POSTER SESSION 1

15:30–17:30, Monday, September 14

Poster Session | 1 | 2 | 3 | 4 | Schedule at a Glance

ABSTRACT 1284 | POSTER 241

UPSCALING BENTHIC HABITAT CLASSIFICATION FROM DRONE TO SATELLITE IMAGERY USING MULTI-TEMPORAL MACHINE LEARNING

Shallow coastal benthic habitats, including seagrass meadows and macroalgae, are declining at critical rates yet remain poorly monitored at regional scales. Satellite-based mapping offers a scalable alternative, but operational benthic classification at 10-metre resolution demands training strategies that can bridge the spectral gap between fine-scale ecological observations and coarser satellite imagery in optically complex coastal waters. No systematic evaluation of training label sources for this upscaling problem currently exists in the literature.

This study presents, for the first time, a direct comparison of field annotation-based versus automated drone classification-based label generation for Sentinel-2 benthic habitat mapping across five biogeographically diverse European coastal sites spanning 37 to 66 degrees North. Sentinel-2 time series were atmospherically corrected using ACOLITE dark-spectrum fitting and processed into approximately 50 water-adapted spectral features per pixel, including red-edge indices, temporal statistics capturing phenological variability, and EMODnet bathymetry. A supervised classifier was applied with explicit handling of class imbalance inherent to rare benthic habitats such as seagrass.

Drone-based training achieved an average F1-Macro of 90 percent compared to 83 percent for annotation-based training, while providing 50 to 100 times more spatially complete samples with substantially reduced class imbalance. Consistent agreement above Kappa 0.6 was obtained across all sites despite contrasting water optical properties from Norwegian fjords to Mediterranean waters These findings establish drone-derived labels as a scalable and reproducible pathway for regional benthic monitoring, with direct implications for how the ocean optics community approaches ground truth collection for satellite habitat mapping.

Gladys Villegas, RBINS, [email protected], https://orcid.org/0000-0002-4462-0485

Dimitry Van der Zande, RBINS, [email protected]

Mihailo Azhar, Aarhus University, [email protected]

Hege Gundersen, NIVA, [email protected]

Ari-Pekka Jokinen, SYKE, [email protected]

Kasper Hancke, NIVA, [email protected]

Peter Anton Staehr, Aarhus University, [email protected]

Valentina Todorova, IO-BAS, [email protected]

Louise Forsblom, SYKE, [email protected]

Dimitar Berov, IBER-BAS, [email protected]

Todor Lambev, IO-BAS, [email protected]

Nuria Marbà, IMEDEA, [email protected]