Unsupervised Hyperspectral Band Selection via Multimodal Evolutionary Algorithm and Subspace Decomposition.

Area of Science:

• Remote Sensing
• Data Science
• Machine Learning

Background:

• Unsupervised band selection is crucial for hyperspectral image (HSI) dimension reduction.
• Existing methods often overlook the diversity of band subsets and the ordered nature of HSI data.
• This can lead to redundant band selection and suboptimal performance.

Purpose of the Study:

• To propose an unsupervised band selection method that addresses the limitations of existing approaches.
• To explore the diversity of band subsets by seeking multiple solutions.
• To leverage the ordered property of HSI to avoid redundant band selection.

Main Methods:

• Utilized a multimodal evolutionary algorithm for spectral subspace decomposition to find diverse band subsets.
• Incorporated subspace decomposition to focus on the ordered property of HSI, increasing differences between neighbor band subspaces.
• Adopted an integrated utilization strategy for the diverse band subsets to enhance prediction performance.

Main Results:

• Demonstrated the effectiveness of the proposed method on hyperspectral remote sensing datasets.
• Showcased superior prediction accuracy compared to state-of-the-art methods.
• Validated the method's ability to find diverse and representative band subsets.

Conclusions:

• The proposed unsupervised band selection method effectively addresses the diversity and ordered property challenges in HSI.
• The multimodal evolutionary algorithm and subspace decomposition approach yield superior prediction accuracy.
• This method offers a significant advancement for hyperspectral data analysis and dimension reduction.