Evaluating zooplankton indicators using signal detection theory

Ecological Indicators 77 (2017) 14–22

Contents lists available at ScienceDirect

Ecological Indicators journal homepage: www.elsevier.com/locate/ecolind

Original Articles

Evaluating zooplankton indicators using signal detection theory Susanna ∗ , Maiju Lehtiniemi, Laura Uusitalo Finnish Environment Institute, Marine Research Centre, P.O. Box 140, FI-00251 Helsinki, Finland

a r t i c l e

i n f o

Article history: Received 13 October 2016 Received in revised form 30 January 2017 Accepted 30 January 2017 Keywords: Indicator Mean size Abundance ROC AUC Baltic Sea

a b s t r a c t Indicators are used to help managers to conserve biodiversity and guarantee the sustainable use of marine resources. Good indicators are scientifically valid, ecologically relevant, respond to pressures, and it is possible to set target levels for them. Zooplankton is an important link in the food web as it transfers energy from primary producers (phytoplankton) to planktivorous fish, including the commercially important herring in the Baltic Sea. Eutrophication is known to increase zooplankton abundance and decrease the mean size of zooplankton individuals, while particularly herring prefers larger zooplankton as prey. Therefore, both the abundance/biomass and the size structure of the zooplankton community are highly relevant for the functioning of the pelagic food web, and their combination has been proposed as an indicator of the status of the pelagic food web. In this study, we evaluated the indicator performance of zooplankton abundance and mean size in the northern Baltic Sea using signal detection theory and annual zooplankton monitoring data from years 1979–2014. Herring weight-at-age and chlorophyll a levels were used to estimate the reference periods or “gold standard” of the food web status. The sensitivity and specificity of the indicator was evaluated using ROC curves. Thresholds were set and evaluated with positive and negative predictive values for the zooplankton mean size in three sub-basins of the Baltic Sea. The results suggest that zooplankton mean size is able to reflect the state of the food web in the Baltic Sea. Our study also confirms that signal detection theory is useful in evaluating ecological indicators with clear pressures. However, with parameters that are affected by multiple contrary pressures, such as zooplankton abundance, ROC curves cannot offer enough information about the performance of an indicator parameter. © 2017 Elsevier Ltd. All rights reserved.

1. Introduction Human pressures are affecting the marine environment around the world, and to be able to conserve the marine biodiversity and achieve sustainable use and management of marine resources, we need to have knowledge about the state of the ecosystems and how management measures affect them. The EU Marine Strategy Framework Directive (MSFD) aims at promoting sustainable use of the seas and conserving marine ecosystems by requiring the EU member states to attain good environmental status in their marine areas by the year 2020 (MSFD, 2008/56/EC; European Commission, 2008). To be able to monitor the ecosystems, assess their health and implement the MSFD, establishment of several indicators for different ecosystem components is required. Zooplankton is an important link in the aquatic food webs as it transfers energy from primary producers to higher trophic levels, such as planktivorous fish. In the Baltic Sea herring (Clu-

∗ Corresponding author. E-mail address: susanna.jernberg@ymparisto.fi ( Susanna). http://dx.doi.org/10.1016/j.ecolind.2017.01.038 1470-160X/© 2017 Elsevier Ltd. All rights reserved.

pea harengus membras) is an economically important species. The stock sizes have varied depending on the area, and their condition (described as weight-at-age, WAA) has decreased in several areas of the Baltic Sea since the beginning of the 1980s (Flinkman et al., 1998; Cardinale and Arrhenius, 2000; Rahikainen and Stephenson, 2004; ICES, 2015). Decreased salinity correlates with the decrease in WAA. It has been suggested that the bottom-up effect through zooplankton quality and availability may act as a significant controller of herring condition (Flinkman et al., 1998; Lindegren et al., 2011). Herring feeds selectively on larger zooplankton species like copepods Pseudocalanus elongatus and Limnocalanus macrurus and changes in their populations are reflected in herring stock condition (Flinkman et al., 1998; Möllmann et al., 2003; Rönkkönen et al., ¯ et al., 2016). Zooplankton com2004; Rajasilta et al., 2014; Livdane munities are affected by salinity (Vuorinen et al., 1998; Möllmann et al., 2000), temperature (Dippner et al., 2000; Möllmann et al., 2000; Suikkanen et al., 2013) and eutrophication (Hsieh et al., 2011). The zooplankton biomass has been reported to correlate with eutrophication (chlorophyll a) (Pace, 1986) and especially the abundance of smaller grazing zooplankton increases simultaneously with increasing nutrient levels (Hsieh et al., 2011).

