Study area

The Loire catchment is the largest in France, with a drainage area of about 117,000 km², covering 20% of the French metropolitan territory. The Loire River is 1,012 km long and runs through a variety of landscapes and climates from its source at Gerbier-de-Jonc to its estuary at Saint-Nazaire where it flows into the Atlantic Ocean. The upper part of the catchment is located within a forested mountain range (the Massif Central) while agriculture dominates in the downstream areas. The sedimentary basin is mostly occupied by cereal farming and vineyards and the western and south-eastern parts of the catchment being mainly dedicated to livestock (mostly cattle). Annual precipitation varies between 600 and 1,300 mm per year over the catchment and average air temperature ranges from 6 to 12.5°C. The Loire catchment shows a variety of hydrological regimes with eight regimes out of the 12 present in France [32, 33]. Specific discharge ranges from 3 l/s/km² (in the Maine catchment) to 40 l/s/km² (in the Cévennes). Seasonal variation in discharge can be high for some catchments due to very little groundwater supply [33]. The mean interannual discharge of the Loire River in its downstream part (at Montjean-sur-Loire) is 840 m³/s, with flood peaks 5–10 times higher (Banque Hydro; www.hydro.eaufrance.fr). The study area includes the entire Loire catchment except the part under marine influence, downstream of Montjean-sur-Loire (110 kilometers upstream from the estuary) (Fig 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1.

Study area and location of the sampled reaches for (A) fish, (B) macroinvertebrates and (C) both (i.e. “common” dataset) and (D) T-NET hydrographic network. In panels A, B and C, a simplified version of the T-NET hydrographic network is shown. In panel D, the 52,278 T-NET reaches are colored according to their mean annual water temperature (°C). The inserted map in the center locates the Loire catchment within Europe (source: Natural Earth).

https://doi.org/10.1371/journal.pone.0274167.g001

Environmental and biological datasets

Direct vs. surrogate environmental variables.

All environmental variables were selected on the basis of knowledge of their potential influence on fish and macroinvertebrate communities (i.e. literature, expertise). A set of 23 environmental variables were considered, including descriptors of water and air temperatures, discharge, local hydraulic conditions and topography (see Table 1 for the full list of variables). They were described at the scale of the river reach (mean length ± standard deviation = 1.7 ± 1.4 kilometers), which is the spatial unit of the physical-based model Temperature-NETwork (T-NET). This model is based on a hydrographic network of the Loire catchment through 52,278 reaches and use the equilibrium temperature concept [34].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Range of environmental values covered by each dataset.

Average value and range (i.e. [minimum: maximum]) of the 23 environmental descriptors computed on: (1) the entire T-NET network (N = 52,278 reaches), (2) reaches with fish data (N = 1,191), (3) reaches with macroinvertebrate data (N = 1,195), and (4) reaches with fish and macroinvertebrate data (“common” dataset; N = 532). Ta and Tw refer to air and water temperature, respectively. Suffixes m, sd, 10 and 90 refer to the mean, standard deviation, lower 10% quantile and upper 90% quantile, respectively.

https://doi.org/10.1371/journal.pone.0274167.t001

Discharge data were obtained from the semi-distributed hydrological EROS model [35], which simulates daily discharge at the outlet of 368 sub-basins within the Loire catchment (sub-catchments area range from 40 to 1600 km² with mean = 300 km²). Within a sub-basin, daily discharge is extrapolated depending on drainage area. The EROS model performed well on 44 stations at the annual scale with a mean relative bias ± standard deviation of -0.12 ± 10.62%.

Air temperature was derived from SAFRAN, which is a mesoscale atmospheric analysis system for surface variables at the hourly time step using ground data observations at the spatial resolution of a grid of 8 x 8 km [36, 37]. At annual scale the RMSE between observations and corresponding SAFRAN outputs was of 1°C [37].

