The value of soil respiration measurements for interpreting and modeling terrestrial carbon cycling

Abstract

Background

An acceleration of model-data synthesis activities has leveraged many terrestrial carbon datasets, but utilization of soil respiration (R_S) data has not kept pace.

Scope

We identify three major challenges in interpreting R_S data, and opportunities to utilize it more extensively and creatively: (1) When R_S is compared to ecosystem respiration (R_ECO) measured from EC towers, it is not uncommon to find R_S > R_ECO. We argue this is most likely due to difficulties in calculating R_ECO, which provides an opportunity to utilize R_S for EC quality control. (2) R_S integrates belowground heterotrophic and autotrophic activity, but many models include only an explicit heterotrophic output. Opportunities exist to use the total R_S flux for data assimilation and model benchmarking methods rather than less-certain partitioned fluxes. (3) R_S is generally measured at a very different resolution than that needed for comparison to EC or ecosystem- to global-scale models. Downscaling EC fluxes to match the scale of R_S, and improvement of R_S upscaling techniques will improve resolution challenges.

Conclusions

R_S data can bring a range of benefits to model development, particularly with larger databases and improved data sharing protocols to make R_S data more robust and broadly available to the research community.

Introduction

Measures of soil respiration (R_S, the flux of CO₂ between the soil surface and atmosphere) extending over the last 60+ years constitute a geographically and ecologically rich dataset that has improved our understanding of soil carbon (C) cycling. For example, meta-analyses have used these data to draw inferences about the correlation between R_S and its biotic and abiotic drivers (Raich and Schlesinger 1992); the autotrophic and heterotrophic sources of R_S (Bond-Lamberty et al. 2004; Subke et al. 2006); the importance of CO₂ production and biophysical lags to R_S patterns (Vargas et al. 2010a); the response of the global R_S flux to climate change (Carey et al. In review; Bond-Lamberty and Thomson 2010; Wang et al. 2014; Zhou et al. 2016); the temperature sensitivity of soil heterotrophic respiration (Zhou et al. 2009; Wang et al. 2014); and the response of R_S flux to experimental changes in precipitation patterns (Thomey et al. 2011; Vicca et al. 2014). For specific sites, R_S has also been used to calculate biometric-based estimates of carbon exchange in comparison with eddy covariance data (Curtis et al. 2002; Gough et al. 2008; Goulden et al. 2011; Giasson et al. 2013), providing a powerful constraint on carbon-cycle measurements and modeling.

In spite of this progress, the terrestrial carbon cycle remains a remarkably uncertain component of Earth system models (ESMs), an uncertainty that has been primarily attributed to the soil component (Friedlingstein et al. 2006; Todd-Brown et al. 2013; Anav et al. 2013; Friedlingstein et al. 2014; Hoffman et al. 2014). Many fundamental questions about climate change depend on the sensitivity of R_S to global change factors, including whether soils will gain or lose C over the next century (Cox et al. 2013; Todd-Brown et al. 2014), how the residence time of terrestrial C will change (Schuur et al. 2009; Giardina et al. 2014; Carvalhais et al. 2014), and whether mitigation actions can sequester soil C in meaningful quantities, and over meaningful timeframes for human societies (Paustian et al. 2016). Although researchers recognize the central importance of soils to C-climate feedbacks (Ciais et al. 2013), and are aware of the thousands of extant data on seasonal and annual fluxes (Bond-Lamberty and Thomson 2010), R_S data are rarely used for these models’ parameterization or assessment (Shao et al., 2013).

There are several reasons why R_S data are infrequently used. First, many recent efforts to improve the representation of soil C turnover in ESMs have focused on simulations at decadal to millennial timescales, and therefore on improving model fidelity of variables that integrate long time periods, such as soil C stocks and ¹⁴C ages (Todd-Brown et al. 2012; Koven et al. 2013; Wieder et al. 2013; Todd-Brown et al. 2014). In contrast, when daily to annual C fluxes are simulated, there is often a bias towards utilizing eddy covariance (EC) flux data to the exclusion of other data streams partially because of the readily accessible EC data (Williams et al. 2009; Richardson et al. 2010; Kuppel et al. 2012; Keenan et al. 2013; Collalti et al. 2016). Second, while R_S seasonal and annual sums have been compiled (Bond-Lamberty and Thomson 2010; Hashimoto 2012), instantaneous R_S measures from autochambers and survey campaigns have rarely been synthesized (Bahn et al. 2010; Vargas et al. 2010a; Lavoie et al. 2014; Cueva et al. 2015), and are not easily available to modelers because they are not in organized data repositories (see Vargas and Leon 2015 for an exception). These small but geographically widespread datasets are generally in the hands of individual investigators, and are part of what Dietze et al. (2013) refers to as the ‘long tail’ of data. Third, many scientists are not sure how to use R_S data effectively. The relationship between R_S and soil C stocks and turnover is not always conceptually clear, particularly given uncertainties in partitioning R_S sources. R_S also does not match one-to-one with soil heterotrophic respiration (R_H), which is the variable generally desired for modeling soil C turnover, but rather integrates R_H and belowground autotrophic respiration (R_A). The fact that R_S integrates multiple processes—root and microbial activity, along a vertical gradient of soil activity—can make R_S a challenging flux to interpret and represent in mathematical models (Vargas et al. 2011b).

Given the rapid development of new soil C cycle representations in ESMs (Luo et al., 2016), these R_S data represent an underutilized resource, as R_S is an information-rich data stream that occupies an important mid-level spatial scale bridging soil pore-scale processes to ecosystem-scale fluxes (Fig. 1). Other ecosystem-scale measures, such as net primary production (NPP) and net ecosystem production (NEP), similarly integrate multiple processes (e.g., respiration and photosynthesis, from both subcanopy and canopy layers), but nonetheless have been used effectively for model validation and improvement (Hudiburg et al. 2009; Williams et al. 2009; Schwalm et al. 2010; Keenan et al. 2012).

The goal of this review is to highlight the opportunities for utilizing R_s data, justify continued collection of these data, and inspire novel synthesis and modeling activities. We highlight three challenges for utilizing and interpreting R_S, and identify potential solutions. We then discuss how R_S data are useful for model development, and outline approaches for using R_S for data-model fusion (i.e. selecting models and parameters that are ecologically realistic and consistent with a range of data streams) and for model benchmarking (ranking models by how well they match observations). We also discuss quality control and database activities that will make R_S data more robust, useful, and broadly available to the scientific research community.

