A Multi-Scale Satellite Framework for Mapping Posidonia oceanica Using SkySat and Sentinel-2

Posidonia oceanica is a key habitat-forming seagrass species in the Mediterranean Sea, and its spatial distribution is widely used as an indicator of coastal ecosystem status. Despite its importance, large-scale monitoring of submerged vegetation remains challenging due to the limited availability of in situ observations and the spatial constraints of freely available satellite data. These constains often limits the effective application of machine and deep learning approaches in coastal environments.
In this study, we present a hierarchical multi-sensor framework that integrates very high-resolution SkySat imagery with Sentinel-2 data to enable scalable mapping of P. oceanica. A Random Forest classifier was first applied to SkySat imagery and validated using diver surveys and single-beam echo-sounder data, achieving an overall accuracy of 92.3% (κ = 0.89). The validated SkySat outputs were then converted into spatially filtered pseudo in situ reference data, which were used to train a shallow, patch-based convolutional neural network on Sentinel-2 imagery.
Patch extraction at 10 m resolution, combined with targeted data augmentation, reduced spectral mixing effects and improved model robustness. The Sentinel-2 CNN classification achieved an overall accuracy of 89% (κ = 0.81). Depth-stratified validation results show that both Random Forest and CNN models performed best at depths shallower than 15 m, with classification accuracy is perpendicular to water column influence. The research results indicate that to an extent, the high-resolution pseudo-labelling can effectively compensate for limited field data and support regional-scale seagrass mapping using Sentinel-2. The proposed framework provides a transferable and cost-effective approach for operational monitoring of P. oceanica and other submerged coastal habitats using multi-sensor satellite observations