An ensemble high‐resolution projection of changes in the future habitat of American lobster and sea scallop in the Northeast US continental shelf

To address the uncertainty associated with climate‐driven biogeographical changes in commercial fisheries species through an ensemble species distribution modelling (SDM) approach.


| INTRODUC TI ON
Species distribution and abundance are central components of ecological research and critical to both conservation planning and fisheries management (Franklin, 2010). The biogeography of many marine species is currently shifting in response to climate-related changes in water temperatures and other oceanographic conditions (Hollowed, Planque, & Loeng, 2013;Nye, Link, Hare, & Overholtz, 2009;Pinsky, Worm, Fogarty, Sarmiento, & Levin, 2013). Altered biogeography of a species poses several management challenges because changes in species distributions can move stocks in and out of fixed management boundaries (Gaines et al., 2018;Pinsky et al., 2018). Predicting responses of important fish stocks to future climatic conditions is critical to the implementation of adaptive management measures (Hollowed, Curchitser, Stock, & Zhang, 2013). However, likely shifts in biogeography for many species moving into the future remain largely unknown (Tompkins & Adger, 2004). Correlative species distribution models (SDMs), which rely on quantified species-environment relationships to explain and predict spatial distributions of species, have become important tools in adaptive natural resource management (Franklin, 2010). Long-term projections of species distribution shifts are often sensitive to complexities, properties and assumptions associated with an individual SDM algorithm (e.g. how much interaction among variables is considered or how a species' flexible responses along environmental gradients are fitted; Guisan, Thuiller, & Zimmermann, 2017). An emerging consensus recommends that uncertainty associated with climate-driven changes in species distribution can be better addressed through an ensemble SDM approach that can summarize and represent the information obtained across projections based on all SDMs considered (Gama, Crespo, Dolbeth, & Anastácio, 2016;Mędrzycki et al., 2017).
The Gulf of Maine (GOM), Georges Bank (GB), Mid-Atlantic Bight (MAB) and Southern New England (SNE) are parts of the Northeast US continental shelf Large Marine Ecosystem (NEUS-LME) (Townsend, Thomas, Mayer, Thomas, & Quinlan, 2006). Climatedriven changes in the GOM and GB ecosystem structure are a growing concern for its socio-economically important fisheries (Peck & Pinnegar, 2018). The NEUS-LME is undergoing rapid physical change with water temperature increased at an average of 0.04°C/year between 1982 and 2017 (Banzon, Smith, Chin, Liu, & Hankins, 2016;Reynolds et al., 2007; Figure S1). Recent studies showed a similar increasing trend in the GOM and GB fall water temperature during the period of 1968-2013, with faster warming rates in GB and the MAB (Kleisner et al., 2016(Kleisner et al., , 2017. Furthermore, a recent high-resolution global climate model projection suggests that warming of the NEUS-LME associated with the radiative effects of greenhouse gases may be augmented by warming associated with dynamic shifts in the Gulf Stream associated with climate change (Saba et al., 2016). A projected northerly shift in warm Gulf Stream waters may increase quantities of warm slope water entering the Northwest Atlantic Shelf, possibly leading to warming by as much as 4-5°C along the southern portion of the shelf (MAB and GB) and 3.7-3.9°C in the GOM along with an accompanying increase in salinity. Recent observations suggest that such a shift may be occurring (Caesar, Rahmstorf, Robinson, Feulner, & Saba, 2018).
American lobster (Homarus americanus; hereafter lobster) and sea scallop (Placopecten magellanicus; hereafter scallop) support two of the most economically valuable single-species commercial fisheries in the NEUS-LME with lobster valued at over 669 million USD andScallop at over 486 million USD in 2016 (NMFS, 2016).
These species are vital to the economies and social well-being of coastal communities in this area (NMFS, 2016). The biogeography of both American lobster and sea scallop is keenly affected by changes in climatic variables (Kurihara, 2008;Tanaka et al., 2018;Tanaka & Chen, 2016;Torre, Tanaka, & Chen, 2018. The latest benchmark assessment showed that the lobster stock in the southern range is severely depleted due to several environmental factors (e.g. climate-driven recruitment failure and shell disease; ASMFC, 2015). Recent studies also showed that observed and projected northern shifts in the distribution and habitat of American lobster (Greenan et al., 2019;Pinsky et al., 2013;Stanley et al., 2018). There have also been studies that have linked changes in the distribution of scallop in response to abiotic factors (benthic temperature, benthic salinity, bottom slope) and biotic factors (sea star predation; Lowen et al., 2019). Several species in the NEUS-LME are thought to be depleted due to a failure to recover from intense overfishing along with the southern extent of the species' range resulting from persistent warming conditions (Pershing et al., 2015;Wahle, Dellinger, Olszewski, & Jekielek, 2015). Therefore, long-term changes in the climate regime in the NEUS-LME are expected to greatly impact lobster and scallop habitat suitability (Caputi, Lestang, Flusher, & Wahle, 2013). Providing ensemble projections of climate-driven habitat suitability for these species is of great interest to stakeholders, policymakers and fishery management bodies.
While the effects of climate change are complex and diverse, the impacts on fisheries can be grouped into two general categories: changes to stock biomass or productivity, and changes to stock distribution, each of which poses different management challenges (Brander, 2009;Gaines et al., 2018). This study focuses on the latter, changes to stock distribution, which affects where fish can be caught and who has access to them over time.
To evaluate potential climate change impacts on lobster and scallop fisheries, we considered bottom temperature and bottom salinity as proxies of species habitat suitability. These ecologically relevant and readily available variables have been shown to be  (Kleisner et al., 2016;McHenry, Welch, Lester, & Saba, 2019;Rheuban, Kavanaugh, & Doney, 2017;Tanaka et al., 2018;Torre, Tanaka, & Chen, 2019). The objective of this study is to provide an ensemble projection, generated through a range of different SDMs, of the spatiotemporal changes in habitat of two most important commercial fish stocks in the NEUS-LME, lobster and scallop, in response to projected changes from the high-resolution climate model described by Saba et al. (2016). This study also provides a critical step towards establishing long-term adaptive management measures for lobster and scallop in the NEUS-LME.

