enabling super-resolution through fusion with multispectral images.
• Current fusion methods struggle with high-dimensional spectral data, leading to ill-posed regression problems.
• Real-world HSIs exhibit local low-rank properties, offering a potential solution.

Purpose of the Study:

• To develop a robust hyperspectral super-resolution method that overcomes the limitations of existing techniques when dealing with high-dimensional spectral data.
• To exploit the locally low-rank nature of hyperspectral images for improved data fusion.
• To enhance the spatial resolution of hyperspectral images through effective fusion with multispectral data.

Main Methods:

• Proposed a local approach by partitioning hyperspectral images into spatial patches.
• Applied data fusion independently within each patch to maintain a low-dimensional subspace.
• Utilized local dictionary learning with endmember induction algorithms for super-resolution.
• Explored sliding windows and binary partition trees for defining local regions.

Main Results:

• Demonstrated improved hyperspectral super-resolution performance, particularly for data with large subspace dimensionality.
• Successfully addressed the ill-posed regression problem by processing data in locally low-rank patches.
• Validated the effectiveness of the proposed local fusion approaches using synthetic and semi-real datasets.

Conclusions:

• The proposed local approach effectively enhances hyperspectral super-resolution by leveraging local low-rank properties.
• Partitioning images into patches and processing them locally resolves issues associated with high spectral dimensionality.
• The method offers a significant advancement in fusing low-resolution hyperspectral data with high-resolution multispectral imagery.