Bathymetry, KdPar, sea surface temperature, fetch, current velocity and other variables were used as predictors. Random forest models were fitted and evaluated using N=393 Zostera marina presence-only data. 10 sets of N pseudo-absences were randomly drawn to consolidate 10 alternative presence/pseudo-absence datasets. 79 occurrences (20% of N) and 79 pseudo-absence were set aside for independent model validation. To capture results sensitivity to pseudo-absences, 10 Random Forest models were fitted on each dataset. Coastwide extensive rasters of the predictors were used to predict habitat suitability along the French Atlantic and English Channel coastline using each model. An Ensemble modelling approach was used to produce a single mean habitat suitability estimate using a weighted mean across individual model predictions based on goodness-of-fit as assessed by the Area Under the receiver-operating Curve (AUC). Standard deviation around the mean habitat suitability index corresponds to the standard deviation across the 10 model estimates. Habitat suitability estimates were classified into binary 0/1 predictions according to model-specific threshold using the validation dataset and the specificity=sensitivity optimisation criteria. The dominant mode across the ten model binary estimates is presented in the binary habitat suitability index raster. In a final step, as Zostera marina cannot settle on hard bottoms, rock polygons from EMODnet Geology’s seabed substrate product were used to post-process the raster products as follows: in cells where rock is present, the two habitat suitability indices were set to 0, and the standard deviation set to NA.

Presence records: Zostera marina occurrence datasets were collated across a range of local studies completed by different organisations (e.g. Ifremer, French Office of Biodiversity, Marine Nature Parks, universities) in the context of different ecological monitoring initiatives (Natura 2000 network, OSPAR, CARTHAM project, REBENT network etc.). All observations (surface and point) regardless of the sampling technique (e.g. aerial photos, diving etc.) were considered over the recent period, i.e. 2000 to 2016 (older records were discarded).

Environmental variables: eight covariates were used to fit the Random Forest models as described below.

-depth: EMODnet Bathymetry 100m DTM

-KdPAR (diffuse attenuation coefficient of downwelling photosynthetically available radiation, proxy of water transparency): EMODnet Seabed Habitat data product, spatial resolution 250m, averaged values (temporal coverage 5 years -2005-2009, temporal resolution 1 day)

-PAR (photosynthetically active radiation): EMODnet Seabed Habitat data product, spatial resolution 4000m, averaged values (temporal coverage 5 years -2005-2009, temporal resolution 1 day)

-Fetch: Burrows Wave fetch GIS layers for Europe, spatial resolution 100m (https://doi.org/10.6084/m9.figshare.8668127)

-Chlorophyll-a, phosphate, nitrate, dissolved oxygen: Ifremer MARC archive, spatial resolution 4000m, averaged values (temporal coverage 7 years -2012-2018, temporal resolution 1 day)

-air temperature: Arome – Météo France archive, spatial resolution 1000m, averaged values (temporal coverage 7 years -2012-2018, temporal resolution 1 hour)

-Tidal current velocity, water temperature, salinity: Ifremer MARC archive, spatial resolution 500m, averaged values (temporal coverage 6 years, 2010-2015, temporal resolution 1 hour)

-Wave energy: Ifremer Homère archive, spatial resolution 300m, Percentile 90th values (temporal coverage 6 years, 2010-2015, temporal resolution 1 hour)