Hyperspectral Tensor Completion Using Low-Rank Modeling and Convex Functional Analysis.

Area of Science:

• Remote Sensing
• Machine Learning
• Functional Analysis

Background:

• Hyperspectral imaging (HSI) is vital for material identification due to unique spectral signatures.
• Remotely acquired HSIs often suffer from low data purity and incomplete observations.
• Hyperspectral tensor completion (HTC) is crucial for processing corrupted 3-D HSI data.

Purpose of the Study:

• To develop an efficient HTC method using John ellipsoid (JE) topology.
• To address the challenge of JE computation requiring complete HSI data, which is unavailable in HTC settings.
• To improve the performance of hyperspectral data analysis and subsequent applications.

Main Methods:

• Decoupling the HTC problem into computationally efficient convex subproblems.
• Adopting John ellipsoid (JE) topology for hyperspectral analysis.
• Developing a novel algorithm for hyperspectral tensor completion.

Main Results:

• The proposed method achieves state-of-the-art performance in hyperspectral tensor completion.
• The algorithm effectively resolves the dilemma of JE computation with incomplete HSI data.
• Demonstrated improvement in land cover classification accuracy using the completed HSI tensor.

Conclusions:

• The novel HTC method offers a computationally efficient and effective solution for corrupted hyperspectral data.
• The integration of JE topology enhances hyperspectral analysis capabilities.
• The approach significantly improves the accuracy of downstream applications like land cover classification.