Climate-induced warming imposes a threat to north European spring ecosystems

Introduction

Water temperature is a primary determinant of freshwater ecosystem functioning and the distribution, composition and diversity of aquatic biota (e.g. Vannote & Sweeney, 1980). There is a plethora of studies showing that global warming has caused a detectable increase in surface water temperature, and the ecological effects of warming are already evident (Woodward et al., 2010; Jeppesen et al., 2012). Recently, the impacts of climate change on groundwater systems have received increasing attention and the mechanisms of climate-induced depletion of groundwater have been thoroughly reviewed (Green et al., 2011; Taylor et al., 2013; Kløve et al., 2014). Intuitively, one might expect the thermal regime of groundwater systems to have remained relatively unharmed by global warming (Luce et al., 2014) as they are often several tens or even hundreds of metres below earth's surface. However, several recent studies suggest that this is not the case and groundwater resources may be equally vulnerable to climate-induced warming as are surface waters (Kurylyk et al., 2014a,b; Menberg et al., 2014). The effects of warming are manifested more quickly in small and shallow aquifers, while larger aquifers provide a more stable thermal regime governed by geothermal processes. Despite the growing body of literature confirming climate-induced warming of groundwater resources, its impacts on ecosystem-level changes in groundwater-dependent ecosystems (GDEs) are hitherto poorly understood (Kløve et al., 2014).

Nonthermal springs provide a specific case of GDEs. They are fed by a continuous flux of groundwater that provides a unique, thermally stable and cold-water habitat for a diverse aquatic biota. Springs and their terrestrial surroundings support a high level of rarity and endemism and are therefore often considered hotspots for regional biodiversity (Cantonati et al., 2012). Springs are, however, threatened globally by multiple anthropogenic activities such as habitat loss and degradation, and groundwater abstraction and pollution (Laini et al., 2012; Glazier, 2014). Global warming is emerging as yet another threat to GDEs: the very existence of these fragile ecosystems is determined primarily by the availability of low-temperature groundwater, and even minor shifts in their thermal regime may have significant consequences for the biodiversity and functioning of spring ecosystems (Glazier, 2012; Kløve et al., 2014).

Most studies on the biodiversity impacts of climate change in freshwater ecosystems have suggested considerable changes in abundance and distribution of, for example, stream macroinvertebrates, with cold-water specialists generally suffering the greatest losses (e.g. Durance & Ormerod, 2007; Domisch et al., 2011). Nevertheless, local species richness (α diversity) is often predicted to increase substantially (Brown et al., 2007; Rosset et al., 2010), while at the same time, the between-site component of diversity (β-diversity) may be reduced. This has been observed in, for example, glacier-fed streams where the reduced meltwater contribution causes habitat homogenization and, consequently, reduction in regional (γ) diversity (Brown et al., 2007; Jacobsen et al., 2012). As these landscape-level diversity responses will be difficult to tackle experimentally, natural temperature gradients have been suggested as ‘sentinel systems’ of the ecological impacts of climate change, particularly those occurring at the community level (Woodward et al., 2010).

We studied long-term changes in water temperature of cold-water springs from 1968 to 2012 using time-series data from 66 Fennoscandian springs. Our aim was to examine whether spring water temperatures have increased during the study period and whether the warming rates are regionally coherent or driven by local environmental characteristics. We then used a multitaxon niche modelling approach to relate future changes in thermal regime to distributional patterns of spring biota (benthic invertebrates and aquatic bryophytes), and to demonstrate, based on the three alternative global warming scenarios (IPCC, 2013), the potential consequences of groundwater warming to GDE biodiversity. We specifically expected that species dependent on cold-water thermal refugia (spring specialists) would suffer a biodiversity loss as a consequence of increasing water temperature (see Glazier, 2012), while headwater generalist taxa should exhibit a much weaker response to increased temperature. As a consequence, the overall species richness of spring biota may not vary much, whereas the species composition of spring-dwelling invertebrate and bryophyte communities may change substantially through time as spring water temperatures increase.

