Water samples for stable isotope analysis and chlorophyll content were collected during NOAA Shelf-wide Research Vessel Surveys coordinated by the Oceans and Climate Branch in Narragansett, RI between 2000 and 2005 and 2010–2013 (Table 1). Cruises occurred several times per year over the continental shelf from Cape Hatteras, North Carolina to Cape Sable, Nova Scotia, and a subset of stations, encompassing the full geographic area, were selected during each cruise for stable isotope sampling. Aspects of the 2000–2005 data, which we refer to as Phase I, were presented in McKinney et al. (2010). The Phase I dataset was combined with the Phase II (2010–2013) data to add strength to the spatial analysis presented in this paper, and resulted in a dataset of more than 340 particulate nitrogen and carbon stable isotope values. Carbon isotope data (δ¹³C) were only available for the Phase II dataset. Sample stations spanned the range from the Gulf of Maine to Cape Hatteras and most were unfixed (not in the same location, Figure 1). To supplement our analyses, we matched monthly North Atlantic Oscillation (NAO) index values with the months when research cruises were conducted (Table 2).

For surface sample collection, an on-board flow-through seawater line was used to collect at least 200 ml of water which was first passed through a 300 μm mesh screen to remove zooplankton, and then through a pre-combusted glass fiber (GF/F, GE Whatman, gelifesciences.com) filter. For deep samples, a 10 L Niskin bottle (General Oceanics Inc., Miami, FL USA) was used to sample the bottom of the water column (to a maximum of 500 m) and the water captured by the Niskin was treated in the same manner as the surface seawater. Filters were frozen while the research cruise was underway. They were then dried in the lab in a 65°C oven for a minimum of 24 h. Once fully dry, filters were pelletized and analyzed on an Isotope Ratio Mass Spectrometer for δ¹⁵N and δ¹³C.

For surface chlorophyll-a (hereafter referred to as chlorophyll) samples, 200 ml of water was collected from the flow-through seawater line, passed through a 300 μm mesh screen and then passed through a GF/F glass fiber filter. The glass fiber filter was then placed in 7 ml of 90% acetone and stored in the freezer for 24 h (JGOFS Protocols, 1994). The extracted chlorophyll was read using a Turner Designs 10-AU fluorometer (Turner Designs, San Jose, CA USA). A solid standard was used to correct for instrument drift each time a reading was taken. For deep samples down to a maximum of 500 m, a 200 ml sample was drawn from the Niskin bottle used for the isotope sample and the water captured was treated in the same manner as the surface seawater. To determine the chlorophyll a level, the reading of the blank 90% acetone was subtracted from the extracted sample reading and multiplied by the solid standard calibration value and the ratio of volume of water filtered to the volume of acetone used to extract the chlorophyll.

The δ¹⁵N and particulate nitrogen (PN) content was determined using an Isoprime 100 Isotope Ratio Mass Spectrometer interfaced with a Micro Vario Elemental Analyzer (Elementar Americas, Mt. Laurel, NJ). The nitrogen (δ¹⁵N) isotope composition was expressed as a part per thousand (permil) deviation (‰) from air, while the carbon (δ¹³C) was referenced to Vienna PeeDee Belemnite where δX = [(R_sample − R_standard)/R_standard] × 10³, X is δ¹⁵N or δ¹³C, and R is the ratio of heavy to light isotope (¹⁵N: ¹⁴N, ¹³C: ¹²C). Samples were analyzed randomly in batches of ~30. We used laboratory working standards of cystine, urea, graphite, dogfish, and blue mussel tissues to check for instrument drift in each run and to correct for instrument offset. Instrument precision is better than 0.3‰. The PN content on the filter was calculated by comparing the peak area of the unknown sample to a standard curve of peak area vs. standard N content. To determine the PN concentration, the N content of the material retained on the filter was divided by the volume filtered.

The full dataset that was used for all subsequent statistical analyses is open-access and can be found at: http://tinyurl.com/gu9r34e.

Random forest modeling of the full Phase II dataset (including surface and bottom data) was used to explore the spatial structure as well as the potential environmental drivers of isotope values (Table 2). Random forest is a machine learning technique that aggregates a multitude of decisions trees in order to obtain a consensus prediction of the response variable (Breiman, 2001). Each tree model is constructed based on a recursively bootstrapped sample of the full data and fit using a subsample of the predictor variables. The out-of-bag (OOB) data, or cases left out of the model, are then used as an unbiased estimate of error and also to estimate variable importance. This process is repeated numerous times; specifically, we created 1000 trees for each of our final models. Thus, this ensemble approach avoids overfitting the final model to the data. All random forest modeling was conducted in R v. 3.2.4 (R Core Team, 2016) specifically using the randomForest package (Liaw and Wiener, 2002). Predictor variables were developed to assess the potential drivers of temporal and spatial stable isotope variation (Table 2).

