Zooplankton data and sampling areas

The data originate from national Finnish, German, Latvian, Lithuanian and Swedish monitoring programs in the Baltic Sea (Table 1; Fig 1); all data are publicly available from several databases (Swedish Meteorological and Hydrological Institute, SHARK database: www.smhi.se; Baltic Sea mesozooplankton dataset: http://kodu.ut.ee/~riina82/; Data Center of the German Maritime and Hydrographic Agency: http://www.bsh.de/en/, and COPEPOD: http://www.st.nmfs.noaa.gov/copepod/). Sampling locations represent good geographic coverage for coastal and open sea areas in the Gulf of Bothnia, Northern, Central, Southeastern and Southern Baltic proper, and Gulfs of Finland and Riga. Due to considerable variation in sampling frequency between the monitoring programs, the data are restricted to the average values observed during the summer period (June-September) as the most represented in all datasets. This is also the period of the highest plankton productivity as well as predation pressure on zooplankton (S1 Fig) [28–30]. The length of the time series used in this analysis varied from 6 to 51 years (Fig 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Map of the Baltic Sea indicating sampling sites for zooplankton datasets used in this study.

Each dataset is represented by a single circle; when several stations contributed to a dataset, the circle shows the approximate middle of the sampled area. See Table 1 for description of sampling sites and sampling frequencies.

https://doi.org/10.1371/journal.pone.0158326.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Time coverage for zooplankton data and reference periods based on the existing EQR for Chlorophyll a (RefCon_Chl) and fish body condition (RefCon_Fish) for each dataset.

See Table 1 for details.

https://doi.org/10.1371/journal.pone.0158326.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Details for the data sets provided by national monitoring programs for indicator testing; deviations in sampling methods from the HELCOM guidelines (i.e., WP2, 100-μm mesh size) are indicated.

See S1 Table for details.

https://doi.org/10.1371/journal.pone.0158326.t001

The indicators

A set of putative indicators (Table 2) was selected based on existing literature and discussions within Zooplankton Expert Network (ZEN) supported by HELCOM and pelagic biodiversity group (HELCOM CORESET 1). The following metrics of zooplankton communities were evaluated as indicators of change in fish feeding conditions and food web properties caused by eutrophication.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Indicator description, calculation principles, and rationale.

Only species/groups that are included consistently in the zooplankton analysis were used for calculations.

https://doi.org/10.1371/journal.pone.0158326.t002

Approaches for defining reference periods and boundaries

A fundamental difficulty when using indicators is setting reference conditions. The reference condition can be based on existing reference areas or populations that are in a pristine state, historical records that date back to the time when anthropogenic pressures are considered as being low/absent, or a modeling using related variables with known reference condition to derive the reference state for the variable in question [61]. The main difficulty is the lack of sites that are not currently affected by human activities and data that date back to such reference periods; this holds true also for Baltic zooplankton. Therefore, it would not be feasible to follow this approach for establishing a reference condition for zooplankton as well as for many other ecological groups in the Baltic Sea.

Alternatively, a period within existing time series can be selected to define a reference state when the food web structure was not measurably affected by eutrophication or representing good fish feeding conditions. To define the reference conditions, existing GES for eutrophication-related variables and fish stocks may be applied. This, however, is complicated by the occurrence of sudden changes in the structure and function of the food web, i.e. regime shifts, that have been identified in the Baltic Sea system, including zooplankton [13,62], although we know little about proximate causes and timing of such shifts from areas other than the Central Baltic Sea. Some of these abrupt changes have been linked to eutrophication and fishing, whereas others were related to the climatic and hydrographic conditions. Since the GES boundaries should be in line with the prevailing physiographic conditions and climate, the existence of time periods with different stable states should be acknowledged in the selection of a reference period.

