Seagrass Trends.

Seagrass trends were reconstructed following ref. 4. Here, we analyze a subset of sites that included data for the decade 2000 to 2010 (n = 395 sites, with site defined as a location with at least two observations between 2000 and 2010). The Temperate North Atlantic West bioregion had the highest number of sites (n = 121 sites), while the Temperate North Pacific had the least (n = 14). All other bioregions had between 41 and 66 sites (SI Appendix, Fig. S1). For sites with time series containing at least three measurements spanning 2 or more years (n = 344), trends were estimated for each site and bioregion using hierarchical generalized additive models (GAMs), as described in ref. 4. Seagrass meadows can be highly dynamic in nature, so we used GAMs because of their ability to fit complex nonlinear relationships in data, allowing us to fit and estimate trends in seagrass extent. The complexity of the site-level smoothing term, k, varied depending on the length of the time series at each site, allowing more complex terms for sites with longer time series. Site was included as a random effect, allowing for site-specific intercepts (59). We reconstructed trends for the entire length of the available time series (from 1950) and only used predictions from the decade 2000 to 2010 for further analysis. We predicted seagrass extent at the start and end of the decade (i.e., extent in the year 2000 and 2010) by simulating from the posterior distribution (GAMs are a type of empirical Bayesian analysis) of the parameters (1,000 samples). We estimated specific rate of change (yr⁻¹) over time interval, t (10 y in this instance), from the initial to final estimated areas (Ai and Af, respectively) as follows:

For sites with only two observations (also spanning 2 or more years, n = 51), we calculated the specific rate of change using the same formula and assumed this rate was constant for the entire decade. Rates were extracted for each site and were used as the response variable in our predictive models. GAMs were fit in R using the package mgcv (60). Overall, our reconstructions accurately captured trends in seagrass extent for a decade (SI Appendix, Fig. S4).

We grouped sites on a five-point ordinal scale based on trajectories for the period 2000 to 2010: rapidly declining, slowly declining, stable, slowly increasing, and rapidly increasing (Fig. 5). This classification aligned with error rates in seagrass meadow estimates [∼10% (61)], and the reasonably even distribution of sites by categories improved the statistical power of our tests. Sensitivity analyses indicated that parameter estimates from the best model were not sensitive to the rates used to classify trajectories (SI Appendix, Fig. S5). The direction of the parameter estimates did not change, and the estimates themselves were only slightly affected. In each case, turbidity variability was consistently selected as the pressure with the highest probability of an effect.

Trajectory categories used in analyses. Each site was grouped into one of five trajectory categories based on the direction and magnitude of decadal extent change (trajectory).