Challenge 1: reconciling soil and tower fluxes

The R_S-R_ECO mismatch

At many EC tower sites, particularly in forests, studies have reported periods when R_S is consistently higher than ecosystem respiration (R_ECO) (Phillips et al. 2010; Thomas et al. 2013; Giasson et al. 2013; Speckman et al. 2015). This is biophysically impossible: R_S is a major component of whole-ecosystem respiration in forests, but it cannot be higher than R_ECO, which also includes stem, leaf, and other aboveground autotrophic respiration (Bolstad et al., 2004; Gough et al. 2008; Ohkubo et al., 2007; Tang et al. 2008). This irregularity indicates either inconsistent measurement footprints, or persistent measurement biases in R_ECO, R_S, or both (Fig. 2).

This challenge was first identified nearly 20 years ago, with R_S measurements reported to be higher than R_ECO in boreal forests (Lavigne et al. 1997). While R_S has measurement errors and scaling issues (see Table 1 and Challenge 3), systematic underestimation of R_ECO is now believed to affect practically all EC sites (Schimel et al. 2008; Aubinet et al. 2012). For instance, Luyssaert et al. (2007) conducted a global synthesis of forest C flux and pools, and showed that non-closure of ecosystem C budgets was the rule rather than the exception. As much as 60 % of the CO₂ thought to be taken up by forests based on EC-based GPP was unallocated to plant biomass. Indirect evidence also comes from EC energy balance studies, which have shown that more energy enters EC sites than can be accounted for by latent and sensible heat losses, suggesting that some CO₂ emissions might also be missed (Falge et al. 2001; e.g. Barr et al. 2006; Papale et al. 2006; Franssen et al. 2010; Foken et al. 2011; Stoy et al. 2013).

One common explanation is the influence of advection on net ecosystem exchange (NEE) measurements using EC. This implies that we should have more confidence in high R_ECO values, while lower values should be rejected as they have higher probability of being affected by advection or other systematic errors (Van Gorsel et al. 2007). Several studies in forest ecosystems have attributed R_ECO-R_S mismatch to EC measurements that are affected by topography, which influences advection and thus calculation of NEE (Kutsch et al. 2008; Phillips et al. 2010). In tall plant canopies, nocturnal fluxes can also be difficult to measure because of poorly constrained storage of CO₂ within canopies, assumptions for low turbulence filtering (i.e., u* threshold), and decoupling of eddy covariance systems from subcanopy processes (Goulden et al. 1996; Aubinet and Feigenwinter 2010; Thomas et al. 2013; e.g. Alekseychik et al. 2013).

Mismatch may also occur when R_ECO is much larger than R_S, for example as reported by Giasson et al. (2013) during winter at Harvard Forest. In this case, the large difference between R_ECO and R_S indicated high rates of aboveground respiration, approximately twice as large as soil respiration. This is improbable, given that deciduous trees dominate the site. Again, the difficulty in reconciling EC and soil measures may result from poor vertical air mixing. For quality assurance, time periods with low wind or turbulent conditions are generally filtered from NEE time series, but if high winds ventilate the snowpack where CO₂ from soil respiration has been accumulating, then such windy periods could produce over-estimates of R_ECO.

The tendency for nocturnal NEE (and hence R_ECO based on common approaches) to be underestimated results in artificially high estimates of gap-filled annual net ecosystem production (NEP), and can also cause ecosystem responses to global change factors to be missed. In a dramatic example, Speckman et al. (2015) compared chamber and tower-based estimates of R_ECO over seven growing seasons during a bark beetle infestation that reduced live wood biomass by 85 % at the GLEES Ameriflux site in Wyoming (Fig. 3). Over that period, EC-based estimates of R_ECO did not change, while chamber-based estimates of R_ECO declined by 35 %. Importantly, the mismatch between tower and chamber-based measurements was not dependent on turbulence, and did not diminish when more restrictive turbulence filtering criteria (u*) were used. However, the mismatch did diminish over time with increasing tree mortality, leading the authors to conclude that the loss of canopy improved coupling of flows between the subcanopy and within-canopy air space.

In addition to measurement limitations, there are algorithmic limitations to estimating R_ECO with EC. Importantly, R_ECO and GPP are not directly measured–they are inferred as the result of partitioning NEE. Most partitioning approaches rely on the assumption that NEE at night, when photosynthetically active radiation is low or zero, does not have GPP, and that the drivers of R_ECO (primarily temperature) and GPP (primarily light) differ. Most partitioning methods also assume that nighttime R_ECO is a good proxy for daytime R_ECO, or vice versa. However, linkages between photosynthesis rate and soil autotrophic and heterotrophic activity (Tang et al. 2005; Carbone et al. 2007; Bahn et al. 2009; Vargas et al. 2012), diurnal patterns of soil moisture available to soil heterotrophs (Baker et al. 2008), and diurnally shifting tower footprints (Xu et al. 2017), all lead to potential mismatch between R_S and R_ECO. Furthermore, recent methods for estimating daytime respiration using light-response relationships (Laslop et al. 2010) or stable-isotope approaches (Wehr et al. 2016), have shown that R_ECO is lower during day than at night, due to photoinhibition of leaf respiration. By providing daytime measurements of R_ECO, these methods also allow partitioning of GPP and R_ECO in situations with low nocturnal turbulence, or with advection, will hopefully yield better fit between R_S and R_ECO in the future.

The preceding examples demonstrate that EC, while a powerful approach, has particular limitations and biases. Franks et al. (1997) cautioned against calibrating models exclusively with EC flux data, a point reiterated by Medlyn et al. (2005) and Keenan et al. (2013). Nevertheless, there is a heavy reliance on EC estimates of GPP and R_ECO for model development and benchmarking, a practice that amounts to fitting models to models.

Opportunities to close the R_S-R_ECO gap

Given that simultaneous chamber-based R_S measurements are already collected at many EC sites, it would be relatively easy to use comparisons of R_ECO and R_S to flag potentially low-quality EC data (Tang et al. 2008; Phillips et al. 2010; Giasson et al. 2013). Based on Ameriflux sites contributing R_S data the Ameriflux database or to the Soil Respiration Database (Bond-Lamberty and Thomson 2010), we estimate R_S data are available for at least 37 sites.

While continuous measurements are ideal, even intermittent data provide valuable constraints on R_ECO. First, R_S data constitutes an independent measurement, which is important for methodology validation. Second, R_S measures are not subject to systematic biases at night due to low turbulence conditions (although R_S chamber bias can occur under high turbulence conditions, e.g. Takle et al. (2004)). Third, R_S is directly measured rather than inferred. Fourth, the accuracy of soil chamber measurements is technically straightforward to measure using a sand column with a known flux rate, a type of calibration check that is not possible for EC (Martin et al. 2004; Pumpanen et al. 2004; Risk et al. 2011). Laboratory calibration has the potential to make R_S data, at least in theory, more accurate than R_ECO. For instance, Pumpanen et al. (2004) compared twenty different R_S chambers on a sand column through which a known CO₂ flux rate was established. They found substantial variation based on chamber design, ranging approximately ±35 % from the true flux rate, indicating that there is unrealized potential to reduce R_S uncertainty by using sand columns more widely for routine soil chamber calibration. Such systems can also be used to test chamber responses to wind and changes in soil moisture under laboratory conditions. Sand column calibration cannot account all sources of error and bias, such as soil collar damage roots and hyphae (Heinemeyer et al. 2011), but provides considerable constraint on instrumentation uncertainty (Table 1).

