Domain-Aware Adversarial Domain Augmentation Network for Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Classifying hyperspectral remote sensing images across diverse scenes presents a significant challenge.
• Leveraging historical labeled data (source domain) for unseen samples (target domain) requires models with strong generalization capabilities.

Purpose of the Study:

• To propose a novel single-domain generalization (SDG) network, DADAnet, for cross-scene hyperspectral image classification (HSIC).
• To enhance the generalization ability of models trained on limited source domain data.

Main Methods:

• DADAnet utilizes a two-stage approach: adversarial domain augmentation (ADA) and task-specific training.
• ADA employs a progressive adversarial generation strategy with a domain-aware spatial-spectral mask (DSSM) encoder and a two-level contrastive loss (TCC) to create diverse augmented domain samples.
• Task-specific training involves supervised learning on both source and augmented domain data.

Main Results:

• DADAnet achieved high overall accuracies of 80.69%, 63.75%, and 87.61% on three different hyperspectral image datasets.
• The proposed method outperformed existing domain adaptation (DA) and domain generalization (DG) techniques.

Conclusions:

• DADAnet demonstrates superior performance in cross-scene hyperspectral image classification.
• The domain-aware adversarial domain augmentation strategy effectively improves model generalization for unseen target domains.