Methods

2.1 Study area

Switzerland (46° 57’ 04’’ N 7° 26’ 19’’ E - 41,285 km²) has a variety of landscapes and habitats with the Alps acting as a separation line. The climate conditions vary regionally due to the mountainous influence; from an intra-alpine dry and continental climate regime in the Western and Eastern Alps to an oceanic climate at low-elevation (Plateau) and high elevation (Northern Alps, Jura Mountains). The Southern Alps are dominated by relatively mild and dry winters and warm humid summers. During the hottest month (July), the range of the mean monthly temperature varies from 2.7°C (Arctic) to 21.9°C (Mediterranean) according to altitude (Zimmermann and Kienast, 1999; MeteoSuisse, 2013). July rainfall exhibits important spatial variations and ranges from 359 mm to 2594 mm.

More than a third of the Swiss territory is used for agriculture and another third is covered with forest. Lakes, rivers, unproductive vegetation and surface areas without vegetation cover about a quarter of the territory. Urbanization which used to occupy 7

% of the land is now growing quickly to the detriment of agriculture (OFS 1992/1997).

41

Fig. 1: Distribution of the 1,402 localities within the 6 main biogeographical Swiss regions (Gonseth et al. 2001). JU = Jura Mountains, PL = Plateau, NA = Northern Alps, WA = Western Alps, EA= Eastern Alps, SA = Southern Alps.

2.2 Charophyte species data

The aquatic ecosystems of the country were investigated with the main objective of evaluating the extinction risk of the charophyte species growing in Switzerland (Auderset Joye and Schwarzer, 2012). Sites to be explored in the field were selected from a stratified sample of 21,092 aquatic ecosystems represented by ponds (45%), lakeshore segments (10 %) and breeding sites for amphibians of national importance (45 %). The stratification of the set of 21,092 localities was based on biogeography (6 regions), altitude (6 classes; Fig.1) and the proportion of agricultural land used in the catchment area (3 categories). The localities finally chosen for investigation were randomly selected from the dataset with a double constraint: selecting at least 4 sites in each of the 108 strata. The sites were selected within historical sites (when existing) known to have harbored charophytes in the past. Lastly, 1,402 localities were surveyed during 4 years (2006-2009). Only 387 localities were colonized by charophyte species during this fieldwork. The data analyzed here included presence/absence for the 20 species recorded in the 1,402 localities (Table 1).

42 species modelled are in bold fonts. Nomenclature according to www.algaeBASE.org

Abbreviation Occurrence Frequency Degree of threat in the Red List Switzerland

Nitella opaca (Bruzelius) C.Agardh niopa 39 2.8 VU

Nitella syncarpa (Thuill.) Chevall. nisyn 11 0.8 EN

Nitellopsis obtusa (Desv.) J Groves niobt 46 3.3 NT

Tolypella glomerata (Desv.) Leohn. toglo 12 0.9 EN

2.3 Selection of environmental predictors

A conceptual framework based on the ecophysiological and biophysical processes that govern the relationships among species and their environment can be used to choose potential environmental variables that describe species distributions (Austin, 2007). This framework recognizes indirect, direct and resource variables. Temperature and rainfall are considered as direct variables, while resource variables are those which are consumed by organisms, e.g. nitrogen for plants (Austin and Smith, 1989; Guisan and Zimmerman, 2000). This conceptual general model works for terrestrial plants but rainfall, for example, acts indirectly through water-level changes and their consequences (light, temperature) on aquatic plants.

43

Predictors were carefully chosen from the GIS available variables (at 25 m resolution) which were considered to have a high potential to explain charophyte species distribution. They belong to four domains: climate, rock type, land-use and waterbody size. For each domain we designated a subset of one or several candidate explanatory variables. The variables finally selected show relatively low correlation to each other (Table 2). In order to avoid problems of collinearity between predictors we restricted their number to 8 (Table 3).

Table 2: Pearson correlations between the variables used to model charophyte species (n=1402). The full names of the variables are given in Table 3.

AGRIBV

-0.46 FORETBV

0.087 0.12 CaCO3 _BV

-0.34 0.65 0.045 FO200

0.58 -0.35 0.046 -0.55 AGR200

-0.042 0.40 0.092 0.22 -0.19 Tju

0.066 -0.11 0.17 -0.078 0.1 -0.29 Preju

-0.02 -0.14 -0.0081 -0.12 -0.085 -0.009 -0.17 log(area)

44

Table 3: Statistics of the environmental variables used for the predictions (21,092 localities).

