Connectivity, persistence, and loss of high abundance areas of a recovering marine fish population in the Northwest Atlantic Ocean

Abstract In the early 1990s, the Northwest Atlantic Ocean underwent a fisheries‐driven ecosystem shift. Today, the iconic cod (Gadus morhua) remains at low levels, while Atlantic halibut (Hippoglossus hippoglossus) has been increasing since the mid‐2000s, concomitant with increasing interest from the fishing industry. Currently, our knowledge about halibut ecology is limited, and the lack of recovery in other collapsed groundfish populations has highlighted the danger of overfishing local concentrations. Here, we apply a Bayesian hierarchical spatiotemporal approach to model the spatial structure of juvenile Atlantic halibut over 36 years and three fisheries management regimes using three model parameters to characterize the resulting spatiotemporal abundance structure: persistence (similarity of spatial structure over time), connectivity (coherence of temporal pattern over space), and spatial variance (variation across the seascape). Two areas of high juvenile abundance persisted through three decades whereas two in the northeast are now diminished, despite the increased abundance and landings throughout the management units. The persistent areas overlap with full and seasonal area closures, which may act as refuges from fishing. Connectivity was estimated to be 250 km, an order of magnitude less than the distance assumed by the definition of the Canadian management units (~2,000 km). The underlying question of whether there are distinct populations within the southern stock unit cannot be answered with this model, but the smaller ~250 km scale of coherent temporal patterns suggests more complex population structure than previously thought, which should be taken into consideration by fishery management.

