Climate changes affect aquatic ecosystems by altering temperatures and
precipitation patterns, and the rear edges of the distributions of cold-water
species are especially sensitive to these effects. The main goal of this
study was to predict in detail how changes in air temperature and
precipitation will affect streamflow, the thermal habitat of a cold-water
fish (the brown trout,
Water temperatures are a primary influence on the physical, chemical and biological processes in rivers and streams (Caissie, 2006; Webb et al., 2008) and, subsequently, the organisms that live completely or partially in the water. Temperature is a major feature of the ecological niche of poikilothermic species (e.g. Magnuson and Destasio, 1997; Angilletta, 2009) and a key factor in energy balance of fish. It affects the rate of food intake, metabolic rate and growth performance (Forseth et al., 2009; Elliott and Elliott, 2010; Elliott and Allonby, 2013). It is also involved in many other physiological functions, such as blood function and reproductive maturation (Jeffries et al., 2012), reproductive timing (Warren et al., 2012), gametogenesis (Lahnsteiner and Leitner, 2013), cardiac function (Vornanen et al., 2014), gene expression (White et al., 2012; Meshcheryakova et al., 2016), ecological relationships (Hein et al., 2013; Fey and Herren, 2014), and fish behaviour (Colchen et al., 2017).
Natural patterns of water temperature and streamflow are profoundly linked with climatic variables (Caissie, 2006; Webb et al., 2008). Therefore, stream temperature is strongly correlated with air temperature (Mohseni and Stefan, 1999), whereas streamflow has a complex relationship with precipitation (McCuen, 1998; Gordon et al., 2004). In addition, atmospheric temperature influences the type of precipitation (rain or snow) that occurs and the occurrence of snowmelt; conversely, river discharge is also a main explanatory factor of water temperature for some river systems (Neumann et al., 2003; van Vliet et al., 2011). Furthermore, geology affects surface water temperatures by means of groundwater discharge (Caissie, 2006; Loinaz et al., 2013), influenced by the aquifer depth (shallow or deep) and the water's residence time (Kurylyk et al., 2013; Snyder et al., 2015).
Climate change is already affecting aquatic ecosystems by altering water
temperatures and precipitation patterns. Stream temperature increases have
been documented over the last several decades over the whole globe, such as
in Europe (e.g. Orr et al., 2015, documented a mean increase in stream
temperature of 0.03
The rear edge populations (sensu Hampe and Petit, 2005:
“populations residing at the current low-latitude margins of
species' distribution ranges”) of a cold-water species are especially
sensitive to changes in water temperature, in addition to reductions in the
available habitable volume (i.e. streamflow). The rear edge is the eroding
margin of the range where lineages mix, the genetic drift and local
adaptations increase, and droughts put populations under stress. The impact
of water temperatures on the distribution of salmonid fish is well documented
(e.g. Beer and Anderson, 2013; Eby et al., 2014); however, the combined
effects of rising stream temperatures and reductions in streamflow remain
relatively unexamined, with some exceptions (e.g. Wenger et al., 2011;
Muñoz-Mas et al., 2016). Jonsson and Jonsson (2009) predicted that the
expected effects of climate change on water temperatures and streamflow will
have implications for the migration, ontogeny, growth and life-history traits
of Atlantic salmon,
Logical framework of the study.
The objective of this study is to predict how and to what extent the availability of suitable habitat for the brown trout, a sensitive cold-water species, will change within its current natural distribution under the new climate scenarios through a study of changes in streamflow and temperature and their interactions. Specifically, in this paper, we (i) assessed the effects of both streamflow and geology on stream temperature; (ii) predicted the changes in streamflow and stream temperature implied by the climate change scenarios used in the 5th Assessment Report (AR5) of the IPCC; and (iii) assessed the expected effects of these changes on trout habitat aptitude. To this end, hydrologic simulations with M5 model trees coupled with non-linear water temperature models at the daily time step were fed with high-resolution, downscaled versions of the air temperature and precipitation fields predicted using the most recent climate change scenarios (IPCC, 2013). The effects of basin geology on the stream temperature models and on the estimated changes in thermal regimes were studied. Finally, the changes in the thermal habitat of trout were assessed by studying the violation of the tolerable temperature thresholds of the brown trout.
The logical framework followed is summarized in Fig. 1. First, the daily global climate models output presented by the IPCC were downscaled to the study area. Then, the obtained local climate models output were applied to generate simulations of streamflow and water temperature. The results are daily values that can be used for the assessment of fish habitat suitability and availability.
River network and location of the study sites (water temperature data loggers), with details regarding lithology. The grid depicts the actual occurrence of brown trout in Spain.
The procedure yielded results in the form of continuous time series, but they are presented for two time horizons: the year 2050 (H-2050) and the year 2099 (H-2099). The values for these horizons correspond to the average of the values of the different variables for the decades 2041–2050 and 2090–2099, respectively.
