the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Revisiting and attributing the global controls on terrestrial ecosystem functions of climate and plant traits at FLUXNET sites with causal networks
Haiyang Shi
Geping Luo
Olaf Hellwich
Alishir Kurban
Philippe De Maeyer
Tim Van de Voorde
Abstract. Using statistical methods that do not emphasize the systematic causality to attribute climate and plant traits to control ecosystem function may produce biased perceptions. We revisit this issue using a Bayesian network (BN) capable of quantifying causality. Based on expert knowledge and climate, vegetation, and ecosystem function data from the FLUXNET flux stations, we constructed a BN containing the causal relationship of 'climate-plant trait-ecosystem function'. Based on the sensitivity analysis function of the BN, we attributed the control of climate and plant traits to ecosystem function and compared the results with those based on Random forests and correlation analysis. The main conclusions of this study include: BN can be used for the quantification of causal relationships between complex ecosystems and climatic and environmental systems, and enables the analysis of indirect effects among variables. The control of ecosystem function by climate variables (especially mean temperature and mean vapor pressure deficit) may have been underestimated previously, and the mechanism of indirect effects of climate variables on ecosystem function through plant traits should be emphasized in future studies. Further inclusion of temporal information in BN holds promise for improving the analysis of lagged effects and interactions and feedback effects between variables.
Haiyang Shi et al.
Status: final response (author comments only)
-
RC1: 'Comment on bg-2022-191', Anonymous Referee #1, 26 Oct 2022
Review: Revisiting and attributing the global controls on terrestrial ecosystem functions of climate and plant traits at FLUXNET sites with causal networks
1. General comments
The authors attempt to build causal links between plant traits, climate and ecosystem functions by constructing a Bayesian Network (BN), where links between traits and functions are based on expert knowledge, while the climatic variables are informed by the model. The authors then reevaluate the relative importance of plant traits and climate in determining ecosystem functions through a sensitivity analysis based mainly on FLUXNET data. Building on this they argue that climate indirectly affects ecosystem functions via its control on plant traits. We agree that, from an ecological perspective and considering the increasing availability of data, exploring different methods to analyze the interactions climate-vegetation involved in the ecosystem functions is a relevant and meaningful research topic. However, the paper is missing an appropriate model validation (making it difficult to evaluate the robustness of the results) and suffers from reproducibility issues as some important methodological points and choices require clarification or additional information. We start by providing some major concerns followed by minor comments in order of appearance.
2. Specific comments
- The data used in this study relies on a database collected by Magliavacca et al. (2021). In the cited paper, there are three complementary variables to quantify water use efficiency (G1, WUEt and uWUE); Rb was calculated in terms of both mean and max (95th percentile) values; and aCUE data is also available. The authors do not specify the criteria they followed when choosing which variables to include in the BN among the available variables in Magliavacca et al. (2021). Why was uWUE used and not the other water use efficiency metrics? How sensitive are the results to this choice? Why was the calculation of Rb unique in using the mean – compared to the other ecosystem function variables? If the network was based on expert knowledge, why to exclude aCUE which is explicitly included in the expert framework (Fig 1)? What would be the effect of adding CUE as an extra constrain? How sensitive is the network to its inclusion?
- In the BN (Figure 1), the causal relationships plant traits – ecosystem functions were assigned based on expert knowledge (Reichstein et al, 2014). Then, the climatic variables and the respective causal relationships were added: how were the links climate-plant traits and climate-ecosystem functions determined? Considering that BN based on expert knowledge rely heavily on the prior understanding of the processes, the approach used to assign these links should be clearly stated in the methods.
- Each plant-trait and ecosystem-function variable used in this study has a clear equivalent in the expert knowledge frame (upper panel in Figure 1) (Reichstein et al., 2014), except for AGB – which is unique in the authors model. Please provide more information about the assignment of links to this variable. How did you link Gsmax with AGB? Why not to link LAImax to AGB? AGB is confounded with wood density, height and other size metrics as plant diameter. Which variable on Reichstein’s frame is being represented with AGB? How was the confounding controlled for? Can the authors ensure no circularity was added to the framework? This is because AGB from Globbiomass is inferred from models and algorithms that use as input remote indicator variables that are correlated with many of the other variables in the authors network. Finally, AGB from Globbiomass is subject to large error – how was this error controlled for?
