A Classifier Adaptation and Adversarial Learning Joint Framework for Cross-Scene Coastal Wetland Mapping on Hyperspectral Imagery

Coastal wetlands play a crucial role in maintaining global ecological security and biodiversity. Hyperspectral image (HSI), with its rich spectral information and high spatial resolution, has emerged as a powerful tool for fine-grained classification of coastal wetlands. However, cross-domain classification of HSI is challenging due to high dimensionality, redundant bands, small inter-class differences, large intra-class discrepancy across domains, and limited labeled data caused by high annotation costs. In addition, coastal wetlands exhibit significant dynamic changes, such as seasonal and tidal variations, which lead to large differences between scenes and further increase the difficulty of cross-scene classification. In recent years, domain adaptation (DA) techniques have shown great promise in addressing the issues of limited labeled samples and distribution shifts across domains in HSI analysis. To tackle the cross-scene classification problem in coastal wetlands, this paper proposes a novel domain adaptation framework that integrates classifier adaptation with adversarial learning to enable effective knowledge transfer from a labeled source domain to an unlabeled target domain. Specifically, we design a domain-invariant feature extractor composed of a spatial perturbation-aligned autoencoder (SPAA) and a long-range spatial–spectral attention (LSSA) network to enhance cross-domain feature representation. To enhance the semantic discrimination in the target domain, we propose a cross-domain cross-entropy loss to achieve semantic alignment from source to target domain. A dual-classifier structure is introduced as a discriminator, where the cross-domain cross-entropy loss facilitates the adaptation of the dual classifiers by encouraging consistent and discriminative predictions on target samples. Extensive experiments conducted on two representative coastal wetland hyperspectral datasets demonstrate that the proposed method achieves superior performance and exhibits strong generalization capability in cross-scene classification tasks.