Identifying the location and importance of spawning sites of Western Baltic herring using a particle backtracking model

Abstract

The recruitment success of some herring stocks fluctuates strongly, and apparently, success is often already determined during the early life stages, i.e. before metamorphosis. In studying the survival of early life stages and its affecting factors, particularly those during the egg stage, it is crucial to examine the processes at the spawning sites, which often cannot be explored directly. A recent decline in the recruitment of Western Baltic spring-spawning herring (WBSSH) increases the urgency of filling the knowledge gap for this stock, especially because one bottleneck in the recruitment seems to occur before hatching. We examined the successful 2003–2009 spawning sites of WBSSH in the main spawning ground, the Greifswalder Bodden lagoon. Instead of using common techniques such as diving or underwater videography, which are usually unsuitable for mapping large areas, we applied a model approach. We tracked herring larvae at length 6–10 mm, recorded by larval surveys during March–June of the respective years, back to their hatching sites using a Lagrangian particle backtracking model. We compared the spawning areas identified by the model with the results of earlier field studies; however, we also analysed variations between years, larval length groups, and different applied growth models, which are needed to define hatch-dates. Although spawning sites could not be identified with high precision because of the strong diffusion in the area studied, results indicate that larvae up to 10 mm length are caught near their hatching sites. However, the location of successful spawning sites varied largely between years, with the main hatching sites situated in the Strelasund and the eastern entrance of the lagoon. This may reflect variations in spawning-site selection or quality. A better knowledge of the locations and relative importance of, and the processes occurring on, the different spawning sites will provide an important contribution to the sustainable management of this commercially valuable herring stock.

backtracking, Baltic Sea, Clupea harengus, hatching sites, lagoon, larval transport, spring spawning

Introduction

The identification of spawning sites of exploited fish stocks is particularly important for recruitment and conservation in fisheries management. This knowledge can be used to protect sensitive habitats (Hammer et al., 2009), establish marine protected areas, or assess recruitment by conducting larvae and egg surveys in relevant areas. Spawning grounds, especially for migratory fish species, are usually assumed to have relatively fixed locations. Spawning-site fidelity has been discussed for Atlantic and Pacific herring and is considered an important component in the structuring of herring populations (Sinclair and Tremblay, 1984; Wheeler and Winters, 1984; Blaxter, 1985; Flostrand et al., 2009). However, some long-term studies contradict these hypotheses, revealing a significant variation in the spawning grounds utilized [e.g. Dickey-Collas et al. (2001) for the Irish Sea; Munk and Christensen (1990) and Schmidt et al. (2009) for the North Sea; and Hay et al. (2009) off British Columbia]. Identifying spawning sites of migratory species, such as herring, and their relative importance remains a challenge. Various methods are used, although each has its limitations. For demersal spawning herring, common techniques include diving and underwater videography (Scabell, 1988; Aneer, 1989; Kääriä et al., 1997; Hammer et al., 2009), which are unsuitable for mapping large areas. Other methods include remote sensing of spawning substrata, such as macrophytes, and the incorporation of other spawning-site characteristics (Hammer et al., 2009; Näslund et al., 2011). However, because macrophytes might persist only during the spawning season or for a limited number of years, spawning-site selection can have significant intra- and interannual variations, limiting the representativeness of such studies. Recently, particle backtracking has gained attention for its ability to identify spawning sites of fish stocks (Christensen et al., 2007, 2008). In such an application, larvae are tracked back in simulated flowfields as parameterized drifters, from areas of known occurrence or catch. Although this method allows rapid examination of the location of successful multiyear spawning sites, the precision of estimating locations may be limited. Contrary to expectation, dispersed larval distributions cannot be converged to a common origin, because dispersive and stochastic processes must be considered in the context of the model simulations. In this regard, Christensen et al. (2007) distinguished between an ideal “inverse time simulation” and a “reverse time simulation”, where dispersion is taken into account and hatching sites were thus identified as more or less accurate spatial probability distributions. Despite this limitation, reverse time simulations helped locate spawning sites of sandeels in the North Sea and are probably also useful in cases where larvae are released into a steady flow regime. Their applicability, however, remains to be tested in diffusion-dominated circulation systems.