Table 1 Sources of uncertainty for chamber-based R_S and EC-based R_ECO measurements

We suggest using R_S-R_ECO comparison as an alternative evaluation of EC data quality, in addition to energy balance closure and more traditional metrics such as u*. (One of the largest uncertainties associated with this, upscaling chamber measurements to the EC tower footprint, is discussed below in Challenge 3.) There are several examples of utilizing R_S to detect low-quality nocturnal EC fluxes, and to construct bottom-up, chamber-based estimates (combining soil, stem and foliage respiration) to gap-fill R_ECO under periods of flow decoupling and advection (Van Gorsel et al. 2009; Zeeman et al. 2012; Thomas et al. 2013; Alekseychik et al. 2013; Speckman et al. 2015). Alternatively, Thomas et al. (2013) were able to salvage a large portion of nocturnal EC measures at a site prone to advection by summing above-canopy fluxes with fluxes inferred from a sub-canopy system and soil chambers.

Ultimately, uncertainty analyses are likely to reveal sites and periods where EC R_ECO fluxes are difficult to reconcile with other techniques. Therefore, it may be necessary to move away from the customary approach of gap-filling and partitioning EC flux data, toward more use of the best observations—EC data from optimal conditions, or chamber-based data—for model evaluations.

Challenge 2: beyond autotrophic and heterotrophic partitioning

Terrestrial ecosystem models rarely simulate R_S explicitly, but instead simulate components of R_H and R_A. In some cases, aboveground and belowground components of R_A may be pooled into a single flux (e.g. CLM: Oleson et al. 2010), roots may be lumped with aboveground biomass, as is the case in some simple models (e.g., DALEC: Williams et al. 2005 or SIPNET: Braswell et al. 2005), or respiration flux components may be output in ways difficult to combine into an estimate of modeled R_S that is truly comparable to observed R_S. For this reason, and because R_H is used in conjunction with NPP to estimate annual terrestrial C storage (NEP, see Fig. 1), total R_S is often not as useful to modelers as its partitioned components. But partitioning R_S is non-trivial, requiring experimental manipulations that are prone to artifacts (Hanson et al. 2000; Kuzyakov 2006; Subke et al. 2006). Most partitioning experiments, whether they utilize isotopes or bulk fluxes, require physically separating roots from native soil, and doing so alters moisture, temperature, oxygen levels, and C source-sink dynamics from the undisturbed system, all of which likely influence R_A and R_H estimates (Nickerson and Risk 2009; Phillips et al. 2013; Snell et al. 2014). Furthermore, there is ample evidence that microbial ‘priming’, i.e. enhancement of microbial decomposition due to root growth, is likely the rule rather than the exception, making R_H and R_A conjoined fluxes that can only be artificially disentangled (Zhu and Cheng 2011; Dijkstra et al. 2013; Qiao et al. 2014; Finzi et al. 2015).

Fig. 1

Schematic of ecosystem C budget components. Gross primary production (GPP) is allocated to above and belowground plant growth (net primary production, NPP) and respiration (R_A). Total soil respiration (R_S) is the sum of belowground R_A and heterotrophic respiration (R_H). Total ecosystem respiration (R_ECO) includes R_S plus aboveground R_A. Net ecosystem exchange (NEE) is the CO₂ flux measured by the eddy covariance method, and generally reported on a 30 min basis. Cumulative annual NEE is called net ecosystem production (NEP) and represents the net increase or decrease in ecosystem carbon storage. NEP can also be estimated from biomass-based measures of plant growth and chamber-based measures of R_H

Fig. 2

Comparison of R_ECO (Level 4 FLUXNET estimates) and R_S at four forested eddy covariance sites. R_S > R_ECO for significant periods at Harvard Forest, Wind River, and Willow Creek. By contrast, at Mary’s River site a subcanopy flux system was used to identify and gap-fill nights with decoupling, improving R_ECO estimates (Thomas et al. 2013)

Fig. 3

Comparison of chamber and tower based estimates of R_ECO at the GLEES Ameriflux site during a massive bark beetle die-off. a Comparison of R_ECO measured by EC (R_EC), and by summed chamber measures of stems (R_W), foliage (R_F), and soil (R_S). Chamber and tower mismatch correlated with canopy leaf area index (LAI). Reprinted with permission from Speckman et al. (2015)

Because of these obstacles, are there alternatives to traditional partitioning approaches that can be used as model constraints? One well-established approach for using aggregate, unpartitioned R_S to constrain component processes is the Total Belowground Carbon Flux (TBCF) approach, originally proposed by Raich and Nadelhoffer (1989). This method indirectly estimates plant belowground C allocation (R_A + belowground NPP, see Fig. 1) using measurements of R_S and litter fluxes. The TBCF approach makes several major assumptions to justify indirect estimation of plant allocation, namely that most C leaves the soil through respiration (i.e. dissolved C losses and erosion are negligible), and that net changes in standing litter and SOM are negligible. The TBCF approach has been criticized for poor correlation with direct observations in managed forest systems (Gower et al. 1996), but nevertheless has generated important insights into plant allocation when applied at a global-scale (Litton et al. 2007; Litton and Giardina 2008). In addition to employing TCBF,we argue there are more opportunities to use measured aggregate R_S to validate the sum of the separately modeled R_H and R_A. Furthermore, while partitioning studies will continue to be important, there are also opportunities to improve on them to obtain much more detail, for instance by applying techniques that allow more than two sources of R_S to be resolved.

Partitioning R_S – what has been learned?

Researchers have attempted to partition R_S into discrete sub-fluxes for decades (Coleman 1973), with a few studies almost a hundred years old (Lundegårdh 1927; Turpin 1920). Attributing R_S to its source components furthers our mechanistic understanding of the many processes underlying and controlling this large C flux (Luo and Zhou 2006). In addition, after ecosystem-scale theories of C and energy dynamics were developed (Kira and Shidei 1967; Odum 1969), isolating the heterotrophic component of R_S allowed, in conjunction with a measure of net primary production, computation of ecosystem carbon gain or loss (NEP = NPP – R_H, see Fig. 1). Finally, once it was realized that R_S components might respond differently to environmental drivers and climate change (Boone et al. 1998), partitioning experiments were critical to draw generalized inferences in these areas (Wang et al. 2014).

