Deep Learning-Based Multitemporal Mapping of Posidonia oceanica from High-Resolution Multispectral Satellite Imagery

Seagrass meadows of Posidonia oceanica are key coastal habitats in the Mediterranean, yet they remain diffcult and costly to survey frequently with traditional field methods. This thesis develops and evaluates a remote sensing and deep learning workflow that enables routine, spatially explicit mapping of P. oceanica from freely available satellite imagery. The work aligns with the broader aims of the POSEIDON initiative to support European directives through reproducible, non-destructive, georeferenced monitoring of priority marine habitats. The study focuses on the Capo Testa Marine Protected Area in northern Sardinia. Level-2A Sentinel-2 multispectral images from 2015 to 2024 were processed and used to train semantic segmentation models. Two architectures were considered, U-Net and DeepLabv3, each coupled with ResNet encoders. Model performance was assessed on held-out data. The resulting classifier achieved an overall accuracy of about 88% for discriminating P. oceanica from spectrally similar substrates and water. Multi-temporal classification and post-classification change analysis indicate a net change in mapped meadow extent of roughly 8% over 2015–2024, with spatially heterogeneous patterns that are consistent with a combination of local anthropogenic pressures (for example anchoring and coastal works) and climate-related stressors (warming, turbidity variability, and storm exposure). The proposed pipeline is cost effective, scalable, and transferable to other Mediterranean sites. It can underpin operational monitoring and planning while reducing dependence on intensive field campaigns. The study also highlights the value of targeted ground truth through field visits, underwater photography, and photogrammetry to calibrate and validate models and to refine change attribution.