Observational benchmarks inform representation of soil organic carbon dynamics in land surface models
Abstract. Representing soil organic carbon (SOC) dynamics in Earth system models (ESMs) is a key source of uncertainty in predicting carbon climate feedbacks. Machine learning models can help identify dominant environmental controllers and their functional relationships with SOC stocks. The resulting knowledge can be implemented in ESMs to reduce uncertainty and better predict SOC dynamics over space and time. In this study, we used a large number of SOC field observations (n = 54,000), geospatial datasets of environmental factors (n = 46), and two machine learning approaches (Random Forest (RF) and Generalized Additive Modeling (GAM)) to: (1) identify dominant environmental controllers of global and biome-specific SOC stocks, (2) derive functional relationships between environmental controllers and SOC stocks, and (3) compare the identified environmental controllers and predictive relationships with those in Coupled Model Intercomparison Project phase six (CMIP6) models. Our results showed that diurnal temperature, drought index, cation exchange capacity, and precipitation were important observed environmental controllers of SOC stocks. RF model predictions of global-scale SOC stocks were relatively accurate (R2 = 0.61, RMSE = 0.46 kg m−2). In contrast, precipitation, temperature, and net primary productivity explained > 96 % of ESM-modeled SOC stock variability. We also found very different functional relationships between environmental factors and SOC stocks in observations and ESMs. SOC predictions in ESMs may be improved significantly by including additional environmental controls (e.g., cation exchange capacity) and representing the functional relationships of environmental controllers consistent with observations.
Kamal Nyaupane et al.
Status: open (until 23 Jun 2023)
- RC1: 'Comment on bg-2023-50', Lorenzo Menichetti, 08 May 2023 reply
Kamal Nyaupane et al.
Kamal Nyaupane et al.
Viewed (geographical distribution)
The study is interesting because of the general perspective it gives, but it contains also quite some self-evident remarks and some related misunderstandings. It also lacks context, since most of the advancements in modeling in the last 50 years (starting from the temperatures and soil water content relationships with microbial activity) are not considered but you are still making recommendations to modelers.
In order to make your statements, you need first to study and briefly review (connecting your results to them) how these processes are represented in ESM, to be honest it sometimes seems you have just a vague idea.
The main of these remarks are all the discussions about introducing in models any variable related to soil moisture or temperature. These relationships are very well known, we know that those are the main factors influencing SOC decomposition and models consider this already quite well.
Using drought instead of precipitations is no improvement. Most SOC models (or maybe all) consider soil moisture simulating it based on precipitations and evapotranspiration (with more or less refined water balance), and have relatively refined response functions of decomposition responding to soil moisture. When the soil moisture falls, the microbial activity in the models (often represented by the kinetics) reduces. This is the main impact of drought on SOC. Most parametric decomposition models go further, representing also a decrease in activity towards the end of the curve, when soil gets close to saturation (this time due to lack of oxygen). You could for example start from the review by Moyano et al., 2013 to have an overview of the discussion about moisture. Concerning temperature effects, you could check first the Lloyd and Taylor 1993, a good citaton classic. But even if temperature is between the two the easier bit, there’s still discussion going on (for example Ratkowski). And here we are still just considering first order kinetic models, representing these interactions as external forcing variables as a scaling of the kinetics. There are much more complex models that represent these effects internally to the model itself, for example representing the effect of soil texture on moisture or explicitly considering diffusion of nutrients.
If drought works better in your model than precipitation the main reason (given that you are using models with potentially "infinite degrees of freedom", at least I personally define ML models like that when I use them, being quiote illiterate on the topic) is probably that you are not including evapotranspiration in your model (while most/all ESM will probably due, calculating the soil water balance) while drought contains, in some sense, information about that too.
Also your speculation about the causes of why drought was so important in your model (like 262-264) are not so convincing. For sure it might be that there is also an effect on inputs, but those should be considered in your model(s) already by using NPP. While it is very well known (and how much, numerically) that soil moisture affects microbial activity.
Summarizing, you are probably asking the wrong questions to your model(s). Saying that temperature related and moisture related indicators (whatever those are, since your model has ”infinite” degrees of freedom), is extremely self evident since half a century. While asking to your model how much would it matter to include also some edaphic parameter, and which one, one a global scale for predictions, that would be an interesting question to read about.
Line 293-294: you don’t need this kind of study to demonstrate different controls of moisture for different ESM. You can simply read which functions they rely upon, if those functions are different (they are) then the controls will be different. I think you should study the main functions available for that, and which function has been implemented in which model.
Your conclusions seem off track. Line 310-311: I would say that there’s no disagreement, all SOC models are using temperature and moisture of the microbial environment to control decomposition, those processes are well taken care of (better than using drought alone). Different models will of course rely on different variables to represent the same processes, the fact yours relies on diurnal temperature instead of soil temperature or daily temperature does not allow you to make inferences on other models, it depends on the functions they use. But they all agree that we need to represent the impact that the water present in the microbial micro-environment has on kinetics, and the effect that temperature in such micro-environment has also on the kinetics.
So, concluding, the study could have some potential but it requires a much better and extensive work on documenting the state of the art in detail. Understanding how the problems you talk about are already dealt with in models (and I mean at the level of the single functions) will also help you to repurpose your conclusions.
I also suggest you to shift your focus a bit, you are probably having a bit too ambitious goals (of making a big impact on modeling). Your approach is interesting for me (I am a modeler myself) because it offers insights on processes that ok, we know well in principle, but still they vary in different environments, there might be interactions with einvironmental factors changing the relationships, and so on. The global perspective of your study is interesting already, even in case you won't revolutionize anything.
I have minor (but still important) concerns about validation too. How did you trained your RF models? Can you ensure that the validation is completely independent? For example if you used Caret to train the metaparameters, you might have a spillover of the training in validation (because you select the metaparameters with the crossvalidation results). Another big issue would be to ensure that the data points of each fold of the crossvalidation are not correlated with any data point in the training. For example if you have more propfile from one single site, some would end in validation some in training (for each fold), injecting information from validation into training. If you are selecting instead at the site level (or if it corresponds to the data point level) it’s all fine.
Concerning the GAM, how did you validate them? You say you used Rˆ2, but based on which dataset did you calculate it?
In general, please be extremely specific about your validation approaches, in particular discussing why you believe there is no spillover of information between training and validation and why the two are supposed independent.
You also need to describe better the study on the ESM data. What are the SOC data you mention on line 156? Are those simulated data or measured?
Moyano, Fernando E., Stefano Manzoni, and Claire Chenu. “Responses of Soil Heterotrophic Respiration to Moisture Availability: An Exploration of Processes and Models.” Soil Biology and Biochemistry 59 (April 2013): 72–85. https://doi.org/10.1016/j.soilbio.2013.01.002.
Lloyd, J., and J. A. Taylor. “On the Temperature Dependence of Soil Respiration.” Functional Ecology 8, no. 3 (June 1994): 315. https://doi.org/10.2307/2389824.