Seasonal variability as a source of uncertainty in the West Siberian regional CH₄ flux upscaling

The field experiments were carried out during the 2011 summer period. For comparability, these measurements were taken in the same season as the majority of the measurements in a previous study (Glagolev et al 2011): 9–12 July, 26–28 July and 20–21 August. Methane fluxes were measured at two test sites in the subtaiga zone of Western Siberia: 'Biaza' (56.50°N, 78.31°E) a topogenous fen, and 'Chuvashi' (56.41°N, 78.80°E), a soligenous poor fen, according to a Russian mire classification scheme (Masing et al 2010). Detailed test site descriptions are in appendix A.

2.2.1. CH₄ flux measurements

2.2.2. Data analysis

2.2.3. Mire ecosystem concept

2.2.4. The methane emission period model

The inventory concept (as well as other inventories, for example, Bartlett and Harriss 1993, Christensen et al 1995, Wang et al 2005) accounts for seasonal variations of methane emissions by dividing the year into three periods (Suvorov and Glagolev 2007). The methane flux is first assumed to increase from winter to summer and then after peaking it regularly decreases (figure 1). In the modeled version a step function is used to represent this seasonality, where the flux is

• (a)  
  zero: from the beginning of the year to the day τ corresponding to the beginning of the methane emission period (MEP);
• (b)  
  the flux median for each mire ecosystem types in each climatic zone: from τ to τ + MEP, corresponding to the end of MEP;
• (c)  
  zero: from τ + MEP to the end of the year.

Zoom In Zoom Out Reset image size

The parameters defining the MEP were chosen on (1) the basis of the biologically active period whose start is determined from the date when mean daily air temperature becomes higher than 10 °C and ends the when mean daily temperature becomes lower than 0 °C (Rihter 1963) and (2) the analysis of seasonal variations in methane flux from several south taiga wetlands. For the south taiga zone, the MEP was calculated as the ratio of the total emission in the snow-free period (calculated by a numerical integration of the methane emission curve during the snow-free period) and the median flux from the same wetland and was about 10% longer than the biologically active period (Suvorov and Glagolev 2007). Because the MEP is derived from the median value of the flux measurements and the integrated annual flux, asymmetry in the flux measurement distribution would not distort the use of the median (Glagolev et al 2010). The MEP values for other zones were calculated assuming that their MEP value is proportionately larger than the biological activity period by as much as the MEP value is larger than biological activity period in the south taiga.

Fluxes from the two field sites are reported in appendix C and table 1. At the fen site, the CH₄ fluxes were found always higher compared to the poor fen site (the median across the summer season at the fen site is 3.20 mgCH₄ m⁻² h⁻¹). The fen site fluxes showed a clear seasonal trend, with the occurrence of the lowest fluxes at the beginning of July followed by an increase in the each measurement period's median flux until the end of August. The water level was above the moss surface during most of the measurement periods at this site except for the beginning of July when it was slightly below the moss surface. At the poor fen site the average CH₄ fluxes were about one-third lower (the median for all seasons is 1.25 mgCH₄ m⁻² h⁻¹) than the fen site and showed the highest median values in the second session of measurements in the end of July. There was more scatter in these data during the final measurement period, with the highest fluxes (2.83 mgCH₄ m⁻² h⁻¹) measured then. The water level in the poor fen site remained 10 cm below the moss surface and never reached the surface during the full measurement period.

Typically, the uncertainty of the individual measurements was around ±0.2 mgCH₄ m⁻² h⁻¹ for both sites, with the highest uncertainty contributed by scatter in the measured gas concentrations. The temporal variability within the test site was estimated as the standard deviation of the average across the three measurements taken in a row (i.e., within 1.5 h) and was also around 0.2 mgCH₄ m⁻² h⁻¹.