A number of meta-analyses and syntheses have attempted to distill what has been learned from decades of partitioning experiments. Hanson et al. (2000) gave a qualitative assessment of partitioning methods and found that published studies exhibited wide variability, but on average about half (46 %) of the annual R_S flux was root-derived. Using significantly larger data sets, Bond-Lamberty et al. (2004) and Subke et al. (2006) found that as R_S increases, autotrophic contributions increase more than heterotrophic contributions; this pattern is generally consistent across biomes and different ecosystem ages; and there are no significant differences between partitioning methods (Kuzyakov 2006). Finally, two recent meta-analyses have explored how climate warming (Wang et al. 2014) and the interactive effect of multiple global change factors (Zhou et al. 2016) affect R_S and its components. Importantly, these studies have found differences in response timing and magnitude, with R_H exhibiting a sustained response to warming versus a non-significant R_A response. They have also shown that complex interactive effects between factors such as warming, elevated CO₂, and nitrogen addition can even reverse the direction of response (Zhou et al. 2016).

An interesting question is how and when global relationships for partitioning R_S can be used, particularly as more than a decade has passed since the original global analyses of R_H, and almost an order of magnitude more studies have been published partitioning R_S. To examine this question, we took data from a global soil respiration database (Bond-Lamberty and Thomson 2010), computed R_H via two well-known published equations (Bond-Lamberty et al. 2004; Subke et al. 2006), and computed the resulting error relative to the known, measured R_H flux in the database (Fig. 4). From 386 studies, only 121 reported R_H fluxes within ±10 % of the equation-derived value, while large errors occurred in ecosystems with anomalously-low R_H values (but note that these high percentage errors represent very low absolute numbers); no relationship with biome, ecosystem type, or time since disturbance was evident. This emphasizes the caution researchers must use when applying such generalized equations to singular locations, in particular in areas with high nitrogen deposition (Janssens et al., 2010) or suspected imbalances (from e.g., recent disturbance) in the soil carbon cycle (Harmon et al. 2011). Because direct measurements of R_A and R_H are have not been collected for many C-budget study sites (Luyssaert et al. 2007; Litton and Giardina 2008), there may be temptation to downscale global equations where no better measurements are available. However, C-budget relationships derived from large aggregates of data can add considerable error when applied at local scales (Kicklighter et al. 1994; Gower et al. 1996).

Fig. 4

Error arising from application of global relationships to partition sources of soil respiration. Data from a global soil respiration database, filtered to studies reporting both R_S and R_H from unmanipulated ecosystems. R_H was computed from R_S following Bond-Lamberty et al. (2004) and compared to reported R_H; dashed lines show ±10 % error. Data based on Subke et al. (2006) are almost identical and not shown. These global relationships predict that the R_H fraction of R_S decreases as R_S increases, but do not allow for sites with high R_S and very low R_H (at upper left; the graph cuts off a few very high-error sites)

Opportunity: advancing partitioning to consider more than two sources

Partitioning experiments have generated important insights, but designing experiments around a two-source model of R_S is both a conceptual and physical limitation. There are many cases in which it is useful to partition R_S into three or more sources, for instance to distinguish decomposition of recent soil C additions from that of native soil C and R_A (Whitman and Lehmann 2015), or to separate out the contribution of mycorrhizae from roots and free-living microbes (Heinemeyer et al. 2007; Heinemeyer et al. 2011; Phillips et al. 2012; Barba et al. 2016). As climate and land-use change disrupt equilibrium soil conditions, it becomes more important to partition microbial decomposition of older, stored C in addition to labile C (Wutzler and Reichstein 2007). By applying multiple partitioning approaches simultaneously–for instance physically separating roots as well as monitoring the isotopic composition of each component (Czimczik et al. 2006; Hopkins et al. 2013; Phillips et al. 2013), using different-sized micropore mesh treatements (Moyano et al. 2007; Heinemeyer et al. 2007), using two gas tracers simultaneously (Risk et al. 2013b), or combining single tracer experiments in which different C sources are isotopically labelled in each experiment (Whitman and Lehmann 2015)–it becomes possible to characterize more than two components of R_S.

Until recently, many researchers simply equated R_A with being from ‘recent’ C sources, and R_H with being from ‘older’ C sources, as depicted in Fig. 5a (Högberg et al. 2001; Hahn et al. 2006; Ogle and Pendall 2015). However, a number of studies have demonstrated that the C age of R_A and R_H varies at seasonal to interannual timescales as depicted in Fig. 5b. For instance, compiling root-respired ¹⁴CO₂ data from three temperate forests, Hopkins et al. (2013) showed a tendency for R_A to include substantial C from previous years early in the growing season, transitioning to current-year C by the end of summer (Fig. 6). Similarly, Lynch et al. (2013) showed that after the cessation of a free air CO₂ enrichment experiment, carbohydrates stored by trees during the experiment continued to contribute to root respiration for at least two years after the isotopically distinct CO₂ source was gone. Czimczik et al. (2006) found that as boreal forest stands matured following wildfire, root respiration tended to contain more stored C. Examining soil-respired ¹⁴CO₂ in a deciduous forest in Northern Wisconsin, Phillips et al. (2013) found evidence for gradual microbial transition from previous-year to current-year substrates over the course of a growing season. Fresh root exudates appeared to be a preferred microbial substrate. In an arctic tundra ecosystem, Hartley et al. (2012) also found evidence for seasonal microbial switching, but found that mid-summer plant activity stimulated microbial decomposition of older soil organic matter (i.e. positive priming).

Fig. 5

Combining ¹⁴C monitoring or ¹³C pulse-chase labeling with physical partitioning provides observations that are more useful for testing model C pool structures. a An isotope mixing model for partitioning R_S, which assumes new C is from R_A and old C is from R_H. b By using physical partitioning methods such as root exclusion while also monitoring the isotopic composition of each component, dynamic ages have been observed at seasonal to interannual timescales. In this example, both R_H and R_A include more ‘new’ C as the growing season progresses. c Age determinations of R_A and R_H can be used to test models simulating C as discreet pools with characteristic turnover times. In this example, five soil C pools and two plant C pools are represented, and turnover of both very old and new C increased during summer

Fig. 6

Growing season patterns for the difference in Δ¹⁴C between root respiration and the atmosphere (ΔΔ¹⁴C), which is used to infer C age, for three temperate Ameriflux forest sites. Howland forest (solid triangles), Harvard Forest deciduous site (open squares), and Harvard forest evergreen site (open diamonds). Error bars are the larger of the standard deviation of replicates of the propagated error of ΔΔ¹⁴C calculations. Reprinted with permission from Hopkins et al. (2013)

Collectively, these studies suggest that R_S should be represented by more than two sources, with both R_A and R_H containing varying levels of fresh photosynthetic products and stored C pools throughout the year (Hopkins et al. 2013; Phillips et al. 2013), and over timescales of ecological succession (Czimczik et al. 2006). Using conventional assumptions for partitioning R_S with ¹⁴C (R_H = ‘old’ and R_A = ‘recent’, Fig. 5a) could lead to overestimates of R_H early in the growing season when root respiration utilizes older stored C, and underestimates of R_H late in the growing season, if microbes have switched to recent C substrates.