| Study area
This study area covers the majority of the NEUS-LME (38.0°-45.0°N and 75.0°-67°W; Figure 1). The NEUS-LME waters are encompassed by the Gulf Stream to the southeast and the U.S. coast to the northwest. These waters are comprised of mixed slope and shelf waters and can be divided into several relatively distinct regional subsystems but are all interconnected to some degree by the Labrador Current which flows southward towards the equator (Townsend et al., 2006). This study was structured around important management zones for lobster and scallop (Figure 1) to incorporate a spatial scale relevant to management interests, as well as to encapsulate projected biological activity from each of these distinct subregions in the NEUS-LME.

| Study species
American lobster are large, mobile, cold-blooded, marine decapods and undergo migration to maintain exposure to optimal biophysical conditions (Caputi et al., 2013). American lobster are ectothermic and climate change can have a pervasive bottom-up influence throughout its larval and post-larval life stages (Caputi et al., 2013;Quinn, 2016;Steneck & Wahle, 2013). Water temperature has a significant impact on lobster's life history and ecology (e.g. recruitment, behaviour and distribution (ASMFC, 2015). For example, the species undergo seasonal migrations likely regulated by temperature change as opposed to selecting for optimal habitat on an instantaneous basis. American lobster can be found spanning a range of temperatures, from −1 to 26°C (Quinn, 2016), while a preference for a narrower range, 12-18°C, and avoidance of temperature below 5°C and above 19°C has been demonstrated through laboratory experiments (Crossin, Al-Ayoub, Jury, Howell, & Watson, 1998). While the increase in water temperature may not negatively affect the availability of thermally suitable habitat, warmer temperature has been linked to the increased prevalence of epizootic shell disease, ESD caused by chitinolytic bacteria (Groner et al., 2016;Hovel & Wahle, 2010;Maynard et al., 2016).
Scallop are bivalve mollusks of the family Pectinidae. In contrast to lobsters, sea scallops are largely sedentary, especially during the adult phase of their life history (Shumway & Parsons, 2016). The species' abundance and distribution are influenced by a multitude of habitat characteristics and ocean currents that interact to control larval settlement and survival into the adult stage, whereas water temperature plays an important role in regulating the distribution of scallop  through larval movement and post-larval survival/mortality (Hart & Chute, 2004;MacDonald & Thompson, 1985a, 1985bStokesbury & Himmelman, 1995;Thouzeau, Robert, & Smith, 1991;Wildish & Saulnier, 1992). Sea scallop occurs mainly at depths ranging from 15 to 110 m throughout its range but can be found in shallower water in the northern part of its range, where it has been reported at depths up to 2 m (Naidu & Anderson, 1984). Adults show optimal growth at a temperature between 10-15°C, with temperatures above 21°C being lethal, and prefer full-strength seawater ~35 ppt, with salinities of 16.5 ppt or lower being lethal (Stewart, 1994). Common predators for juvenile sea scallops in the NEUS-LME include the sea star (Astropecten americanus) and the rock crab (Cancer spp.). These species contribute significantly to scallop mortality and can thus impact their distribution and abundance (Hart, 2006).

