<p>Accurate infrared ship segmentation is vital for maritime surveillance but is severely hampered by costly annotation and the large visible-to-infrared domain gap. Existing unsupervised domain adaptation methods often fail, crippled by the generation of unreliable pseudo-labels. To address this problem, we propose a data-agumented domain adaptation network (DADANet) for infrared ship semantic segmentation, whose core innovation is a targeted data migration strategy that rectifies flawed pseudo-labels. The network first identifies low-confidence regions in the unlabeled target domain. Then, it implants corresponding labeled pixels from the source domain into these areas, creating high-quality mixed-domain images with corrected labels. Subsequent self-supervised training on these augmented images enables the network to learn domain-invariant features, achieving a substantial leap in segmentation accuracy. Experiments on custom and public datasets confirm DADANet significantly outperforms all state-of-the-art approaches, validating its effectiveness.</p>