CViT Weakly Supervised Network Fusing Dual-Branch Local-Global Features for Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral image (HSI) classification is vital for analyzing spectral data.
• Noisy labels in HSI datasets degrade the performance of deep learning models.
• Existing deep learning methods often sacrifice feature representation for noise resistance.

Purpose of the Study:

• To develop a robust and accurate HSI classification network resistant to noisy labels.
• To enhance feature learning capabilities while maintaining computational efficiency.
• To improve the generalization ability of HSI classification models.

Main Methods:

• Proposed a Convolutional Vision Transformer (CViT) Weakly Supervised Network (CWSN).
• Employed a lightweight 1D-2D two-branch network for spatial-spectral feature extraction.
• Utilized a CNN-Vision Transformer cascade for fusing local and global features.

Main Results:

• The CWSN demonstrated strong anti-noise capabilities on benchmark HSI datasets.
• Achieved superior classification accuracy compared to existing methods.
• Showcased robustness and versatility with both clean and noisy training sets.

Conclusions:

• The CWSN effectively addresses the challenge of noisy labels in HSI classification.
• The proposed network offers a robust and versatile solution for accurate HSI analysis.
• This approach balances feature representation and noise resistance for improved performance.