the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
What information on contemporary terrestrial carbon dynamics can an integration of current geospatial data confidently provide?
Abstract. The terrestrial carbon (C) cycle is a key global biogeochemical cycle, linked to climate change and regulation, provisioning and regulating ecosystem services and underpinning human livelihoods. To understand climate risks and to support the development of effective climate mitigation strategies requires rigorous, uncertainty bounded, systemic information on C dynamics and their underpinning ecological processes. Significant effort has been invested into attempting to quantify current C stocks and exchanges of C. These data are often incomplete in space and time, contain poorly characterised errors, and lack internal consistency checks, making the construction of a reliable global C budget a significant challenge. Here, we ask what information on contemporary terrestrial carbon dynamics can an integration of current geospatial data confidently provide?
We present a global (0.5 × 0.5 deg), multi-decadal (2003–2024) analysis of the terrestrial C-cycle at a monthly time step. Using a Bayesian model-data fusion framework (CARDAMOM) we calibrate a model of terrestrial ecosystems (DALEC) integrating diverse ecological spatio-temporal observations (leaf area, absorbed photosynthetically active radiation, gross primary production, woody biomass, soil C), and forcing (meteorology, atmospheric CO2, burned area and forest loss). CARDAMOM propagates observational uncertainties through to the retrieved DALEC parameters (i.e. ecosystem properties) and simulates C states and fluxes for each of 55,246 pixels across the vegetated land surface. The resulting analysis is systemic (i.e. consistent between states and fluxes) and can support national and international policy development (e.g., Global Carbon Project).
Our analysis shows that current global multi-decadal datasets are inadequate to provide > 95 % confidence on the sign of net exchange across 91 % of the vegetated land surface. At this confidence level we can identify 0.23 % of vegetated land is a net source, with 0.75 % a net sink and 7.3 % as neutral. Although we note that at 68 % confidence we the area identified as likely a sink increases to 30 % without change to the area identified as neutral or a source. We show that uncertainty in net C exchange is strongly correlated with the dynamics of wood (r = 0.8) and soil organic matter (r = 0.7) but poorly correlated with gross fluxes which have historically been a focus of research such as photosynthesis (r = 0.1). New satellite missions such as BIOMASS have more accurate wood biomass sensing, and so should enhance confidence in sink/source detection in the next five years. Soil C dynamics data remain a major challenge, with research and investment required to provide a global constraint on soil C stock changes.
- Preprint
(27647 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2026-3014', Anonymous Referee #1, 13 Aug 2026
The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2026/egusphere-2026-3014/egusphere-2026-3014-RC1-supplement.pdfCitation: https://doi.org/
10.5194/egusphere-2026-3014-RC1 -
AC1: 'Reply on RC1', Thomas Smallman, 10 Sep 2026
The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2026/egusphere-2026-3014/egusphere-2026-3014-AC1-supplement.pdf
-
AC1: 'Reply on RC1', Thomas Smallman, 10 Sep 2026
-
RC2: 'Comment on egusphere-2026-3014', Anonymous Referee #2, 14 Nov 2026
The authors use the CARDAMOM Bayesian model–data fusion framework to calibrate the DALEC model globally by assimilating MODIS LAI and fAPAR, annual MODIS GPP, woody biomass time series, and priors on initial soil C and Ra:GPP. They report global and RECCAP-regional C budgets, trends, NPP allocation and residence times, evaluate against OCO2-MIPv10 NBE and GFEDv5/GFAS fire emissions, and conclude that current data cannot determine the sign of NBP with >95 % confidence over ~91 % of the vegetated land surface, with uncertainty in ΔSoil and ΔWood dominating NBP uncertainty. This topic fits the journal well. My main concern is that I still have the question posed in its own title after reading it a couple of times.
The title asks “what information the integration of current geospatial data can confidently provide on contemporary terrestrial carbon dynamics”. Answering this question requires comparing what is known with the data against what is known without it (i.e., prior model predictions). The manuscript reports only the posterior predictions: a set of posterior budgets and their confidence intervals. From the posterior alone a reader cannot tell whether a wide interval reflects data that carry little information, data whose information was diluted by the assumed observation errors, the choice of uniform prior ranges and EDCs, model structure, or the fully-correlated aggregation choice. Nor can a reader tell which data stream constrained which parameter or output, or whether the assimilation actually improved the simulation relative to an unconstrained (prior) model.