The Greifswalder Bodden (GWB), a shallow lagoon of the western Baltic Sea south of Rügen Island, is an example of such an environment (Figure 1; total surface area of GWB: 514 km²; Schiewer, 2008), i.e. drift and circulation patterns can change rapidly as a result of frequent changes in prevailing wind conditions, particularly in its direction (Bauer et al., 2013). This lagoon is further considered to represent the main spawning ground of Western Baltic spring-spawning herring (WBSSH) (Jönsson and Biester, 1981; Biester, 1989) and therefore has been monitored continuously during the entire spawning period (March–June) since 1977 in the framework of the Rügen Herring Larvae Survey (RHLS; Oeberst et al., 2009a). This assumption is supported by the fact that estimated annual abundances of 20 mm larvae in the GWB, obtained from RHLS, correlate well with later recruitment estimates, such as the abundance of age 1 and 2 herring, derived from acoustic surveys, and VPA-derived estimates of age 0 recruitment (Oeberst et al., 2009a). The time-series of the index calculated from the survey, however, reveals strong fluctuations in the recruitment of WBSSH. In the period 2004–2008, a continuous decline was recorded, with an average annual decline of 15–35%, the causes of which remain unknown (ICES, 2011). In contrast to the abundance and distribution of larvae, little attention has been paid to the processes taking place on spawning sites, specifically the mortality of eggs and related effects on WBSSH recruitment. As one bottleneck in the recruitment appears to occur before hatching (Polte et al., 2013), such an investigation appears relevant but requires a detailed identification of the location of spawning sites. Potential macrophyte spawning grounds are present along almost all edges of the pan-shaped GWB lagoon but are not used uniformly for depositing eggs (Figure 1; Hammer et al., 2009). It is currently unclear why certain macrophyte assemblages are used as spawning grounds and others are not. The specific location of the macrophyte beds may be decisive for egg survival because of, inter alia, current and wave-surge exposure (e.g. oxygen supply or mechanical destruction). The latest studies of WBSSH spawning sites were conducted by Scabell (1988) and Hammer et al., (2009); however, they provided no information on potential intra- and interannual variations in spawning habitat abundance and utilization. Such variations are likely to exist for WBSSH because the spatial distribution of early larval stages in GWB varies significantly within and between years. By applying a particle backtracking model, we therefore focus on examining (i) the location of successful spawning sites of WBSSH during 2003–2009, and (ii) spawning-site fidelity during this period. We compare the results with previous field investigations and discuss the suitability of particle backtracking modelling to perform these tasks.

Figure 1.

GWB area with circulation model domains (upper right corner; BSM, Baltic Sea model; WBSM, western Baltic Sea model; GWBM, Greifswalder Bodden model), a detailed view of the GWBM, RHLS stations (blue dots), and strata (I–V, KB, Kubitzer Bodden; B, Baltic Sea), reported WBSSH spawning sites by Scabell (1988; green and red) and macrophyte coverage (light green), mapped by aerial photography in 2009, adapted from Hammer et al. (2009). Map adapted from Bauer et al. (2013).

Material and methods

We examined hatching sites of all WBSSH larvae of size classes 6–10 mm sampled weekly at up to 35 stations during RHLS between March and June in the period 2003–2009. For this purpose, we applied an offline Lagrangian particle tracking model to track simulated larvae back to their hatching sites.

