IMIS

The purpose of this study was to predict how climate change could affect the species richness of consumers lower in the food web (planktivores), compared to those intermediate (benthivores) and higher up (piscivores).  In brief, a collation of trophic interactions spanning the northeast Atlantic shelf seas has been applied to define feeding guilds by grouping fish species size classes (we use the taxonomic level of Gobiidae and Ammodytes for taxa that are not consistently identified to species) that have prey taxa in common, and whose prey differentiate them from other predator guilds based on cluster analysis. Taxon-specific size categories were defined as: <3 cm as larvae; small juvenile fish between 3 cm and half of length at maturity; juvenile-medium fish from half of length at maturity to length at maturity; medium fish from length at maturity to half-maximum length; and all remaining larger fish as large. Taxon-specific length at maturity and maximum length (i.e., asymptotic length at infinity) were estimated using the R package Fishlife (Thorson et al., 2017).

Environmental projections were derived from the coupled marine ecosystem models POLCOMS/NEMO-ERSEM. We focus on two emission scenarios (representative concentration pathways, RCPs) developed for the Intergovernmental Panel on Climate Change (IPCC)’s fifth phase of the Coupled Model Intercomparison Project (CMIP5): RCP4.5, the “stabilization scenario”, characterised by medium emissions and high mitigations, and the “no mitigation policy” scenario RCP8.5, derived from high fossil fuel emission and low mitigations. We use projections of temperature, salinity, pH, nitrate, phosphate, dissolved oxygen, current velocity, chlorophyll, gross primary production, non-living organic carbon, zooplankton carbon concentration and secondary carbon production by zooplankton in our predictive habitat models. For temperature, salinity, current velocity, dissolved oxygen, and pH both surface and bottom mean annual averages were considered, and in the case of temperature, also the difference between bottom and surface values, to account for stratification. For chlorophyll, gross primary production, non-living organic carbon, phosphate, nitrate, zooplankton carbon concentration and secondary carbon production by zooplankton the total across the water column was used, rather than surface or bottom values. We also include depth (from the General Bathymetric Chart of the Oceans GEBCO; www.gebco.net, at 15 second resolution), distance to coast and substrate composition (median grain size and percentages of mud, sand and gravel from Wilson et al., 2018) to capture key spatial gradients that affect habitat suitability for fish. 

All environmental data was processed onto a 10 km by 10 km grid, and because an annual mean of, e.g., temperature, does not capture the environmental variability that ultimately determines the thresholds within which biota must survive; for temperature, salinity, pH, oxygen and current speed we also include the standard deviation of the 12 monthly means in each year, for all locations within a radius of 75 km of each grid cell, in order to provide a measure of spatio-temporal heterogeneity. For the environmental variables where surface and bottom values were extracted, we use sea surface values to model habitat suitability for planktivores, which are largely pelagic species, and seabed values for the benthivores and piscivores which are largely demersal species. Pairwise Pearson correlation coefficients were computed for the set of environmental variables used to model the habitat of planktivores and of non-planktivores separately, to assess multicollinearity. Variables were removed if they correlated with another >0.7.

Additionally we provide data on model performance and presence/absence thresholds, in the file "Model performance and threshold.xlsx". Model performance is measured via the area under the curve (AUC) of the Receiver Operating Characteristic (ROC) plot, the AUC of the Precision-Recall (PR) plot (He & Garcia, 2009), and the Miller slope (Miller et al., 1991). We assess model performance in both space and time. Spatial performance was analysed via 8-fold block cross validation using the R library “blockCV” (Valavi et al., 2019). Temporal performance of the models was assessed by training a new model that excluded the last 5 years of survey data (2015 to 2019), using the resulting model to predict to those 5 years, a “novel time period” for that model. 

These data were produced for the study in the article "Climate change affects the distribution of diversity across marine food webs", by Murray S. A. Thompson, Elena Couce, Michaela Schratzberger and Christopher P. Lynam, in Global Change Biology (in press).