Process-based proxy of oxygen stress surpasses indirect ones in predicting vegetation characteristics

General approach

We compiled a database with site information on species composition, groundwater level and soil type from a large number of sites in the Netherlands. For each site, we computed the mean indicator value of the vegetation for moisture regime, F_m [–] (Witte et al., 2007), considered as the response variable of the vegetation to the soil oxygen status (Figure 1). This variable was regressed against three different proxies of oxygen stress, described in the succeeding discussions.

Daily groundwater levels were interpolated (Figure 2: A; Figure 3) from biweekly measurements. From these groundwater levels, existing indirect measures of oxygen stress, namely mean spring groundwater level [MSL (m above soil surface)] and sum exceedence values [SEV (m day)], were calculated. We considered root respiration stress [RS (kg O₂ m⁻² 10 day⁻¹)] due to anoxic soil conditions as a process-based measure. RS includes the most direct vegetation response to oxygen stress, as respiration is the first plant process that is affected by low oxygen availability in the root zone. For our process-based measure, first daily soil moisture and temperature (Figure 2: B and C, respectively; Figure 3C) were simulated with the dynamic Soil–Water–Atmosphere–Plant (SWAP) model (Van Dam et al., 2008) for the unsaturated zone. Then, the daily respiration reduction (Figure 2: D) was simulated with a recently developed model for oxygen stress to plant roots (Bartholomeus et al., 2008b), which involves macro-scale and micro-scale diffusion, as well as the plant physiological demand of oxygen (Figure 4).

Next, the relationships thus obtained were used for a test of the applicability of these measures outside their calibration range. Thereto, we predicted F_m for the Flemish region in Belgium and compared prediction errors for MSL, SEV and RS. Finally, to investigate how the three measures behave under different future climates, we predicted F_m for four climate scenarios and investigated differences in F_m predicted on the basis of RS versus those by MSL and SEV.

Groundwater and vegetation characteristics data

We used vegetation plots from the datasets of Runhaar (1989) (188 plots) and the Netherlands State Forest Service (Hommel et al., 2007) (178 plots) to establish relationships between proxies for oxygen stress and F_m. The plots originate from a wide range of terrestrial vegetation types differing in succession stages, soil type, soil moisture regime, nutrient availability and soil pH. None of the investigated plots had been under the influence of a major change in hydrological conditions. All plots were representative of natural habitats in the Netherlands, a small and flat country with a temperate climate that has only small spatial differences in meteorological conditions. The mean summer and winter temperature is 289 K and 276 K, respectively, the mean yearly precipitation is 770 mm, and the mean yearly reference evapotranspiration [according to Makkink (1957)] is 563 mm.

Fortnightly measurements of groundwater level data were available in or immediately next to each plot, but only for specific periods and for a limited number of years. However, long time series of groundwater levels are required to define generally applicable relationships, i.e. relationships that are not biased by temporal deviations in meteorological conditions (Bartholomeus et al., 2008a). Therefore, the groundwater level series of all plots of both data sources were extended to the period 1971–2000, which is representative of the actual climatic conditions (Knotters and Van Walsum, 1997; Bartholomeus et al., 2008a) and interpolated to daily values with the Menyanthes impulse response software (Von Asmuth et al., 2002). Menyanthes transforms precipitation and evapotranspiration series (impulse) into groundwater level series (response) (Bartholomeus et al., 2008a) (Figure 3A, B). Daily local meteorological data on air temperature, precipitation and reference evapotranspiration [according to Makkink (1957)] were available from the Royal Netherlands Meteorological Institute (KNMI) from 1970 onwards. We omitted plots without soil descriptions, plots with variable soil surface levels and plots for which a reliable simulation of soil moisture conditions was impossible. Overall, 145 plots were used for further analysis. The same time series of groundwater levels were used to compute MSL, SEV and RS (Figure 2).

