Obstacle detection is a crucial component in unmanned surface vehicles for collision prevention and unnecessary stopping at false detections. Autonomous vessels are a rather unexplored area compared to autonomous ground vehicles, thus there are much fewer annotated datasets for training modern obstacle detectors. Since manual acquisition of ground truth segmentation data is time consuming and expensive, a viable alternative is training with low supervision. In the master's thesis, we focus on evaluation of unsupervised domain adaptation methods, which use an annotated source dataset and a target dataset without annotation. We test four modern adaptation methods: Intra doman adaptation, Fourier domain adaptation, Instance matching and Bidirectional learning. We perform analysis on complete WaSR  method, which is currently state-of-the-art in the field of semantic segmentation on water domain, and on a reduced WaSR method version, without additional regularizations. Our analysis shows, that on reduced WaSR, Fourier domain adaptation gets the best F-measure, which outperforms original WaSR trained without adaptation by over 6%. We then test the same adaptation method on original WaSR and discover, that the Fourier method underperforms the complete reference network for approximately 7% F-measure, and outperforms for approximately 3% if we use WaSR method with only IMU.