Discharge statistics were translated into direct hydraulic statistics (i.e. reach width, water depth and current velocity) using the hydraulic geometry models from Morel et al. [38]. These models improve the classical hydraulic geometry equations of Leopold and Maddock [39] and relate hydraulics to discharge using power laws. To date, the models of Morel et al. [38] provide the best estimates of reach hydraulics at a given river discharge in the absence of detailed field hydraulic measurements. These models were available for a hydrographic network of the Loire catchment (Réseau Hydrographique Théorique, RHT) [40] coarser from that of the T-NET model. Therefore, we made a correspondence between RHT and T-NET river reaches, so that 49% of T-NET reaches had corresponding RHT reaches. For T-NET reaches without RHT correspondence, the hydraulic geometry coefficients were extrapolated from the nearest downstream T-NET reach with RHT correspondence.

Water temperature data were obtained from T-NET, which simulates daily water temperature for each reach of the entire river network of the Loire catchment (Fig 1D). Comparison of annual average values (i.e. the time-scale of this study) between simulations and observations at 72 stations (2010–2014) showed good performance of the T-NET model with mean bias ± standard deviation of 0.2 ± 0.82°C and mean root-mean square error (RMSE) = 0.7°C.

Based on the daily T-NET, SAFRAN and EROS outputs, we computed the interannual average of four summary statistics (i.e. annual mean, lower 10% quantile, upper 90% quantile and standard deviation) for water temperature, air temperature and discharge, respectively, in order to describe the thermal and hydrological regimes of each river reach over the period 1990–2010. Three summary statistics (i.e. annual mean, lower 10% quantile and upper 90% quantile) were calculated for hydraulic variables. Water temperature and hydraulics descriptors were considered as direct variables whereas discharge and air temperature descriptors were considered as surrogate variables. The mean slope of the river reach and the distance to the estuary, which are known to influence aquatic communities, were also considered. The mean slope of the river reach was extracted from the digital elevation model IGN BD ALTI® 25x25 m [41]. The distance to the estuary was obtained by summing the length of all T-NET reaches located downstream of the reach of interest. Metrics of discharge, hydraulic variables and slope were log-transformed prior to analyses to approach normality.

Data analysis

Non-metric multidimensional scaling (NMDS) was used to examine the organization of aquatic communities at the scale of the entire Loire catchment. This iterative ordination method is based on a ranked dissimilarity matrix between statistical samples (here T-NET reaches) and is adapted to ecological datasets including rare and low abundance taxa, especially when data from different biological groups are considered [47]. In addition, NMDS performs well in highlighting key organizational patterns at the community level, without placing too much emphasis on patterns specific to certain species or taxa. First, for each biological dataset, a dissimilarity matrix based on genera or species Wisconsin-transformed CPUE was built using Bray-Curtis dissimilarity among T-NET reaches. The best ordination was selected after 200 iterations from a random initial ordination. We assessed the quality of the NMDS using the stress value, which is a measure of the goodness-of-fit between the ranks of the observed dissimilarities and ordination distances. When stress value is lower than 0.2, the ordination is considered correct (accurate if stress values are lower than 0.1) [48].

Pairwise ordinations between biological groups were then compared using a Procrustes analysis. This method assesses the similarity of two factorial spaces and thus here the spatial concordance between biological groups. Procrustes analysis was preferred to a Mantel test because of its greater statistical power [49]. The Procrustes sum of squares m², also called concordance value, was used to measure the similarity between two ordinations [22]. It ranges from 0 to 1 with low m² values indicating a higher degree of similarity. Because Procrustes analyses must be conducted on ordinations performed with the same samples, NMDS were reproduced considering only the 532 T-NET reaches common to all datasets (i.e. fish, macroinvertebrate and “common”). Finally, a fourth NMDS was computed based on the 23 environmental descriptors (Table 1), for the 532 reaches common to all three biological datasets. This fourth ordination was then compared to each biological ordination using Procrustes analyses to assess the concordance between environmental gradients and the organization of each biological group.

