The authors created a new 10-m land cover map for the circumpolar Arctic based on the combined use of Sentinel-2, Sentinel-2, and DEM and multiple-source training samples. It is challenging to map land cover over the Arctic due to the limited number of snow/ice-free observations. Yet, the current manuscript does not well demonstrate the land cover product is promising, particularly for the validation and the comparison. I have also lots of questions regarding the method. Please see my detailed comments below.

(1) All the samples filtered by the SOM may be helpful to improve the classification, but it will select similar samples for each category. The validation data, with the same resources as training data, will overestimate the accuracy of the land cover map. The samples from flux tower sites may be OK, because of the real independence to the training samples, but the number is too limited, and serval important land cover types are lacking, such as cropland, forest, snow/ice (Table S2). Here I highly recommend the authors create new and independent validation samples stratified on the Arctic land cover map (see Olofsson et al. 2014, RSE, good practices).

(2) The description of the sample is not clear, i.e., how many samples were in the study area (L 133), how many changed pixels were removed (L 134), how did the authors determine the land over type for the pixels within the 90-m buffer (L 135), what was the strategy (L 139), etc. The strategy has a 40% maximum proportion and a 3% minimum proportion during the training data selection. How did the authors apply the strategy to “each” used land cover product, such as NLCD, Canada land cover, and GlobeLand30? This strategy was originally designed for selecting the training samples, that will be directly used to train the final random forest model; however, when the lichen/moss class was interpreted, how did the authors know its proportion? By the way, how did the authors immigrate the category of the existing products to Table 1, which are so different?

(3) The approach of comparing the new map to existing land products is not objective. For example, why did the authors focus on the agreement between the new map and the existing products, rather than focusing on the accuracy? How did the authors align the different definitions of land cover between products? What are the means for DM, DS, and DD? What this comparison analysis delivered to audiences is a huge disagreement among the maps, and so what? How about the accuracy of each product? At least, the authors can provide the accuracy against the samples from flux tower sites.

Other comments:

L 110: What were discarded from the Sentinel-2 QA band? Snow/ice observations were excluded?

L 157: The number of each category should be given as well.

Section 2.4.2 Man-made impervious surface mapping: As far as I understand, the impervious surface from CAMI-2020 has the highest priority, and the authors classified the other 9 land cover types only. If that, I suggest clarifying that the classification method did not include the impervious surface, like in Section 2.4.3.

The manuscript combined the multisource and multitemporal remote sensing imagery to develop the new circumpolar Arctic land cover product for circa 2020, containing 10 land-cover types especially for the tundra. The dataset is great and can be used to improve earth system modelling. However, I think there are several problems should be solved.

1. In the second paragraph of the introduction, author stated that ‘but their strength is limited by the over simplification of classification schemes’ and used the MCD12Q1 to explain the limitations of global land-cover products. In my opinion, the statement might be inaccurate, the ESA CCI_LC, GlobeLand30, and three global 10 m land-cover products (ESA, ESRI and dynamicworld) can capture the land-cover information with various classification systems or spatial resolutions. Therefore, I suggested that the authors further added the necessity of developing CALC-2020 when there were so many global land-cover products.
2. In the third paragraph of the introduction, the reasons why authors used multi-source remote sensing data to develop the CALC-2020 map was not clear. Actually, the combination of multisource remote sensing observations has been demonstrated to improve the land-cover classification in so many studies because of importing new information. Namely, the combination of optical, SAR and terrain features was not particularly novel in this study.
3. Authors used the FAST sample library and three land-cover products to derive the primary training samples and then used visual interpretation to collect lichen and moss, however, how to identify the graminoid tundra and shrub tundra samples, which was not explained in the method.
4. Authors used the SOM method and further used the threshold of 75% to remove the wrong samples, why chose this threshold? Can you give the number of training samples for each land-cover class?
5. In section 2.4.3, authors mentioned that “Cloud-free Sentinel-2 observations were first interpolated in each pixel at an 8-day time step using penalized cubic smoothing splines”, is it mean these cloudy or snow pixels were interpolated by cubic splines method? Can you provide the details and the results of the time-series sentinel-2 observations after interpolating?
6. The CALC-2020 was individually produced for each country using RFC models, how to ensure the spatial continuity between the transition areas over two neighboring countries because of using different RFC models?
7. The comparisons between CALC-2020 with three global maps were unfair for global land-cover products, because the CALC-2020 only focused on the circumpolar Arctic while the global land-cover maps emphasized the comprehensive performance over the globe.
8. In the two comparison figures of Figure 6 and 7, I doubt the difference mainly came from the different definition for the same land-cover class especially in the Figure 7.
9. I suggest to remove the statistics figure in Figure 9b, because it cannot provide useful information for the CALC-2020 map.
10. Figure 10 about the feature importance is interesting and vital for developing the CALC-2020, I think the author should explain in more details why the topography had highest importance while the optical features were unimportant.
11. I suggest that authors shared the validation data in the Data availability

              Line 334, Figure 30 should be 10.