The study by Li et al. entitled “Global natural wetland methane emissions (2000-2025)” uses various bottom-up and top-down estimates of methane emissions from wetlands for the period 2000-2020 in conjunction with a machine-learning approach to model global emissions for the period 2000-2025 and simultaneously project global methane emissions from wetlands for the period 2021-2025. Based on these newer estimates, wetland emissions and growth rates are presented and discussed in relation to latitude, regions and season. Their findings suggest that methane emissions from wetlands in the northern mid-to-high latitudes have increased in recent years. Hence, this machine-learning approach provides the opportunity to model more recent developments of wetland methane emissions, able to investigate possible feedback mechanism and emission changes earlier than global data networks can usually achieve.  

This manuscript is a valuable contribution to the research community, and I recommend it for publication in ESSD once the authors have addressed the comments and suggestions raised in this review.

General comments:

The authors should improve the resolution of all figures to at least 300 dpi. After improved resolution, Figures 1, 3, should be increased in size, so that small details can be seen.

Is it possible to write the definitions of all parameters including a “_” with small letters instead (e.g., CH_{4, pred})?

The gap between top-down and bottom-up approaches for wetland emissions was only recently narrowed by improvements of simulations, adding additional parameters, such as including inland freshwater systems, inundation, accounting for double counting, etc. (e.g., Saunois et al., 2025). The authors state that those are not included in the simulations (line 150). What error is expected regarding those missing emissions?

The uncertainties of the emissions are reported with a CI value of 95%. Highest uncertainties for wetland CH4 emissions correlate with highest emitting regions (e.g., Zhu et al., 2025). This is not reflected by the CI, as can be seen in Figure S1 (0-10%). Did the authors consider uncertainties from the spread of ensemble simulations? Did all BU and TD estimates provide global coverage, or is there a bias in spatial data coverage, influencing uncertainties? Remapping coarser products on finer resolution; have the authors just replicated the mean and uncertainties, or did they used a scaling factor (particularly important for the uncertainty)? The manuscript would benefit from more discussions regarding uncertainties and uncertainty propagation.    

Methane experienced a remarkable growth during 2020-2022, coinciding with an unusual La Nina event (triple-peak, Hasan et al., 2022, Nisbet et al., 2023), also supported by observations in the tropics (e.g., Ort et al., 2026; Balasus et al., 2026). Figure 5 shows a strong increase of GMB emission estimates towards 2020, mostly visible in tropical regions, which the XGBoost predictions cannot reproduce. Simultaneously, the Covid-19 pandemic is discussed to had a global influence on CH4 oxidation capacity though the decreased NOx emissions (e.g., Stevenson et al., 2021). However, as Covid-19 influences most likely are not considered by the XGBoost predictions, neither are unusual ENSO phenomena, aren’t they? Sections 3.3, and 3.4 would benefit from more detailed discussions about limitations of the XGBoost predictions, and comparisons of the results with other studies, to better quantify the value of the results provided by this machine-learning approach.

With this machine-learning approach, predictions from 2021-2025 were performed. Why did the authors decide for this time period only, and not included climatological predictions further into the future?

Specific comments:

• Title: The title should be reconsidered to match the content of the manuscript and the journal requirements more. The machine-learning approach should be somehow mentioned. What about “A machine-learning approach to estimate global wetland methane emissions to recent years (2000-2025)” or “Extending global wetland methane emissions to more recent years using a machine-learning approach (2000-2025)”?

• Abstract: I would remove the “z-scores” in the Abstract (too detailed).

• Introduction: Wetlands for BU and TD approaches have the largest uncertainties of all methane emissions. I miss this important information in the introduction, alongside with suitable literature (e.g., Zhu et al., 2025).  

• Line 143: Please specify this statement: “sparsity of training data”. Do you mean from satellites, in-situ observations, both? Then, this is only true for some regions, e.g., the tropics and high-latitudes. Wouldn’t this be also an issue for the data used for this study, as it is based on in-situ and satellite data?  

• Lines 140-158: What makes the machine-learning approach presented in this study superior, or how does it differ from other data-driven modeling approaches (e.g., Yuan et al., 2024)?

• Lines 190-194: The GMB dataset used is a combination of BU wetland biogeochemical models and TD inversion products. The TD inversions include observations from satellites and surface observations, which were then used to estimate posterior wetland emissions. However, the authors state here that they used anthropogenic emission inventories for each inversion. It is not clear to me if the authors therefore only include wetland emissions, or also emissions from anthropogenic or other natural sources. Please specify more, how the posterior flux estimates were achieved.