Determination of R_A and R_H ages by combining physical partitioning and isotopic monitoring also provides observations to test models with more than two discreet C pools (e.g. Parton et al. 2010; Sierra et al. 2012). For instance, Fig. 5c represents hypothetical contributions of five discreet soil carbon pools ranging in turnover time from less than one year to millennial timescales, and two plant C pools with annual and decadal turnover times, that could give rise to the C ages shown in Fig. 5b. Such large ranges in ages of forest soil and root C have been established by measurements of plant, litter, and soil Δ¹⁴C (Gaudinski et al. 2000; Trumbore 2000), and have been used to develop soil C models with excellent predictive power. For instance, Sierra et al. (2012) showed that C turnover times inferred from plant, litter, and soil Δ¹⁴C measurements in 1996 were able to accurately predict gradual changes in Δ¹⁴C measured in R_S over the following decade.

Figure 5 also demonstrates why the relationship between R_S and soil C stocks can be unclear to many, and why modelers have found it difficult to use bulk (unpartitioned) R_S data to test soil C pool models. Currently, the most abundant and widely-used datasets used for soil C model testing do not include in situ R_S, but instead are from litter bag decay experiments (Bonan et al. 2013), laboratory incubations, soil C stock sizes, and ¹⁴C ages of solid soil C (Jenkinson et al. 2008; Koven et al. 2013). Laboratory incubations and litter bag experiments have demonstrated that a large fraction of soil C turns over slowly, and solid soil ¹⁴C data indicates the timescale of the slowest C pools can be on timescales of century to millennia (Trumbore 2009). But R_S is dominated by relatively young C that turns over in a year or less (Gaudinski et al. 2000), and that may contribute very little to the soil organic C pool. Collectively this suggests R_S can vary substantially from year to year with little impact on soil carbon aggradation or decay (Sierra et al. 2012). It is impossible to know from bulk R_S data, or even from estimated R_H, when a change in the contributions of long-lived ‘old’ soil C pools has taken place, unless age estimates of the fluxes can also be determined.

However, isotopic age estimates of R_H can allow changes in SOC turnover to be detected years to decades before they would be measurable from sampling solid SOC (Gaudinski et al. 2000; Hartley et al. 2012). For instance, Sierra et al. (2012) provided compelling ¹⁴C evidence that global change factors, including warming and N-deposition, resulted in a transfer of C from long-lived to rapid-turnover pools in soil experiments at Harvard Forest. Moving forward, it would be advantageous for more studies to combine ¹⁴C or pulse-chase ¹³C monitoring with physical partitioning approaches, to provide a baseline for detecting climate change impacts on organic matter decomposition. Furthermore, there is some urgency to apply ¹⁴C measurements soon, given the anticipated loss of ¹⁴C bomb-C determinations in coming decades (Graven 2015).

Opportunity: evaluate abiotic sources of soil CO₂

There is also a relatively unexplored opportunity to utilize partitioning experiments to evaluate abiotic sources of soil CO₂. The general assumption in most soil flux studies is that only biological communities near the soil surface are responsible for the observed CO₂. This is of course a simplification, and CO₂ in the soil profile may originate from a variety of sources, including soil pedogenic or liquid water carbonates (Cerling et al. 1991; Serrano-Ortiz et al. 2010; Shanhun et al. 2012; Risk et al. 2013b), methane oxidation (Romanak et al. 2012; Mills et al. 2013), or volcanic activity (Werner and Brantley 2003; Viveiros et al. 2008; Beulig et al. 2016). Soil sinks for CO₂ may also exist, for example from carbonate dissolution (Shanhun et al. 2012) and C-fixation by chemoautotrophic microbes (Beulig et al. 2016). Though these abiotic and anaerobic mechanisms are probably quite common and contribute at some scale to R_S, they are often studied only under specialty situations and we lack widespread estimates of the abiotic flux contribution. At individual sites, however, the contribution can be large: for instance, Rey et al. (2014) determined that volcanic CO₂ emissions accounted for almost 50 % of net ecosystem carbon balance at an arid steppe site in Spain.

Various noble gas tracers and isotopic methods are available for tracing CO₂ origin (Wilkinson et al. 2010), but there are also simpler approaches for abiotic partitioning that can readily be integrated into existing study sites. Researchers can 1) increase their use of geochemistry and gas ratios, 2) mine continuous temporal records more effectively, and/or 3) add gradient sensors or measurements. For example, the ratios of two common gases can help infer source contributions based on known stoichiometric relationships (Fig. 7).

Fig. 7

Mixing lines for biological respiration and methane oxidation (left) and example data for different idealized environments, each with strong source dominance (right) except for oil and gas thermogenic seeps, which are shown here as being mixed with biological contributions. Modified from Romanak et al. (2012)

What would a generic biotic-abiotic R_S measurement system look like? It would measure CO₂ plus another useful accessory such as O₂, at the surface (flux), and potentially other gases at a limited number of depths (concentrations) along with in-situ diffusivity measurement. Because aerobic respiration involves a theoretically constant ratio of CO₂ production to O₂ consumption, measuring deviations from this ratio can allow quantification of anaerobic respiration, or wetting events that preferentially store or advect CO₂ over O₂ (Turcu et al. 2005; Angert et al. 2015). Oxygen sensors are widely available, and although high accuracy is required for this application, they could presumably be added to autochambers (Liptzin et al. 2011). The system would likely need to run as often as possible, both to avoid the problem of bias from sparse data (Creelman 2015), but also to capture periods when biotic activity is negligible, so that other interesting background processes (e.g. geologic seepage, carbonate dissolution, wind ventilation of soil CO₂) are brought into focus.

Additionally, R_S studies do not often make full use of spatiotemporal possibilities to evaluate abiotic drivers. Non-growing season data are particularly well suited for identifying the presence of continuous abiotic activity such as volcanic contributions that may otherwise be masked (Risk et al. 2013a; Rey et al. 2014). Also, with the growth of temporally rich data from autochamber systems, spectral or frequency-based analytical techniques (e.g., frequency filtering, wavelet) become useful (Vargas et al. 2010b; Heinemeyer et al. 2012). While not currently common, they may also help identify abiotic background drivers of R_S that may vary on unusual timescales for biological processes (i.e., independently of moisture, temperature, and phenology). Finally, from a spatial perspective, subsurface concentration gradient studies can be extremely useful in helping identify layer-by-layer processes (Tang et al. 2003; Davidson et al. 2006; Vargas and Allen 2008; Phillips et al. 2012; Maier and Schack-Kirchner 2014), and basal fluxes that originate from below the soil profile. It would not be surprising if much of the new understanding in R_S over the coming years comes from gradient-based techniques, because of its ability to vertically resolve respiration across gradients of soil age, substrate quality, and O₂ concentration, and because of the potential for parameterizing physical-biogeochemical models.

