Methane emissions from soils: synthesis and analysis of a large UK data set

Introduction

Methane (CH₄) is the second most important trace gas contributing to the radiative forcing of the atmosphere. The largest term in the global CH₄ budget is the emission from wetland soils, as a product of anaerobic decomposition of soil organic matter by methanogenic micro-organisms (Gorham, 1991; Lelieveld et al., 1993; Denman et al., 2007), although global-scale estimates of this vary widely (13–160 Tg yr⁻¹, Petrescu et al., 2010). Soils can also act as a significant sink for CH₄, via oxidation by methanotrophic bacteria, and the net efflux is the balance between production and oxidation (Dalal & Allen, 2008; Lai, 2009). Methane emissions are expected to increase in future in a warmer and wetter climate, and in response to changes in land management, particularly restoration and re-wetting of peatlands (Denman et al., 2007; Waddington & Day, 2007). With emission reduction targets now in place, measures to reduce greenhouse gas emissions through land management are being investigated as a potential means to mitigate climate change. Driven by this, there is a strong demand from policy-makers for simple emission factors, which quantify the effect of land management activities on the net emission of greenhouse gases, but as yet these are poorly developed (Couwenberg et al., 2011; Eggleston et al., 2011). As part of this, we require a better understanding of how the emissions of CH₄ (as well as CO₂ and N₂O) respond to changes in environmental factors (Segers, 1998; Conrad, 2009). Also, given predictions of climatic change, there is a need to know how CH₄ fluxes will respond in the long term, and emissions may increase in a warmer and wetter climate. Modelling work has shown the potential for positive feedback with climatic change (Gedney et al., 2004), particularly in arctic and boreal regions (Koven et al., 2011).

Measurements of CH₄ flux are most commonly made using static chambers, whereby a small area of the land surface (typically ca. 0.1 m²) is enclosed, and the rate of change in methane concentration measured in a sequence of gas samples, analysed subsequently on a gas chromatograph. It is commonly observed that these small-scale measurements are very variable in space and time, making their interpretation in terms of processes difficult. This is because (i) the balance between CH₄ production and oxidation is a complex interaction, influenced by biotic diversity and activity, soil water, oxygen concentrations within the soil, redox potential and transport processes within the soil–water–plant–atmosphere (diffusion, ebullition and transport via aerenchyma and xylem), and (ii) the driving variables are themselves heterogeneous in time and space, and often difficult to measure appropriately. If these chamber measurements are to be used to estimate fluxes at larger scales or over longer time periods, a robust method for scaling up is required which accounts for the key environmental driving variables and their interactions, that can be calibrated over an appropriate range of conditions, and with a suitably large sample size (Baird et al., 2009a). Such a method is needed if CH₄ emissions from natural wetlands are to be included appropriately in IPCC greenhouse gas inventories (IPCC, 2003; Eggleston et al., 2011). Models of CH₄ flux range from simple empirical or statistical models (e.g. Couwenberg et al., 2011), to detailed mechanistic models which explicitly represent many of the processes involved, including CH₄ production, oxidation, transport via diffusion, ebullition and plant stems (Arah & Stephen, 1998; Walter et al., 2001). Because of the computation time and coarse spatial resolution in regional- and global-scale models, relatively simple models of CH₄ emissions are better suited in this context.

Here, we have collated a large number (4831) of existing chamber measurements of soil CH₄ flux made in the United Kingdom, covering a range of vegetation types, and including both mineral and organic soils. The overarching aim of this study was to analyse this data set in terms of a number of co-located measurements of environmental and ecological variables, to assess the appropriate degree of complexity for modelling. Given the explanatory variables that are generally available, we aimed to derive a parsimonious model with the least requirements for environmental inputs, suitable for calculating national-scale CH₄ emissions from ecosystems, or inclusion within a regional-scale land surface model (e.g. Clark et al., 2011). As an example application of such a model, we then derived a simple emission factor for peatland restoration based on the modelled responses to water table changes.

