Nonparametric Coupled Bayesian Dictionary and Classifier Learning for Hyperspectral Classification.

Area of Science:

• Machine Learning
• Remote Sensing
• Signal Processing

Background:

• Hyperspectral classification requires effective feature representation.
• Dictionary learning methods aim to represent spectral data efficiently.
• Existing methods often struggle with automatic dictionary size selection and adaptive feature learning.

Purpose of the Study:

• To develop a principled Bayesian approach for discriminative dictionary learning coupled with a linear classifier for hyperspectral classification.
• To enable automatic inference of dictionary size and adaptive learning of atom-class associations.
• To improve classification accuracy by jointly learning the dictionary and classifier.

Main Methods:

• Utilizing Gaussian Process priors for dictionary smoothness and multivariate Gaussians for classifier parameters.
• Employing Beta-Bernoulli processes for joint inference of dictionary and classifier.
• Implementing Gibbs sampling for posterior distribution inference and deriving analytical expressions.
• Solving simultaneous sparse optimization problems for spectral representation.

Main Results:

• The proposed approach effectively learns a discriminative dictionary and a linear classifier.
• Automatic inference of dictionary size and adaptive learning of atom-class associations were achieved.
• Demonstrated competitive classification performance on benchmark hyperspectral images compared to state-of-the-art methods.

Conclusions:

• The nonparametric Bayesian framework offers an effective solution for discriminative dictionary learning in hyperspectral classification.
• Jointly learning the dictionary and classifier enhances spectral analysis and classification accuracy.
• The method's ability to adaptively learn dictionary-class associations is a key advantage.