Sources: (1) measured in GIS; (2) Zimmermann and Kienast (1999); (3) CSCF=Swiss Center for Faunal Cartography; 4) Swiss Federal Statistical Office, modified by Allenbach et al. (2010).

type code Definition

Range within study area 1stquartile Median Mean 3rd quartile transformation source

Size log(area) Waterbody surface area (m²) 2 - 866464 3 159 2957 833 Log10 1 Climate Tju Mean temperatures in July

(1960-1990), in °C 2.71- 21.9 10.3 17.1 14.6 18 none 2 FORETBV Proportion of forest in the drainage

basin (watershed), in % 0 – 100 0 18 24 38 none 4

AGR200 Proportion of farmland in the surrounding area (r = 200m) ), in % 0 – 100 7 38 39 67 none 4 FO200 Proportion of forest in the

surrounding area (r = 200m) ), in % 0 – 100 0 6 20 35 none 4

a The 72 land use categories (OFS 1992/1997 were aggregated into 3 categories: agriculture: 71 to 89, forest: 11 to 19 and urban: 27). (Spatial resolution 25 x 25 m).

Climate variables were preferred to altitude because they are more direct and known to constraint species distribution. We chose mean temperature and July precipitation as the most representative predictors for climate because charophyte species are strongly influenced by high temperature and droughts (Casanova and Brock, 1990; 1999; Bonis et al, 1995, 1996; Rey–Boissezon and Auderset Joye, 2012). The temperature of the colder months was supposed to be less relevant as many charophyte species seem to be not sensitive to cold conditions (Krause, 1997; own observations). Mean annual temperature and July radiation were not used as predictors because they were highly correlated to mean July temperatures.

Rainfall has a direct effect on plants but is also indirectly linked to physical processes (erosion, flooding, etc). In Switzerland, July is the hottest month of the year with a highly

45

variable precipitation. July precipitation (Preju) is the sum of rainfall during this month (mean 1960-1990). Mean July temperature (Tju) was calculated by averaging the temperatures of the month over 30 years (1960-1990). Climatic predictors are those calculated by Zimmermann and Kienast (1999). Because the species occurrence currently recorded is the result of recent warming, it would have been better to have more recent climate data (1980-2010). However, it is just a temperature lag: if the warming was of 2°C in the last decades, the increase took place in the same direction at every locality. This means that even if the absolute values obtained here are a little low, the species-environment relations stay mostly the same over time. The same reasoning applies to the landuse data.

The means by which the lands are managed have an impact on natural ecosystems and thus components of land-use were utilized as indirect predictors for species distribution. A review of the literature suggests that human transformation of land-cover and land-use are key drivers of the loss of biodiversity and ecosystem services (Haines-Young 2009). Intensification of agriculture is likely to be followed by the decrease of species richness primarily caused by increasing disturbance, increasing loss of valuable habitat, use of chemicals and fertilizers (Benton et al. 2003). High nutrient concentrations increase phytoplankton biomass leading to a decrease in water transparency of eutrophicated waterbodies which can probably be related to the decline of charophytes (Baastrup-Spohr et al. 2013). We assumed that ecosystems surrounded by intensive agriculture have reduced chances to host charophytes compared to ecosystems located in a more pristine environment.

Likewise, we presumed the proportion of land covered by forest to be positively correlated to the occurrence of charophyte species. Woodlands are expected to have the opposite effects as cultivated land: forests retain nutrients from agricultural activities or from deposition and therefore help to preserve low levels of eutrophication and a good water quality (Baattrup-Pedersen et al. 2012, Holmberg et al. 2013).

We used the proportions of land used by agriculture and covered with forest at the scale of the catchment area and the scale of the surroundings of the localities (200 m) because we expected different spatial scales to contribute differently in explaining species distribution.

46

The original nation-wide land-use data for Switzerland are derived from aerial photographs at a 100 m resolution (OFS 1992/1997) but the data used here were interpolated at a 25 m resolution (Allenbach et al. 2010). The calcareous content of parent material (CaCO3_BV) was retained as a geological predictor. Calcium is a direct variable and was proved to be a discriminatory factor for charophyte species (Stroede 1933, Olsen 1944). The digital layer of CaCO3 content was derived from the geotechnical map of Switzerland (OFS 2003) using an expert classification (J. Ayer, University of Neuchâtel, Switzerland). A semi-quantitative variable representing the proportion of calcareous content within each type of parent material was created (class 1: low content to class 5: high content).