The relative seasonal variability (calculated as the ratio between the standard deviation of the median flux of each measurement session to median fluxes across the whole season) were 26% and 8% for the poor fen and fen, respectively. The significantly higher seasonal flux variations in the poor fen derive from seasonal fluctuations of the water table level. In the second half of July precipitation in this region was high (about 120 mm) and resulted in a ∼40 mm rise in the water table of both study sites (see table 1). At the end of August the hydrological conditions were changed: the poor fen with the better drainage became slightly dryer, while the inundation level of the weakly drained fen continued to grow. Thus, the water table level of the poor fen varied more widely across the measurement period. Since both mires have similar regression lines between CH₄ flux and water table level, their hydrological differences resulted in a higher variability of methane emissions in the poor fen.

Regression analysis revealed that it was possible to construct a multidimensional model with independent and significant parameters for all microsites as well as separately for the fen and poor fen (figure 2). The model for all microsites (F_a, mgCH₄ m⁻² h⁻¹) contains WTL and T₁₅:

$$\begin{eqnarray*} &&F_{\mathrm {a}} = 5.13\pm 0.91 + (0.086\pm 0.009) \mathrm {WTL}\nonumber \\ &&\quad\qquad -\, (0.19\pm 0.07) T_{15};\quad\qquad R^{2} = 0.83,\quad\qquad \nonumber \\ &&\mathrm {RMSE} = 0.45~{\rm mgCH}_{4}~{\rm m}^{-2}~{\rm h}^{-1}. \end{eqnarray*}$$

The model of CH₄ fluxes in the fen (F_f, mgCH₄ m⁻² h⁻¹) also contains WTL and T₁₅ as explanatory variables:

$$\begin{eqnarray*} &&F_{\mathrm {f}} = 4.74\pm 0.98 + (0.11\pm 0.02) \mathrm {WTL}\nonumber \\ &&\quad\qquad -\, (0.17\pm 0.08) T_{15};\quad\qquad R^{2} = 0.77,\quad\qquad \nonumber \\ &&\mathrm {RMSE} = 0.42~{\rm mgCH}_{4}~{\rm m}^{-2}~{\rm h}^{-1}. \end{eqnarray*}$$

For the poor fen (F_pf, mgCH₄ m⁻² h⁻¹), WTL and pH_mean are the key explanatory variables:

$$\begin{eqnarray*} &&F_{\mathrm {pf}} = (0.07\pm 0.02) \mathrm {WTL} + (0.45\pm 0.06) {\mathrm {pH}}_{\mathrm {mean}};\quad\qquad \nonumber \\ &&R^{2} = 0.92,\quad\qquad \mathrm {RMSE} = 0.43~{\rm mgCH}_{4}~{\rm m}^{-2}~{\rm h}^{-1}. \end{eqnarray*}$$

Zoom In Zoom Out Reset image size

A simple one-dimensional regression reveals that for both sites as well as for each site separately WTL is the single best predictor of emissions (where the median flux for the three temporal replicate measurements is used; see table 2). For the poor fen site, the peat water chemical characteristics (EC and pH) are also significant. Conversely, for the fen site only T₅ and T₁₅ are also significant. The highest deviations from the WTL regression curve for the fen site (i.e., the two significantly lower values) may be caused by the relatively small coverage of vascular plants (10–20%) compared to the other points whose coverage is more than 30%. For the poor fen site, the substantially higher and lower residuals were correlated with the highest and the lowest values of pH_max, respectively.

The seasonal variation of our methane flux data is 8% for the fen with weak drainage and 26% for the poor fen site with relatively good drainage. Similar calculations using various published data (Heikkinen et al 2002, 2004, Friborg et al 2003, Ding and Cai 2007) give the seasonal variation of 11–30% (see table 3). For the fen intermediate flark site (Heikkinen et al 2004) the relative seasonal variability is 11% and for non-degraded and slightly degraded polygon tundra sites (Sachs et al 2010) it was 21% and 16%, respectively. Because other fen sites even in different climatic zones also have seasonal variability sufficiently less than 30%, this low variation may be presumed to be a general pattern for fens. In their general characteristics these mires are similar—they all are fens on watersheds with relatively shallow peat depth and weak drainage (caused by clay parent rock or permafrost). In general the bell-shaped curve of methane emission for Western Siberia mires has two main contributors, which are the similarly shaped seasonal precipitation and air temperature curves (Heikkinen et al 2004, Naumov et al 2007, Naumov 2011). Such a pattern is typical for all regions with a continental climate.

