Operational Inflexibility

Daily bathymetric movements of cod (Beamish 1966), their use of visual cues to avoid trawls and their potential effects on survey catches (Konstantinov 1964) have long been known to scientists and fishermen. To avoid adding this variability to index-based surveys and with it reducing the power to detect change, catches are standardized to daylight hour. The inevitable consequence, however, is that International Bottom Trawl Survey (IBTS) cod abundance data are comparable only to data collected under the same environmental conditions. Events not under our control such as annual differences in light conditions not due to daily periodicity such as turbidity or cloud cover would equally affect the cod index. The latter effects are likely to be less significant at night, but we are unable to evaluate performance improvements in the absence of night-time collections. Furthermore, Shephard et al. (2015) provide examples of conflicting standardizations hindering potentially efficient multiple indicator data collection. Solving the problem by abandoning historic time series and developing new multiple indicator time series with more compatible standardization merely treats the symptom (a specific problematic constraint, i.e. where or when we sample) but not the cause (being methodologically constrained and operationally inflexible).

Uncontrolled Changes

Where standardization is not conflicting, potentially uncontrollable behavioural or ecosystem changes may still introduce biases into the temporal perspective. Many cod stocks undertake seasonal migrations (Metcalfe 2006; Neat et al. 2014). It therefore matters where and when monitoring takes place. Indexed-based fisheries monitoring samples at the same place at the same time of the year to minimize the effect of migration on catchability, ensuring abundance estimates between years are comparable. A number of authors have, however, discovered relative changes in the spatial distribution of cod and other species, either because the composition of substocks has changed or because parts of the population have differentiated in their migration behaviour and/or its timing in response to environmental or ecosystem cues (Brander 1994; Hedger et al. 2004; Blanchard et al. 2005; Perry et al. 2005).

Tagging studies (Robichaud & Rose 2004; Righton et al. 2007; Tamdrari et al. 2012) describe changes in the migration behaviour of cod and can even statistically link these to changes in environmental conditions. The considerable investment in investigating (tagging) and monitoring (IBTS) cod has resulted in short-term qualitative estimates of changes in the distribution of cod populations and long-term quantitative abundance estimates of the cod population that are biased. It is not possible to combine these data sources to derive an unbiased estimate of the trend in cod distribution, because sampling was conducted independently at incomparable spatio-temporal scales.

Correlative Relationships

Many documented fish stock-recruitment trends are based on correlations between abundance estimates, environmental variables (temperature and NAO index are common examples) and anthropogenic pressures such as fishing effort. However, the majority of the correlations, though statistically significant, may be in fact coincidental (Myers 1998). Larking (cited in Dickson, Pope & Holden 1974) described the situation as: ‘virtually any set of stock-recruit data is sufficiently variable to inspire untestable hypotheses about the effects of trends in environments, especially with the wealth of meteorological and oceanographic data that can be mined for real and fortuitous correlations’. Vice versa, some undeniably causal relationships between abundance and exploitation may be masked in correlations. For example, the asymmetry in response rate of cod abundance to increases and decreases of fishing pressure is masked by response lags either through mixing spatially or recruitment dynamics temporally. In the absence of process information provided by integrated monitoring, correlation analyses will undesirably tend to produce more false positives and false negatives than statistical evaluation would suggest.

For ecosystem monitoring to service the needs of the ecosystem approach and ultimately ecosystem-based management, it must causally relate the effects of anthropogenic pressures and environmental variability on the ecosystem and the services it provides to society, while taking account of the complexities in those relationships. Future monitoring must therefore be able to provide the means to test causality (Hjermann et al. 2007) and not rely on correlative analyses.

Aggregation

Engelhard, Righton & Pinnegar (2014) relate the change in the distribution of the cod stock in the North Sea to a combination of fishing effort and temperature increases. They conclude that the population had shifted northwards in response to climate change and eastwards in response to fishing pressure. To reach these conclusions, cod abundance, temperature and fishing effort were all averaged at the annual North Sea level and smoothed. Both processes remove variance at the finer spatial and temporal scales. Similarly, Tett et al. (2013) intentionally smoothed data, explicitly ignoring the variation at the local scale, to relate ecosystem states on the regional scale. Aside from confusing correlations with causality described earlier, averaging or smoothing data further complicates the identification of causal effects and increases the chance of random correlations because it reduces the true variability, potentially to monotonic trends that, by their very nature are statistically correlated. Similarly, aggregation leads to a loss of contrast in both dependent and independent variables. This reduces the ability to detect critical effects of change on ecosystem components and potentially underestimates the magnitude of the impacts.