Susanna et al. / Ecological Indicators 77 (2017) 14–22

The central Baltic Sea ecosystem has undergone a trophic cascade where the dramatic reduction of cod (Gadus morhua), caused by increased fishing pressures and decreased salinities, has increased the abundance of its main prey sprat (Sprattus sprattus) (Casini et al., 2008; ICES, 2012). The system has shifted from a coddominated system to a clupeid-dominated system. The change can also be seen in the northern Baltic Sea plankton community which has shifted towards a food web structure with smaller sized organisms, resulting in less energy available for grazing zooplankton and planktivorous fish (Suikkanen et al., 2013). Gorokhova et al. (2016) proposed an indicator that reflects the status of the zooplankton community, and thereby the status of the food web functioning. The idea is that zooplankton should be abundant enough to effectively graze phytoplankton and its mean size should be large enough to guarantee effective energy transfer to the higher trophic levels. Zooplankton mean size and total abundance is an indicator that refers especially to the MSFD descriptor 4 which deals with the food webs (MSFD, 2008/56/EC; European Commission, 2008). It reflects the food web status of the Baltic Sea area as it indicates both the fish feeding conditions and the eutrophication of the sea (Simm et al., 2014; Gorokhova et al., 2016). This indicator has also been adopted into the core set of indicators of the Baltic Marine Environment Protection Commission (HELCOM, 2013; Gorokhova et al., 2016). Creating accurate measures of good environmental status for different marine ecosystem components should be considered carefully (Mee et al., 2008; Borja et al., 2012). The definitions of a good ecological indicator include the following characteristics: the indicator should be easy to measure and sensitive to a particular pressure, its response should be predictable, well-known, and have low natural variability, and the information the indicator offers should have clear management implications, i.e. points towards a clearly identified management action (Dale and Beyeler, 2001; Rice and Rochet, 2005; Queirós et al., 2016). Well established and tested indicators guarantee the basis for efficient management decisions. In this study, we present an application of the signal detection theory to evaluate the sensitivity and specificity of zooplankton indicator parameters mean size and total abundance in four areas of the Baltic Sea. ROC (receiver operating characteristics) curves and AUC (area under curve) values are produced to demonstrate the behavior of the indicator, particularly its sensitivity and specificity. Signal detection theory has been widely used in medical research (Murtaugh, 1996) and it has recently been applied in the evaluation of ecological indicators as well (Hale and Heltshe, 2008; Chuˇseve˙ et al., 2016). It is also a useful tool for deliberate threshold-setting for analyzed parameters, and we propose thresholds of a good environmental status for zooplankton mean size to be utilized in environmental decision-making.

2. Material and methods 2.1. Study area The Baltic Sea is a brackish water basin with mean depth of 55 m and area 422 000 km2 . The only connection to the North Sea is through the narrow Danish straits. There are both horizontal and vertical salinity gradients and a permanent halocline at a depth of 60–80 m (Leppäranta and Myrberg, 2009). Our study area covers the most northern parts of the Baltic Sea: the Bothnian Bay, the Bothnian Sea, the Åland Sea and the Gulf of Finland (Fig. 1) The Bothnian Bay and the Bothnian Sea are partly separated basins from the rest of the Baltic Sea as the southern part of the Bothnian Sea is separated by a shallower sill and the water exchange from the main basin is slow. The whole Baltic Sea shows clear signs of eutrophica-

15

tion (Fleming-Lehtinen et al., 2015) and the Gulf of Finland, which has a direct connection to the main basin has also been suffering from anoxia during the last decades (Carstensen et al., 2014). The Baltic Sea ecosystem is very sensitive to changes as most of the species are from marine or limnic environments and thus occur at the limits of their distribution. The catchment area is large compared to the sea area and the food web structure is simple compared to ocean food webs in general (HELCOM, 2010a). 2.2. Data collection The long-term zooplankton data has been collected by the Finnish Institute of Marine Research and Finnish Environment Institute in connection with the HELCOM COMBINE monitoring program. The data have been collected and analysed for 1979–2014. Seven off-shore monitoring stations are included in this study: one in the Åland Sea and two in the Bothnian Bay, Bothnian Sea and Gulf of Finland. We used samples taken in August as this time had the best data coverage over the study years. Zooplankton was collected by vertical tows of the WP-2 plankton net (mesh 100 ␮m) and preserved in 4% buffered formaldehyde. The sampling was performed according to the COMBINE manual (HELCOM, 2010b). The zooplankton were identified to the lowest possible taxonomic level and the mean size of zooplankters was calculated using the standard species and stage specific wet weights (Hernroth, 1975) with modifications by HELCOM ZEN expert group for certain species. The nauplii were excluded as the mesh used does not sample them comprehensively. The zooplankton total abundance and mean size were calculated for each station for each year. The data from the two stations in each sub-basin was averaged and the total abundance and mean size were calculated for each four sub-areas. 2.3. Signal detection theory Signal detection approach is suitable for dichotomous situations where there are only two possible outcomes. For example the ecological condition can be expressed as “acceptable” and “unacceptable” levels and signal detection approach allows to evaluate, how well an indicator with continuous values can reflect these levels (Murtaugh, 1996). A recent application comes from Chuˇseve˙ et al. (2016) who analyzed Benthic Quality Index (BQI) and its response to eutrophication. Signal detection theory uses a matrix that includes the real status of the ecosystem, and the indicator values or predictions (Table 1). The positive indicator response refers to a situation where an indicator detects the bad environmental condition and hence gives an “alarm” or “a positive signal”, whereas negative indicator response means that there is nothing to alarm about. The table can be used in multiple ways: sensitivity, i.e. the true positive rate (TPR) is the probability of a positive indicator response (unacceptable environmental condition) given that the true condition of environment is positive (TPR=TP/(TP+FN)). Specificity, i.e. the true negative rate (TNR), is the probability of a negative indicator response (acceptable environmental condition) given that the true condition is negative (TNR=TN/(TN+FP)). In addition, the analysis enables the assessment of positive predictive value (PPV) and negative predictive value (NPV); i.e. if the indicator predicts a positive outcome, what is the probability that the true status is positive, and similarly, if the prediction is negative, what is the probability that the true status is negative. These are calculated as PPV=TP/(TP+FP) and NPV=TN/(TN+FN). PPV and NPV take into account the prevalence of real positives in the data set, which affect the probabilities of different outcomes (Murtaugh, 1996). Finding a good indicator threshold often involves a trade-off between the false positive and false negative rates: as an extreme example, if we classify everything as negative, there will be no false