Methods

Field sites and measurements

The measurements analysed here were made at 21 field sites across the United Kingdom (Fig. 1; Table 1). At most of these sites, measurements were made approximately monthly over one or more years, covering all seasons. The exceptions were Glensaugh, Grenstein, Hafren, House O'Muir and Ullapool, where measurements were made in only one or two campaigns. At each site, a similar measurement procedure was carried out, as described below. Where there are important differences in the details of the methodology (chamber size, gas sampling method, etc.), these are listed in Table 1. A cylindrical PVC collar was inserted into the soil and left in place for a number of weeks or months. On each sampling occasion, a lid was sealed on top, and left in place for up to 2 h, but more commonly 30 min to 1 h. Samples were removed by syringe through a 3-way tap or rubber septum, and analysed on a gas chromatograph, together with replicates of three or four standard gases with known concentrations. For each sequence of gas samples from a chamber, the flux (F, in μmol m⁻² s⁻¹) was calculated as:

(1)

where dC/dt is the rate of change in concentration (C, in μmol mol⁻¹) with time (t, in s), ρ is the density of air in mol m⁻³, V is the volume of air within the chamber in m³, and A is the surface area within the chamber in m². dC/dt was calculated from the time sequence of C with t, either from the mean value of dC and dt or by linear regression.

Table 1. Location of field sites, number of measurements available, the sampling and fitting methods used, and published references to the data or site

Site	Lat	Lon	Soil class	Sampling method used	Fitting method used	References (to site or data set)
Auchencorth	55.79 N	3.24 W	Organic	1	1	Drewer et al. (2010)
Clocaenog	53.08 N	3.38 W	Organo-mineral	2	1	Unpubl.
Cow Park	55.85 N	3.22 W	Mineral	1	1	Jones et al. (2005)
Easter Bush	55.86 N	3.21 W	Mineral	1	1	Unpubl.
Forsinard	58.37 N	3.97 W	Organic	4	2	Gray (2007)
Glensaugh	56.90 N	2.52 W	Organo-mineral	1	1	Unpubl.
Grenstein	52.73 N	0.81 E	Mineral	1	1	Unpubl.
Hafren	52.63 N	3.95 W	Mineral	1	1	Unpubl.
House O'Muir	55.86 N	3.25 W	Mineral	1	1	Unpubl.
Lincoln	53.32 N	0.55 W	Mineral	1	1	Drewer et al. (submitted)
Loch More	58.39 N	3.60 W	Organic	1	2	Macdonald et al. (1998)
Migneint A	52.99 N	3.80 W	Organic	2	2	Unpubl.
Migneint B	52.99 N	3.80 W	Mineral	2	2	Unpubl.
Migneint C	52.99 N	3.80 W	Organic	3	1	Unpubl.
Moor House	54.65 N	2.45 W	Organic	5	1	Ward et al. (2007)
Peaknaze A	53.47 N	1.91 W	Organo-mineral	2	1	Unpubl.
Peaknaze B	53.47 N	1.91 W	Organic	2	2	Unpubl.
Peaknaze C	53.47 N	1.91 W	Mineral	2	2	Unpubl.
Tadham	51.16 N	2.81 W	Organic	5	1	Lloyd (2006)
Ullapool	57.91 N	5.17 W	Organic	1	1	Unpubl.
Whim	55.76 N	3.27 W	Organic	1	1	Sheppard et al. (2004)

• In sampling method 1, the chamber was 38 cm in diameter and 23 cm high, sample volume was 20 mL, two samples were taken over 30–90 min, gas samples were stored in vials or tedlar bags, and analysis was by gas chromatograph (GC, 5890 series II; Hewlett Packard, Santa Clara, CA, USA). In sampling method 2, the chamber was 20 cm in diameter and 10 cm high, sample volume was 20 mL, three samples were taken over 30 min, gas samples were stored in vials, and analysis was by GC (Clarus 500; Perkin Elmer, Cambridge, UK). In sampling method 3, the chamber was 30 cm in diameter and 41 cm high, sample volume was 20 mL, 2–4 samples were taken over 0.5–2 h, gas samples were stored in vials or tedlar bags, and analysis was by GC (Clarus 500; Perkin Elmer). In sampling method 4, the chamber was 63 cm in diameter and 29 cm high, sample volume was 20 mL, four samples were taken over 30 min, gas samples were stored in vials or tedlar bags, and analysis was by GC (5890 series II; Hewlett Packard). In sampling method 5, the chamber was 30 cm in diameter and 35 cm high, sample volume was 20 mL, two samples were taken over 1–2 h, gas samples were stored in Exetainers, and analysis was by GC (Clarus 500; Perkin Elmer). In fitting method 1, the mean values of dC and dt were used to calculate dC/dt in Eqn 1; in fitting method 2, the slope of a linear regression between C and t was used.

