Unsupervised Hyperspectral Band Selection Using Spectral-Spatial Iterative Greedy Algorithm.

Area of Science:

• Remote Sensing
• Computer Vision
• Data Science

Background:

• Hyperspectral band selection (BS) is crucial for reducing data dimensionality in hyperspectral remote sensing images (HSIs).
• Existing searching-based BS methods often fail to fully leverage inherent spatial and spectral information, limiting their effectiveness.
• There is a need for unsupervised BS methods that integrate both spatial and spectral prior information.

Purpose of the Study:

• To propose a novel unsupervised band selection method, the Spectral-Spatial Iterative Greedy Algorithm (SSIGA).
• To address the limitations of existing methods by effectively exploiting spatial and spectral prior information in HSIs.
• To improve the performance of BS for classification applications of HSIs.

Main Methods:

• SSIGA employs K-means clustering with balanced cluster size constraints for spectral information processing.
• A K-nearest neighbor graph is constructed for each cluster to facilitate efficient local search.
• An objective function evaluates band discriminability and redundancy using Fisher score, superpixel segmentation, information entropy, and mutual information.

Main Results:

• SSIGA demonstrates superior performance compared to state-of-the-art methods on three real HSI datasets.
• The algorithm effectively utilizes spatial and spectral information for band selection.
• Experimental results validate the efficacy of the proposed objective function and local search strategy.

Conclusions:

• The proposed SSIGA is an effective unsupervised band selection method for HSIs.
• SSIGA achieves superior performance by integrating spectral and spatial information.
• The method offers a promising approach for dimensionality reduction in hyperspectral image classification.