We evaluated two alternative strategies for setting reference conditions for the indicators tested. The first approach was to use a long-term average for an entire dataset and corresponding variance and to evaluate deviations from the variability boundaries; this approach is particularly relevant if the time series are very short. The second approach was based on basin-specific reference conditions for (1) chlorophyll a concentrations (RefCon_Chl) representing eutrophication state with no measurable effects on grazers in the food web, and (2) fish feeding conditions (RefCon_Fish) representing food web structure supporting adequate nutrition for zooplanktivorous fish. To define RefCon_Chl, we used existing assessment for eutrophication in the sub-basins of the Baltic Sea [63,64]. To define RefCon_Fish, we used data on growth and stocks of young herring and sprat to identify periods of good feeding conditions for zooplanktivorous fish in the relevant ICES subdivisions. Herring and sprat are dominant species both in the commercial fishery and as zooplanktivores in the Baltic Sea, playing a crucial role in the food web functioning across the sea basins [13,14] and depending on zooplankton availability, particularly during summer (S1 Fig). Therefore, any impact on food availability and population recruitment of these species would affect the ecosystem performance.

Detection of changes in time-series of indicators using control charts

The principles of process control are well-established in the area of production and operations management [65]. Process control makes use of control charts to determine if the underlying distribution of a measurable variable is undergoing a shift. A control chart uses information about the process variation to examine if the process is moving beyond the expected stochastic variability stated as desirable tolerance limits. If the process is in control, then subsequent observations lie within the limits. The hypothesis that the process is in control is rejected if the observations fall outside the limits. As a test statistic, control charts employ the controlling mean (μ) and specify control limits of n × standard deviations (σ) above and below the mean or the confidence intervals (CI). The baseline (or reference) conditions are represented by μ that can be defined for a selected period or the entire dataset. The time series of the selected metrics of zooplankton community structure were analyzed with combined Shewhart and cumulative sum (CuSum) control charts using SixSigma module in STATISTICA 8.0 (StatSoft, USA). The Shewhart control chart provides enhanced detection of sudden deviations, whereas CuSum methods detect persistent small changes in observed processes or periods when the long-term mean changes [65]. The control charts have been recommended as a tool to interpret environmental monitoring data and to detect abnormal deviations in time series [66,67], including fish [68,69] and zooplankton [70,71].

A factor to consider when interpreting control charts is the control limit values, which is a function of the variability of the data, and, thus, reflects the statistical power to detect a deviation from the baseline. In this study, the upper and lower control limits (UCL and LCL, respectively), were defined as either 99%-CIs around the mean values (for μ based on an entire dataset), or using a conservative approach of ±3σ and ±5σ for Shewhart and CuSum control limits, respectively (for μ based on either RefCon_Fish or RefCon_Chl) [67–69]. The determination of whether an indicator was beyond the expected limits was carried out over the evaluation period for each data set. Most of the datasets with >12 years of observation tested with Kolmogorov-Smirnov normality test were found to deviate significantly from the normal distribution; particularly, MMB and the community ratios. Therefore, the indicator values were Box-Cox transformed, and all downstream analyses were carried out on the transformed data; all z-scores were normally distributed (p > 0.2 in all cases). Missing values were predicted by Eigen-Vector Filtering [72].

For each indicator and dataset, once a controlling mean (μ_i) and standard deviation (σ_i) have been specified based on the chosen baseline period, indicator values (x_i,t) within the time-series were standardized to z-scores (z_i,t) as: (1) As standardized values, z-scores enable direct comparison of changes for different sites and variables, irrespective of their absolute values.

To implement our two approaches for setting reference conditions, we specified the μ_i and σ_i of the underlying normal distribution parameters for constructing the control charts. In the first approach, we used all data available (i.e., all years of the monitoring period, including the most recent year). In the second approach, a window of the data corresponding to the selected reference period (Fig 2) representing:

1. RefCon_Chl that was defined using a period with environmental quality ratio (EQR) >0.67 and historical data on chlorophyll-a [64,65]; this indicates in-GES state in the system, and
2. RefCon_Fish that was set using periods of successful foraging in the relevant ICES subdivisions, when both fish growth assessed as weight-at-age, WAA, or other condition indices; [23,73], and stocks were relatively high [14,74]. Recently, Ljunggren et al. [37] suggested that WAA could be used as a proxy for food availability to relate feeding conditions to fish recruitment in coastal areas of the northern and central Baltic Sea.

