Synopsis

The manuscript describes the creation of a new land cover classification system designed specifically for quantifying methane emissions from Arctic and Boreal regions. This science has been plagued by the complexity of methane biogeochemistry and the landscape heterogeneity of the region. Thus, estimates of northern high-latitude emissions carry large uncertainty. The authors have somewhat reinvented the wheel by designing a new model grid by combining expert knowledge and machine learning-based techniques to estimate fractional cover of 19 classes within a half-degree grid. They use schemes to harmonize land cover types to reduce double counting issues with bottom up emission budgeting. The large collaborative effort demonstrates that this technique could be widely accepted and highly useful for tracking methane emissions from the rapidly changing north. The article is well written and very near its final form. I can only recommend very minor changes to help improve the clarity and impact of the manuscript. 

General Comment

This manuscript was written as a companion to Kuhn et al. 2021 (BAWLD-CH4: Methane Fluxes from Boreal and Arctic Ecosystems), which contains more quantitative methane information for the BAWLD model. Kuhn et al. 2021 is heavily cited in this manuscript, however a better connection to this paper could be made in the introduction and/or the conclusion. This would increase the impact and utility of both papers. A paragraph could be warranted to more clearly and explicityly illustrate/bridge how the two papers are companions to one another. This could also potentially improve on the lack of quantitative methane information in this manuscript. I was somewhat expecting a new pan-arctic/boreal annual CH₄ estimate based on the BAWLD modeling in either this manuscript or Kuhn et al. 2021, however no global estimate was produced. This seems to be the potential a-priori gridded data that was suggested as highly useful to the inverse modelling community.

Specific Comments

Lines 181-184: Are the 53 variables available in all grid-cells? Do some grid cells contain more/less variables? Probably best to specify/clarify in the text.

Section 2.2.1: Could be really helpful to quickly provide some well-known real geographic examples for some of the main wetland classes.

Line 280: Add comma after “As such”

Line 356: Add an “a” in between “have” and “high”

Line 442: Suggest changing the comma to a period and starting a new sentence with “We henceforth…”

Figure 6: Minor suggestion: add the word “Legend” to the color key to quickly differentiate this wheel from the others. Maybe in the center? Bold? Maybe with a box around it? Take it or leave it, but it is confusing at first. At first glance, I interpreted the legend as a global distribution.

In this manuscript, the authors presented a new dataset (BAWLD) of wetlands, lakes, rivers, and other land-cover types for the boreal and arctic areas. Although many land cover data have been proposed, I agree that this dataset has advantages in its comprehensiveness and expert assessment. Namely, this dataset would surely contribute to improve accuracy of methane emissions from this region, especially in terms of separation of wetland and freshwater sources. Therefore, I found enough merits to publish this manuscript.

              I have two minor caveats. First, I could not understand the reason why the authors chose the spatial resolution of 0.5 degree? I know this resolution has been standard for global terrestrial models, but it may be difficult to capture spatial heterogeneity due to topography and micro hydrometeorology in this area. Indeed, several land-cover maps such as GLC2000 (used as an input data of random forest model in this study) have a spatial resolution of about 1km. One possible option may be to provide several data files with different spatial resolutions: e.g., 1km as a full resolution and 0.5 degree as an aggregated resolution. Second, I know that the Global Lake and Wetland Dataset (GLWD, Lehner and Döll 2004), which contains multiple types of wetlands and lakes, has been used in several studies. However, the authors rarely mentioned about this dataset and used it only for river detection. For example in Fig. 4 and 4S, the authors did not include the GLWD into their inter-data comparison. I recommend making a comparison or discussion with the GLWD (and other data, if necessary) to clarify the advantage of the BAWLD. For example, the explicit separation of characteristic types such as permafrost wetlands and yedoma lakes look a clear advantage for data user working in this area.

              The manuscript gave full descriptions of the dataset. Although descriptions of individual wetland and lake types look lengthy, it may be useful for data users. Similarly, the authors provided a plenty of figures and descriptions as the supplementary file. The wetscape, derived from the wetland and lake data, can be excessive and unnecessary, but I agree that it is implicative. Finally, I recommend the manuscript is acceptable after minor revisions.

Technical points

Page 19 Line 13: Please note that Bohn et al. (2015) conducted a model intercomparison study on CH4 emissions in the West Siberia Lowland including the Ob River floodplains.

Page 30 Line 748: Several records in References lack the information on journal name. For example, Bastviken et al. (2004) was published from Global Biogeochemical Cycles. Please check also other records.

The manuscript addresses the very important topic of estimating the extent of different wetland types in northern latitude regions. As the authors mention, these regions are particularly affected by climate change and accurately estimating the extent of wetland types can help reduce the uncertainty associated with methane emissions, which is currently very high. In summary, the authors produce a state-of-the-art dataset that can be readily used by experimental scientists and modelers from various disciplines. The paper is also well structured and overall easy to follow. I recommend publication after minor revisions, especially given the impact of the work.

The only major point that I suggest the authors is to discuss more thoroughly is the spatial resolution of the dataset. 0.5x0.5 is relatively coarse, while land surface models are now moving towards much finer resolutions. So, I think it would be beneficial to discuss more what were the limiting factors for having to work with this resolution. This might inform next steps that can be taken in order one day have a finer resolution gridded dataset.

I also suggest expanding Table 1. More information could be included in this Table, such as the spatial and temporal resolution of the data sources.

One aspect that could be made clearer is how the expert assessment was integrated with the overall procedure to derive the dataset. How was this information used? For example, it was not clear (at least to me) whether this information was integrated into the modeling or not. The authors could expand a bit section 2.3 and provide more details in the introduction.

Lastly, I wonder what the authors think of other machine learning approaches that can take advantage of relatively high-resolution satellite images (computer visions tasks using convolution neural networks) and whether these approaches might prove useful to improve spatial resolution and details of BAWLD or similar databases.