As stated above, the OOB data were used to robustly estimate the variable importance. For each regression tree, the OOB data were used as test data to quantify mean square error. Also, for the subsampled variables of the respective tree, each variable was iteratively randomly permuted and the change in prediction was recorded. Relatively small changes in mean square error would indicate that the variable had relatively little impact on the overall model function. Conversely, large changes indicate that a particular variable had large impacts on the model's predictive ability.

Finally, to look for relationships between δ¹⁵N and depth in the subsurface samples of the Phase II data, nonlinear regression was used.

To create a visualization of continuous isotopes concentrations across the study area, we used kernel interpolation with barriers with the Phase I and Phase II surface data only. This approach is similar to a local polynomial interpolation that allows non-Euclidean measurement of distances between nearest neighbors. Specifically we used the generalized coastline layer as a barrier which forced an as-the-fish-swims distance measure. This analysis was conducted in ArcGIS v. 10.2 (ESRI, 2013). Because results of random forest analyses indicated that temporal variables (year, season) were not the primary drivers of the observed spatial patterns in δ¹⁵N and δ¹³C, we used all available surface data in the interpolations.

The final random forest models were built using exclusively Phase II data. The second phase of the data collection added carbon isotopes and chlorophyll concentrations, which allowed for an additional model beyond the nitrogen isotopes. These models included 158 Phase II data points that had all measurements for response and predictor variables. The carbon model explained 65% percent of the variance and has a mean-squared error of 1.96, while the nitrogen model explained 57% of the variance and had a mean-squared error of 2.70.

Chlorophyll concentration was the most important predictor of δ¹⁵N values, with higher chlorophyll associated with higher δ¹⁵N values (Figure 4A). The other stronger predictors of δ¹⁵N values were water column depth, position, particulate N, and distance to shore (Figure 4A). In general, subsurface water samples had higher δ¹⁵N than surface values; the highest values measured (>10‰) were from the deeper water samples (Figure 2). While subsurface δ¹⁵N values were significantly higher with depth (F = 16.6, P < 0.001), the relationship was coarse (R² = 0.27), where δ¹⁵N values at the shallowest depths (~20 m) ranged from 4 to 12‰ and from about 4–18‰ among the deepest samples (~500 m).

The NAO index was the most important predictor of δ¹³C values, far beyond the other variables (Figure 4B), but this relationship appears to be primarily driven by high δ¹³C values from a research cruise (DE1102) conducted in a month with a very low NAO index (−2.25, Figure 5). These higher δ¹³C values were also positively correlated with the PN content of the filters, where PN was the second strongest predictor of δ¹³C values on the shelf. While year was also a strong predictor of δ¹³C values, this was likely also driven by the 2010 data, which had, on average, higher δ¹³C values than the other research cruises (as well as a very negative NAO index).

Both high chlorophyll concentrations and high particulate N were indicators of higher productivity and, as suggested by this study, greater δ¹⁵N values. Our kriged image of all available surface nitrogen isotope data associated the highest surface values with near-shore areas and lower latitudes. Similar differences between higher nearshore and lower shelf edge δ¹⁵N values were seen in fish and invertebrates off the Newfoundland and Labrador coasts, northeast of our study area (Sherwood and Rose, 2005). Given that this portion of the Canadian coast was minimally developed, the authors didn't attribute the higher nearshore δ¹⁵N values to anthropogenic sources. Rather, they suggested that the observed differences may have been attributable to different productivity regimes, potentially associated with a range of processes, like the upwelling of nutrients (Sherwood and Rose, 2005). Similar patterns were also observed in our study area when we considered only the Phase I data (2001–2005; McKinney et al., 2010). McKinney et al. (2010) hypothesized that higher nearshore δ¹⁵N values are partially driven by continental and estuarine runoff. Estuaries and large rivers have been estimated to contribute between 464 and 627 × 10⁹ moles y⁻¹ to the continental shelves (<200 m deep) of the North Atlantic Ocean (Nixon et al., 1996), which was thought to be more than 20 times the potential contribution via biological N fixation (~20 × 10⁹ moles N y⁻¹). However, N upwelled from the continental slope was estimated to be much greater than all other sources combined, contributing between 2000 and 6000 × 10⁹ moles y⁻¹ (Nixon et al., 1996). Both terrestrial runoff and upwelled slope N may be contributing to the higher nearshore δ¹⁵N values.