16

Susanna et al. / Ecological Indicators 77 (2017) 14–22

Fig. 1. Map of the study area. Black dots implicate the sampling stations.

Table 1 The observation matrix. Each cell would include the number of observations falling into that class.

Real Positives (Bad environment condition) Real Negatives (Good environment condition)

Predicted Positive (Indicator outcome)

Predicted Negative (Indicator outcome)

True Positives (TP)

False Negatives (FN)

False Positives (FP)

True Negatives (TN)

positives, but very many false negatives, as all the positive instances are also classified as negative. It may be that a single optimal threshold does not exist, but a choice must be made between equally good thresholds, one producing more false negatives and the other more false positives. ROC (receiver operating characteristics) curves demonstrate the classification accuracy of an indicator in terms of sensitivity (TPR) and specificity (TNR) (Murtaugh, 1996). ROC curve is a normalized coverage plot where the distributions of different classes (actual

number of true positives and negatives and false positives and negatives) have been taken into account (Flach, 2012). In a ROC curve, sensitivity is on the y-axis and specificity on the x-axis. The ROC curve of a perfect indicator would go straight up along the y axis all the way, and then to the right (Fig. 2). So the more the ROC curve is bounded to the upper left corner, the higher is the separating power of an indicator. As this is often not achievable, the best classification accuracy is achieved by choosing a threshold that is closest to the upper left corner in the ROC plot. However, as there is a direct

Susanna et al. / Ecological Indicators 77 (2017) 14–22

17

Bothnian Sea 1981–1991, Åland sea 1983–1991 and Gulf of Finland 1979–1988) were derived from the weight-at-age (WAA) of the Baltic herring (Rahikainen and Stephenson, 2004) and ICES stock assessment (ICES, 2015). As especially the diet of larger herring includes also nektobenthos (mysids, amphipods, polychaetes), we chose to use the age class of 4-year-old herring which feed only on zooplankton (Casini et al., 2004). ROC curves and AUC values were produced for both indicator parameters against both gold standards (chlorophyll a and herring WAA) in all four study areas separately. The AUC value was considered acceptable when it exceeded 0.7 and excellent when 0.8 (Hale and Heltshe, 2008). The threshold-setting was done by selecting the cut point where the sum of sensitivity and specificity values was the highest. PPV and NPV were calculated for each proposed indicator parameter threshold to demonstrate the actual performance of a parameter. All data analysis and graphs were produced using R 3.1.2. software and package pROC (Robin et al., 2015). Fig. 2. Example of a ROC curve. The more the curve is bounded to the upper left corner, the better the indicator is performing. A straight line in the middle would mean that an indicator has no sorting power over different environmental conditions. The AUC statistic gives the proportion of the area that is under the curve. A perfect indicator that goes right through the upper left corner would have the whole plot area under the curve, i.e. AUC = 1. An indicator that has no sorting power (represented by a diagonal line) would have AUC = 0.5.

