Locality Preserving Projection Based on Endmember Extraction for Hyperspectral Image Dimensionality Reduction and...

Area of Science:

• Remote Sensing
• Computer Vision
• Signal Processing

Background:

• Spectral variability in hyperspectral imagery affects dimensionality reduction precision.
• Existing methods like Locality Preserving Projection (LPP) can be sensitive to these variations.
• Accurate feature extraction is crucial for subsequent analysis, such as target detection.

Purpose of the Study:

• To propose an improved dimensionality reduction method, EE-LPP, for hyperspectral imagery.
• To mitigate the impact of spectral variability on LPP's weighted matrix calculation.
• To enhance target detection accuracy in hyperspectral data.

Main Methods:

• Vertex Component Analysis (VCA) for endmember spectra extraction.
• Spectral Angle Distance (SAD) to determine pixel spectral similarity.
• Constructing a weighted matrix based on spectral similarity for neighboring pixels.
• Applying LPP principles for dimensionality reduction using the constructed weighted matrix.

Main Results:

• The EE-LPP method effectively extracts endmember spectra and constructs a robust weighted matrix.
• Low-dimensional features derived from EE-LPP preserve essential characteristics of the original hyperspectral data.
• Experimental results show improved target detection accuracy compared to traditional methods.

Conclusions:

• EE-LPP offers a significant improvement over standard LPP for hyperspectral image analysis.
• The method successfully addresses spectral variability challenges in dimensionality reduction.
• EE-LPP provides a valuable tool for enhancing hyperspectral target detection applications.