High organic inputs explain shallow and deep SOC storage in a long-term agroforestry system &ndash; combining experimental and modeling approaches

Cardinael, Rémi; Guenet, Bertrand; Chevallier, Tiphaine; Dupraz, Christian; Cozzi, Thomas; Chenu, Claire

doi:https://doi.org/10.5194/bg-15-297-2018

Articles | Volume 15, issue 1

https://doi.org/10.5194/bg-15-297-2018

© Author(s) 2018. This work is distributed under
the Creative Commons Attribution 3.0 License.

https://doi.org/10.5194/bg-15-297-2018

© Author(s) 2018. This work is distributed under
the Creative Commons Attribution 3.0 License.

Articles | Volume 15, issue 1

Research article

|

15 Jan 2018

Research article |

| 15 Jan 2018

High organic inputs explain shallow and deep SOC storage in a long-term agroforestry system – combining experimental and modeling approaches

Rémi Cardinael, Bertrand Guenet, Tiphaine Chevallier, Christian Dupraz, Thomas Cozzi, and Claire Chenu

Download

Final revised paper (published on 15 Jan 2018)
Supplement to the final revised paper
Preprint (discussion started on 12 Apr 2017)
Supplement to the preprint

Interactive discussion

Status: closed

AC: Author comment | RC: Referee comment | SC: Short comment | EC: Editor comment

- Printer-friendly version

- Supplement

RC1: 'Highlight the successful model validation', Thomas Wutzler, 19 Apr 2017
- AC1: 'Reply to RC1', Rémi Cardinael, 25 Jun 2017
RC2: 'review', Anonymous Referee #2, 28 May 2017
- AC2: 'Reply to RC2', Rémi Cardinael, 25 Jun 2017

Peer-review completion

AR: Author's response | RR: Referee report | ED: Editor decision

ED: Reconsider after major revisions (29 Jun 2017) by Andreas Ibrom

AR by Rémi Cardinael on behalf of the Authors (30 Jun 2017) Author's response Manuscript

ED: Referee Nomination & Report Request started (03 Jul 2017) by Andreas Ibrom

RR by Anonymous Referee #2 (18 Jul 2017)

RR by Thomas Wutzler (22 Jul 2017)

Suggestions for revision or reasons for rejection

The presentation of the manuscript improved compared to the previous version, but not enough. Some required additional analysis has not been done. With the better representation several issues became more clear that need to be tackled. The manuscript should only be published with more serious work on reacting on the review comments.

The study reports changes across the entire soil profile. However, changes seem to be most important in the top 20cm and in the tree alleys (See strange legend in Fig 4 bottom with a change in magnitude for one level). When reading only about the aggregated results, it implied to me different conclusions than reading the full paper in detail. The authors did some small notes in reacting on review requests for these details but did very little adjustments in presenting the overall results. I suggest reporting and discussing results independently in a) tree-alleys, b) top 20cm in rows c) below 20cm d) aggregated space.

The study claims performing a Bayesian analysis because it applies prior information on model parameters. However, exact prior information is still not presented (e.g. diagonal values of P_b in eq. 34) let alone motivated sufficiently. Moreover prediction uncertainty is not presented and reasons for not doing that are very weak (see detailed comments). Uncertainties and full probability distributions are essential components of a Bayesian analysis. I again require to present and motivate the priors and to discuss predictive uncertainties. I still recommend using log-transformed parameters (and hence log-normal prior information) where prior parameter ranges span several orders of magnitude.

The presentation of the numerical solution is now sufficient to be understood. There is still works to do on notation (see detailed comments). Why did the authors not check that the resolution of the spatial grid is sufficient? Why should the spatial modeling grid be regular? In the used Soetart's ode.1d code there is no such constraint. With looking at the results, I again strongly recommend repeating the analysis with a finer vertical resolution at the top 30cm.

I can see the point in the current quantifying the priming effect with transferring the decomposition of the priming model to the agroforestry plot with a parameter fixed to conditions of the control (L 627ff). However, I do not agree in the interpretation. To me it is not an absence of priming but is the priming effect as it worked in the control. I still have difficulties in accepting the approach. For me it does not make sense to apply the priming-explicit model with parameter fixed to the control as a no-priming base scenario. Also the results of this run with stock increases of 60t/ha in 18 years are not reasonable (Fig. 6). I am still curious on the comparison of the current quantification to the (to my opinion more straightforward and without much work obtainable) model comparisons I suggested in the previous review involving comparison of priming-explicit vs. no-priming model variants. Therefore, I cannot follow the numbers in abstract of priming effects reducing potential SOC storage by 75 to 90%. (In addition one should differentiate between tree rows/alleys and different depth)

Detailed comments

Table 5 caption: “the prior values that minimized J(x)”. Why is the prior information different for different model variants? I hope that you did not try/optimize for different prior information means. Prior information is defined as an information that is independent of the observations and the calibration. Please add an additional column with mean and standard deviation of the prior parameter pdf.

Fig. 4. I is still interesting to compare depth profiles across control, tree row, and alley. I agree that the closeness of the values makes a combined plot hard to read. But it is even harder to compare across facets. Please, make an attempt to display at least the data or one variant of the prediction in one plot.

Fig 5. First I though the [10,11] entry in legend of the bottom subplot was an error of omitting the decimal point, but the change of magnitude and the non-continuous scale are deliberate. Please, make this really obvious and discuss.

Eq 12 (and other eq.): Notations (here d for \partial) dFoc/dt and dFoc_t,z,d/dt are ambiguous. Eq. 14 for dFoc/dt involves terms specific to t,z,d. I suggest renaming all dFoc/dt to dFoc_t,z,d/dt and renaming all previous dFoc_t,z,d/dt to dec_Foc_t,z,d as it only describes decomposition/mineralization. Similar for HSOC.