- More information is required regarding the specific criteria taken into account when defining the discretization thresholds. It is mentioned in the text (Lines 113-114) that the “meanings of the thresholds” were considered, but it is not completely clear how the thresholds were chosen nor what their meanings are. Ideally, BN models developed using different discretization methods should be considered and compared. If the results of such models are different, the choice of one method over the other should be justifiable (see e.g. Nojavan et al., 2017 - Comparative analysis of discretization methods in Bayesian networks).
- Not all FLUXNET stations used in Magliavacca et al. (2021) have data available for all the variables. Only 94 out of 203 sites have data regarding vegetation structure, N%, LAImax, Hc and AGB. This means that there is a substantial amount of missing data in the model. Please report the missing fractions in the manuscript, or another indicator of the amount of missing data treated with the Expectation-Maximization method that is mentioned in the text (Line 131). Also, how robust are the results to the imputation methods used? It is critical to show that the results are not dependent on data imputation when a large amount of data is missing.
- One core question arising from this study is: how to show that the artificial-rules-based model can reveal the real rules compared with a data-driven model? It is, therefore, necessary to show validation results in the paper. What validation methods did the authors use? how does this validation result compare to typical standard seen in similar studies? If a validation (i.e. k-fold), and robustness checks are done and reported, then the model results can be interpreted with more certainty. However, with the available data in the current version of the manuscript, the model performance cannot be assessed. We believe, that even though the BN was calculated for categorical data instead of continuous data, there is always a need to show that the model predicted well through model validation techniques. Validation results are also important for comparison/verification with future ecological-knowledge-based models.
- Finally, to prevent confusion and over interpretation of the result, the authors should acknowledge that BN are not necessarily causal networks, they are essentially a set of conditional (in)dependencies that factorize the joint probability distribution of all the variables. Causal deductions hence may not be made (Ramazi et al., 2020 - Exploiting the full potential of Bayesian networks in predictive ecology).
3. Technical corrections
3.1. Introduction
- Since the main focus of this study are ecosystem functions, a clear and concise definition of this term is needed in the Introduction section. Also, it would be useful adding some supporting references regarding the theory linking the functional traits included in the paper and Reichstein et al.’s frame.
- There are some terms used along the paper referring to climate variables, vegetation structure variables, or ecosystem functions. These terms change along the manuscript. Try to use a consistent terminology. Examples of these terms are:
Line 27: “complex ecosystems”, “environmental systems”.
Line 37: “environmental conditions”
Line 60: “environments”
Line 158: “ecosystem service functions”
Line 271: “ecosystem systems”- Lines 66-67: The paper by Gregorutti et al. (2017) is used to support the statement that IMP-based attribution can be unreliable when the aim is explaining systematic causality. The cited paper does not discuss systematic causality. Please support this idea with appropriate references.
3.2. Methodology
- Line 95: “Climatic variables:…”
- Table 1: Though the detailed methods for the ecosystem functions’ calculation is in Magliavacca et al. (2021), provide a complete summary for each variable in the column “Approach”. E.g. if the percentile used in the calculations will be reported, report it for all the variables consistently (GSmax – 90th percentile). The use of medians instead of means for some of the variables may also be important information.
- Line 119: Figure 1 is presented as a column with three panels. It is not clear what the author refers to when pointing to the lower left panel.
- Lines 142-147: This sentence is quite long. Try to split it or make it shorter.
3.3. Results
- Line 154: Incomplete sentence at the end of this line: “… SWin, VPD, and showed…”
- Figure 3: In the text, some information is extracted from this figure regarding the correlations climate vs. ecosystem functions and ecosystem functions vs. ecosystem functions. More information can be extracted from this figure regarding the correlations ecosystem function vs. plat traits and plat traits vs. climate.
- Figure 4: Can the display of this figure be improved? E.g. locating the tables in a more equidistant layout.
- Lines 195-196: When looking at Figure 1, it is not clear what the author is referring to with the loop of Tair controlling LAImax. What variables are included in this loop? This word should be used with caution since it could be taken as equivalent to a feedback.
3.4. Discussion
- The interactions among plant traits, climate and ecosystem function variables may be complex when high-order effects are considered. Does the author think these effects play an important role when trying to explain the causal links to ecosystem functions? If these effects are not considered in this study, this limitation should be stated in the Discussion section.
