Joint Learning of Correlation-Constrained Fuzzy Clustering and Discriminative Non-Negative Representation for...

Area of Science:

• Remote Sensing
• Computer Vision
• Data Science

Background:

• Hyperspectral band selection is crucial for managing high-dimensional data in hyperspectral imaging (HSI).
• Existing clustering methods struggle with the curse of dimensionality inherent in HSIs.
• There's a need for advanced methods to select informative and representative bands.

Purpose of the Study:

• To develop a novel hyperspectral band selection method (CFNR) that overcomes the limitations of existing approaches.
• To integrate feature learning and clustering for improved band selection performance.
• To enhance the reliability of hyperspectral image classification through better band selection.

Main Methods:

• A joint learning framework combining graph regularized non-negative matrix factorization (GNMF) and constrained fuzzy C-means (FCM).
• Clustering is performed on learned feature representations, not original high-dimensional data.
• A correlation constraint is introduced to ensure similarity between neighboring band clustering results.

Main Results:

• CFNR successfully learns discriminative non-negative representations for bands.
• The method effectively utilizes the intrinsic manifold structure of HSIs.
• Experimental results show CFNR outperforms state-of-the-art methods on five real hyperspectral datasets.

Conclusions:

• CFNR provides a more informative and representative band subset compared to existing methods.
• The proposed method significantly improves the reliability of hyperspectral image classification.
• CFNR offers a robust solution for effective hyperspectral band selection.