the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Resolving scale-variance in the carbon dynamics of fragmented, mixed-use landscapes estimated using Model-Data Fusion
David T. Milodowski
T. Luke Smallman
Mathew Williams
Abstract. Many terrestrial landscapes are heterogeneous. Mixed land cover and land-use generate a complex mosaic of fragmented ecosystems at fine spatial resolutions with contrasting ecosystem stocks, traits and processes, each differently sensitive to environmental and human factors. Representing spatial complexity within terrestrial ecosystem models is a key challenge for understanding regional carbon dynamics, their sensitivity to environmental gradients, and their resilience in the face of climate change. Heterogeneity underpins this challenge due to the trade-off between the fidelity of ecosystem representation within modelling frameworks and the computational capacity required for fine-scale model calibration and simulation. We directly address this challenge by quantifying the sensitivity of simulated carbon fluxes in a mixed-use landscape in the UK to the spatial resolution of the model analysis. We test two different approaches for combining EO data into the CARDAMOM Model-Data Fusion (MDF) framework, assimilating time series of satellite-based Earth Observation (EO) derived estimates of ecosystem leaf area and biomass stocks to constrain estimates of model parameters and their uncertainty for an intermediate complexity model of the terrestrial C cycle. In the first approach, ecosystems are calibrated and simulated at pixel-level, representing a "community average" of the encompassed land cover and management. This represents our baseline approach. In the second, we stratify each pixel based on land-cover (e.g. coniferous forest, arable/pasture etc.), and calibrate the model independently using EO data specific to each stratum. We test the scale-dependence of these approaches for grid resolutions spanning 1° to 0.05° over a mixed land-use region of the UK. Our analyses indicate that spatial resolution matters for MDF. Under the "community-average" baseline approach biological C fluxes (GPP, Reco) simulated by CARDAMOM are insensitive to resolution. However, disturbance fluxes exhibit scale-variance that increases with greater landscape fragmentation, and for coarser model domains. In contrast, stratification of assimilated data based on fine-resolution land-use distributions resolved the resolution dependence, leading to disturbance fluxes that were approximately double the baseline experiments. The differences in simulated disturbance fluxes were sufficient to drive alternative interpretations of the terrestrial C balance: in the baseline experiment the live C pools suggest a strong C sink, whereas in the stratified experiment, the live C pools were approximately in steady-state as the C gains from NPP were balanced by losses due to the higher simulated harvest fluxes focused in conifer woodlands. We also find that stratifying the model domain based on land-use leads to differences in the retrieved parameters that reflect variations in ecosystem function between neighbouring areas of contrasting land-use. The emergent differences in model parameters between land-use strata give rise to divergent responses to future climate change. Accounting for fine-scale structure in heterogeneous landscapes (e.g. stratification) is therefore vital for ensuring the ecological fidelity of large-scale MDF frameworks. The need for stratification arises because land-use places strong controls on the spatial distribution of carbon stocks and plant functional traits, and on the ecological processes controlling the fluxes of C through landscapes, particularly those related to management and disturbance. Given the importance of disturbance to global terrestrial C fluxes, together with the widespread increase in fragmentation of forest landscapes, these results carry broader significance for the application of MDF frameworks to constrain the terrestrial C-balance at regional and national scales.
David T. Milodowski et al.
Status: final response (author comments only)
-
RC1: 'Comment on bg-2022-160', Anonymous Referee #1, 21 Sep 2022
General comments:
This paper demonstrates the impact of using both finer-scale and categorically refined representations of land surface heterogeneity on modeled carbon stocks and fluxes. The authors present a case study over a region with four dominant land use types to demonstrate that modeling the ecosystem response of each land use type separately, and aggregating the results, does not always yield the same result as modeling the aggregate ecosystem response of the region. The authors document differences in simulated carbon stocks and fluxes, and derived parameters characterizing the ecosystem function, among the different approaches and resolutions tested.
The hypotheses are interesting and well-explored by the experiments chosen, and the results are important. By documenting the sensitivity of the data assimilation framework to the spatial scale and categorization of the data inputs, this work highlights ecological assumptions embedded in standard usage of this and similar models that may undermine their ability to investigate questions of ecological function and future response. Raising this issue is a useful contribution.
However, some of the framing and concluding statements in this paper assert improvements in ecological fidelity or simulated carbon fluxes due to stratification without validating this with any outside data. It would be great if possible to include some validation, such as comparison with outside data to validate derived parameters (e.g., residence times) or carbon/water flux data from flux towers. If this isn’t possible, I recommend adopting new language in your framing and conclusions to focus on the sensitivity you demonstrate and make less claim to ecological fidelity or improved representations of carbon fluxes. Section 4.3 is informative and some of the context outlined there could be brought out in the framing and goals at the outset.