Citation: https://doi.org/10.5194/bg-2022-191-RC1 -
AC1: 'Reply on RC1', Haiyang Shi, 14 Dec 2022
The comment was uploaded in the form of a supplement: https://bg.copernicus.org/preprints/bg-2022-191/bg-2022-191-AC1-supplement.pdf
-
AC1: 'Reply on RC1', Haiyang Shi, 14 Dec 2022
-
RC2: 'Comment on bg-2022-191', Anonymous Referee #2, 28 Nov 2022
Review of the manuscript ‘Revisiting and attributing the global controls on terrestrial ecosystem functions of climate and plant traits at FLUXNET sites with causal networks’
1. General comments
The study describes exemplarily the construction of a network linking plant traits and climatic drivers not only with a statistical background but by taking into account causal linkages. Using a Bayesian Network (BN), expert knowledge is introduced to evaluate the causal effects of climate variables for ecosystem functions. The main achievement and argument is that this type of analysis goes beyond usual statistical relationships which often fail to reveal indirect effects and trade-offs. Although this approach is appealing and from an ecological point of view very promising, the manuscript does not provide a proper validation of the method. The increasing availability of data such as collected within the FLUXNET community hopefully will further trigger new ways of exploring the connection of environmental conditions with the evolving plant community and at the same time allow to test and consolidate analysis tools. Here, the method would benefit from better methodological clarification, description of data use, validation and presentation of the results which are detailed below. The paper needs major revisions before publication.
2. Specific comments
- The text includes various repetitions when stressing that the new method is superior to usual analyses. Please be more concise when making this point (e.g. in introduction and discussion) or more specific when certain aspects are described in detail (e.g. in results). The re-occurring statement is not strengthening the argument.
- Although the data base for the BN is given in table 1 in detail, the choice of the variables does not become clear. Which variables were taken into account and why? Some variables are taken as is and some averaged. Please state as well the temporal resolutions of original variables and averages (why mean values and not medians?). Also the choice of the intervals for discretization (right column in table 1) is not motivated – please provide more detail and reasoning.
- The interesting part of constructing the BN in section 2.2 is not transparent. On which basis is the expert knowledge extracted from Reichstein et al. (2014) and how is it transferred to the BN? When the authors main agenda is to promote their new analysis method, it would be good to give more insights in the process of finding the linkages that should be considered.
- The result section would benefit from a better description of the results of both methods. Reducing the text with general statements should give enough space for guiding through figure 4 and highlighting the benefits of the second approach. How do you motivate this statement when e.g. comparing the results for AGB in the BN-plant-trait-climate in comparison to the BN-plant-trait?
- One major concern is a validation. A presentation of a data-driven method without a validation can hardly be recommended for publication. Please not only provide one but also make clear which data are used for building the model, getting the results and performing the validation.
3. Technical remarks
L 20 and 31: The term ‘emphasized’ seems not appropriate in this context. Please be more specific what you mean here.
L 36: ‘Changes in climate change’ is misleading – please modify.
L 64: The sentence is very long and could be split into two.
L 67: Also very long sentence which makes me wonder, if you assume all relations in these systems to be causal, which they are of course not. Please clarify.
L 96: Including the cumulative soil water index means that a variable is chosen which is already the result of precipitation and evapotranspiration. How do you deal with the interdependency of the variables?
Fig. 2: please explain the black dots in the figures.
L 142: Another very long sentence on a complex issue. A stepwise approach would increase readibility.
L 163: How do you evaluate the compilation as being ‘successful’? Which criteria are fulfilled?
Fig. 4: Values and text in the figure are very small. Why did you choose ‘?’ as a separator between mean and standard deviation?
L 190: As an example for the wish for a better presentation of the results please give more reasoning for the statement that climate variables ‘showed a role beyond plant traits’. Without an understandable link to the results shown, a sentence like this is appropriate in the discussion.
L 224: The methods described in the caption and the text should be moved to the methods section! Here, please elaborate more on the explanation of the very valuable figure 6.
L 281: The idea of extending the causal linkages to the temporal dimension is intriguing but opens the problem of non-independent variables. Do you have an idea how to treat causally linked and dependent variables in this approach?
L 308: Although the conclusions are free to mention related issues which are not part of the study, I would recommend to replace this last point e.g. by the importance of your findings for the modeling community.
Citation: https://doi.org/10.5194/bg-2022-191-RC2 -
AC2: 'Reply on RC2', Haiyang Shi, 14 Dec 2022
The comment was uploaded in the form of a supplement: https://bg.copernicus.org/preprints/bg-2022-191/bg-2022-191-AC2-supplement.pdf
-
AC2: 'Reply on RC2', Haiyang Shi, 14 Dec 2022
Haiyang Shi et al.
Haiyang Shi et al.
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 583 | 138 | 18 | 739 | 8 | 10 |
- HTML: 583
- PDF: 138
- XML: 18
- Total: 739
- BibTeX: 8
- EndNote: 10
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1