A list of moisture indicator values for plant species tailored to the Netherlands and the Flemish region in Belgium was used to compute the arithmetic mean moisture indicator value F_m for each plot. Witte et al. (2007) compiled this list of indicator values from published ecological groups for vascular plants (Witte, 2002; Runhaar et al., 2004), mosses and liverworts (Dirkse and Kruijsen, 1993) and Characeae (Van Raam and Maier, 1993). All plant species present in a plot were used to calculate its average vegetation characteristic in terms of mean indicator value F_m. Following the findings of Käfer & Witte (2004), no weight was given to species abundance. The indicator values range from one, for species from aquatic systems, to four, for species from extremely dry systems (illustrated in Figure 1).

We validated the relationships between proxies of oxygen stress and vegetation characteristics obtained for the Netherlands with data from the Flemish region in Belgium (Huybrechts et al., 2009), neighbouring the Netherlands. Detailed data on groundwater levels and species composition of the vegetation were available for each plot, and F_m was derived for each plot. Similar as for the Netherlands, the groundwater level series were interpolated to daily values and extrapolated to the period 1971–2000 with Menyanthes. All proxies were calculated for each plot. In contrast to the dataset from the Netherlands, only relatively coarse soil information was available for each plot, which introduced noise in the measures where soil properties are involved, i.e. SEV and RS.

Indirect proxies of oxygen stress

Process-based proxy of oxygen stress

Climate scenarios and resulting changes in moisture regime

To test the behaviour of the three proxies under changing climatic conditions, we considered four climate scenarios for the year 2050 (Van den Hurk et al., 2006) to predict changes in MSL, SEV and RS and subsequently in F_m. These national scenarios, developed by the KNMI, are based on general circulation model simulations published in the Fourth Assessment Report of the IPCC (Solomon et al., 2007) and include changes in temperature, precipitation and reference crop evapotranspiration (Table 1). The G and W scenarios comprise a +1 K and +2 K global temperature increase, respectively, but with unchanged air circulation patterns in summer and winter. The G+ and W+ comprise the same global temperature increases, but the ‘+’ indicates changed air circulation patterns (Van den Hurk et al., 2006). We transformed time series from 1971 to 2000 of these three climate variables to the 2050 climate, using transformation software supplied by the KNMI (Bakker and Bessembinder, 2007). In addition, on the basis of the data by Kruijt et al. (2008), we decreased transformed reference evapotranspiration values by 2% to account for the higher water use efficiency of plants at increased CO₂ levels in 2050.

These future precipitation and evapotranspiration predictions were used to simulate future groundwater levels in Menyanthes (see 2.2). Subsequently, MSL and SEV were calculated. Future soil moisture and soil temperature profiles were created with SWAP for each scenario, using as input the future air temperature, precipitation, reference crop evapotranspiration and groundwater level series. Finally, future respiration stress RS was simulated on the basis of these soil moisture and soil temperature profiles.

Statistical analysis

We defined relationships between each of the explanatory variables (MSL, SEV and RS) and F_m through regression. First, we checked whether a linear regression model was suitable. If not, we tested non-linear regression equations. The optimal regression equation was selected on the basis of the Akaike information criterion. Non-linear regression models that were conceptually incorrect, because of local extremes or the exceedence of F_m limits within the measurement and predictions range, were not considered. For comparison of the predictive power among explanatory variables, we applied both the Pearson correlation coefficient r between observed and predicted F_m values and the root mean squared error (RMSE) of the predictions.

For the Flemish validation data, we tested whether observed F_m values deviated from predicted F_m values for each relationship (i.e. MSL, SEV and RS vs observed F_m). To do so, we analysed whether the slope and intercept of the relationship between observed and predicted F_m deviated significantly (P < 0.05) from slope 1 and intercept 0.

Relationships between proxies for oxygen stress and F_m

The relationship between MSL and F_m was best described by a sigmoid regression model (Figure 5A), as it levels off at both ends of the F_m scale, i.e. at deep and at shallow groundwater levels (Witte and Von Asmuth, 2003). MSL did not account for differences in F_m values at sites with relatively deep groundwater levels because, with given meteorological conditions, the moisture regime at such sites is completely determined by soil texture.