In this manuscript, the "information" is not really measured. It is inferred from a couple of proxies, such as the width of the posterior confidence intervals (e.g. the uncertainty:flux ratios, L275–281), histogram overlap with the assimilated data and priors (Sect. 3.1), spatial correlations between pixel-level CI widths (Fig. 7), and within-pixel ensemble correlations offered as hypotheses (Fig. A3). None of these compares the posterior with a reference state of knowledge, so none quantifies what the data added. The authors may need to develop or adopt a formal metric of information contribution, such as the relative information approach of Weng and Luo (2011). Overall, this research should examine the following questions as usual data assimilation studies: (1) how much the parameters were constrained; (2) what information each assimilated dataset contributed; and (3) how much the simulated carbon states and fluxes improved as a result. All of these variables should be expressed through a formal information metric.
Major suggestions:
1. Show how the parameters were constrained (prior vs posterior) by the assimilated datasets
Table A1 gives only the uniform prior ranges. The posterior parameter distributions are never shown against those priors, so the reader cannot see which ecosystem properties were actually learned from the data. The only parameter-level results are spatial correlations among posterior medians (Figs. A1–A2) and the statement that posteriors overlap the assimilated priors for LCA, Ceff, Ra:GPP and initial soil C (L246–247). A high overlap with the prior is ambiguous: it is consistent both with good agreement and with no update at all.
2. Quantify what information each assimilated dataset provided
The paper's policy-relevant conclusions concern the value of particular data streams (e.g. that BIOMASS-type wood data will improve sink/source detection while GPP and LAI matter less, L20–23, L295–299, L413–418). These are inferred from spatial correlations between posterior CI widths (Fig. 7), which is an indirect and, I think, unreliable route.
There are some examples from previous studies, including those from the authors of this manuscript. Richardson et al. (2010) assimilated different combinations of stock and flux data into DALEC at Howland Forest. Adding woody biomass increment and soil respiration both narrowed the predictive intervals and reduced bias in the forward run, while no data stream constrained fine-root or soil C dynamics. Smallman et al. (2017) compared retrievals using repeated, single and no woody biomass observations. Repeated biomass reduced the uncertainty of retrieved ecosystem traits by up to 50 % and reduced uncertainty and/or bias in all C pools, not only wood. At global scale, Peylin et al. (2016) assimilated three data streams in turn into ORCHIDEE and reported the effect of each step. Applying the same logic in this study would make the manuscript's claims testable.
3. Develop or adopt a formal metric of information contribution
To answer the title question, those results need a common, quantitative currency. I strongly recommend that the authors develop or adopt a formal measure of information contribution. A directly applicable precedent is Weng and Luo (2011, Ecol. Appl., 21, 1490–1505, https://doi.org/10.1890/09-1394.1). They quantified the relative information contributions of model and data to forecasts of forest C pools, using a Shannon information index computed from the probability density functions of pool sizes. In their framework, the state of no knowledge is the uniform PDF over the prior range. The model's contribution is the information gained from the uniform PDF to the PDF simulated by the unconstrained model. The data's contribution is the further information gained from that PDF to the posterior after assimilation.
4. Demonstrate how far the simulations improved
The evaluation (Sect. 3.1) shows that the posterior is consistent with assimilated and independent data, but not that it is better than it would have been without assimilation. For example, the NBE overlap of 60 % with OCO2-MIPv10, should be compared with the same statistics from the prior ensemble. Without this, it would be difficult for the reader to tell whether the data improved the NBE estimate. Additionally, it would be better to show prior vs posterior fit to the assimilated LAI, fAPAR, GPP and wood, e.g. a change in log-likelihood or RMSE than only showing the posterior overlaps.
5. Soil carbon dynamics
No information on soil C dynamics is assimilated: only a time-invariant prior on the initial soil C stock. It is therefore expected that ΔSoil is poorly constrained, and its uncertainty is set largely by the uniform prior on soil turnover (spanning almost two orders of magnitude, Table A1) and by the EDCs. This does not make the conclusion wrong, but the manuscript presents it as a finding about the global carbon cycle rather than as a consequence of the experimental design. The authors should show how much the soil-stock prior actually reduces Soil carbon uncertainty, and separate "no data were available" from "data were available but uninformative".