It is less a criticism of the methods themselves designed to demonstrate changes in time. Rather, we are concerned that the use of such unrelated data collections is inefficient and unsuitable for the application of the ecosystem forcing researchers to address policy questions with inappropriate methods. The inability to investigate ecosystem change at appropriate spatio-temporal scales will underestimate uncertainties in future predictions (Planque, Bellier & Loots 2011) and limit the capacity of managers to respond appropriately to change.

The State View – Space And Time (Monitoring Where and When)

The ecosystem state is the result of a large number of cumulative interactions amongst the physical environment, the biota and humans. Jakobsen & Ozhigin (2011) show how much pertinent information on environmental and biological states is available for the Barents Sea. A joint Norwegian and Russian survey in the area since 2004 (see ICES 2012a for a description) has confirmed much of the previous ecosystem understanding of the Barents Sea described by Jakobsen & Ozhigin (2011) based on a compilation of data. A description of a sequence of states of dominant ecosystem components is sufficient to develop monitoring and modelling approaches.

Capelin and herring are the dominant pelagic fish in the Barents Sea ecosystem and use the Norwegian coastal waters as a nursery (Johannesen et al. 2012). Both species move northwards with ontogeny but despite a sizeable overlap in the distributions, they maintain a persistent east–west gradient in the proportions of each species. After 1–2 years in the Barents Sea, herring return to the Norwegian Sea while adult capelin remain, annually migrating north to the summer feeding grounds as the ice melts and returning south to overwinter in the warm water ingress from the Atlantic. Barents Sea cod is one of the dominant consumers of the production through capelin and when the capelin stock increases, the cod stock tends to follow. In the eastern Barents Sea, under more typically arctic conditions, polar cod tend to dominate numerically and are the main consumers of zooplankton production (Wienerroither et al. 2013), leading to a more trophically compressed system. Looking at the life-history characteristics of these fish and the locations they inhabit (Wienerroither et al. 2011, 2013), it is apparent that the environmental characteristics of the water masses (Olsen et al. 2010; Johannesen et al. 2012) define the spatial and temporal patterns of species distributions. The correlation between demersal species distributions and the water masses suggests that the pelagic processes are the primary determinants structuring both, demersal and pelagic communities in the Barents Sea. It should therefore be possible to design ecosystem monitoring based on the spatio-temporal dynamics of these easily identifiable water masses irrespective of whether the interest is in demersal or pelagic species.

Bogstad, Hauge & Ulltang (1997) developed a spatio-temporal raster for modelling pelagic ecosystem interactions in the Barents Sea. The raster was based on the life-history stages and migration patterns of ecologically dominant fish species and the environmental and ecological interactions that determine their dynamics. The raster would equally well describe the distribution of many of the less dominant species, including those examined by Wienerroither et al. (2011, 2013). This is because the less dominant processes operating at smaller spatial scales are invariably correlated or nested within dominant hydrological or ecological processes operating at large spatial scales.

Defining areas or times in which environmental conditions and biological states are comparable (i.e. temporal or spatial strata) greatly simplifies the monitoring approach because we do not have to sample everywhere all the time. Using a common stratification is efficient because it is irrespective of the ecosystem descriptors or indicators to be assessed and will provide time series of standardized ecosystem information in line with the MSFD reporting requirements. At the same time, commonality in the balance of states infers commonality of processes within strata, allowing ecosystem processes to be quantified even when they are poorly understood at present. Given the common scales of variability shared by ecosystem processes, Stommel's (1963) work on monitoring efficiency can be extended from the single to the multi-metric case. This reduces the complexity of the monitoring design and the need to a priori define the metrics in relation to the design with only a small loss of precision. The generalization is necessary, because monitoring, unlike an experiment has to both provide information on changes over time of specific metrics as well as be adaptable to address future questions.