Description of the data logger (thermograph) sites, specifying given name, UTM coordinates (Europe WGS89), altitude (metres above sea level), code of the nearest temperature meteorological station with suitable time series for this study (AEMET: Spanish Meteorological Agency), orthogonal distance between the data logger and the meteorological station, number of recorded days for stream temperature and characteristic geological nature (lithology) of the data logger site (the last of which was obtained from IGME, 2015). Bold letters indicate sites associated with the gauging stations.
In total, 31 sites in 14 mountain rivers and streams inhabited by brown
trout were chosen with the aim of encompassing a diverse array of geological
and hydrological conditions in the centre of Spain (between the latitudes of
39
River regime patterns for the different gauging stations. The flows are expressed as percentages of the mean annual flow, and the months (horizontal axis) are ordered from January to December.
The land cover type is mainly pine forest in all of the studied basins
(
Official stations used (meteorological and hydrological), variables, length of time series used and geographical position. AEMET: Spanish Meteorological Agency; CHD: Water Administration of Duero Basin; CHT: Water Administration of Tagus Basin; and CHJ: Water Administration of Júcar Basin.
Hydrological data characterize the streamflow regimes as extreme winter/early spring (groups 13 and 14 in the classification of Haines et al., 1988). However, the hydrographs show a west-to-east smoothing gradient (Fig. 3). This smoothing is associated with the carbonate rocks, whereas greater seasonality is associated with the igneous and detrital geological materials.
At each study site, water temperatures were recorded every 2 h throughout the year using 31 Hobo® Water Temperature Pro v2 (Onset®) and Vemco® Minilog data loggers located at several sites along the studied rivers and streams (Table 1). Loggers were tested for malfunctions before being deployed, and they were placed in areas not exposed to direct sunshine (Stamp et al., 2014). Meteorological data were obtained from nine thermometric and 15 pluviometric stations of the Spanish Meteorological Agency (AEMET) network, and data from 10 gauging stations (from the official network of the Spanish Water Administration) were obtained to model the streamflows. The AEMET thermometric stations that lie closest to the stream temperature monitoring sites and have at least 30 years of data between 1955 and the present were selected. The selected pluviometric stations were those located within the upstream river basin or near the corresponding gauging station (Table 2). The air temperature and precipitation data from AEMET were tested to assess their reliability by applying a homogeneity test. This test is based on a two-sample Kolmogorov–Smirnov test, and it marks years as possibly containing inhomogeneous data. In the second phase, the marked years are matched against the distribution of the entire series to determine if they contain true inhomogeneities, searching for possible dissimilarities between the empirical distribution functions. Only reliable series were used. The locations of the stations did not change in the studied period.
Data from nine global climate models associated with the 5th Coupled
Model Intercomparison Project (CMIP5) were used, namely BCC-CSM1-1, CanESM2,
CNRM-CM5, GFDL-ESM2 M, HADGEM2-CC, MIROC-ESM-CHEM, MPI-ESM-MR, MRI-CGCM3,
and NorESM1-M (Santiago et al., 2016). These models provided daily data to
simulate future climate changes corresponding to the Representative
Concentration Pathways RCP4.5 (a stable scenario) and RCP8.5 (a scenario
including a pronounced increase in CO
Pourmokhtarian et al. (2016) note the importance of the use of fine downscaling
techniques. Thus, a two-step analogue statistical method (Ribalaygua et al., 2013)
was used to downscale the daily climatic data, specifically the maximum and
minimum air temperatures and the precipitation for each station and for each
day. For both air temperature and precipitation, the procedure begins with
an analogue stratification (Zorita and von Storch, 1999) in which the
A systematic error is obtained when comparing the simulated data from the climate models with the observed data. Such errors are inherently associated with all downscaling methodologies and climate models, which usually introduce bias into their outputs. To eliminate this systematic error, the future climate projections were corrected according to a parametric quantile–quantile method (Monjo et al., 2014), which was performed by comparing the observed and simulated empirical cumulative distribution functions (ECDFs) and linking them using ECDFs obtained from the downscaled European Centre for Medium-Range Weather Forecasts ERA-40 reanalysis daily data (Uppala et al., 2005).
As a result, for each climate change scenario, the daily maximum and minimum air temperatures (which were used to infer the mean air temperature) and precipitation were obtained for each climate model, and the whole dataset were used as inputs to simulate the runoff and water temperatures under these climate change scenarios.
