New Techniques and Data for Understanding the Global Soil Respiration Flux

Abstract

Soil respiration (R_s; the soil surface‐to‐atmosphere CO₂ flux) has been measured in the field for decades, but only recently have we begun to assemble and leverage these small‐scale but extensive data. Recently, Zhao et al. (2017, https://doi.org/10.1002/2016ef000480) applied a novel artificial neural network model to the problem of estimating the global R_s flux and understanding its variations between regions and biomes. Their results point to a convergence in estimates of global R_s, and the power of leveraging the long record of observed R_s in global ecosystems, but also to uncertainties about soils' response to climate change. It will take a combination of long‐term studies, data syntheses, modeling intercomparisons, and probably a new generation of sampling networks and experiments to fully resolve these questions.

Introduction

The soil‐to‐atmosphere CO₂ flux (“soil respiration”; R_s) that results from both root and microbial respiration constitutes a large part of the terrestrial carbon cycle (Le Quéré et al., ). The R_s flux is thought to be increasing (Hashimoto et al., ) due to changes in climate, land cover/disturbance, and perhaps other biogeochemical cycles (Hursh et al., ), but the magnitude, variability, sensitivity, and sources of global R_s remain poorly constrained.

Starting with Schlesinger (), estimates of regional to global R_s fluxes have generally used either simple upscaling of land cover mean rates, or linear regressions driven by climate data and land area. This has been necessary because we have no reliable way of measuring R_s at areas larger than ~0.1 m⁻² (Bond‐Lamberty et al., ), unlike, for example, net carbon exchange or net primary production, for which robust larger‐scale measurement methods exist. Only recently have we begun to leverage these small‐scale but extensive and informative data (Phillips et al., ). Although such chamber‐scale measurements have been assembled into openly available databases (Bond‐Lamberty & Thomson, ), the sparse coverage and significant errors associated with such spatially limited observations challenge attempts to estimate and characterize larger‐scale R_s fluxes, and the statistical approaches above have typically explained only 30–40% of observed variability. This limits our ability to draw inferences about current soil carbon dynamics and, potentially, the robustness of the inferred land carbon sink (Le Quéré et al., ).

Study Results

Zhao et al. () take this problem another step forward by applying an artificial neural network (ANN) model to the problem. ANNs are a type of machine learning (Jordan & Mitchell, ) in which one or more layers of artificial neurons form a mapping between inputs (in this case, potential drivers of R_s such as biome types, climate, and other ecosystem stocks and fluxes) and outputs (annual R_s). The output of each artificial neuron is calculated by a nonlinear function of its inputs, and the nodes have weights and threshold values that are adjusted as learning—the gradient‐based optimization algorithm minimizing errors in the output layer—occurs. In the Earth and environmental sciences, ANNs have been used to model and gap‐fill eddy covariance flux data (Knox et al., ; Moffat et al., ) and for remote sensing image classification (Sedano et al., ), but have had very limited previous use in soil respiration (e.g., Song et al., /9).

Combining this novel ANN approach with updated, spatially detailed driver data sets, Zhao et al. () focus on prediction of R_s and the degree to which this carbon flux varies systematically between biomes (boreal forests, temperate forest, grasslands, savannas, etc.). They find that (i) the ANN was notably more successful than simpler approaches, explaining 50–60% of R_s variability, with no evidence of overfitting; (ii) that the limited amount of available R_s data from some biomes (e.g., drylands and deserts, for which few published R_s data exist) hampered the ANN's ability to predict at the biome level; and (iii) tropical forests dominate the global flux. They estimate a global flux of 93.3 Pg C/year increasing by 0.04 Pg C annually, implying a temperature sensitivity of 1.6 Pg C/°C, and a moisture sensitivity of 0.5 Pg C per 10 mm increase in mean annual precipitation. Interannual variability was low (~6%) at the global level and in forests (2–3%) but much higher in deserts, tundra, and wetlands. Finally and interestingly, they find that high‐latitude (boreal and Arctic) R_s exhibits significantly different dynamics than in other biomes.

Emerging Understandings

This study points to a number of interesting points, and broader understandings that are emerging. First, different upscaling and/or modeling techniques applied to the observational data are converging (Figure ) on a global R_s estimate of 91–94 Pg C/year (Xu & Shang, ) increasing at an average rate of 0.04–0.1 Pg C/year. It is important to note however that (i) most of these estimates start from a common global data set (Bond‐Lamberty & Thomson, ), and thus are not truly independent, and (ii) Jian et al. () challenge this consensus, arguing that annual data are not adequate for global R_s modeling and thus the global flux is considerably lower. Such a lower flux would be more consistent with early indirect estimates (Schlesinger, ) but pose consistency problems with, for example, higher global gross primary production estimates of 150–175 Pg C/year (Anav et al., , and references therein). If this 91–94‐Pg C/year R_s estimate proves robust, however, it—in conjunction with measurements from remote sensing platforms from, for example, the OCO‐2 mission—puts better constraints on the inferred land carbon sink (Le Quéré et al., ), the sensitivity of terrestrial ecosystems to land use and climate change, and provides a robust benchmarking data set for global models (Friedlingstein et al., ).