trade-off between the sensitivity and specificity of the indicator, this may need to be taken into account as well, and the user may want to choose a threshold with, e.g., a weaker overall accuracy but higher sensitivity. AUC value measures the area under the ROC curve. The AUC value 0.5 indicates that an indicator has no sorting power over the different environmental conditions, whereas AUC value 1.0 would indicate a perfectly functional indicator (Swets, 1988). As implied in Table 1, in principle the signal detection theory requires that the true status of the measured system is known. However, as that is often impossible, a “gold standard”, the best available information about the true state, can be used instead (Murtaugh, 1996). In ecological research, this can be an environmental variable that is in a good status and closely linked to the variable of interest, or an expert opinion of which values represent the good status (defined, e.g. through the years when the system was in a good status). For the zooplankton community status, chlorophyll a and herring growth were chosen as the gold standards, as these neighboring food web components are closely related and affect each other strongly (Gorokhova et al., 2016). The chlorophyll a reference years (Bothnian Bay 1979–1987, Bothnian Sea 1979–1985, Åland Sea 1979–1985 and Gulf of Finland 1979–1983) are based on the target levels of EQR values stated in HELCOM’s eutrophication report (2009); the years reaching the target level of 0.67 are considered to be in a good condition. For Gulf of Finland the EQR of chlorophyll a did not actually reach the target level in any of the years so we used the earliest years available which were closest to the target levels. While the phytoplankton, zooplankton, and planktivorous fish statuses are expected to correlate, other factors than phytoplankton and planktivorous fish definitely affect the status of the zooplankton community i.e. salinity and temperature. However, these years represent the “best estimate” of the years when zooplankton community has been in good status concerning the next trophic levels this indicator is targeting. Violations of this assumption, if any, can decrease the performance of the indicator in the present analysis. Herring weight-at-age was also used as a gold standard of the good environmental status as both zooplankton mean size and total abundance were assumed to enhance the fish feeding conditions (Gorokhova et al., 2016). The years (Bothnian Bay 1981–1991,

3. Results The indicator parameters zooplankton mean size and total abundance were analyzed for all study areas separately as regional differences are considerable. Zooplankton mean size exceeded the acceptable AUC threshold in response to the herring WAA reference and is therefore able to reflect the status of the food web in three out of the four tested areas (Fig. 3). The only area where the performance of the mean size indicator parameter did not reach the acceptable value was the Bothnian Sea. In the Åland Sea the AUC value (0.862) was considered excellent. Zooplankton mean size did not show as clear response to chlorophyll a reference, as only in one area (Åland Sea) it could reach the acceptable value (Fig. 4). The zooplankton total abundance against chlorophyll a reference exceeded the acceptable AUC value in the Bothnian Sea and Åland Sea (Fig. 5) but not in the Bothnian Bay and the Gulf of Finland where therefore the indicator response is considered poor. When tested against the herring WAA, the zooplankton abundance exceeded the acceptable value only in the Bothnian Bay (Fig. 6). The proposed thresholds are presented in Fig. 3 and in Table 2, where also the negative (NPV) and positive predictive values (PPV) calculated based on the prevalence of the unacceptable environmental condition are shown. Alternative thresholds were proposed for area to demonstrate the variation of specificity, sensitivity, NPV and PPV in different thresholds. 4. Discussion Our results confirm that the mean size of zooplankton reflects the status of the food web and the condition of herring in the northern parts of the Baltic Sea. Mean size showed a response to the weight-at-age of herring by exceeding the acceptable value (AUC ≥ 0.7) in three out of four study areas. These results are in accordance with Gorokhova et al. (2016) who also concluded that zooplankton mean size is a good measure of fish feeding conditions in all other areas except the Bothnian Sea. We also tested the mean size against the eutrophication pressure as eutrophication should increase the abundance of smaller zooplankton, but this link was not clear as the mean size was able to reflect the eutrophication status only in the Åland Sea. Total abundance of zooplankton did not have a clear response to eutrophication or herring weight-at-age in the study areas based on signal detection theory. The salinity has decreased in the Bothnian Bay and the Bothnian Sea which has led to an increase in the abundance of the limnic copepod Limnocalanus macrurus. Limnocalanus sp. is an important food source for herring and can improve herring condition ¯ et al., 2016). The recent increase in (Rajasilta et al., 2014; Livdane

18

Susanna et al. / Ecological Indicators 77 (2017) 14–22

Fig. 3. ROC (receiver operating characteristics) curves and AUC (area under curve) values of zooplankton mean size response to herring weight-at-age status. Proposed thresholds (with specificity and sensitivity values) are also printed in the figures where the AUC exceeds the acceptable value (>0.7).

Table 2 Proposed thresholds for zooplankton mean size for the areas that reached the acceptable AUC value. Alternative thresholds were proposed for each area to demonstrate the differences in the features of the thresholds. Zooplankton mean size Area

Threshold for the mean size (mg)

Prevalence of the unacceptable environmental condition

Specificity

Sensitivity

NPV (%)

PPV (%)

Bothnian Bay Bothnian Bay Bothnian Bay Åland Sea Åland Sea Gulf of Finland Gulf of Finland