To assess the influence of the previously described environmental variables on the organization of aquatic communities resulting from the NMDS, we perform a redundancy analysis (RDA). RDA were used as an extension of linear regression allowing to handle the multiple response variables and to overcome the collinearity inherent to the calculated metrics (i.e. mean, lower 10^th quantile, upper 90^th quantile and standard deviation) [50]. First, for each biological group (i.e. fish, macroinvertebrates and common) and each environmental variable (i.e. water temperature, air temperature, hydraulic variables and discharge), a RDA was performed with the three NMDS axes as response variables and groups of 4 metrics of the corresponding environmental variable, as explanatory variables. Second, to compare the effect of direct vs. surrogate variables, a variation partitioning in RDA was performed per biological group with the three NMDS axes as response variables and all metrics of direct or surrogate variables as explanatory variables (i.e. metrics of water temperature and hydraulic variables for direct variables and metrics of air temperature and discharge for surrogate variables). The respective explanatory power of each variable (coefficient of determination R²) for each biological group was then assessed [51]. For the hydraulic variables, for which we computed a total of nine metrics (for each of the three variables, i.e. current velocity, water depth and reach width), all combinations of four out of nine metrics were tested, to keep the same number of explanatory variables and the same degree of freedom between the RDA. To minimize the influence of spatial autocorrelation inherent to these datasets on the statistical assessments, RDA was combined with a Monte-Carlo procedure. We first selected all reaches contained in 70% of the sub-basins (i.e. among the 368 sub-basins considered in the EROS model) as calibration dataset to perform a RDA. This entire procedure was repeated 999 times. To assess both the model explanatory performance and the significance of the difference in average R² values between groups of environmental variables, the overlap between confidence intervals around the R² was evaluated. Confidence intervals was calculated as 2×standard deviation. The predictive performance of the model was then also assessed on the remaining 30% of sub-basins (see S1 Appendix for details).

Finally, we tested the sensitivity of the results obtained by RDA to uncertainties in environmental variables derived from physical-based models (i.e. water and air temperatures, discharge and hydraulic variables). To do so, we replicated the cross-validation procedure, except that each value of water temperature, air temperature, discharge and hydraulic variables was modified by adding a randomly sampled error scaled to the uncertainties provided for each environmental model (see S1 Appendix for details).

Data formatting and statistical analyses were performed using the R software [52]. The vegan R package [53] was used for NMDS and RDA computation and the maptools R package [54] was used for GIS data manipulation.

Main environmental determinants of the organization of aquatic communities in the Loire catchment

The NMDS performed on the three biological datasets (i.e. fish, macroinvertebrates and common) highlighted the role of the longitudinal gradient in structuring aquatic communities at the extent of a large catchment. The NMDS based on fish densities closely transcribed this longitudinal gradient, the relationships with the environmental variables showing the opposition between thermal and hydraulic-hydrological variables (increasing from upstream to downstream) and topographic variables (decreasing). Fish communities varied along this axis, from salmonid-dominated communities in the upstream reaches to cyprinid-dominated communities in the downstream areas, with a concomitant increase in species richness and total density. The resulting organization of fish communities is consistent with that generally described in the literature [15, 20]. The relationship between the upstream-downstream environmental gradient and macroinvertebrate densities was less clear but still appeared on the second NMDS axis. All R² between the macroinvertebrate ordination and the environmental variables were slightly lower than those found for fish, which indicates that the higher variability of the macroinvertebrate community was slightly less explained by our set of environmental variables. This suggests that macroinvertebrate communities are not as strongly structured by direct variables and the longitudinal gradient at this spatial extent as fish communities are. Key differences in the ecology of fish and macroinvertebrates may explain this result. For example, the greater ability of fish have access to more diverse habitats and can therefore occupy habitats more closely aligned with their ecological requirements [26, 55] Overall, for the three biological NMDS, we did not observe clustered groups of reaches or any other patchy patterns, although the sampling stations covered very distinctive geographical entities (e.g. Massif Central, Loire valley). This supports the idea that succession between communities is gradual and continuous, with transitions where species with different ecological preferences co-exist, rather than being bounded (as formulated in the River Continuum Concept) [18].