Although process-based physical models are considered the standard hydrological models, flexible data-driven machine learning techniques are gaining popularity because they can be based solely on precipitation and temperature (Shortridge et al., 2016) and can be automatized to perform multiple simulations. Therefore, the prediction of the streamflows under the climate change scenarios was performed with data-driven hydrological models developed using the M5 algorithm (Quinlan, 1992). M5 has been shown to have skill in modelling daily streamflow (Solomatine and Dulal, 2003; Taghi Sattari et al., 2013), including in studies involving climate change (Muñoz-Mas et al., 2016), and it is sufficiently fast to deal proficiently with larger datasets (Quinlan, 2017)
Mathematically, M5 is a kind of decision tree that, instead of assigning a single value to each terminal node, assigns a multi-linear regression model (Quinlan, 1992). Therefore, the dataset is hierarchically divided into homogeneous parts and a multi-linear model is adjusted to every part (Hettiarachchi et al., 2005). In this regard, each node, and the corresponding multi-linear regression model, is specialized in particular areas of the data set, such as peak flows or base flows, to name the extremes (i.e. it is a piece-wise linear model with each part dedicated to a particular hydrologic condition) (Taghi Sattari et al., 2013). Based on the multi-linear models at the terminal nodes, M5 allows extrapolation, in contrast with other machine learning techniques that have demonstrated little or no extrapolation ability (e.g. random forest or multilayer perceptron) (Hettiarachchi et al., 2005; Shortridge et al., 2016).
The M5 hydrological models were developed in R (R Core Team, 2015) with the
Following previous studies, the M5 hydrological models were trained by employing the daily, monthly and quarterly data lags of historical precipitation and air temperature collected at meteorological stations within or nearby the target river basins as input variables (Table 2) (Solomatine and Dulal, 2003; Taghi Sattari et al., 2013; Muñoz-Mas et al., 2016). These three groups of variables were intended to reflect the causes of peak, normal and base flows. The study encompassed several rivers and streams that may have different hydrologic behaviours; therefore, the starting set of input variables, which was afterwards subset, was larger than that used in other studies (Solomatine and Dulal, 2003; Taghi Sattari et al., 2013; Muñoz-Mas et al., 2016). The daily variables included the precipitation and air temperature from the current day to the 15th previous day (16 variables in total). The monthly variables were calculated using the moving average for the 12 previous months (12 variables in total), and the quarterly data were calculated from the moving average for the current month to the 24th previous month (8 variables in total). Consequently, the daily variables overlapped with the current month variable, and the first four quarterly variables overlapped with the monthly data. In the end, 72 variables were gathered, 36 each for air temperature and precipitation.
The whole set of input variables may be relevant for some river systems
(Shortridge et al., 2016), although it may cause M5 to overfit the data in
others (Schoups et al., 2008). Therefore, the ultimate variable subset was
optimized following the forward stepwise approach (Kittler, 1978). This
approach relies on iteratively adding input variables (one at a time) while
the performance on the test data set improves and stopping (i.e. selecting a
smaller subset of the input variables) as soon as the performance stagnates
or degrades. However, the classical forward stepwise approach may cause
consideration of unrelated variable sets (i.e. disjoint precipitation and air
temperature variable lags). To address such potential inconsistencies, the
optimization began by testing the precipitation-related variables and only
tested the air temperature variables for lags coinciding with those
precipitation-related variables that were already selected. No precautions
were taken regarding correlations among inputs (D. P. Solomatine, personal
communication, 2016), and the forward stepwise approach sought to maximize
the Nash–Sutcliffe efficiency (NSE) index (which ranges from
The daily data were analysed monthly and seasonally using the following
statistics: minimum flow (
To assess the significance of the streamflow trends throughout the century,
Sen's slope was used (as implemented in the
Finally, the variation in the patterns of the monthly mean streamflow was
studied by means of the Ward hierarchical clustering implemented in the
Stream temperature (
A blockwise non-parametric bootstrap regression (Liu and Singh, 1992) was used to estimate the parameters of both the modified Mohseni models (with and without streamflow), and residual normality and non-autocorrelation were checked with the Shapiro test and Durbin–Watson test. Moreover, the 7-day lag PACF (partial autocorrelation function) was obtained. These calculations were performed using R. A 95 % confidence interval was calculated for each parameter. Performance was quantified using two indicators: the residual standard error (RSE) and the Nash–Sutcliffe efficiency index (NSE). The Bayesian information criterion (BIC) and the Akaike information criterion (AIC) were used to test the eight-parameter models (Terms 1 and 2 of Eq. 1) against the five-parameter models (Term 1 of Eq. 1).
This model can be classified as semi-physically based. It has some advantages over machine learning methods, such as classification and regression trees (De'ath and Fabricius, 2000) or random forests (Breiman, 2001), because the model parameters imply a mechanistic interpretation of how process drivers act, yielding a higher transferability (Wenger and Olden, 2012). These features make of this model an advantageous option for our goals.
Different classes of thermal thresholds for emerged trout classes found in the literature. The type of experiment differentiates the experiments with controlled (laboratory) and uncontrolled (wild) temperature. Latitude of the experiments' location is shown.