0.0510 0.0333 0.0230 0.0125 0.0133 0.0113 0.0095

0.69 0.69 0.69 0.79 0.79 0.71 0.71

0.56 0.78 1.00 0.83 0.67 0.80 0.90

0.95 0.70 0.35 0.91 0.96 0.75 0.63

83 54 41 71 80 57 0.50

83 88 100 95 92 90 0.94

the abundance of L. macrurus may explain the growth in the herring stock in the Bothnian Sea but as suggested by Lindegren et al. (2011), a strong increase in the biomass of sprat in recent years may have strengthen the interspecies competition and explain why the condition of herring has not improved as much. However, the recent slight increase of the herring condition may imply that the zooplankton mean size has only little variation between the reference years and other study years. This variation cannot be detected by the ROC curve and thus zooplankton mean size does not respond clearly to the herring WAA in the Bothnian Sea. Analyzing indicators with the signal detection theory makes it easier to evaluate the sensitivity and specificity of the indica-

tor (Rice, 2003). The response of an indicator to an environmental pressure should be clear (Dale and Beyeler, 2001) and if the known pressures are used as the gold standard, the pressure and its acceptable and unacceptable levels should be known. Setting the targets is usually the challenging part in indicator development. Indicators can have two types of errors: they might miss the bad environmental status or they might give a lot of false alarms, which can become very costly for the society if actions are taken every time based on indicator values (Rice, 2003). Signal detection theory is a very useful tool in evaluating these risks with negative and positive predictive values. The distribution of the real positive and negative examples in the data affects

Susanna et al. / Ecological Indicators 77 (2017) 14–22

19

Fig. 4. ROC (receiver operating characteristics) curves and AUC (area under curve) values of zooplankton mean size response to chlorophyll a status. Proposed thresholds (with specificity and sensitivity values) are also printed in the figures where the AUC exceeds the acceptable value (>0.7).

the assessment: for example in the Gulf of Finland with threshold 0.0113 mg of zooplankton mean size, the negative predictive value (NPV) is rather low; when the indicator shows a good environmental condition, it is correct with probability of only 57% (NPV) (Table 2.). This is caused by the scarcity of the reference years of the herring WAA (gold standard) in the dataset, which affects the probability of detecting a good environmental condition. However, when the zooplankton mean size shows environmental condition below a good status, it is correct in 90% of the cases, so it is better at detecting a bad condition than a good one. This means that the indicator may miss the good environmental status but it gives a relatively low number of false alarms. To be able to say that the zooplankton mean size actually indicates good environmental status, we should have the data and status assessments covering several years. At the moment, the MSFD requires the assessment of marine environmental status every six years, which should be enough for a reliable status assessment with this indicator. To be able to avoid the problems caused by a low number of reference years the option would be to compare the values to other values obtained from a pristine environment, as then it would be possible to get more examples from the good years. Unfortunately, this is impossible in many cases, like in the case of unique environment such as the Baltic Sea. The zooplankton abundance did not have as clear response as the zooplankton mean size and it was able to reflect the status of chlorophyll a and thus eutrophication only in two out of four tested

areas (Bothnian Sea and Åland Sea) and the herring WAA only in the Åland Sea. This was not a totally unexpected result since the assumed link between chlorophyll a and zooplankton abundance has not been fully clear in earlier studies either; Casini et al. (2008) showed that the chlorophyll a concentration correlated negatively with the zooplankton biomass. The amount of chlorophyll a has increased significantly in the Gulf of Finland from the 1970s to the early 2000s (Fleming-Lehtinen et al., 2008), at the same time when the total zooplankton abundance has decreased (Suikkanen et al., 2013). Even though biomass of phytoplankton (and thus chlorophyll a) has been shown to respond to eutrophication (Hsieh et al., 2011), it seems that the increasing temperature might actually have a bigger impact on phytoplankton dynamics. The sum of nitrate and nitrite nitrogen concentrations have decreased in the Gulf of Finland (no change in phosphatephosphorus) but the chlorophyll a concentration has increased (Fleming-Lehtinen et al., 2008). Cyanobacteria are phosphorus limited so they are good competitors in environments where nitrogen is scarce. Cyanobacteria are low quality food for zooplankton and thus do not provide good basis for zooplankton growth (Gulati and Demott, 1997). There is a top-down control of zooplankton as total zooplankton biomass correlates negatively with clupeid abundance (Casini et al., 2006) and decreases with increasing sprat abundance (Casini et al., 2009). The forces of eutrophication and predation pressure on zooplankton abundance work in opposite ways (eutrophication

20

Susanna et al. / Ecological Indicators 77 (2017) 14–22

Fig. 5. ROC (receiver operating characteristics) curves and AUC (area under curve) values of zooplankton total abundance response to chlorophyll a status. Proposed thresholds (with specificity and sensitivity values) are also printed in the figures where the AUC exceeds the acceptable value (>0.7).

