Predictive Seagrass Habitat Model
Detenbeck, N. AND S. Rego. Predictive Seagrass Habitat Model. U.S. Environmental Protection Agency, Washington, DC, EPA/600/R-15/003, 2015.
We developed an approach for creating statistical models to predict seagrass presence/absence at the scale of individual grid cells (10 m x 10 m) as well as a number of endpoints relative to seagrass presence/absence along transects perpendicular to the shoreline: presence/absence, relative frequency along transects, and minimum and maximum depth of occurrence. Unlike many of the statistical models that have been developed to predict seagrass occurrence, ours take into account the nonrandom distribution of seagrass (patchiness), nonlinear effects of light availability, salinity (as an indicator of total N gradients, and sediment organic carbon, and parameter interactions. We tested this modeling approach using data for distribution of seagrass (Zostera marina) in Narragansett Bay, RI. Models developed were most robust for predictions of shoreline occurrence rather than at the 10 m x 10 m grid scale. Based on model results, multiple factors affect seagrass success. The minimum depth of occurrence is influenced by both wave energy (wave mixing depth) as well as particle size, with finer sediments more susceptible to disturbance by wave action. The maximum depth of occurrence is influenced by both transparency and sediment organic carbon, an indicator of past eutrophication. Depth limits decrease as sediment organic carbon increases due to increased energetic demands for seagrass to counteract effects of increased toxicity of anaerobic sediments (e.g., sulfide toxicity). Our models allowed us to distinguish multiple modes of action for nitrogen effects on seagrass distribution: 1) shading by phytoplankton affecting Secchi depths, 2) shading and/or competition by periphyton and macroalgae mediated by nitrogen concentrations, and 3) effects of sediment organic carbon on minimum light requirements. Incorporation of historic distribution of seagrass patches did not improve model predictions, suggesting that patches may exist in a state of dynamic equilibrium with a lag time for recovery of disturbed sites. Our models detected significant differences in the probability of seagrass occurrence by shoreline even after we factored out the effect of site characteristics. This could be explained by hysteresis effects related to tidal currents. Tidal currents in Narragansett Bay are strong enough to resuspend fine sediments, thus limiting establishment of new seagrass patches, but not strong enough to damage established seagrass beds. Specific restoration measures such as co-restoration of shellfish beds (to reduce suspended sediments and effects of tidal currents and wave action) or use of existing or constructed coastal barriers to limit effects of wave action and tides might improve probabilities of initial colonization success and the initiation of positive feedback effects. Our predictive models have multiple potential applications: identification of aquatic life use zones for setting nutrient criteria for areas of potential seagrass habitat, prioritization of areas and strategies for seagrass restoration, and projection of potential benefits of management actions. Adaptive management will need to take into account different projections for short-term versus long-term recovery due to the multi-decadal persistence of organic carbon in sediments and effects on minimum light requirements for seagrass.
Restoration of ecosystem services provided by seagrass habitats in estuaries requires a firm understanding of the modes of action of multiple interacting stressors including nutrients, climate change, coastal land-use change, and habitat modification. We explored the application of generalized linear mixed models (glmms) and generalized additive mixed models (gamms) to describe the simple and interactive effects of environmental factors on the distribution of a common seagrass, Zostera marina, in Narragansett Bay, Rhode Island. We used a random shoreline effect to account for “founder” (random colonization or extinction) effects. We provide several strategies to overcome three challenges in developing empirical species distribution models to describe and predict seagrass distribution in estuaries: the fine-scale patchiness of seagrass distributions with attendant problems of spatial autocorrelation; the large areas of interest for model development and application entailing significant memory demands for modeling; and the potential co-variance of multiple interacting factors affecting seagrass. We developed a spatial framework describing the coordinates of spatial autocorrelation in estuarine systems, with the main axis parallel to the shoreline and a secondary axis perpendicular to the shoreline. We demonstrated an approach to incorporate a term for residual autocorrelation in glmms first introduced by Crase et al. (2012). To account for anisotropy in the system, we calculated zonal averages of residual errors within rectangular boxes oriented parallel to the shoreline along the longer main axis. We successfully dealt with covariance of influential factors by centering variables, by using multiple strategies to describe the interaction of the light environment and wave energy with depth, and by excluding correlated variables where necessary. We predicted seagrass distribution at the scale of 10-meter grid cells, as presence/absence or average presence/absence associated with shoreline locations spaced at 10-meter intervals, and minimum or maximum depth of distributions at those locations. Prediction of seagrass absolute or average presence/absence at shoreline locations was very robust, with area-under-the-curve (AUC) values associated with Receiver Operating Characteristic (ROC) curves of 0.95 – 0.98 following 10-fold cross-validation of models. Random shoreline effects varied over several orders of magnitude, probably tied to the distribution of tidal currents. For the model predicting seagrass presence/absence at the grid cell scale, the most influential predictor is Secchi depth, followed by (in order): shoreline isolation, sediment percent total organic carbon, sediment type, and salinity. The least influential variable is water depth greater than average wave mixing depth. For the model predicting presence of seagrass at shoreline locations, the most influential predictor is sediment type, followed by sediment percent total organic carbon (at low Secchi depth), then salinity (as an indicator of downstream gradients in water column total nitrogen). As demonstrated in other recent studies, sediment total organic carbon interacts with light availability by increasing energy requirements and the light compensation point for seagrass. For all shorelines combined, our model predicts that following a 40% reduction in TN loads (and concentration) the colonized area would increase from 12% of area in the 0 to 5 meter depth zone to about 63% of area in the short term and slightly more over subsequent decades as sediment organic carbon recovers. Finally, we provide data sources for application of this approach to other U.S. estuaries, with much of the data available through EPA’s Estuary Data Mapper application (www.epa.gov/edm).