Biotope (macrofaunal assemblage) map and associated confidence layer based on grab and core data from 1976 to 2020

notPlanned

The modelled layer for macrofaunal assemblage is based on a random forest modelling of point sample data from the OneBenthic *(OB, `https://rconnect.cefas.co.uk/onebenthic_dashboard`_/)*dataset, largely following the methodology in Cooper et al. (2019), but with an expanded dataset covering the Greater North Sea and including data from the EurOBI (`https://www.eurobis.org/`_) data repository. Of the 44,407 samples within OB, we selected a subset of 31,845 for which data were considered comparable (i.e. sample acquired using a 0.1 m2 grab or core, processed using a 1 mm sieve and not taken from a known impacted site). Colonial taxa were included and given a value of one. To take account of potential differences in taxonomic resolution between surveys, macrofaunal data were aggregated to family level using the taxonomic hierarchy provided by the World Register of Marine Species (`https://www.marinespecies.org/`_). This reduced the number of taxa from 3,659 to 750. To address spatial autocorrelation in the data, and in keeping with the previous approach, samples closer than 50 m were removed from the dataset, reducing the overall number to 18,348. A fourth-root transformation was then applied to the data to down weight the influence of highly abundant taxa. Data were then subjected to clustering using k-means. A species distribution modelling approach, based on random forest, was then used to model cluster group (i.e. macrofaunal assemblage or biotope) identity across the study area (Greater North Sea). Cross-validation via repeated sub-sampling was done to evaluate the robustness of the model estimate and predictions to data sub-setting and to extract additional information from the model outputs to produce maps of confidence in the predicted distribution, following the approach described in Mitchell et al. (2018). The cross-validation was done on 10 split sample data sets with 75% used to train and 25% to test models, randomly sampled within the levels of the response variable to maintain the class balance. The final model output was plotted as the cluster class with the majority vote of all 10 model runs. An associated confidence map was produced by multiplying map layers for 1) the frequency of the most common class and ii) the average probability of the most common class. Model outputs are used in the OneBenthic Layers Tool ( `https://rconnect.cefas.co.uk/onebenthic_layers/)`_. Cooper, K.M.; Bolam, S.G.; Downie, A.-L.; Barry, J. 2019. Biological-based habitat classification approaches promote cost-efficient monitoring: An example using seabed assemblages. J. Appl. Ecol. 56:1085–1098. `https://doi.org/10.1111/1365-2664.13381`_ Mitchell, P.J., Downie, A.-L., Diesing, M. How good is my map? 2018. A tool for semi-automated thematic mapping and spatially explicit confidence assessment. Env. Model. Softw. 108, 111–122. `https://doi.org/10.1016/j.envsoft.2018.07.014`_ .. _`https://rconnect.cefas.co.uk/onebenthic_dashboard`: https://rconnect.cefas.co.uk/onebenthic_dashboard .. _`https://www.eurobis.org/`: https://www.eurobis.org/ .. _`https://www.marinespecies.org/`: https://www.marinespecies.org/ .. _`https://rconnect.cefas.co.uk/onebenthic_layers/)`: https://rconnect.cefas.co.uk/onebenthic_layers/) .. _`https://doi.org/10.1111/1365-2664.13381`: https://doi.org/10.1111/1365-2664.13381 .. _`https://doi.org/10.1016/j.envsoft.2018.07.014`: https://doi.org/10.1016/j.envsoft.2018.07.014