The community organization of the aquatic groups studied was essentially explained by the same set of environmental variables. The most influential determinants were water temperature (especially mean and quantile 90), hydraulic variables (i.e. water depth, reach width and current velocity) and slope associated to the distance to estuary. This finding is consistent with the literature on the topic. First, water temperature, especially warm temperature, limits the spatial distribution of aquatic species according to their thermal tolerance and oxygen requirements [28, 56, 57]. Nonetheless, while being frequently recognized in the literature, few studies have shown the key role of water temperature in structuring aquatic communities at large scale using water rather than air temperature data. Second, hydraulic conditions represent an important local environmental filter that constrains species depending on their hydraulic habitat preferences, also influencing other habitat characteristics (e.g. substrate size, type of vegetation) [58, 59].

Direct vs. surrogate environmental drivers of the organization of aquatic communities

We found that direct environmental variables (i.e. water temperature and hydraulic descriptors) explained significantly better the organization of aquatic communities at the scale of the Loire catchment than surrogate ones (i.e. air temperature and discharge). This corroborates our first hypothesis assuming that direct variables better reflect the actual conditions experienced by aquatic organisms than surrogate ones, therefore better explaining their community organization. We found significant differences between water and air temperature in their ability to explain the organization of fish and macroinvertebrate communities. Studies generally agreed that water temperature is highly correlated with air temperature, especially when the spatial resolution of the study is large [28]. As water temperature measurements are rarely available at large extent, air temperature data are therefore often used as a proxy in large extent ecological studies (e.g. [5, 60, 61]). However, air temperature does not capture all the variability of water temperature. This is highlighted by the ranges of values taken by the two kind of metrics at the scale of the whole T-NET network (see Table 1), with for instance mean air temperature ranging from 6.1 to 12.5°C whereas mean water temperature ranges from 5.4 to 16.2°C (the differences being even higher for the values of lower 10% quantile and upper 90% quantile). Consequently, air temperature does not substitute well for water temperature under certain conditions, such as in the most upstream reaches or near groundwater inputs [28, 34]. For all three biological datasets, the relationships between the NMDS ordinations and the discharge were weaker than with the corresponding hydraulic variables. This supports the idea that the resolution at which discharge metrics have been computed was not the most appropriate to correctly transcribe the relationships between organisms and discharge, unlike hydraulic metrics, which describe habitats at a finer grain. Indeed, hydraulic metrics provide additional dimensions relatively to discharge, because they better reflect local geomorphological variations that influence habitats [38, 62, 63]. Finally, when the direct environmental variables were considered together, they still better explained the organization of fish and macroinvertebrate communities than the full set of surrogate environmental variables. Moreover, the differences between direct and surrogate variables were even greater for the common dataset. These two results reinforce the hypothesis that the combination of several biological groups associated to the use of direct environmental variables is most suitable for accurately assessing the environmental-organism relationships in lotic hydrosystems. As far as possible, we recommend using several biological groups and direct variables to study spatial patterns of aquatic communities. Nonetheless, surrogate variables remain valuable when direct variables are not available, especially at large scale where the differences in explaining the organization of aquatic communities can be slight as found in the present study. In particular, direct variables are likely to perform better at finer spatial extents and may be more interesting for environmental management as they are more directly linked to actual levers of action. These differences were further reduced with respect to the predictive power of both variables. Therefore, the choice of using direct or surrogate environmental variables may also depend on whether the objective is to explain and understand the organization of aquatic communities and its main drivers, or rather to predict spatial patterns under different scenarios (e.g. climate change).