Geology determines the residence time of deep groundwater in the aquifers
underlying streams (Chilton, 1996), and residence times influence discharge
temperatures. To explore the relationships between thermal regimes and
geology, a stratified study of both the geology classes of the parameter
values was completed by means of a
The variation in the patterns of the monthly mean stream temperature was
studied by means of cluster analysis of the temperature increases
corresponding to H-2050 and H-2099 for the RCP4.5 and RCP8.5 scenarios
(using Ward's hierarchical clustering as implemented in the
Several tolerance temperatures and thermal niche limits have been described
for brown trout (Table 3). The realized niche must reflect energetic
efficiency: spending long periods above that threshold makes animals less
efficient competitors, and their performance decreases critically (Magnuson
et al., 1979; Verberk et al., 2016). Thus, we focused our study on the realized
thermal niche. The elected threshold for this study was the occurrence of
DMST values above 18.7
Once DMST was modelled, the frequency of events of seven or more consecutive
days above the threshold per year (times above the threshold, TAT
To assess the general trend in thermal habitat alterations at the middle
(H-2050) and the end of the century (H-2099), the TAT
The number of sampling sites and their distribution in the Cega, Pirón
and Lozoya rivers (Fig. 2, Table 1) permit the longitudinal interpolation
and extrapolation of the predicted water temperatures to study the
relationships between the annual average DMST and altitude (strong
correlations were detected between these quantities;
All variables and abbreviations are summarized in the Appendix A. An overview of the uncertainty issue is given in Appendix B.
Under the climate change scenarios, all the meteorological stations will
experience noticeable temperature (DMAT) increases through the century. As
might be expected, this trend is steeper for the RCP8.5 scenario, especially
in summer, though it is also noticeable in winter to a lesser extent (annual
trends are shown in Fig. 4; the seasonal results are shown by location in
Figs. S1 to S24 in the Supplement). The air temperature variations
will run parallel to one another in the two scenarios until mid-century, when
the RCP8.5 scenario predicts a similar trend and the increases decrease
under the RCP4.5 scenario; the annual change in temperatures for RCP4.5
fluctuates between 2 and 2.5
The change in the annual precipitation (mm day
Changes in mean air temperature and total annual precipitation related to climate change for the nine general climate models and the two climate change scenarios for the all the studied meteorological stations.
Predicted monthly mean specific flow in H-2050 and H-2099 for the
RCP4.5 and RCP8.5 scenarios. Shaded areas indicate decadal fluctuations.
Triangles show significant negative or positive trends (Sen's slope
In general, decreases in flow will occur throughout the century, but the degree of change will vary among the sites. Stations located in the western (Tormes) and eastern (Ebrón) extremes of the study area will experience an increase in flow by 2099 after decreasing in the mid-21st century. Lozoya will suffer the most intense flow decreases, followed by Pirón and Cega-Lastras, Tagus and Gallo, and Cabrillas. These patterns of change in flow regimes are predicted to be linked to a west-to-east longitudinal gradient; climate change is expected to have less of an influence on discharge at the western stations and Ebrón (in the far eastern portion of the study area).
The hydrological models performed well; all of them achieved NSE values
Seasonal significant changes of flow variables in percentage (DJF: winter; MAM: spring; JJA: summer; SON: autumn) in H-2050 and H-2099, and RCP4.5 and RCP8.5 scenarios.
Statistically significant (
According to the predictions, the most significant changes in flow regimes will occur at the gauging stations of Cega-Lastras and Lozoya in H-2050 (Table 4, Fig. 5). In H-2099, most sites will experience strong flow reductions, even in seasons where seasonal increases in flow are predicted (e.g. Ebrón and both stations in the Tormes River) (Table 4, Fig. 5). Significant annual runoff reductions in H-2050 will occur at five of the stations, increasing the occurrence of significant losses at 9 out of the 10 sites in H-2099 (i.e. all stations except Ebrón). The most important decreases in every variable and throughout the century were predicted for the stations in the middle Cega Basin and the Tagus Basin. A significant increase in the number of days with no flow was predicted for Cega-Lastras, Pirón and Gallo.
The cluster analysis of gauging stations based on seasonal variations in the
flow regime revealed the importance of careful examinations at the local
level, since hydrological behaviour is a consequence of both macroclimatic
and mesoclimatic conditions. A geographical pattern is recognizable when the
actual flow regime (2006–2015) is seasonally clustered (Fig. 6). Analysing
the deviations in this geographical pattern by scenarios and horizons, the
different gauging stations can grouped according to the seasonal behaviour
of the flow changes (Fig. 7a). For the RCP4.5 scenario in H-2050
(agglomerative coefficients, a.c.
Gauging stations clustered by the current normalized seasonal
streamflow regime (agglomerative coefficient, a.c.
Study sites clustered by the predicted change ratios of the seasonal
mean streamflow (gauging stations) and by the predicted increase in the
monthly mean stream temperature (
The inclusion of the streamflow component improves model performance at 12 out of the 28 study sites (Table 5). In the remaining 16 cases, either no convergence of values was observed in the regression process or the obtained values did not improve the results, as the streamflow component (Term 2 of the equation) is virtually zero at the other sites. The five-parameter model was used in these remaining 16 cases. The calculated parameters and the performance indicators (RSE and NSE) of the models are shown in Table 6, and daily mean stream temperatures estimated by the climate change models are given in the Supplement (Dataset S3). Performance was high in all the cases except in Pirón 5 where NSE was low.