To investigate trends in accumulated changes for each indicator in question, a decision-interval CuSum (DI-CuSum) was calculated by recursively accumulating positive and negative deviations separately with two statistics: (2) for positive deviations (one-sided upper CuSum), and (3) for negative deviations (one-sided lower CuSum), with S_{i = 0} = 0 [74]. This scheme is particularly suitable for indicators showing either a positive or negative response and has, for example, been applied for analysis of cod stocks in the North Sea [69]. The k value is the allowance value in the process expressed in z units of the mean shift one wishes to detect, i.e., deviations smaller than k are ignored. The default choice of k = 0.5 was applied here, which is considered appropriate for detecting a 1-σ shift in the process mean [75].

Detecting monotonous trends and shifts in zooplankton community structure

For the analysis of trends and sudden shifts, we used z-scores calculated for the entire data period. For each indicator, the non-parametric Mann—Kendall test for a monotonic downward or upward trend was applied. Chronological clustering was used to identify homogenous time intervals for zooplankton community structure in each dataset. Chronological clustering produces groups of sequential years, defined by connectedness and a clustering sensitivity parameter α. Identification of breakpoints is usually investigated using different α values for the same connectedness level [76]. To detect sudden shifts in our indicator time series, we used z-scores calculated for the entire data period and the Euclidean distance function to calculate the dis(similarity) between years using software package Brodgar (www.brodgar.com). The main breakpoints in the time series were calculated using α values 0.01, 0.05 and 0.1 and a constant connectedness level set at 50%.

Logistic regressions linking indicators to the reference periods

Standard binomial regression (logistic regression) was used to evaluate a binary response variable (1 when in-control vs. 0 when out-of-control years) for RefCon_Fish and RefCon_Chl as a function of the indicators, utilizing a minimum number of variables. The indicator values were calculated using a baseline for the entire observation period (i.e., using the first approach). The reference and the out-of-control years were selected based on the indicator behavior with RefCon_Fish and RefCon_Chl baselines (i.e., using the second approach), so that if any of the indicators exceed CuSum control limits for a given year and a given type of the reference state, it was assigned as an out-of-control year. Thus, these regression models described the probability of falling inside the RefCon_Fish or RefCon_Chl. In other words, we investigated what indicators were the most informative for predicting whether zooplankton community is within GES. As some indicators would be correlating due to the nature of their calculation, Pearson correlation analysis were used for z-scores (entire data period) to evaluate possible redundancy of the indicators as predictors. When selecting predictors, multicollinearity was explored using the correlation analysis results as well as regression diagnostics and Variance Inflation Factor (VIF scores; [77]). In the final models, none of the VIF were ≥ 3, which is well below the cut-off point of 10; thus, the models have not been degraded by collinearity. To identify whether for neighboring areas have similar behavior of the indicators in relation to RefCon_Fish or RefCon_Chl, we included dataset as a categorical variable in the regressions. When dataset was found non-significant, the resulting regressions were defined as applicable for more than one area. The scaled deviance and the Pearson χ² were used to evaluate the model fit, and the overall best model was determined using Akaike's Information Criteria (AIC). The AIC selects models with high likelihood while penalizing for additional parameters, such that the best model has the smallest AIC. When a more complicated model was not significantly different (p > 0.05) than a simpler model with a similar AIC value, the simpler model was chosen as the best model. Percentages of correct classification cases and odds ratio were used to access prediction accuracy; the classification cut-off was set to 0.5.