. Presently, the Canadian Atlantic halibut commercial fishery is managed as two discrete stocks, the larger of the two is referred to as the Scotian Shelf-southern Grand Banks management unit (DFO, 2015a), and encompasses the waters from the Gulf of Maine off of southwest Nova Scotia to southeastern Newfoundland and Labrador. The smaller stock is contained in the Gulf of St. Lawrence (DFO, 2015b).
The Canadian Atlantic halibut fishery was largely unregulated until 1988, when the management units were established, and a minimum legal length for commercial harvest was introduced to maximize sustainable yield (Neilson & Bowering, 1989;Trumble et al., 1993).
Although admittedly data limited, the definition of "one large stock" for the Scotian Shelf-southern Grand Banks management unit was based primarily on tagging results which indicated the intermixing of Newfoundland and Scotian Shelf halibut, but not those in the Gulf of St. Lawrence Neilson, Bowering, & Frechet, 1987; Figure 1) and supported partially by a recent study on electronically tagged halibut in the Gulf of St. Lawrence (Le Bris et al., 2017). Additionally, length at age comparisons (Bowering, 1986) suggested differences in demographic rates.
There is a long history of halibut exploitation on the Atlantic coast.
Caught first as bycatch in the Atlantic cod fisheries, there are records of abundant halibut in the near shore of Massachusetts Bay during the early 1800s (Grasso, 2008). The directed fishery began there in the 1830s, but sequentially moved further offshore to maintain catch rates. Effort then shifted northward throughout Canadian waters, again moving sequentially as stocks were depleted, eventually exploiting populations as far north as the Davis Strait of the Canadian Arctic in 1866 (Grasso, 2008;Trumble et al., 1993). The halibut fishery continued in Canadian waters throughout the early 1900s, with core fishing areas in southwest Nova Scotia, along the continental shelf edge south of Newfoundland and near the Gully, a deepwater canyon on the edge of the Scotian Shelf (now a marine protected area), and in the Gulf of St. Lawrence (McCracken, 1958; Figure 1). Decades of fishing pressure culminated in large declines of halibut and the fishery through the 1990s (Trzcinski & Bowen, 2016).
Halibut landings in Atlantic Canada have been steadily increasing since the early 2000s (DFO, 2015a(DFO, , 2015bTrzcinski & Bowen, 2016) where the Scotian Shelf-southern Grand Banks Atlantic halibut management unit was certified by the Marine Stewardship Council in 2013 (Martell, Vincent, & Turis, 2013). This stands in stark contrast to the status of halibut at the southern end of the species' range in the United States where the commercial halibut fishery has been under moratorium since 1999 (Department of Commerce (DC), 1999 (Shackell, Frank, Nye, & den Heyer, 2016). This suggests that halibut dynamics occur on a much smaller scale than currently assumed and that US halibut have never fully recovered from historical overfishing (Shackell et al., 2016). Recent (den Heyer et al., 2012;Kanwit, 2007;Seitz, Evans, Courtney, & Kanwit, 2016;Seitz, Farrugia, Norcross, Loher, & Nielsen, 2017) and historical tagging analyses (McCracken, 1958) provide support for the hypothesis that halibut may exist as a series of local populations.
The importance of understanding population spatial structure was acknowledged by fisheries scientists during the late 1800s; however, since the 1960s, fisheries science has focused on the development of quantitative methods to assess the amount of fishable biomass (Stephenson, 2002), with an underlying assumption that any locally depleted population would be replenished by neighboring areas (e.g., Svedäng, Cardinale, & André, 2001). While it is possible for recolonization to occur (Corten, 2013), when there is a lack of recovery in local areas, it suggests that recolonization from elsewhere is not a simple density-dependent response and may be attributed to philopatric behavior (Svedäng et al., 2001;Svedäng & Svenson, 2006) or changes in population structure or demographic rates (Payne, 2010), or that recolonization rates cannot counter excessive fishing pressure (e.g., Shackell, Frank, & Brickman, 2005). Lack of recovery of species, such as Northern cod, has contributed to renewed interest in spatial stock structure in fisheries science (Cadrin, Kerr, & Mariani, 2014) as commercial fishing can erode local concentrations when spatial structure is ignored. Fishing patterns need to be evaluated with respect to stock structure so that smaller, more vulnerable concentrations and associated habitat are not overfished.
Halibut have been experiencing population growth, supported by a period of high recruitment in Canadian waters since the early 2000s (DFO, 2015a(DFO, , 2015bTrzcinski & Bowen, 2016). This upward trend is in contrast to the NW Atlantic continental shelf's documented ecosystem shift from one dominated by large-bodied groundfish to one abundant in invertebrates (Frank, Petrie, Choi, & Leggett, 2005;Shackell & Frank, 2007;Shackell, Frank, Fisher, Petrie, & Leggett, 2010;Worm & Myers, 2003). Here, we present evidence of spatial structure in juvenile halibut in the NW Atlantic through three and a half decades of commercial exploitation by building on evidence that halibut are not habitat-limited (Shackell et al., 2016) and that the majority of tagging studies suggest local residency. We also consider that the lack of halibut's recovery in the United States is likely due to historical local overfishing (Grasso, 2008).
Here, our goal was to explore the spatial structure of juvenile halibut abundance, an index of fisheries recruitment, using a hierarchical borders, the model identified areas of relatively higher abundance that were persistent over time but the abundance of juvenile halibut varied among regimes. We argue that the protection of persistent high abundance areas may have contributed to the recovery of this stock and that sustainable management will need to consider stock structure.

| Data
Data from 27 fishery-independent research trawl surveys con-  (Table 1). Date, latitude, longitude, bottom temperature, depth, abundance (number), and biomass (weight in kg) were recorded for each set. Abundance was standardized within each research survey to account for variations in set duration and distance sampled (Ocean Biogeographic Information System (OBIS), 2014). Annual estimates of stratified mean abundances were used to display regional-scale time series trends.

