Physics-Guided Time-Interactive-Frequency Network for Cross-Domain Few-Shot Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Signal Processing

Background:

• Few-shot learning (FSL) and domain alignment are used in hyperspectral image classification (HSIC) to address data scarcity and distribution issues.
• Existing cross-domain FSL methods overlook crucial frequency priors that enhance category discrimination and knowledge transfer.

Purpose of the Study:

• To propose a novel physics-guided network (PTFNet) for cross-domain few-shot HSIC.
• To simultaneously extract frequency priors and spatial features using a time-interactive-frequency module.
• To enhance the diversity of physical attributes and imitate domain shifts.

Main Methods:

• A lightweight time-interactive-frequency module (TiF-Module) extracts frequency and spatial features.
• A spectral Fourier-based augmentation module (SFA-Module) decouples frequency priors and enhances distribution diversity.
• Physics consistency loss regularizes embeddings, guiding the network to learn transferable knowledge.
• Multiorientation homogeneous prototypes and an uncertainty-rectified random walk strategy improve classification accuracy, especially for boundary pixels.

Main Results:

• The proposed PTFNet demonstrates prominent performance on four public datasets.
• The method effectively extracts and utilizes frequency priors for improved HSIC.
• The approach enhances knowledge transfer in cross-domain few-shot settings.

Conclusions:

• PTFNet offers a pioneering solution for cross-domain few-shot HSIC by integrating frequency priors.
• The network successfully addresses limitations of existing methods by leveraging physics-guided principles.
• The results validate the effectiveness of the proposed modules and strategies for robust hyperspectral image classification.