Building a random forest machine learning model for Net Ecosystem Exchange estimation in agricultural fields using discontinuous atmospheric Eddy Covariance measurements by Camille Pery et al.
This is my second review of the manuscript following revision. The authors have clearly put substantial effort into addressing the previous comments, and the manuscript has improved in terms of structure, clarity, and presentation. The study addresses an interesting and relevant question, and the combination of eddy covariance measurements, wind-sector partitioning, and machine-learning-based reconstruction has the potential to provide a useful framework for investigating carbon dynamics in heterogeneous agricultural landscapes.
However, several of my previous concerns remain insufficiently addressed, particularly those related to uncertainty quantification. Throughout the manuscript, uncertainty estimates are derived from bootstrap RF ensembles, yet many conclusions are subsequently framed in terms of confidence, robustness, precision, and well-constrained annual carbon budgets. I remain unconvinced that the reported uncertainty estimates support these stronger claims. Given the extensive reliance on gap-filled data and the large differences in observational support among sectors, it remains difficult to assess the confidence that can reasonably be placed in the reported NEE, GPP and Reco estimates and in the ecological interpretations derived from them.
Major comments
The manuscript presents uncertainty estimates derived from an ensemble of 100 bootstrap Random Forest models and refers to the resulting intervals as confidence or prediction intervals. However, these intervals appear to represent only the variability among RF model realizations and therefore capture only one component of the overall uncertainty associated with the reconstructed NEE time series. Several potentially important sources of uncertainty do not appear to be included, including: uncertainty in the EC flux measurements themselves, including u* filtering sensitivity, uncertainty associated with footprint attribution and sector delineation, uncertainty arising from assigning observations to specific wind sectors, and uncertainty associated with the subsequent NEE partitioning procedure.
This issue is particularly important because many sector-period combinations rely heavily on reconstructed rather than directly observed fluxes (Table 1), with gap-filled proportions frequently exceeding 90% and reaching up to 98%. Consequently, the reported annual NEE budgets are driven largely by model predictions, making a comprehensive treatment of uncertainty essential for assessing the robustness of the conclusions. More generally, I am left wondering to what extent the annual sector-specific NEE estimates can be interpreted as robust ecological quantities given the very large differences in sample size among sectors and the extremely high proportion of reconstructed data in several cases. The manuscript places considerable emphasis on differences in annual NEE among crops and sectors; however, it is not yet clear whether these differences are sufficiently constrained by observations or are primarily a product of the modelling framework. Given the extremely high proportion of reconstructed data for several sectors, please provide evidence that the resulting annual NEE estimates are robust to the modelling choices employed. While the modelling framework is carefully developed, it remains unclear how robust annual cumulative NEE estimates can be when they are derived from such limited observational support. The authors should discuss this limitation more explicitly and justify why annual sector-specific carbon budgets remain reliable despite the very high degree of extrapolation required.
In addition, GPP and Reco are derived from RF-reconstructed NEE, yet no uncertainty estimates are reported for these quantities. It is unclear how uncertainty in the NEE reconstruction propagates through the light-response partitioning approach. A possible solution would be to apply the partitioning procedure to each bootstrap RF realization separately and derive confidence intervals for annual GPP and Reco from the resulting distributions. I therefore encourage the authors to clearly define what is and is not captured by the reported uncertainty intervals, discuss the uncertainty sources that are not represented, and, where possible, propagate RF reconstruction uncertainty through to the final NEE, GPP and Reco estimates. Without this information, it is difficult to assess the confidence that can be placed in the reported differences among sectors and land-use types.
I am also concerned about the treatment of wind direction during reconstruction. Wind direction is included as a predictor and justified as accounting for directional heterogeneity in source areas. However, when reconstructing sector-specific time series, wind direction is randomly assigned from within the sector boundaries whenever the actual wind direction falls outside the sector. Please report the variable importance of wind direction and evaluate the sensitivity of reconstructed NEE to this random assignment. If predictions are insensitive to the selected wind-direction value, this should be demonstrated quantitatively. If they are sensitive, the rationale for assigning random wind-direction values requires further justification.
The explanation provided in Sect. 4.1 for the differing behaviour of the global and sector-specific models relies heavily on the assumption that wind direction is only weakly captured by the global RF model because it competes with other predictors during tree construction. However, this interpretation appears speculative and is not directly supported by the analyses presented. I encourage the authors to provide additional evidence supporting this claim, for example through the variable importance of wind direction in the global model, partial dependence analyses, or other diagnostics illustrating the extent to which the global model is able to recover sector-specific structure from the wind-direction predictor.