pressure either increasing (Pace, 1986) or decreasing (Suikkanen et al., 2013) the abundance of zooplankton and predation pressure decreasing the abundance) so the evaluation of a good status of zooplankton abundance is difficult. During the summer, top-down control on zooplankton is stronger than bottom-up (Casini et al., 2008). The increased clupeid predation pressure may have resulted in decreased zooplankton abundance which may also affect the cyanobacterial blooms (Urrutia-Cordero et al., 2015). Zooplankton abundance is a good example of a parameter that is difficult to evaluate with signal detection theory because the approach expects a dichotomous situation and here the parameter has multiple dimensions. This is also why we did not propose any thresholds for zooplankton abundance even in those areas where it exceeded the acceptable level. This also means that the zooplankton abundance as an indicator of the environmental status needs to be interpreted with care and only in conjunction with the neighboring food web components. The zooplankton indicator proposed for the Baltic Sea (Gorokhova et al., 2016) is a two-dimensional indicator, in which both parameters should show a good status for the indicator to reach the best possible value. This means that the abundance is always evaluated as part of the “bigger picture”, avoiding some of the pitfalls discussed above. If we assume that a well-functioning food web would imply that all components (phytoplankton community, zooplankton abundance and mean size, and herring growth) were in good status at the same time, it would seem reasonable to test the zooplankton

variables against a gold standard that has both the chlorophyll a and herring WAA in good status at the same time. This was however not done, as the good status was not always observed at the same time in chlorophyll a and herring WAA, and the fraction of good environmental status data would have been very small. This implies that the dynamics of the pelagic food web are more complicated than a binary all-good or all-bad situation, and it is important to understand the role of zooplankton and the various zooplankton community metrics in this puzzle. In signal detection approach, the gold standard or reference is turned from a continuous response to a binary one to represent either a “good” or “bad” environmental status. This might, in some cases, result in a loss of information when continuous values are simplified into 1’s and 0’s. By doing this, we also make a decision on how to define a good environmental status. In medical research this is easier; the patient either has or does not have the disease, but in ecology the changes are usually not this simple and responses may be linear or have multiple stable conditions. On the other hand, for example the MSFD has only these two definitions for the environmental status so the method corresponds to its requirements relatively well. It has to be noticed that in complex ecological systems, setting a gold standard is challenging, and it should not be interpreted as “the best possible scenario” or “best possible environmental status” but rather as a state that is assumed “good enough”. As the indicator parameters are usually affected by mul-

Susanna et al. / Ecological Indicators 77 (2017) 14–22

21

Fig. 6. ROC (receiver operating characteristics) curves and AUC (area under curve) values of zooplankton total abundance response to herring weight-at-age. Proposed thresholds (with specificity and sensitivity values) are also printed in the figures where the AUC exceeds the acceptable value (>0.7).

tiple environmental factors, the identified gold standard is always a compromise and may change as more information accumulates.

5. Conclusions Our analysis with signal detection theory demonstrated that the mean size of zooplankton is able to reflect the state of the food web and the energy transfer to higher trophic levels in the Baltic Sea and we recommend that it will be a part of the status assessments of the Baltic Sea. Indicators that include both structural and functional aspects of ecosystem, may give more information about the ecosystem health (Rombouts et al., 2013). The zooplankton total abundance is used to calculate the zooplankton mean size which responds to eutrophication and also to predation. Mean size is thus a great measure of food web balance and the measure of zooplankton abundance may support it. Signal detection theory is not a right method to analyze and determine the threshold values for zooplankton abundance as it has a two way response: on the other hand clupeid pressure decreases it and eutrophication increases or decreases it. However, together with the mean size of zooplankton, the abundance is an important parameter in defining both the grazing rate and the food availability for planktivores, and therefore should be part of ecosystem assessments as has been suggested (Gorokhova et al., 2016).

Acknowledgements This study was partly funded by Maj and Tor Nessling Foundation (Grant No. 201600194) and DEVOTES (DEVelopment Of innovative Tools for understanding marine biodiversity and assessing Good Environmental Status) project funded by the European Union under the 7th Framework Programme, ‘The Ocean of Tomorrow’ Theme (Grant Agreement No. 308392), www.devotes-project. eu. This work was also partly supported by the BONUS BIO-C3 project that were supported by BONUS (Art 185), funded jointly by the EU, and Academy of Finland. We also want to thank Ville Karvinen for the map.

References Borja, Á., Dauer, D.M., Grémare, A., 2012. The importance of setting targets and reference conditions in assessing marine ecosystem quality. Ecol. Indic. 12, 1–7. Cardinale, M., Arrhenius, F., 2000. Decreasing weight-at-age of Atlantic herring (Clupea harengus) from the Baltic Sea between 1986 and 1996: a statistical analysis. ICES J. Mar. Sci.: J. Conseil 57, 882–893. Carstensen, J., Andersen, J.H., Gustafsson, B.G., Conley, D.J., 2014. Deoxygenation of the Baltic Sea during the last century. Proc. Natl. Acad. Sci. 111, 5628–5633. Casini, M., Cardinale, M., Arrhenius, F., 2004. Feeding preferences of herring (Clupea harengus) and sprat (Sprattus sprattus) in the southern Baltic Sea. ICES J. Mar. Sci.: J. Conseil 61, 1267–1277. Casini, M., Cardinale, M., Hjelm, J., 2006. Inter-annual variation in herring, Clupea harengus, and sprat, Sprattus sprattus, condition in the central Baltic Sea: what gives the tune? Oikos 112, 638–650.