At this study's poor fen site, where water continuously drains from the mire during the summer period, the methane emissions generally follow the changes in WTL. This relationship has also been discovered for the poor fen site in Friborg et al (2003) and the bog site in Ding and Cai (2007). As an example, the seasonal dynamics from Friborg et al (2003) were successfully predicted with a WTL model (Bohn et al 2007). But if the drainage from a fen is very poor, then there is often no significant correlation between WTL and CH₄ flux and only temperature acts as a control of emission (e.g., Heikkinen et al 2004, Sachs et al 2010). Thus based on the available data two types of seasonal emission dynamics could be defined:

• (1)  
  with a significant correlation between CH₄ fluxes and WTL on a seasonal time scale as typical for the raised bogs and transition (or soligenous) poor fens; the relative seasonal variability is about ±30% (Friborg et al 2003, Ding and Cai 2007);
• (2)  
  without a significant correlation between CH₄ fluxes and WTL on a seasonal time scale as typical for the poorly drained fens; the relative seasonal variability is ±(10–20)% (Heikkinen et al 2004, Sachset al 2010).

The sites in our study also are in agreement with this pattern (see below in section 4.4). It additionally should be noted that local weather conditions and climatic anomalies can change this general pattern. However, the two divergent seasonal flux–WTL patterns should encourage the inclusion of detailed wetland type classification schemes into large-scale modeling efforts because different wetland types can have opposite responses to changing climate characteristics.

An estimate of seasonal variability can be compared with other sources of variability to determine the greatest contributor to uncertainty in the regional flux estimate. This comparison can help to generate more effective field campaigns to improve the quality of regional flux estimates. Contributions of other types of variability were determined using measurement data from a methane emissions database (Glagolev et al 2011) and elsewhere and are synthesized in table 3. The potential influence of diurnal variability was not analyzed in contrast with studies from China (Ding and Cai 2007), because significant evidence of CH₄ flux diurnal variations have not been found in previous West Siberian studies (Heyer et al 2002, Sabrekov et al 2011).

Unfortunately, there are not enough data to estimate variability among different mires properly. This type of variability is especially hard to determine in the field because fluxes in different places should be measured at more or less the same time. With the unknown strength of inter-mire variability and the high contribution of local heterogeneity in flux upscaling (table 3), we conclude that for regional flux estimates the spatial coverage of a field investigation is more critical than temporal coverage. We can then advise as a best strategy that field campaigns should investigate more sites with only a few days of observations per site. There may be diminishing benefits to conducting more than 50–70 flux measurements on each site, since according to table 3 this number is adequate to represent spatial variability within mire ecosystems and a greater number of measurements is not necessary. Improvements in constraining site-to-site variations in flux strength should lead to improvements of the regional flux estimate because they allow the aggregated flux estimate to incorporate spatial variations within the ecological diversity of mires. It will help to remove the highest uncertainties in the regional flux estimate despite the continued difficulty in accounting for interannual (including climate-related trends) and interseasonal variations.

The fact that spatial variability is of more importance than temporal variability implies that such factors as local hydrology, geochemistry, and plant community are more important than seasonal variations in environmental conditions. These factors can vary among different mires within one climatic zone (Glagolev et al 2010, 2012) and also within a single mire (Wagner et al 2003, Kutzbach et al 2004, Kao-Kniffin et al 2010), and hence should be evaluated with a spatially oriented sampling approach. These local factors may also be particularly susceptible to climate changes that induce shifts in vegetation community and regional hydrology. Furthermore, an emphasis on greater spatial, rather than temporal, coverage during field investigations may allow the detection of methane emission hot spots that can have proportionately large impacts on the regional estimate. This approach may generate a trade-off with 'hot moments' of flux emissions, such as through ebullition, although these events are often also missed in temporally intensive campaigns (Stamp et al 2013).

