Part one: trends and insights from climate change experiments
Studies which have experimentally applied temperature treatments to ecological communities in the period 2000–2012 were reviewed using Web of Science (accessed 1 July 2012 to 1 November 2012, using the keywords climate change with; experiment or experimental or manipulation or warming). Studies which used natural gradients such as altitude and latitude were deliberately excluded, as they do not directly manipulate environmental conditions. We also excluded studies of single species, which includes a large body of literature from studies of adaptive capacity to evolutionary genetics. Because there were relatively few freshwater studies, the literature review was extended for freshwaters only to include the time period 1995–2012. This resulted in 110 studies (Supplementary materials S1), 66 from terrestrial environments, 23 from marine settings and 21 from freshwaters. Those studies were classified into a priori defined ‘generations’ of experiments, each of which treats temperature in different ways (Table 1).
Table 1. Review and classification into generations of community climate change experiments 2000–2012 (terrestrial and marine) and 1995–2012 (freshwater) which involved temperature manipulations (excluding other physical and chemical manipulations). For definitions of the ‘generations’ of studies, see the main text. A number of studies (with percentages of the total in brackets following) are shown. Individual papers are shown in Table S1
|Generation||Effects on mean||Effects on variability||Incorporates extreme events?||Number of studies found|
|Fixed mean||Increase||Large reduction||No||3 (4.5%)||15 (65.2%)||5 (23.8%)|
|Fixed minima||Increase||Small reduction||No||7 (10.6%)||0 (0%)||0 (0%)|
|Fixed increment||Increase||No effect||Some||52 (78.8%)||8 (34.8%)||15 (71.4%)|
|Extreme event studies||Increase||Increase||Yes||4 (6.1%)||0 (0%)||1(4.8%)|
Generation one: fixed mean experiments
Fixed mean experiments represent the simplest treatment possible and apply temperature treatments at a stable level over the length of the experiment. Most often these take the mean temperature of current conditions and add an increment to it to generate a new mean temperature, which is then applied as the treatment (compare Fig. 1a,b). Some of the studies listed in Table 1 (e.g. Beisner et al. 1996; Mitchell & Lampert 2000) compared fixed temperature treatments, others (e.g. Petchey et al. 1999; Fox & Morin 2001) compared a constant to a warming treatment. These types of experiments underestimate the effects of climate change as they do not include the ‘event effect’ component in the treatment. The warming treatments in these experiments are also associated with a reduction in temperature variability, potentially confounding any results.
Figure 1. Conceptual diagram of generations of temperature treatments used in climate change experiments. (a) baseline temperature (natural or current scenario), (b) fixed mean (temperature set to a fixed value), (c) fixed minima (temperature has a fixed minimum), (d) increment (fixed increment is applied to natural variability), (e) extreme event (extreme event is superimposed on natural variability), (f) down-scaled climate model (temperature is determined by weather scenarios generated from down-scaled climate model). Dashed lines indicate maximum, mean and minimum temperatures. The black arrow indicates when experimental treatments are applied.
Download figure to PowerPoint
Generation two: fixed minima experiments
Fixed minima or maxima experiments have commonly been applied in warming experiments in the field. Experiments using substrate warmers inserted into the forest floor are an example of this type of approach (Melillo et al. 2002), as are experiments which re-radiate heat during the night to reduce night time minimum temperatures (e.g. Lloret et al. 2005). While able to prevent the coolest temperatures occurring, and having some warming effects on cool to moderate temperature days, this approach cannot affect the warmest days or generate high temperature extremes. Effectively, these treatments generate fixed minimum temperatures. The effect is to increase mean temperatures but to reduce variability, although not to the extent of fixed mean experiments (Fig. 1c).
Generation three: increment studies
Increment studies apply a temperature treatment while retaining natural variability in temperature. Most often, these treatments are applied as a fixed increment (for example, + 3.5 °C) over natural conditions. These experiments have the advantage that they incorporate many of the natural features of weather, for example, one warmer than average day is more likely to follow another than it is to follow a colder than average day. Overall, these studies increase mean temperatures while retaining the variability which is typical of current climates (Fig. 1d). For example, Yvon-Durocher et al. (2010) used twenty mesocosms in southern England and warmed ten of these by 3–5 °C above ambient conditions. Similarly, Perdomo et al. (2012) applied a 6 °C increment to moss patches in the field. These types of experiments cannot incorporate features such as predicted climates where, for example, winter becomes warmer but spring becomes cooler. Nor do these kinds of experiments take into account changes in the climate variability. As such, they may underestimate the effects of climate change in some systems.