22

Susanna et al. / Ecological Indicators 77 (2017) 14–22

Casini, M., Lövgren, J., Hjelm, J., Cardinale, M., Molinero, J.-C., Kornilovs, G., 2008. Multi-level trophic cascades in a heavily exploited open marine ecosystem. Proc. R. Soc. Lond. B: Biol. Sci. 275, 1793–1801. Casini, M., Hjelm, J., Molinero, J.-C., Lövgren, J., Cardinale, M., Bartolino, V., Belgrano, A., Kornilovs, G., 2009. Trophic cascades promote threshold-like shifts in pelagic marine ecosystems. Proc. Natl. Acad. Sci. 106, 197–202. ˙ R., Nygård, H., Vaiˇciut ¯ e, ˙ D., Daunys, D., Zaiko, A., 2016. Application of Chuˇseve, signal detection theory approach for setting thresholds in benthic quality assessments. Ecol. Indic. 60, 420–427. Dale, V.H., Beyeler, S.C., 2001. Challenges in the development and use of ecological indicators. Ecol. Indic. 1, 3–10. Dippner, J.W., Kornilovs, G., Sidrevics, L., 2000. Long-term variability of mesozooplankton in the Central Baltic Sea. J. Mar. Syst. 25, 23–31. European Commission, 2008. Directive 2008/56/EC of the European Praliament and of the Council establishing a framework for community action in the field of marien environmental policy (Marine Strategy Framework Directive). Off. J. Eur. Union, 12–24, L232 http://publications.europa.eu/code/en/en-130100. htm. Flach, P., 2012. Machine Learning: the Art and Science of Algorithms That Make Sense of Data. Cambridge University Press, pp. 396. Fleming-Lehtinen, V., Laamanen, M., Kuosa, H., Haahti, H., Olsonen, R., 2008. Long-term development of inorganic nutrients and chlorophyll ␣ in the open northern baltic sea. AMBIO: J. Hum. Environ. 37, 86–92. Fleming-Lehtinen, V., Andersen, J.H., Carstensen, J., Łysiak-Pastuszak, E., Murray, C., Pyhälä, M., Laamanen, M., 2015. Recent developments in assessment methodology reveal that the Baltic Sea eutrophication problem is expanding. Ecol. Indic. 48, 380–388. Flinkman, J., Aro, E., Vuorinen, I., Viitasalo, M., 1998. Changes in northern Baltic zooplankton and herring nutrition from 1980 to 1990: top-down and bottom-up processes at work. Mar. Ecol. Prog. Ser. 165, 127–136. Gorokhova, E., Lehtiniemi, M., Postel, L., Rubene, G., Amid, C., Lesutiene, J., Uusitalo, L., Strake, S., Demereckiene, N., 2016. Indicator properties of baltic zooplankton for classification of environmental status within marine strategy framework directive. PLoS One 11, e0158326. Gulati, R., Demott, W., 1997. The role of food quality for zooplankton: remarks on the state-of-the-art, perspectives and priorities. Freshw. Biol. 38, 753–768. HELCOM, 2009. Eutrophication in the baltic sea—an integrated thematic assessment of the effects of nutrient enrichment and eutrophication in the baltic sea region. In: Balt Sea Environ Proc. Helsinki Commission. HELCOM, 2010a. Ecosystem health of the baltic sea: HELCOM initial holistic assessment. Baltic Sea Environ. Proc., 122. HELCOM, 2010b. Manual for Marine Monitoring in the COMBINE Programme of HELCOM. Helsinki Commission. HELCOM, 2013. HELCOM Core Indicator of Biodiversity Zooplankton Mean Size and Total Abundance. Hale, S.S., Heltshe, J.F., 2008. Signals from the benthos: development and evaluation of a benthic index for the nearshore Gulf of Maine. Ecol. Indic. 8, 338–350. Hernroth, L., 1975. Recommendations on methods for marine biological studies in the Baltic Sea: mesozooplankton biomass assessment. Baltic Mar. Biol. 10, 1–32. Hsieh, C.H., Sakai, Y., Ban, S., Ishikawa, K., Ishikawa, T., Ichise, S., Yamamura, N., Kumagai, M., 2011. Eutrophication and warming effects on long-term variation of zooplankton in Lake Biwa. Biogeosciences 8, 1383–1399. ICES, 2012. Report of the ICES Advisory Committee 2012. ICES Advice (Book 8, 158 pp). ICES, 2015. Report of the Baltic Fisheries Assessment Working Group (WGBFAS): 14–21 April 2015, ICES HQ, Copenhagen, Denmark. ICES CM 2015/ACOM:10. 826 pp. Leppäranta, M., Myrberg, K., 2009. Physical Oceanography of the Baltic Sea. Springer, Chicester, UK, pp. 378.