| Survey data
The data available for the analysis are a multi-decade scientific bot-

| Environmental information and climate projections
Projected oceanographic conditions used in this study were developed using the delta method (Fogarty, Incze, Hayhoe, Mountain, & Manning, 2008;Hare et al., 2012). The delta method is commonly used for future climate projection, which relies on the difference between future climate anomalies and a baseline regional climatology (historical climate condition). The delta method can remove the climate model projection biases (e.g. drift) and provide a simple and robust projection of future climate conditions (Hare et al., 2012).
The historical bottom temperature and salinity climatologies within the NEUS-LME were developed using high-resolution, quality-controlled monthly means from the Northwest Atlantic regional

| Ensemble species distribution modelling algorithm
Ensemble SDMs for lobster and scallop were calibrated using both presence-absence data and environmental data collected by the available bottom trawl survey programmes ( Figure S2). The environmental variables used for the ensemble lobster and scallop habitat modelling were directly obtained from the scientific bottom trawl survey dataset (1984-2016; Figure S2). We used bottom temperature, salinity and depth that were available at each tow location (Tanaka & Chen, 2016;Torre, Tanaka, & Chen, 2018 (Table S2). For model tuning, the BIOMOD tuning function was used, which uses tuning functions from the CARET R-package to tune GBM, ANN, GAM, MARS, GLM and CTA (Kuhn, 2008), and ENMEVAL R-package to tune Maxent (Muscarella et al., 2014). 10fold split sampling (90% training data and 10% test data) was used to evaluate the models.
Once SDMs were fitted with optimized parameters, all SDMs were run three times each using a randomly chosen 80% of the presence-absence data, with the remaining 20% of the data being used to cross-validate model results. A balance of three runs per each SDM was struck to limit computational demands while still achieving stable results (Thuiller et al., 2016). Two SDM evaluation criteria, receiver operator curve (ROC) and the true skill statistic (TSS), were calculated through cross-validation and used to assess the performance of each algorithm, with higher values for each metric being an indication of higher model skill (Hill, Gallardo, & Terblanche, 2017;Mi, Huettmann, Guo, Han, & Wen, 2017). For both lobster and scallop, best-fitting SDM performance was evaluated against predetermined thresholds (TSS > 0.5 and ROC > 0.8; Hill et al., 2017;Mi et al., 2017 where A i denotes the habitat suitability (probability of presence) a single run of one of the 10 SDM algorithms; TSS i denotes the true skill statistic score received by that run; and n is the total number of all runs of all algorithms to be included in the final ensemble model.   Habitat suitability values from the first 10 years (modelled years 1-10; 1st sample) and last 10 years (modelled year 71-80; 2nd sample) were treated as two empirical distributions at every projection grid across the study area (0.1°, n = 10,497). The KS test was used to assess the null hypothesis that these two distributions are equal (e.g.

| Projections of future habitat
grid-level KS estimates with p < .05 indicating that the distribution of two habitat suitability samples from the first and last 10 years of the projected 80 years are significantly different). given more weight on agreement ratio. For a given management area, a projected trend was considered likely (agreement ratio above 66%) or unlikely (agreement ratio of less than 33%). Furthermore, the agreement ratio from both weighted and unweighted individual ensemble members was compared to assess the robustness of projected habitat suitability changes within each management area.