| Approach: spatiotemporal model
To explore the abundance and distribution of juvenile Atlantic halibut, we employed a Bayesian hierarchical spatiotemporal model fol-  (Table 1).
For spatial analyses of large datasets, SPDE models are efficiently fit using a Gaussian random field (GRF), a discretely indexed spatial process. The GRF models spatial dependence on a mesh and estimates the parameters of the covariance function (i.e., the Matérn covariance function; Rue & Held, 2005; see also Lindgren, Rue, & Lindstrom, 2011 for theory and proofs; and www.r-inla.org). Following Carson and Mills Flemming (2014; in press), we estimated connectivity (coherence of temporal pattern over space) and spatial variance (variation across the seascape) of juvenile halibut abundance across the modeled domain. In addition, a temporal autocorrelation term in the model allowed us to estimate the persistence (similarity of spatial structure over time) of the areas of abundance.
The response variable of interest is juvenile halibut abundance (number) at each location in each time period. All true zero values in the original data were removed from the dataset, but were represented in the spatiotemporal model as nonpositives in the mesh (see Figure 2, smaller triangles where data are positive). As nonpositive values are computationally less demanding, this enabled us to run the analysis.
The model has the following general form; where E(Y s,t ) is the mean of the expected response of Y at location s at time t, ξ(s,t) represents the spatiotemporal latent GRF (a random effect), f j (c j (s,t)) are smoothed functions of the covariates, while j refers to jth of a total n covariates where depth and temperature were the tested covariates. The mean of the response variable, E(Y(s,t)), is mapped by a canonical link function (i.e., log link) to a linear predictor, η(s,t), by the generalized additive model framework (Hastie & Tibshirani, 1990). The spatiotemporal latent GRF, ξ(s,t), represents the cumulative effect of all unmeasured latent factors. The characteristics of this spatiotemporal random effect comprise the spatial and temporal covariance structure of the model (Rue et al., 2009).
In addition to defining the domain and building the mesh, it was also necessary to determine the appropriate statistical distribution of the response variable, the importance of covariates, and the inclusion of the latent field to the model. The error distributions explored included Poisson, negative binomial, and Gaussian. The spatiotemporal covariance structure was also tested, examining whether it was temporally invariant (a single spatial field), and whether the time periods were autocorrelated or independent (see Cosandey-Godin et al., 2015).

| Persistence
Through this process of model building and evaluation, the autocorrelation of the abundance and location of halibut through the three time periods (t = 3) was tested to determine whether the locations persisted from one time period to the next, as measured by the AR (1) term. The AR(1) term, a first-order autoregressive model using the previous time step to predict the most recent, is interpreted as the a parameter, or persistence, which ranges from −1 to 1.

| Goodness of fit
The goodness of fit of the various candidate models was compared

| Connectivity and spatial variance
R-INLA allows estimation of spatial characteristics. Connectivity, or ρ, is conventionally interpreted as the distance at which the spatial covariance of the field decays to 0.13 (Cameletti, Ignaccolo, & Bande, 2011;Cameletti, Lindgren, Simpson, & Rue, 2013). When the connectivity parameter, ρ, is large, it means that the covariance decays slowly in space, that is, the temporal pattern is similar over a large spatial scale. The units of connectivity are degrees Latitude, which are converted to distance in kilometers (km) for added interpretation. Similarly, σ 2 , spatial variance is a relative index of the differences in amplitude across a seascape, and is scaled to the linear scale of the predictors (i.e., log juvenile halibut abundance).
A large spatial variance indicates a large amplitude in the overall field. When ρ and σ 2 are reported as parameters of the model, they provide a description of the multivariate normal distribution of the mean of the response, after accounting for the variables which are explicitly included. Our model, incorporating three fisheries management regimes, has one ρ (index of connectivity) and σ 2 (spatial variance), in addition to one a (persistence) parameter. Bayesian credibility intervals indicating the probability that the parameters lie within a specific range were also examined. To examine the connectivity, spatial variance, and persistence of juvenile Atlantic halibut areas of abundance through the three time periods, these random latent fields were plotted as maps.
F I G U R E 2 Triangulated mesh. Constrained refined Delaunay triangulation mesh. The smaller triangles of the grid are where there are more data points. It is formed of triangles, and the vertices are called nodes (here there are 2,120)

| Time series of abundance indices
Halibut were captured most often along the continental shelf, associated with a deeper edge, and also within the GSL (Figure 1).