The relationship between SEV and F_m (Figure 5B) showed a vertical asymptote at SEV ≈ 0. This relationship was best described by a power function. Most important is the shallow slope for higher SEV values, representing high levels of oxygen stress. This implies that the proxy apparently does not differentiate much in the range of moderate to severe oxygen stress.

The relationship RS versus F_m (Figure 5C) was best described by a linear regression model, with the F_m values rather evenly distributed across the whole range of RS. The predictive power of the relationship was high and comparable with those based on MSL and SEV. Thus, all three measures performed equally well within the chosen calibration range, but the type of relationship was very different with the relationship between F_m and RS being the most straightforward.

Validation of relationships between proxies for oxygen stress and F_m

Figure 6 shows the simulated MSL, SEV and RS for our Flemish validation data. Compared with the Dutch calibration set, the validation data showed additional noise around the regression line for all three proxies. However, there were strong differences among the proxies. Large deviations in the MSL–F_m graph especially occurred at sites with relatively deep groundwater levels (appr. MSL < −0.50 m + s.s.), where the observed F_m values of the Flemish region are systematically lower than those of the Netherlands. This can be ascribed to the inclusion of more loamy soils with deep groundwater levels in the validation set. It is very likely that the additional scatter for SEV and RS is caused, at least partly, by the coarse soil data for the validation data. However, the differences for SEV are larger than for RS. Thus, despite the relatively coarse soil data for our validation data, RS is the only proxy for which predicted F_m values did not significantly deviate from the observed ones (Figure 6).

Vegetation predictions for different climate scenarios

All climate scenarios include an increase in temperature and extreme rainfall events (Table 1). Consequently, like in the experiments of Thompson and Fick (1981) and Tsukahara and Kozlowski (1986), root oxygen stress was expected to increase. However, in contrast to the process-based measure RS, in none of the indirect measures, increased oxygen stress was apparent (Figure 7). This difference was also reflected in the predicted F_m values (Figure 8). The predictions based on MSL and SEV were higher (i.e. drier) than those based on RS for each scenario. For instance, for all climate scenarios, a plot with a predicted F_m²⁰⁵⁰(RS) of 1.8 appeared to have a predicted F_m²⁰⁵⁰(MSL) and F_m²⁰⁵⁰(SEV) of 2.1. Hence, using RS as predictor consistently led to more oxygen stress, especially for sites with a shallow groundwater table.

Process-based versus indirect relationships between oxygen stress and vegetation

Even though the two indirect proxies (MSL and SEV) and the process-based proxy for oxygen stress (RS) performed equally well in describing vegetation characteristics for the Netherlands under the current climate, only RS performed well in the vegetation validation. Our process-based proxy thus outperforms indirect ones, as relationships based on the latter seem only valid for the region in which they were determined and calibrated. This is most likely due to the fact that the use of root respiration reduction, which is the most direct plant physiological response to oxygen stress, allows the inclusion of all environmental drivers, including climate, in the estimates.

Assumptions on model parameters, schematization of relevant processes and numerical discretizations in our detailed simulation of soil moisture, soil temperature and soil oxygen (Figure 2) potentially introduces errors. Therefore, adding process knowledge may be at the cost of the goodness-of-fit of derived relationships. Despite this drawback, the explained variance of the relationship between F_m and RS is comparable with the explained variances of the relationships between F_m and MSL and SEV, which indicates that RS is a robust predictor for F_m. This is also confirmed by the strong predictive power of RS across the complete range of F_m values, which indicates that we considered the main processes that are relevant to vegetation composition. The fact that respiration stress is also a discriminating factor at relatively dry sites (F_m = 3–4) (Appendix A) demonstrates that it captures the entire moisture gradient. It is reasonable to expect, however, that drought stress will be the primary driver for vegetation characteristics at dry, groundwater-independent sites (Runhaar et al., 1997) and that respiration stress is only related to this driver in an indirect manner, i.e. through soil texture (Appendix A).