Specific comments are below:
Comment 1: Section 2.3-2.4: It would be helpful to see a paragraph at the end of the methods discussing the various spatial scales at play, and how these are integrated into the model pixel in each case. When working with a 0.05 degree (~5km) model resolution but imposing constraints on biomass at 100m, on a soil type at 250m, an LAI at 300m, and a timber harvest at 30m, how are these aggregated across the pixel?
Comment 2: As a follow on from comment 1, the description of study area emphasizes gradients in temperature and precipitation over topographic features within the 3x3 grid (Lines 128-131), which justifies testing surface resolutions down to 0.05 degree, but then the model runs use 0.5x0.5 forcing data in each case. How do the authors expect this to relate to the amount of scale-dependent variation seen across the model runs?
Comment 3: Section 3.1 The calibration metric, RMSE/sigma, could be further explained. It sounds as though smaller values are desirable here, but if this is a comparison to inherent observational uncertainty, I don’t immediately see why <1 is a good thing. Please make this a bit clearer.
Comment 4: Table 1, Table A2: Additionally, the values of the calibration metric do not proceed monotonically with the shift in resolution. It would be helpful if the authors could explain (or speculate) why this is, especially in the context of the stated goal of improving ecosystem representation by going to smaller scales. A response to this could connect to a response to comments 1 & 2— how does the scale of the input data impact how well things are lining up in the model (applying the right processes to the right initial conditions) at different resolutions?
Comment 5: Line 275 and Figure A2: Please explain why the coniferous woodland has a strong seasonal cycle of LAI which reaches zero in the winter. The black dots in Figure A2 top left panel suggest that this oscillation is present in the earth observation data, but Scotland is not known for its deciduous conifers. Does snow blanketing the tree canopies, masking out the greenness or making the canopies indistinguishable from the ground, cause this seasonality in the Copernicus LAI product, or is this considered ecologically realistic for your region? If it is unrealistic, does this matter to the resulting biomass trends— for instance, did the authors test a different (presumably more realistic) oscillation bottoming out at ~3?
Comment 6: Section 3.3: Regarding the differential response to disturbance flux, it would be helpful if the authors emphasize earlier on that this arises from a mismatch in applying the disturbance to the correct land cover type when using the aggregated pixels. I see the authors do come to this in lines 358-361 but would appreciate it earlier. Section 2.3.5 could be a good place to explain how the authors imposed the disturbance flux in each case, so it is ultra-clear why this difference in how the disturbance is allocated to each land use type arises between the two cases.
Comment 7: Section 3.3: Line 319: “it is evident that stratification leads to preservation of ecological information across resolutions” and similar statements throughout; suggest to qualify these statements, e.g., “as encoded in the observations available to the model”. Getting back to the conifers acting like deciduous trees— it is important to tread carefully with caveats that the observations themselves come with many assumptions, and may not always represent ecological fidelity. The authors make this caveat in section 4.3 line 395-400, but it would be helpful to keep it at the forefront throughout.
Comment 8: Section 4.3 line 395-320: Great points. It would help satisfy reader curiosity if the authors could delve a bit into these deficiencies (Zhao et al 2020, Heiskanen et al 2012) as relevant to the datasets they are using, and discuss how the deficiencies might impact their results. This comment has substantial overlap with comment 5.
Comment 9: Line 406-415: This is a very strong point. It would be great to go further and see the authors chart out a bit what is needed to actually do these improvements in process representation— what is the to-do list? How will improvements be verified?
And a few minor technical comments below:
Comment 10: Table 1 last CWood stratified row, shifted numbers
Comment 11: Figure 6: What does the color of dots (blue vs green) in the right panels mean? A legend would help
Comment 12: Figure A2-A3: A legend would help here also.
Citation: https://doi.org/10.5194/bg-2022-160-RC1 -
RC2: 'Comment on bg-2022-160', Anonymous Referee #2, 22 Sep 2022
This paper demonstrates the importance of accounting for fine-scale structure in heterogeneous landscapes to ensure ecological fidelity in modeling carbon dynamics. The authors designed two different approaches with model-data fusion to constrain estimates of model parameters and their uncertainty, and compared the difference in simulated carbon dynamics by these approaches with varying spatial scales over a mixed land-use region of the UK. The paper is overall well-written, and the experiment is well-designed. However, I have serval concerns regarding the methods and conclusions. Please see my comments below.
Major concerns:
The model-data fusion framework (CARDAMOM) assimilated remotely sensed LAI and aboveground biomass (line 158), but soil organic carbon (SOC) extracted from SoilGrid2 was “used to set a prior constraint on the initial SOC stock. Is SOC a constraint in CARDAMOM? To set the initial SOC stock, how did the initial SOC be set for the baseline and stratified cases and for each spatial scale? Did the authors separately set the initial SOC for each sub-pixel type and constrain the sum/average of sub-pixel types to the SOC value derived from SoilGrid2? Please provide more information about this.