Challenge 3: better upscaling and downscaling

The final challenge we address is in how to bridge R_S measurements to spatial and temporal scales matching those of other ecosystem C measurements and model outputs. One of the largest uncertainties associated with comparisons between R_ECO and R_S (Challenge 1, above) is the upscaling of chamber measurements to the footprint of the EC tower. The difference in ground measurement areas between individual soil collars (0.01–0.1 m²) and a 30 m forest EC tower (~1 km²) give a scaling factor of 10⁵–10⁶. Similarly, gap-filling monthly survey measurements to infer an annual total requires estimating over 99 % of the time period. These numbers suggest that high spatial variability of R_S is an additional plausible reason that R_S is often not matched well with R_ECO.

Upscaling often entails taking a simple mean of measurement locations distributed across a tower footprint, and some studies have performed power calculations to derive the minimum acceptable number of R_S measurements (Rodeghiero and Cescatti 2008). More rarely, regression models relating R_S to the spatial distribution of its main environmental drivers, including roots, microbial substrates, soil temperature and soil moisture, or soil depth (Tang and Baldocchi 2005; Saiz et al. 2006; Savage et al. 2008; Martin and Bolstad 2009), are used. Often, however, other easily measured soil variables have limited capability to predict spatial variation in R_S (Stoyan et al. 2000; Allaire et al. 2012; Kreba et al. 2013). For instance, at Harvard Forest, Giasson et al. (2013) found no more correlation among neighboring collars than among distant collars. In fact, they showed that variation due to spatial variability was as great as the variation due to experimentally imposed manipulations, such as soil heating, rain exclusion, selective harvest, and N additions.

Multiple studies have discussed the challenge of representing spatial heterogeneity of R_S (e.g. Stoyan et al. 2000) where geostatistical techniques have been used to demonstrate the challenge of measuring R_S from the plot to the landscape scale (Kosugi et al. 2007; Teixeira et al. 2013). One of the reasons for this lack of predictive ability is the tendency of R_S to have “hotspots” and “hot moments”, locations or periods of time with disproportionally high R_S emissions compared to surrounding conditions (Kim et al. 2012; Leon et al. 2014). (sensu McClain et al. 2003) suggested that these occur when episodic hydrologic flow paths deliver substrates and nutrients in abundance to locations of high biological activity. However, Leon et al. (2014) showed that hotspots may not be predictable, as they may appear in areas where no evidence of high metabolic activity was present during antecedent measurements. Furthermore, Leon et al. showed that the location and intensity of hot spots can change seasonally, and can be correlated with different environmental variables in different seasons (Fig. 8).

Fig. 8

Spatial patterns of R_S generated by ordinary kriging for the dry season (a) and wet season (b) in a Mediterranean shrubland. The figure shows the emergence of a hot spot (i.e., area of high R_S) during the wet season (orange area in panel b). Reprinted with permission from Leon et al. (2014)

This spatial variability makes measuring R_S representatively and robustly, and using these measurements to compare and evaluate R_ECO derived from EC, more challenging. One way researchers can help to identify the presence of hot spots and hot moments to other data users is to report the full probability distribution function (or at least the second and third moments—variance and skewness) of observed soil respiration in addition to site means (Lavoie et al. 2014). Ideally, complete data sets that include full time series of raw data and all sampling locations would be made available in archived repositories, allowing users to define hot spots and hot moments using different algorithms, and re-analyze older data as new techniques or insights become available.

Opportunity: ‘smarter’ quantification of site-level R_S

While we do not offer a blanket approach for upscaling, several techniques can provide better site- or ecosystem-level estimates than simple means of chamber locations. If R_S spatial variance and its drivers are well quantified, stratified (as opposed to random) sampling can allow for more efficient measurements and robust upscaling (Rodeghiero and Cescatti 2008). For instance, soil chambers in distinct vegetation types can be weighted by the fractional extent of vegetation types within the EC tower footprint (Giasson et al. 2013). Taking this approach a step further, footprint analysis can be employed to account for movement of the EC tower footprint through time, i.e. applying a time-varying source weight function to individual chambers (Budishchev et al. 2014). In addition, random variability of R_S across space and scales can be simulated with geostatistical techniques such as kriging (Herbst et al. 2012; Allaire et al. 2012; Leon et al. 2014).

In complex terrain, in contrast to typically flat EC sites, empirical models relating R_S to hydrologic factors may provide ‘smart’ upscaling to watershed scales. For instance, Riveros-Iregui and McGlynn (2009) demonstrated that upland accumulated area was a robust predictor of R_S, because both soil moisture and R_S increased at valley bottoms relative to ridges. However, the opposite pattern was found at a different Western U.S. site, where elevation-related cooling increased soil moisture, and thus R_S, at high elevation positions (Berryman et al. 2015). Ultimately, researchers must discern effects that covary with topography such as soil moisture, soil organic carbon (as available substrate for Rs), temperature, and vegetation type to fully understand the biophysical controls on R_S in complex terrain.

R_S measurements frequently must be scaled in time as well as space, as most site-level estimates are derived from sporadic manual measurements, not continuous ones. While seasonal differences in R_S can be large (e.g. Khomik et al. 2006), seasonal patterns tend to be relatively straightforward to account for in gap-filling regression models, absent the effects of disturbances or drought (Borken et al. 2006; Martin et al. 2012; Barba et al. 2016). More difficult to simulate are R_S ‘bursts’ after rewetting or thawing (Kim et al. 2012). This phenomenon, known as the “Birch effect”, is thought to be due to some combination of increased microbial populations, higher metabolism, and changes in the physical protection and hydrological connectivity of the soil (Xu and Baldocchi 2004; Tang and Baldocchi 2005; Oikawa et al. 2014). These bursts present similar challenges as other types of ‘hot moments’, but have more predictable occurrence, and in some cases can constitute an appreciable fraction of annual fluxes. Also difficult to represent are seasonal or diurnal hysteresis of R_S responses to temperature (Kicklighter et al. 1994; Riveros-Iregui et al. 2007; Phillips et al. 2011), and linkages and lags between photosynthesis and respiration (reviewed by Carbone and Vargas 2007; Davidson and Holbrook 2009). Oikawa et al. (2014) proposed a model that could simulate wetting-related pulses and temperature hysteresis, and Zhang et al. (2015)used a CO₂ gas transport model to simulate temperature hysteresis. However, there are presently no standardized gap-filling algorithms or approaches for soil respiration, unlike for EC (Moffat et al. 2007). In the most comprehensive work to date, Gomez-Casanovas et al. (2013) assessed a variety of gap-filling R_S algorithms and found that different methods exhibited different levels of skill with data gaps of various lengths, but rarely produced large systematic biases. Obviously, continuous data are preferable to periodic campaign-style measurements for constructing annual carbon budgets, and systematic biases can be introduced when measurements are made always at the same time of day (Savage et al. 2008; Savage et al. 2009; Gomez-Casanovas et al. 2013). Nevertheless, supplemental periodic measurements with a portable instrument are generally necessary to capture site spatial variability.