By using F_m, we are able to predict the type of plants that grow at a site. If desired, F_m can be translated into the occurrence probabilities of specific species groups (Witte et al., 2007). Instead of F_m, also plant traits may be used to characterize the vegetation (Suding et al., 2008). From a scientific viewpoint, the use of measurable plant traits that link directly to the most important drivers of vegetation responses would be ideal to relate vegetation characteristics to soil oxygen. However, because there are many physiological responses to oxygen stress, it is difficult to combine all possible adaptations into a single variable. In addition, information on physiological adaptations to oxygen stress for all representative plant species is lacking. Because such information is difficult and laborious to obtain, it is unlikely that it will become available in the near future. Considering both the strong relationship between oxygen stress and F_m and its proven applicability outside the measuring range, it seems that the concept of indicator values is, at least for the time being, the best alternative for plant traits.

Applicability of proxies for oxygen stress for climatic projections

Our analysis illustrates that there are differences between vegetation predictions for future climatic conditions that are based on RS versus those based on MSL and SEV (Figure 8). Our analysis suggests that the indirect proxies of oxygen stress may underestimate the increased oxygen stress under changing climatic conditions at a given future groundwater level dynamics. These differences in calculated oxygen stress occurred despite the fact that the same groundwater levels were used as input. In contrast to RS, especially the highest values for MSL and SEV (thus highest oxygen stress) are almost equal to those for the actual climatic conditions (Figure 7). Consequently, both MSL and SEV predict higher F_m values (i.e. drier vegetation) than RS (Figure 8). However, especially at shallow groundwater levels, large shifts in oxygen stress are expected with climate change, as the combination of a higher soil moisture content and an increased temperature will lead to increased oxygen stress (e.g. Sojka et al., 1972; Drew, 1983). The differences among F_m²⁰⁵⁰(RS), F_m²⁰⁵⁰(MSL) and F_m²⁰⁵⁰(SEV) were not caused by the choice of the non-linear regression models for MSL and SEV, because for each of these measures, an alternative design by a linear regression without asymptotes based on F_m < 3.0 led to highly similar differences in predicted F_m values.

The differences in predicted vegetation characteristics may have major consequences. For example, according to our findings, differences of 0.5 on the F_m scale are expected. Under the actual climate, this comprises a difference in MSL of about 0.25 m, which in turn determines the occurrence versus the absence of, for example, Molinea meadows or Sphagnum dominated bogs (Runhaar et al., 2003). Therefore, using process-based or indirect measures in habitat distribution models could lead to systematic and ecologically significant differences in the evaluation of the impact of climate change on vegetation.

We based our current climate relationships on a combination of two very detailed datasets, both of which were originally specifically compiled to analyse soil moisture–vegetation relationships (Runhaar, 1989; Hommel et al., 2007). Unfortunately, such a specific combination of accurate data is currently unavailable for different climates, obstructing the possibilities for validating the predicted F_m values for such climates. Nevertheless, there are several arguments that support the assumption that RS actually leads to more reliable vegetation predictions than MSL or SEV. First, the oxygen consumption of and supply to plant roots strongly depend on soil temperature, a variable that has been incorporated only in RS (2.4.2). Second, both MSL and SEV require assumptions on the starting date and length of the growing season, which are to some extent arbitrary and which will change with the changing climate. Third, extreme precipitation events are known to have a major effect on vegetation characteristics, which is expressed in RS, hardly in MSL or SEV. Fourth, unlike RS, the relationship of MSL and that of SEV versus vegetation characteristics have to be recalibrated for all new environmental conditions (as shown by Figure 6), which is impossible for a future climate.

On the basis of these arguments, we argue that vegetation predictions based on relationships that have relatively low levels of causality (i.e. indirect measures) should not be used to predict the vegetation response under future climatic conditions. As the predictions based on process-based and on indirect proxies differ, our analysis ratifies the need for process-based relationships in habitat distribution models. We advocate using RS as process-based measure to predict vegetation characteristics at groundwater-dependent systems, as it is based on relevant plant physiological, soil physical and climate-dependent processes.

