Predator transitory spillover induces trophic cascades in ecological sinks

Abstract

Understanding the effects of cross-system fluxes is fundamental in ecosystem ecology and biological conservation. Source-sink dynamics and spillover processes may link adjacent ecosystems by movement of organisms across system boundaries. However, effects of temporal variability in these cross-system fluxes on a whole marine ecosystem structure have not yet been presented. Here we show, using 35 y of multitrophic data series from the Baltic Sea, that transitory spillover of the top-predator cod from its main distribution area produces cascading effects in the whole food web of an adjacent and semi-isolated ecosystem. At varying population size, cod expand/contract their distribution range and invade/retreat from the neighboring Gulf of Riga, thereby affecting the local prey population of herring and, indirectly, zooplankton and phytoplankton via top-down control. The Gulf of Riga can be considered for cod a “true sink” habitat, where in the absence of immigration from the source areas of the central Baltic Sea the cod population goes extinct due to the absence of suitable spawning grounds. Our results add a metaecosystem perspective to the ongoing intense scientific debate on the key role of top predators in structuring natural systems. The integration of regional and local processes is central to predict species and ecosystem responses to future climate changes and ongoing anthropogenic disturbances.

Ecological connectivity and spatiotemporal linkages among communities and ecosystems, as well as the interactions between regional and local processes, are fundamental aspects in ecology (1, 2) for the understanding of ecosystem functioning on a broader landscape (or metaecosystem) scale (3, 4). In particular, cross-habitat fluxes of organisms (mediated by passive transport or active migration) occur between natural habitats and across natural/anthropogenic systems, at various geographical scales (5–9). In heterogeneous landscapes, a particular case of cross-habitat fluxes is represented by source-sink dynamics in which a population from a productive source habitat maintains by migration populations in sink habitats where local reproduction is insufficient to compensate for mortality (10, 11). These processes are sometimes mediated by spillover, where an increase in population abundance in the source may result in a fast colonization of sink habitats (12, 13). The directional flow of organisms across systems boundaries has the potential to influence local food-web dynamics by coupling the trophic dynamics (e.g., predators and prey) of different habitats and ecosystems, as shown by theoretical and empirical studies (4, 14). In marine ecosystems, field examples of spillover effects have been presented mainly in terms of the effects that fishery no-take areas have on the external exploited fraction of the target populations (15, 16). Evidence of spillover impacts on the whole structure of pelagic marine ecosystems has until now not been documented in a natural source-sink dynamic context; with this study, we provide such an example linked to the management of marine resources.

Herein we applied a sequential modeling setup (17) using generalized additive models (GAMs) on a dataset of multiple trophic levels (from predatory fish to phytoplankton) and hydrological factors collected in the Baltic Sea during the past 35 y (Fig. 1 and Fig. S1). We show that the structure of a semi-isolated pelagic ecosystem [the Gulf of Riga (GoR); Fig. 1 and SI Materials and Methods] during summer is shaped by a series of top-down effects (i.e., a trophic cascade) (18), propagating from the neighboring Baltic Main Basin (MB). In the GoR the trophic cascade is modulated by the occasional occurrence of the main predatory fish of the Baltic Sea, the cod (Gadus morhua), which under specific conditions spill over from the MB.

Fig. 1.

Structural changes in the MB and GoR ecosystems during the past 35 y. (A) Changes in cod biomass and spatial distribution in the MB (source habitat for cod). (B) Changes in the food web of the GoR (sink habitat for cod), as indicated by time series of cod biomass index, herring abundance, zooplankton, and phytoplankton. The vertical dashed lines indicate the period of maximum cod population size and range of distribution in the MB that triggered the spillover into the GoR. The scale bar next to the distribution maps is in relative values.

