Understanding the environment is crucial for autonomous navigation, and panoptic segmentation serves as a key enabler by providing a comprehensive representation of both objects and their surroundings. This work presents a robust panoptic segmentation model for Unmanned Surface Vehicles (USVs) in maritime environments using the Mask2Former framework. Unlike traditional segmentation methods, which often falter in water-based scenes due to reflections, waves, and varying illumination, Mask2Former leverages a transformer-based architecture with a ResNet50 backbone, pixel decoder, and masked attention-based transformer decoder to enhance object recognition. Evaluated on the LaRS dataset, our model achieves a Panoptic Quality (PQ) of 29.250, with strong background segmentation performance (Stuff PQ: 35.301) but lower accuracy in instance-level recognition (Things PQ: 5.353). Compared to conventional models such as Panoptic FPN and UPSNet, Mask2Former demonstrates superior performance in background segmentation while offering the advantage of a unified architecture. These results highlight the potential of transformer-based models for maritime perception tasks. Future work will focus on improving instance recognition through data augmentation and fine-tuning to ensure reliable obstacle detection and navigation planning for real-world USV deployments.