Surveys in Nova Scotian waters (the Scotian Shelf and Gulf of
Maine) had the highest percent occupied, 16%, while the United States had the lowest, 2% (Table 1).

| Model selection
First models were tested using alternative likelihood families, temporal structures, covariates, and inclusion of the latent effect. The best performing models had a spatiotemporal random field using a firstorder autoregression model that yielded, a, the persistence parameter (Tables 2 and 3). The latent field ξ(s,t) of each time period (group) was a function of the value of the one previous (i.e., correlated). The DICs and CPOs of models using the Poisson error distribution were consistently better than those for other error distributions. The covariates were largely insignificant most likely because the model focussed on nonzero values. As per the resulting lowest DIC and highest CPO, bottom temperature and the latent variable were retained in the penultimate model (Table 2); η(s,t) = ξ(s,t) + Bottom Temperature. A smoothing function was not applied to bottom temperature.
Visual analysis of the areas of abundance through the three periods illustrates that juvenile halibut distributions have changed over time (Figure 4). Examining the model results, connectivity (ρ) showed that, on average, nodes in the random field domain were significantly spatially correlated up to 250 km (2.25° Latitude) (Table 3; Figure 4).
The spatial variance (σ 2 ) was also quite low (0.12) indicating a relatively flat field (Table 3)

| Spatiotemporal model results
The analysis over three time periods is an initial broad look at the spatiotemporal trends in juvenile halibut abundance in the NW Atlantic. Mean Bottom Temperature (C) G 1980 1985 1990 1995 2000 2005 2010  100 120 140 160 180 200 Year Mean Depth (m) Two areas of persistent abundance were identified, both on Scotian Shelf: (i) Southwest Nova Scotia and (ii) the Gully (see Figures 1 and 4).
Less persistent high abundance areas were also found in the GSL and NL, although the results may be confounded by differences in surveys. In