Generation four: extreme event studies
The most recent examples of climate change studies explicitly include extreme events in some fashion. In terrestrial studies, experimental enclosures have been exposed to drought, night heat waves and extreme rainfall scenarios to assess effects on primary productivity (e.g. Fay et al. 2000; Beier et al. 2004). These approaches do not seem to have been applied in freshwater studies of the effects of temperature. Dang et al. (2009) applied an increased diel temperature variation to stream mesocosms and assessed impacts on detrital decomposition, but this experiment exposed the system to a cyclic series of extreme events rather than periodic events. A number of freshwater studies have assessed the effects of drying as an extreme event (Leberfinger et al. 2010; Ledger et al. 2011), but none to date have considered extreme temperature events such as heat waves explicitly, as shown in Figure 1e. Extreme event studies increase means and variability in temperatures, but do not replicate changes in the timing or duration of extreme events.
Part three: using down-scaled climate models to generate experimental climate change treatments
It is now possible to generate experimental treatments which are based on the predictions of global climate change models for large scale climate phenomena, but down-scaled to generate hourly weather scenarios. Two types of approaches (dynamical and statistical) are normally used to take information from global climate models (GCMs) (~ 100 km resolution) to be applied at higher resolutions that are more meaningful to local ecological scales (see Wilby & Wigley 1997; for a review). These approaches have been widely used in hydrology, but not directly in ecological experiments (Wilby & Dawson 2012). GCMs typically have coarse temporal (monthly) and spatial resolution and are most useful at these scales. Experimental treatments for ecological studies need predictions at relatively fine spatial and temporal scales. These need to incorporate increases in mean temperatures, but also increased variability and increased frequency of extreme events, such as heat waves and extreme rainfall events, and more subtle impacts such as changes in cloud cover. For example in Figure 1f, prolonged extreme high temperature events (‘heatwaves’) appear in the treatment based on predictions from a GCM.
In our example, we sought to generate a climate change treatment to apply to indoor experimental stream flumes to assess climate change impacts on temperate Australian stream benthic communities. We wanted to compare responses to conditions representative of mid-summer over the last decade, to mid-summer conditions predicted to occur under a climate change scenario for 2100. The controllable variables in the flumes were temperature, rainfall (as flow velocity) and light intensity. We carried out the down-scaling process for one future time (2100) and one time of year (60 days in summer), using a single model and one emissions scenario (A1B scenario, predicting a year 2100 carbon dioxide concentration of 700 ppm (IPCC 2000). However, more complex experiments could generate treatments for other years, times of year or emissions scenarios. In addition, multi-model ensembles could be used to capture the uncertainty in climate predictions resulting from structural differences in the GCMs as well as uncertainty due to variations in initial conditions or model parameterisation (Semenov & Stratonovitch 2010). It is important that these weather time series are not averaged in a multiple ensemble as the resultant time series will lose its statistical variation. Rather the key here is to ultimately generate multiple weather time series treatments (ensembles) that are applied experimentally so that the ecological results are robustly replicated.
Our strategy was to use the information contained in a GCM output, which projects how climate may evolve under future scenarios over the following centuries, and apply that to the local scale. We then merged this data with statistical information from real historical observations and applied that to the changed climate from the GCM to a time series at daily resolution using a ‘weather generator’ (see below). We used the MIROC (Model for Inderdisciplinary Research on Climate) global climate model outputs available from the Center for Climate System Research (CCSR), University of Tokyo (http://www.ccsr.u-tokyo.ac.jp/) as the basis for our generation of the temperature treatment data. The model has a spatial resolution of 1.4 degree in longitude, 0.5–1.4 degree in latitude, and 43 vertical levels in the medium-resolution version. We chose this model because it has performed well for the Australian climate (Pitman & Perkins 2008). Data were extracted from the CMIP3 (http://www-pcmdi.llnl.gov/ipcc/about_ipcc.php) archive which is a repository for climate models that were used in preparing the IPCC Fourth Assessment Report (http://www.ipcc.ch/). We extracted the air temperature variable (TASA1) from the run ‘sresb1atmmotasmiroc3_2medres’ to demonstrate the method. This file was for the A1B scenario with a carbon dioxide concentration in the year 2100 of 700 ppm. Further information on climate change scenarios can be found at www.ipcc.ch/pdf/special-reports/spm/sres-en.pdf. We extracted data for the grid cell closest to Melbourne Airport, Australia (37.67 °S 144.83 °E) for the 21st century.
To generate weather data, we entered the GCM data into the LARS-WG (Long Ashton Research Station Weather Generator) stochastic weather generator (http://www.rothamsted.ac.uk/mas-models/larswg.php) (Semenov et al. 1998). LARS-WG is a model simulating hourly time series of daily weather at a single site, which can generate long time series of weather conditions for a particular site, and includes extreme weather events, such as extreme daily precipitation and long dry spells or heat waves (Semenov et al. 1998). LARS-WG has been well validated in diverse climates around the world (Semenov et al. 1998). It utilises semi-empirical distributions for the lengths of wet and dry day series, daily precipitation and daily solar radiation. The seasonal cycles of means and standard deviations are modelled by finite Fourier series of order three and the residuals are approximated by a normal distribution (http://www.rothamsted.ac.uk/mas-models/download/LARS-WG-Manual.pdf).