Additional comments
• Was soil moisture and temperature recorded separately for the three different crop fields? If not, how can their inclusion as sector-specific predictors be justified? The manuscript argues that NDVI could not be used because the fluxes represent footprint-integrated measurements rather than pure-field observations. However, a similar argument could potentially be made for soil temperature and soil moisture if these variables were measured only at a single location.
• Why use a day/night indicator when PAR is already included as a predictor?
• Was mtry tuned or sensitivity-tested? If not, why was the default heuristic adopted rather than selecting mtry based on predictive performance?
• The description of the OOB validation procedure is unclear. Because the analysis includes both the internal RF bootstrap used to generate OOB predictions and an additional layer of B = 100 bootstrap-resampled RF models, please clarify how RMSE, MAE and R² were calculated. Were performance metrics computed separately for each bootstrap model and then averaged, or were predictions aggregated across bootstrap models prior to calculating the performance metrics?
• The uncertainty intervals seem to characterize bootstrap model variability, yet differences among modelling choices produce much larger changes in annual NEE estimates. Please clarify what uncertainty is and is not captured by the reported intervals.
• Figure 8 presents uncertainty at the sector level only. Since the main conclusions of the manuscript are ultimately discussed at the crop level (wheat, mixed grain and grassland), I suggest including a corresponding summary figure showing the reconstructed NEE and associated uncertainty for the aggregated crop categories, particularly for the combined wheat sector (4+5+6). This would help readers assess the robustness of the crop-level conclusions and better relate the reported uncertainties to the final annual carbon budgets discussed in the manuscript. This also applies to the discussion, which is hard to follow as references are made towards sectors rather than land cover.
• Paragraph starting on line 623: It is unclear what is meant by "precise" in this context. The manuscript does not quantify uncertainty in annual cumulative NEE estimates, but only RF bootstrap uncertainty. Therefore the statement that some annual NEE estimates are more or less precise does not appear to be demonstrated.
• I am struggling to understand the purpose of Sect. 4.3. The Kowalski partitioning is applied to the RF-reconstructed NEE, therefore the comparison between partitioning-derived NEE and RF-derived NEE does not appear to be a comparison of independent estimates. Please clarify what information is gained from this comparison and in what sense the two methods can be considered independent.
• Section 4.4 starting on line 709: I find this section overly speculative. The mixed-grain and hybrid estimates are repeatedly described as footprint-contaminated and not fully representative of the target parcels, yet the discussion proceeds to attribute the observed patterns to specific mechanisms (land-use history, productivity, microbial processes, respiration dynamics, etc.). These interpretations appear stronger than the evidence presented.
Minor comments
There are several minor grammatical and stylistic issues throughout the manuscript (e.g. line 224, line 251 and several others).
Line 55: the chain of reasoning would benefit from additional references.
Line 78: VPD has already been defined previously.
Lines 119-120: GPP and Reco have already been defined.
Line 138: Please specify which species comprised the mixed-grain crop.
Line 140: Did the wheat field also remain bare after harvest, similar to the mixed-grain field?
Lines 151-153: EC and NEE have already been defined.
Lines 179-187: This addition is useful, but I recommend rewriting the paragraph slightly to improve readability.
Line 189: Please correct the manufacturer information in parentheses.
Line 220: The sentence beginning with “This parameter” should be rewritten.
Line 226: By “strong meteorological disturbances”, are the authors referring to instrumental artefacts or physiologically plausible flux responses?
Figure 3: I recommend increasing the font size for readability. The authors may also consider moving this figure to the Supplementary Information.
Line 293: The sentence following the reference should be rewritten.
Line 300: The wording “pure agricultural parcels” is unclear. I assume the authors refer to the three crop parcels described above, but the wording could be improved.
Line 389: NEE, GPP and Reco are redefined although they were introduced previously.
Line 415 (and Section 3.4): Soil respiration was not investigated in this manuscript and should be excluded from the results section.
Line 446: This sounds like a repetition of the methods.
Line 502: From this point onward it would help to provide uncertainty estimates for GPP and Reco.
Line 544: The title is not formulated as a question and should be reworded.
Line 603: I do not agree with the phrasing that the cumulative annual NEE value for the wheat parcel represents a “genuinely well-constrained” value, given the uncertainty quantification issues discussed above.
Line 607: Adding the references to Kljun et al. (2015) and Kumari et al. (2020) here does not seem appropriate when discussing the authors' own data.
