This manuscript proposed an efficient method to produce the 10-m global impervious surface areas (GISA-10m) based on the existing ISA maps, Sentinel-1/2 images, and OSM data. Compared to existing global GISA products, GISA-10m can provide higher spatial resolution while keeping higher accuracy. The inter-comparison with existing datasets demonstrated the superiority of GISA-10m. Analysis of ISA on rural and urban areas further revealed the urbanization level and landscape of different countries in more details. In particular, an interesting point of GISA-10m is that it is able to delineate the area of roads across the world, making GISA-10m valuable for relevant urban studies. In general, this manuscript is well presented and makes novel contributions. However, some issues should be clarified to improve this manuscript. Specific comments include the following aspects:

1) In Section 3.1.3, the authors used 200 trees for training the random forest classifier, while the effect of the number of trees is not analyzed. Besides, the key parameter, e.g., the number of features used for training each tree, is not clarified. Please provide this information for better understanding.

2) Line 90: do you mean by "operating by"?

3) Line 103: relevant reference should be provided to support “the terrain distortion caused by the combination of two orbits”.

4) L120: it should be "Landsat 8".

5) L121: I found both "GLCFCS" and "GLC_FCS" in the manuscript. Please explain.

6) Line 201: the original OSM data are provided in vector form. When this data was converted to 10-m raster, whether the majority rule was applied? The majority rule refers to “a pixel (10m × 10m) was labelled as ISA if more than half of its area was cover by ISA, otherwise it was identified as NISA”. Please clarified this issue.

7) L295: why the total number of visually interpreted samples was 10800 when 200 samples were selected in 59 grids? Please check.

8) Section 3.3: it is better to move this section to Section 5.1, since the detailed discussion has been presented in Section 5.1.

9) Figure 6: is it possible to compare the continental accuracies of other datasets presented in Table 3? The comparison at continental level may give a clear difference of different datasets.

10) Figure 8: not much information in it.

11) Figure 11: it is better to use (a), (b), (c) to distinguish each subgraph.

12) Figure 17: this figure is not clear enough for presenting 30 grids. It is suggested to add legend and put this figure to the supplementary materials.

13) Line 372: “extracted” or “detected”?

14) Table 9: whether test samples used in Table 9 are from visually interpreted samples? Please clarify this.

The mapping of 10-m impervious surfaces at the global scale using multiple geodata sources is interesting. The authors applied temporal-spatial-spectral-geometrical rules to generate samples, and validation of the results is comprehensive and adequate. They also attempted to delineate the spatial distribution of impervious surface in urban and non-urban areas. The manuscript fits the journal's scope and the dataset is valuable, which is suitable for publication in ESSD. However, the paper still has some flaws (see my comments below) which should be further clarified or discussed before acceptance.

My major concern lies in the completeness and correctness of the OSM data. How about the effect of the geographic bias in spatial distribution of OSM data? More analysis is needed to discuss this issue.

Line 10: “global ISA mapping” should be “global ISA datasets”

Line 21: “refined OSM data” -> “OSM data”.

Line 80-85: The GISA-10m dataset attempted to further delineate road regions from the ISA. This should be mentioned in the introduction and abstract.

Line 152: “multiple sources” is not clear, and can be modified as “multi-source datasets”.

Figure 1. It would be better to label each step, e.g., "Step 1. Training sample generation".

Line 157. The authors selected the GlobeLand30 in 2010 but chosed other data (e.g., GISA and FROM-GLC) in 2016. Would the temporal gap between these data impact the quality of training data?

Line 171. How did you define edge pixels? I think the edge pixels are different between 30-m and 10-m images, as a non-edge pixel in a 30m image may be edge pixels in a 10m image. Could you clarify this issue?

Line 197. Why buildings with area less than 100 m2 were excluded?

Line 210. Why did the authors remove the OSM samples intersected with those from other global datasets?

Line 235. Please explain why these features were chosen.

Line 286. How many RF models were built?

Line 293. How did the authors select the ISA test points? If the points were mostly located in urban areas, it might bias the assessment result. Could you provide the ISA density around these ISA points?

Figure 9. It's interesting to see the accuracy in rural and arid areas. How about urban areas?

Line 379. How did you divide the rural and urban areas?

Line 380. What do you mean by Global ISA?

Figure 14. The title of subgraph seems incorrect.

Figures 16 and 17 may be moved to the supplements.

Line 523. "difference" or " differences"

Line 500: “distinguish well ISA from NISA” -> “distinguish ISA from NISA effectively”