Results and Discussion

The Baltic cod population rapidly increased in the late 1970s to record levels, due to a combination of particularly favorable hydrological conditions for reproduction, and relatively low levels of exploitation (19, 20). At the same time, the cod expanded its distribution into the northern MB, and spilled over into adjacent areas such as the GoR (Fig. 1A). The appearance and rise of cod in the GoR is demonstrated by scientific survey data and documented by the start of a cod fishery as shown by a progressive increase of commercial catches (Fig. 1B and Fig. S2). Cod is not able to successfully reproduce in the GoR due to the low salinities; its occurrence in the GoR is hence mediated by active migration of juvenile and adult individuals, as well as by passive larval dispersal, from the MB (21). The MB acts therefore as a source of cod that, during periods of high abundances, expands its distribution area and colonizes the sink habitat of the GoR.

From the mid 1980s, the Baltic cod population rapidly decreased and eventually collapsed (Fig. 1A), owing to overfishing and adverse hydrological conditions for its reproduction (19, 20). Concomitant with the decline in population size, the cod distribution contracted back to the southern Baltic Sea, where the hydrological conditions were more suitable for spawning and reproduction (i.e., high salinity and oxygen). As a result, the cod quickly disappeared from the GoR, and the commercial catches in this area dropped rapidly to zero (Fig. 1B and Fig. S2). The cod biomass index in the GoR was almost fully explained in our model by cod biomass in the MB (deviance explained = 84.3%; Table 1, Fig. 2, and Fig. S3), providing quantitative evidence of the link between the two ecosystems at the top of the food web. Fishing mortality on the GoR cod has certainly accelerated its local decline when the immigration from the source area of the MB ceased.

Table 1.

Results of the GAMs for each trophic level of the GoR

GAMs	Predictors	GCV	r2 adjusted	Deviance explained, %	n	df	F	P	Difference deviance explained, %
Cod biomass, model A	Cod biomass MB					2.39	60.17	<0.0001	—
	Salinity, summer–autumn*
	Temperature, summer–autumn*
	Final model	0.398	0.83	84.30	35
Herring abundance, model B	Cod biomass					1.00	93.65	<0.0001	20.37
	Fishing pressure					1.00	8.61	0.007	2.48
	Zooplankton, spring†
	Temperature, spring†					1.00	5.14	0.031	1.40
	Final model	0.021	0.92	92.80	32
Herring size, model B1	Herring abundance					1.65	27.70	<0.0001	—
	Temperature, spring–summer‡
	Salinity, spring–summer‡
	Final model	0.015	0.64	65.20	36
Copepod biomass, model Ca	Herring abundance					2.08	8.34	<0.001	—
	Chl. a biomass, summer
	Temperature, summer
	Salinity, summer
	Final model	0.129	0.38	41.80	35
Cladoceran biomass, model Cb	Herring abundance					1.00	6.73	0.015	22.39
	Chl a biomass, summer
	Temperature, summer					1.85	4.12	0.023	27.69
	Salinity, summer					2.83	5.37	0.005	45.47
	Final model	1.458	0.50	58.50	35
Chl a biomass, model D	Copepod biomass, summer					1.00	7.33	0.011	30.14
	Cladoceran biomass, summer
	River runoff§					1.00	5.75	0.023	27.75
	Temperature, summer					1.00	6.02	0.020	28.95
	Final model	0.098	0.36	41.80	33

Difference deviance explained indicates the deviance contribution of the predictors in each model, which was estimated based on the percent difference in explained deviance of the final models after deletion of one predictor at a time while keeping all the rest (i.e. with replacement; ref. 22). GCV, Generalized Cross Validation. Fig. S3 shows the partial effects of the predictors in each model.

*GoR hydrological conditions after cod spawning in the MB.

^†Copepod Eurytemora affinis, main prey for herring larvae, and temperature are taken during the herring spawning period.

^‡Hydrological conditions during the main herring growth seasons.

^§River runoff and nutrients concentration (dissolved inorganic nitrogen and dissolved inorganic phosphorus, both separated and summated) were separately tested as bottom-up forces, the second being not significant.

Fig. 2.

