All experiments were carried out in 2013–2014 in the Volgermeerpolder (52°25’17”N; 4°59’35”S), the Netherlands, a newly constructed wetland on top of a former waste dump site that was completed in 2011. Multiple basins ranging from 550 to 1600 m2 were created in this wetland to initiate peat development on top of a sand-covered geomembrane. Basins were formed by clay dikes and the sand substrate in some basins was complemented with clay or organic sludge. The application of different mixtures of sand, clay and organic sludge resulted in a range of organic matter fractions in the sediments (0.01 to 0.23, Fig. 1a, Table 1 and Online Resource 1) in the different basins. The 18 basins used in this study were fed either with rain water (collected in a dedicated basin), or with nutrient-rich surface water from surrounding agricultural fields. Water levels were kept at 60 ± 15 cm above the sediment surface. As a result of the combination of different sediment mixtures with the two water regimes, the basins each developed a unique mineral composition with resulting sediment food quality and availability of micronutrients, macronutrients and trace elements. All experiments and measurements in this study were performed in 2013/2014, three years after the construction of the wetland was completed. During the initial years vegetation in the basins developed depending on the sediment and water composition. Basins with sand sediments remained mainly unvegetated or covered by submerged species (77 ± 7%, mean ± SD), with a helophyte coverage of only 14 ± 6% after 3 years. Basins with addition of clay or organic sludge showed higher coverage by helophytes (40 ± 23 and 71 ± 16%, respectively). Typha angustifolia was the dominant species in basins with added clay sediment, covering around 26 ± 21% of the basins, while Typha latifolia was the dominant species in sediments with added organic sludge (65 ± 22%) (Harpenslager et al. 2018).
Starting three years after construction, various physico-chemical characteristics of surface water (SW), pore water (PW) and sediment (SED) were measured several times in one year, as described in Overbeek et al. (2018, details in Online Resource 1).
Measurements of surface water temperature (T), electrical conductivity (EC) and pH were taken in October – December 2013 and 2014 and April – June 2014 at 10 cm below the water surface using a HQ40D portable meter (HACH-Lange, Tiel, the Netherlands). Surface water samples were taken in November 2013 and February, May, July and December 2014 and filtered before further analysis in the laboratory, using Whatman mixed cellulose ester filters ME24. Pore water samples were taken at the same time as surface water samples at 15 cm depth in the sediment using vacuum syringes attached to ceramic soil moisture cups (Eijkelkamp, Giesbeek, the Netherlands). Alkalinity for surface water of unfiltered samples was determined by titration down to pH 4.2 using an auto-burette with accurately determined titer (ABU901, Radiometer, Copenhagen, Denmark, or Metrohm 716 DMS Titrino, Metrohm Applikon, Herisau, Switzerland). Carbon dioxide (CO2) and bicarbonate (HCO3−) were measured on an infrared gas analyser or high-temperature combustion total organic carbon (TOC) (IRGA, ABB Analytical, Frankfurt, Germany, or TOC-V CPH, Shimadzu, Kyoto, Japan), using unfiltered samples. Nitrate (NO3−), ammonium (NH4+), dissolved organic nitrogen (DON), soluble reactive phosphorus (SRP), potassium (K+) and sodium (Na+) were measured in filtered samples on an auto-analyser (AA3 system, Bran & Luebbe, Norderstedt, Germany, or San ++ system, Skalar, Breda, the Netherlands). Chloride (Cl−), calcium (Ca2+), total iron (Fe), total manganese (Mn), total phosphorus (P) and total sulphur (S) were measured in filtered samples using inductively coupled plasma spectrometry (ICP-OES iCAP 6000, Thermo Fisher Scientific, Waltham, MA, USA, or Optima 8000DV, Perkin Elmer, Waltham, MA, USA).
Sediment samples were pooled from five subsamples per basin using the top 10 cm in February 2014 and the top 5 cm in June 2014, and stored at 4 °C until further analyses. Fraction organic matter (OM) was determined using loss on ignition (LOI, 4 h at 550 °C). Percentage nitrogen (N), carbon (C) and sulphur (S) were measured on an elemental analyser (Carlo Erba NA1500, Thermo Fisher Scientific, Waltham, MA, USA, or Vario EL cube, Elementar, Hanau, Germany). Phosphorus readily available for uptake by vegetation (Olsen_P) was determined using extraction with 0.5 M NaHCO3.
For all measured characteristics, except for temperature, yearly averages were calculated per basin since no seasonal differences were observed within basins. To give an indication of variation between basins yearly averages and percentiles of all basins combined are presented in Table 1 (see Online Resource 1 for values per basin).
As described in Overbeek et al. (2018), the decomposer community of the Volgermeerpolder has developed on top of the mineral sand bed from inocula introduced during the construction and during the pumping of water from the surrounding areas. Furthermore, the open basins serve as a refuge for waterfowl that may have introduced microorganisms and macroinvertebrates.
