Evaluation and optimisation of the soil carbon turnover routine in the MONICA model (version 3.3.1)

Aiteew, Konstantin; Rouhiainen, Jarno; Nendel, Claas; Dechow, Rene

doi:https://doi.org/10.5194/egusphere-2023-760

Preprints

https://doi.org/10.5194/egusphere-2023-760

Preprints

06 Jun 2023

| 06 Jun 2023

Status: this preprint is open for discussion.

Evaluation and optimisation of the soil carbon turnover routine in the MONICA model (version 3.3.1)

Konstantin Aiteew, Jarno Rouhiainen, Claas Nendel, and Rene Dechow

Abstract. Simulation models are tools commonly used to predict changes in soil carbon stocks. Prior validation is essential, however, for determining the reliability and applicability of model results. In this study, the process-based biogeochemical model MONICA (Model of Nitrogen and Carbon dynamics on Agro-ecosystems) was evaluated with respect to soil organic carbon (SOC) using long-term monitoring data from 46 German agricultural sites. A revision and parameterisation of equations, encompassing crop and fertiliser-specific C contents and the abiotic factors of soil temperature, soil water and clay content, were undertaken and included in the model. The modified version was also used for a Morris elementary effects screening method, which confirmed the importance of environmental and management factors to the model’s performance. The model was then calibrated by means of Bayesian inference using the Metropolis-Hastings algorithm. The performance of the MONICA model was compared with that of five established carbon turnover models (CCB, CENTURY, C-TOOL, ICBM and RothC). The original MONICA model systematically overestimated SOC decomposition rates and produced on average a ~17 % greater mean absolute error (MAE) than the other models. The modification and calibration significantly improved its performance, reducing the MAE by ~30 %. Consequently, MONICA outperformed CENTURY, CCB and C-TOOL, and produced results comparable with ICBM and RothC. Use of the modified model allowed mostly adequate reproduction of site-specific SOC stocks, while the availability of a nitrogen, plant growth and water submodel enhanced its applicability compared with models that only describe carbon dynamics.

Received: 17 Apr 2023 – Discussion started: 06 Jun 2023

Download & links

Konstantin Aiteew et al.

Status: open (until 20 Oct 2023)

Post a comment Subscribe to comment alert

CEC1:
'Comment on egusphere-2023-760', Juan Antonio Añel, 31 Jul 2023 reply

Dear authors,

Unfortunately, after checking your manuscript, it has come to our attention that it does not comply with our "Code and Data Policy".

https://www.geoscientific-model-development.net/policies/code_and_data_policy.html
Firstly, you have archived the MONICA code on GitHub. However, GitHub is not a suitable repository for scientific publication. GitHub itself instructs authors to use other alternatives for long-term archival and publishing, such as Zenodo. Therefore, please, publish your code in one of the appropriate repositories, and reply to this comment with the relevant information (link and DOI) as soon as possible, as it should be available for the Discussions stage.
Please, note that you should not include in your manuscript a "Software availability" section. The section for the information that you include there is the "Code availability" section.
Secondly, it is not enough that you state in the "Data availability" section that it is possible to obtain the input data from the "State Office for Mining, Energy and Geology of the Lower Saxony state government". This information is hardly useful for anyone that wants to get access to the data. You should include full details about the webpage, email addresses, persons or processes that anyone needs to follow to obtain them. Ideally, you should store a copy of the data that you have used in a private Zenodo repository. In this way, the dataset continues to be private, and we are sure that it is properly stored, and you would have a DOI to include in your manuscript and cite.
Also, you say, "Additional data is freely available under the CREATIVE COMMONS ATTRIBUTION 4.0 INTERNATIONAL LICENSE"; however, you do not provide details on the data you refer to. You must clarify it and be sure that such data complies with the policies of our journal.
I should note that, actually, your manuscript should not have been accepted in Discussions, given this lack of compliance with our policy. Therefore, the current situation with your manuscript is irregular. In this way, if you do not fix the above-mentioned problems, we will have to reject your manuscript for publication in our journal.
Juan A. Añel
Geosci. Model Dev. Executive Editor