Line 642: The manuscript describes the RF reconstruction and the light-response partitioning as two "independent methods", but the partitioning is applied directly to the RF-reconstructed NEE time series. The agreement between cumulative NEE estimates therefore represents an internal consistency check rather than an independent validation of the GPP/Reco decomposition.
Line 672: I do not fully understand this statement. The manuscript appears to compare cumulative NEE derived from the Kowalski partitioning with cumulative NEE from the RF reconstruction, yet the partitioning is applied to the RF-reconstructed NEE time series. Could the authors clarify exactly which datasets are being compared and how the agreement is evaluated?
Line 674: Throughout the discussion, the authors argue that the annual NEE estimates differ in precision, confidence and robustness among sectors. If so, discussing and comparing the absolute NEE values among cropping systems is concerning, particularly when some estimates are explicitly described as less reliable because of footprint mixing.
Line 697: The meaning of "some residual caution" is unclear.
Line 704: It is unclear what the authors mean by interpreting the mixed-grain estimate "qualitatively rather than quantitatively". If the estimate is not considered an acceptable field-specific quantitative carbon budget, it is unclear how conclusions regarding its underlying controls (e.g. productivity, soil respiration, GPP/Reco balance) can be drawn. |
This manuscript presents a novel gap-filling method for CO₂ fluxes measured by a single eddy covariance tower located in the midlle of four fields with different vegetation types. The proposed approach relies on machine learning techniques, specifically a random forest algorithm, to reconstruct fluxes when wind direction does not originate from the target area. The random forest model is trained using a set of environmental variables, including temperature, radiation, soil moisture, as well as information related to crop management practices and the day of year. This approach shows strong potential for reconstructing CO₂ fluxes, as well as energy fluxes, in heterogeneous landscapes using measurements from a single flux tower, and is therefore of great interest for extending the spatial representayiveness of Eddy Covariance.
However, although the method is promising and deserves particular attention, the manuscript would benefit from more details regarding the technicality of the method. It would in particular benefit from the computation of the uncertainties in reconstructed/gap-filled CO2 flux in each wind sector. This woul allow to assess the statistical significance of the differences in CO2 fluxes observed in each wind sector (specific to a vegetation/soil type). The authors could also try to use spatially explicit data to evaluate if the observed CO2 flux dynamics is sound. The use of NDVI or EVI from Sentinel data may be of interest. Due to the lack in uncertainty and signficance analysis, in its current form, the discussion remains somewhat speculative regarding the reconstructed CO₂ fluxes for the plots surrounding the tower.
I would therefore recomend major revision by integrating uncertainties and significance analysis, before interpreting and discussing potential differences between crop types.
Specific comments:
L155-164: in equation ρ should be "dry" air density (ρd). and there is a conversion from g to mol missing . Also give units of w and s.
L190: why using a time delay with colocated CO2/H2O and sonic instruments?
L200: the u* threshold is usually adapted to local conditions. Jusitify the use of a single value.
Figure 3 (L219): you could select only neutral conditions here to have a closer relationship with the roughness z0.
L244-250: this is somewhat redundant with previous paragraph.
L265-270: the temporal subdivision, and all parameters in the RF appraoch, should be explicited in a table (in appendix eventually).
L280: could you explicit what is a qualitative day-night indicator?
L282: PAR also captures seasonal changes in radiation regimes? Is this not redundant?
Section 2.3.3 : in the predictive model, a measure of the leaf area index would be very valuable and would integrate spatial variations. I would recomend trying using an NDVI or EVI high resolution satellite product (from Sentinelle data).
Secion 2.3.4: this is the critical point in this manuscript. I would suggest the authors to compute an uncertainty for the RF reconstructed CO2 fluxes. This would be necessary to interpret the results and determine if significant differences can be retreived from each wind sector. I am not an expert in RF modelling but I guess some uncertainty could be computed. At least from the comparison between model and observed data uncertainties could be deduced.
L314-318: I would recommend avoiding any use of GPP here as NEEday is not GPP. This paragraph could be rephrased to clarify that NEE was simply split between day and night
3.1 section: The Figures 7 and 8 would benefit from showing the difference between the modelled and observed CO2 flux rather than each flux on top of each other. Actually the two may be shown : the modelled fluxes and the difference with observations. This graph should include uncertainties in the modelled fluxes.
The rest of the results and the discussion provides valuable arguments but would really need to have uncertainties to evaluate if the differences discussed are significant or not, especially when comparing annual budgets.