Is Cwood in the Results section the aboveground biomass or total woody carbon that includes both aboveground and belowground? Line 190, “Cwood pool is also a reservoir for non-woody structural tissues, for example in areas covered by crop and pasture.” Does DALEC also have a woody carbon pool for pasture? Did the authors also infer belowground woody carbon for crop and pasture based on the allometric relationship in Eq. 3?
In table 1, it seems that Cwood and LAI are underestimated (i.e., negative bias) in both baseline and stratified cases across spatial scales. What are the possible reasons for it? Maybe the parameters were not fully constrained?
Several arguments in the conclusion section sound a bit misleading to me, e.g., “failure to account for sub-pixel ecosystem heterogeneity within MDF inversions leads to bias in the flux estimates”, “stratification improves flux estimates”, and “ecological fidelity of the calibrated model parameters is enhanced”. The differences in RMSE and bias between baseline and stratified cases are not very significant, and sometimes the biases (absolute values) are even greater in stratified cases (e.g., -584 gCm-2 in baseline Cwood and -627 gCm-2 in stratified Cwood at 0.05deg scale). More validations of model estimates and constrained parameters should be included to draw such conclusions. Regarding “stratification improves flux estimates”, the authors might want to say that “stratification reduces flux uncertainties”. If yes, however, lines 287-289 already demonstrate that the reduced uncertainty is a result of assuming independence between strata, and uncertainty with full correlation across strata is comparable to the baseline uncertainty. Please clarify these arguments.
I was confused about the reason for invariance in biogenic fluxes with respect to both resolution and method. Does this indicate the biogenic processes in DALEC are (almost) linear ecological processes? The linearity in GPP estimation could be possible, but I am not sure if Reco should have a similar pattern.
Minor:
Line 123, should be 30,000 km2?
Line 293, how does DELAC simulate fire effects? Why fire was negligible?
Citation: https://doi.org/10.5194/bg-2022-160-RC2 -
RC3: 'Comment on bg-2022-160', Anonymous Referee #3, 10 Oct 2022
Dear authors,
I feel like this study and methodology of cardamom represents a major advancement in model calibration. It is particularly exciting to see a framework that could run autonomously using earth observation data. The reproducible nature of this data fusion and calibration process, when coupled with the Bayesian methodology for error estimation (and propagation) could provide more iterable forecasts of carbon or other ecological processes. The question of how spatial heterogeneity plays out in cellular automaton models with single calibrations is crucial to future forecasting and I appreciate the focus on both natural and anthropogenic disturbances. Further, I appreciate the authors’ efforts to address the problem of Jenkin’s inequity and the role scale selection plays in informing the forecasting and responding. I think the authors' use of stratification by land cover represents a relatively straightforward, logical, and widely available method by which to create more representative models.
However, I feel that some of the conclusions may overreach their results (particularly without independent evaluation). While this methodology has the possibility to improve the estimation of fluxes (etc.), it is not validated to have done so against measurement. The paper presents a gap in understanding to what degree model performance was improved while making some strong claims of the level of ecological fidelity it is able to preserve. I feel this study is both novel and relevant. I would like to see the authors either change the language and address more of the existing limitations of this study or provide further validation to some of the authors' larger claims. I look forward to receiving your response and want to thank you for conducting great work.General comments
The problem of Jenkin's inequity in landscape or earth systems modeling is a valid and underrepresented viewpoint. However, the answer the authors' model provides can not solve this. Raising this in the introduction raises the idea that authors’ methods will be solving or improving on the current structure. Please address in the discussion whether authors feel results provide further proof for Jenkin's inequity or whether they work to address it. This paper does not clearly quantify the advantage of picking one scale (sub-degree +LUC ) in a scale variant system. Some may seem likely or self-evident but would need to be proven. For example, the effect of Jenkin's inequity might be quite similar from the cellular to sub-degree+LUC scale when compared to the explicit plant scale. The authors do however do a great job quantifying that there is an amount of scale variance in this model. Quantification and discussion of how this scale variance impacts forecast, when compared to observable phenomena, would provide a lot of support to this paper.If I understand the authors’ methods correctly, for the baseline model the authors parametrize each pixel separately and in the stratified version the authors separately parameterize each pixel and each pixel's land cover. Is there no information shared across pixels or land cover? Given that land covers would presumably share ecological properties, what is the advantage of not using a hierarchical Bayesian, with priors informed by the larger population of land covers or a bayesian mixed model approach? Please either clarify the decision to make this choice or discuss further the limitations of the separate pixel approach.
Given that none of your parameters seem to directly map onto disturbance, is your model capturing the heterogeneity on the landscape, or overfitting a model? Perhaps some of my concern comes from a lack of understanding of how harvest or fire operates in this model (see below). Harvest would be inversely correlated with the likelihood of future harvest at certain temporal scales and correlated at others. At the scale of a few decades recovering stands would likely (though not exclusively) experience an increase in GPP as forests regrow. Given the stationarity of your parameterization, how does your parameterization constrain such instances?
