This paper produces a global building height dataset, which represents a substantial workload, and the dataset itself holds significant value.

My key concerns are: (1) the methodological innovation of this work can be better clarified; and (2) given that this data set completely relies on commercial satellite data, whether updates can be sustained in the future and whether other researchers can replicate the results remain unclear.

Page 4, Line 101: You mention that current raster-scale building height data often suffer from low resolution and poor quality. However, state-of-the-art raster-scale building height data can already achieve resolutions of 2.5 meters (e.g., Cao et al. A deep learning-based super-resolution method for building height estimation at 2.5 m spatial resolution in the Northern Hemisphere), and many of them, after contour optimization, rival instance-level products in structural detail. Compared to these studies, where does the advantage of your dataset lie?

Page 6, Line 130: What type of LiDAR did you use—airborne or spaceborne? Which countries are covered? This information is essential, even if it appears in supplementary materials. Furthermore, Figure 1 shows no 3D labels for Africa or South America. Did you rely on training from other continents and generalize to these area? How was the accuracy of this extrapolation validated?

Section 3.1: PSR is a commercial satellite, can other researchers replicate this study at low cost or update the results in the future?

Section 4.3.1: When constructing the training set, the access time of all labels and imagery should be emphasized.

Page 6, Line 130: You mention adding an extra FCN head. From which layer of UperNet is this FCN head connected—PPM block or fused layer? Were ablation studies conducted to demonstrate accuracy improvements? Otherwise, the addition of an FCN head seems arbitrary, especially since the pyramid structure already captures deep semantic information. Furthermore, if an additional head for supervision is necessary, FCN is an overly simplistic choice.

Figure2 Why were two separate models chosen for height and contour estimation? Numerous studies have shown that joint training is more efficient and improves accuracy.

Page 8, Line 180: You trained the model using labels with randomly added noise, but the patterns of this artificial noise may differ significantly from the actual noise introduced by the model itself. How effective is the resulting denoising model in practice? Furthermore, since there are already superior denoising models available—such as those based on adversarial learning or diffusion models—wouldn't these approaches be more suitable alternatives?

Page 9, Line 205: According to Figure 1, 3d samples from mainland China appear extremely limited (only three cities). Moreover, the data source is not specified. To my knowledge, building height labels in China typically only include floor counts, not precise meter-level measurements. What specific processing was applied?

Page 10, Line 251: Additional buildings in auxiliary data could easily be false positives. While you attempt to remove false positives in your own data, how do you ensure that false positives in auxiliary sources do not compromise the final results?

Page 11, Line 277: Asia has the largest number of buildings, yet according to the paper and Figure 1, it has the fewest 2D and 3D labels. Could this affect model performance in Asia?

Table 3: Instance-level RMSE should theoretically be much lower than 3-meter raster resolution, yet anomalies appear in Asia, Africa, and South America. Please explain. The RMSE for building height estimation in Oceania is reported as merely 1.5 meters. However, existing studies suggest that current building label data itself contains inherent inaccuracies. If this 1.5-meter error is potentially smaller than the intrinsic error of the reference labels themselves, the validity and meaningfulness of such accuracy evaluation become scientifically questionable.

The study provided high-quality, consistent, and global building data in 2D and 3D form, which are helpful to the urban management and Sustainable Development Goal. The proposed method is innovative and dataset is of high accuracy. However, there are still some problems that deserve to solve before publications.

(1) In the related work part, the author mentioned more details about data and methods used for building footprint and height estimation, including SAR/InSAR data, deep learning methods.

(2) More details about the PSR data should be introduces in section 3.1, including acquisition time, data quantity, etc. The detailed description of LiDAR data used in this study is lacking. Is it the satellite-based data or the ground-based data?

(3) To reduce the calculation amount, the author could extract areas with building and remove areas without building in advance.

(4) The LiDAR data, satellite imageries and extracted polygons may have some biases. The author could introduce more about rectification and the elimination of these biases. Meanwhile, I wonder if there are LiDAR data, why should the author use the deep learning method?

(5) The author presented the estimation error of the building height and volume. I wonder if the accuracy of building footprint in this study can be estimated.

(6) There are many data in the section 5.1 and the author should consider their relations. For example, the RMSE of the building height and building volume are 8.9m and 586.8m3/m2. For simple calculation, the estimation error of the building footprint is about 66m2 per 100m2.

(7) The author should simply discuss the reason for the performance difference of the proposed method in different continents.

(8) The section 5.4 and section 6 could be merged in to a new discussion part to improve the readability. The high correlation with grid population data can be seen as a kind of precision verification.

(9) There are still some grammatical and lingual problems, and authors should make a thorough revision.