A Hyperspectral Image Classification Framework with Spatial Pixel Pair Features.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) are prevalent in hyperspectral image (HSI) classification, primarily focusing on spectral information.
• Spatial consistency in HSI has been underexplored, limiting classification performance.
• Pixel Pair Features (PPF) offer a novel approach to integrate spatial information into HSI classification.

Purpose of the Study:

• To propose an improved pixel pair feature, Spatial Pixel Pair Feature (SPPF), for enhanced HSI classification.
• To develop a flexible, multi-stream CNN framework for HSI classification leveraging SPPF.
• To evaluate the effectiveness of the proposed SPPF and CNN framework on public HSI datasets.

Main Methods:

• Introduced Spatial Pixel Pair Features (SPPF) by restricting pixel selection to immediate neighbors of the central pixel, enforcing stronger spatial regularization.
• Developed a multi-stream, late-fusion CNN classification framework compatible with various sub-network designs.
• Conducted experiments on three publicly available hyperspectral image datasets to validate the proposed method.

Main Results:

• The proposed SPPF effectively integrates spatial and spectral information, outperforming traditional methods.
• The multi-stream CNN framework demonstrated superior performance compared to competing approaches.
• The method achieved high classification accuracy without requiring extensive network configuration tuning.

Conclusions:

• The proposed SPPF-based HSI classification framework offers a significant advancement in leveraging both spatial and spectral information.
• The flexible multi-stream CNN architecture provides a robust and adaptable solution for HSI classification tasks.
• The findings highlight the importance of spatial context in HSI analysis and classification.