Reply

Citation: https://doi.org/10.5194/egusphere-2023-760-CEC1
- AC1:
  'Reply on CEC1', Konstantin Aiteew, 03 Aug 2023 reply
  
  Dear Juan A. Añel,
  thank you very much for drawing attention to the shortcomings in the manuscript.
  Since I'm not a developer of the MONICA model, I have no rights to the intellectual property of the model. My contribution consisted only of minor modifications to the code. In consultation with my colleagues, we decided to only make the entire source code available to the reviewers to avoid any legal difficulties. Therefore, as suggested I published the model code under a restricted access with Zenodo.
  https://doi.org/10.5281/zenodo.8211378
  The source code without the modifications is published by the actual developers and I used the section "Software availability" to describe the original model and the project website. In the revised manuscript I will rewrite this point and include it to the "Code availiability" section. I still recommend to visit the model website https://github.com/zalf-rpm/monica for model description.
  As suggested, I did the same with the data. The whole dataset is published under restricted access with Zenodo.
  https://doi.org/10.5281/zenodo.8211468
  I will give more information about the data used in the "Data availability" section. However, as far as I know, obtaining of the long-term monitoring data from Lower-Saxony, Germany is only possible if you work for the Federal Republic of Germany.
  Data availability
  The described long-term soil monitoring data is available from the State Office for Mining, Energy and Geology of the Lower Saxony state government. Restrictions apply to the availability of these data, which were used under license for this study. Data are available from the State Office for Mining, Energy and Geology of the Lower Saxony state government from the website: https://www.lbeg.niedersachsen.de/boden_grundwasser/bodenmonitoring/bodendauerbeobachtung/das-boden-dauerbeobachtungsprogramm-von-niedersachsen-572.html and mail: bodenkundlicheberatung@lbeg.niedersachsen.de Additional data for evaluating soil temperatures and moisture contents is available from the Thünen Institute of Climate-Smart Agriculture from the website: https://www.thuenen.de/en/institutes/climate-smart-agriculture and mail: ak@thuenen.de, from the University of Hohenheim from the website: https://spurengas.uni-hohenheim.de/en/108520 and mail: ruser@uni-hohenheim.de.
  Konstantin Aiteew
  Thünen-Institut for Climate-Smart Agriculture
  
  Bundesallee 65
  
  38116 Braunschweig (Germany)
  
  Reply
  
  Citation: https://doi.org/10.5194/egusphere-2023-760-AC1
  - CEC2: 'Reply on AC1', Juan Antonio Añel, 03 Aug 2023 reply
    
    Dear authors.
    MONICA, as stated in its GitHub repository, is published under the MPL-2.0 license. This means that you have the right to copy and redistribute the code. Therefore, the ZENODO repository must be public and include both the original source code and the modified or added code you have contributed.
    
    We can understand the problems with the data from the Lower Saxony state government. However, I recommend you contact them and ask to release the dataset open to everyone.
    
    However, we can not accept the restrictions for the data from the Thünen Institute and the University of Hohenheim. The first one is your institution, so what precludes you from publishing the data is unclear. The second one is a university; again, it is unclear what legal reason forbids a university from publishing data.
    
    Therefore, your manuscript continues to be unsuitable to be considered for publication in our journal. Please solve the issues with the code and release the data or provide evidence about what legal reason or regulation forbids you to do it.
    Regards,
    Juan A. Añel
    Geosci. Model Dev. Executive Editor
    
    Reply
    
    Citation: https://doi.org/10.5194/egusphere-2023-760-CEC2
    
    AC3: 'Reply on CEC2', Konstantin Aiteew, 04 Aug 2023 reply
    
    Dear Juan A. Añel,
    We will contact all creators/owners of the data and try to find an agreement to publish the data. However, this could take a while, since various stakeholders are involved and we are currently in the holiday season.
    With kind regards,
    Konstantin Aiteew
    
    Reply
    
    Citation: https://doi.org/10.5194/egusphere-2023-760-AC3
RC1:
'Comment on egusphere-2023-760', Lorenzo Menichetti, 01 Aug 2023 reply

General comments:

The study is quite robust and well-executed. I like the use of uniform priors and the validation approach (not considering cross-validation but a fully isolated validation set).

Statistical methods are well described concisely and clearly, and applied in a rational and consequential way.