Particle backtracking model

To simulate larval dispersal, it is crucial to consider larval behaviour, because vertical migrations can affect the speed and direction of larval drift. However, information on specific diel vertical migration patterns of WBSSH larvae in the GWB is lacking, and cannot be transferred easily from other study areas because they differ considerably between herring stocks and regions, and are further likely to be size-specific (Schnack, 1972; Johannessen, 1986; Munk et al., 1989). Recent field investigations in the GWB indicate that early stages of WBSSH larvae exhibit a near-surface distribution (P. Polte, pers. comm.) that is in agreement with observations from other shallow-spawning herring stocks (Stevenson, 1962; Johannessen, 1986). We therefore assumed that larvae are released immediately to the water column after hatching and remain in the upper 6 m. This depth approximately equals the average depth of the GWB and also the lower depth limit of its macrophyte coverage (Scabell, 1988; Hammer et al., 2009). We disregarded possible vertical migration patterns. Instead, we assumed larvae to encounter a depth-averaged flow caused by the pronounced vertical mixing in the lagoon. The offline particle backtracking model used (Gräwe and Wolff, 2010) was therefore forced by calculated hourly depth-averaged flowfields of the upper 0–6 m water column of the GWB area. Appropriate flowfields for model years 2003–2009 were obtained from a three-dimensional, triple-nested circulation model (Figure 1), where the innermost nested model domain covers the GWB area with a horizontal resolution of 180 m and 16 vertical layers. A more detailed description of the model set-up and validation is given in Bauer et al. (2013).

For every sampling record (287–353 per year), we released larvae—parameterized as Lagrangian drifters—in simulated flowfields at the time of sampling and at the grid point closest to the sampled survey station. We set the number of seeded particles constant to 100 000, regardless of the number of larvae caught during the specific sampling event, to achieve a sufficiently resolved picture of the larval dispersal from each release, and so the possible origin of larvae. Based on trial runs, this number seemed appropriate to the 35 039 total grid points of the model domain. Particle tracking results were stored as concentration fields with a temporal resolution of 1 h.

Growth model

We defined the tracking duration of larvae as the time (in hours) needed to reach the hatching sites. However, the specific age, and hence the hatch-dates, of sampled larvae was unknown. The tracking duration was therefore defined as a function of larval growth. Modelling larval growth is one of the most difficult tasks of advanced individual-based models (IBMs; Hauss and Peck, 2009). As larval growth rates exhibit high spatio-temporal variability and are size- and stock-specific, a detailed knowledge of growth-affecting factors (e.g. food density and size), their spatio-temporal variation, and larval physiology is required. However, for WBSSH larvae, this information is rare. Therefore, instead of applying a more detailed IBM, we used a comparative approach of different growth models. First, we used a temperature-dependent, mean-daily growth model for WBSSH larvae (Oeberst et al., 2009b) given as

$$G = 0.011 + 0.037\,T\quad (r^2 = 0.51).$$

(1)

Here, G is the growth rate in mm d⁻¹ and T the local temperature experienced by larvae. Second, we applied a constant-growth model with two different growth rates (0.2 and 0.3 mm d⁻¹). Based on additional model runs in which particle temperature exposure was recorded, we estimated temperature-dependent larval growth rates of different releases. Here, only one example particle was released, disregarding differences in the temperature exposure among particles, so reducing the computational effort. Growth rates, estimated in this way, can vary considerably during the spawning season, a known feature of herring larval ecology, reaching rates of 0.38–0.57 mm d⁻¹ at 10–15°C. Although these rates are still in the range of reported rates (Oeberst et al., 2009b), they might be too high for early larval stages. Busch et al. (1996) described growth rates of WBSSH yolk-sac larvae as ranging between 0.11 and 0.17 mm d⁻¹ in the early spawning season (March) and between 0.25 and 0.38 mm d⁻¹ in the later spawning season (May), which is in accordance with findings of other studies of spring-spawning herring larvae (Checkley, 1984; Oeberst et al., 2009b). For example, Henderson et al. (1984) described growth rates of 0.18 mm d⁻¹ for yolk-sac larvae and 0.43 mm d⁻¹ for post-yolk-sac larvae of Thames Estuary and Blackwater herring. In contrast, both constant-growth rates are relatively conservative estimates for fast-growing, spring-spawning herring larvae and are therefore used as comparative values.