• Line 196: Do all BU and TD approaches provide global data coverage? If not, please discuss possible data biases.

• Lines 251-253: What is the reason for this defined cut-off? Is it the highest X^th percentiles?

• Lines 263-269: What is the exact difference here between the two test data periods and the validation dataset? Please specify their roles more. And have you tested different choices of test and training data periods? Did those choices make a big difference on the outcome for the 2020-2025 period?

• Line 295: Were latitudinal differences considered by calculating the grid-cell areas?

• Line 338: Is R²=0.65+-0.003 here correct for the global monthly mean? Comparing this value with Figure S3a, the mean would lie more around 0.9, or are there strong outliners reducing the mean? Then, it might be better to mention this, and where they are geographically, or adding the median and standard deviation/percentiles as well, so that this skewness is clear. Or what is the difference in data shown between Fig. 1a and Fig. S3a? Same for the following R² mentioned in the text. Is the CI for the mean value here an averaged over all CI’s? What about the spread across the means?

• Lines 369-371: Please include here the global mean wetland CH4 emissions for the predicted and the GMB CH4 over the test periods. What is their estimated CI/uncertainty? Does the global difference of 2.27 Tg yr^-1 exceed the associated uncertainties?

• Line 392: Define “z-score” briefly and what this value represents. Or refer to literature defining the z-score.

• Lines 414-420: Explain the definition of significance in the text as well.

• Figure 4: Is there a reason, why the total emissions are plotted in logarithmic scales? It does not seem to be needed regarding their values ranging from mostly 0-30 Tg yr^-1. Would it be possible to include the emission changes between 2021-2025 compared to 2000-2020, for all regions on, e.g., a second y-axis? As the authors are mainly referring to those differences in the text.

• Lines 440-452: The authors are referring to two La Nina events (within the training data) and one El Nino event (in the prediction) here. How certain are the authors that the XGBoost prediction can represent the variability of ENSO phenomena frequencies?

• Figure 5: Could linear regression lines be included here to verify the trends (slopes)? Furthermore, linear regression lines for the separated time periods (2000-2020; 2021-2025) could show the difference in trend development for each period more visually, supporting the discussions in the text.

• Conclusion: At present, the conclusion is limited mainly to the results of the seasonal analysis of the methane growth rate, which, in my opinion, is not the only aspect of this work. The model structure and performance evaluation should also be briefly mentioned here to place the results in a broader context. Furthermore, it should be discussed whether there is room for improvement or whether this model is also suitable for other applications, apart from CH4 emissions from wetlands.

Technical comments:

• Please change throughout the whole manuscript all units Tg/year into Tg a^-1, or Tg yr^-1 (Tg CH₄/year, accordingly).
• Line 82-85: Shorten sentence for easier reading flow. Exchange “and applying the framework to” with “applied on”. Put “ at monthly 1°x1° resolution” into next sentence (you anyhow repeat this information here).
• Line 86 – 88: Make a new sentence out of the content of what you write within the brackets (R² and RMSE).
• Line 88-89: Change to “model estimates, including 22 process-based, and 13 atmospheric inversion estimates, paired with …”
• Line 90: Change “While” to “Our results show that”.
• Line 92: Change “, this” to “. This”.
• Line 96: I would suggest “The predicted emissions are able to capture …”.
• Line 99: Don’t you mean 2000-2020 instead of 2000-2025?
• Line 101: Include “modeled” before “dataset”.
• Line 115: Change “lags” into “a delay”.
• Line 122: This sentence has a lot of “ands”. Maybe “that regulate methane production, alongside with oxidation within the sediment and water column, to be eventually released to the atmosphere.
• Line 185: Change to “The TD inversion products”.
• Line 209: missing bracket after “index”
• Line 209: Please define LAI precisely before using this abbreviation.
• Line 223: Split into two sentences: “… transport. Their relevance …”
• Line 231: “XGBoost” is used here the first time. Please define it properly. Or move this whole part to Section 2.2.2.
• Line 235: Correct “section” to “Section”.
• Line 245: Please adjust Eq. 1 to a proper equation form (e.g., use mo_sin and mo_cos; use , etc.). Why aren’t you use just “m” for months, as it is more common in literature.
• Line 252: Change “of 1x20-15 kg CH4/m2/s“ to “≥ 10 - 15 kg CH₄ m^-2 s^-1”. And remove “or greater”.
• Line 264 + 265: Change “cover” to “covers”.
• Lines 286-292: Almost the exact same sentence is repeated twice.
• Line 294: change unit to “kg CH₄ m^-2 s^-1”.
• Line 295: “m²”
• Line 319: Include “them” after “compared”
• Line 320: Please state earlier in the manuscript, that you defined the 18 geographic regions similar to Saunois et al., 2025. Best when you mention them the first time.
• Line 332: Don’t you mean Figures 5, S6, S8?
• Line 398: Remove “this study”.
• Line 435: Make sure the caption of the Figure is on the same page than the Figure.
• Line 449: Aren’t you referring to Figures S6, S7 here?