For the entire array of sites (
By the end of the 21st century, the predicted average increase in the mean
annual stream temperature among the sites is 1.1
Bayesian (BIC) and Akaike (AIC) information criteria values for the stream-temperature models with five and eight parameters.
Parameter values of the stream temperature models for every
thermograph site (
The values of the model parameters showed different behaviours depending on
the lithology found in each basin, which thus influences the thermal response
to climate change. The thermal amplitude is greater at sites underlain by
igneous bedrock (
Among the eight-parameter models (
Under the RCP4.5 scenario,
Distributions of the stream temperature model parameter values
(
Distributions of
The results of the cluster analysis of the monthly mean stream temperatures
revealed a highly homogeneous aggregation of sites for the different
combinations of horizons and scenarios, given that the thermal responses of
the rivers and streams are tightly linked with lithology (Fig. 7b). The
carbonate sites from the Cabrillas stream (in the east) and Pirón 3
(which is strongly influenced by a calcareous spring) form a group of sites
that shows low thermal amplitude and in which
The predicted flow reductions lead to notable increases in water
temperature. The effect of streamflow variation on stream temperature is
analysed at the following sites: Tormes 2, Tormes 3, Pirón 1, Cega 1,
Lozoya 1 to 4, Cabrillas, Ebrón 1 and Vallanca 1 and 2. These are the
sites at which the eight-parameter model improves upon the five-parameter model. In
all cases, differences in stream temperature between the five- and eight-parameter
models are found, and summer flow reductions lead to increases in stream
temperature, increasing DAT, TAT
Increases in TAT
For all of the sites at which the influence of streamflow on stream
temperature was revealed, the eight-parameter model estimates higher values of
maximum annual DMST than the five-parameter model. The maximum annual DMST
calculated by the eight-parameter model is 3.6
Maximum daily mean stream temperature (
The length of the thermal habitat of trout will undergo important reductions
due to the rises in water temperatures and the increase in the extent of the
warm period. In the predictions for H-2050, the 18.7
By the end of the century (H-2099), the most notable increases in TAT
Continuous modelling of water temperature by means of the interpolation of model parameters along the Cega, Pirón and Lozoya rivers and the application of the model to DEM data predicts relevant losses of thermal habitat, which will affect up to 56, 11 and 66 % of the lengths of these streams, respectively. In the Cega and Pirón rivers, the habitat loss is expressed relative to the proportion of total stream length where trout currently dwell (98 and 77 km in the Cega and Pirón streams, respectively). In the Lozoya River, the loss is predicted to occur in the reach (20 km) immediately upstream of a large reservoir (the Pinilla reservoir), which produces a total disconnection of the stream. The losses in maximum usable habitat will shift the current downstream limit of the trout distribution from 820 up to 831 m a.s.l in the Pirón River, from 730 up to 830 m a.s.l. in the Cega River, and from 1090 up to 1276 m a.s.l. in the Lozoya River. In the particular case of the Cega River, a window of usable thermal habitat is also predicted to occur upstream from this altitudinal range (from 913 up to 1050 m a.s.l.).
Our downscaled results predict greater air temperature increments than the original IPCC (2013) results. These higher temperatures may lead to increased ecological impacts (Magnuson and Destasio, 1997; Angilletta, 2009) caused by the combination of rising water temperatures and decreasing stream flows. The results from the AR5 of the IPCC and its annex, the Atlas of Global and Regional Climate Projections (IPCC, 2013), suggest that droughts are unlikely to increase in the near future for the Mediterranean area. However, air temperatures are expected to rise, subsequently increasing evapotranspiration. As a consequence, the available water in rivers and streams will be reduced. Regional studies have used coarser resolutions than ours, which may be appropriate for their goals (e.g. Thuiller et al., 2006). However, they may be insufficient when more local predictions are needed, as does our study, which treats geographically confined, stream-dwelling trout populations. Therefore, fine downscaling techniques like those applied in this study must be used when high-resolution, detailed predictions are needed.
This study predicts significant but diverse streamflow reductions during the
present century. At the regional level, a reduction in water resources is
expected in the Mediterranean area (IPCC, 2013). Milly et al. (2005) predicted a
10–30 % decrease in runoff in southern Europe in 2050. In another
global-scale study, van Vliet et al. (2013) predicted a decrease in the mean flows
of greater than 25 % in the Iberian Peninsula area by the end of the
century (2071–2100), using averages for both the SRES A2 and B1 scenarios
(Nakicenovic et al., 2000). Our results predict mean flows that are similar to
that value (
More specifically, the predictions for the RCP4.5 scenario show flow reductions that range from negligibly small to significant (up to 17 %). Under the RCP8.5 scenario, significant reductions become more widespread, ranging up to 49 % of the annual streamflow losses. Our results also predict a relevant increase in the number of days with zero flow for some stations in the detrital area under this scenario (RCP8.5). The predicted streamflow changes are compatible with those obtained in previous studies, although these studies were performed at larger scales (Milly et al., 2005; van Vliet et al., 2013). The apparent differences between the streamflow reductions estimated in this study and those obtained by Milly et al. (2005) and van Vliet et al. (2013) (who report lower flow reductions than those given in the present study) might be caused by the regional focus of their predictions (the entire Iberian Peninsula), whereas ours are focused on mountain reaches.
