Dimensionality Reduction of Hyperspectral Imagery Based on Spatial-Spectral Manifold Learning

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Graph embedding (GE) methods are used for hyperspectral imagery (HSI) dimensionality reduction.
• Challenges in GE include selecting appropriate neighbors and utilizing spatial information in HSI.
• Existing methods struggle to effectively integrate spatial and spectral information for HSI classification.

Purpose of the Study:

• To propose an unsupervised dimensionality reduction algorithm for HSI classification.
• To develop a novel method that enhances feature extraction by optimizing graph construction.
• To improve the classification performance of hyperspectral imagery through spatial-spectral information fusion.

Main Methods:

• Preprocessing HSI data using a weighted mean filter (WMF) to reduce noise.
• Developing a spatial-spectral combined distance (SSCD) to select effective spatial-spectral neighbors.
• Adjusting reconstruction weights based on spatial relationships to enhance manifold reconstruction.

Main Results:

• The proposed spatial-spectral manifold reconstruction preserving embedding (SSMRPE) method extracts discriminant features.
• SSMRPE demonstrates improved classification performance on PaviaU and Salinas hyperspectral datasets.
• Experimental results show SSMRPE outperforms existing state-of-the-art methods in HSI classification.

Conclusions:

• SSMRPE offers an effective approach for unsupervised dimensionality reduction in HSI.
• The method successfully fuses spatial and spectral information for enhanced feature representation.
• SSMRPE significantly improves the accuracy of hyperspectral image classification.