We applied the respiration model to a hypothetical reference vegetation to obtain a site characteristic, independent of the actual (adapted) vegetation. The use of a reference vegetation improves the applicability of models in which stress measures are implemented, especially in predicting climate change effects (Dyer, 2009). The yearly average simulated soil respiration rates (for our reference vegetation, defined as a temperate natural grassland) were 1.47 and 1.06 kg O₂ m⁻² year⁻¹, based on potential and actual root respiration, respectively. These simulations correspond reasonably well to measured average values for temperate grasslands (1.18 kg O₂ m⁻² year⁻¹; Raich and Schlesinger, 1992) and respiration ranges across different terrestrial biomes (0.1 to 3.5 kg O₂ m⁻² year⁻¹) in general (Raich and Schlesinger, 1992).

For the calculation of RS, we selected the period with the highest cumulative respiration reduction for each year. Focussing on these extreme events may be crucial to capture climate change effects on vegetation. For instance, suboptimal soil oxygen conditions do not necessarily impair the vegetation, because normal metabolism of plants is flexible, responding to moderate fluctuations in environmental changes (Gaspar et al., 2002; Körner, 2003). Therefore, the events that deviate most from the average conditions, i.e. the extremes, will have most impact on the vegetation (Chapin et al., 1993; Knapp et al., 2002; Weltzin et al., 2003; Van Peer et al., 2004; Bokhorst et al., 2007). To account for these extremes, we selected for each simulation year the 10-day period with highest reduction in plant metabolic functioning, i.e. in respiration reduction, because a period of 10 days of oxygen stress has been shown to hamper the plant metabolism already (Huang et al., 1998). However, using a period of, for example, 7 or 14 days led to highly similar F_m²⁰⁵⁰(RS) values for each of the climate scenarios [average absolute difference 0.01 (±0.01)]; thus, the proxy is robust for the period chosen.

Application in ecological prediction models

By including the relevant processes and the use of a site-independent reference vegetation, our novel proxy RS to characterize oxygen stress will be applicable to vegetation types and regions that differ from those considered in this research. The process-based approach makes superfluous the calibration of the RS versus F_m relationship in other regions, as shown by Figure 6. By defining the stress of a reference vegetation, specific information about the actual vegetation characteristics at a site is not required. This increases the applicability of our approach, because databases that combine the required environmental data and actual vegetation information are currently lacking (4.2).

Given the demonstrated ecological significance of process-based relationships, incorporating them into habitat distribution models is advised. The calculation of RS requires additional simulations compared with MSL and SEV, which may be considered undesirable or unfeasible. For the calculation of actual root respiration, we therefore constructed repro functions that reproduce the behaviour of the full respiration model Appendix B(Appendix B). The application of these repro functions simplifies and speeds up the calculations considerably. Still, the use of repro functions does not diminish the amount of soil data that are needed as input for the simulations. However, even when only coarse approximations of soil type, soil moisture and soil temperature are available, RS can be calculated. The model predictions thus obtained correspond quite well to the predictions based on the detailed simulations (Appendix B).

Conclusions

We have demonstrated that our process-based explanatory variable for oxygen stress surpasses indirect ones used so far to predict vegetation characteristics. We further suggest that indirect measures may be insufficient to analyse the effect of climate change on vegetation composition. Essential parameters that influence the occurrence of oxygen stress and that are expected to change in the future climate – higher temperatures and extreme rainfall events – are insufficiently accounted for. Hence, we propose a process-based relationship between oxygen stress and vegetation characteristics in terms of respiration stress. Our novel measure of root oxygen stress is an important quantitative step following the recommendations of previous studies (Franklin, 1995; Guisan and Zimmermann, 2000) to define relationships with high levels of causality that are indispensable for robust vegetation predictions.