Variability of the indicator values

For all indicators, there was considerable variability across datasets (Fig 3), particularly for Cla/Cop and MMB. The observed median values for these indicators (non-transformed) spanned 64- and 41-fold range, respectively, with the highest values recorded in the BIOR datasets and the lowest in ÅlandFI and K32-41 datasets. The least variable median values were the percentages (CB% and MMB%), with maximal differences of 3- and 5-fold respectively) and CB, for which a ~5-fold difference was observed between the highest (BIOR) and the lowest (ÅlandFI) values.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Variability of zooplankton indicators in the analyzed datasets.

The datasets are indicated as Station. Box-and-whiskers show median, 25 and 75% percentiles, min and max values. Asterisks (*: p< 0.05, **: p < 0.01, and ***: p < 0.001) indicate significant deviations from Gaussian distribution using the Kolmogorov-Smirnov statistics applied to the data sets with ≥18 years of observations. Shaded columns indicate datasets that are <12 years and thus not eligible for normality testing. See Tables 1 and 3 for the details on the indicators and datasets.

https://doi.org/10.1371/journal.pone.0158326.g003

The Box-Cox transformation significantly decreased CV% values in most of the indicators (Wilcoxon signed rank test, p < 0.015; S2 Fig). For the transformed data, the highest variability (CV%>50%) was found for the ratios (RotCla/Cop and Cla/Cop), particularly in the short datasets (J56-K18 and K32-41), whereas the lowest CV% values (13–34%) were observed for TZA, MeanSize and MMB, particularly in the relatively long datasets (i.e., Landsort, BIOR and Bornholm; CV% < 20%). For most indicators, the CVs% estimated for the RefCon_Chl and RefCon_Fish periods were less variable, with MeanSize (<26% in all datasets) and TZB (<32% in all datasets) having the lowest variability (S2 Fig).

Control charts for indicators with baselines estimated for the entire data sets

TZA and TZB.

In all datasets, TZA and TZB occasionally exceeded both UCL and LCL (Table 3; S3A and S3B Fig). During the 1960–1980s, low TZA values with out-of-control periods of >5 years were detected in the Gulf of Bothnia and Åland Sea (BoBFI, BoSFI, and ÅlandFI) as well as in the offshore areas of the Baltic proper (EGB-BIOR and Landsort). These periods were followed by generally high TZA in 1990–2000s, with values often exceeding the CuSum-UCL in the same datasets. As a result, significant increasing trends for TZA (BoSFI, ÅlandFI and EGB-BIOR) and TZB (BoSFI and EGB-BIOR) were observed for the entire period (S3A Fig). The opposite TZA and TZB trends occurred in the Gulf of Riga (GoR-BIOR), with a significant overall decrease in both variables. A significant TZB decrease was also found for GoFFI, Askö, and Bornholm (S3A and S3B Fig). Moreover, periods of low out-of-control TZA and TZB values were observed for GoFFI and BIOR data around the middle of the last decade, although these deviations were less pronounced than in the 1980s (S3A and S3B Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Summary of CuSum analysis (S3 Fig) for all indicators and datasets with >12 years of observations.

Periods when UCL or LCL (bold) were violated for >3 consecutive years are shown; the first two digits are omitted for simplicity.