Length- and time-dependent hatching probabilities

In addition to growth rates, hatch-lengths are necessary to estimate larvae hatch-dates. The latest study of the hatch-lengths of WBSSH larvae was conducted by Klinkhardt (1986) (Figure 2), which indicated that hatch-lengths can differ by up to 2 mm and follow a normal distribution (Kolmogorov–Smirnov test, p = 0.2875, D = 0.071).

Figure 2.

Hatch-length distribution of WBSSH larvae in April and May 1983, redrawn from Klinkhardt (1986). The solid line indicates fitted normal distribution.

To account for this variability, we did not consider the hatch-lengths, and hence the hatch-dates of larvae, as a fixed length or time but as normally distributed. We weighted the position of larvae at a specific time-step according to the probability that the larvae length at this time-step represented the actual hatch-length. To obtain the time-specific hatching probability, we fitted a normal distribution to the hatch-length distribution postulated by Klinkhardt (1986) and subsequently described it as a function of time (Figure 2). The median hatch-date was defined as the time at which larvae reached the median hatch-length |$t(\tilde L)$|⁠, whereas the hatch-date variance was further estimated by the length of minus sigma t (L_−σ) and plus sigma t (L_+σ); σ represents the standard deviation of the hatch-date distribution, giving the variance in time-steps as

$$\sigma ^2 = \displaystyle{{(t(L_{ - \sigma } ) - t(\tilde L))^2 + (t(L_{ + \sigma } ) - t(\tilde L))^2 } \over 2}.$$

(2)

This procedure was applied to every release and implies that the width of hatch-date distributions of the temperature-dependent growth model could differ significantly between releases, because the defined distribution estimates |$t(\tilde L)$|⁠, t (L_−σ), and t (L_+σ) were reached at different time-steps, depending on ambient water temperatures. Therefore, slow-growing cohorts (releases) demonstrated a broader hatch-date distribution than fast-growing cohorts. In contrast, the distribution parameters (median and variance), and therefore distributions of all evaluated cohorts, were constant for both constant-growth models. From these probability distributions, we calculated hatching probabilities P_t for all time-steps t between the minimum and maximum (5.5 and 7.3 mm) described by Klinkhardt (1986), using the probability density function. We standardized these estimates by the sum of all calculated probabilities, so that the sum of all probabilities of incorporated time-steps was adjusted to 1.

Subsequently, we used time-dependent hatching probabilities as a weighting factor for the concentration fields. The hatching site H of larvae of a specific length group L, caught at the specific station S and time T, was defined as

$$H_{L,S,T} = \sum\limits_t C_{L,t} P_t ,$$

(3)

where C_L,t gives the concentration field of the particle distribution at time-step t, and P_t denotes the probability of the time-step to represent the hatch-date.

Particle spreading

Depending on the degree of dispersion of a larval patch, the number of particles can vary significantly between grid points. Typically, this degree increases with time. As a consequence, aggregated distributions, and therefore early positions, are more accurate and so are given greater weight than dispersed patches, which occurred at later time-steps. However, this poses a problem because both early and later particle positions may be equally likely or unlikely, according to the hatch-length distribution (Figure 2). To minimize this effect and to facilitate a comparison of aggregated and dispersed distributions, we log-transformed particle concentration fields, after increasing them by 1 to avoid negative results. As a result, the information of particle spreading was preserved, but the range of particle concentrations per grid point could not exceed values >11.52 [∼log(100 001)].

