Changes in ecosystem resilience detected in automated measures of ecosystem metabolism during a whole-lake manipulation

Significance

Large changes can occur when ecosystems cross certain thresholds. Crossing such thresholds poses a challenge to ecosystem management because the positions of the threshold are uncertain and change over time. However, as an ecosystem approaches a threshold its resilience declines, resulting in changes in system dynamics that increase variance and autocorrelation. Calculating these statistics requires frequent and sustained sampling efforts. Our study detected an approaching threshold by computing the statistical indicators from data collected by automated sensors, which are far less labor-intensive than comparable manual methods. Thus it may be feasible to monitor for approaching ecosystem thresholds using automated methods. This finding highlights a powerful use of modern sensor technology.

Abstract

Environmental sensor networks are developing rapidly to assess changes in ecosystems and their services. Some ecosystem changes involve thresholds, and theory suggests that statistical indicators of changing resilience can be detected near thresholds. We examined the capacity of environmental sensors to assess resilience during an experimentally induced transition in a whole-lake manipulation. A trophic cascade was induced in a planktivore-dominated lake by slowly adding piscivorous bass, whereas a nearby bass-dominated lake remained unmanipulated and served as a reference ecosystem during the 4-y experiment. In both the manipulated and reference lakes, automated sensors were used to measure variables related to ecosystem metabolism (dissolved oxygen, pH, and chlorophyll-a concentration) and to estimate gross primary production, respiration, and net ecosystem production. Thresholds were detected in some automated measurements more than a year before the completion of the transition to piscivore dominance. Directly measured variables (dissolved oxygen, pH, and chlorophyll-a concentration) related to ecosystem metabolism were better indicators of the approaching threshold than were the estimates of rates (gross primary production, respiration, and net ecosystem production); this difference was likely a result of the larger uncertainties in the derived rate estimates. Thus, relatively simple characteristics of ecosystems that were observed directly by the sensors were superior indicators of changing resilience. Models linked to thresholds in variables that are directly observed by sensor networks may provide unique opportunities for evaluating resilience in complex ecosystems.

In the 20th century, changes in land use, nutrient mobilization, and species invasion altered Earth’s ecosystems more than in any previous century of human history (1). The changing climate and expanding human population and consumption are likely to exacerbate ecosystem change in the future. Ecosystem changes have significant effects on human well-being through benefits that people receive from nature. To assess and anticipate environmental changes, observation networks are expanding for hydrology, biogeochemistry of air and water, land cover, and other relevant features of ecosystems (2–5). The observed variables are often designed to address trends in particular resources or pollutants, such as water flow, contaminant concentrations, ecosystem metabolism, carbon storage, or living resources. Although some applications of environmental observations are straightforward, rapid expansion of sensor networks and information management and analysis (6–8) may create new opportunities for detecting, anticipating, or forecasting fundamental changes, such as ecosystem regime shifts.

Ecosystem thresholds are associated with large changes that can involve a significant loss of resources or trigger the restoration of desired conditions in other cases (9). Thresholds are uncertain, change over time, are not preceded by obvious changes in state, and consequently are difficult to anticipate. However, an ecosystem near a threshold has low resilience, and therefore is relatively sensitive to small perturbations and recovers slowly from these perturbations (10). As a result, statistics related to variance or autocorrelation could be indicators of declining resilience or approaching thresholds (10, 11). Experimental tests of declining resilience in living systems are rare (12–17), especially in large-scale field settings. In terrestrial ecosystems, such as forests and grasslands, changing resilience has been measured by analyzing the spatial patterns of vegetation maps obtained by satellite remote sensing (18, 19). In other cases, resilience has been assessed using time series data (17, 20), although acquisition of appropriate data is challenging. Resilience indicators for time series, such as variance, spectral power, and autocorrelation, require high-frequency and long-term datasets. The intensity of sampling needed to acquire such datasets can be costly, and this cost is amplified if multiple variates must be monitored. Modern sensor technology permits automated, in situ, high-frequency, long-duration, and real-time data collection that is less labor-intensive than comparable manual methods (6). Although automated sensors are promising for meeting intensive data requirements, it is not known whether variables measured by such sensors are suitable for assessing resilience or thresholds in complex ecosystems under field conditions.

Sensor data are often used to estimate metabolic rates (e.g., primary production) at the ecosystem scale (21–23). Primary production and respiration (R) are fundamental ecosystem variables closely related to carbon balance, and they have been the focus of extensive research in diverse ecosystems for many decades. Furthermore, metabolism is closely tied to primary producer biomass and life form, which have been used to detect changing resilience in both terrestrial (18, 19, 24) and aquatic systems (13, 17, 25). So far, however, it is unknown whether metabolism can be used to detect changes in ecosystem resilience.