In terms of methods, process-based hydrological models are often preferred for climate change studies (Van Vliet et al., 2012). However, they can be overly complicated and require excessive data inputs, which may also lead to over-fitting of the data (Zhuo et al., 2015). Constraining further predictions to within the training domain is a rule of thumb for machine learning studies (Fielding, 1999), although extrapolation is rather common (Elith and Leathwick, 2009). Therefore, taking into account the extrapolation that occurs towards lower flows, which are over-represented in the training dataset, we consider the magnitude of the extrapolation acceptable, and we consider the values, although they are not exempt from uncertainty, to be reliable.
The model we present in this study showed good performance. Bustillo et al. (2013)
recommended the assessment of the impacts of climate change on river
temperatures using regression-based methods like ours that rely on logistic
approximations of
However, we also sought to identify relationships between thermal regime and
other environmental variables besides air temperature and streamflow, such
as geology. Bogan et al. (2003) showed that water temperatures were uniquely
controlled by climate in only 26 % of 596 studied stream reaches.
Groundwater, wastewater and reservoir releases influenced water temperatures
in the remaining 74 % of the cases. Loinaz et al. (2013) quantified the
influence of groundwater discharge on temperature variations in the Silver
Creek Basin (Idaho, USA), and they concluded that a 10 % reduction in
groundwater flow can cause increases of over 0.3 and
1.5
A wide range of models is described in the literature, and each such model has its strengths and weaknesses. Arismendi et al. (2014) hold that regression models based on air temperature can be inadequate for projecting future stream temperatures because they are only surrogates for air temperature, whereas Piccolroaz et al. (2016) argued that the adequacy depends on the hydrological regime, type of model and the timescale analysis. Their main objections to regressive methods arose when modelling reaches of regulated rivers, but this is not our case. In addition, our model improves the models that were tested in both studies (Arismendi et al., 2014; Piccolroaz et al., 2016). Performance indicators of our models produce good results, showing that the models are sufficiently competent. We show that our model implicitly integrates the effect of other factors, such as geology and flow regime by means of its parameters. A fine mechanistic solution to the modelling issue could need prohibitive methods (Kurylyk et al., 2015), losing the advantages that make attractive the model (input data easy to get). Therefore a compromise between improved precision and increased cost must be met.
The behaviour and dynamics of the parameters offer a promising research field. Their analysis may help to introduce new parametrization criteria to avoid the risk of ignoring the effect of climate warming on groundwater (subsurface water and deep water), for instance. The thermal sensitivity of shallow groundwater differs between short-term (e.g. seasonal) and long-term (e.g. multi-decadal) time horizons, and the relationship between air and water temperatures does not necessarily reflect this difference. This variability should be taken into account in order to avoid underestimating the effects of climate warming (Kurylyk et al., 2015).
Regression models are substantially site-specific compared to deterministic approaches (Arismendi et al., 2014). However, the parameters of these regression approaches are still physically meaningful, and these models require fewer variables that can limit the applicability of more complex models in areas where data are scarce. Consequently, the value of this type of model is its applicability to a large number of sites where the only available data describe air temperatures (and precipitation and streamflow to a lesser extent). On the other hand, our results show that predictions can improve when streamflow is included in the water temperature model, although some streams show little or no sensitivity to the introduction of streamflow into the model. However, the lack of sensitivity is not necessarily due to the absence of the influence of flow on the water temperature but rather to its minor relevance compared to other sources of noise. Thus, when flow data are available, it may be recommended to use the more complex eight-parameter model to predict the effects of climate warming. This conclusion is especially applicable to lithologically sensitive basins, such as those underlain by igneous rocks.
The predicted increase in water temperature will be substantial at most of
the study sites. The annual mean rates of change will increase with time.
Stewart et al. (2015) predict an increase of 1–2
Brown trout are sensitive to changes in discharge patterns because high-intensity floods during the incubation and emergence periods may limit recruitment (Lobón-Cerviá and Rincón, 2004; Junker et al., 2015). In the Iberian Peninsula, the trout distribution is mainly concentrated in mountain streams, where extreme discharges during winter are expected to increase (Rojas et al., 2012). These extreme discharges will likely affect trout recruitment negatively. Thus, the predicted changes in the hydrological regime can subject brown trout populations to more variable conditions, which may occasionally present some populations with insuperable bottlenecks. Trout are polytypic and display an adaptable phenology and rather high intra-population variability in their life history traits that might allow them to show resilience to variations in habitat features (Gortázar et al., 2007; Larios-López et al., 2015), especially in the marginal ranges (Ayllón et al., 2016). However, despite these strong evolutionary responses, the current combination of warming and streamflow reduction scenarios is likely to exceed the capacity of many populations to adapt to new conditions (Ayllón et al., 2016). Consistent with regional predictions (Rojas et al., 2012; Garner et al., 2015), significant flow reductions are expected during summertime in most of the studied rivers and streams at the end of the century, and this may mean, in turn, the reduction in the suitable habitat (i.e. the available water volume) (Muñoz-Mas et al., 2016). Finally, the increase in extreme droughts, which involve absolute water depletion, in certain reaches of the streams may be critical for some trout populations.
