<p>Cross-domain vessel re-identification (Re-ID) aims to transfer knowledge from labeled source domains to unlabeled target domains. However, two key challenges hinder the effectiveness of existing pseudo-label-based cross-domain Re-ID methods: first, the degradation of clustering quality due to inaccurate classification of target domain categories in open-set environments; second, the interference caused by false pseudo-labels from difficult samples during model training. To address these challenges, we propose a cross-domain Re-ID method via density-guided sample clustering (DGSC) and density-guided progressive training (DGPT). Our method guides the model to progressively optimize recognition performance by leveraging sample distribution density in high-dimensional space, thereby enhancing model’s adaptability. Specifically, DGSC optimizes evaluation metrics based on the distribution density of cluster samples in the target domain to determine the optimal number of target domain categories, while simultaneously incorporating clustering loss to improve both clustering accuracy and initial vessel pseudo-labels. After determining the number of clustering categories, DGPT guides the model to train progressively from dense to sparse samples and from easy to difficult samples according to their density distribution. This progressive training strategy gradually optimizes the model’s ability to distinguish challenging boundary samples, further enhancing cross-domain vessel Re-ID adaptability. Extensive experiments demonstrate that our proposed method outperforms state-of-the-art methods in cross-domain vessel Re-ID tasks.</p>