https://doi.org/10.1371/journal.pone.0158326.t003

Control charts with baselines estimated for RefCon_Fish

Control charts with baselines estimated for RefCon_Chl

Abrupt shifts in community structure revealed by chronological clustering

The earliest breakpoint was detected for EGB-BIOR in the late 1960s. The change was related to the upward shift in total zooplankton stocks, including both copepods and cladocerans, with prevalence of cladocerans (Table 3; S3A, S3B, S3C, S3D and S3H Fig). In the early 1980s, a breakpoint was identified in the Åland Sea, albeit only at α levels of 0.05 and 0.1 (Table 4). Similar to the EGB-BIOR dataset, the increased total abundance due to the increased stocks of cladocerans and rotifers and, consequently, declining percentage of copepod biomass was responsible for this shift in the ÅlandFI (S3A, S3D, S3G and S3H Fig). More profound and significant changes in community structure were detected during the mid-1990s in the northern Baltic proper and the Gulf of Finland, but also in the eastern Gotland basin and Bornholm when using α levels of 0.05 and 0.1 (Table 4). In the late 1990s, the breakpoint in the Bornholm was identified at all α levels. These changes were related to declining zooplankton stocks with concomitant changes in the community structure. The structural changes were, however, different between the offshore northern Baltic together with the Gulf of Finland, where cladoceran biomass declined, and coastal northern Baltic proper together with Bornholm basin, where the decline was mostly attributed to the copepod biomass (Table 3; S3C, S3D and S3H Fig). Finally, in the early 2000s, the less pronounced shifts in the Bothnian and Åland seas and in the Bornholm were identified with lower α levels (Table 4).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Shifts in zooplankton community structure (year) that were detected for different α levels using the indicator time series (only entire datasets with more than 12 years of observations were considered).

No community shifts were detected for the Bothnian Bay (BoBFI) data. Years that were consistently detected at all levels of α are in bold.

https://doi.org/10.1371/journal.pone.0158326.t004

Covariation among the indicators

There were significant correlations among the indicators, with substantial differences among the datasets (S2 Table, S4 Fig). Due to the nature of the indicator calculations, the strongest and ubiquitous correlations were observed for RotCla/Cop, Cla/Cop and CB%, which are the indicators reflecting proportions of copepods and cladocerans, the two major groups contributing to the community biomass. As cladocerans contribute heavily to MMB and MMB%, these indicators were also strongly correlating with the indicators reflecting proportions of cladocerans and copepods. TZA and TZB exhibited moderate to strong intercorrelations in all datasets except BoBFI, whereas Pearson r for MeanSize was most variable, with significant moderate to strong positive correlations observed for CB and/or CB% in BoBFI, BoSFI, Askö, and Bornholm datasets.

Linking indicators to the reference conditions

Indicators that were significantly associated with in-control zooplankton community state for RefCon_Chl and RefCon_Fish were identified by logistic regressions (Table 5). MeanSize was a significant predictor in >50% and 75% of the best-fit models for RefCon_Chl and RefCon_Fish, respectively. The best-fit models that did not include MeanSize frequently identified a combination of TZA and TZB as negative and positive predictors, respectively (Table 5, models 7, 12–13), implying a positive effect of MeanSize on the response variable. The latter was always observed in the alternative models, albeit with lower fit. The main difference between the sets of the models for RefCon_Chl and RefCon_Fish was that the latter consistently included variables describing total zooplankton stock size (TZA, TZB or both) as positive predictors, unlike the former that included variables related to community structure (percentages of main groups and/or ratios). Also, in RefCon_Fish models, in contrast to the RefCon_Chl models, both TZA and TZB effects were predominantly positive.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Winning logistic models for prediction of zooplankton community structure being in the reference state for RefCon_Chl and RefCon_Fish.

When equally strong models were found for the same dataset, their AIC values are provided. Only significant models are shown; significant effects are in bold. Correct classification percentage and odds ratio are used for model accuracy evaluation.

https://doi.org/10.1371/journal.pone.0158326.t005

The overall prediction accuracy of the models was 65–88% and 66–86% for the RefCon_Chl and RefCon_Fish models, respectively (Table 5; S5 Fig). The sensitivity, i.e., the proportion of cases (years) correctly identified by the model as being within the reference conditions, was similar between RefCon_Chl and RefCon_Fish models. By contrast, specificity, i.e., the proportion of cases correctly identified as being outside the reference conditions, was significantly higher in the RefCon_Fish models, with lower between-model variability (72–86%; Wilcoxon matched-pairs signed rank test; p < 0.004; S5 Fig).