In addition to estimating a site average R_S it is necessary to estimate the associated uncertainty (Reichstein and Beer 2008). While uncertainty of individual R_S chambers can be quantified readily (Pumpanen et al. 2004), describing site-level uncertainty is more difficult, as it includes uncertainty due to spatial variability in R_S, gap-filling assumptions, and the goodness of fit of the statistical model used for scaling (Lavoie et al. 2014). A growing body of research is indicating consistencies in the statistical features of R_S in both time and space, work that may provide simplifying solutions for uncertainty quantification. For example, at Willow Creek Ameriflux site (Lavoie et al. 2014), at Harvard Forest (Savage et al. 2008), and at multiple sites around the world (Cueva et al. 2015), the standard deviation of R_S measured at multiple locations within the ecosystems was heteroschedastic, increasing linearly with site average R_S. Similarly, in the temporal domain, the random error of R_S (defined as the deviation between measures made at one location under as-similar-as-possible conditions) was found not only to scale with R_S, but to scale with the same slope across a large number of sites, and even in different ecosystems (Savage et al. 2008; Lavoie et al. 2014; Cueva et al. 2015).

Opportunity: downscaling tower measurements

In addition to scaling up R_S to the EC site level, there are emerging opportunities for decomposing EC fluxes to capture spatial variability within the tower footprint, perhaps more in line with the spatial scale of R_S. Flux tower based R_ECO estimates may suffer from systematic bias due to constantly changing flux footprint source area with time, as a function of wind direction, surface structural properties, and atmospheric stability (Schmid, 2002). However, techniques that take advantage of this footprint variability can be used to map spatial variation and provide gridded estimates of R_ECO.

The Environmental Response Function (ERF) approach (Metzger et al. 2013) is one such approach to direct attribute the measured flux to the actual source area of contribution and use that information to appropriately scale tower fluxes to other measurements such as chambers . As applied to towers (Xu et al. 2017), the ERF techniques estimates minute-to-minute variations in the flux footprint to compare measured fluxes against covarying spatial and temporal drivers. The comparison allows a statistical model to be built to map fluxes and their uncertainty across space, using a machine learning approach. With ERF, down-scaled NEE observations can then used to derive a re-upscaled, gridded R_ECO and directly compared to point-level Rs chamber measurements. To date, ERF has only been applied to direct fluxes (NEE), but a new opportunity arises to use ERF in partitioning and improving comparison of daytime R_ECO estimates and R_S.

Using R_S for model development

Aside from the impetus to use as many data sources as possible for data-model synthesis, field R_S data have the tantalizing potential to provide a critical bridge across scales and domains. In R_H models, there is a significant scale gap between lab incubations and litter decomposition experiments, and the field- and global-scale questions that are being asked of these models. R_S measurements scaled to site and ecosystem levels (see Challenge 3) could reduce some of this gap and bridge key spatial-temporal heterogeneities in the R_H model. In addition, NEE is currently the only measurement routinely used for model development which crosses both soil and plant domains. Having additional information from R_S for model benchmarking would allow modelers to better identify poorly-performing sections of land carbon models (Luo et al. 2012).

The few studies that have directly tested the value of R_S for model fidelity also suggest the importance of these data. Keenan et al. (2013) rated the utility of data streams for data-model synthesis, and found R_S data were one of the most valuable measurements for reducing uncertainty of forest productivity estimates. They concluded that most model improvement is seen with a limited number of measurement types, and speculated that the most valuable measurements provide information about turnover of C pools at both fast and very long timescales. Similarly, Zobitz et al. (2008) tested whether R_S could be simulated by a forest C model using only NEE and soil C stock data for parameterization, and found that the model could not be satisfactorily constrained, implying that Rs brings distinct and unique information to any model-data assimilation exercise.

We see three types of model improvement activities that could utilize R_S data: (1) data-model synthesis where the aim is reducing uncertainty of parameter estimates, (2) preventing models from predicting biologically implausible states, and (3) benchmarking a range of models to common sets of observations (Kelley et al. 2013). In Table 2 and below we give examples using R_S data for each of these.

Table 2 Uses of R_S data for terrestrial model improvement

Data-model synthesis techniques attempt to identify and select models, parameters, and uncertainty that are more parsimonious or consistent with the entire range of collected observations. These approaches, which can be Bayesian in nature, attempt to propagate a probabilistic set of consistent parameters forward, allowing for direct estimation and partitioning of sources of uncertainty (LeBauer et al. 2013; Dietze et al. 2013). One advantage of these approaches is that simple assumptions often made on scaling, gap-filling, and single variable regression can be removed if model processes are able to simulate components or model outputs can be “sampled” like the original observation. For R_S, this means that infrequent survey measurements can be assimilated without requiring separate gap-filling procedures.

Particularly relevant for R_S, data-model synthesis allows observations to be used even when they do not match one-to-one with modeled variables. For instance, incorporating R_S in data-model syntheses with a forest C model was found in one instance to improve estimation of NEE (Keenan et al. 2013), and in another instance to reduce the uncertainty in heterotrophic respiration and the size of soil carbon pools (Keenan et al. 2012). Direct model-data integration would require terrestrial C models to simulate belowground R_A, which many do not. Autotrophic respiration is usually split into growth and maintenance respiration processes, but rarely into separate above- and below-ground components (Medvigy et al. 2009; Oleson et al. 2010; Parton et al. 2010; Dunne et al. 2012). Pushing for the development of allocation schemes that distinguish belowground R_A in global scale models would provide an elegant solution to the comparison problem, but it is by no means the only option.

In addition to reducing parameter uncertainty, another approach for data-model synthesis is to focus on preventing parameters and states that are not biologically plausible (Bloom and Williams 2015). A common problem for C models is equifinality: they can produce similar annual fluxes with different underlying parameters and process representations. Poor model parameterization is likely to translate to weak predictive power under changing environmental conditions. Perhaps the most powerful use of R_S data is as a diagnostic tool to identify implausible parameterization of R_H or R_A, as a form of ecological dynamic constraint (Bloom and Williams 2015). For instance, to check on the parameterization of R_A, one could use R_S as a co-constraint with aboveground measures of plant growth, such as biometric or remotely-sensed measures of NPP. To check on the parameterization of R_H, one could use R_S and a co-constraint with C stock or ¹⁴C data, thereby encompassing turnover processes at immediate to long timescales.