Materials and methods

Water temperature data

We compiled water temperature data from springs in Finland and Sweden using three data sources: Finnish Groundwater Monitoring Network, Geological Survey of Finland and Swedish Groundwater Monitoring Program. Data were initially screened to retain only sites with temperature measurements available for at least 20 years, which resulted in 66 springs (20 from Finland, 46 from Sweden). At maximum, the data spanned 45 years (1968–2012), with a mean length of 35 years (Fig. S1). Study springs were mostly on crystalline bedrock, and the parent aquifers were mainly Quaternary glaciofluvial soil formations. Only a few springs in Sweden were related to karstic geology. The springs are geographically evenly distributed, and the data are therefore well representative of north-temperate, boreal and subarctic ecoregions (Fig. 1). Each spring was visited one to 12 times a year (Fig. S1) by local environmental authorities, and water temperature was measured from the spring pool or the spring outlet between 08:00 and 16:00 h using a calibrated multiuse probe or a portable thermometer.

Data analyses

Trends in water temperature were analysed using the Theil–Sen's method to estimate trend slopes, and the nonparametric Mann–Kendall test was used to determine the statistical significance of monotonic trends (Helsel & Hirsch, 2002). Annual median values calculated across all site-specific annual temperature measurements were used as the response variable. To assess the sensitivity of springs to water temperature change, we related site-specific Theil–Sen's slopes (i.e. warming rates) to site characteristics using linear regression models. Spatial autocorrelation was explored by calculating spatial correlograms and Moran's index (I). Moran's I ranges from −1 to +1 where positive values indicate strong positive autocorrelation and zero values suggest no autocorrelation or random distribution. Expected Moran's I and statistical difference between observed and expected Moran's I were derived from permutations (n = 999) using the Moran I function in the ape R package (Paradis et al., 2004).

Next, for each study year, we calculated mean values for spring and air temperatures across all study sites; only data after 1977 were used as earlier observations were mainly from southern Swedish springs, causing a bias in mean temperature estimates. We used multiple linear regressions to relate mean annual spring water temperature to regional air temperature and radiative forcing (RF, W m⁻²) (Ramaswamy et al., 2001). Due to the lack of detailed local data on soil properties and groundwater temperatures at different depths, we were constrained to using linear regression models instead of process-oriented (Taylor & Stefan, 2009) or quasi-empirical approaches (Kurylyk et al., 2013). However, similar regression models have been used successfully to forecast Central European groundwater temperatures (Figura et al., 2014). The feasibility of simple regression models for estimating future groundwater/spring water temperature is largely determined by the length (and other features) of the calibration period (see Figura et al., 2014). In our data, calibration period was 34 years and included two notable air temperature regime shifts (mid 1980s and 1990s); we thus consider our approach suitable for our study.

Annual air temperatures and RF values of the corresponding year and of the three preceding years were used as candidate predictors of spring water temperature in a stepwise forward selection procedure where the predictor variables were included in the final model if they improved model fit according to Akaike's information criterion. We then used the final model to predict future spring water temperatures (2015–2086) based on three alternative up-to-date greenhouse gas concentration trajectories provided by the Intergovernmental Panel on Climate Change (IPCC, 2013). The selected scenarios represent low (RCP2.6), intermediate (RCP6) and high (RCP8.5) emission scenarios. Air temperatures from 2015 to 2086 were generated based on average values across 28 different Coupled Model Intercomparison Project (CMIP5) climate simulation models that were produced recently by the SETUKLIM project of the Finnish Meteorological Institute (K. Ruosteenoja: Finnish Meteorological Institute, pers. comm.). Predicted changes in global mean radiative forcing were adopted from Meinshausen et al. (2011), and the RF estimates for the selected climate scenarios were obtained from http://www.pik-potsdam.de/~mmalte/rcps/.

We related the species richness of non-spring-dependent (hereafter, headwater generalists) and spring-dependent (spring specialists) bryophytes (Ulvinen et al., 2002) and macroinvertebrates (Ilmonen et al., 2009) to early summer water temperatures. We used generalized additive models (GAM; Hastie & Tibshirani, 1986) assuming Poisson error distribution to model richness response along the thermal gradient. GAM are a semi-parametric smooth maximum-likelihood extension of the generalized linear models (GLM) and are increasingly used by ecologists to describe nonlinear and nonmonotonic relationships between ecological responses and the environment (Guisan et al., 2002).