High spatial variability of zooplankton and its indicators

Ecologists and managers have long recognized the challenges imposed by the inherent spatial complexity of aquatic communities for indicator development within MSFD. While studies on long-term zooplankton dynamics in some Baltic areas have received considerable attention [22, 28, 40, 59; 78], the temporal variability of zooplankton across the sea remains much less understood. The need for indicators that would be equally applicable in different areas highlights this concern. Our examination of various metrics reflecting zooplankton community dynamics revealed high variability among the basins of the Baltic Sea, typical for marine estuaries [35]. Total zooplankton stocks were highest in the Baltic proper and the adjacent Gulfs of Riga and Finland, mostly due to the higher contribution of cladocerans but also greater copepod stocks (Fig 3). The cross-Baltic variability of the mean zooplankter size reflects relative contribution of both large copepods (e.g., Limnocalanus macrurus in the Bay of Bothnia) and large size classess of cladocerans (e.g., Evadne nordmanni and Bosmina maritima in the Gulf of Riga and the eastern Gotland basin). These differences emphasize the need for non-taxonomic zooplankton indicators that would represent common features regarding the food web functioning in the pelagia, i.e., maintaining energy raceways from primary producers to higher trophic levels.

Trends and shifts in zooplankton community structure

Both monotonous trends (S3 Fig) and sudden shifts (Table 4) were detected during the past decades in virtually all datasets; moreover, these changes for specific indicators varied in their direction and timing among the datasets. In all areas, except the Bothnian Bay, chronological clustering identified shifts in zooplankton community structure and stock size, with the earliest shift observed in the late 1960s, and the most profound pan-Baltic changes occurring in mid- to late 1990s (Table 4); the latter is in agreement with the best documented regime shift in the central Baltic [79]. However, the taxonomic and structural changes underlying these shifts differed among the areas; as a result, most indicators showed both increases and decreases over time depending on the dataset. The most consistent trends were observed for MeanSize that significantly decreased in the northern Baltic proper and the adjacent areas (Åland Sea and the Gulf of Finland) as well as in the Bornholm basin (S3 Fig) due to decreasing stocks of larger copepods, such as Limnocalanus macrurus in the north and Pseudocalanus spp. in Bornholm basin, respectively [78]. The observed absolute decrease of the mean zooplankter body weight varied from 18% in the Bornholm to 57% in the Åland datasets. Such profound and consistent throughout the ecosystem decline in the body size of pelagic grazers and fish prey have strong implications for both grazing capacity of zooplankton community and fish feeding conditions.

Detecting out-of-control periods with control charts

The combination of Shewhart and CuSum control charts provides a useful tool in the analysis of both sudden deviations and persistent small changes in zooplankton metrics. In each data set, at least one indicator was found to cross control limits regardless whether the acceptable background variability of the indicator was based on the entire dataset or on the reference period only (Table 3). When the baseline variability was set based on the entire dataset, the indicators that violated their control limits most frequently were MeanSize, TZA, MMB%, MMB and the community ratios (Table 3). The datasets with the highest number of the indicators violating their control limits were GoFFI, Landsort, Bornholm and EGB-BIOR (Table 3). When reference periods based of the existing EQR for chlorophyll and fish condition were used in setting up the baseline variability, the number of the violations were fewer for most of the indicators compared to those identified when using the entire dataset variability. The most characteristic violations were observed for MeanSize. When outside any reference period, MeanSize crossed lower control limit in all datasets except BoSFI and GoR-BIOR; this was observed more frequently for the RefCon_Fish- than for RefCon_Chl-based evaluations (Table 3). Behavior of the indicators reflecting total zooplankton stock (TZA and TZB) differed among the areas, with values indicating suboptimal fish feeding conditions in the coastal areas of the western and eastern Baltic proper and the Bornholm basin. The differences were also apparent with regard to the eutrophication degree, with TZA, TZB and MeanSize values indicating increased total abundance or decreased biomass in combination with decreased body size in zooplankton in most of the datasets (Table 3). The behavior of indicators reflecting contribution of copepods, small-sized zooplankters and community ratios was also informative, albeit only for some datasets, depending on the area-specific community structure and natural prevalence of cladocerans or copepods (S3 Fig). Most of the violations occurred in 1990s, and, in many cases, the deviations from in-control state have escalated dramatically toward the end of the time series and never returned to the baseline levels, particularly in the RefCon_Fish-based evaluations (e.g., MeanSize, Cla/Cop, CB%, MMB%; Table 3). Notably, in the longest dataset (EGB-BIOR, ~50 years), the most pronounced period of sub-GES zooplankton state appeared to occur in the 1960s, when both TZA and TZB were exceptionally low, reflecting low stocks of copepods and cladocerans (S3 Fig). While we can only speculate about the driving forces behind the increase in zooplankton stocks during the late 1960s to 1970s, it is clear that this increase and the following period of high zooplankton stocks coincided with good feeding conditions for herring and sprat [38,79,80]. Finally, as with any other biological data, the uncertainty related to sampling and data analysis comparability over decades remains important for compiling long-term data sets. With regard to the EGB-BIOR data, one should keeep in mind that the sampling and analysis methods behind this dataset deviated most from the HELCOM guidelines that were followed more closely by other laboratiories (Table 1, S1 Table), which complicates the interpretation of the indicator trends.