Comparing modeled and observed seasonal dynamics of R_S provides another means for bounding plausible conditions. There are some situations (e.g. winter dormancy) where R_A is all but shut down and we would expect R_S to be roughly equal to R_H. In these cases, relatively straightforward direct comparison of R_S and R_H is likely reasonable. More creative comparisons could be done by examining relative shifts of R_S across an environmental gradient where either variations in R_H or R_A are expected to dominate, or where belowground R_A is expected to be proportional to some other measurement like leaf R_A. Ideally competing hypothesis would be codified in model formulations, leading itself to formal data-model integration and model selections. Clearly, this is a study-specific approach that requires a deep understanding of the mechanisms being investigated, but has the potential to stimulate creative uses of data-model synthesis techniques.

Building on these ideas, R_S data can also play a valuable role in benchmarking multiple models against common sets of observations. Luo et al. (2012) summarized several key qualities that benchmark data should have. First, benchmark data should reflect properties that are fundamental to all ecosystems, such as C uptake and residence time. As discussed above (see section 2.3), R_S provides a critical data source for evaluating both soil and plant C residence time. It is particularly powerful for evaluating residence time at the timescale of manipulative experiments, and when measured in combination with isotopic dating methods. Second, benchmark data should be selected to reduce equifinality. R_S provides a focus on soil processes that may be obscured in EC data or remotely-sensed data products, and is thus particularly useful for detecting weaknesses in belowground parameterization. R_S data are also abundant and geographically widespread, another important quality for reducing equifinality in global-scale models. Third, benchmark data should be reliable. As discussed above (Challenge 1), R_S measurement systems have the benefit of validation under laboratory conditions, and are likely more reliable than EC-derived estimates of R_ECO.

Luo et al. (2012) also noted that land models can be evaluated on the basis of seasonal patterns rather than absolute values. R_S data are particularly useful for calibrating soil models at seasonal timescales, since soil C stock and even litterbag data integrate decomposition processes occurring over timescales ≥1 year. For example, Fig. 9 shows seasonal patterns in R_H and R_ECO simulated by the Community Land Model (CLMv.4.0) in comparison to observed R_S and R_ECO from the a pine forest site in Oregon (US-ME2, model data from Hudiburg et al. 2013). Even after site-specific tuning of the model, seasonal patterns had considerable mismatch with observations. Modeled fluxes were strongly suppressed by seasonal summer drought. As a result, R_H peaked in winter rather than during the growing season, and modeled R_ECO peaked two to three months early, demonstrating fundamental problems in representation of soil and plant activity.

Fig. 9

A comparison of R_H and R_ECO modeled with CLM v.4.0 to chamber R_S and EC-based R_ECO (Level 4 data) at the AmeriFlux Metolius intermediate pine site (US-ME2). Modeled R_ECO peaked two to three months earlier than observations, missed a large portion of the growing season, and is decoupled from R_H, which peaked in winter rather than during the growing season. Model output is from Hudiburg et al. (2013)

Conclusions and future directions

R_S data are a rich source of information on soil responses to land use and climate factors, and are important for advancing our understanding of the terrestrial C cycle. However, to a large extent R_S data have not been integrated with modeling efforts nor with analyses of related C fluxes and pools. We argue that in field studies, R_S data are particularly useful for:

• 1) Identifying and potentially gap-filling low-quality nighttime EC data. R_S data are completely independent of, and not subject to the same nighttime biases as EC. They therefore provide a lower plausible bound for R_ECO and more generally an additional EC data quality metric. R_S measures can also be independently validated such as with sand-column flux generators. For these reasons, R_S measurements should be considered important components of EC site instrument arrays, particularly at forest sites with complex canopies.

• 2) Determining soil responses to non-equilibrium conditions over experimental time frames. Given the slow turnover of the majority of soil C pools and high spatial variability, it is often difficult to detect statistically significant changes in soil C stocks in response to manipulations of land use or climate variables. Measurements of R_S and partitioned R_H components therefore provide an important means to detect experimental treatment effects on soil C turnover.

• 3) Identifying sources of R _S . While partitioning R_S into autotrophic and heterotrophic components is valuable in its own right, applying physical partitioning in combination with isotopic monitoring makes it possible to evaluate contributions from recent and old C by both plants and microbes, as well as abiotic CO₂ sources, under field conditions.

• 4) Quantifying spatial variability within tower footprints. Soil flux measurements can be used to identify hot spots and hot moments, evaluate spatial variables that control R_S, and also evaluate the representativeness of EC tower measurements within the larger domain.

In addition, R_S data can be used in the following ways to improve terrestrial C cycle models:

• 1) Improving data reliability. R_S provide directly-measured respiration fluxes, in contrast to inferred EC fluxes based on partitioning NEP, and can be used to reduce the weight given to low quality EC data during data assimilation procedures.

• 2) Testing multiple-pool soil C models. When measured in combination with partitioned fluxes (R_H and R_A), or ¹⁴C abundance, R_S data can be utilized for calibration of complex soil C models that represent ‘new’ and ‘old’ C contributions to both R_H and R_A, or additional C sources, such as from geologic sources.

• 3) Preventing biologically implausible states. When used as a joint constraint with soil C stock data, R_S can improve modeled soil C turnover across immediate to very long timescales. When used as a joint constraint with plant growth data, R_S can provide a plausibility test for evaluating R_A and plant allocation.

• 4) Addressing model equifinality. R_S data can help to ensure that belowground processes are well-constrained in complex terrestrial models. Because R_S crosses scales and plant and soil domains, it can reduce the uncertainty of several related variables in data-model synthesis techniques, including NEE, R_A, and R_H, and soil C stocks.

• 5) Providing data for model benchmarking. Because R_S data are abundant, geographically widespread, and generally reliable, they have qualities that make them well-suited for benchmarking models.

Taking advantage of the full potential of R_S measurements will require a concerted effort to make R_S data more accessible and intercomparable among sites. Many measurement ‘best practices’ and data processing QA/QC procedures have been suggested, but need to be more broadly adopted by the terrestrial ecology community (Bahn et al. 2008; Heinemeyer et al. 2011; Gomez-Casanovas et al. 2013; Lavoie et al. 2014; Maier and Schack-Kirchner 2014; Cueva et al. 2015). While the Global Soil Respiration Database (Bond-Lamberty and Thomson 2010) has made an extensive amount of seasonally and annually averaged Rs data available, similar efforts are needed to gather and make publicly available instantaneous flux measurements from survey campaigns and autochambers (McFarlane et al. 2014). Ideally, efforts to catalog both raw and processed R_S data would allow comparison with contemporaneous EC tower fluxes and other soil C data, including SOC stocks, isotopes measurements, and related biogeochemistry data. These efforts can help to usher new global syntheses, and progress in both measurement and modeling realms.