The objective of this study was to determine if automated measures of ecosystem metabolism could be used to detect declining resilience in a lake approaching a regime shift. We deployed automated sensors in Peter and Paul Lakes from 2008 to 2011. Peter Lake experienced a regime shift due to an experimentally induced trophic cascade, and Paul Lake was an unmanipulated reference lake. At the start of the experiment, the lakes were in different stable states: The food web of the manipulated lake was dominated by planktivorous fishes and contained few largemouth bass (Micropterus salmoides), and the reference lake was dominated by bass and contained few planktivorous fishes (Fig. 1). In one stable state, which we refer to as the “planktivore-dominated state,” the number of adults in a bass population is small and unable to recruit because their young are consumed or outcompeted by other fishes. In the alternative stable state, which we call the “bass-dominated state,” a large population of adult bass limits the planktivore population. These two sets of feedbacks form the mechanistic basis for the alternative attractors (26–28). We added bass to the manipulated lake to induce a trophic cascade, pushing the manipulated lake toward the bass-dominated state that characterized the reference lake throughout the experiment. This critical transition shifts the ecosystem from a planktivore-dominated state to a bass-dominated state when bass abundance passes a critical threshold (27). Resilience declines as the ecosystem approaches this critical threshold. The aim was to approach the unknown critical bass abundance slowly so that the ecosystem would be in transition long enough to assess indicators of changing resilience. Bass were added to the manipulated lake in several small stocking events over the course of 4 y: 12 bass were added on day 189 of 2008, 15 on days 169 and 202 of 2009, none in 2010, and 32 on day 174 of 2011.

Fig. 1.

Conceptual diagram outlining the experimental design and hypotheses. The first two rows are ball and cup diagrams, where the balls represent the state of the system and the steepness of the cups is related to the stability of the system: A ball nestled in a deep cup is stable. In 2008, the manipulated lake (red) is stably situated in a basin characterized by few bass and the reference lake (blue) is stable with many bass. In 2008, the lakes are in different stable states but the manipulated lake begins to change as bass are added to it. The third row shows one variable representing the state of the two systems over time (e.g., chlorophyll concentration), and the fourth row shows leading indicators over time (e.g., autocorrelation time). In 2009 and 2010, the manipulated lake is not yet bass-dominated but the system becomes unstable and the leading indicator rises. The star indicates the day of the “first alarm” of the approaching regime shift, as computed from the leading indicators by the quickest detection method. The elevated leading indicator and the first alarm precede the transition of the manipulated lake to a bass-dominated state. In 2011, both lakes are in a stable, bass-dominated state.

Leading indicators, such as variance or autocorrelation, are statistics computed from time series of ecosystem variables that are affected by an approaching threshold. Trophic cascades affect gross primary production (GPP) and algal biomass, which, in turn, influence other metabolic rates [R and net ecosystem production (NEP)] and other variables associated with metabolism [e.g., pH, dissolved oxygen (DO, percent saturation)]. Automated measurements of these six variables [derived estimates = GPP, R, and NEP; direct measurements = chlorophyll-a concentration (Chl-a), pH, and DO] can be made more frequently and easily than their manually measured counterparts. Previous analyses of this experiment found that the variance and autocorrelation of manually collected time series, including algal biomass, increased before day 230 of 2010, which was when the food webs of the two lakes became similar (17, 29, 30). Here, we address whether or not automated measurements of three directly measured variables and three derived estimates of metabolism could be used to detect the loss of resilience in the manipulated ecosystem. We expected that the warning statistics computed from the time series of these six quantities would signal the approaching threshold before the manipulated lake became bass-dominated but that there would be no such signal in the reference lake (Fig. 1).

Results and Discussion

Surprisingly, the calculated rates of metabolism (GPP, R, and NEP) did not provide consistent signals of the approaching threshold (Fig. 2 and Fig. S1). We used the quickest detection (QD) method to define the day of first alarm (DoFA), the day when a statistical indicator, SD or autocorrelation time (AcT), first became high enough to signal a nearby threshold (Materials and Methods). No alarms of the approaching threshold were detected in GPP. Alarms were detected in the AcT of R and NEP (DoFA in 2008 and 2009, respectively), but neither of these signals was corroborated by a change in SD.

Fig. 2.