Combining indicators to predict GES

We demonstrated the diagnostic yield of the putative indicators using logistic regressions that identified indicators for predicting whether zooplankton community is within variability typical of RefCon_Chl or RefCon_Fish conditions. Each of the 25 regression models achieved a high level of statistical proficiency with three or fewer predictive variables. The model reliability was moderate to high, 65–88% and 66–86% for the RefCon_Chl and RefCon_Fish models, respectively (Table 5). The model sensitivity was similar between RefCon_Chl and RefCon_Fish models, whereas model specificity was significantly higher in the RefCon_Fish models, with lower between-model variability (~80%; S5 Fig). Thus, the RefCon_Fish models can predict equally well cases both within and outside the fish feeding reference conditions whereas the RefCon_Chl models are reliable for predicting in-GES cases. In all models, MeanSize was the most common significant predictor, contributing to >50% and 75% of the models for RefCon_Chl and RefCon_Fish, respectively (Table 5). In all cases, the probability of falling outside the reference state increased with decreasing body size of a zooplankter in the community. Alternatively, a combination of TZA and TZB as negative and positive predictors, respectively, was observed, implicating a positive effect of MeanSize on the probability of zooplankton community being in the reference state. In some basin-specific models, also other metrics, such as copepod biomass (CB), the contribution of copepods to the total zooplankton biomass (CB%), the biomass-based community ratios or contribution of microphagous groups (MMB%), were equally good or even superior. For example, CB was a single positive predictor for fish feeding conditions in the Bornholm basin, whereas MMB% was a single negative predictor for the coastal northern Baltic proper (Askö); both models had high classification accuracy (Table 5).

Overlapping RefCon_Chl and RefCon_Fish conditions

Delineating zooplankton indicators of the in-GES state between the RefCon_Chl and RefCon_Fish is particularly challenging because these periods largely overlap in our datasets (Fig 2). As a result, the zooplankton datasets used to train the models, overlap as well. Moreover, one has to keep in mind that the reference periods that represent pelagic food web “not measurably affected by eutrophication” were defined using existing EQR for chlorophyll. In the Baltic Sea, the documented chlorophyll data extend back to the 1980s, rarely 1970s, which means that they are not likely to cover truly non-eutrophic conditions [61,81]. Thus, for all datasets (except, perhaps, the Gulf of Bothnia), the baselines corresponding to RefCon_Chl are, in fact, likely to represent mesotrophic to eutrophic communities that were typical for the Baltic Sea in the 1970s-80s [82]. Therefore, in addition to the constraints related to the data availability, the overlap between these conditions may have occurred in the Baltic Sea, similar to other systems, where moderately eutrophied conditions were beneficial for fish production [83,84]. The latter can explain why the indicators for RefCon_Chl and RefCon_Fish are similar. However, in the RefCon_Fish models, the indicators describing total zooplankton stock size (TZA, TZB or both) contributed as positive predictors, unlike the RefCon_Chl models where indicators related to community structure (i.e., percentages of main groups and/or ratios) were more significant (Table 5). These differences in the relative importance of the predictors are indicative of the structural and functional properties of a food web with high energy transfer efficiency vs. food web not measurably affected by eutrophication [85].