The environmental variables studied here only partially explained the organization of fish and macroinvertebrate communities. The spatial resolution of the study was the T-NET reach (~1.7 km on average; N = 52,278) and we believe that such accuracy, density and coverage at the extent of a large hydrographic network (cumulative length of the river system = 88,058 km) made it highly suitable for precisely investigating the organization of aquatic communities in the Loire basin. Setting the study at a broad spatial extent allows capturing organism responses to major environmental gradients arising from climate, geology or topography [64]. Nevertheless, some processes occurring at a markedly different spatial scale may be blurred at such large spatial extent [65, 66]. For instance, large spatial extent may include distinct biogeographic units, which are known to be driven by other factors such as historical events (e.g. major glaciations) and dispersal capacity of organisms [67–69]. While the catchment scale is now widely recognized as appropriate for examining the organization of fish communities and for management and conservation measures (e.g. Huet [12]; Miranda et al. [70]; Oberdorff et al. [69]), the appropriate extent at which macroinvertebrate communities should be studied remains under debate [71]. For example, Hawkins and Norris [72] and Johnson et al. [73] agreed that the organization of macroinvertebrate communities is better explained by local habitat factors (e.g. substrate, aquatic vegetation) than coarser ones (e.g. land-use in the upstream drainage area, position of the stream in the hydrographic network) whereas Townsend et al. [74] found the opposite. Overall, it seems that the organization of aquatic communities is ruled by the interplay of different spatial scales, with broad-scale processes influencing those occurring at finer scales [17, 58, 73, 75, 76]. One way to capture these scale interactions would be to build hierarchical models able to account for several scales together (e.g. Villeneuve et al. [17]). Here, the NMDS based on macroinvertebrate densities suggested the role of processes occurring at two levels: catchment scale, governed mainly by thermal and hydrologic regimes, and reach scale at which hydraulic variables show important variations. However, explanatory variables that may be in interaction with variables already considered may have been overlooked, and their inclusion could improve our understanding of the organization of aquatic communities. Furthermore, limitations in the modelling of environmental variables may also influence their contribution in explaining the organization of aquatic communities. The sensitivity analysis performed (S1 Appendix) showed that the results were robust to the uncertainties arising from the physical-based environmental models. Nevertheless, we acknowledge that the T-NET and EROS models, from which the water temperature and discharge data used in this study are derived, do not account for all human-induced alterations of streams, notably physical barriers (e.g. dams, weirs) that are known to alter the hydrological and thermal regimes of rivers [77, 78]. Although the Loire basin is acknowledged as one of the last wild river in Europe because of a limited impact of human activities on its hydrological regime, further developments to integrate anthropogenic uses would be interesting [33].

Concordance between the organization of fish and macroinvertebrate communities and the value of a cross-taxa approach

The concordance between the three NMDS ordinations (i.e. fish, macroinvertebrates and “common”) was very weak. This result highlighted that fish and macroinvertebrate communities do not seem to share very similar spatial patterns at the extent of the Loire catchment. Thus, while being driven by the same set of environmental variables, both communities seem to respond differently to environmental gradients. This reinforces the assumption that macroinvertebrate and fish communities are sensitive to environmental gradients but at a different spatial extent. The weak concordance between their community organization confirms the relevance of studying fish and macroinvertebrates together to identify generalizable distribution patterns of aquatic communities instead of considering only one or the other. Most studies that used a similar approach (i.e. multivariate ordination and concordance assessment) have also found weak concordances between biological groups at spatial extents similar to the Loire catchment (e.g. fish and macroinvertebrates in Kilgour and Barton [79]; many biological groups in the review from Heino [24]; diatoms, macroinvertebrates, macrophytes and fish in Johnson and Hering [80]). In addition, some studies have compared the concordance between taxonomic groups at different spatial scales. Dolph et al. [23] and Paavola et al. [22] found greater concordance in the organization of fish and macroinvertebrate communities at broader spatial extent(i.e. similar to the Loire catchment) whereas Bae et al. [81] found the opposite. The first two studies showed that fish and macroinvertebrate communities were concordant at broad geographical scale, over study areas of similar size to that of the Loire catchment (i.e. 101 stations within two ecoregions in Finland in Paavola et al. [22] and 670 stations in Minnesota, USA in Dolph et al. [23]). However, there are noteworthy differences between these studies and the present one. Paavola et al. [22] focused on boreal aquatic communities, which live under much colder conditions and have lower richness and abundance. Dolph et al. [23] based their results on concordance between taxonomic groups using species-poor North American communities with a total of 5 fish species and 15 macroinvertebrates taxa. Moreover, in both studies, the broad spatial extent is constituted by grouping small sampling areas from different distant sub-basins, thus not considering a single large catchment that would include the full range of environmental conditions. Furthermore, the weak concordance found between the ordination based on the “common” dataset (i.e. combining fish and macroinvertebrate sampling) and the other ordinations indicated that neither fish nor macroinvertebrates were more representative of the organization of the combined aquatic communities. The influence of environmental drivers differed: macroinvertebrates showed weaker correlations with most environmental variables than fish, despite our second hypothesis. The concordance between the environmental ordination and the “common” ordination was higher than with fish and macroinvertebrate ordinations, which is certainly due to a better embedding of the environmental gradient through the biological gradient captured by the NMDS combining fish and macroinvertebrate data (i.e. richer and more complex communities). This observation was also confirmed by the slightly higher R² found between most explaining environmental variables and the “common” ordination. This supports our third hypothesis that combining different biological groups improves the fit with the main environmental gradients, even though it increases the biological variability and complexity.