Results

Site means for methane flux and the explanatory variables are shown in Table 2. CH₄ fluxes showed strongly left-skewed distributions at all sites, dominated by low, positive (i.e. emission) fluxes (Fig. 2a). Overall, the mean and median fluxes were 6.3 and 0.23 nmol m⁻² s⁻¹, with lower and upper quartiles at −0.1 and 4.5 nmol m⁻² s⁻¹. Sites with mineral soils tended to have small emissions or low rates of net uptake, with site means between −0.3 and 0.7 nmol m⁻² s⁻¹ (Fig. 2). Sites with organic soils had much larger emissions, with means between −0.3 and 27.4 nmol m⁻² s⁻¹. Sites with organo-mineral soils were intermediate, though the sample size was much smaller.

Table 2. Table of means for methane flux () and explanatory variables by site

Site			z water	θ	T soil	pH	z peat	C soil	ρ soil	A er	S pp1	n
Auchencorth	0.4	1.6	−13.1	0.91	9.0	4.4	0.5	33.0				595
Clocaenog	0.1			0.32	8.1	3.9	0.1	15.1				130
Cow Park	0.1			0.45	12.2	6.4	0.0	13.9				290
Easter Bush	0.1			0.30	10.2	5.0	0.0	12.1				704
Forsinard	15.9		−8.7	0.36	9.9	4.0	2.9	189.9		28.6	0.8	615
Glensaugh	2.2	2.5		0.91	7.3	4.5	0.7	0.6				18
Grenstein	0.7	1.2		0.60	3.0				0.22			69
Hafren	−1.4	2.9		0.06	8.6	4.5	0.6	0.5				29
House O'Muir	−0.3	3.5		0.73	12.0	6.0	0.0	5.0	0.27			80
Lincoln	−0.1	1.7		0.23			0.0	8.3				229
Loch More	27.4		−5.6	0.93	14.5		4.0	264.0				188
Migneint A	16.9	0.1	−10.1		8.0	4.3	2.0	130.9				251
Migneint B	0.0	0.4			8.4	4.4	0.2	14.8				261
Migneint C	12.7		−5.8		12.6	4.0	2.0	132.0				160
Moor House	11.3	2.1	−10.3	0.90	7.3	3.8	1.5	85.0		32.4	−0.4	208
Peaknaze A	0.0			0.49	9.9	4.1	0.1	5.0				75
Peaknaze B	5.5	0.3	−15.0		8.3	3.9	2.3	153.0				249
Peaknaze C	−0.3	0.4	−1.0		8.4	4.1	0.2	13.7				239
Tadham	−0.4		−32.7		11.1	6.4	2.0	132.0				197
Ullapool	7.2	2.0		0.77	8.1	4.5	0.7	0.4	0.14			15
Whim	22.2		−2.6	0.95	8.0	3.6	6.0	396.0	0.10	48.5	0.5	229

• The explanatory variables are: soil CO₂ efflux (, μmol m⁻² s⁻¹), water table depth (z_water, cm), soil moisture (θ, volumetric water content as a fraction), soil temperature (T_soil, °C); soil pH, peat depth (z_peat, m), soil carbon stock (C_soil, kg C m⁻²), soil bulk density (ρ_soil, g dry mass cm⁻³), the percentage cover of aerenchymatous species (A_er, %), and the first axis (S_pp1, dimensionless) of an ordination of species cover. n is the number of CH₄ flux samples.