| RE SULTS
For both lobster and scallop, 16 and 21 tuned SDM runs met the performance thresholds (TSS > 0.5 and ROC > 0.8; Table S3). ANN (2nd run), GLM (all runs), and SRE (all runs) did not meet the performance thresholds and were rejected for lobster, while GLM (all runs) and SRE (all runs) were rejected for scallop (Table S3). The prediction accuracies of the selected SDM runs for both lobster and scallop were considered acceptable by meeting both TSS > 0.5 and ROC > 0.8 thresholds (Tables S3 and S4). Additionally, based on spatially stratified cross-validation, we determine that both overfitting and spatial bias within our modelling framework are negligible ( Figure S6). GBM

| D ISCUSS I ON
Ensemble SDMs are increasingly being used in ecology as they address a major challenge associated with SDM algorithm selection, that can have a large impact on projections (Araújo & New, 2007;Buisson, Thuiller, Casajus, Lek, & Grenouillet, 2010;Forester, Dechaine, & Bunn, 2013). Studies have shown that projections based on a single SDM, out of the myriad biostatistical approaches currently available, can have enough variability to cause misinterpretation of even a simple application (Araújo & Luoto, 2007;Pearson et al., 2006). For example, Pearson et al. (2006)  This study provided an ecosystem-wide projection of lobster and scallop habitat suitability changes in response to simulated 1% year -1 increase in global atmospheric CO 2 concentration, which was characterized by more than 1°C increase in the average bottom temperature in the areas of high lobster and scallop abundance ( Figure   S4). For lobster, the key findings include a marked decrease in fall inshore habitat suitability and a statistically significant increasing trend in habitat suitability in the deeper GOM-GB area ( Figure 5 and  and found that many stocks exhibited a poleward shift with a concurrent increase in depth. We analysed the ratio of agreement between individual model projections to provide a useful measure of uncertainty in the species habitat suitability projections ( Table 2). Assessment of individual SDM agreement in habitat suitability trend identified management areas with robust changes, areas with small changes or areas where models disagree or a combination of those. Statistically weighted ensemble means generally offers a more useful estimate of changes in species habitat suitability distribution as it is less sensitive to outliers. The trends in the ensemble habitat suitability projections were statistically significant in 10 out of 12 management areas (Table 2). However, we showed that there were significant differ- Modelled habitat suitability in this study should be interpreted as a proxy for probability of presence (occupancy) as opposed to actual lobster and scallop habitat suitability, given that measured catch was affected by some niche dimensions and processes not explicitly included in the predictors (e.g. territorial occupancy occurring at smaller scales) Torre et al., 2018). Furthermore, projected species habitat suitability changes in this study should be viewed as a potential change in occupancy of a species due to changes in bottom tem-  Figure   S11). Salinity is a physiologically important environmental variable for marine species, and it can directly influence broad-scale species distribution patterns in nearshore waters and estuaries (e.g. lobsters in Long Island Sound; Tanaka & Chen, 2015;Watson III, Vetrovs, & Howell, 1999). For scallop, as large portions of the study area are well within the species' physiological tolerance (29.9-35.6 ppt), the observed scallopsalinity relationship is likely related to other variables that are structured along salinity gradients in a given area (e.g. origin of water mass; Torre et al., 2019). As is often the case with species-environment modelling, certain variables may be functioning as surrogates for factors directly controlling species distribution through physiological mechanisms (Araújo & Peterson, 2012;Austin, 2007). Regional predator-prey interactions such as sea star predation on sea scallops can also have a significant effect on the distribution of the species at smaller scales (Hart, 2006;Hart & Chute, 2004;Lowen et al., 2019). Furthermore, we used latitude and longitude as proxy variables to capture a wide range of covarying bio-climate factors such as spatiotemporally variable fishing pressure and larval supply (Guernier, Hochberg, & Guégan, 2004;Shumway & Parsons, 2016;Tanaka et al., 2018;Wikgren, Kite-Powell, & Kraus, 2014). As a result, certain areas with high-quality habitats may have a lower probability of detected presence. Another potential limitation of our modelling approach is that the interpolation of survey-derived environmental data masks the scale at which fine-scale habitat selection (active or passive) is occurring for each species. These are important points to consider in future studies. This study focused on evaluating the changing species habitat suitability over large a large spatial scale. Our dataset reflects more than 30 years of aggregated species occurrence-environment relationship characterized by broad spatial and temporal ecological ranges, which can reduce uncertainty in the subsequent projection effort by allowing models to incorporate more complete species' realized niche. As more comprehensive environmental data becomes available in the future, a further detailed ensemble SDM approach could include additional variables such as pH, dissolved oxygen, predator-prey and other food-web interactions to capture a more comprehensive representation of the biogeography of lobster and scallop (e.g. Bio-ORACLE http://www.bio-oracle.org/).
Our study can contribute to the assessment of exploited fishery resources in a rapidly changing environments such as the NEUS-LME. For example, our study can contribute to lobster and scallop assessment by improving the effectiveness of survey efforts and the precision of stock assessment models.
Fishery-independent surveys are a critical component of stock assessment as they provide spatial and temporal information about lobster and scallop stocks (ASMFC, 2015;Johnson et al., 2015). However, the effectiveness of the survey depends greatly on both the bias and precision of abundance estimates (Mier & Picquelle, 2008), and geographically uneven change in lobster and scallop catch could lead to inefficient survey design and the allocation of sampling effort. Even when intensive sampling efforts are conducted, resource limitations often preclude the acquisition of adequate spatial or temporal coverage to capture an entire range of available habitat, or species distribution, which TA B L E 1 Linear trends in the ensemble habitat suitability projections by species, season and management area  et al., 2015). Incorporating climate-driven habitat availability into stock assessments may also improve the model fittings by defining different modelling time periods with respect to these processes . One accomplishment of this study is to provide an ensemble projection of the future habitat availability of the lobster fisheries by synthesizing multiple surveys (Figure 3).
Several studies have documented the climatic impact on the species' biogeography (e.g. Kleisner et al., 2017;Pinsky et al., 2013) based on the offshore survey data alone, which does not cover the inshore waters where over 95% of lobster catches were re- Commercial marine fisheries are complex socio-ecological systems that support social well-being and global food security (FAO, 2018). The ecological, economic and social value of fisheries depends largely on the biomass of fish stocks, with fishing pressure being the main driver of resource status (Hilborn & Walters, 1992). Species distributions are influenced by many interacting biotic and abiotic processes, which can manifest as highly unpredictable occurrence-environment relationships (Boulangeat, Gravel, & Thuiller, 2012;Merow et al., 2014). While there are numerous examples of climate-forced distribution shifts, there has been little progress incorporating population regime changes into stock assessments and management outputs (Link, Nye, & Hare, 2011;Smith, Sameoto, & Brown, 2017