The seasonal variability of methane emissions needs analysis from some other points of view, including whether the step function approach is justified for climatic zones outside the south taiga. In this respect a comparison between the most northern (tundra) and the most southern zone (forest steppe) of West Siberia can be instructive. For the tundra zone the comparison can be based on measurements carried out near Vorkuta in the European part of Russia near the border with West Siberia (67.38°N, 63.37°E) (Heikkinen et al 2002, 2004). These studies indicate MEP values (95 days) similar to the MEP of 101 days used in Glagolev et al (2011). Wetlands of north-eastern China (47.58°N, 133.52°E) are close to the forest-steppe mires of Western Siberia by their climatic conditions (see the methane flux measurements in Ding and Cai 2007). For these mires the calculated value of MEP ranges between 195–210 days; thus slightly higher than the MEP of 196 days taken for forest steppe in Glagolev et al (2011). Therefore we can conclude that the MEP values reported in our previous regional flux estimate are reliable and therefore median is here a more statistically robust characteristic than the mean.

Another critical question is whether the method generates an overestimation of the regional flux because the median flux is calculated for the months with the most intensive methane emissions and is further multiplied by the MEP as calculated from flux data from the full snow-free period. The majority of the measurements used to calibrate the model (Glagolev et al 2011) were carried out during the period from late June to early September. In reality, methane fluxes become higher than the mean winter flux earlier (from April to beginning of June, depending on climatic zone) and decrease again only after the measurement period (i.e., from September to November, also depending on climatic zone). To determine this excess seasonal methane flux variation, data from other studies were analyzed. On a preliminary basis, our data was linearly interpolated on a daily basis to compensate for data gaps in different parts of the emission season. Other studies have suggested that the ratio of median fluxes for the entire snow-free period to the period from the end of June until the beginning of September was 0.8 for tundra mires (Heikkinen et al 2004), 0.65 for south taiga mires (Maksyutov et al 1999) and 0.54 for the Chinese analog of Siberian forest-steppe mires (Ding and Cai 2007). The reduction of this ratio from north to south probably results from the longer period of ideal methane production in the more southern zones. Thus, the artificial overestimation caused by measurements conducted in the most intensive periods of methane emission may be 20–45%, with greater errors in more southern climate zones.

Another question, arising from lack of flux data in spring and autumn periods, is the reliability of the estimate of seasonal variations. Of course when the wider time range is considered a higher level of seasonal variation is obtained: methane emissions starts from near-zero fluxes in the spring and decrease to near-zero in autumn. It is important to mention that our work is not about seasonal dynamics of methane emissions directly but about the influence of these variations on the regional flux estimate. The changes of the total warm season flux (which is the area under the curve of season emission) are determined due to uncertainties deriving from the seasonal dynamics of methane emissions. All of the data for the previous regional flux estimate (Glagolev et al 2011) were obtained for the late-June–late-August period and so the estimate of seasonal variability should be calculated within this time. According to the MEP concept the exact variation in fluxes from spring to autumn is not important, rather it is important how fluxes vary within the period for which our flux data was obtained. Of course, this period should not be very short. The summer flux is the largest part of the warm season flux (as described above it composes 65–90% of the total flux depending on the climatic zone). Thus, a lack of flux data in the spring and autumn periods is less critical for our analysis. But, for example, if the flux and seasonal variability had been measured only in July, it would become critical for the total warm season flux and variability estimation because the July flux is not the biggest part of the warm season flux. Calculating the potential overestimation in this case would also be statistically unfounded.