Finally, to obtain the total weighted contribution of all hatching sites per year H_Y, all hatching areas of a specific year estimated this way were summarized after being multiplied by the related number of larvae caught at the specific release station and time N_L,S,T, thus accounting for the success of different spawning sites:

$$H_Y = \hbox{ln(}H_{L,S,T} + 1\hbox{)}N_{L,S,T} .$$

From the results obtained, we examined intra- and interannual variations of estimated hatching sites, along with related larval drift rates and distances, and compared them with results of previous studies of WBSSH hatching sites conducted by Scabell (1988) and Hammer et al. (2009). The visualization of estimated hatching sites of larvae, however, requires a selection of modelled results. Therefore, the results presented in the following sections focus mainly on model year 2006 and 8 mm larvae. We chose the length of 8 mm because it is close to hatching but does not overlap the hatch-length distribution, and corresponding larval abundances appear less patchy than those of smaller length groups. For this length group, the relative importance of different spawning strata (Figure 1) was calculated, and potential links between the location of successful hatching sites and the N20 recruitment index were examined.

Sensitivity of the flowfield

Apart from the application of different growth models affecting tracking durations, the sensitivity of results obtained was tested relative to the flowfields utilized. In the first sensitivity experiment, we assumed that particles only experience the flowfield of the upper 3 m, not of the upper 6 m. This indicates the importance of the direct effect of wind. In a second set of sensitivity experiments, we investigated mesoscale effects of the flowfield on the particle spreading. This could be done because the applied circulation model used to generate the flowfields is able to resolve mesoscale eddies by virtue of its high resolution of 180 m. However, the generation of these eddies can be seen as a random process, and a single model run is only one realization of such a stochastic process. To account for such randomness, the initial conditions of the ocean model were perturbed, while still applying the same atmospheric and boundary forcing, resulting in different flowfields. By calculating three of these perturbed flowfields, we could quantify the effect of eddies on the transport and spreading of the particles. Sensitivity runs were conducted only for 2006.

Computational effort

In summary, we investigated 62–311 sampling events per model year, length group, and growth model, resulting in 21 093 evaluated release events of Lagrangian tracers and an additional 7031 separate releases to assess temperature-dependent growth rates. To reduce the related computational effort, model runs were conducted only once for all length groups caught at a specific sampling station and time. This required the simplified assumption that larval drifting characteristics do not differ for small larvae (≤10 mm). Therefore, we defined the tracking duration of every model run to last at least until larvae of 10.5 mm (the largest length group evaluated) had reached the lower hatch-length of 5.5 mm, regardless of the applied growth model. In this way, the number of Lagrangian model runs could be reduced significantly to only 2146 each.

Drift rates and drifted distances

To further improve our understanding of the wind-dominated larval dispersal in such semi-enclosed lagoons as the GWB, we examined drift rates of larvae and their linear distance to hatching sites. To obtain drift rates, we calculated drifted distances of particles and defined them as the length of drifted trajectories from their release stations (catch positions of larvae). For this purpose, the Lagrangian particle backtracking model was extended by a “drift-distance module”, recording the covered distance of particles per time-step. Thus, we derived hourly, weekly, and monthly larval drift rates, specified as the distance covered within the given interval, from each Lagrangian model run. In addition, we calculated linear distances of particles to their release positions (catching sites of larvae). To account for the non-uniform, cloud-like spreading of particles, we applied results obtained from model runs conducted under the Eulerian framework. We calculated linear distances of every grid point of the model domain to the particle release stations. The resulting distance matrix was weighted by the time-dependent spatial distribution of particles, giving the time-dependent distances of particles to their release sites.

Results

Larval abundances obtained from the narrow grid of survey stations reveal a strong intraannual variation in the amount and location of larvae caught (Figure 3, left panel). The spawning season of WBSSH is relatively short, with the peak of larval production commonly occurring within a few weeks. Especially small larvae (≤10 mm), which are the focus of this study, appear more patchy than evenly distributed. In 2006, two spawning peaks occurred, one during calendar weeks (CWs) 17–18 and one in CW 21 (Figure 3). High abundances of larvae were detected in the eastern and western parts of the lagoon, including the adjacent Strelasund.

Figure 3.

