Semisupervised classification of hyperspectral images with low-rank representation kernel.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral image (HSI) classification often faces challenges due to limited labeled samples.
• Unlabeled data is abundant and can potentially improve classification performance.
• Existing kernel methods may struggle with effectively incorporating unlabeled information.

Purpose of the Study:

• To propose a semisupervised deformed kernel function for hyperspectral image classification.
• To leverage both labeled and unlabeled data for improved classification accuracy.
• To address the common issue of limited labeled samples in HSI applications.

Main Methods:

• A novel semisupervised kernel function is developed using low-rank representation and considering local data geometry.
• The kernel combines a standard radial basis function kernel (using labeled data) with a low-rank representation kernel (using all data).
• The method is evaluated with a support vector machine classifier.

Main Results:

• The proposed method effectively utilizes unlabeled information to overcome the limitations of scarce labeled samples.
• The low-rank representation kernel outperforms kernels derived from traditional clustering methods.
• Experimental results on two HSI datasets demonstrate superior classification performance.

Conclusions:

• The proposed semisupervised kernel function offers an effective approach for hyperspectral image classification.
• Incorporating unlabeled data via low-rank representation significantly enhances classification accuracy.
• This method provides a valuable tool for HSI analysis where labeled data is limited.