Transductive Few-shot Learning With Enhanced Spectral-Spatial Embedding for Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Few-shot learning (FSL) is crucial for hyperspectral image (HSI) classification due to high labeling costs.
• Effective feature embedding is vital but challenging for HSIs with rich spectral-spatial information.
• Transductive FSL models often outperform inductive ones by utilizing query set statistics.

Purpose of the Study:

• To develop a transductive FSL framework (TEFSL) for enhanced spectral-spatial embedding in HSI classification.
• To leverage limited prior information effectively for improved classification performance.
• To address the challenge of feature embedding in HSIs within a few-shot learning context.

Main Methods:

• An attentive feature embedding network (AFEN) with a channel calibration module (CCM) was designed for informative feature extraction.
• A meta-feature interaction module (MFIM) was employed for adaptive co-attention between support and query features.
• An iterative graph-based prototype refinement scheme (iGPRS) was proposed for transductive test-time adaptation.

Main Results:

• The proposed TEFSL framework demonstrated superior performance on four standard HSI benchmarks.
• The model achieved high accuracy with limited labeled samples (1-5 per class).
• Experimental results validate the effectiveness of the enhanced spectral-spatial embedding and transductive adaptation.

Conclusions:

• The TEFSL framework effectively exploits spectral-spatial information for few-shot HSI classification.
• The proposed methods, including AFEN, MFIM, and iGPRS, significantly improve classification accuracy.
• The study highlights the potential of transductive FSL for reducing data acquisition burdens in HSI analysis.