The Process View – Dynamics and Variability (Monitoring Why and How)

Ecosystem processes, comprising complex material cycles and flows of energy, link the biotic and abiotic elements of ecosystems. The state of the ecosystem is the integral of these processes over time. Any particular state variable in the system is characterized by the balance between the upstream and downstream process rates (e.g. recruitment and mortality for fish populations). Changes in this balance will result in a change of state but state alone is not indicative of a process. However, by monitoring states over time, it is possible to gain an ecosystem process view and to understand the important processes affecting ecosystem state. This principle and the importance of linking the data through modelling have been demonstrated by ecosystem research in the Barents Sea and the Baltic Sea over the last 20 years.

Hjermann et al. (2007) examined the ecosystem processes that link capelin, herring and cod. They developed a simple statistical multi-species model, based on the most recent and more detailed time-series data, to predict cod recruitment. They then extended their model back in time, substituting abundance of species about which no information exists with a function of the abundance of species for which long time-series data are available. The innovative approach considered the value of the different information sources so that recent highly temporally resolved samples were more influential in some structural parts of their model, whereas the historical lower resolution data provided the long-term view. In this way, time series can be maintained throughout the transition from species-specific to ecosystem monitoring. This maintains existing time series of indicators where necessary, including fish, eutrophication etc., while ensuring that monitoring remains flexible to meet future demands of policy, and responsive to advances in ecosystem understanding.

Loeng & Drinkwater (2007) qualitatively summarized the links between environment and organisms and the trophic dynamics of the Barents Sea and Norwegian Sea ecosystem in terms of predation and competition amongst dominant fish species. Rather than regarding the ecosystem as static or in equilibrium, they took the variation in ecosystem dynamics and its known or hypothesized causes into consideration. Their ecosystem overview is notable for the realization that it is the variation in processes that helps their significance to be understood. The understanding is more qualitative than quantitative because it relies on outputs from different models dealing with different units and feedback loops that do not function across model boundaries (Travers et al. 2007). Working in the Baltic Sea, Köster et al. (2001) made significant strides towards a more quantitative integrated ecosystem understanding. They showed how both ecological and environmental processes combine to create recruitment variation in Baltic Sea cod. Importantly, they demonstrated that the effects are not additive so that the same pressure can have opposing effects under contrasting conditions. Building on this advanced ecosystem understanding, Möllmann et al. (2008) developed a conceptual model for trophic pathways in the Baltic Sea, considering anthropogenic, ecological and environmental effects and their interactions simultaneously.

The need for conceptual models that help organize complex information on system components and interactions is obvious. Collecting information on ecosystem components independently of the processes that connect them can only work if by chance the spatio-temporal scales of these collections are commensurate with the scales at which key ecosystem processes operate. Experience with the Barents Sea and the Baltic Sea ecosystem has been useful to the management of key ecosystem components because it is based on sound conceptual models that articulate key system components and their interactions at scales relevant to management. However, attempts to apply these principles to modelling of the North Sea ecosystem have not progressed beyond the lower trophic levels (Moll & Radach 2003; Gibson, Atkinson & Gordon 2006; Travers et al. 2007). Greater ecosystem complexity is frequently cited for the slow progress (Anderson 2005), but this appears to be only part of the problem. North Sea modelling work has been based on highly standardized time series which are spatially and temporally incompatible and hence require aggregation. Consequently, there usually is insufficient contrast in the aggregated data to appropriately quantify key processes. Inevitably, investigation of high-profile topics such as regime shifts or the effects of temperature and fishing activity on cod (for example Kempf, Floeter & Temming 2009; Engelhard, Righton & Pinnegar 2014) and the wider ecosystem (Kröncke 2011; Tett et al. 2013) have reverted to large-scale correlative analyses that cannot address the important question as to why observed ecosystem changes occurred or what can be done to minimize the impacts at scales relevant to management. A consideration of the underlying processes can fulfil those aspirations (Roff & Evans 2002).

The Process View – Prioritization of Processes and Samples (Monitoring What and What with)

Knowledge of the spatial distribution of multiple states can be used to infer commonality of processes so that monitoring becomes efficient. Ecosystem modelling informs on how to assess states to maximize gains in understanding ecosystem processes. However, the conclusions still leave us sampling everything. How do we decide what to measure and how?