Scatterplots of methane flux against the main explanatory variables are shown in Fig. 3 and Figs S1 and S2. Some pattern is apparent with most variables, and this is clearest when the data are averaged at site and sub-site level. The strongest relationships are with soil carbon, soil moisture and the measures of plant species composition. Where relationships appeared nonlinear, transformations were examined to find a more linear relationship with methane flux. However, only in the case of soil moisture did a transformation give a consistently better fit to the data across all averaging levels, where θ⁴ had the closest linear relationship with methane flux; this transformation was used in the statistical modelling.

Figure 4 and Table 3 show that the adjusted r², effectively the percentage of variance explained by the regression model, increases with the extent of averaging; the model fit is poorest with the raw data and best when the site means are analysed. Several variables show reasonably close relationships with methane flux, particularly soil carbon, peat depth, soil moisture and plant species composition, though the ranking varies between averaging levels. Table 3 shows that the size of the data set (without missing values) varies widely between variables, so an exact like-for-like comparison is not possible. Also, much of the variation in explanatory variables is confounded, so caution is needed in ascribing cause and effect. Some variables showed surprisingly poor relationships with methane flux when analysed in this simple way, such as water table, soil temperature, CO₂ efflux, pH and bulk density.

Table 3. Results of univariate linear regression, showing the variance explained by each explanatory variable, and the sample size available (n)

Variable	Adjusted r2				n
	All data	Plot	Sub-site	Site	All data	Plot	Sub-site	Site
C soil	26	48	58	72	4764	279	54	21
z peat	26	50	59	74	4764	279	54	21
S pp1	24	44	60	33	1016	85	17	3
θ	14	41	40	25	3000	254	46	21
θ 4	13	59	71	30	3000	254	46	21
A er	11	29	39	33	1016	85	17	3
pH	7	30	13	23	4171	235	45	21
Bulk density	5	13	14	0	292	28	16	5
z water	5	25	23	15	2517	130	24	10
T soil	4	3	9	0	3950	302	54	21
	0	4	4	1	2055	173	26	12

• The explanatory variables are: soil carbon stock (C_soil, kg C m⁻²), peat depth (z_peat, m), the first axis (S_pp1, dimensionless) of an ordination of species cover, soil moisture (θ, volumetric water content as a fraction), the percentage cover of aerenchymatous species (A_er, %), soil pH, soil bulk density (g dry mass cm⁻³), water table depth (z_water, cm), soil temperature (T_soil, °C), and soil CO₂ efflux (, μmol m⁻² s⁻¹).

All sub-sets of variables were used in a multivariate regression procedure to identify which combinations of variables were most useful in explaining the variation in methane flux. The best-fitting models with each number of terms are shown in Table 4, ranked by their Akaike Information Criterion (AIC; Burnham & Anderson, 2002), a measure which ranks competing models according to goodness-of-fit with a penalty associated with the number of parameters (here, equivalent to the number of explanatory variables used + 1). Other similar measures (Mallow's C_p and Bayesian Information Criterion) were examined but gave essentially the same rankings. Comparisons among models becomes difficult when there are missing values for some variables, so results are presented for each averaging level, with and without a restriction on the data sub-set to those with species composition data.

Where species composition data were available, the first PCA axis, S_pp1, came out as the best-fitting univariate model at all averaging levels. The percentage cover of aerenchymatous species, A_er, gave additional information, and was the second or third term to be added to the model. Using these two terms alone could account for up to 53% of the variance in methane flux, although n was always smaller when restricted to samples with species data. Without the restriction of including species data, peat depth was the best single predictor of methane flux, accounting for 74% of the variance at site level, and 33% in the raw data. According to the AIC, the parsimonious models included soil temperature, soil moisture and soil carbon, although the best combinations of these depended on the averaging level. When a unique identifier for each plot was included as an extra term in the model, the adjusted r² was generally increased by a further 10%. In other words, 10% of the variance was explained by something specific to each unique location, but not captured by the explanatory variables measured here.