Some hypotheses may explain the differences between biological groups found here. First, the macroinvertebrate dataset included 186 genera against 43 species for fish. Consequently, communities were more diverse for macroinvertebrates than for fish (on average 43 genera per reach for macroinvertebrates vs. 10 species per reach for fish). The observed CPUE were also more variable for macroinvertebrates (CPUE mean = 3,851 ± 3,121 specimens per 100 m² per T-NET reach) than for fish (CPUE mean = 97 ± 140 specimens per 100 m² per T-NET reach). This could lead to a higher variability to be explained by the same set of environmental variables for macroinvertebrates than for fish. Furthermore, the extent of the spatial distribution of macroinvertebrate genera was on overall more variable than for fish species, with ubiquitous genera occurring along the whole longitudinal gradient (e.g. the genus Baetis was present in more than 95% of the reaches sampled; S2 Table) and other genera showing very particular ecological preferences and restricted occurrences within the Loire catchment (e.g. the genus Leptophlebia was present in less than 2% of the reaches sampled, mainly in upstream reaches; S2 Table). In contrast, fish ordination has highlighted typical ‘indicator’ species like S.trutta or T.thymallus whose ecological requirements are well known and which are found in specific parts of the Loire catchment. Considering the high number of genera in the macroinvertebrate dataset and the large extent of the area studied, it seems plausible that the Loire catchment includes distinct sub-basins, each having distinct macroinvertebrate biogeographic patterns. The organization of macroinvertebrate communities would thus be less driven by the upstream-downstream environmental gradient. In contrast, the longitudinal organization of fish assemblages seems to apply to the entire catchment. Second, the identification of macroinvertebrates was limited to the genus level, which may group species with different ecological preferences [72, 82]. Previous studies had shown that family-level data can be sufficient to examine the main spatial patterns in macroinvertebrate communities, although more precise patterns can only be detected with a finer level of taxonomic identification [83]. In addition, the sensitivity of non-metric ordination methods to the identification level of macroinvertebrates is known to be reduced under the family level [55, 79]. According to these studies, the use of the genus level seems to be a good compromise between practical limitations (i.e. field surveys, time allocated to taxa identification) and appropriate identification level to examine ecological patterns. Third, fish typically use larger areas of stream to complete their life cycle than macroinvertebrates: above the kilometer for most fish species vs. a few meters for most macroinvertebrate genera [26, 79]. Both groups can therefore be sensitive to the same environmental variables but not at the same spatial grain [84]. Macroinvertebrates are more sensitive to local determinants (sensu habitat variables) and show stronger responses to local perturbations than fish, especially due to their lower dispersal and avoidance abilities in their larval form [58, 85, 86]. This implies that the identification of patterns in the organization of macroinvertebrate communities at large extent might be blurred by local processes and disturbances acting at fine grain. Fish are probably less affected by these fine-grained factors as they are able to move to suitable habitats when habitat alteration is local and temporary. Stochastic processes of colonization dynamic (e.g. organisms’ drifts, flying macroinvertebrates) may also play a more important role for macroinvertebrates than for fish [26].