Résumé
Semi-supervised domain adaptation methods aim to generalize from a labeled source domain to a scarcely labeled target domain, despite potential distribution shifts. This challenge becomes even more complex when the source and target data differ in modality representation, such as when acquired using sensors with varying characteristics. In remote sensing, for example, data may differ in acquisition modes (e.g., optical vs. radar), spectral properties (e.g., RGB vs. multispectral), or spatial resolution. This scenario, known as Semi-Supervised Heterogeneous Domain Adaptation (SSHDA), introduces additional challenges due to the heightened distribution shift caused by modality heterogeneity. This work builds upon SHeDD (Semi-supervised Heterogeneous Domain Adaptation via Disentanglement), a recently proposed method for SSHDA, adapting it to address a more challenging scenario where the target domain has extremely limited labeled data. By introducing a novel adversarial module to compel the domain-invariance of features and incorporating a mean teacher approach to improve pseudo-labeling, the enhanced framework demonstrates robust and stable performances. Empirical evaluations on a well-known remote sensing benchmark for scene classification involving heterogeneous data (radar-multispectral) show that the proposed method outperforms both its predecessor and state-of-the-art competitors. Our code is publicly available at https://github.com/giu-guarino/A-SHeDD.