BIOLOG GN2 plates (BIOLOG Inc., Hayward, CA, USA) were used to determine functional microbial community composition and activity in the sediments three years after wetland construction. BIOLOG GN2 plates contain 95 wells with single simple carbon substrates and one control well (for an overview of all carbon sources see Garland and Mills 1991), specifically designed to make rapid community-level physiological profiles which best differentiate gram negative bacteria (Garland 1997). For each basin, the top 5 cm of five sediment samples taken in June 2014 were pooled together to get a representative sample per basin and stored at 4 °C until further processing the next day. Sediment samples were diluted with sterilized demi water to obtain a dilution of 1:7, shaken by hand for a minute to detach bacteria from sediment particles and centrifuged at 1000 g for 15 min, after which the supernatant was diluted with sterilized demi water to a final dilution of 1:87 (adapted from Hench et al. 2004). BIOLOG plate wells were inoculated with 150 μl bacterial suspension and incubated in the dark at 15 °C to simulate natural conditions. Absorption was measured at 590 nm every 24 h for 7 days (VersaMax microplate reader, Molecular devices, Sunnyvale, USA). The absorbance for individual wells was corrected for background absorbance by subtracting absorbance of the control well and subsequently considering negative values as zero. Average Well Color Development (AWCD), a measure of microbial activity, was calculated according to Garland (1997), by averaging all 95 corrected response well absorbances. For Community Metabolic Diversity (CMD), a measure for microbial diversity, corrected absorbance values were converted to binary data (presence/absence) using a threshold absorbance of 0.25 (Garland 1996). The maximum slope in the sigmoidal response curve for AWCD as well as CMD appeared after three days of incubation, providing the highest distinctiveness between samples. Therefore, measurements taken after three days of incubation were used in further analyses.
Macroinvertebrates were sampled six months after the start of the experiment at the same time as the standing litter biomass (May 2014, see below) and identified to family-level, except for chironomids and oligochaetes, which were identified to tribe and class level, respectively. This taxonomic information was used to estimate the representation of functional feeding groups (FFGs). It was assumed that individuals found in the basins all originated from source populations in the surroundings, and that data on these source populations could therefore be used to determine the functional feeding guilds (FFGs) of the collected individuals in our study without determination to species level. To this purpose, FFGs were determined for all macroinvertebrate species found in an area stretching 5 km around the Volgermeerpolder in the years 2000–2015 (data provided by local water authority, http://hnk-water.nl) using the database from Schmidt-Kloiber and Hering (2015) (Online Resource 2). Subsequently, weighted averages, using the number of times a species was sampled by the water authority in all sampling locations at all times together as weight, were calculated for all FFG fractions per taxonomic family present in the surrounding area. For chironomids and oligochaetes, weighted averages were calculated per tribe and class, respectively (Online Resource 2). The calculated FFG distribution from the source population was assigned to the sampled individuals. When a sampled individual belonged to a taxonomic family which was not present in the source population, the FFG distribution was assumed to be equal to the one given by Schmidt-Kloiber and Hering (2015). Weighted averages of FFGs per sample were calculated accounting for number of individuals sampled per taxonomic level, without correction for size per individual. Gatherers (GAT), shredders (SHR), miners (MIN) and grazers (GRA) are all active in decomposition, therefore those four FFGs together were labeled as detritivores (DET).
To separate the direct effects of environmental variables on decomposition rates of aboveground litter from the indirect effects of litter quality itself, we used common reed (Phragmites australis) as a single standard substrate. Common reed is a common wetland plant with easily distinguishable leaves and culms, therefore making it easy to construct quantifiable substrates for decomposition measurements. By constructing bundles of vegetation, keeping larger fragments together, instead of using litterbags (e.g., Petersen and Cummins 1974; Triska and Sedell 1976; McArthur et al. 1994), we avoided negative influences caused by using litterbags such as altering microclimates within the litterbags or exclusion of certain sizes of decomposer organisms (Boulton and Boon 1991; Bradford et al. 2002; Kurz-Besson et al. 2005; Bokhorst and Wardle 2013).