Calculated hatching sites in 2006 for 8 mm larvae, and related larval abundances (from survey data) recorded during RHLS in different CWs. Left panel: total abundance (number m⁻²) of WBSSH larvae at a given CW in 2006 for each survey station. Other panels: calculated hatching sites described for two different flowfields (0–3 and 0–6 m depth-integrated) and three different growth models (temperature-dependent growth, constant growth at 0.2 and 0.3 mm d⁻¹). The median temperature-dependent growth rates are given for each CW.

Results from backtracking simulations of larvae of different size classes found during several years in GWB and the Strelasund demonstrate that most of these larvae originate in this area and not in the surrounding Baltic Sea (Figures 4 and 5). Drift rates of larvae normally range between 0.13 and 0.45, 0.18 km h⁻¹ on average, and are generally below 1.5 km h⁻¹. Therefore, weekly drifted distances, the length of trajectories, account for an average of 28.1 km and a maximum of 65 km (Figure 6a). Despite the comparatively great distance drifted, larvae remain closer to their release stations, on average within a linear distance of 5.1 km after 1 week and 11 km after 1 month of backtracking (Figure 6b). The linear distance to release positions can be expressed as a saturation function, with 11 km representing the asymptote of average distances to catch positions. Therefore, larvae originate close to catch positions. However, owing to the proximity of many potential spawning sites, the specific hatching sites cannot easily be determined from the sampled larval distribution. On a larger scale however, particularly the Strelasund and the eastern part of the lagoon (with its transition to the Baltic Sea) appear to feature pronounced hatching sites. Simulations conducted for several spawning seasons (2003–2009) confirm these results, though they reveal that both areas are of varying importance (Figures 5 and 7). Further, the time-series reveal no trend in the location of successful spawning sites but a pronounced variability, which, however, seems unrelated to the observed recruitment decline of WBSSH (Figure 7).

Figure 4.

Calculated hatching sites in 2006 and annual larval abundances (from survey data) of five different size classes (6–10 mm larvae). Left panel: total abundance (number m⁻²) of WBSSH larvae in 2006 for each survey station and length group. Other panels: calculated hatching sites described for three different growth models (temperature-dependent growth, constant growth at 0.2 and 0.3 mm d⁻¹).

Figure 5.

Calculated hatching sites of 8 mm herring larvae and related annual larval abundances (from survey data) during 2003–2009. Left panel: total abundance (number m⁻²) for each survey station. Other panels: calculated hatching sites described for three different growth models (temperature-dependent growth, constant growth at 0.2 and 0.3 mm d⁻¹).

Figure 6.

Box-and-whisker plots, showing the time-dependent drifted distance (a) and linear distance to catch positions (b) of larvae. The line indicates the linear relationship between absolute drifted distance and duration.

Figure 7.

Relative contribution of the seven spawning area strata defined in Figure 1 to the overall annual spawning success in the GWB area, measured as abundance of 8 mm herring larvae, 2003–2009 (sum of each column = 100%). Strata: KB, Kubitzer Bodden and adjacent Baltic Sea; I–V, RHLS survey strata (Oeberst et al., 2009a), comprising the Strelasund (I), Northwestern (II), Southwestern (III), Southeastern (IV), and Northeastern (V) GWB; B, Baltic Sea. Lower panel: N20 recruitment index as derived from RHLS and used in the ICES assessment (Oeberst et al., 2009a; ICES 2011).

Estimated hatching sites obtained from different growth models and length groups are in broad agreement, although tracking durations could differ significantly. Owing to the warming of the water, daily temperature-dependent growth rates rapidly exceed the two constant-growth rates within each spawning season. As a result of often shorter drift duration, estimated hatching sites are more confined when derived from temperature-dependent growth models than when obtained from constant-growth models. However, all applied growth models indicate that larvae could also originate in areas where spawning appears unlikely, particularly in the centre of the lagoon where the main spawning substratum, macrophytes, is missing (see Figure 1 for comparison). Effects of the applied flowfields (0–3 and 0–6 m) on the location of hatching sites are not substantial (Figure 3) and even imperceptible for perturbations in the mesoscale of the outer Baltic Sea (not shown).