In general, this is a rather good study of a potentially quite useful model, quite balanced in terms of complexity/capabilities (and I might, for example, want to use it for perennials, where feedbacks are more important).
Your priors are thin, though, even if uniform. A broader prior choice would probably yield different results, particularly regarding the model biases you found (like overestimating water content, for example) and model performance results. Your prior choice influences the results quite heavily. Do you have anything to add in this respect?

Compared to the simpler models, I expect better absolute performances (considering the best sets) of MONICA with broader priors but a lot higher indetermination and uncertainty in predictions.

You could introduce or discuss the reasons for working with such a small range of priors (for example, feasibility ranges of parameters due to physical constraints). A more detailed description of prior assumptions and the reasons behind each choice, since they greatly influence the results, could improve the MS.
I also do not understand how the discussion initiated in the introduction about the possible advantage of MONICA in representing fertility feedback mechanisms is developed in the rest of the MS. That part should be connected better with the results, discussion, and method description. At the same time, these feedbacks are one of the most exciting characteristics of a more complex model compared to decomposition models and one of its biggest potential issues when it comes to model determination and predictive uncertainty. And while reading the introduction, I was expecting this topic to be developed more in-depth.

Specific comments
Introduction

Nothing wrong here, but as a suggestion, you could develop more as a central question the potential increase in indetermination of a more complex model (with interactions, which seem nonlinear) like this one compared to simple decomposition models. Adding nonlinearities to consider feedback (if I understand the model correctly) is the main potential issue and advantage of going more complex on top of data requirements.
Line 155-160: so you are using linear allometric ratios with no intercept, right? This could deserve some discussion if you like (the approach works well within certain limits, which are likely including your test sites anyway since you don't seem to have anything extreme)
Line 232: define better the squared range between 1% and 5%. Does this range vary with different parameters?

Reply

Citation: https://doi.org/10.5194/egusphere-2023-760-RC1
- AC2:
  'Reply on RC1', Konstantin Aiteew, 04 Aug 2023 reply
  
  Dear Lorenzo Menichetti,
  thank you very much for the assessment and the suggestions for improvement of the manuscript.
  We selected the prior values based on the observed upper and lower bounds of each parameter to ensure that each calibrated value is within the realistic, observed range. It is highly probable, that with a different prior probability distribution, deviating values could be achieved. However, it was particularly important to us that the parameter values were based on realistic and observed values. Based on that, we could go more into detail explaining the priors in the manuscript. Aside from that, when considering the performance of the model, we would not expect that the calibration could significantly improve the model performance. For comparison, we also applied a simple hill-climbing algorithm to estimate optimal parameter values and didn’t achieve a better performance. We would also argue that complex biogeochemical models tend to perform weaker compared to simple C-turnover models and we were surprised that the MONICA model performed comparably to the simple models.
  If considering the overestimation in the soil water submodel, we didn’t evaluate the feedback between plant growth and soil water contents, but we assume that there should be at least a minor effect of the plant growth parametrisation on soil water changes. We don’t think that the parametrisation of the C-turnover submodel had a significant effect on modelled soil water. We also used the default parametrisation of the soil water submodel, so there could be a general bias in the parametrisation. However, this part needs further evaluation and we are not sure if this should be evaluated/discussed in this manuscript.
  We also like the suggestion, to further clarify the difficulties of complex models, that they are inherently black boxes that make understanding the various factors/feedbacks difficult. Therefore, we used a sensitivity analysis, to estimate the effect of plant growth (if that is what you mean with fertility feedback mechanisms) and other factors on soil carbon changes. But we agree that this aspect could be developed further in the manuscript.
  Line 155-160: We used the allometric functions to estimate the C-Inputs from the harvested yields, on the one hand for the simple C-models and second for comparison with the modelled C-Inputs of the MONICA model.
  Line 232: This range seems quite arbitrary, but it was used in another publication and it worked quite well to achieve a random walk in the parameter space. Basically, each step for generating new candidate values was limited by 1 – 25 % of the parameter space. We assume the formulation we used was to complicated. We try to explain this aspect a bit better in the revised version.
  Konstantin Aiteew
  Thünen-Institut for Climate-Smart Agriculture
  
  Bundesallee 65
  
  38116 Braunschweig (Germany)
  
  Reply
  
  Citation: https://doi.org/10.5194/egusphere-2023-760-AC2
  - RC2: 'Reply on AC2', Lorenzo Menichetti, 07 Aug 2023 reply
    
    The discussion is getting pretty interesting for me!
    