Values of leading indicators (columns) computed from calculated rates of metabolism (rows) over time. Red lines are the manipulated lake, and blue lines are the reference lake. Vertical dashed lines separate years (2008–2011). Bass were first added to the manipulated lake on day 189 of 2008, and the regime shift completed near day 230 of 2010. Stars indicate the DoFA, computed from a particular leading indicator–variable combination using the QD method. Daily averages of calculated metabolism rates (millimoles of O₂ per cubic meter per day) were used to compute the leading indicators in the first (NEP), second (GPP), and third (R) rows.

In contrast to the calculated rates of metabolism, automated measurements of the three directly measured variables (Chl-a, pH, and DO) detected the approaching threshold more than a year in advance of its arrival (Fig. 3 and Fig. S1). These three directly measured variables performed similarly: Most of the DoFAs were in 2009 (the DoFA in the AcT of pH was in 2008, and the DoFA in the SD of DO was in 2010). The DoFA for the SD of Chl-a, pH, and DO coincided with increases in low-frequency variability in the manipulated lake relative to the reference lake (Fig. 4). The DoFAs in the directly measured variables were as early as the DoFAs in manual samples of chlorophyll (Fig. S1), which we analyzed as a benchmark because its SD and autocorrelation are known to be indicators of this trophic cascade (17, 29, 30). Additionally, the directly measured variables were better correlated with manual daily chlorophyll concentrations than the calculated rates of metabolism, which had relatively noisy time series (Figs. S2 and S3).

Fig. 3.

Values of leading indicators (columns) computed from directly measured sensor variables (rows) over time. Red lines are the manipulated lake, and blue lines are the reference lake. Vertical dashed lines separate years (2008–2011). Bass were first added to the manipulated lake on day 189 of 2008, and the regime shift completed near day 230 of 2010. Stars indicate the day of the first alarm, computed from a particular leading indicator–variable combination using the QD method. Daily averages of measurements made by automated sensors were used to compute the leading indicators in the first [chlorophyll (Chl), micrograms per liter), second (pH), and third (DO, percent saturation) rows.

Fig. 4.

Time series of the log-ratio of the variance spectrum of the manipulated lake to that of the reference lake, computed from automated high-frequency (5-min) observations of Chl-a (A), pH (B), and percent saturation of DO (C). The colors represent log-ratio, such that warm colors indicate the manipulated lake is more variable at that frequency than the reference lake. Note that the log-ratio of variance (color) is scaled differently for each variable. The vertical axis is the log₁₀(frequency) of the signal computed within the rolling window, such that a fortnightly variance signal is at −3.61, a weekly variance signal is at −3.30, a daily variance signal is at −2.46, and an hourly variance signal is at −1.08. The stars indicate the first alarms computed from the SDs of the daily averages of each variable (Fig. 2). The horizontal placement of the stars indicates the DoFA, and the vertical placement is at the daily frequency. Bass were first added to the manipulated lake on day 189 of 2008, and the regime shift completed near day 230 of 2010. Vertical dashed lines separate years (2008–2011).

In summary, not all variables perform equally well when monitoring for changes in resilience. We detected consistent signals of declining resilience in directly measured variables that were associated with metabolism; we did not see consistent signals in the derived estimates of metabolism. Some regime shifts may involve mechanisms that obscure signals of declining resilience in aggregated system variables like NEP (31), but measurement error or other sources of noise could have a similar effect. In lakes, estimates of GPP, R, and NEP often exhibit high day-to-day variability (32, 33) and are often poorly correlated with potential driver variables (34, 35). Sensor measurements are subject to measurement errors related to spatial heterogeneity and other processes (32, 36, 37). These measurement difficulties affect metabolism estimates (GPP, R, and NEP) because these rates are calculated by fitting models to sensor data, and therefore incorporate uncertainty from both measurement and model errors (21, 36). The directly measured variables are subject only to measurement error. Therefore, additional noise in the model-derived estimates of metabolism may have obscured the signals of declining resilience that were detectable in the directly measured variables.

Our findings suggest that monitoring for changes in the resilience of complex systems like lakes requires a careful choice of variables, based on theory as well as field trials. Depending on the particular ecosystem under study, certain variables might be expected to signal several kinds of critical transitions. For example, aquatic ecosystem regime shifts that are driven by nutrient loading, organic matter loading, and changes in top predator abundance all affect metabolism (16, 38–40). Therefore, variables related to metabolism might be useful to monitor. However, even variables with similar drivers can perform differently when used to compute resilience indicators. We found that direct measurements of variables related to metabolism were more accurate than estimates of metabolism for signaling changes in resilience.