Schematic representation of the effects of MB cod through the food web of the GoR. The thickness of the arrows indicates the relative deviance contribution of the predictors in each model (i.e., on each trophic level A–D and on herring body size B1). Red, top-down effects; blue, bottom-up and hydrological effects. The direction of the interactions is also indicated. Table 1 shows the statistics of the models, and Fig. S3 the partial effects of the predictors in each model. Cladoceran model is not shown because of the lack of relation with phytoplankton.

The effects of the MB cod spillover and contraction propagated down the GoR food web, as suggested by the sequential modeling of each trophic level (Table 1, Fig. 2, and Fig. S3). The invasion and successive disappearance of the top-predator cod in the GoR was paralleled by a twofold decrease and then dramatic increase in the population of its main pelagic prey (i.e., herring; Fig. 1B). The variations in herring abundance were explained in our models mainly by changes in the cod biomass index, although fishing and spring temperature also have a significant effect (Table 1, Fig. 2, and Fig. S3)—the latter likely through enhanced recruitment (23). The relatively low herring population at the beginning of the 1970s was likely due to high fishing pressure (Fig. S2). The enduring high herring population from the early 1990s, which resulted despite an increase in fishing pressure, can instead be attributable to predation release from cod intertwined with an increase in spring water temperature of nearly 2 °C (Fig. S1).

Herring is the major zooplanktivore in the GoR ecosystem (SI Materials and Methods), and its population size was inversely correlated to the summer biomass of both copepods and cladocerans (Table 1 and Figs. 1B and 2). In addition to quantitative evidence of top-down regulation on zooplankton, temperature and salinity were also significant predictors for the cladocerans in our models (Table 1 and Fig. S3), confirming the importance of hydrological variability on zooplankton dynamics (24). Mean body size (weight at age 3) of herring was also negatively related with herring abundance, evidencing a strong density-dependent response to the changes in herring population size (Table 1, Fig. 2, and Fig. S3). Increased feeding competition within the enlarged herring population has likely been the mechanism mediating the depletion of zooplankton resources and the decrease in herring body size during the past three decades (Fig. 1B).

Top-down control was even detectable at the phytoplankton level [indicated by chlorophyll a (Chl a)], although its strength was minor compared with the higher trophic levels. This attenuation of top-down control at lower trophic levels conforms to what was found in other field and experimental studies (25). Though a significant negative effect of copepods on phytoplankton was unveiled in our model (26), temperature and river Daugava runoff were also important drivers (Table 1, Fig. 2, and Fig. S3), the latter likely subsidizing the semienclosed GoR ecosystem with nutrients (27). River runoff in GoR is eventually driven by large-scale climatic processes acting on precipitation and ice dynamics (28). The impact of top-down forcing on the summer phytoplankton biomass was also indirectly evidenced by the Secchi depth, an indicator of water transparency, which covaried positively with the copepods (Fig. 1B, Table S1, and Fig. S3).

The propagation of predator spillover and contraction effects through the GoR ecosystem was supported by the detection of indirect top-down regulation on trophic levels two or three steps below the predator. Specifically, cod biomass in the MB was a major predictor of herring abundance (negative effect), zooplankton biomass, herring body size, Secchi depth (positive effects), and phytoplankton biomass (negative effect) in the GoR (Table S2 and Figs. S4 and S5). These relationships unveil the occurrence of indirect mutualisms between trophic levels and reinforces the evidence of trophic cascades propagating across systems mediated by changes of cod distribution in the Baltic Sea landscape.

The dynamic between the MB and the GoR may represent an empirical example of a temporally heterogeneous source-sink dynamic (29), whose existence and persistence is regulated by the size and geographical distribution of the top-predator population in the source area. Cod fishery and habitat conditions, in terms of salinity and oxygen levels in the source habitat, are likely the major factors affecting the connectivity between the two systems, by limiting cod population size and distribution range (19). The GoR can be considered for cod as a true sink (29) or absolute sink (30), where in the absence of immigration from the source areas of the MB the cod population is not able to persist due to the lack of suitable spawning habitats. The linkage between the two systems is mainly mediated by active dispersal of juveniles and adult fish into the GoR when cod population size in the MB is high. In this case the GoR acts as nursery area for juveniles (31) and feeding area for both juveniles and adults of cod (21). During periods of high cod population size, its main pelagic food resource in the MB (i.e., the fish sprat) is depressed (17), and thus cod may search for alternative foraging habitats, as the GoR. This response would conform to the density-dependent ideal free-distribution theory (32), which suggests that as population size increases, individuals spread into less-favorable habitats. Also under these circumstances, however, some back-migration of cod individuals from the GoR to the MB also occurs, especially related to spawning (21), potentially stabilizing the source population and hence the source-sink dynamic (10).