Discussion

The larvae of WBSSH caught in the GWB area during weekly surveys were found to originate close to their catch positions. In this context, drifted distances (length of the drift pathway) appear remarkably great. This discrepancy can be attributed to the great wind-induced variability of flowfields in the area studied, described by Bauer et al. (2013), causing larvae to drift in an oscillating manner and thus be retained. The location of estimated hatching sites obtained from different growth models and flowfields (0–3; 0–6 m with or without perturbations in the mesoscale of the outer Baltic Sea) is in relatively good agreement, further highlighting the degree of larval retention and emphasizing that the position of larvae in the water column as well as the exact drift duration are less important. In contrast, the spatial heterogeneity of located hatching sites of different length groups appears more pronounced, particularly between length groups 6–8 and 9–10 mm. These differences may reveal uncertainties of the model but are more likely related to differences in the initial larval distribution, as local abundance peaks clearly diminish and larvae are distributed more homogeneously with increasing size. This in turn could be a consequence of spatially inhomogeneous and density-dependent larval mortality rates. It is likely that the differences in the larval distribution obtained from larval surveys result in a different weighting of identified hatching sites. As a result of the latter, hatching sites of different size classes may not necessarily overlap. However, it is clear that the estimate of hatching sites is also affected by larval dispersion as stated by Christensen et al. (2007), causing significant uncertainties, particularly for older and hence longer-drifted larvae. Therefore, back-calculated hatching sites appear more as diffusive clouds than precise locations. Accounting for the uncertainties mentioned above, the results indicate that, in some cases, spawning may also occur in the deeper central part of the lagoon. This appears unlikely as the area is not currently covered by macrophytes, the main spawning substratum, except a small group of boulders or dropstones, the “großer und kleiner Stubber” (Hammer et al., 2009); rather, the area is characterized by soft sediments (Katzung, 2004). However, owing to this exception and the sporadic availability of larger boulders (Figure 1), which can also serve as spawning substratum (Klinkhardt, 1996; Hammer et al., 2009), successful hatching in the central lagoon cannot be excluded entirely. In fact, fishers report spawning in deeper parts of the lagoon, because gillnets set in the area are often covered by spawn, a behaviour previously described by Scabell and Jönsson (1984). Still, the low probability of the occurrence of hatching sites in shallow water and small coves, estimated from particle backtracking experiments, is striking, because these areas are well covered by macrophytes and were previously identified as important spawning sites. This may reveal a methodological problem with the backtracking approach, because dispersion effects often act as a one-way street for drifting larvae, making it easy to leave but not to enter isolated or high-energy environments. The latter refers especially to shallow waters, which are much more reactive to changes in windforcing than the calmer and deeper central part, intensifying larval dispersion. Results of back- and forward-tracking experiments therefore must not lead to the same results (Christensen et al., 2007). Despite these constraints, the results give a rough idea of the location of spawning grounds in the GWB area. Generally, they highlight the importance of the western (Strelasund) and eastern entrances to the lagoon. Both of these areas are migration channels of adult prespawning herring to the inner GWB, of which the Strelasund is particularly of varying importance (Jönsson and Biester, 1981; Jönsson and Richter, 1993). Larval abundances in the Strelasund also demonstrate significant variability. Based on extensive tagging experiments in the 1980s, and in agreement with the model results presented here, Jönsson and Richter (1993) described that spawning takes place almost immediately after herring enter the lagoon by its eastern entrance. It is interesting that, there and within the western entrance (the Strelasund), currents are generally pronounced. WBSSH spawning-site selection may therefore be triggered not only by the availability of spawning substrata (macrophytes) but also by a certain exposure to currents, which are likely to provide sufficient oxygen and avoid egg sedimentation. Such a preference for spawning-site characteristics has been reported for other herring stocks (Haegele and Schweigert, 1985). Although migration patterns of WBSSH have been described as relatively constant over