Global Prototype-Driven Contrastive Partial Adaptation in Hyperspectral Image Classification

Résumé fourni par la source

When labels of hyperspectral images (HSIs) are difficult to obtain, domain adaptation (DA) techniques can employ a knowledge transfer strategy to annotate unlabeled target datasets. However, in real-world scenarios, the source domain often contains abundant, well-annotated data, while the target domain has a smaller set of label categories. In these cases, partial DA (PDA) methods are required to achieve HSI classification (HSIC). Existing methods rely on adversarial learning frameworks and estimate class weights by averaging the predicted probabilities of target samples, but fail to capture interclass discriminability and the underlying structural information within the samples. To address this issue, this article proposes a global prototype-driven contrastive partial adaptation (GPDCPA) network that leverages contrastive learning (CL) to better learn interclass discriminability. Importantly, a new weighting mechanism is designed based on the similarity between target samples and source global prototypes, enabling adaptive reweighting of source samples to mitigate interference from irrelevant categories. Meanwhile, the computation of the target soft prototype and the online pseudolabel self-updating strategy provide more reliable semantic guidance for contrastive alignment. Furthermore, a prototype-level cross-domain contrastive constraint enforces global semantic consistency, while intradomain prototype margin learning and prototype CL jointly enhance instance discriminability and preserve structural relationships. Experiments and analyses on three datasets demonstrate the effectiveness and robustness of the proposed GPDCPA algorithm.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.