The manuscript presents the China Building Rooftop Area (CBRA) dataset, which provides national-scale pixel-level information on individual building rooftop distribution and multi-annual dynamics from 2016 to 2021. The authors proposed an interesting and novel method for extracting high-resolution production of building rooftop from Sentinel images that could potentially reduce data acquisition costs. The study is well-structured, and the results demonstrate characteristics and superiority over previous production. The paper could be a valuable contribution to the society of urban remote sensing in terms of both the novel methodology and production. However, some revisions are necessary before accepting it for publication.

Major comments:

1. The one of the key contributions in this manuscript is the usage of a spatial generalization (SG) loss to increase the generalization capacity of a larger geographical region. However, as shown in figure 9, 12, 13, there exist some block-like result. It seems the SG loss could not perform well in these regions. Are they all caused by a resolution of less than 2.5 meters so the model can't distinguish them? The authors should provide further justification of the importance of the SG loss, including its superiority compared to directly utilizing cross-entropy.
2. While the SG loss might be reasonable for regions with missing high-resolution supervised information, the authors assign a "built" land cover type to each building rooftop reference as described in line 260, which might potentially confuse the original high-resolution information in my personal viewpoint. Therefore, a detailed description of the pipeline is required when both high-resolution and low-resolution references are available.
3. By comparing with the 90 cities BRA data, the authors claimed that the results of the 90 cities BRA data are spatially inconsistent than the CBRA. It is necessary to ensure that the supervised data is similar between the CBRA and the 90 cities BRA to rule out any differences that could be causing this inconsistency.
4. The paper highlights the limited availability of datasets covering the entire China. However, there exist some BRA datasets derived from sub-metric aerial images that cover specific regions. It would be valuable to show a comparison of the CBRA dataset with other small yet region-specific BRA datasets to gain more insights.
5. The multi-annual dynamic world product is presented as a probability map. Therefore, it is necessary to provide a more detailed technical description of the threshold, which might be used to binarize it in a meaning way.

Minor Comments:

1. The introduction and background sections have some repetitive descriptions. For example, the second paragraph could be shortened for there already has a detailed description in Section 2.2.
2. line 60: please provide the source of the statistics about the urbanization rate and the population structure of China.
3. line 159: “apple” should be “apply”
4. line 220: please unify the description of “arcgis”
5. Figure 5: what is the meaning of the four rectangle and arrow in subfigure (C)?
6. Figure 6: the direction of blue arrow and red arrow in the legend should be the same.
7. line 268: correct "red backward arrow" to "red arrow"

Dear authors,

The authors used the STSR-Seg method to develop a novel BRA dataset covering 2016-2021 with a temporal interval one-year, it has been demonstrated to have better performance than Zhang et al. (2022) results and covers the whole China (rural and urban areas), with an overall accuracy of 82.85%. The manuscript has been well-written and clearly organized. However, I still have several concerns about the current manuscript as follows:

1. The generating training samples should be greatly improved. For example, are the sample size of 200 and standard deviation of 150 reasonable? The land-cover label of each training coordinate came from which dataset (the spatial resolution of the coordinate represent 10 m spatial resolution?). As for the reference year, why you randomly assigned during 2016-2021? And you stated ‘we further rebalance the gathering samples to make sure that the built type is the majority by thresholding’, how to achieve the goal?

2. The authors used the spatiotemporal learning method to achieve the aim of downscaling and independently generate the CBRA map in each year, but the block effect is still obvious in your results. To further demonstrate the feasibility of the proposed method, can you use the Landsat image (30 m) as an example to generate the CBRA at 2.5? I believe it would make more sense (only a suggestion).

3. In section 4.4, the authors used a lot of thresholds (0.5 and 0.2) according to their prior knowledges, however, as for a national-scale mapping method, the authors should give the analysis of why these thresholds are reasonable.