    I am surprised you expect a more complex model to perform worse than a simpler one. Based on that, I would expect the opposite; the more complex a model is, the more freedom it has to fit the information content of your calibration data. I often borrow the bias-variance trade-off concept from machine learning, at least conceptually, and it usually suits all my models (as a concept): simpler model > better fitting my training but more risk of overfitting.
    
    I tend to work with broad priors and disregard the physical meaning of parameters (to some extent, of course, like no negative values or orders of magnitude off), so usually, my calibrations have quite some freedom.
    
    In your case, if you find that a more complex model performs worse unless it is because of specific reasons connected with the model structure (like it misses some pieces. In this case, it doesn't, compared to a compartmental model). You have broad enough priors; I would say this depends on the characteristics of the training vs. validation datasets. I imagine your more complex model is overfitting; in other terms, it ends up accommodating the information content of the training too much and loses generalization power since, for some reason, your validation has different information, like some regular bias (which, for soil, is not particularly surprising).
    
    If with an UNCONSTRAINED hill climbing algorithm, you get worse results from a more complex model; I would say that unless you just hit some local optima and the algorithm got stuck in it (for testing this, try starting from different initial values), then it should be either because of a bias between training/validation or some structural reasons.
    
    Was it an unconstrained calibration, in that case? Or did you still specify limits for the parameter ranges?
    
    A compartmental SOC decomposition model has a few possible structural reasons, particularly for models with more than 3-4 pools (usually enough to describe most SOC dynamics). One is model initialization: if, for some reason, one of your models is getting too much slower pools, this will, of course, impact predictions. In this case, though, you could find heteroskedasticity in the errors over time since this model bias would differ. Another one is different forcing functions: the main difference between first-order compartmental SOC decomposition models is in the reduction functions, the rescaling of the kinetics with temperature or moisture or some other edaphic property. For the rest, they are all very similar. It could be an interesting discussion to go deeper into this topic, checking (provided that the models are initialized in the same way) possible differences in the model response functions. Maybe one of your models has better climate or edaphic scaling than others, which would explain fitness differences.
    
    With this, I do not mean to criticize your manuscript, which is already at a very good level of detail. These kinds of discussions are usually not well developed in the literature, according to me, and I know these questions are some kind of personal interest from my side, not extremely common among environmental modelers. So I am not going forward in it to say your MS shouldn't be published to the next step, but I am indeed interested in the discussion.
    