the years, changes in the locations of successful hatching sites are evident. This could be the result of spatially inhomogeneous mortality rates but could also indicate respective changes in spawning-bed utilization and thus contradict considerable spawning-site fidelity for WBSSH. Although this certainly remains a question of geographical scale (Hay et al., 2001), variations in hatching-site location could also reflect changes in spawning-site quality. Addressing this question appears to be of particular interest because it could help identify factors responsible for WBSSH recruitment fluctuation. It is therefore recommended to quantify spawning-site selection, e.g. as the amount of spawn per spawning site per year. Studies could be conducted on both lagoon inlets and focus on the emergence of immigrating spawners and later to larval abundances. The uncertainties mentioned above demonstrate the limitations of backtracking models, precisely “reverse time simulations”, as a tool for identifying spawning sites. It is not possible to determine well-defined hatching sites in diffusion-dominated circulation systems, even when using early larval stages. The complex geographical and hence hydrological structure of the lagoon studied further impedes this task. Another approach would be to conduct forward particle-tracking experiments from diverse spawning grounds and subsequent comparisons with the larval abundances measured. This may help solve some of the problems (Christensen et al., 2007), e.g. the one-way streets of shallow waters and coves found around the lagoon. The results further illustrate the importance of larval mortality in comparing hatching sites for several length groups. More advanced approaches should address these effects, particularly density-dependent and spatio-temporal changes in larval mortalities, as well as larval behaviour, because it can affect the precision of estimated larval dispersion patterns. Here, changes in the drifting characteristics during the larval development should be considered, especially buoyancy and vertical migration patterns (Schnack, 1972; Johannessen, 1986). In this context, the yolk-sac stage appears especially important, because it can last for a considerable time (4.4–11.6 d for WBSSH larvae; Busch et al., 1996). However, larval behaviour during this stage remains poorly understood, and studies indicate significant differences to later larval stages, with the vertical distribution of yolk-sac larvae being less variable (Johannessen, 1986), and more related to the seabed, owing to their relatively poor swimming performance and greater specific weight (Pilz, 1986). These factors could increase larval retention at the spawning sites and therefore significantly reduce the real “drifting period”, and thus larval dispersion. A better understanding of these processes will improve the backtracking of larvae considerably, especially of early stages (<8 mm), which overlap with the hatch-length distribution. In this context, changes in the hatch-length distribution, particularly within years, caused by differences in encountered incubation temperatures (Blaxter and Hempel, 1963; Busch et al., 1996), should also be considered because they can result in variations of more than 2 mm. Related inaccuracies are therefore likely to have more serious effects on the backtracking results than the precision of the growth model. Related to the suggestions mentioned above, the detailed data on larval abundances in the GWB provide a good opportunity for the further development, and thus improvement, of larval-particle (back-)tracking models. In the light of this, the approach presented here can be seen as another step towards a better understanding of herring larval ecology in shallow lagoons in the Baltic, after the significant insights gained by Bauer et al. (2013) on larval retention.

The results of this study also demonstrate that the importance of the location of successful hatching sites varies largely. Although the precision of determining the relative importance of spawning sites retrospectively within GWB seems to be sufficient, the results of this study cannot be used to predict their importance in future. Further, the results do not reveal a link between the geographical position of important spawning sites and spawning success in a specific year. It can be concluded therefore that protective measures should target not only a particular spawning site but rather the lagoon's entire shallow-water area, and so the vast majority of WBSSH spawning sites.

Acknowledgements

The research was partly financed by and conducted as part of the Fehmarn Belt Science Provision Project. Additional financial support by the Anker Stiftung, Dassow, is gratefully acknowledged. The work of UG was funded by the BMBF of Germany through grant number 01LR0807B.

Author notes

Handling editor: Claire Paris