Predicting climate change effects on wetland ecosystem services using species distribution modeling and plant functional traits

Wetlands provide multiple ecosystem services, the sustainable use of which requires knowledge of the underlying ecological mechanisms. Functional traits, particularly the community-weighted mean trait (CWMT), provide a strong link between species communities and ecosystem functioning. We here combine species distribution modeling and plant functional traits to estimate the direction of change of ecosystem processes under climate change. We model changes in CWMT values for traits relevant to three key services, focusing on the regional species pool in the Norrström area (central Sweden) and three main wetland types. Our method predicts proportional shifts toward faster growing, more productive and taller species, which tend to increase CWMT values of specific leaf area and canopy height, whereas changes in root depth vary. The predicted changes in CWMT values suggest a potential increase in flood attenuation services, a potential increase in short (but not long)-term nutrient retention, and ambiguous outcomes for carbon sequestration. Electronic supplementary material The online version of this article (doi:10.1007/s13280-014-0593-9) contains supplementary material, which is available to authorized users.


Properties of the Swedish National Wetland Inventory (VMI) data and details on species distribution modeling (SDM) with MaxEnt
The VMI was a nationwide effort to survey the Swedish wetlands below the alpine region, directed by the Swedish Environmental Protection Agency (Gunnarsson and Löfroth 2009). A total of 35 000 sites were included and mapped using aerial photographs. More than 4 000 wetlands were surveyed in the field, generating plant species lists including estimates of plant abundance (on a scale of 1 to 3, from single individuals to dominant).
Only large wetlands were surveyed (> 10 ha in Southern Sweden, > 50 ha in the North), leaving presumably large numbers of small or ephemeral wetlands uncatalogued. Methodology was standardized, but observer's bias is hard to avoid in an effort involving dozens of researchers. Also, since the main aim was description of vegetation types, not all species lists are complete, as own point field surveys revealed.
Within field surveyed wetlands, sub-objects of differing characteristics (hydrology, vegetation type) were delineated and assigned to more detailed wetland types. As species lists and wetland type classification are associated with these sub-objects, our analyses are based on this level, i.e. on spatial scales smaller than the main wetland objects.
These points are typical weaknesses of large datasets like the VMI. Fortunately, species distribution modeling methods are designed to deal with exactly such data.

MaxEnt
We treated occurrence data as presence only, acknowledging that absence from a species list did not necessarily imply true absence. For such data, MaxEnt is the method of choice (Philips et al. 2006, Elith et al. 2006). The method is robust also to the limitation to relatively few occurrence points throughout a species' range (emission of small wetlands) and potential sampling bias (field inventories of high nature value wetlands) (Elith and Leathwick 2009). MaxEnt does not project occurrences into geographical space, but directly into n-dimensional environmental space, thus overcoming issues of spatial autocorrelation (Elith et al. 2011). To counter potential sampling bias in terms of environmental space, we used all inventoried VMI wetlands as a targeted background (Philips and Dudík 2008).
As we modeled current and future distributions of species across the whole extent of Sweden, regional predictions for the Norrström Drainage Basin (NDB) are expected to be robust to weaknesses in primary data as well as the lack of data for surrounding countries (for species with ranges that extend further south).
Predictor variables were pre-selected both based on ecological relevance and driven by data (i.e. strong contribution to models across all species run with a full predictor set). Models were run with default values for parameters, applying linear and quadratic features only. Clamping was avoided to allow for extrapolation to new combinations of environmental drivers. 5-fold cross-validation was performed and model results averaged.
Climatic variables were obtained from Worldclim at a resolution of 0.5° (~630 m) (Hijmans et al. 2005). For the future scenario we used predictions by 2070 of the HadGEM2-AO model at an intermediate emission scenario (RCP 6.0). For the Norrström Drainage Basin (NDB), the model predicts an increase in mean annual temperature of +2.85°C, and a decrease in precipitation of -6.7%. A digital elevation model (DEM) was constructed from 50m resolution elevation data from the Swedish Land Survey (Lantmäteriet 2010). Soil and bedrock data at 1km resolution were obtained from the Swedish Geological Survey (SGU 2013). pH point measurements (n = 20 733) from the Swedish National Forest Inventory (SLU 2013) were interpolated at a resolution of 50 m using a random forest model, with soil type, bedrock type, elevation and land cover as predictors. All rasters were resampled to a common resolution of 500 m with either mean or mode algorithms using GDAL version 1.10.1.

Scaling relationships
The relationship between the estimate of abundance from the SDM approach and the potential relative biomass of each species in the community can be inferred by using general allometric scaling relationships between species height, individual biomass and density. Specifically, height (m)

Wetland distribution in the NDB
In terms of SDM, spatial heterogeneity of wetland type distribution and species occurrence points is not problematic. Regarding the potential for ES delivery however, wetland distribution in the landscape matters (Mitsch and Gosselink 2001).
The VMI coverage of NDB wetlands (Fig. A1, A-C) appears to be in general agreement with wetland area recorded in Swedish land cover data (Fig. A1, D). In the VMI data, the distributions of different wetland types follow a similar pattern, except for relatively higher numbers of riparian wetlands in the lowlands surrounding lakes Hjälmaren and Mälaren (Fig. A1, C), consistent with a higher percentage of inland marshes there (recorded as CORINE land cover types, Table A1).
In the NDB, a gradient of increasing human population density and percentage of agricultural area towards south and east (Fig. A1, E) coincides with fewer wetlands in the landscape ( and relatively little agriculture. Differences between region 2 and region 1 cannot be explained by human population pressure, with the exception of potentially more intensive forestry towards the west. The very low percentage of clay/silt soils and the higher relief might possibly explain why there are relatively fewer wetlands in region 1 (tables S1, S2; Fig. A1, F). According to CORINE land cover data, however, peat bogs should cover a larger area in region 1 (table S1); hence, regional sampling bias in the VMI data might partially cause this pattern.
To quantitatively estimate the regional distribution of ES potential, the VMI data should preferably be integrated with other, higher resolution datasets. A quantitative link from landscape characteristics to wetland type distribution, in conjunction with local vegetation, could then be used to model the spatial distribution of ES potential. While this was not the focus of the present study, it constitutes a natural extension on the regional scale. Species list and trait data for hydrophytes.  (Table S5).  Moor H., K. Hylander, and J. Norberg (2015)  Note that changes in species proportions of community biomass are relatively larger where species lists are

Predicted change in relative biomass of species (vascular plants)
shorter. This drawback of our approach might be less pronounced with more extensive species lists, but we believe our species selection covers those species that collectively dominate community biomass. To resolve this issue, field data on realized relative biomass across both field and shrub layers (and potentially even mosses and trees) would be required.

Fig. S3
Strongest predicted change in abundance distribution of bryophytes in the NDB for the three wetland types. Blue indicates decrease, red increase. The scale shows change in suitability relative to current suitability (between 0 and 1).  57 192.48 197.74 196.55