P. australis was collected from fresh stands growing in a ditch in the Volgermeerpolder at the end of the growing season in September 2013 and dried for several weeks at room temperature in the laboratory. Part of the litter was oven dried at 60 °C to determine the conversion factor between air-dry and oven-dry weight, as well as C, N and S content of the litter (n = 4, 45.0 ± 0.2, 1.7 ± 0.1 and 0.30 ± 0.03% (mean ± SD), respectively). Collected reed material consisted of equal amounts of stems and leaves and therefore both were tied together to 10 g bundles in equal proportions. The conversion factor between air-dry and oven-dry weight was used to calculate initial oven-dry weight of the air-dried bundles to be able to compare them to oven-dry weight at the time of collection. Litter bundles were placed in the field in November 2013 using a random block design with four blocks of 1 × 1 meter each placed in an area of 5.5 × 5.5 m. Each block contained two litter bundles for collection at two different times, resulting in eight litter bundles per basin in total. The bundles within a block were placed about 40 cm apart, and secured to the sediment using pins. Upon placement of the bundles, handling loss was determined to be ~2%. One bundle per block was retrieved after 6 months, while the other half was collected after 12 months by gently but quickly lifting the bundle from the sediment using a net to include macroinvertebrates. Litter bundles and macroinvertebrates were transported to the laboratory in sealed plastic bags and stored at 4 °C until further processing the next day, as described in Overbeek et al. (2018). Litter was gently rinsed and sieved using a mesh size of 1 mm to exclude sediment particles and dried for approximately three days at 60 °C, after which remaining litter mass was determined. The weight difference between standing litter biomass at the start of the experiment (corrected for handling loss) and at time of retrieval was considered to be decomposition and expressed as fraction of the oven-dry start weight to get the fraction of decomposed aboveground litter (Frac_D6 and Frac_D12 for fraction aboveground litter loss after 6 and 12 months, respectively).
Predictor and Response Variables
Linear models (multiple regression, assuming Gaussian errors) were formulated, fitted and validated to establish the relative importance of measured predictor variables for explaining response variables in the early stages of our newly constructed wetland (Burnham and Anderson 2002). To post-process the models, all measured predictor variables were grouped in order to determine the relative importance of specific variable groups, defined as presence of these groups in the model ensemble, in explaining the response variable. Variables were grouped to water (SW), sediment (PW + SED), microbes (CMD + AWCD as measured by BIOLOG GN2 plates) or macroinvertebrates (abundance and FFGs, Online Resource 1), using yearly means for water and sediment variables. Next to these categories, predictor variables were also grouped in macronutrients (all SW and PW N + P), sediment food quality (SED OM + C + N + C_N), or other (see Online Resource 1 for an extensive list). To explain variation in (1.) functional community composition of microbes, (2.) activity of microbes, and (3.) fraction detritivores in macroinvertebrates, these three variables were separately taken as the response variable with all remaining separate variables from water and sediment variable groups after variable selection (see “Variable Selection” section and Online Resource 3a) as potential predictor variables in model formulation. When using (4.) fraction aboveground litter loss after 6 months of decomposition or (5.) fraction aboveground litter loss after 12 months of decomposition as a response variable, all remaining separate variables from all variable groups after variable selection (see “Variable Selection” section and Online Resource 3a) were present as possible predictor variables.
The models were validated in two steps: (1.) by cross-validation, applying fitted models to predict unseen-data measured at the same time-period (both after 6 or 12 months of decomposition, resulting in R2val), and (2.) by extrapolation, applying fitted models (based on data at 6 months of decomposition) to predict unseen-data at 12 months of decomposition (resulting in R2val_t2_with_t1).
The different response variables were observed in different experimental layouts: macroinvertebrate and decomposition observations were laid out in 4 randomized blocks per basin while microbial data were observed only once per basin. For that reason the different response variables were validated differently: microbial data was validated using leave-one-out cross-validation (LOOCV), macroinvertebrate and decomposition data was validated using blocks for cross-validation: using three blocks to predict the fourth.
Formulation of the linear models took place in two steps: (1.) selecting a subset from all predictor variables, using regression trees, and (2.) finding the best linear models using all-possible-subsets regression.
In the first step, regression trees were built using all available scaled predictors. For Frac_D and DET as response variables, where four-fold cross-validation was applied, the regression trees were based on the different combinations of three out of four blocks, after which all potentially important variables resulting from the four regression trees were combined into one list of potentially important predictor variables per response variable. For CMD and AWCD as response variables, a regression tree was fitted only once to the entire data set (see Online Resource 3a for an overview of the resulting potentially important predictor variables per response variable). In the second step, all possible combinations of up to four variables were made from the set with potentially important predictor variables, excluding combinations of variables with an absolute correlation higher than 0.7, and including interactions between number of macroinvertebrate individuals and FFGs. Only those models with an AIC-difference less than 2 from the model with the lowest AIC-value were considered adequate and retained. For the resulting model ensemble the importance and effect of each predictor variable was determined, as well as mean adjusted R2 (R2adj), with variable importance representing the fraction of models in the model ensemble in which the variable was present. The predictors were also aggregated into variable groups (Online Resource 1), and the relative frequency at which each variable group was represented in the model ensemble was used to express variable group importance.
All analyses were performed in R (R Core Team 2015), using functions from the packages plyr, reshape, rpart and ggplot2 (Wickham 2007; Wickham 2009; Wickham 2011; Therneau and Atkinson 2018).