Literature:

Zhu et al., 2025, DOI: 10.1088/1748-9326/adad02

Yuan et al., 2024, https://doi.org/10.1038/s41558-024-01933-3

Hasan et al., 2022, https://doi.org/10.3389/fclim.2022.1001174

Nisbet et al., 2023, https://doi.org/10.1029/2023GB007875

Stevenson et al., 2021, DOI: 10.5194/acp-2021-604

Ort et al., 2026, DOI: 10.22541/essoar.177248180.00166712/v1

Balasus et al., 2026, https://doi.org/10.5194/egusphere-2025-6251

The manuscript “Global natural wetland methane emissions (2000-2025)” uses a machine- learning emulator to estimate monthly methane emissions from global natural wetlands for the period 2000-2025. The emulator was trained on 35 GMB model estimates for 2000-2020 and the results show that the emulator, XGBoost models, reproduced both the magnitude and spatial pattern of the 2000-2020 GMB estimates very well. The XGBoost was able to extend the global emissions beyond 2020 to 2025 (2021-2025) with high confidence. The results from the machine-learning technique highlight multiple spatial and temporal patterns in the global methane emissions data. Since efforts to synthesize global methane emission estimates using BU and TD models lead to lags of multiple years, the paper proposes use of the XGBoost model to monitor global methane emissions from wetlands in almost "real time" and at lower computational costs.

I highly recommend it for publication in ESSD once the authors have addressed the following minor comments and suggestions.

Specific comments:

• Title: The title may be rephrased to “Global methane emissions from natural wetlands (2000-2025) using a machine-learning approach”
• Line 126: Insert “extent” in the sentence......Hydrology, often summarized by water table depth, soil moisture and inundation extent, controls oxygen...
• Line 127: Insert “CH₄” in the sentence .... thereby influencing wetland CH₄ production and oxidation (Cui et al., 2024; .........
• Lines 180-183: Could you please clarify on the advantages of the fact that all BU models compute wetland extent internally? Like you mentioned that contributes to inter-model spread. If a reliable dataset of the wetland extents, maybe from simulations by one specialized model, was used for all BU models, would that help to reduce the spread? Inaccurate representation of wetland extents has also been associated with failures of process-based models (Lin et al., 2024 - Nature Communications vol. 15: 10894).
• Line 252: Rewrite the mean flux of CH₄. It is not clear, but I believe it should read 1x10^-15 kg CH₄/m²/s. Is this value 30,000 or 300,000 times lower than the global mean of 157 Tg/yr?
• Line 271-273, Line 277, Line 290, Line 337-339: Can you clarify on which predictor variables were identified and retained as important after the Boruta feature screening? If the key predictor variables changed from one cell to another, how reproduceable is this step? Can the key predictor variables be grouped based on geographical regions (e.g., southern vs northern hemisphere) or latitudinal bands?  
• Line 342: The sentence “Mean R² is higher in regions with higher wetland CH₄ emissions....” sort of contradicts Line 340-341 because as stated here the northern hemisphere (NH) as a geographical region emits more CH₄ than the southern hemisphere (SH), yet the R² value is higher in the SH than in the NH.
• Line 364: Are the Sudd wetlands in South Sudan within the Congo Basin?
• Lines 459-466: Hydroclimatic drying conditions have been associated with regions that experienced significant decreases in emissions (see Lines 418-428, Lines 449-457), what prompted significant increases (in 9 regions) and no changes (in 4 regions) in emissions? Can this be made clear?
• Lines 494-499: The tropics show emission growth rate of almost zero but with high uncertainty. Is this high uncertainty, also observed using the machine-learning approach, associated with use of limited training data or poor understanding of key biogeochemical and hydroclimatic drivers such as inundation extents in tropical wetlands?

please see attached