Boosting Few-Shot Hyperspectral Image Classification Through Dynamic Fusion and Hierarchical Enhancement.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Few-shot learning (FSL) is crucial for hyperspectral image classification (HSIC) to reduce reliance on extensive labeled data.
• Existing HSIC methods often use fixed-size patches and neglect central pixel information, leading to inefficient feature utilization.
• Limited exploration of feature correlations weakens expressiveness and hinders cross-domain knowledge transfer in current FSL for HSIC.

Purpose of the Study:

• To propose a novel FSL framework for HSIC that addresses limitations in feature extraction and fusion.
• To enhance feature representation and knowledge transfer by incorporating dynamic fusion and hierarchical attention mechanisms.
• To improve the discriminability of features through specialized loss functions.

Main Methods:

• A robust feature extraction module combining small and large patches with a central-pixel-guided dynamic pooling strategy for patch-to-pixel fusion.
• A support-query hierarchical enhancement module utilizing intraclass self-attention and interclass cross-attention.
• Intraclass consistency and interclass orthogonality loss functions to improve feature separability.

Main Results:

• The proposed method achieves substantial improvements in classification accuracy across four benchmark hyperspectral datasets.
• The dynamic fusion strategy enables more comprehensive and robust extraction of ground object information.
• Hierarchical enhancement and specialized losses effectively improve feature representation and discriminability.

Conclusions:

• The novel framework significantly enhances few-shot hyperspectral image classification performance.
• Dynamic fusion and hierarchical enhancement are effective strategies for improving information utilization and knowledge transfer.
• The method offers a promising solution for accurate HSIC with limited labeled samples.