MEDFATE 2.8.1: A trait-enabled model to simulate Mediterranean forest function and dynamics at regional scales
Abstract. Regional-level applications of dynamic vegetation models are challenging because they need to accommodate the variation in plant functional diversity, which requires moving away from broadly-defined functional types. Different approaches have been adopted in the last years to incorporate a trait-based perspective into modeling exercises. A common parametrization strategy involves using trait data to represent functional variation between individuals while discard taxonomic identity, but this strategy ignores the phylogenetic signal of trait variation and cannot be employed when predictions for specific taxa are needed, as in applications to inform forest management planning. An alternative strategy involves adapting the taxonomic resolution of model entities to that of the data source employed for large-scale initialization and estimating functional parameters from available plant trait databases while adopting alternative solutions for missing data and non-observable parameters. Here we report the advantages and limitations of this second strategy according to our experience in the development of MEDFATE (v. 2.8.1), a novel cohort-based and trait-enabled model of forest dynamics, for its application over a region in the Western Mediterranean Basin. First, 217 taxonomic entities were defined according to woody species codes of the Spanish National Forest Inventory. While forest inventory data were used to obtain some empirical parameter estimates, a large proportion of physiological, morphological, and anatomical parameters were mapped to measured plant traits, with estimates extracted from multiple databases and averaged at the required taxonomic level. Estimates for non-observable key parameters were obtained using meta-modeling and calibration exercises. Missing values were filled using imputation procedures based on trait coordination, taxonomic averages or both. The model properly simulated observed historical basal area changes, with a performance similar to an empirical model trained for the same region. While strong efforts are still required to parameterize trait-enabled models for multiple taxa, estimation procedures can be progressively refined, transferred to other regions or models and iterated following data source changes by employing automated workflows. We advocate for the adoption of trait-enabled population-structured models for regional-level projections of forest function and dynamics.
Miquel De Cáceres et al.
Status: final response (author comments only)
- RC1: 'Review of gmd-2022-243', Harald Bugmann, 17 Dec 2022
- RC2: 'Comment on gmd-2022-243', Nikolaos Fyllas, 13 Jan 2023
- AC1: 'Comment on gmd-2022-243', Miquel de Cáceres, 14 Mar 2023
Miquel De Cáceres et al.
Miquel De Cáceres et al.
Viewed (geographical distribution)
This manuscript tackles the considerable challenge of parameterizing a highly detailed, process-based model of forest dynamics for 217 taxa towards an application at the regional scale (ca. 30'000 km2; Catalonia, Spain). It is a methodologically interesting and valuable contribution that I think should be published, notwithstanding my concerns detailed in the paragraphs below.
First, the manuscript tries to make the point that highly detailed models that distinguish a very large number of very detailed ecophysiological processes are useful (or perhaps even needed?) for faithful assessments of climate change impacts on forest dynamics. I am hesitant to accept this. The description of MEDFATE in this paper shows that it is based on multiple assumptions that are highly uncertain. Hence, different assumptions (i.e., different formulations) could have been chosen, leading on the one hand to different parameter requirements and on the other hand to different model behavior at least at some 'intermediate' level (i.e., between the ecophysiological processes and what is ultimately shown in Fig. 3, i.e. ecosystem-level NPP and wood volume). Thus, we cannot really tell whether the model is getting the right answer for the right reasons, or whether errors are compensating each other internally, thus actually enhancing the uncertainty in future projections rather than decreasing them. In this context, it is somewhat marring that the simpler variant of the model is generally closer to the measurements in what the authors call model "evaluation" (although the data are not independent of those used for the parameterization; more on that under the "Specific comments" below). I would have appreciated if the authors had used both model variants for the future projections - they mention computational reasons for not doing so, but even a 1728 h (= 72 days) simulation effort might be undertaken as a "production run" for such a paper. It would be quite instructive to see whether the two variants agree under future conditions.
Second, while I admire the effort and creativity employed by the authors to come up with parameter values (a total of ca. 25'000 parameter values had to be estimated!), I am concerned about the vast number of parameters that had to be estimated by imputation based on partly shaky assumptions (cf. below). And using a default value across all taxa is a somewhat desperate assumption to begin with. Such things induce a lot of uncertainty in the model on top of those mentioned above re. process formulations. So how much signal is actually gained by using a highly resolved approach when the projections are blurred by uncertainties in both process formulations and parameter estimates? It would appear to me (but this is not coming as a surprise) that going to this level of detail is not appropriate, and simpler models (such as the IPM...) should be used.
Third, I think for the manuscript to be more convincing, the authors would need to do a better job re. model validation, particularly regarding the "intermediate" levels in the model, to show that it does capture ecosystem structure and function and their dynamics reasonably well. It is only then that we could "trust" the regional-scale projections. In the current stetup, the manuscript does not provide any model validation in the strict sense, which I think is unfortunate.
Specific comments (referred to by line number)
65: It would be relevant to expand a little bit on what "sufficient" actually is - in the present sentence, this is a claim whose scope is hard to assess. It appears that the starting point for the authors is that a detailed treatment of the energy, water and carbon balance is a pre-requisite for any vegetation model, and demography is "nice to have". One could view things exactly the other way around when focusing on the regional scale (cf. the statements on l. 50-51, and the application of the model to project wood volume).
72-74: The nature of "trait-enabled" should be phrased in a more comprehensible manner for readers who are not familiar with the concept. I would write something like "...are 'trait-enabled', in the sense that their parameters can be linked quantiatively to easily measurable plant traits for which large-scale data bases exist." - and if you should disagree with such formulation, you would need to explain even better because I wouldn't have understood either...
