Single-Source Frequency Transform for Cross-Scene Classification of Hyperspectral Image.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Cross-scene classification of hyperspectral images (HSI) using domain generalization (DG) is a growing research area.
• Existing DG methods for HSI often rely on data manipulation to create richer samples, but struggle with effectively mining complex HSI features.
• This limitation hinders the performance of newly generated HSI samples in cross-scene classification tasks.

Purpose of the Study:

• To propose a novel single-source frequency transform (SFT) method for enhancing domain generalization in HSI classification.
• To address the insufficient mining of complex features in existing HSI DG methods.
• To improve the diversity and reliability of generated HSI samples for more effective cross-scene classification.

Main Methods:

• Introduced a single-source frequency transform (SFT) framework for domain generalization.
• Developed frequency transform (FT) to learn dynamic attention maps in the frequency space, filtering components to enhance feature diversity.
• Incorporated balanced attentional consistency (BAC), based on class activation maps, to improve the reliability of newly generated HSI samples.

Main Results:

• The proposed SFT method demonstrated superior performance in cross-scene HSI classification compared to state-of-the-art approaches.
• Experiments on three public HSI datasets showed significant accuracy improvements.
• The method achieved up to 5.14% higher accuracy than the second-best performing method.

Conclusions:

• The novel SFT method effectively improves domain generalization for HSI cross-scene classification.
• The combination of FT and BAC enhances feature representation and sample reliability.
• This approach offers a promising direction for advancing HSI analysis in diverse scenarios.