Lindegren, M., Östman, Ö., Gårdmark, A., 2011. Interacting trophic forcing and the population dynamics of herring. Ecology 92, 1407–1413. ¯ Livdane, L., Putnis, I., Rubene, G., Elferts, D., Ikauniece, A., 2016. Baltic herring prey selectively on older copepodites of Eurytemora affinis and Limnocalanus macrurus in the Gulf of Riga. Oceanologia 58, 46–53. Möllmann, C., Kornilovs, G., Sidrevics, L., 2000. Long-term dynamics of main mesozooplankton species in the central Baltic Sea. J. Plankton Res. 22, 2015–2038. Möllmann, C., Kornilovs, G., Fetter, M., Köster, F.W., Hinrichsen, H.H., 2003. The marine copepod, Pseudocalanus elongatus, as a mediator between climate variability and fisheries in the Central Baltic Sea. Fish. Oceanogr. 12, 360–368. Mee, L.D., Jefferson, R.L., Laffoley, D. d. A., Elliott, M., 2008. How good is good? human values and europe’s proposed marine strategy directive. Mar. Pollut. Bull. 56, 187–204. Murtaugh, P.A., 1996. The statistical evaluation of ecological indicators. Ecol. Appl. 6, 132–139. Pace, M.L., 1986. An empirical analysis of zooplankton community size structure across lake trophic gradients1. Limnol. Oceanogr. 31, 45–55. Queirós, A.M., Strong, J.A., Mazik, K., Carstensen, J., Bruun, J., Somerfield, P.J., Bruhn, ˙ R., Muxika, I., Nygård, A., Ciavatta, S., Flo, E., Bizsel, N., Özaydinli, M., Chuˇseve, H., Papadopoulou, N., Pantazi, M., 2016. An objective framework to test the quality of candidate indicators of good environmental status. Front. Mar. Sci. 3. Rönkkönen, S., Ojaveer, E., Raid, T., Viitasalo, M., 2004. Long-term changes in Baltic herring (Clupea harengus membras) growth in the Gulf of Finland. Can. J. Fish. Aquat. Sci. 61, 219–229. Rahikainen, M., Stephenson, R.L., 2004. Consequences of growth variation in northern Baltic herring for assessment and management. ICES J. Mar. Sci.: J. Conseil 61, 338–350. Rajasilta, M., Hänninen, J., Vuorinen, I., 2014. Decreasing salinity improves the feeding conditions of the Baltic herring (Clupea harengus membras) during spring in the Bothnian Sea, northern Baltic. ICES J. Mar. Sci.: J. Conseil 71, 1148–1152. Rice, J.C., Rochet, M.-J., 2005. A framework for selecting a suite of indicators for fisheries management. ICES J. Mar. Sci.: J. Conseil 62, 516–527. Rice, J., 2003. Environmental health indicators. Ocean Coast. Manage. 46, 235–259. Robin, X., Turck, N., Hainard, A., Tiberti, N., Lisacek, F., Sanchez, J.-C., Müller, M., 2015. Package pROC. https://cran.r-project.org/web/packages/pROC/index. html. Rombouts, I., Beaugrand, G., Artigas, L.F., Dauvin, J.C., Gevaert, F., Goberville, E., Kopp, D., Lefebvre, S., Luczak, C., Spilmont, N., Travers-Trolet, M., Villanueva, M.C., Kirby, R.R., 2013. Evaluating marine ecosystem health: case studies of indicators using direct observations and modelling methods. Ecol. Indic. 24, 353–365. Simm, M., Kotta, J., Jänes, H., 2014. Mean weight and total biomass of zooplankton as a core indicator of biodiversity of the Marine Strategy Framework Directive: an example of the Gulf of Riga. Estonian J. Ecol. 63, 232–241. Suikkanen, S., Pulina, S., Engström-Öst, J., Lehtiniemi, M., Lehtinen, S., Brutemark, A., 2013. Climate change and eutrophication induced shifts in northern summer plankton communities. PLoS One 8. Swets, J., 1988. Measuring the accuracy of diagnostic systems. Science 240, 1285–1293. Urrutia-Cordero, P., Ekvall, M.K., Hansson, L.-A., 2015. Responses of cyanobacteria to herbivorous zooplankton across predator regimes: who mows the bloom? Freshw. Biol. 60, 960–972. Vuorinen, I., Hänninen, J., Viitasalo, M., Helminen, U., Kuosa, H., 1998. Proportion of copepod biomass declines with decreasing salinity in the Baltic Sea. ICES J. Mar. Sci.: J. Conseil 55, 767–774.