In their review of larval transport mechanisms on the Northwest Atlantic continental shelf, Epifanio and Garvine (2001) identified the importance of buoyancy-driven flow and wind-driven upwelling circulation on larval transport. They describe north-to-south nearshore surface currents along the coast of the northeastern United States that transport fresher estuarine water south. This, combined with northward wind-driven upwelling, which brings nutrient-rich water into the photic zone, serve not just to transport larvae, but to support nearshore productivity (Epifanio and Garvine, 2001; Townsend et al., 2006). Deeper water tends to be more nutrient rich and to have higher δ¹⁵N values (e.g., Altabet, 1988). Our measurements of deeper water PM were consistent with this, as δ¹⁵N values tended to be higher at depth. The challenge in interpreting relative contributions of N from continental runoff and upwelling via δ¹⁵N values is that both are characterized by high δ¹⁵N values. Thus, while we can suggest that the higher δ¹⁵N values observed closer to shore may be attributable to some combination of these sources, we cannot parse out relative contributions. Contributions from both of these relatively higher nutrient sources would result in high isotope values close to shore, as well as high chlorophyll and PN concentrations. Our observed lower δ¹⁵N values north of Cape Cod and particularly in the Georges Bank are consistent with previously reported ranges, and may reflect differences in nitrogen dynamics related to temperature, or possibly differences in rates of upwelling (Fry, 1988; Wainright and Fry, 1994).

Relationships between carbon isotopes and productivity are fairly well established in both estuaries (Oczkowski et al., 2010, 2014) and the open ocean (Ostrom et al., 1997; Graham et al., 2010), where higher rates of primary production are associated with higher δ¹³C values. The random forest results were consistent with this, in which the NAO index and particulate N were the two biggest predictors of δ¹³C values. While higher amounts of particulate N indicate higher stocks of plankton in the water column, the link to NAO is more tenuous. The June 2012 data had, on average, higher δ¹³C values than the other cruises and coincided with a particularly strong NAO index (Figure 5). A very negative NAO phase is typically associated with colder and wetter, often snowier, weather in this region, which in turn can result in a greater than average spring freshet, but also with drier conditions in non-winter months. However, overall, hydrologic associations with the NAO are complex and difficult to generalize (Bradbury et al., 2002a,b; Pociask-Karteczka, 2006). In general, if δ¹³C values track production (e.g., Oczkowski et al., 2010, 2014), and production is impacted by the NAO, then a link between δ¹³C values and NAO phase is logical. However, we do not have the data to establish such a link and any such relationships are speculative.

In their review of open ocean isoscapes, Graham et al. (2010) observed that higher-latitude seas were often characterized by more negative δ¹³C values than lower latitudes. They correlated these spatial trends to lower photosynthetic rates associated with the colder regions. Our analyses also indicated a similar relationship between δ¹³C and geographic region. While admittedly coarse, results of kriging all available surface carbon isotope data show higher values closer to shore and in the mid-Atlantic region. These higher nearshore δ¹³C values may be reflecting relatively higher productivity supported by continental nutrient runoff and shoaling or upwelling of deeper, more nutrient rich waters, as described in association with higher δ¹⁵N values.

There are clear similarities in spatial trends in the kriged surfaces for δ¹⁵N and δ¹³C values in particulate matter, where the highest values were observed closer to shore and south of Cape Cod. This is an area of (relatively) enhanced primary productivity that is supported by nutrients from continental runoff and upwelled from bottom waters. However, it is important to emphasize that it is not appropriate to use these surfaces to establish isoscapes or to identify specific values as characteristic of a region. The random forest modeling results indicated that neither distance from shore, nor water column depth, were particularly strong predictors of stable isotope values. This dataset challenges the assumption that offshore δ¹⁵N and δ¹³C values are uniform and suggests that we need to use caution when assuming characteristic isotopically low (δ¹⁵N < 6‰, δ¹³C < −20‰) shelf contributions to an estuary.

This relatively long time series of about twice yearly sampling of stable isotope (δ¹⁵N, δ¹³C) values from the Gulf of Maine to Cape Hatteras was undertaken to capture robust spatial patterns. Our sampling does not capture episodic events like spring blooms and warm core rings, which can have dramatic, if ephemeral, effects on δ¹⁵N and δ¹³C values. For example, while others have measured seasonal shifts in δ¹³C values associated with diatom blooms on Georges Bank, which is in the Gulf of Maine, our sampling did not capture them (Fry and Wainright, 1991; Wainright and Fry, 1994). As another example, detailed documentation of δ¹⁵N values in particulate nitrogen in warm core rings, which were generally observed at the edge or to the east of our study area, demonstrate a shift of up to 8‰ associated with stratification in these rings (Altabet and McCarthy, 1985). Even more dramatically, δ¹⁵N values of up to 40‰ were observed in the particulate N collected from the center of these rings. Despite this, our results can provide some insight into large scale stable nitrogen and carbon isotope dynamics in continental shelf waters, and hence help fill the data gap between estuarine and open ocean studies. While tentative, the possible link between δ¹³C values and large-scale climate oscillations (e.g., the NAO) is intriguing and suggests that spatially and temporally explicit stable isotope measurements made in surface phytoplankton could record global perturbations.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.