Single-Source Domain Expansion Network for Cross-Scene Hyperspectral Image Classification.

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Cross-scene hyperspectral image (HSI) classification requires models trained on a source domain (SD) to generalize to an unseen target domain (TD).
• Real-time processing demands direct model transfer without target domain retraining, posing a significant challenge in HSI classification.
• Existing methods often struggle with domain shift, leading to performance degradation when applied to new environments.

Purpose of the Study:

• To develop a domain generalization method for effective cross-scene HSI classification.
• To ensure reliable and efficient transfer of HSI classification models from SD to TD.
• To address the need for real-time HSI processing without requiring target domain data for training.

Main Methods:

• A Single-source Domain Expansion Network (SDEnet) is proposed, utilizing generative adversarial learning.
• A generator with semantic and morph encoders employs an encoder-randomization-decoder architecture, incorporating spatial and spectral randomization for domain expansion.
• Supervised contrastive learning in the discriminator and adversarial training optimize the network for domain-invariant feature learning.

Main Results:

• SDEnet successfully generates an extended domain (ED) that bridges the gap between SD and TD.
• The method learns class-wise domain-invariant representations, improving intra-class sample consistency between SD and ED.
• Experiments on HSI and multispectral image (MSI) datasets demonstrate superior performance compared to state-of-the-art techniques.

Conclusions:

• SDEnet provides a robust solution for cross-scene HSI classification under domain generalization constraints.
• The proposed approach enhances model reliability and effectiveness for real-time HSI analysis.
• The method's superiority is validated across multiple datasets, highlighting its practical applicability.