Learning Dual Geometric Low-Rank Structure for Semisupervised Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Graph-based semisupervised methods for hyperspectral image classification often rely on the cluster assumption.
• This assumption is challenged by mixed pixels, where recorded spectra combine multiple materials, leading to classification errors.

Purpose of the Study:

• To develop a novel semisupervised classifier for hyperspectral images that addresses the challenge of mixed pixels.
• To improve classification accuracy by exploring both global spectral and local spatial geometric structures.

Main Methods:

• A geometric low-rank representation (GLapLRR) based graph is proposed to assess spectral-spatial affinity of mixed pixels.
• This method reveals the global low-rank and local spatial structure of hyperspectral images.
• The constructed graph exhibits spatial-spectral geometry description, robustness, and low sparsity.

Main Results:

• The proposed geometric low-rank Laplacian regularized semisupervised classifier demonstrates superior performance on three real hyperspectral datasets.
• The method achieves more accurate classification of mixed pixels compared to existing approaches.
• Effectiveness is particularly pronounced when only a small number of labeled instances are available.

Conclusions:

• The developed GLapLRR-based graph effectively captures the spatial-spectral geometry of hyperspectral data, enhancing mixed pixel classification.
• The proposed method offers a robust and accurate solution for semisupervised hyperspectral image classification, outperforming traditional techniques.