According to our study, the potential for cod to reestablish in the GoR is related to processes operating in the MB. If the cod population in the MB recovers to high levels, its distribution range may expand northward, and potentially reinvade the GoR. For this to occur, enduring low fishing pressure and favorable hydrological conditions for cod in the source habitat are necessary. Under this potential scenario, and to aid cod persistence in the GoR and other marginal habitats, fishing mortality should be constrained at the local scale to remain well below the influx rate from the MB.

Resolving how and under which circumstances local systems respond to in situ processes vs. external forcing is crucial for understanding ecological dynamics in a metaecosystem context (1, 3). We have shown that extreme variations in a heavily exploited predator population, coupled with large variations in its distribution range, have cascading repercussions not only on its main distribution area (17, 33, 34) but also in adjacent systems, which usually are not under the trophic control of the top predator (for examples in terrestrial systems, see refs. 12 and 13). Therefore, high cod population levels can, through transitory spillover, link food webs that would otherwise be functionally nearly isolated. This finding provides empirical evidence that dispersal as a regional process may act to alter local predator–prey dynamics, prey competitive interactions, and food web structure (2). Our study also adds a metaecosystem perspective to the debate on the key role of top predators in structuring marine ecosystems (35).

Cross-system fluxes may play an important role in linking food-web dynamics across natural habitats, and investigations of their variations are crucial to forecast species’ and ecosystems’ responses to climate changes, fisheries, and eutrophication. The integration of landscape and food-web ecology is therefore central in the management of exploited resources and in ecosystem conservation.

Materials and Methods

Data on Cod Population from the Main Basin.

Time series of cod biomass (ages 2+) in the Baltic MB at the start of the year was calculated with an extended survivors analysis (XSA) using commercial landings and scientific surveys from the MB. This is the standard methodology used within the International Council for the Exploration of the Sea (ICES) stock assessment framework (19). Maps of cod distribution were created using standardized catch per unit effort (catch per hour of trawling) data collected during the Bottom International Trawl Survey (36) and previous research surveys. Three distribution maps were produced, corresponding respectively to the periods before, during, and after the high cod population of the early 1980s.

Data from the Gulf of Riga.

Data on cod commercial catches in the GoR were obtained from the former Latvian Fish Resources Agency. A fishery-independent index of cod biomass (catch per hour) in the GoR was calculated using the eelpout survey performed in the GoR by the same agency since 1974. Herring biomass and abundance at the start of the year were calculated using XSA stock assessment models (19). Time series of herring mean weight at age 3 was available from ICES (19). Weight at age 3 was used because it is the first age-class fully reproductive. Herring fishing pressure was estimated using the ratio of commercial catches to biomass (37). The zooplankton species used include the main copepods Eurytemora affinis and Acartia spp., and the indigenous cladocerans Bosmina longispina maritima, Evadne nordmanni, and Podon spp. These species constitute ∼90% of the zooplankton consumed by herring in the GoR (38). The invasive cladoceran Cercopagis pengoi, occurring in our samples since 1997, was not included because of its potential multiple effects on different trophic levels, which may confound the analyses and make the interpretation of the results difficult: C. pengoi is prey for herring and a predator of other zooplankton species, which are in turn prey for herring and predators of phytoplankton (39, 40). However, the structural changes in the GoR ecosystem started before the appearance of this plankter in our time series. Zooplankton data in the GoR were provided by the former Latvian Fish Resources Agency. Details on zooplankton sampling procedure and identification can be found in a previous work (41). Water temperature (°C) and salinity [practical salinity units (psu)] were averaged over the 0 to 50- and 0 to 20-m depth. Chl a (mg/m³) was averaged over a 0 to 20-m depth and nutrients (mmol/m³) over a 0 to 50-m depth. Runoff (m³/s) from the Latvian main river Daugava was used as proxy for the nutrient load to the GoR (27).