Nonlinear regression was used to test some commonly used models, and the goodness-of-fit measures and model parameters are shown in Table 5; for brevity, results are only shown for two averaging levels, but the pattern is similar elsewhere. The results show that the exponential function of temperature alone has little explanatory power for these data. Water table also performs poorly for the raw data, and is less effective than soil moisture at sub-site level. When the exponential response to temperature is combined with a power function for soil moisture and a linear function for soil carbon, the explanatory power is much better, and is similar to or better than that from the best sub-sets linear modelling.

Discussion

Our analyses show that some of the expected relationships are present in these data, and up to ca. 74% of the variance in methane flux can be explained at site level. However, there is a considerable amount of noise in the data, and more than half the variance in the raw data is unexplained. This is similar to other studies which have applied statistical approaches, and the primary relationships with explanatory variables are similar. In peatlands in Canada and Finland, several studies on individual sites have found poor correlations between CH₄ flux and temperature, but reasonable relationships with water table depth (r² typically ca. 0.5 using seasonal means and log transformation, Liblik et al., 1997 and references therein). Bubier (1995a) found that r² was increased to 0.84 by adding mean temperature at mean water table depth, water chemistry (conductivity, pH, Ca), tree cover and herbaceous plant cover to the regression model. Macdonald et al., 1998 found similar relationships with temperature and water table in mesocosms from Northern Scotland under controlled conditions, and found a clear effect of plant-mediated transport via the stems of bogbean Menyanthes trifoliata L. Greenup et al. (2000) also found a strong relationship between below-ground biomass of an aerenchymatous sedge species and CH₄ emission in a UK bog. At sites in Germany, Couwenberg et al. (2011) report a close relationship between the number of leaves of two aerenchymatous species and CH₄ emission. Kettunen et al. (2000) found reasonable plot-specific relationships with temperature (0.33 < r² < 0.72) at a fen site in Finland; water table depth had less explanatory power, probably because its range of variation was small. The fitted model performed poorly when the data were pooled across plots, and did not explain short-term variations well. Laine et al. (2007) found very similar results in a blanket bog in Ireland. In addition, they measured the leaf area of aerenchymatous species, but found little relationship with CH₄ flux.

There are several likely causes for the unexplained variation in the raw data. First, there is typically a substantial amount of measurement error in static chamber flux measurements, which adds some random noise to the data (Levy et al., 2011), as well as possible systematic errors (Christiansen et al., 2011). Levy et al. (2011) estimate typical 95% confidence intervals of ±20% for individual flux measurements, using Monte Carlo error propagation. Second, CH₄ can be transported by ebullition (i.e. via rising bubbles of gas in soil water), rather than diffusion. This will add a random term to the measured flux, if transport of bubbles is not as a continuous stream of small bubbles, but episodic and largely stochastic, influenced by short-term wind-driven surface pressure fluctuations which are unpredictable (Baird et al., 2004). Third, the net efflux of methane is a complex interaction between several underlying processes, and several of the variables driving the underlying processes are either not measured, or poorly represented by the independent variables which are commonly measured. To give some examples, water table depth or soil moisture are used as surrogate variables for soil O₂ concentration, and the extent of the anaerobic zone. However, Askaer et al. (2010) show, using high-resolution measurements using planar O₂ optodes, that water table depth can be a poor proxy for anaerobic status at small spatio-temporal scales, as aerobic zones were observed below the water table, and vice versa, especially when the water table was dynamic. The correspondence becomes much better when the water table depth remains more stable, and at larger spatial scales. The rate of methane production is affected by soil redox potential and the availability of terminal electron acceptors (primarily Fe³⁺, Mn⁴⁺, SO₄²⁻, NO₃⁻). These are not commonly measured, so variation due to these will be unaccounted for; at best, they may correlate with soil pH or pollutant deposition as a proxy. The availability of carbon substrates and the methanogenic microbial biomass and diversity are generally assumed not to be limiting, but there is some evidence to the contrary (Segers & Kengen, 1998). Furthermore, there are two distinct metabolic pathways which can produce methane, (based on either acetate or hydrogen/CO₂ substrates, and associated with different bacterial taxa), and their response to environmental variables and relative abundance may differ (Cadillo-Quiroz et al., 2006; Lai, 2009). Soil carbon stock and peat depth will provide some proxy for the availability of carbon substrates, as well as the extent of anaerobic conditions, but this will not include short-term variations.

