Dimension Reduction Aided Hyperspectral Image Classification with a Small-sized Training Dataset: Experimental...

Area of Science:

• Remote Sensing
• Machine Learning
• Computer Vision

Background:

• Hyperspectral images (HSI) offer rich spectral information but generate large, redundant datasets.
• This data volume increases computational complexity and demands extensive labeled training data for classification.
• Acquiring sufficient training samples for HSI analysis is often difficult and costly.

Purpose of the Study:

• To address challenges of large data volume and limited training samples in HSI classification.
• To explore classical dimension reduction algorithms for enhancing HSI classification efficiency and accuracy.
• To evaluate feature selection and extraction methods in conjunction with Support Vector Machine (SVM) classification.

Main Methods:

• Applied feature selection techniques like mutual information and minimal redundancy maximal relevance.
• Employed feature extraction methods including Principal Component Analysis (PCA) and Kernel PCA.
• Augmented a baseline Support Vector Machine (SVM) classification model with these dimension reduction techniques.

Main Results:

• Principal Component Analysis (PCA) demonstrated superior performance in reducing spectral bands and feature dimensionality.
• The proposed dimension reduction methods significantly decreased computational complexity.
• Improved classification accuracy was observed with PCA on small training datasets compared to standard SVM.
• Similar classification performance to standard SVM was achieved on large datasets with substantially reduced computation time.

Conclusions:

• Dimension reduction techniques, particularly PCA, are effective for optimizing HSI classification.
• These methods enhance classification performance and efficiency, especially when training data is limited.
• The approach is suitable for real-time HSI applications and scenarios with scarce labeled data.