Not all ecosystem processes are equally important. By focusing on processes that quantitatively dominate the energy flow through the system, we can capture the vital signs of the ecosystem. For example, primary productivity (of either photic or chemical origins) forms the basis of all ecosystems. Understanding how much is being produced and how this production is influenced by the timing and magnitude of the ice melt in the Barents Sea (Slagstad & Wassmann 1996) or the extent of stratification and the amount of suspended particulates in the North Sea (Tett et al. 1993) is essential to understanding those specific ecosystems. The relative rates of dissipation of this production through the ecosystem are also highly informative regarding the importance of subsequent ecosystem processes. When most of the primary production is consumed by the pelagic components, benthic components perform a smaller role in structuring the ecosystem. In such instances, it is unlikely that the role of benthic systems can be understood without understanding pelagic systems. In contrast, systems that are top-down-structured will need to more heavily rely on estimating the variation in the consumption of top predators and the pathways by which the primary production reaches them.

The links between some processes are more obvious than others. For example, the link between phytoplankton productivity and zooplankton productivity is stronger and more obvious than the link to demersal productivity (Snelgrove 1999; Renaud et al. 2008). Understanding the process sequence and how processes interact helps to strategically identify a subset of processes that inform on, or contribute to, the estimation of those processes that are not intensely monitored. A balance of monitored processes across the system maximizes the cumulative information samples provide and reduces the risk of failing to detect changes in the system. Model interpretations can and should be included in the estimation of states to further reducing monitoring requirements in cases where processes or sequences of processes can be modelled reliably and quantitatively. Complete replacement of monitoring with modelling, however, is not advisable, as modelled estimates are only ever as good as the models themselves. Unusual and unlikely events invariably lead to new and improved ecosystem understanding. Deep Sea vents, for example, would not have been discovered without sampling, since then current models predicted that conventional energy inputs were insufficient to maintain benthic communities (Roff, Taylor & Laughren 2003) and those models would not have changed without the discovery.

Every sampling methodology has a unique combination of costs, operational characteristics and contributions to the estimation of ecosystem processes. To determine the most effective use of monitoring resources, we must address the following questions: What processes does the sample inform on? How important is this process in the context of the ecosystem under consideration? How much does it cost? How do we balance or trade-off the collection of one type of information over another? What cheaper, less restrictive or more informative alternatives exist for assessing a particular process? For example, in benthic monitoring we might consider whether sediment particle size data are necessary to our ecosystem understanding beyond the classification of predominant habitat types. At the subregional level, particle size distribution appears to be spatially more heterogeneous and temporally more stable than the associated community structures (Barents Sea: Künitzer et al. 1992; Carroll et al. 2008; Renaud et al. 2008; North Sea: Clark & Frid 2001; Kröncke et al. 2011). McBreen et al. (2008) consider sediment particle size as insufficient to describe benthic communities in the Irish Sea and suggest that sediment types are correlated with other environmental variables known to structure benthic communities. Are those environmental variables, including tidal currents and wind-driven disturbance more ecologically relevant, easier and more cost-effective (i.e. not requiring a vessel to collect samples) to measure? Assuming particle size information is still required, should we use Hamon grabs, NIOZ corers or dredges to collect this information? Corers maintain the structure of the sample so can also inform on geochemical processes in the sediments. The Hamon grab can be deployed in coarser sediments and has a fixed sample volume which means it provides more consistent additional information on infaunal communities. Dredges are the quickest to deploy and are most robust, but provide only an unquantified portion of the sediment and infauna. Evaluating trade-offs between indicator requirements by assessing what is necessary, as opposed to ideal, but maximizing the benefit for other indicators makes monitoring efficient.

Much of the existing monitoring data does not readily lend itself to answering these questions quantitatively, but we do have enough conceptual knowledge to rank the major ecosystem processes. Adequate sampling methodologies to quantify these processes are also in place. Initially, the ranking of importance is unlikely to be uniform across all disciplines. However, a recurrent quantitative evaluation of the monitoring data as part of the ecosystem assessment process will provide the necessary information to achieve future consensus regarding the importance of processes. Effects of optimizing the monitoring can be isolated from future changes in the ecosystem because the monitoring programme is flexible. In consequence, impacts of poorly informed decisions are reduced compared to monitoring programmes that rely entirely on consistency for assessing change.