Species compositional responses to increasing water temperature were modelled using RIVPACS-type multitaxon niche predictive models. RIVPACS is a widely used statistical tool to assess the biological integrity of freshwater ecosystems, and it provides a site-specific prediction of the biota to be expected in the absence of anthropogenic stress (Moss et al., 1987). The observed number of taxa (O) is compared to expected (E) values, which provides an estimation of taxa loss (O/E) compared to reference status. Recent adaptations of the RIVPACS approach enable the calculation of a measure of Bray–Curtis compositional dissimilarity (BC index; Van Sickle, 2008) that complements O/E and summarizes the taxon-specific disparity between observed and expected assemblages, including both the loss of expected and the occurrence of unexpected taxa. Thus, BC index better considers the occurrence of rare taxa and is therefore more suitable as an assessment metric for the entire community and not only the most common taxa.

We built RIVPACS models to assess the impacts of different warming scenarios on bryophyte and macroinvertebrate species composition in 208 Finnish headwater springs and streams (see above). Prior to building the models, biological data were randomly subdivided into calibration (n = 158) and validation (n = 50) sites to calibrate the models and to verify model accuracy and precision, respectively. We developed separate RIVPACS models for macroinvertebrates and bryophytes. First, we grouped the sites according to their biological similarity with flexible-β clustering algorithm (β = 0.6) using the Jaccard dissimilarity measure for presence/absence data. We then manually truncated the dendrograms to produce biologically homogenous site groups. Next, we used random forests (RF) models (Breiman, 2001) to account for the biological clustering and to predict the probability for a site to belong to each of the biologically defined groups as a function of geographical coordinates and early summer (late May–early June) water temperature. We applied the RF models to predict the probability of cluster membership (G) for all study sites (j). We then calculated the probability of capture (p) for each taxon (i) following the standard RIVPACS. Expected number of taxa (E) was obtained by summing the product of site probabilities (G_j) and probabilities of taxa captures (p_ij) as follows:

For evaluating model goodness, we then calculated the ratio of the number of observed (O) to expected (E) taxa (O/E) for each site. Distribution of O/E ratios among the 50 validation sites was used to evaluate model performance compared to a null model where all sites are allocated to one group, yielding a single E for all sites. We calculated averages and standard deviations (SD) for the O/E ratio; a successful model should have mean O/E close to unity with low standard deviation. Species with low capture probabilities (P < 0.5) are commonly excluded because rare species tend to decrease model accuracy and precision (Van Sickle et al., 2007). In the model evaluation phase, we therefore used P ≥ 0.5 for macroinvertebrates, thus only focusing on the most common taxa; however, P ≥ 0.5 may be inappropriate for less diverse communities such as bryophytes (Jyväsjärvi et al., 2011). Thus, a lower P value (0.35) was applied for bryophytes.

RIVPACS models were then used to predict future species composition in the 208 study sites by holding geographical coordinates constant but modifying water temperature to mimic temperature increase by 1–4 °C (in increments of 1.0 °C). BC indices (see above) were calculated for each site from the output of the RIVPACS models as follows:

where O is the observed occurrence and p is the predicted probability of occurrence of taxon k. Here, also the rarest taxa (p < 0.5 or <0.35; see above) were included. Differences between the present and modelled (future) BC index were plotted against current water temperature to evaluate the intensity of species compositional change along the thermal gradient. All statistical analyses were performed using R program (R Core Team, 2014).

Results

A great majority of springs (91%) showed an increasing trend in water temperature (Fig. 1), and 82% of these trends were statistically significant (P < 0.05) (see Fig. S2 for site-specific trends). Theil–Sen's slope values varied from −0.05 to 0.16, the mean warming rate across all sites being 0.034 °C per year. Average Theil–Sen′s slope for air temperature data was slightly higher (0.050). Moran's spatial autocorrelation (I) for the Theil–Sen's value was 0.156, differing significantly (P = 0.004) from expected (−0.015). This suggests modest spatial autocorrelation in the data. However, no distinct spatial pattern was detectable in the distribution of the Theil–Sen's slope values and the slopes varied irrespective of, for example, latitude (R² = 0.004, P = 0.588). Neither were the warming rates related to spring discharge (R² = 0.009, P = 0.541), suggesting that springs with both low and high discharge are equally vulnerable to warming. This implies that warming has occurred in the aquifer, not only in the spring pool. There was, however, a negative relationship between the warming rate and surface area of the parent aquifer (R² = 0.178, P = 0.007; Fig. 2).