4. In term of temporal-optimization, authors used the temporal checking method (proposed by Li et al. (2014)) for optimize the multitemporal CBAR results. Did the authors consider whether classification errors would be introduced into the optimized results?

5. As for the validation, I think the comparisons with impervious surfaces products (CLCD, GAIA) might not make sense. Instead, they should pay more attention on the comparison with CBA dataset and add more BAR dataset (such as regional dataset) as comparison dataset.

6. The quantitative analysis in Section 5.1.1 is interesting, can you give the reasons why the BAR dataset achieved higher accuracy in rural areas than that of the urbans.

7. The figure 10 illustrated that the CBAR dataset has obvious advantages than BRA dataset in their previous study. However, why the BRA has high geometry accuracy in top left corner of the Figure 10c and suffers obvious offset in the central areas (red circle) over such local area.

8. The temporal analysis in Section 5.2 should be strengthen, as the increased/decreased BRA is great small than the stable BRAs, combining the stable and changed BRAs for analyzing the accuracy metrics cannot illustrate the performance of proposed method in the temporal dimension. For example, you can use the changed validation points to analyze the changed CBRAs.

9. Figure 18 is interesting, and two enlargements intuitively show the good performance of the CBAR. I suggest the authors added the high-resolution imagery over two enlargements to make the analysis more intuitive. In addition, why the North China Plain (especially in Henan province) shows such a marked increase? An enlargement in Henan should be added.

10. In Line 383, the Radoux et al. (2014) mainly emphasized the spatial heterogeneity, while this part focuses on the temporal consistency, so the reference might be incorrect.

The authors describe an effort to create building footprint data for all of China. Their dataset is a raster dataset at 2.5m resolution, derived from 10m Sentinel-2 data. They use a deep learning approach involving super-resolution segmentation, allowing to downscale the information from 10m to 2.5m resolution.

The contribution is timely and very relevant, as it tackles several gaps in the global data landscape on human settlements and built-up areas: 1) The created data covers China (including its rural areas), unlike other data products; 2) The dataset is multitemporal (2016-2021) which is rare, allowing for the assessment of built-up growth and shrinkage due to demolition etc.

The paper is well-written and structured. It is very detailed and includes a thorough accuracy assessment against other datasets including a comparison to global datasets available at lower spatial resolution, including datasets from different sources, and also involves hand-crafted validation data and manual checks. The obtained accuracy estimates are quite high and promising.

As I cannot judge the quality of the deep learning framework, I have mostly minor comments, as well as some comments on the data themselves and a request for clarification on the accuracy assessment.

Comments on the data:

• Empty raster datasets such as CBRA_2016_E113.5_N51.3.tif or CBRA_2016_E76.0_N33.8.tif should be excluded from the dataset.
• Until looking at the data, I was very positive towards this manuscript. However, I then had a look at a small selection of the data, and was quite surprised to see very coarse “blobs” delineating settlement areas, rather than mapping “rooftop areas” as the dataset suggests (example in the figure below). I zoomed into 3-4 regions, and most seem to be finer-grained than these blobs and actually delineating individual buildings / rooftops. However, the authors should be transparent and also show such an example in their manuscript, to highlight that the method does not seem to work well everywhere – and possibly provide an explanation for this. Looking at this specific example, I don’t think it is defendable to call this “rooftop areas” – this is a quite generalized settlement area, slightly more refined than the GHS-BUILT 10m dataset, shown for comparison.

• Another comment on the data: I would anticipate much wider usage of the data if they were provided as vector data (i.e. polygon objects describing each building) rather than raster data. The fact that the authors provide 2.5m-raster data still leave a major chunk of processing work to the user. While there are applications where the fine-grained raster data is useful, most applications will be based on vector data. The authors correctly mention the vectorization step as “future work”, but I would like to raise the discussion here if it would be beneficial to do this at this point, or otherwise provide a vectorized version of the data in the near future – just as some “food for thought”.