87: I think it is not quite fair to state that "all these approaches ignore taxonomic information". I think all of them do consider them, although in widely different form. Your approach differs from these other approaches by being much more consistent in linking species parameters to plant traits at different levels of taxonomic resolution, rather than to impose one level of resolution across the model (as e.g. the 'classical' DGVMs with their PFTs are doing).
90-91: "trait coordination" and later also "trait syndromes" or "functional syndromes": I presume your are referring to "trade-offs" between traits here? If so, pls use this more common term. And if you mean something else, pls explain better.
279-290: The assumptions about the causes of tree mortality (starvation and desiccation) are a bit shaky. Mortality formulations are highly likely to be pivotal for model behavior, and hence they require careful thought and parameterization. I am not convinced that the assumption of "30% of maximum" for both processes is appropriate. Furthermore, I was surprised to read that turgor loss (to 30% of max symplastic water content) is a key cause of mortality in MEDFATE - to the best of my knowledge, wilting is easily reversible, but plant hydraulic failure is a syndrome that is more difficult to overcome (although I disagree with the authors' interpretation of the Choat et al. 2018 paper in that regard, but this is not the major point here). So why should I trust these formulations and their parameterization?
315: The authors later (in the Discussion) acknowledge the problems with using the SoilGrids database. We have tried multiple times to use that database for estimating soil properties for simulation studies, and have always given up because of truly strange results, taking resort to other sources of data. Wouldn't there be a better data product for Catalonia?
338-341: So the SNFI data were used for model "tuning". This means that what is shown in section 5.1 is not a "model evaluation" in the sense of a validation. These results have been forced, which strongly reduces their usefulness for demonstrating the skill of the model. I think alternative model validation data would be needed. I am aware that this may be everything but easy. However, this is a major weakness of the manuscript.
362: I wouldn't use this example, as RGGS is not explained anywhere in the main paper. At the same time, however, I appreciated this very example because it shows how uncertain the parameterization procedures used here actually are: shade tolerance (sensu Niinemets & Valladares) is a highly aggregated concept that is pragmatically useful in some models, but probably not in highly detailed ones. Why should shade tolerance be related to these traits, and even in a linear manner? We have recently evaluated whether we can relate P88 (or P50) from the TRY database to the species' drought tolerance (according to Niinements & Valladares). The correlation is essentially non-existing, because the two concepts relate to vastly different levels of integration, and actually no relationship should be expected to begin with.
460 (Tab. 3): Why are there different entires under "Mean observed" for the basic vs. advanced model version (and again different for the IPM)? This is not explained anywhere, and the 'naïve' reader would expect the observations to be independent of the model resolution?!
578-580: This assertion is only valid if the processes are captured in the right way, which may not be the case (see general comments on structural uncertainties in the model). I suggest being more careful here.
599-600: Even if one follows this conclusion (which I do not really, for all the reasons mentioned above), one easily arrives at the question why a very complicated model like MEDFATE is actually used if the goal is indeed to "only" project regional standing timber volume. Simpler approaches would be equally suitable, would be inflicted with fewer uncertainties, and are even likely to provide more robust results. So why go with MEDFATE in this case? See also lines 635, where many more variables are mentioned that are of (potential) interest for forest management planning - but none of them are shown here, and the question arises how accurately they are projected by MEDFATE.
The manuscript is generally easy to read and understand, but it should still undergo a careful linguistic check before it is being published. I am pointing out examples of the kind of issues below, this is not a comprehensive list.
63: "impact these" -> "impact of these".
70: remove "into model parameters" (simply not needed). Then "while it is known that" -> "as".
83: "consists in" -> "consists of".
91: "within- and among-": remove dashes.
110: Not clear how one can add processes to TWO preceding models. Below, two different additions ("basic" and "advanced") are explained that relate to ONE preceding model version. Pls clarify (consider using the terms "version" or "variant" when refering to these entities, maybe this would help).
196: "conforms" -> "constitutes" ???
222: replace "and" by a comma (note that this is another documentation of high structural uncertainty in the model, cf. comments further above).
271, 272: "feedback" -> "feed back" (the verb is used here, not the noun).
272: add a comma before "and" (a full sentence = main clause is starting here).
292: remove "but" (grammatically incorrect, not needed).
303: "includes a strong" -> "includes strong".
309: remove "different" (simply not needed).
310: number of SNFI plots in inventory #2 is not given here. Why?
369: "aim make" -> "aim to make".
395: "found" -> "find".
432-433: Why not between the second and the fourth inventory? This would have provided a longer time series and thus better insights into model behavior. The shorter the 'evaluation' period (when the model is initialized with the data from the first inventory, and this is what was done here, I suppose), then the more difficult it is for any model to be wrong - because the initial data is most likely the strongest predictor of its skill, not the model-internal processes. Hence this 12-year (?) period is a really short snippet for a model test, on top of the non-independence of the data (cf. comments further above).
507: "First": there is no "second" further below. Re-consider the structure of this paragraph.
554: "account" -> "amount".
560: "intra-specific can" -> "intra-specific variation can" (?).
575: "focus" -> "foci".
596: "prediction forest" -> "prediction of forest".
599: "confidence on" -> "confidence in".
614: "this equations" -> "these equations".
615: "dessication" -> "desiccation" (also elsewhere).
623: "perform under" -> "perform well under".
650: "think MEDFATE can be" -> "think that MEDFATE is".