Conclusions and Future Directions

On the basis of our operational definition of GES, we propose mean body mass of zooplankter in the community (MeanSize) in combination with zooplankton stock measured as either abundance (TZA) or biomass (TZB) to be applicable as an integrated indicator within the Descriptor 4 in the Baltic Sea. These metrics performed best in predicting zooplankton being in-GES when considering all datasets evaluated and can be integrated as a single two-dimensional indicator representing the mean size and total stock (MSTS) of zooplankton (Fig 4). The rationale for MSTS is as follows. High standing stocks of zooplankton composed by larger organisms have a higher capacity for transfer of primary production to fish production (i.e., higher energy transfer efficiency). By contrast, the dominance of small-sized organisms indicates the prevalence of microbial prey and thus inefficient energy transfer due to losses in microbial loop. Thus, abundant zooplankton with high mean individual size would represent both favorable fish feeding conditions and high grazing potential. All other combinations of zooplankton stock and individual size would be suboptimal and imply food web limitations regarding energy transfer from primary producers to higher trophic levels and poorer food availability for planktivorous fish. Of course, these conclusions are based on the zooplankton data representing only the growth period (June–September); a further evaluation is required to understand indicator properties of zooplankton during other seasons. Nevertheless, our evaluation suggests that in several basins of the Baltic Sea, such as Åland Sea, northern and southern Baltic proper and major gulfs (Gulf of Finland and Gulf of Riga), zooplankton communities currently appear to be out-of-GES, being comprised by smaller zooplankters and having lower total abundance or biomass compared to the communities during the reference conditions. However, the changes in the taxonomic structure underlying these trends vary widely across the sea basins.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Conceptual diagram for MSTS, a two-dimensional indicator comprised by total stock (TZA or TZB) and MeanSize.

The green area represents in-GES condition, orange areas represent sub-GES conditions where only one of the two parameters is adequate, and the red area represents sub-GES conditions where both parameters fail.

https://doi.org/10.1371/journal.pone.0158326.g004

In future work, MSTS should be validated for all Baltic Sea areas where zooplankton monitoring is conducted. In particular, its applicability in the Western Baltic Sea, much of the Eastern, South-Eastern, and Southern Baltic, as well as southern and eastern parts of the Gulf of Finland, needs to be tested before MSTS can be applied in these assessment units. In addition to MSTS, region-specific supporting indicators can be considered based on our findings showing the adequate performance of other community metrics in the binary models. Also, temperature- and salinity-induced changes in MSTS need to be evaluated and, if relevant and significant, they need to be accounted for in the indicator-based assessment of eutrophication effects on pelagic food webs. Moreover, the outcome of MSTS-based assessment needs to be cross-checked with other food web indicators within Descriptor 4 as well as eutrophication status according to current ecological status assessment in the specific assessment units.

Methodologically solid monitoring programme is a prerequisite to interpret, evaluate and predict the state of the zooplankton communities. To this end, harmonization of sampling and analytical methods for zooplankton analysis in the Baltic Sea is essential for the application of the unified indicator approach across the sea basins. For MSTS validation, particularly valuable are long data series with high taxonomic resolution that allows accurate individual size and total biomass assessment. For data-poor areas, a cross-basin surveillance program would help establishing indicator baselines by identifying neighbouring data-sufficient areas with similar communities. However, the ultimate goal of the monitoring is to generate sufficiently long data sets using consistent methods and gears, if we are to detect meaningful changes in zooplankton communities and to provide scientific advice on ecosystem management.