Learning Hierarchical Spectral-Spatial Features for Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral image (HSI) classification requires robust feature extraction.
• Existing methods often struggle to simultaneously leverage both spectral and spatial information effectively.
• Deep learning approaches show promise but require careful architecture design for HSIs.

Purpose of the Study:

• To propose a novel deep hierarchical model for hyperspectral image classification.
• To develop a spectral-spatial feature learning (SSFL) method that robustly extracts HSI features.
• To enhance classification accuracy, particularly in scenarios with limited training samples.

Main Methods:

• A spectral-spatial feature learning (SSFL) unit is designed to combine spectral and spatial feature extraction.
• Multiple SSFL units are stacked to form a deep hierarchical spectral-spatial network (SSN).
• Kernel-based extreme learning machine (KELM) is integrated for pixel classification within the SSN framework.

Main Results:

• SSN effectively exploits both discriminative spectral and spatial information simultaneously.
• Experiments on benchmark HSI datasets demonstrate SSN's superior performance.
• SSN achieves higher classification accuracy (overall, average, and kappa coefficient) compared to state-of-the-art methods, especially with small training datasets.

Conclusions:

• The proposed SSN model offers an effective deep hierarchical architecture for HSI classification.
• SSN's ability to integrate spectral and spatial features leads to improved robustness and accuracy.
• The method shows significant advantages in low-sample-size classification scenarios for hyperspectral imaging.