Emerging sensor technology and networks are improving capabilities for environmental monitoring (6, 8). These technologies hold considerable promise for data-intensive and spatially extensive studies of ecosystem processes, such as metabolism, ecosystem stability, resilience, and threshold detection. Organizations like the Global Lake Ecological Observatory Network (4), the National Ecological Observatory Network (2), and the Ocean Observatory Initiative (5) reflect the well-established and growing interest in using automated sensors that are similar to those used in this study. Our results suggest that considerable research is needed to determine the monitoring strategies and variables that are most useful for measuring and comparing resilience in a wide range of ecosystems.

The susceptibility of an ecosystem to changing drivers or random events depends on the characteristics of critical thresholds (41). Crossings of some thresholds are unwanted, such as loss of a coral reef or rangeland, whereas crossings of other thresholds are deliberately sought, as in restoration of productive vegetation on degraded lands (42, 43). Moreover, fundamental progress in ecology requires better understanding of thresholds and the tempo of change in ecosystems. Thus, we expect that field studies of resilience and thresholds will continue to be important in ecology as well as in environmental sciences in general. Emerging technology for observing time series data and for detecting warnings of change will make growing contributions to this field.

Materials and Methods

Fish Communities.

At the beginning of the experiment, Peter Lake had a small population of the piscivorous largemouth bass (M. salmoides), and the fish community was dominated by small fishes, such as the pumpkinseed sunfish (Lepomis gibbosus), golden shiner (Notemigonus crysoleucas), central mud minnow (Umbra limi), fathead minnow (Pimephales promelas), brook stickleback (Culaea inconstans), and dace (Phoxinus spp.). To strengthen the planktivore dominance of Peter Lake, we added 1,200 golden shiners to Peter Lake on day 149 of 2008. Conversely, Paul Lake was in a bass-dominated state throughout the experiment, with only a small number of L. gibbosus present. No manipulations were made in Paul Lake.

Automated Sensor Data.

Peter and Paul lakes were each monitored with two YSI multiparameter sondes (model 6600 V2-4) fitted with optical DO (model 6150), pH (model 6561), optical Chl-a (model 6025), and conductivity-temperature (model 6560) probes. Sensor measurements were made at 0.7 m every 5 min and were calibrated weekly. Wind speed and mixed layer depth (z_mix; the shallowest depth below which temperature changes by at least 2 °C m⁻¹) data supplemented other sensor data as input to calculate ecosystem metabolism.

Statistical Indicators.

Three statistical indicators were computed for each of the variables using programs written by the authors in the R programming language (44). Increasing SD (45), increasing AcT (15, 20), and increasing spectral ratio (variance spectrum of the manipulated lake relative to that of the reference lake) at low relative to high frequencies (46) are signals of an approaching threshold. SD and AcT, −log_e(autocorrelation)⁻¹ (15), were computed for manual samples of Chl-a (micrograms per liter), daily averages of the three directly measured variables collected by automated sensors [pH, DO (percent saturation), and Chl-a (micrograms per liter)], and the three calculated metabolism rates (GPP, R, and NEP; millimoles of O₂ per cubic meter per day) (SI Text). Spectral ratio was computed only for the three direct measures of metabolism that were collected by automated sensors. Indicators were computed in 28-d rolling windows. We expected to detect signals of low resilience in the manipulated lake in the form of the three indicators increasing before the regime shift in late 2010. We expected no signals in the reference lake.

QD.

To determine when the change in an indicator first became large enough to constitute a signal of low resilience, we used the QD method for detecting state changes (47–49). In our application, the QD method uses two probability distributions for an indicator: a baseline state f(x_i,t), where the indicator has low values (far from the tipping point), and a critical state g(x_i,t), where the indicator has a high value (close to the tipping point). Indicator x in lake i on day t (x_i,t) follows probability density f(x_i,t) for the baseline state and g(x_i,t) for the critical state. The likelihood for the indicator being in either state is updated each time the systems are observed, and the ratio of these likelihoods is g(x_i,t)/f(x_i,t) = Λ_t. The detection statistic for the QD method is R_t, and R_t is updated as new Λ_t arrive: R_t = (1 + R_t−1) × Λ_t. When the indicator is more likely to be in the critical state than in the baseline state, Λ_t is greater than 1, which increases the rate at which R_t rises. The DoFA is when R_t first reaches a predefined threshold, A. We used the same formulation for calculating the parameters of f(x_i,t) and g(x_i,t) for all variables and indicators, and over broad ranges, the DoFA was not sensitive to the value of these parameters or to our choice of A (Figs. S4–S7 and SI Text). The QD method is particularly useful for detecting approaching thresholds because observations after the shift to the new state are not needed to signal the alarm.