L 649: Suggest modifying “importance of of priming” to “importance of additional priming due to tree row vegetation and trees”

L 654: I do not agree with “correspond to the absence of priming due to trees”. To my opinion it corresponds to “same magnitude of the priming effect as in the control”. See may general comments on my difficulties with the current priming effect quantification.

L 567: Please write more specifically what is “the associated uncertainty”. In the previous sentence the known is “These parameters”. The 93 soil samples estimate the uncertainty of observed C stocks. How did you use the uncertainty of the observed stocks in the optimization?

L 605: Manzoni et al 2012 is not a good citation for BIC. Please go back to the original literature cited therein.

Eq. 21: I still do not understand the reasons for using an unusual variant of BIC when you defined your Likelihood by the first term in eq. 21. In effect the current BIC variant uses some different Likelihood that ignores observation uncertainty used in matrix R in eq. 21. It might be an adequate Likelihood for the Manzoni 2012 studied system, but I cannot see how it applies to your case. It would have been very easy to report the requested standard BIC formulation alongside your variant.

L 836: Why do the correlations hinder you to compute prediction uncertainty? They are important to consider in Bayesian analysis. The straightforward way would be an MCMC run on eq. 21 to obtain a proper sample from posterior parameter distribution. Perhaps more easy: From the curvature at the optimum you already got and report correlations of a first approximation of a multivariate normal of the parameters posterior distribution. You can draw samples from this distribution and do (say ~1000) forward model runs to compute 95% confidence interval on predictions.

L 902ff: I suggest moving the part on differences in soil temperature and soil moisture and the sensitivity study to the part where you describe what you have not done “simplifying assumptions”. Instead of strongly concluding that OC inputs drive SOC storage I recommend formulating more weakly that in this study OC inputs are sufficient to explain the differences in stocks. (You did not compare to explicitly modelling the differences in temperature and moisture with the agroforestry)

L 996: Its a result section. “results were not significantly different” ? Hence, ...

Hide

ED: Reconsider after major revisions (23 Jul 2017) by Andreas Ibrom

AR by Rémi Cardinael on behalf of the Authors (14 Nov 2017) Manuscript

ED: Referee Nomination & Report Request started (17 Nov 2017) by Andreas Ibrom

RR by Thomas Wutzler (23 Nov 2017)

Suggestions for revision or reasons for rejection

I thank the authors for taking into account my concerns.

Regarding the differences between topsoil and tree alleys, the results are now clearer. Thanks for the new Fig. 2In the replies. This shows nicely that the differences are almost entirely in topsoil and only for the tree row. (but also reinforces my concerns for numerical errors in topsoil because of the strong gradients compared to the used resolution)I think this figure is necessary and should be added to the manuscript, because it helps to clarify the results and sets the context for comparing agroforestry with control. The wording in the abstract needs to reflect this too (L 34ff).

Regarding depth resolution, Fig. 1 in reply dampens my concerns on numeric error, as it suggests some robustness. Please, add some discussion to the manuscript on this point in addition to the reply.

Regarding the priming effects, it is kind to omit the interpretation form the manuscript that I was not able to follow.

Regarding the prior information, I am still confused and have still concerns, see below.

You seem not be interested in prior and posterior parameter distributions. To my reading, rather, a prior-like term is used to regularize the cost function. Hence, I suggest to avoid the term Bayesian. The result stay valid also with a gradient based optimization.

In the reply on my qeustions regarding the piror you write about starting parameters for the gradient search instead of prior knowledge on parameters. That is why I am still confused. Starting parameters are a very different thing from priors on parameters. Its good practice to vary the starting values of a gradient search, but its conceptually wrong to randomly vary prior parameter values (aside from testing whether your prior has a significant impact on the posterior.)
Did you vary also the prior parameter values in x_b and the P-Matrix (eq. 34) during optimization? To my opinion an optimized prior would be conceptually not acceptable and I would request to redo the optimization with fixed values of x_b and P. This is true for both, a Bayesian analysis and for a regularization of the cost function. Still you need to document values of P (at least diagonals). Or you may omit the prior term in eq. 34 altogether, if you do not need a regularization. If you are not able to redo the optimization, please, clearly describe what you have done with the priors (x_b and P in eq. 34), maybe in an appendix. Although, varying priors is conceptually wrong, you may be confident that it does not have a big effect on results and does not impair conclusions.

Please, be precise when speaking about prior parameter values in the cost function (x_b) versus starting values for the optimization.

Tab 5: If the optimization is done at original (not log) scale, I suggest to also present the results at original scale.

L34: The wording suggests that the stocks in the agroforestry plot are increased across all the area and all the depth. While this is true for a total comparison, I strongly suggest rewording so that the facts from new Fig 2 (mainly topsoil, mainly tree alleys) become clearer.

Hide

RR by Anonymous Referee #2 (27 Nov 2017)

ED: Publish subject to minor revisions (review by editor) (27 Nov 2017) by Andreas Ibrom

AR by Rémi Cardinael on behalf of the Authors (01 Dec 2017) Author's response Manuscript

ED: Publish as is (04 Dec 2017) by Andreas Ibrom

AR by Rémi Cardinael on behalf of the Authors (04 Dec 2017) Author's response Manuscript

Download

Article (4832 KB)
Full-text XML

Short summary

The introduction of trees in an agricultural field modifies organic matter (OM) inputs to the soil (litterfall, root litter), the microclimate, and the stabilization and decomposition processes of OM. These changes could affect soil organic carbon (SOC) storage, but the importance of each process is not well known. In a long-term agroforestry trial, we showed that SOC storage could be explained by high OM inputs to the soil but that enhanced decomposition could also have reduced this potential.