Accuracy assessment, comparison:

Table 4: Why is there only recall reported for the rural scenes, whereas for the urban scenes you report recall, F1, Iou, OA? And why you do not report Precision in both cases? This need to be done and is standard for an accuracy assessment. Of course the reader could calculate the precision based on recall and F1, but please provide Precision, recall, OA, F1, IoU for both the rural and the urban scenario. No rationale is provided for only reporting recall in the rural scenario.

Moreover, it is unclear how the accuracy estimates for the Global Human Settlement Layer (GHSL) as reported in Fig. 4 were produced. Which of 10m GHS-BUILT data products was used? There is either the GHS_BUILT_S_E2018_GLOBE_R2022A_54009_10_V1_0 dataset, or the GHS-BUILT-S2 dataset. Both are continuous, with the former reporting the 10m built-up fraction, and the latter reporting the built-up probability. Please provide the following information: Which of the datasets was used? And how were these continuous data thresholded in order to carry out a binary (2-class) agreement assessment? I.e., what cut-off value was used, and how was this cut-off value derived?

The observed accuracy drop from urban towards rural is typical for settlement mapping, please place your work in the context of the literature, e.g. by citing Leyk et al. 2018, or Kaim et al. 2022.

The authors use the overall accuracy for their accuracy assessment. However, it is well-known that OA yields biased results in the case of imbalanced class distributions (see Shao et al. 2019, Uhl & Leyk 2022) for a recent in-depth study. Such class imbalance is typically the case for built-up vs not built-up assessments, in particular in rural areas. Under the light of this potential bias, please add some sentences critically evaluation the magnitude of the OA values obtained. That being said, I appreciate the authors also report IoU and F-1.

Fig. 16: Legend should be swapped – the blue should be on the left, and red on the right, also in Fig. 18a.

Fig. 17 b and c: I don’t understand what is the difference between panel b) and c), besides the different visualization technique. Please clarify. Moreover, I don’t think the pie charts are a good choice here. They don’t show the change over time. Please use a layer plot for b) just, as you did  in c).

Fig. 18: the green color used to show demolition is different in the map and in the legend.

Fig. A2, caption: Figure A2: The probability density distribution. …. Of what???

Minor comments:

Please provide a rationale for using the term “rooftop area” instead of “building footprint area”  or “built-up area”.

Line 75: What means “and F1 score of 2.5 m,” …. I don’t understand what the authors mean here.

Line 106: No need to define an acronym for state-of-the-art; the term is only used twice in the paper.

Line 159: “apple” ?

Line 161: Please explain what you mean by “geographical offset”.

155-165: nice transition and justification for the contribution of the paper.

Table 1. Nice overview on existing datasets.

Fig. 2: Please include the GHS-BUILT dataset here, from the Global Human Settlement Layer, e.g., the GHS-BUILT-10m built up layer. This will provide a nice overview on recent work at a global scale, and highlight the merit of your work.

Caption Fig. 2: “cloud under” change to “cloud cover under”

Fig. 4: Text is very small, please increase font size, and decrease spacing between lines; this way, the space can be used more efficiently.

Fig. 9 – caption: “Comparison of the CBRA and the other dataset” – please name the “other dataset”.

References:

Leyk, S., Uhl, J. H., Balk, D., & Jones, B. (2018). Assessing the accuracy of multi-temporal built-up land layers across rural-urban trajectories in the United States. Remote sensing of environment, 204, 898-917.

Uhl, J. H., & Leyk, S. (2022). A scale-sensitive framework for the spatially explicit accuracy assessment of binary built-up surface layers. Remote Sensing of Environment, 279, 113117.

Kaim, D., Ziółkowska, E., Grădinaru, S. R., & Pazúr, R. (2022). Assessing the suitability of urban-oriented land cover products for mapping rural settlements. International Journal of Geographical Information Science, 36(12), 2412-2426.

Shao, G., Tang, L., & Liao, J. (2019). Overselling overall map accuracy misinforms about research reliability. Landscape Ecology, 34, 2487-2492.