Analysis of spatial patterns in biomass across geographic scales
We analyzed the spatial patterns in demersal fish biomass for relationships with environmental and anthropogenic drivers at three spatial scales (ecoregion n = 21, subdivision n = 45 and grid cell n = 1083, Appendix S2, Figure S2.1), and using average demersal fish biomass data from three time periods (1990-1995, 2000-2005 and 2010-2015; note that all data is used in the time series analysis). For the ecoregion and subdivision scale, we used a structural equation model (SEM), which is a multivariate analysis to describe a network of causal relationships (Grace 2006). The network links were inspired by recent modelling predictions of demersal fish biomass based on a trait-based food-web model (Petrik et al. 2019; van Denderen et al. 2021). As such, we hypothesized that relationships between net primary production and demersal fish community biomass are mediated by pelagic and benthic secondary production. Following the rationale laid out in the Introduction and Figure S2.2, we further hypothesized that demersal fish biomass declines with increasing temperature, fishing exploitation and the mean trophic level of the community. Lastly, we expected seafloor depth to have an indirect effect on demersal fish biomass by changing the flux of detritus to the benthos (Figure S2.2). SEM analyses were performed using the package ‘Lavaan’ in R (Rosseel 2012). Since ecoregions/subdivisions varied in their characteristics, multiple sensitivity analyses were performed to test for potential effects of differences in e.g., the number of grid cells and sampled depth range (Appendix S3, Figure S3.1-2).
We further analyzed spatial changes in demersal community biomass at the grid cell level. This analysis was not done with SEM, as we expect our hypothesized causal structure to differ at more local spatial scales, i.e. effects of fisheries may vary with depth and prey productivity within each region. We used wavelet-revised model regression (Carl & Kühn 2010) to explain finer-scale “within region” variability in demersal fish biomass from the same set of predictor variables included in the SEM. Wavelet-revised model regression is designed for regular grid-based data while accounting for spatial autocorrelation and non-stationarity (i.e. spatial autocorrelation may vary across regions). This may be important for fish distributions due to biotic and abiotic differences across marine regions that could affect fish movement patterns, e.g. Windle et al. (2010). The spatial analysis at the individual grid cell level was done for the same three time periods as the SEM. Both biomass and the exploitation rate (catch/biomass) were log10 transformed. Since a few grid cells had zero catch, we added a small quantity (1 kg per km2 per year) to avoid taking the log of zero. Model fits were assessed using the Akaike Information Criterion (AIC) and the model with the lowest AIC was selected as best candidate. When other candidate models had a difference of 0–2 AIC units, we concluded that models were essentially equivalent and the model with the fewest parameters was selected. The analyses were performed using the package ‘spind’ in R (Carl et al. 2018).
The set of environmental variables used as predictors in both the SEM and wavelet-revised model regression was compiled from several sources. Seafloor depths were measured in 96% of the survey hauls and extracted for the remaining hauls, using the haul coordinates, from bathymetric data per 1/12° grid from the ETOPO1 Global Relief Model with sea ice cover (Amante & Eakins 2009). Temperatures were estimated using the COBE sea surface temperature data per 1° grid and year (www.esrl.noaa.gov/psd/data/gridded/data.cobe.html). Data on bottom temperatures was not available for the entire time series but was used to verify some of our results (Appendix S3, Figure S3.3-4). Fish mean trophic level (MTL ), describing the biomass-weighted mean trophic level of the community, was calculated from the survey data using species-specific trophic level information (Froese & Pauly 2018; Beukhof et al.2019). Fishing exploitation rates were estimated by dividing annual fisheries catch of demersal fish with demersal fish survey biomass. Fisheries catch data, available on a 30-minute spatial grid, were obtained from Watson (2017) and estimated as the sum of fisheries landings, illegal, unregulated and unreported catch and discards at sea. Net primary production was obtained from the cafe algorithm using MODIS data per 1/6° grid and averaged between 2005 and 2010 (http://www.science.oregonstate.edu/ocean.productivity) (Silsbeet al. 2016). Estimates of pelagic and benthic secondary production were based on output of GFDL’s Carbon, Ocean Biogeochemistry and Lower Trophics (COBALT) ecosystem model from a climatology of the global earth system model (ESM2.6) representative of the contemporary ocean under 1990 greenhouse gas concentrations (Stock et al.2014, 2017). Simulated mesozooplankton biomass and productivity in ESM2.6 broadly captures observed and estimated contrasts across Large Marine Ecosystems (Stock et al. 2017), and the energy available to fish through this pelagic pathway can be estimated as mesozooplankton production not consumed by other mesozooplankton (Zflux ). ESM2.6-COBALT also simulates the detrital flux that reaches the seafloor, which is used as a proxy for benthic secondary production (Dflux ). For all the predictor variables described above, we obtained an estimate per area and year, averaged for each time period, with the exception of net primary production and pelagic and benthic secondary production where a fixed mean value was used due to data limitations and uncertainties in the estimated values over time.
In order to compare and explore the robustness of our empirical SEM results, we used a trophodynamic model to predict demersal fish biomass for each subdivision and ecoregion. We compared these with the observed estimates using linear regression and obtained the explained variance (R2) and Root Mean Square Error (RMSE). In the trophodynamic model, modified from Stock et al . (2017), we assumed that energy flux into the fish community is in equilibrium with the fisheries harvest out of the community after accounting for food chain length variations and trophic transfer efficiency. Demersal fish biomass B in each region i can then be estimated by dividing the flux with the observed fisheries exploitation rate (ER ):
\({B_{i}=(D}_{flux,i}{\times\text{TE}_{i}}^{\text{MTL}_{i}-1}+p_{i}\times Z_{flux,i}{\times\text{TE}_{i}}^{\text{MTL}_{i}-2.1})/\text{ER}_{i}\)eq. 1
Following the approach of Pauly & Christensen (1995) and Stock et al. (2017), zooplankton were assigned to trophic level 2.1 and detritus to 1, such that the number of trophic steps separating the zooplankton/detritus flux (Zflux andDflux ) from the fisheries catch was estimated byMTL minus 2.1 or 1. We further assumed that only part of the zooplankton production is available to demersal fish and this fraction is proportional to p , which is estimated as the fraction of demersal fish catch relative to total fish catch in each region. The value of p was obtained from Watson (2017). The final parameter is the trophic transfer efficiency (TE ). This parameter controls the decay of energy between trophic levels and was varied from 0.05 to 0.15 (Eddy et al. 2021). Additionally, we varied the thermal sensitivity (Q 10) of TE from 0.2 to 2.5:\(\text{TE}_{i}=TE\bullet{Q_{10}}^{\frac{T_{i}-10}{10}}\) ,with T being the average temperature in each region i .
Analysis of temporal biomass variation in ecoregions
We examined whether the changes in demersal fish biomass with temperature, as observed in the SEM and WMR spatial analyses, also drive biomass changes in ecoregions over time. We estimated the influence of temperature on demersal community biomass and production by fitting different recursive biomass and surplus production models to the data (see Table 1 for model details). We used different models to vary how temperature may affect the demersal fish community. In each model, ecoregion was included as a random effect and fishing catch was treated as an offset. The temperature term was centered on the mean temperature per ecoregion (obtained from the COBE sea surface temperature data but see Figure S3.4) to limit our analysis to temperature changes within each region. We scaled biomass and production to the maximum biomass per ecoregion. We evaluated each model with/without a temperature term using AIC, where a model with temperature was considered the most parsimonious if it is at least 2 AIC units lower.