A Semantic Seafloor Mapping Methodology for Enhanced Monitoring of Posidonia Oceanica Meadows

We propose a fully autonomous system for real-time semantic mapping of Posidonia oceanica (Po) meadows using an autonomous underwater vehicle (AUV) equipped with onboard visual and acoustic sensors. The system integrates computer vision, machine learning, semantic occupancy mapping, and reactive motion planning to enable the AUV to characterize seafloor habitats without any prior knowledge of the environment. By processing visual data in real time, the vehicle continuously updates a georeferenced semantic map of the seafloor and adapts its trajectory to maximize coverage of the meadow's spatial extent, improving efficiency relative to traditional preplanned survey strategies such as lawn-mower patterns. The mapping process is achieved entirely from onboard sensing, without reliance on surface infrastructure. Field trials conducted in a real-world coastal scenario demonstrated the system's robustness and mapping accuracy, with the resulting seagrass distribution maps closely matching aerial imagery. The AUV exhibited adaptive behaviors such as speed modulation, recovery from perception failures, and safe autonomous termination in shallow areas. This work showcases the feasibility of deploying perception-driven AUVs for scalable, cost-effective semantic mapping and monitoring of Po meadows and other benthic habitats.