Model fit was much improved when the raw data were averaged in time and over increasingly larger spatial scales, and there were a number of probable reasons for this. If there is substantial random measurement error in the flux data, this will tend to cancel out as more measurements are averaged. Similarly, there will be some random measurement error in the independent variables which will again tend to cancel out with averaging. Furthermore, the co-location of measurements of independent and dependent variables is not exact, and some independent variables were only measured at site or sub-site level, so some mis-match at plot level is inevitable. Several of the independent variables vary only on relatively long time scales (e.g. soil carbon, species composition, pH), so cannot account for short-term temporal variation. Only a small number of independent variables vary on short time scales (temperature, soil moisture, CO₂ efflux, etc.), and these do not fully account for the variations in methane flux. Averaging removes this unexplained short-term temporal variation.

The results show that plant species composition is the best single predictor of mean CH₄ flux, where data are available. This may be because of direct effects of particular plant species (because some species transport CH₄ or provide suitable substrates for the production of CH₄), as well as indirect effects (because plant species composition is a good indicator of environmental conditions, and effectively integrates past conditions over the long term). For example, Wamelink et al. (2005) showed that plant species composition was a very good predictor of soil pH in a large European data set. Bryophytes are particularly sensitive to water table position, and their use as predictors of CH₄ flux has been noted previously (Bubier, 1995b; Dias et al., 2010). The results suggest that there is an effect of plant species composition beyond that of transport via aerenchyma, as the multivariate PCA axis has substantially better explanatory power, and it seems likely that both direct and indirect effects are important. Using plant species composition may be useful in predicting CH₄ fluxes at large scales, where vegetation survey data are available. Further validation of this approach is being investigated, and this could form part of a method for obtaining national-scale emissions of CH₄ from soils. Plant species composition is hard to include in ecosystem or biogeochemical models, and at best is approximated by some plant functional type classification (e.g. Couwenberg et al., 2011). However, these classifications are relatively crude, and our mechanistic understanding is still incomplete (De Deyn et al., 2008; Kip et al., 2010).

Of the abiotic variables, soil carbon and peat depth have the most explanatory power. Again, a component of this will be a direct effect, as deep peats have a larger pool of carbon substrate for methanogenesis, but indirect effects probably predominate. Deep peats exist because the waterlogged conditions which inhibit decomposition also provide the extensive anaerobic environment which permits methanogenesis. Both soil carbon and peat depth were included in the analyses, even though the two are so closely linked, as the former was generally calculated from the latter at peatland sites. However, their main difference lies in how they represent the transition between mineral and organic soils: soil carbon gives a continuous transition, whereas peat depth (having a zero value in mineral soils) shows a ‘broken stick’ response, with a sharp delineation between mineral and organic soils. The latter commonly fits the CH₄ flux data slightly better. Interestingly, there is no apparent influence of decomposition rate, measured as soil CO₂ efflux, on CH₄ flux. It might be expected that this would provide a measure of the size of the labile carbon pool and substrate availability, and this effect is included in some models (Clark et al., 2011). It may simply be that the measurements here are not sufficient to detect this effect, and more co-located CO₂ and CH₄ flux measurements would be required to quantify this rigorously.

