A Semi-Supervised Reduced-Space Method for Hyperspectral Imaging Segmentation

Area of Science:

• Remote Sensing
• Image Analysis
• Computer Vision

Background:

• Hyperspectral images (HSI) offer rich spectral information for diverse applications.
• High dimensionality of HSI necessitates advanced, efficient processing algorithms.
• Existing segmentation methods struggle with the complexity of HSI data.

Purpose of the Study:

• To develop a novel semi-supervised multilabel segmentation method for hyperspectral images (HSI).
• To improve the efficiency and accuracy of HSI data analysis.
• To address the computational challenges posed by high-dimensional HSI data.

Main Methods:

• A semi-supervised approach combining linear discriminant analysis (LDA), a spectral similarity index, and a random walk model.
• LDA is used for feature space projection to maximize class separation and reduce dimensionality.
• A random walk model on a weighted graph assigns pixel-wise probabilities based on spectral distances and similarity to labeled regions.

Main Results:

• The proposed method effectively reduces computational burden by retaining informative features.
• Achieved accurate multilabel segmentation of HSI benchmark datasets.
• Demonstrated improved class separation and efficient processing compared to traditional methods.

Conclusions:

• The developed semi-supervised method offers a robust solution for HSI segmentation.
• This approach enhances the utility of HSI in various scientific and practical domains.
• The method provides a computationally efficient way to extract detailed spectral information from HSI.