Minor comments
1. L60-61: CARDAMOM updates parameters and initial conditions (Table A1). Please say so here.
2. L84-90: Please give the Gelman–Rubin threshold and the burn-in and thinning details. Is 100 samples from the second half of each chain enough to represent the tails used for 95 % CIs?
3. L131-135: Deforestation/degradation removal fractions are fixed. Please give the values, their source, and the fraction of removed biomass sent to litter/soil, since LUC is ~0.7 PgC/yr in the budget.
4. L148-149: Distributing GFW annual loss evenly through the year will smear the seasonality of LUC and fire-related clearing. Please comment on its effect on the monthly NBE comparison.
5. L168-172: Assimilating annual GPP removes the seasonal constraint. Please note that LAI and fAPAR are the only seasonal constraints, and discuss whether this matters for the monthly evaluation against OCO2-MIP.
6. L201-208: Please define histogram overlap formally (equation), including how the independent ensemble range is converted to a distribution (e.g. uniform between min and max of two fire products?).
7. L283–284: The 10 gC m-2 yr-1 threshold for "neutral" is arbitrary. Please justify it and show sensitivity.
8. L320–322 and L447-449: Low R2 from linear regression on mean meteorology does not show that MRT is controlled by "factors implicitly captured in the assimilated data". It may equally reflect poorly constrained parameters (i.e. posterior medians near the prior center).
9. Table 1: Please clarify how NBE and NBP are related to NEE, fire and LUC (e.g. NBP = 6.0 but −NBE = 6.7), and the sign conventions.
References
Kaminski, T. and Rayner, P. J.: Reviews and syntheses: Guiding the evolution of the observing system for the carbon cycle through quantitative network design, Biogeosciences, 14, 4755–4766, https://doi.org/10.5194/bg-14-4755-2017 , 2017.
Kaminski, T., Scholze, M., and Houweling, S.: Quantifying the benefit of A-SCOPE data for reducing uncertainties in terrestrial carbon fluxes in CCDAS, Tellus B, 62(5), 784–796, https://doi.org/10.1111/j.1600-0889.2010.00483.x, 2010.
Keenan, T. F., Davidson, E. A., Munger, J. W., and Richardson, A. D.: Rate my data: quantifying the value of ecological data for the development of models of the terrestrial carbon cycle, Ecol. Appl., 23(1), 273–286, https://doi.org/10.1890/12-0747.1, 2013.
MacBean, N., Peylin, P., Chevallier, F., Scholze, M., and Schürmann, G.: Consistent assimilation of multiple data streams in a carbon cycle data assimilation system, Geosci. Model Dev., 9, 3569–3588, https://doi.org/10.5194/gmd-9-3569-2016, 2016.
Peylin, P., Bacour, C., MacBean, N., et al., A new stepwise carbon cycle data assimilation system using multiple data streams to constrain the simulated land surface carbon cycle, Geosci. Model Dev., 9, 3321–3346, https://doi.org/10.5194/gmd-9-3321-2016, 2016.
Richardson, A. D., Williams, M., Hollinger, D. Y., Moore, D. J. P., Dail, D. B., Davidson, E. A., Scott, N. A., Evans, R. S., Hughes, H., Lee, J. T., Rodrigues, C., and Savage, K.: Estimating parameters of a forest ecosystem C model with measurements of stocks and fluxes as joint constraints, Oecologia, 164, 25–40, https://doi.org/10.1007/s00442-010-1628-y, 2010.
Smallman, T. L., Exbrayat, J.-F., Mencuccini, M., Bloom, A. A., and Williams, M.: Assimilation of repeated woody biomass observations constrains decadal ecosystem carbon cycle uncertainty in aggrading forests, J. Geophys. Res. Biogeosci., 122, 528–545, https://doi.org/10.1002/2016JG003520, 2017.
Weng, E. and Luo, Y.: Relative information contributions of model vs. data to short- and long-term forecasts of forest carbon dynamics, Ecol. Appl., 21(5), 1490–1505, https://doi.org/10.1890/09-1394.1, 2011.
Citation: https://doi.org/10.5194/egusphere-2026-3014-RC2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 248 | 146 | 55 | 449 | 47 | 48 |
- HTML: 248
- PDF: 146
- XML: 55
- Total: 449
- BibTeX: 47
- EndNote: 48
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1