TA B L E 2
Weighted and unweighted agreement ratio for projected habitat suitability changes over the future 80 years. A linear trend from every individual ensemble member SDM projection was classified as either an increase (positive slope with p < .05), decrease (negative slope coefficient with p < .05) or no change (p ≥ .05). Agreement ratio of 1 indicates that all individual SDM projections exhibited the same trend (increase, decrease or no change). Unweighted agreement ratio indicates that all individual SDM projections were considered equally, while the weighted agreement ratio indicates that individual SDMs with higher skills were given more weights on agreement ratio (assigned weights are shown in Figure S2). Number of accepted model runs; lobster (n = 16), scallop (n = 21) Nicholson, 2011). The first step is for management authorities to identify areas of biogeographical changes with reliable projections and associated uncertainties and establish adaptive management strategies to cope with fisheries impacted by ecosystem change.
For example, our ensemble future habitat projections with uncertainty estimates can increase the long-term effectiveness of marine protected areas and spatially explicit catch quotas to reduce pressures on fish stocks that are expected to experience further habitat degradation. Responding to biogeographical changes in natural resources requires a tool that can synthesize large amounts of information and policies that are appropriately adaptive and adequately informed by high-quality projections.
Our study proposed an ensemble means to infer the potential future habitats, based on high-resolution climate data, of two economically important fisheries resources in the NEUS-LME. Through providing a method to alleviate issues associated with variability in ecological predictions across a wide range of currently available SDMs, ensemble modelling approaches provide the distinct advantages of reducing error in projections and providing a more reliable estimate of uncertainty (Araújo & New, 2007;Buisson et al., 2010).
Additionally, the use of geographically comprehensive survey data can reduce bias in the subsequent modelling efforts. Thus, the modelling framework developed in this study adds quality projections of spatiotemporal trends in the distribution of lobster and scallop, which constitute a critical step to improving the management of these species.

DATA AVA I L A B I L I T Y S TAT E M E N T
Both the field survey data and climate model output used in this study were provided by third parties. While we are not able to share the data and model output provided to us, all sources used in this study are available upon request from the following contacts; 1.
The bottom trawl survey data were collected by the Atlantic States Marine Fisheries Commission (State survey data) and the NOAA Northeast Fisheries Science Center (Federal survey data), which may be contacted at jkipp@asmfc.org and james.manning@noaa.gov; 2.
Bottom temperature and salinity output from NOAA GFDL's CM2.6 global climate model, please contact vincent.saba@noaa.gov.