Fig. 1. Historical estimates of global soil respiration (Rs), including that of Zhao et al. () (in red). Each point shows a different calculated global Rs, with 95% confidence interval (error bars) and temporal trend (dashed line) if available. Points labels: (1) Schlesinger (), (2) Raich and Schlesinger (), (3) Raich et al. (), (4) Raich and Potter (), (5) Peng and Apps (), (6) Chen et al. (), (7) Hashimoto (), (8) Bond‐Lamberty and Thomson (), (9) Adachi et al. (), (10) Xu and Shang (), (11) Hashimoto et al. (), (12) Zhao et al. (), (13) Jian et al. ().

Second, we should welcome new techniques applied to existing data sets, specifically acknowledging the potential of new numerical/algorithmic methods to synthesize existing observations for better quantification and understanding of large‐scale dynamics. Previous examples include wavelet coherence analysis applied to R_s (Vargas et al., ), formal meta‐analysis to draw inferences about the R_s response to nitrogen addition (L. Zhou et al., ), machine‐learning Random Forest algorithms used to generate high‐resolution soil maps (Hengl et al., ), and model tree ensemble trained on remote sensing, climate, and land use data (Jung et al., ). New approaches should be rigorously tested for bias and predictive ability; one concern about ANNs, for example, is that they may not make robust predictions outside of the domain of their training data (Richardson et al., ). Zhao et al. () are careful in this area, limiting their predictions and inferences in biomes with limited data.

Third, as is pointed out by Zhao et al. () and many previous authors, the extant R_s data available remain sharply limited in particularly critical climate‐ and human‐relevant regions such as high‐latitude ecosystems, wetlands, tropical forests, and mountain regions. Over the last 15 years, many new publications from Asia (particularly China) have greatly expanded the available published R_s data (Figure ), but there remain troublingly few observational data from South America and Africa (Epule, ), even though Zhao et al. () estimate that tropical forests and savannas comprise one third of the global R_s flux. A formal global R_s sampling design and prioritization, based for example on multivariate spatiotemporal clustering or other data mining techniques (Bond‐Lamberty et al., ; Hoffman et al., ), would be a significant step forward. In a related vein, Xu and Shang () recently discussed global sampling priorities and underrepresented ecosystems.

Fig. 2. Source of published soil respiration observations over time. Data are from the SRDB (Bond‐Lamberty & Thomson, ), accessed from https://github.com/bpbond/srdb.

Unanswered Questions and Future Work

The results from Zhao et al. () emphasize, to me, three critical areas in which progress must be made. First, understanding how R_s temperature and moisture sensitivities vary in time and space (Hursh et al., ; Liu et al., ; T. Zhou et al., ), and the degree to which “hot spots” and “hot moments” in space and time (Leon et al., ) might affect our sampling priorities at scales from the collar to international network (Bond‐Lamberty et al., ). Apparent temperature sensitivity (Q₁₀) may vary with factors such as temperature and biome type (Reichstein et al., ; T. Zhou et al., ), but most Earth system models use a globally constant Q₁₀ (Anav et al., ). A useful step forward in this area might be more formally comparing and reconciling observed temperature and moisture sensitivities with global syntheses of soil carbon losses in warming and drought experiments (Carey et al., ), using the resulting understanding to develop more mechanistic and robust R_s formulations in global models (Todd‐Brown et al., ).

Second, this study supports previously tentative suggestions (Bond‐Lamberty & Thomson, ) that high‐latitude ecosystems' R_s fluxes seem to be responding differently to climate change. The unique characteristics—remoteness, extensive permafrost and peatlands, frequent wildfires, and high soil carbon densities—in many boreal and Arctic regions mean that they have a large potential climate feedback (Koven et al., ; McGuire et al., ), but one that is extremely poorly constrained (Schuur et al., ). For example, Arctic soils are thermally buffered by snowpack and exhibit strong temperature gradients across the freeze‐thaw boundary, and thus exhibit high apparent temperature sensitivities that are poorly captured by models (Koven et al., ). Such biome‐specific processes are important to capture in modeling frameworks for robust predictions (Chadburn et al., ).

Finally, there is perhaps no more pressing question in all of terrestrial biogeochemistry than the degree to which soils will respond to future climate change—specifically, the degree to which they may lose (via heterotrophic respiration and, secondarily, erosion) some of their enormous carbon pools to the atmosphere, exerting a feedback effect on the climate. The lability and sensitivity of soils to climate has profound implications across science, management, and policy arenas (Bradford, ). As noted above, R_s is thought to be slowly increasing due to climate change, but whether this results from soil carbon loss, and thus constitutes a climate feedback, is unknown; given the year‐to‐year variability in site‐, regional‐, and global‐scale soil carbon fluxes, it is unclear whether we are even able to observe a putative heterotrophic respiration increase and soil carbon loss. It will take a combination of long‐term studies, data syntheses, modeling intercomparisons, and probably a new generation of sampling networks and experiments to fully resolve these questions.

Acknowledgments

This research was supported by the U.S. Department of Energy, Office of Science, Biological and Environmental Research, as part of the Terrestrial Ecosystem Sciences Program. The Pacific Northwest National Laboratory is operated for DOE by Battelle Memorial Institute under contract DE‐AC05‐76RL01830. As a commentary, this paper does not present new data. The data (collected from published studies) and code necessary to reproduce the Figures are deposited at https://doi.org/10.6084/m9.figshare.6275189.v1.