Statistical Analysis.

To analyze the effect of different predictors on each trophic level (response) of the GoR, we used GAMs (42). GAMs offer the main advantage of being able to model both linearity and nonlinearity between responses and predictors. The following additive formulation was used in the full models:

where a is the intercept, s the thin-plate smoothing spline function (43), V_{i . . .} V_n the predictors, and ε the random error.

The trophic levels considered were cod biomass index, herring abundance, and zooplankton and phytoplankton biomass (as indexed by Chl a) in GoR. We also analyzed the effects of the predictors on herring mean body size (weight at age 3) as proxy for individual growth rate. The predictors for each trophic level were selected based on acknowledged ecological and physiological mechanisms, and typically included top-down (predation driven), bottom-up (resource driven), and hydrological forcing (Table S3).

In the herring body size and zooplankton (prey of herring and thus influencing its growth) models, herring abundance, rather than biomass, was used as top-down predictor because biomass contains, by definition, a growth signal, making response (weight at age) and predictor (biomass) not independent a priori. All variables were log-transformed before analyses.

In the GAM modeling, we retained only the predictors that were statistically significant (final model) using a backward stepwise procedure. From the full model, the nonsignificant predictor with the lowest significance level was excluded at each step and the model run again. This procedure was repeated until all of the predictors were significant. We limited the maximum degrees of freedom acceptable for each term to k = 4. A Gaussian distribution with an identity function was used in almost all of the models because of the normal distribution of the response variables. For the cod biomass index, a quasi-Poisson distribution was used because of its larger flexibility in fitting the data distributions, especially for zero-inflated data. For the herring weight at age model, a gamma distribution was used. We calculated the deviance explained by the final models and the deviance contribution of each predictor selected in the final models. The deviance contribution of the predictors was estimated based on the percent difference in explained deviance of the final models after deletion of one predictor at a time while keeping all the rest (i.e. with replacement) (22). This procedure should give a good estimate of the explanatory power of each covariate as long as the estimated degrees of freedoms of the remaining covariates do not change largely, as was the case. The partial effects of the predictors were linear or quasilinear in all of the models (Figs. S3 and S5), and therefore the sign of the relationships was also provided. Model residuals were inspected using graphical methods (44).

For the lower trophic levels (i.e., plankton), the values of the predictors were taken the same year of the response variables, because we assumed that the high turnover rates of zooplankton and phytoplankton would make them respond promptly (the same year) to changes in the environment (17). For herring population size, the use of a different approach was needed due to the co-occurrence in the population of several cohorts persisting in time (years). Therefore, we constructed the predictors’ time series to represent their potential influence on herring population over the period of existence of a cohort. Because the herring population consists mainly of eight age classes (from 1 to 7, plus the 8+ constituted by all of the ages older than 7 y pooled together) (19), herring abundance at time t is the result of forces that have acted over the previous 8 y (i.e., at years t − 1, t − 2, … , t − 8). However, because in average ∼90% of the herring population is constituted by fish 1–4 y old (19), we used predictors values only from t − 1 to t − 4 in this analysis, which decreased the influence of age classes scarcely represented in the herring population and the loss of too many data points for the analysis. The response (herring abundance) at time t, thus, were related to the 4-y mean (at time t − 1, … , t − 4) of each predictor. The same approach was used in the investigation of temperature effects on Baltic sprat landings 3 y after (45), and in the study of the main factors driving sprat population size in the Baltic Sea (17).

The statistical analyses were performed using the mgcv library of R 2.12.0 (http://www.r-project.org/). The significance level was set to α = 0.05 for all tests.