Deep Network With Irregular Convolutional Kernels and Self-Expressive Property for Classification of Hyperspectral...

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral image (HSI) classification is crucial for analyzing Earth's surface.
• Traditional deep learning methods often lack adaptability and require extensive training for HSI feature extraction.

Purpose of the Study:

• To introduce a novel deep network, DIKS, for adaptive and efficient HSI classification.
• To leverage irregular convolutional kernels and self-expression for enhanced feature representation.

Main Methods:

• Utilized principal component analysis (PCA) and superpixel segmentation to generate irregular convolutional kernels.
• Employed a deep network architecture incorporating multiple convolutional layers and stacked shallow/deep features.
• Applied self-expression theory for clustering and generating discriminative final features.

Main Results:

• The DIKS method demonstrated self-adaptability to HSI data due to its irregular kernels.
• Achieved superior classification performance compared to state-of-the-art algorithms in extensive experiments.
• The approach requires no training operations for feature extraction, offering computational efficiency.

Conclusions:

• The proposed DIKS method effectively extracts multiscale and discriminative features for HSI classification.
• DIKS offers a novel, adaptive, and training-free approach for hyperspectral image analysis.
• This work advances the field of HSI classification through innovative deep learning techniques.