Multiscale Feature-Learning with a Unified Model for Hyperspectral Image Classification.

Area of Science:

• Computer Science
• Remote Sensing
• Artificial Intelligence

Background:

• Hyperspectral image classification demands sophisticated feature extraction for high accuracy.
• Existing methods often struggle with capturing both global and local spectral-spatial features effectively.
• Multiscale feature learning is crucial for interpreting complex hyperspectral data.

Purpose of the Study:

• To propose an advanced architectural paradigm for hyperspectral image classification.
• To develop a unified model that synergistically combines Swin Transformer, Convolutional Neural Network (CNN), and encoder-decoder branches.
• To enhance multiscale feature learning for improved classification accuracy.

Main Methods:

• A unified model architecture integrating three specialized branches: Swin Transformer for long-range dependencies, CNN for localized features, and encoder-decoder for comprehensive analysis.
• Multiscale feature learning leveraging the distinct capabilities of each branch to capture spectral and spatial intricacies.
• Experimental evaluation on publicly available hyperspectral datasets (Xuzhou, Salinas, LK) and comparison with state-of-the-art methods.

Main Results:

• The proposed unified model achieved superior classification performance compared to existing state-of-the-art methods.
• Overall accuracies of 96.87% on the Xuzhou dataset, 98.48% on the Salinas dataset, and 98.62% on the LK dataset were obtained.
• Demonstrated effective assimilation of multiscale spectral and spatial information through synergistic branch integration.

Conclusions:

• The proposed unified model offers a powerful approach for hyperspectral image classification.
• The synergistic integration of Swin Transformer, CNN, and encoder-decoder effectively facilitates multiscale feature learning.
• The model's high accuracy on benchmark datasets validates its efficacy and potential for real-world applications.