Average spring water temperature increased from 4.70 to 5.94 °C between 1978 and 2012 (Fig. 3a). Regional air temperature and radiative forcing (RF) of the respective year contributed significantly (P < 0.001) to mean spring temperature, together explaining 74.2% of water temperature variation (Fig. 3a), suggesting that warming was intimately related to prevailing climatic conditions. Air temperature or RF of the three preceding years did not explain residual variation. We next used the selected regression model to predict future (2015–2086) spring water temperatures using three alternative greenhouse gas concentration trajectories (RCP2.6, RCP6 and RCP8.5). Our model predicted that, given the projected increases in RF and regional air temperature in each of the three scenarios, the increase in mean spring water temperature is likely to range from 0.67 °C (RCP2.6) to 5.94 °C (RCP8.5) by 2086 (Fig. 3b).

According to GAM, species richness of both headwater generalist bryophytes (Fig. 4a) and macroinvertebrates (Fig. 4b) increased with water temperature, with a much stronger pattern for invertebrates (R² = 0.507, F = 50.14, P < 0.001) than bryophytes (R² = 0.08, F = 5.99, P < 0.001). In contrast, species richness of spring specialists decreased significantly (bryophytes: R² = 0.238, F = 10.86, P < 0.001; macroinvertebrates: R² = 0.457, F = 22.39, P < 0.001), the rate of taxa loss greatly intensifying as water temperature exceeds 6 °C (Fig. 4c,d). Such a reduction in cold-stenothermic specialist species is likely to induce notable changes in community composition, and this was indeed shown by the RIVPACS-type multitaxon niche modelling. Flexible-β clustering divided calibration sites into six groups for both organism groups, and the number of sites per cluster varied between 12 and 46 for bryophytes and between 7 and 35 for macroinvertebrates. The percentage of calibration sites correctly classified by random forest models was 77.2% for macroinvertebrates and 65.8% for bryophytes, indicating successful classification for both groups. Our models outperformed the null model in terms of accuracy and precision in predicting the number of expected taxa among validation sites (Table S1). The models showed that an increase of 1 °C will alter community composition in spring-fed and other groundwater-dependent streams with low (<6 °C) current water temperature, but not (or much less so) in sites with less groundwater input and thus higher current water temperature (Fig. 5). A further increase in temperature intensifies community change at the lower end of the thermal range, whereas warmer streams remain practically unaltered (Fig. 5).

Discussion

We used extensive time-series data from northern European springs to show that spring water temperatures have increased consistently during the last four decades. Warming was closely related to prevailing climatic conditions and, given the future climate projections, this increase seems unavoidable in the forthcoming decades. To the best of our knowledge, this is the first study that uses spatially and temporarily extensive field data to show that groundwater-dependent ecosystems are vulnerable to climate-induced warming. The few previous studies that have suggested groundwater temperatures to be increasing in response to climate change have been based on either modelling (Taylor & Stefan, 2009; Kurylyk et al., 2014b) or empirical data from only a few monitoring wells (Figura et al., 2011; Menberg et al., 2014).

Most springs in our data responded rapidly to changes in local climatic conditions. This is due likely to the fact that the aquifers were mostly unconfined glaciofluvial deposits, being relatively small and shallow. Such aquifers are common in the Northern Hemisphere (Banerjee & McDonald, 1975) where also higher-than-average increase in air temperatures is likely to occur in future (IPCC, 2013). Therefore, similar intensive warming of springs has likely occurred in other northern regions as well. The few karst springs in Sweden showed no or weak response to climate. Moreover, our results suggest that springs fed by aquifers with a small recharge area and low storage volume are more prone to climate-induced warming than those associated with larger aquifers. This observation agrees with recent modelling findings that small and shallow groundwater systems are particularly vulnerable to climate-induced warming (Taylor & Stefan, 2009; Kurylyk et al., 2014b).

The predicted changes in mean spring water temperature during the next 70 years based on the intermediate (RCP6) or high-emission (RCP8.5) scenarios were alarming (3.4 and 5.94 °C, respectively), whereas the low-emission scenario (RCP2.6) results in a markedly lower temperature change (0.67 °C). However, although technically accessible (van Vuuren et al., 2011), achieving the emission goals of RCP2.6 requires an unrealistically rapid reduction in global emissions, and the predictions based on the two higher emission scenarios therefore seem more realistic. Repercussions of such a rapid warming of spring water temperature may be more deleterious for GDEs than for other aquatic ecosystems. The functioning of nonthermal spring ecosystems is premised on the availability of cold water, and the spring biota is composed primarily of cold-stenothermic species with strictly defined thermal optima (Glazier, 1991; Barquín & Scarsbrook, 2008). Thus, even slight changes in spring water temperature may trigger local (or regional) extinction of spring-dependent species and alter species composition of spring assemblages. This was indeed shown by our multitaxon niche modelling which indicated that water temperatures exceeding 6–7 °C may induce a significant loss of endemic spring specialists, thus altering spring assemblage composition. Importantly, species data for these models come from the Finnish part of the study area and extrapolating biological modelling results to other regions must be done with great caution.