    Reply
    
    Citation: https://doi.org/10.5194/egusphere-2023-760-RC2
    
    AC4: 'Reply on RC2', Konstantin Aiteew, 18 Aug 2023 reply
    
    Dear Lorenzo Menichetti,
    thank you for your thoughts about model fitting and the interesting discussion.
    The difference between MONICA and RothC (and other simple SOC models used here) is that MONICA quantifies the carbon input from crops by crop growth modelling (which is also affected by simulated nitrogen cycling, soil water balance and other processes) while for RothC we used observed yields to estimate carbon inputs from crop residues.
    Because measured carbon inputs from crop residues were not available, we used observed yields to get an indication of the ability of MONICA to simulate crop development and finally carbon inputs. Although we were able to adequately replicate yields (just as a proxy of modelling carbon inputs from crop residues) some deviances between model and observed yields remain, so that the description of variability between sites is not perfect. One reason here is that some processes are not described by the model (the effect of phosphorus and other nutrients or the effect of crop diseases).
    Contrary, the carbon input estimation method used to derive carbon inputs for RothC. Here we used directly observed crop yields and allometric functions, evaluated for Germany in previous studies (Dechow et al. 2019, Riggers et al 2019).
    Like you mentioned, there is not much difference between decomposition approaches of RothC and MONICA. Therefore, no big differences between models would have been expected. However, we assume that a less good description of carbon input variability between sites by MONICA cannot be compensated by more degrees of freedom during the calibration process, because the carbon input estimation of RothC is based on measured yields, that should correlate with carbon inputs (Bolinder et al. 2007) and therefore better describe the variability of carbon inputs between sites. We assume, that next steps in improving the description of soil organic carbon development, should address crop growth modelling and carbon input estimation by dynamic models. Considering the several dozen parameters already associated with crops in MONICA, this is quite challenging. So, in the end the performance is limited by the working hours that can be invested in the model and the data availability. A simple model has the advantage that it takes significantly less time to fully exploit the potential of the model.
    Based on that, it is probable that with more data and time invested in the crop growth model in MONICA a better performance could be achieved in the SOM decomposition submodel.
    We agree, beside carbon input estimation methods, model initialization and response function might be another reason for different model behavior. In the case of soil decomposition, models come from different regions in the world representing different training data sets, which will affect their sensitivities in other regions of the world. Our data set of long-term observational sites is probably not suitable to analyze the effects of model initialsiation or environmental conditions in more detail. This needs data sets with well-defined conditions (indications of initial pool distributions, carbon inputs, root exudates). Here it would be interesting to force carbon inputs in MONICA by observed (estimated) carbon inputs to make decomposition approaches comparable.
    Currently, MONICA uses a standard distribution for the pools, which can be considered as a weak point of the SOC submodel. Compared with RothC, which can be run in equilibrium mode or short-term mode (with pool distribution based on radiocarbon values). It would be great if it was possible to adjust the pools in MONICA, e.g. based on radiocarbon measurements or land-use history.
    Regarding the hill climbing algorithm. We used a constrained hill climbing algorithm, were we defined specific parameter ranges. We started the algorithm several times with different initial values. It came handy as a relatively simple calibration method that allowed us to compare the performance of the Bayesian calibration.
    Thank you very much and with kind regards,
    the authors
    
    Bolinder, M. A., Janzen, H. H., Gregorich, E. G., Angers, D. A., and VandenBygaart, A. J.: An approach for estimating net primary productivity and annual carbon inputs to soil for common agricultural crops in Canada, Agriculture, Ecosystems & Environment, 118, 29-42, https://doi.org/10.1016/j.agee.2006.05.013, 2007.
    Dechow, R., Franko, U., Kätterer, T., and Kolbe, H.: Evaluation of the RothC model as a prognostic tool for the prediction of SOC trends in response to management practices on arable land, Geoderma, 337, 463-478, https://doi.org/10.1016/j.geoderma.2018.10.001, 2019.
    Riggers, C., Poeplau, C., Don, A., Frühauf, C., and Dechow, R.: How much carbon input is required to preserve or increase projected soil organic carbon stocks in German croplands under climate change?, Plant and Soil, 460, 417-433, 10.1007/s11104-020-04806-8, 2021.
    
    Reply
    
    Citation: https://doi.org/10.5194/egusphere-2023-760-AC4

Konstantin Aiteew et al.

Viewed

Total article views: 346 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
231	93	22	346	4	6

HTML: 231
PDF: 93
XML: 22
Total: 346
BibTeX: 4
EndNote: 6

Views and downloads (calculated since 06 Jun 2023)

Month	HTML	PDF	XML	Total
Jun 2023	81	32	4	117
Jul 2023	42	15	3	60
Aug 2023	91	37	14	142
Sep 2023	17	9	1	27

Cumulative views and downloads (calculated since 06 Jun 2023)

Month	HTML	PDF	XML	Total
Jun 2023	81	32	4	117
Jul 2023	42	15	3	60
Aug 2023	91	37	14	142
Sep 2023	17	9	1	27

Viewed (geographical distribution)

Total article views: 341 (including HTML, PDF, and XML) Thereof 341 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 22 Sep 2023

Short summary

This study evaluated the biogeochemical model MONICA and its performance in simulating soil organic carbon changes. MONICA can reproduce plant growth, carbon and nitrogen dynamics, soil water and temperature. The model results were compared with five established carbon turnover models. With the exception of certain sites, adequate reproduction of SOC stock change rates was achieved. The MONICA model was capable of performing similarly or even better than the other models.


Total:	0
HTML:	0
PDF:	0
XML:	0