| DISCUSSION
Here, we demonstrate that halibut exhibit a spatial structure in the NW Atlantic at a scale smaller than the current halibut stock management units. Our analysis showing statistically independent persistent areas of juvenile halibut abundance has provided important evidence that halibut spatial structure is more complex than previously identified, and has varied over the past three and a half decades. Two high abundance areas were consistently present on the Scotian Shelf throughout the time series, southwest Nova Scotia and the Gully (Figure 4).
Despite widespread increase in halibut recruitment during the last decade, previously identified areas of high juvenile abundance have not re-established in southern Newfoundland (Figure 4c). Connectivity of juvenile halibut is an order of magnitude less than the distance assumed by the definition of the management units (~2,000 km; DFO, 2015a). Halibut have a long history of being removed as bycatch in the cod-directed groundfisheries (Grasso, 2008) and persistent patches of abundance along the shelf edge may identify spatial refuges from cod-directed fishing. Coinciding with the reduction in trawling and cod fisheries in the early 1990s, and the introduction of a minimum legal size for landed halibut, halibut have increased since the mid-2000s and catches in the Canadian trawl surveys have been well above the longterm average (DFO, 2015a, 2015b; see also Figures 3 and 4).
While a substantial amount of knowledge is required to identify a discrete population unit, stock assessments, sustainable seafood classifications (i.e., Marine Stewardship Council), and vulnerable species classifications depend on knowing the demographic rates for identified stocks or subpopulations and the impact of fisheries on them.
Sustainable management and recovery of fish stocks are further complicated when fine-scale differences in local populations are observed such as in the Skagerrak cod (Olsen et al., 2008) and yellowfin sole (Limanda aspera) in the Bering Sea (Bartolino, Ciannelli, Spencer, Wilderbuer, & Chan, 2012). In the Northeast Atlantic, evidence of genetic differentiation provides support for the existence of local populations of Atlantic halibut (Foss, Imsland, & Naevdal, 1998;Haug & Fevolden, 1986;Mork & Haug, 1983). In the NW Atlantic, Reid et al. (2005) found no genetic evidence supporting local populations but acknowledged that in the absence of spawning information and with limited sample size, the power to detect differences was not ideal. Genetically distinguishable spawning populations that mix after spawning may not easily be assigned to a home location (Reid et al., 2005). More research is required to determine the genetic population structure of Atlantic halibut.
Even in the absence of genetic differentiation, high abundance areas are especially vulnerable to overfishing if connectivity between areas is low and fishing pressure is high (Shackell et al., 2005). Rapid population-level declines have been observed in demersal fish on the Scotian Shelf (Reuchlin-Hugenholtz, Shackell, & Hutchings, 2015) and cod in Newfoundland (Hutchings, 1996). Notably, one of the persistent high abundance areas is adjacent to US waters, where lack of recovery in halibut since the 1800s is evidence for local overfishing (Shackell et al., 2016;Seitz et al., 2016).
Tagging data provides evidence of connectivity between Canada and United States with almost 30% of the recaptures of halibut tagged in New England occurring in Canadian waters (Kanwit, 2007).
However, in general, tagging studies suggest that a high proportion of halibut are resident or return seasonally to particular locations, with the majority of halibut captured within 200 km of where they were released, and a small proportion moving large distances across the management unit and beyond (den Heyer et al., 2012;Stobo, Neilson, & Simpson, 1988  ecosystem shifted from large-bodied predators to crustaceans, largely due to overexploitation . The observed decline in groundfish abundance Shackell & Frank, 2007) and body size (Shackell et al., 2010) was followed by a large increase in benthic decapods and other prey species, likely because of predation release (Boudreau & Worm, 2010;Steneck, Vavrinec, & Leland, 2004;Worm & Myers, 2003). In recent years in the NW Atlantic, there have been some signs of groundfish population growth, namely in haddock (Melanogrammus aeglefinus; DFO, 2012), cod (Cadigan, 2016;Rose & Rowe, 2015), and halibut (DFO, 2015a(DFO, , 2015bTrzcinski & Bowen, 2016).
We propose that the persistent areas of high juvenile abundance on the Scotian Shelf were offered protection from commercial fishing, and trawl gear, and that this protection has contributed to the rebounding of this stock.  (Hutchings, 1996), do not spawn in groups (Trumble et al., 1993) which has likely also served as a refuge from fishing pressure.
Persistent communities of juvenile halibut have remained in SWNS and the Gully (Figures 3 and 4), suggesting that these areas are core high abundance refugia and density-dependent habitat selection is occurring. As these preferred high abundance areas becomes increasingly occupied, resulting in limited resources, halibut may begin to occupy less ideal habitat in order to reduce intraspecific competition (Fisher & Frank, 2004;Gaston, 2003). Density-dependent habitat selection is generally assumed to be associated with preferred habitats that are rich in prey. Juvenile halibut up to 30 cm in length feed almost exclusively on invertebrates, those 30 to 80 cm in length feed on both invertebrates and fish, while halibut larger than 80 cm in length feed almost exclusively on fish (Kohler, 1967). It has yet to be examined how juvenile halibut abundance and distribution have covaried with their preferred prey species as the NW Atlantic's ecosystem shifted. Building on the analyses presented here, a closer examination of regions with persistent high juvenile abundance or historically high halibut abundance, such as southeastern Newfoundland (Figure 4), could find evidence of local overfishing, differing demographic rates between areas of high abundance, or other finer scale spatiotemporal dynamics of prey.

| CONCLUSION
Using a statistically powerful tool and more than three decades of standardized groundfish trawl survey data from Cape Cod to the Grand Banks, we identified areas of high juvenile halibut abundance that have persisted through three fisheries management regimes. Other highdensity areas were diminished and, as of yet, have not re-established.
As more research is needed to define subpopulations, and to prevent serial local overfishing in the future, we propose a pragmatic interim approach; one that focuses on protecting important patches where density of fish is persistently high (i.e., high abundance areas), indicating preferred or ideal resources such as prey and a refuge from predators.
These areas could also serve to be refuges from fishing pressure, both naturally, such as a deep water canyon (i.e., the Gully), or intentionally as part of a fisheries management strategy to protect juveniles. Atlantic halibut in Canadian waters are experiencing a period of high recruitment and population growth, a relatively unique and positive trend for a commercial fish in the depleted NW Atlantic ecosystem, and there is an opportunity to manage them in a precautionary way. The history of serial depletion of Atlantic halibut underscores the need to consider stock structure in its management. Given the ecological consequences of serial overfishing, local concentrations of halibut should be safe-guarded until it is known whether the evident spatial structure represents distinct or connected populations. We argue that areas of persistently high abundance of juveniles should receive some protection in order to sustain the recovered population(s) today and in the future.