The predicted increase in winter stream temperatures can affect the sessile phases (i.e. eggs and larvae) of trout development. These phases are very sensitive to temperature changes because it affects their physiology, and because their development is temperature dependent (e.g. Lobón-Cerviá and Mortensen, 2005; Lahnsteiner and Leitner, 2013). Thus, changes in the duration of incubation and yolk sac absorption can affect emergence times and, in turn, the sensitivity of these phases to hydrological regime alterations (Sánchez-Hernández and Nunn, 2016). An increase in stream temperature can also reduce hatchling survival (Elliott and Elliott, 2010). In accordance with the results presented herein, the predicted synergy of streamflow reductions and water temperature increases will cause substantial losses of suitable fish habitat, especially for cold-water fish such as brown trout (Muñoz-Mas et al., 2016).
The increases in threshold violations were important in our simulations. The
duration of warm events (temperature above the threshold value) increased by
up to 3 months at the end of the century in the most pessimistic
scenario (RCP8.5). A continuous analysis of the whole-river response should
be conducted to allow spatially explicit predictions and to identify reaches
where thermal refugia are likely to occur. However, our results suggest that
trout will not survive in these reaches because the persistence of thermal
refugia is improbable or because their extents will be insufficient. In the
Cega, Pirón and Lozoya rivers, important losses of thermal habitat will
occur that could jeopardize the viability of the trout population.
Behavioural thermoregulatory tactics are common in fish (Reynolds and
Casterlin, 1979; Goyer et al., 2014); for instance, some species perform short
excursions (< 60 min in experiments with brook char,
According to our results, streamflow reductions are able to synergistically contribute to the loss of thermal habitat by increasing daily mean stream temperatures. This effect is especially relevant in summer in the Mediterranean area, when the warmest temperatures and minimum flows usually occur. The existence of thermal refugia represents a possible means of fish survival, and the probability for a water body to become a thermal refugium is highly geologically dependent. In our simulations, the sites that are most dependent on deep aquifers (i.e. basins underlain by Mesozoic carbonate rocks) display improved resistance to warming. The habitat retraction at the rear edge of the actual distribution of brown trout is deduced to be geologically mediated.
The mountains of central and south-eastern Spain contain the rear edge of the distribution of native brown trout (Kottelat and Freyhof, 2007). Fragmentation and disconnection of populations by newly formed thermal barriers may aggravate the already significant losses of thermal habitat by reducing the viability of populations and increasing the extinction risk. Thus, the rear edge of the trout population in the Iberian Peninsula might shift to the northern mountains to varying extents depending on the presence of relevant mesological features, such as geology. The calcareous mountains of northern Spain could be a refuge for trout because they combine favourable geology and a relatively more humid climate. As a result of this differential response, the western portion of the Iberian range (which is plutonic and less buffered) will eventually experience more frequent local temperature-driven extinction events, thus producing a greater shift northward, than in the eastern Iberian end of this range, which is calcareous and highly buffered and will remain more resilient to these local extinction events. However, the predicted streamflow reductions may act synergistically, reducing the physical space, and this may jeopardize the less thermally exposed populations. In the Iberian Peninsula, stream temperatures will increase less in the central and northern mountains than in the central plateau, and the increases will be smaller in karstic than in granitic (igneous) mountains. At the same time, the side of the peninsula that faces the Mediterranean is expected to be more sensitive to warming and streamflow reductions than the side of the peninsula that faces the Atlantic. Thus, brown trout populations in the karstic mountains of northern Spain (the Cantabrian Mountains and the calcareous parts of the Pyrenees) are better able to resist the climate warming than the populations farther east in the granitic portion of the Pyrenees (Santiago, 2017). Similar patterns may occur in other parts of southern Europe. Most likely, the less pronounced thermal responses of rivers and streams in the karstic areas will allow for greater persistence of the brown trout population, although changes in streamflow regimes will likely also occur there.
In a study of the major basins of Europe, Lassalle and Rochard (2009) predicted that the brown trout would “lose all its suitable basins in the southern part of its distribution area ([the] Black Sea, the Mediterranean, the Iberian Peninsula and the South of France), but [would] likely to continue being abundant in [the] northern basins”. Almodóvar et al. (2011) estimated that the brown trout will be eradicated over almost the entire stream length of the studied basins in northern Spain, and Filipe et al. (2013) estimated an expected loss of 57 % of the studied reaches in the Ebro Basin in north-eastern Spain. Our study shows important, yet not so dramatic, reductions in the thermal habitat of Iberian brown trout populations in mountainous areas. The number of general climate models used, the reliability of the downscaling procedure, the resolution of the stream temperature and streamflow models, and the method used to study the threshold imply a substantial improvement in detail (Santiago et al., 2016) over previous work. It is reasonable to infer that many mountain streams appear poised to become refugia for cold-water biodiversity during this century (Isaak et al., 2016).