Our results parallel studies from glacier-fed streams where the climate-induced reduction in meltwater contribution has homogenized habitats and reduced regional (γ) macroinvertebrate diversity (Brown et al., 2007; Jacobsen et al., 2012). The few experimental studies conducted in cold-water springs also largely concur with our results. Tixier et al. (2009) investigated the responses of nonbiting midges (Chironomidae) to increased water temperature in a small spring system in southern Ontario. They noticed a significant reduction in chironomid abundance, alteration of community composition and a markedly (though nonsignificantly) lower species richness in a heated half of the spring. Using the same experimental set-up, Bärlocher et al. (2008) noticed that also fungal diversity was lower in the heated half of the spring. Historical data have also revealed climate-induced changes in species occurrence and distribution patterns. For example, Giersch et al. (2015) reported a marked reduction in the distributional range of a rare, cold-stenothermic stonefly in the Glacier National Park, Montana, USA. They suggested that the range reduction has been caused by increased water temperature and decreased glacial masses over the past few decades.

Even after considerable warming, the coldest, groundwater-fed headwater streams likely remain uninhabitable for the majority of headwater generalists. Furthermore, spring-fed streams represent the upmost headwater sites within river networks; they are therefore weakly connected to lower stream reaches and likely to be dispersal-limited (Brown & Swan, 2010). In this regard, warming of north European springs entails only few climate change ‘winners’ but many ‘losers’, opposite to that suggested for alpine freshwater ponds (Rosset & Oertli, 2011) and Central European mountain rivers (Domisch et al., 2011). Accordingly, while some freshwater studies have predicted climate warming to induce a general, and sometimes quite substantial, increase in both regional and local species richness (Rosset et al., 2010), the regional diversity of north European spring biota may not respond very strongly. Nevertheless, change to assemblage composition is likely to be extensive. In this regard, our results fit nicely in the general framework of Dornelas et al. (2014) who showed that a great majority of the 100 time series summarized by them exhibited a significant biodiversity change but not systematic biodiversity loss. They further emphasized that such consistent changes to assemblage composition may result in the formation of novel ecosystems that no longer provide the same services as they did before. Indeed, climate-induced warming of groundwater may not only reduce spring biodiversity, particularly that of endemic species, but also impair the functioning and ecosystem services of GDEs. For example, groundwater warming may affect the solubility and behaviour of, for example, pesticides (Bloomfield et al., 2006) and nutrients (Stuart et al., 2011). Warming also accelerates microbial activity and organic matter decomposition (Bärlocher et al., 2008; Friberg et al., 2009), and reduces oxygen concentration and alters redox conditions (Figura et al., 2014). The ecological impacts of increased spring water temperature may also extend across the land–water interface. Springs are important ecotones that link terrestrial and aquatic ecosystems, and groundwater and surface waters, in a four-dimensional framework (Cantonati et al., 2012). Therefore, any major changes to spring water temperature may have wide-ranging impacts on riparian food webs. Spring-fed streams also provide cold-water thermal refugia for endangered fish species during summer months, and rising groundwater temperatures have raised concern about the availability of such refugia in the future (e.g. Kurylyk et al., 2014c).

Our study provides strong empirical evidence that climate change is warming groundwater and associated ecosystems in the Northern Hemisphere. GDEs are currently threatened by various anthropogenic stressors that have already caused significant loss of biodiversity, and climate-induced warming is likely to intensify the biological impairment of GDEs. The independent and interactive impacts of these multiple, simultaneously operating stressors on ecosystem services pose a significant challenge to groundwater management. The projected loss of endemic spring specialists inevitably leads to regional-scale homogenization of biological assemblages in north European headwater streams, thus posing a serious conservation challenge and a need to manage groundwater and freshwater resources in an integrated effort.