We used the following methodology as per Semenov & Stratonovitch (2010).
- Model Calibration – Observed weather data from Melbourne airport (Australian Bureau of Meteorology site number: 086 282, elevation: 113 m, period: 1990–2009) were analysed to determine the local statistical characteristics of air temperature. This information is stored in two parameter files.
- Model Validation – the statistical characteristics of the observed and synthetic weather data were analysed to determine if there are any statistically significant differences (none found).
- Generation of Synthetic Weather Data – the parameter files derived from observed weather data during the model calibration process were used to generate synthetic weather data having the same statistical characteristics as the original observed data, but differing on a day-to-day basis. We applied our global climate model-derived changes in temperature to the LARS-WG parameter files to generate daily weather for 2090–2100.
- Experimental series – A series of weather (20 years long) is generated based on the changes in global climate (2090–2100) and the January/February period for the 10th year was extracted for use in driving the experimental treatments (Fig. 2). Data were similarly generated for the control period (1990–2010). Probability distribution functions for distributions of minimum and maximum temperatures were generated for 2100 (generated by the simulation) and based on combined data for real weather data from the same region 1990–2000 (Fig. 2). Because we needed water temperature data (rather than the air temperature data generated by the model), a long-run series of historical water temperatures for the study site were used with historical air temperature data from the Melbourne airport weather station to generate a relationship between air and water temperature. It is important to note that this kind of relationships is highly nonlinear (Mohseni et al. 1998) and may be relatively site specific depending on local riparian vegetation and interactions with groundwater, among other factors. As such, experiments which seek to assess impacts on particular freshwater sites will require detailed historical water temperature data.
Figure 2. Example of the potential to downscale climate models to generate climate change treatments. Probability distribution functions illustrate the shifts in the actual and expected distributions of (a) maximum and (b) minimum temperatures for the decade 1990–2000 (based on real data, white striped bars) and 2100 (based on weather simulations from the climate model; grey bars). (c) Modelled temperature series for the first 60 Julian days of 2100.
Download figure to PowerPoint
It should be noted that a stochastic weather generator is not a predictive tool that can be used in weather forecasting, but is simply a means of generating time series of synthetic weather statistically ‘identical’ to the observations. The resulting scenarios can be used as experimental treatments to be compared to controls resulting from ambient conditions or to treatments based on historical weather conditions. We used the variance of the ‘real’ historical data and applied that to the climate scenario to generate a weather series. Here, we generated a single run, as generating repeated simulations then averaging results will remove extreme events from the data.
This kind of experimental data allows the application of highly realistic treatments in experiments that include not only changes in mean conditions but also increased frequency, intensity and duration of extreme events. However, they are challenging to apply outside of highly controlled laboratory conditions. In outdoor conditions, increment studies can superimpose a warming treatment on the background conditions (e.g. Yvon-Durocher et al. 2010; Dossena et al. 2012). With simulated weather, there is the potential that a temperature treatment for a particular day may be cooler than ambient conditions, or may be considerably higher than ambient conditions. Both situations require highly energy intensive equipment to apply the treatments. Although it is possible to apply simulated weather as a treatment in an outside experiment, the approach described in the current paper is most amenable to highly controlled laboratory settings. This has the advantage that it is possible to carry out factorial designs which incorporate other stressors, which has been identified as an important next step in climate change experiments (Wernberg et al. 2012). These experiments require stringent attention to issues of experimental design (Jentsch et al. 2007; Wernberg et al. 2012) but have the potential to generate a much greater understanding of the interactive impacts of changing climate with other stressors.
There is a need to consider the degree to which this kind of highly controlled experiments can be scaled to large-scale real-world conditions. Previous small-scale studies have also tended to concentrate on single species, so when experimental results have not scaled to field outcomes, it is difficult to determine which of these two factors is responsible (Wernberg et al. 2012; Wolkovich et al. 2012). In plant studies, it appears that small-scale experiments may not scale up to large scales because they fail to incorporate complex community-level interactions and therefore underestimate warming impacts (Wolkovich et al. 2012). It is important to recognise the limitations of such small-scale experiments (Carpenter 1996; Underwood et al. 2005). The spatial scale of experiments has been shown to affect the magnitude of responses to treatments in a number of difference systems (see Englund & Cooper 2003 for a review). In particular, open systems that are strongly reliant on landscape-scale processes such as meta-population dynamics may respond differently to changing climate than do systems where local processes predominate (Underwood et al. 2005). Notwithstanding those concerns, manipulations at relatively small scales are likely to be the only way to explore impacts of climate change in a way which incorporates all of the features of predicated future climates (Englund & Cooper 2003). We propose that a suite of approaches including laboratory experiments, use of extreme events within traditional experimental increment studies and field studies of extreme events will be needed to gain a thorough understanding of the likely effects of future climates. Increasingly, frameworks are being suggested for how best to integrate across this suite of data (Denny & Benedetti-Cecchi 2012).