Few-Shot Learning With Class-Covariance Metric for Hyperspectral Image Classification

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Few-shot learning (FSL) and metric-based methods have advanced hyperspectral image classification (HSIC).
• Existing methods face challenges in achieving high performance with limited labeled data.

Purpose of the Study:

• To propose a novel FSL framework (CMFSL) for HSIC to improve performance with few labeled samples.
• To enhance the learning of global class representations and avoid overfitting in novel classes.

Main Methods:

• Developed a class-covariance metric (CMFSL) framework for HSIC.
• Employed a synthesis strategy for novel classes and Mahalanobis distance for label determination.
• Designed a lightweight cross-scale convolutional network (LXConvNet) for spectral-spatial feature extraction.
• Integrated a spectral-prior-based refinement module (SPRM) to emphasize informative bands and mitigate domain shift.

Main Results:

• The CMFSL framework demonstrated superior performance compared to state-of-the-art methods on four benchmark datasets.
• The proposed LXConvNet effectively exploited spectral-spatial information with low computational complexity.
• The SPRM module successfully refined feature extraction and addressed domain shift issues.

Conclusions:

• The CMFSL framework offers a significant advancement in few-shot hyperspectral image classification.
• The combination of CMFSL, LXConvNet, and SPRM provides a robust solution for accurate HSIC with minimal labeled data.