Coordinated vs. Integrated Monitoring: Learning from Experience in the Celtic Sea

We provide an example in relation to a biodiversity indicator to illustrate and contrast coordinated vs. integrated ecosystem monitoring. Phytoplankton abundance and communities vary at fine spatio-temporal scales (fronts and seasonal, respectively) and sample analysis to species level is costly. Standard vertical ring nets provide information consistent with the taxonomic and spatial resolution of historic information. Automated techniques such as flow cytometry (size and pigment composition of individual cells) on a continuous pump system provide a higher spatial resolution and reduce costs but the information is of low taxonomic resolution. This may be sufficient for the purpose of evaluating food chain effects but it does not support biodiversity descriptors that rely on species-level information. Moreover, pump systems are restricted to intake at a certain water depth and plankton vertical distribution is strongly affected by turbulence which varies with weather conditions and cannot be standardized for. The coordinated approach then simply applies the two sampling approaches in parallel and addresses biodiversity and food chains independently, accepting the variance in each and then relates the variable indicators at an aggregate scale. Conversely, the integrated approach accounts for both in a common model of phytoplankton dynamics, using weather conditions as a proxy covariate for turbulence. The value of a pumped sample increases with an increase in turbulence from a biodiversity perspective, suggesting the collection of costly vertical ring net samples can be decreased at times or in areas of high turbulence. Additionally, it can be shown that changes in the plankton composition at a specific depth, which are not due to changes in turbulence, indicate spatial changes in species composition. Using the low cost continuous method to maximize information content in the ring net samples through real-time evaluation, even when the method has little direct benefit in calm conditions, illustrates the benefit of integration. Deploying the ring net when continuous sampling indicates a change in community differentiates integration from co-collection and post-sampling data collation in that it uses the ecological and sampling processes to maximize efficiency.

In integrated monitoring programmes, such information can be linked to other information including satellite chlorophyll data and turbulence estimates from hydrodynamic models, combining the higher spatial and temporal resolution of the satellite data with the higher specificity of the survey data to develop a holistic view of plankton dynamics. A consistent time series of indicators can be developed irrespective of weather conditions. Continuous Plankton Recorder (CPR) data that lacks the spatial resolution outside specific routes or fixed platform collections at key points in the system could enhance the temporal and spatial accuracy of the species information without biasing the results when spatial hydrographic changes affect these samples.

Existing time series can be used to improve the model of plankton dynamics by tuning hind casts in observing system simulation experiments style exercises while simultaneously ensuring consistency with historic data and assessments. The same models can be used to investigate different management options where effects such as nutrient concentrations are known to affect status. The outputs from such evaluations can indicate the spatial and temporal resolutions at which changes are orthogonal, that is the same changes occur in space and time. This informs on the value of specific samples in relation to their costs and allows for improved efficiency of future sampling. Where biases exist, for example observed phytoplankton abundance is consistently lower than expected from model outputs, it is possible to test formally for the most likely cause. Are variations in turbidity in coastal waters a suitable alternative to the current model (difference between satellite and in situ chlorophyll) or is grazing by herbivorous zooplankton a more likely explanation for the low production? If an improved model is warranted, the relative importance of samples is likely to change and this can be incorporated seamlessly into decisions on future monitoring with minimal risk to the consistency of time series.

From Monitoring to a Monitoring Programme

Ecosystem monitoring improves the ability to detect change by maximizing contrast, prioritizing the major ecosystem pathways and permiting us to determine the variability in ecosystem processes. Our understanding of ecosystem processes is critical to ecosystem monitoring because it forms the basis of ecosystem models that can be used to evaluate future scenarios of change, be that ecosystem- or impact-based. A meaningful and efficient ecosystem monitoring programme is therefore much more than the collection of data (Arkema, Abramson & Dewsbury 2006; Plasman 2008). It is a common framework for all activities associated with ensuring responsible use of natural resources. Where data are formally integrated, changes to the programme affect all other aspects of monitoring. If CPR methodology was to change in our phytoplankton example above, or lose cooperation with one of their vessels of opportunity, the impacts would cascade through the monitoring system and change the way data would be collected under the new circumstances. Such changes have to be evaluated carefully in the wider ecosystem monitoring context. The trade-off of integration and greater efficiency is the loss of autonomy of different monitoring efforts, and this requires better communication and greater flexibility amongst monitoring actors than is currently the case. For some governments, greater top-down influence on monitoring activities, changes in monitoring institutions and/or greater cooperation between centres in polycentric management systems will have to be found to implement a coherent integrated monitoring programme.