Composite Kernel of Mutual Learning on Mid-Level Features for Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Machine Learning
• Computer Vision

Background:

• Improving machine learning algorithm performance through model averaging is crucial for generalization.
• Knowledge transfer of generalization information between models is essential for effective performance optimization.
• Hyperspectral classification demands advanced techniques for accurate data analysis.

Purpose of the Study:

• To propose a multiple kernel mutual learning (MKL) method for hyperspectral classification.
• To leverage transfer learning for combined mid-level features to enhance classification accuracy.
• To improve the generalization capability of machine learning models in hyperspectral data analysis.

Main Methods:

• Computed three-layer homogenous superpixels on PCA-transformed hyperspectral images.
• Extracted three mid-level features: sparse reconstructed feature, combined mean feature, and uniqueness.
• Employed a multiple kernel mutual learning framework with transfer learning to combine features and minimize divergence.
• Constructed a combined kernel for optimized sample distance measurement and utilized SVM for classification.

Main Results:

• The proposed method achieved significantly better performance compared to state-of-the-art algorithms.
• Demonstrated the effectiveness of transfer learning in combining mid-level features for hyperspectral classification.
• Validated the approach on real-world hyperspectral datasets, showing superior classification accuracy.

Conclusions:

• The developed multiple kernel mutual learning method offers a robust solution for hyperspectral classification.
• Transfer learning of combined mid-level features is a promising direction for improving model generalization.
• The proposed technique provides a substantial advancement over existing MKL and deep learning methods in hyperspectral image analysis.