Whilst there is strong evidence that pH affects soil bacterial communities in wetland soils (Hartman et al., 2008), the net effect on CH₄ fluxes is less clear. An increase in CH₄ fluxes with declining pH would be expected from most studies on methanogens under controlled conditions; for example, in a study by Taconi et al. (2007) methane production from waste water was increased by 30% when pH was decreased from 7.0 to 4.5. On the other hand, reduced pH in UK semi-natural soils is often associated with (current or historically) elevated inputs of atmospheric sulphate, which has a suppressive effect on methanogenesis (e.g. Gauci et al., 2004), potentially countering a direct pH effect. However, methanotrophs may have a different responses to pH which could complicate the net effect; both methanogens and methanotrophs are likely to have bell-shaped response, but with potentially different optima (Dedysh, 1998). Indeed, the role of methanotrophs in bog systems may have been substantially under-estimated (Kip et al., 2010). Our results suggest a weak but detectable effect of pH, with CH₄ fluxes tending to increase with acidity. Partially, this is attributable to the negative correlation between pH and soil carbon, and the regression when this effect is removed is not statistically significant (P = 0.2). A similar effect is present in the relationship between soil bulk density and CH₄ flux, which is apparent in Fig. 3. Soil carbon stock is calculated using the bulk density values, and the two closely co-vary. When the correlation with soil carbon is removed, bulk density provides little extra predictive information.

Throughout the analysis we have focussed on the explanatory power of the models, rather than predictive power, as the latter is not tested here. Given the complexity of possible feedbacks and interactions between the driving variables and the issues of extrapolating to the long term (Norby & Luo, 2004; Boardman et al., 2011), prediction of response to long-term global change is much more difficult. However, in the shorter term, there is considerable current interest in the impact of peatland restoration activities on greenhouse gas emissions (Waddington & Price, 2000; Laine et al., 2006; Rochefort & Lode, 2006; Eggleston et al., 2011). In the United Kingdom, there is some debate over the merits of restoring water tables by blocking moorland drains (Wilson et al., 2010), which may promote Sphagnum growth and re-initiate peat accumulation (and presumably, net CO₂ uptake), at a potential cost of increased methane emissions (Baird et al., 2009b). There are few pertinent data to quantify these effects (Basiliko et al., 2007; Bussell et al., 2010; Fenner et al., 2011; Urbanova et al., 2011), but the relationships in Fig. 3 could be used as a crude summary of the effect of water table on methane flux, if we assume that the long-term effect of water table manipulation is the same as the relationship observed in natural variations. This indicates an effect of +0.8 (±0.28 SE) nmol CH₄ m⁻² s⁻¹ (or +0.4 g CH₄ m⁻² yr⁻¹) per cm increase in water table. To offset this would require an increase in CO₂ sequestration of 7.3 nmol CO₂ m⁻² s⁻¹ (or 10.1 g CO₂ m⁻² yr⁻¹) per cm increase in water table (assuming that a gramme of CH₄ has the global warming potential of 25 grammes of CO₂ over a 100-year time span). By comparison, the review of studies in the Netherlands, Germany, Sweden and France by Couwenberg et al. (2011) obtained a value of +1.7 g CH₄ m⁻² yr⁻¹ per cm increase in water table, around four times larger than our value. However, their value applied only to sites with aerenchymatous vegetation and mean annual water tables <20 cm from the surface. Including all their data points would give a value more similar to ours, and whether the differences reflect real geographic variation, or merely sampling error in a variable parameter, is hard to discern. We note that there are several reasons why the effects of manipulated and natural variations in water table might be different, certainly in the short term, as the dynamics of the system are complex. Also, as Frolking et al. (2006) point out, the relative impacts of CO₂ sequestration and CH₄ emissions on radiative forcing depend largely on the time scale considered. As the time scale considered increases, the relative importance of CH₄ emissions decreases, and the commonly used time horizon of 100 years is essentially arbitrary.

The analysis of this data set has demonstrated the potential and limitations of commonly measured environmental variables as predictors of soil CH₄ fluxes. The observed variability in instantaneous fluxes will remain difficult to explain or predict whilst the explanatory variables are weak surrogates for the underlying driving variables. However, longer term and larger-scale averages are much better predicted by environmental variables, and notably by soil carbon stocks and measures of plant species composition. Given that these variables are often available in national vegetation and soil surveys, these could have considerable utility in estimating national-scale CH₄ emissions from soils. Detailed vegetation characteristics are generally not represented in land surface models, but given an accurate simulation or prescription of soil carbon stock (or peat depth) and hydrology, these have suitable explanatory power for predicting CH₄ fluxes.