Hypothesis three- could be improved. To test that any two methods of model parameterization will have contrasting parameters is almost by definition true. Further, they will always have divergent projections on some level. Please provide better constraints to make this hypothesis falsifiable or use a more stringent definition of the contrasting and divergening of parameters.
Specific comments
Line 155: Please provide either here, in the results, or in the appendix the results of the MCMC process. Or how your criteria for model convergence. Accepted sample rates, plots of autocorrelation, and hyperparameters provided are all necessary to determine if confidence intervals are reasonable.Line 167: Given the importance of EDCs in determining this you should list them plainly, and discuss the constraint they do or do not provide with regard to the function you are trying to achieve.
See: Buotte, P. C., Koven, C. D., Xu, C., Shuman, J. K., Goulden, M. L., Levis, S., ... & Kueppers, L. M. (2021). Capturing functional strategies and compositional dynamics in vegetation demographic models. Biogeosciences, 18(14), 4473-4490.Line 256 (Disturbance): My apologies if I misunderstand this in other comments. Can the authors please provide greater detail on how disturbance is implemented in the model? This paragraph deals primarily with how it is constrained. There are no direct parameters listed in table A1. If I interpret this correctly, did the authors remove a percentage of tree cover or carbon % to match these data sets? Again, given that this is one of the key differences in the stratified model, a better understanding of the disturbances function in the model is important.
Line 320: Shredding of information implies a specificity not realized here. Information loss is inherent in all models. Without estimating the level of information that is lost, shredding seems overly evocative.
Line 361: While more consistent, what evidence do we have that the prediction is significantly or functionally different from the baseline model? The confidence intervals seem to overlap significantly.
Line 376: I feel this sentence speaks to my larger concerns. There is no way to say that disaggregation ensures the ecological fidelity of a system. Ecology is also scale-dependent. Further, without validation by observation, there is no way to know that the version outperforms the previous version, given that Jenkins inequity would be a property of this scale as well.
Line 385: Different modeling frameworks providing different (though I would not say divergent) outcomes are highly likely. I feel your argument would be improved if you would better quantify or qualify the significance (either statistical or practical) of this level of difference.
Line 410: Given static and statistical parameterization, it would be nice to understand the climate change implications of the stratification approach.Line 436: The more you stratify a single cell, the greater proportion of it would be captured by this edge or gradient space. If the gradient space has unique ecosystem properties, is there a point where further stratification would further miscalibrate the model?
If helpful, see: Cushman, S. A., Gutzweiler, K., Evans, J. S., & McGarigal, K. (2010). The gradient paradigm: a conceptual and analytical framework for landscape ecology. In Spatial complexity, informatics, and wildlife conservation (pp. 83-108). Springer, Tokyo.
Line 459: Again, accounting for subcellular processes at the scale you provide by stratification likely also has a high amount of ecological information loss.
Line 467: While conceptually likely that this provides improved flux estimates, I don’t think you have provided enough validation to show this is true. Reduced parameter uncertainty does not dictate estimation capability. Also, if I understand section 3.1 correctly then the parameter uncertainty is roughly similar, though the means may converge indicating some level of reduced scale variance. That this method reduces scale variance, does not directly imply improved estimation.
Figure 9: I feel that this figure is crucial to your larger argument of scale-dependent outcomes impacting future projections. I feel several aspects of this figure should be revised. Do these model runs represent the median trait estimation or a single draw of the cardamom traits? Please explain in the text. Further, why is the error not propagated here, given the Bayesian approach? This seems crucial to the case that these methods result in fundamentally different models. It is hard for me to understand the implementation (or lack thereof) of disturbance in these forecasts, given that none of the parameters presented would represent that explicitly. See the above comment, some of this may be a misunderstanding of how disturbance works within the model. If the disturbance is only applied top-down, do these projections represent what you captured (that disturbance is highly scale-relevant)?
Citation: https://doi.org/10.5194/bg-2022-160-RC3
David T. Milodowski et al.
Data sets
CARDAMOM driving data and C-cycle model outputs to accompany "Resolving scale-variance in the carbon dynamics of fragmented, mixed-use landscapes estimated using Model-Data Fusion" David T. Milodowski, T. Luke Smallman, and Mathew Williams https://doi.org/10.7488/ds/3509
David T. Milodowski et al.
Viewed
HTML | XML | Total | BibTeX | EndNote | |
---|---|---|---|---|---|
341 | 121 | 15 | 477 | 5 | 6 |
- HTML: 341
- PDF: 121
- XML: 15
- Total: 477
- BibTeX: 5
- EndNote: 6
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1