Mapping benthic habitats using a combination of Sentinel2 time series, GAN upscalers and ground truth (acoustic & video)

Accurate benthic habitat mapping at local scale (<1 km) is a prerequisite for assessing the cumulative ecological impacts of anthropogenic effects such as offshore wind farms. Existing global bathymetry such as GEBCO (~100 m horizontal resolution) are too coarse for landscape ecology approaches, and field surveys remain costly and spatially limited. We aimed to develop a cost-effective, multi-source, multi-resolution pipeline to map bathymetry and benthic habitat at a resolution of 3m. We tested the method on the Saint-Brieuc Bay offshore wind farm (Brittany, France). An acoustic survey has been led on the area. Bathymetric data were then derived from a Sentinel-2 time series (325 images, 2015–2026, bands 1–4) processed through a Random Forest (RF) model trained on acoustic echosounder transects, creating a bathymetric map at 11 m resolution, validated with GEBCO within the wind farm (RMSE = 1.44 m, bias = 0.15 m, R² = 0.33). A Real-ESRGAN super-resolution network then improved resolution to 3 m; accuracy against field-recorded transects remained close between the two resolutions (R² = 0.884 vs 0.874, RMSE = 1.74 vs 1.80 m), confirming minimal accuracy loss from super-resolution. Second, 22 morphometric parameters (e.g. slope, curvature, rugosity) were computed from the enhanced bathymetry using SAGA GIS. Third, a second RF model trained on video ground-truth data (79 stations) has been developed to classify benthic habitats across the whole map. As a preliminary step, a binary substrate model (boulder presence/absence) based on bathymetry and morphometric parameters achieved an AUC of 0.727, supporting the relevance of these predictors; full multi-class habitat classification is currently underway. This integrated workflow demonstrates that combining satellite imagery, AI-based super-resolution and field data enables cost-effective, high-resolution benthic habitat mapping at the scale relevant for landscape ecology.