Seasonal methane flux dynamics are more or less symmetric relative to the maximum of the methane emission for West Siberia mires (this study, Maksyutov et al 1999, Friborg et al 2003 and Heikkinen et al 2004) or for ecologically similar mires from other regions (Ding et al 2004, Ding and Cai 2007). Taking into account this fact and the overestimation described above we can make a conclusion about the time of year when the observation would be most accurate from the seasonal dynamic point of view if it is possible to only observe fluxes on a few days of the year per site. These 'optimal' times are the end of the June through the beginning of the July and the second half of August. Measurements taken in these periods would both exclude the overestimation described above and will account for nearly all seasonal variations in flux strength. While this approach has some uncertainty, if the level of seasonal variability is not higher than 30% then the benefits of higher spatial coverage outweigh the potential error in temporal coverage.

It was observed that CH₄ fluxes increased significantly with increases in the water table. This finding is generally expected because lower water table levels favor aerobic respiration and CH₄ oxidation (Dise et al 1993, Bubier et al 1995). A reduction in methane fluxes with increasing T₁₅ does not fit a directly causal pattern (see Dise et al 1993 and Pelletier et al 2007 for counter observations) but can be caused by evapotranspiration. It has been suggested that greater evapotranspiration under the influence of higher soil temperatures will lower the water table and hence methane fluxes will decline (Moore and Roulet 1993, Ehhalt et al 2001). Also hummocks with lower WTL (and hence CH₄ flux) tend to have higher peat temperatures (see appendix C). Methane emissions in the poor fen are positively correlated with mean pH of the surface water. Because the poor fen is soligenous, the chemical characteristics of the peat water change during the season in response to hydrological events; the lack of such changes in the fen may help preclude a significant correlation with pH there. WTL may be the best predictor of emissions across a landscape because there has not been much variability observed in subsurface temperatures (Heikkinen et al 2004, Sachs et al 2010).

The sensitivity of CH₄ fluxes in subtaiga mires to WTL fluctuations (both in the multidimensional and one-dimensional models) was slightly higher than for mires of the same type in New Hampshire, USA (Treat et al 2007). Probably in the New Hampshire study the reduced fluxes result from a low WTL (30–40 cm below the surface) which results in a larger aerobic zone conducive to methane oxidation (see Kalyuzhnyi et al 2009). Fluctuations of WTL in this range do not play a significant role in the CH₄ flux. For example, in a previous peatland study (Pelletier et al 2007) WTL varied in range −30 to 20 cm, as well as in our investigation, and the coefficients are similar (0.10–0.13).

Generally, in our flux measurements we have both types of variability: seasonal and spatial within the mire ecosystem. To distinguish them the simple one-dimensional regression between CH₄ flux and WTL for each site for each of three sessions is provided (see table 4).

A comparison between the coefficients of determination in table 4 and table 2 reveals that for the fen site on the local spatial scale the correlation between WTL and CH₄ flux is as strong and significant as for whole fen site's combined data (where R² = 0.64, see table 2). Hence for the fen site the spatial variability of WTL is more important than seasonal changes in WTL. For the poor fen the WTL-flux relationship within a single measurement period is insignificant but taken together across the season, the correlation is significant (see table 2). Therefore, for the poor fen site the significance of the WTL-CH₄ relationship derives from correlated seasonal changes in both WTL and CH₄ flux while spatial variations for each measurement session do not generate significant correlation between WTL and CH₄ flux. However, the seasonal and spatial coefficients of determination are very similar, so both these potential sources of uncertainty should be considered.

The higher correlations between CH₄ flux and environmental controls for both sites together rather than for each site separately (see table 2) may be explained by the higher importance of spatial variability in methane flux than seasonal variability—combining the datasets significantly increases the spatial sampling range while similar seasonal characteristics affect each site. The alternate explanation is that these findings are just an artifact of including only two mires that differ in mean flux and WTL values. The following normalization procedure was used to test these ideas: we compute the anomalies of CH₄ flux and environmental variables relative to long-term mean values of these quantities at each site and then compute a two-wetland simple one-dimensional correlation between the flux and driver anomalies. In this approach only the correlation with WTL was significant (the coefficient of determination (R²) is 0.51). So for all variables from table 2 the higher correlations derived from combining the two peatlands are artifacts from combining two quite different mires.