Waterbody area is known to have an indirect and positive effect on plant species biodiversity; large lakes support more species (Rorslett 1991) and large ponds host species absent from the smaller ones (Oertli et al. 2002). In the present study the size of the waterbody corresponds either to the surface delimited by 1 km shoreline and the maximal depth of plants colonization observed for a lake section, or to the entire surface for all other localities such as pond and small lakes. A large waterbody size corresponds to lakes (up to 86 ha) and middle to small size for other type of ecosystems (see Table 3). Waterbody size was logarithm transformed to be used in the models (log(area)).

2.4 Statistical analysis 2.4.1 Species distribution along predictors

To explore species distribution along predictors, data were summarized along each of them. The results are presented as boxplots realized with R software (R Development Core Team, 2011). Only species present in 11 or more localities are shown. A Kruskal-Wallis test was applied to the data for testing whether the samples originate from the same distribution. This test leads to significant results, when at least one of the samples, here the occurrences of each species, is different from the others.

2.4.2 Species Distribution Modeling

As a first step, we fitted a model using 8 predictors to assess the species habitat and predict their distribution across the country. Because species-environment interactions

47

are generally non-linear (Austin, 1987), we chose a non-parametric model able to fit any response curve. Models were built with GAMs (Hastie & Tibshirani, 1990) and calibrated using the GRASP version 3.2 package (Lehmann et al. 2002; http://www.cscf.ch/grasp) for S-plus version 6.2 (Insightful Corp., Seattle, WA, USA). GRASP incorporates selection methods and the possibility of dealing with interactions among predictors as well as GRASP. According to Maggini et al. (2006), the model selection based on cross-validation appeared to be a powerful method among several methods tested, offering the best compromise between model stability and performance (parsimony). We chose a backward stepwise selection and 3 degrees of freedom. To validate the selected models, a 5-fold cross-validation was used (Franklin et al., 2000).

To assess the GAM models performance we looked at several criteria: (i) a statistical evaluation using the area under the curve of a receiver operating characteristic plot (ROC) on the training data set (Fielding and Bell, 1997); (ii) a fivefold cross-validated ROC (cvROC), (iii) the percentage of explained deviance (D2, Guisan et al. 2002); iv) a visual evaluation of response curves and of spatial prediction for each species (not presented here). Predictor contribution was calculated as the percentage of the sum of model contributions as defined in GRASP. For each predictor, model contribution is defined by the possible range of variation on the linear predictor scale (Lehmann et al.

2002). According to Swets (1988) the discrimination ability of the model is poor for a ROC value between 0.5 and 0.7, reasonable for ROC 0.7-0.9 and very good for ROC 0.9-1.

A ROC value remaining high after cross-validation (cvROC) indicates a model with a high degree of stability.

2.5 Applying scenarios of environmental changes

Environmental change scenarios have been used in the literature to evaluate the effects of climate warming or land-use modifications on the species richness of different

48

groups (Steck et al., 2007; Nobis et al., 2009) including aquatic organisms (Rosset et al., 2010; Rosset and Oertli, 2011; Alahuhta et al., 2011). Here we used the resulting species models for predicting the impact of a scenario exclusively climatic, i.e. the estimate of the shift in charophyte species occurrence due to future temperature and precipitation changes in Switzerland.

Over the past 100 years, the average annual temperature in Switzerland has risen and during the past 30 years the warming has accelerated. This trend is set to continue and its extent will depend on the development of the green gas emission (OFEV, 2012). For the period 2012-2050, precipitation levels are predicted to decrease very strongly during the summer. Based on this national forecast, we choose to apply a scenario based on an increase of 2°C in July mean temperatures and a decrease of 15 % in July precipitation. The scenario was operated by recalculating the climate variables on the 21,092 aquatic ecosystems and was applied uniformly, even if we were aware that local small differences probably occur in reality. We used the build model to calculate a new probability of occurrence of a species in each ecosystem. We analyzed the modifications (% change) in charophyte species occurrence by comparing the total probability of sites resulting from our species models with the total probability of sites predicted by the scenario. A Kolmogorov-Smirnov test (p=0.05) was performed to compare the probabilities distribution before and after the scenario application using R software (R Development Core Team, 2011).