Deep Learning-Based Underwater Mapping of Posidonia Oceanica from Satellite Data for Coastal Habitat Monitoring

Abstract. The POSEIDON project aims to develop scalable and repeatable approaches for monitoring Posidonia Oceanica (PO) meadows, a key Mediterranean habitat that supports coastal ecosystem services and long-term blue-carbon storage, yet they are increasingly threatened by warming waters and cumulative human pressures. This work presents a satellite-based workflow for benthic habitat mapping that combines Sentinel-2 multispectral imagery, ancillary bathymetry, and deep-learning semantic segmentation. Sentinel-2 Level-2A data and bathymetry were integrated into multi-band inputs on a common 10 m grid, with analysis restricted to water pixels. A wall-to-wall reference map was generated by harmonising existing habitat products into six benthic classes for supervised model training and evaluation. U-Net and DeepLabv3 architectures with a ResNet backbone were tested for a representative September 2015 scene. The workflow was first assessed in the Culuccia peninsula, where it achieved an overall accuracy of 0.830 and a Kappa coefficient of 0.786. It was then successfully transferred to the Capo Testa - Punta Falcone Marine Protected Area (MPA), where the best-performing configuration reached an overall accuracy of 0.882 and a Kappa coefficient of 0.843. These results show that open-access satellite data combined with robust semantic segmentation models can provide a reliable and non-destructive framework for seagrass mapping in complex coastal environments.