FRFSL: Feature Reconstruction-Based Cross-Domain Few-Shot Learning for Coastal Wetland Hyperspectral Image...

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral image classification (HSIC) is crucial for coastal wetland vegetation identification.
• Environmental complexity and land cover similarity hinder large-scale labeling.
• Cross-domain few-shot learning (CDFSL) offers a solution for limited labeled data in HSIC.

Purpose of the Study:

• To address limitations in existing CDFSL HSIC methods, including prototype deviation and covariate shifts.
• To propose a novel Feature Reconstruction-based CDFSL (FRFSL) algorithm for improved HSIC.
• To develop lightweight and interpretable domain alignment for enhanced classification.

Main Methods:

• A Prototype Calibration Module (PCM) uses Bayesian inference-enhanced Gaussian Mixture Model for reliable query feature selection and prototype reconstruction.
• A Distance Metric Module (DMM) incorporates ridge regression for prototype reconstruction, mitigating covariate shifts.
• Dynamic graph reconstruction and optimal transport theory transform domain alignment into a graph matching problem.

Main Results:

• The proposed FRFSL algorithm effectively addresses prototype deviation and covariate shifts.
• Experiments demonstrate FRFSL's superior performance compared to eleven state-of-the-art algorithms on multiple datasets.
• The developed shared transport matrix algorithm achieves lightweight and interpretable domain alignment.

Conclusions:

• FRFSL offers a significant advancement in hyperspectral image classification for coastal wetlands using CDFSL.
• The method provides a robust solution for scenarios with limited labeled data and domain shifts.
• The approach enhances the practical applicability of HSIC in ecological monitoring and management.