The main findings of this study are as follows: (i) our downscaled results predict greater air temperature increments than the IPCC's averages, from which our estimations were made; (ii) significant but diverse streamflow reductions are predicted to occur during the present century; (iii) the models presented in this study have been shown to be useful for improving simulations; (iv) the predicted increases in water temperature will be influenced to varying degrees by the flow and geological features of rivers and streams; (v) the thermal habitat of brown trout, a cold-water species, will decrease as a consequence of the synergistic effects of flow reduction and water warming; and (vi) the peaks in water temperature and the complete depletion of the river channels will produce local extinctions, although the ultimate magnitude of the effect will be governed by the geological nature of the basins.
Our findings might be useful in planning the prevention and mitigation of the negative effects of climate change on freshwater fish species at the rear edge of their distributions. A differentiation of areas based on their risk level and viability is necessary to set standardized conservation goals. Our results show that trout conservation requires knowledge of both temperature and streamflow dynamics at fine spatial and temporal scales. Managers need easy-to-use tools to simulate the expected impacts and the management options to address them, and the methods and results we provide could provide key information in developing these tools and management options.
Stream temperature data owned by the authors can be found
at
In science, and particularly in hydrological studies, the uncertainty is a matter which requires special attention, and it has led us to include a synthesis of our approach to this problem in this appendix. The uncertainty analysis is a necessary step in assessing the risk level in the applicability of a model (Pappenberger and Beven, 2006). Our aim was to study the viability of the brown trout populations, and we modelled the flow and the stream temperature for this purpose. From a conceptual point of view our approaches were consistent with it.
On data inputs to build the models, uncertainties and inconsistencies are a habitual issue (Juston et al., 2013). The meteorological and hydrological services subject data to their own quality controls but systematic error cannot always be completely controlled (Beven and Westerberg, 2011; McMillan et al., 2012). For this reason, in addition, we tested the input data seeking inconsistencies.
The modelling of the river reaches as one-dimensional elements implies a simplification of the fluvial ecosystem that is generally accepted at this scale (e.g. Viganò et al., 2015; Ahmed and Tsanis, 2016), especially for ecological purposes (e.g. Caiola et al., 2014). Nevertheless, the size of the rivers under study made little or nothing relevant the variations in width and depth (it was verified in the field).
Regarding the parameterization of the models, cross-validation was used to evaluate the uncertainty in these process, and indicators such as the NSE (for hydrological and thermal models), the deviance and the RSE (for thermal models) were calculated. In the case of the thermal model, the functions of distribution of the parameters of the model were built by non-parametric bootstrap, and the mean values were chosen as the most proficient estimators. As results show, parameters tell us about the functional behaviour of catchments (particularly on the effects of the catchments geology on the streams temperature) and this might improve predictions in ungauged basins by better controlling uncertainty (Juston et al., 2013).
Once the models were constructed, It was verified that the overlaps of the
ranges of the model input variables and the ranges of the outputs were
significant (
As said, the inherent uncertainty of the climate predictions according to the scenarios RCP4.5 and RCP8.5 was attenuated by means of the ensemble technique, showing the dispersion of the results by mean of the percentiles in Figs. S1 to S24 (in the Supplement). Beven (2011) exposed his legitimate concerns on the credibility of climate models which fail when are compared with the control period and, consequently, we used ERA-40 reanalysis to control this source of bias with excellent results.
There was a time dependence between the errors of the model and the scope of the prediction, but these errors were only important in the zone of high temperature and low flow, as expected by the physical nature of the climatic variables. Moreover, this is the behaviour of the variables that was our intention to evaluate.
The authors declare that they have no conflict of interest.
We are grateful to the Consejería de Medio Ambiente y O.T. of the
Government of Castilla y León, especially to Mariano Anchuelo,
Fabián Mateo and the forest ranger team of Navafría. Also, we are
indebted to the Sierra of Guadarrama National Park staff, and especially with
Juan Bielva and Ángel Rubio for the temperature data of Lozoya stream.
Juan Diego Alcaraz is the author of the temperature data of the Ebrón
Basin. Valérie Ouellet and two anonymous referees provided valuable
comments that substantially improved the original manuscript. The World
Climate Research Programme's Working Group on Coupled Modelling is
responsible for the 5th Coupled Model Intercomparison Project, and we thank
the climate modelling groups for producing and making available their model
output. Climate change study was
partially funded by the Ministerio de Agricultura, Alimentación y Medio
Ambiente of Spain through the Fundación para la Investigación del
Clima (