This study develops a time-series 30 m building height dataset for China by integrating multi-source remote sensing data, representing a valuable contribution to large-scale mapping of urban vertical dynamics. The dataset has clear potential for urban studies and related applications. However, several issues related to data quality, robustness, and validation—particularly for high-rise buildings and regions with limited reference data—still need to be carefully addressed to strengthen the reliability and usability of the product. The following comments are provided for the authors’ consideration.

1. In Eq. (6), averaging all 1 m pixels (900 pixels) within a 30 m grid mixes building and non-building pixels. Please clarify the rationale for this choice, rather than computing height only over building-covered pixels.
2. The model performs well in cities with available building height samples (e.g., Beijing and Tianjin). However, in cities without reference height samples (Fig. 14), the heights of mid-rise and high-rise buildings are significantly underestimated. This indicates limited spatial transferability and cross-city robustness of the model. Please address this limitation, and explicitly discuss the applicability and associated uncertainty in such regions.
3. The model shows limited capability in predicting high-rise buildings. Please clarify the maximum building height represented in the training samples. Moreover, this limitation may be related to the uneven distribution of building height samples. In that case, more appropriate sample setting strategies or model optimization schemes could be considered. In addition, the height bins in Fig. 4 (e.g., “<10 m” and “>60 m”) should be more clearly defined (e.g., “3–10 m”, “60–X m”).
4. In Fig. 7, the reference building heights are mainly concentrated within 0–8 m. Pixels with height values below 3 m are likely non-building pixels and could be excluded from validation. Moreover, reporting relative error metrics (e.g., rRMSE) would provide a more informative evaluation.
5. In building change areas, the accuracy of estimated height changes should be quantitatively assessed to evaluate whether the model can reliably capture building height dynamics. In stable areas, the temporal consistency of predicted building heights should be quantitatively examined to assess prediction stability.
6. For dataset comparison, it would be more appropriate to focus on comparable 30 m resolution multi-temporal or time-series building height datasets, such as He et al. (2023) and Chen et al. (2025).
7. Building height results are visualized using both classified and continuous color schemes, with inconsistent value ranges across figures (Figure 11, Figure 13, Figure 14, Figure 15). Adopting a unified classified visualization scheme would improve clarity and consistency.
8. The comparison presented in Fig. 10 raises several concerns.

(1) Different building height products adopt different height definitions (e.g., average height, area-weighted average height), and these definitional differences should be explicitly considered. In addition, it is worth noting that some gridded building height products may include only building information, which appears to differ from the definition adopted in this study. As a result, direct resampling and comparison across products may not be strictly comparable.

(2) The reported RMSE of 748.79 m is physically implausible and requires clarification.

(3) Please note that it is unclear whether the same reference building height data were used consistently across Fig. 10b–e. Reference heights above 20 m appear in some panels (b–d) but not in others (e), and zero-height pixels are shown for Ma’s dataset, although such values do not appear in Ma’s original released data.

(4) It is unclear whether this part of the validation is limited to buildings with heights below 45 m.

Overall, Section 4.4 requires careful re-examination.

1. More detailed descriptions of the accuracy assessment and dataset comparison procedures are required, for instance, whether the validation is conducted at the pixel level or using object-based approaches based on building footprints (with height).
2. If possible, providing spatial uncertainty or confidence layers would substantially enhance the usability of the dataset.
3. Please carefully check the reference formatting. The citation of Che et al. (2024) should refer to the final published version, rather than the Earth System Science Data Discussions version.

Overall, this is a valuable and timely study. Long-term, annual building-height reconstruction remains an important gap in urban remote sensing, especially at fine spatial resolution and national scale. The proposed dataset has clear potential to support studies of three-dimensional urban growth and renewal in China. Nevertheless, several methodological and interpretive issues still deserve further clarification. Addressing these points would improve the robustness, transparency, and broader usability of the dataset.

1. On Page 9, Lines 205–210 and Page 10, Lines 217–219, the manuscript states that reconstructed 1 km VV backscatter is used as an independent variable for XGB-ReVV during 1990–2014. Since this VV layer is reconstructed from Landsat-derived variables and terrain/DSM information, its uncertainty may be propagated into the subsequent 30 m building-height prediction. Although the authors report satisfactory performance of the reconstruction model in the Beijing–Tianjin–Hebei region, it remains unclear whether this performance can be generalized to the whole of China, particularly across regions with different urban forms, climate conditions, and surface characteristics. I suggest that the authors add a short discussion on the robustness and potential error propagation of the reconstructed VV variable, and clarify the extent to which the regional validation supports its nationwide application.
2. On Page 7, Lines 148–151, the 2019 reference height data are derived from Baidu Maps building footprints, where floor counts are converted to height using a simple assumption of 3 m per floor. Although this approach is supported by Liu et al. (2021), the cited validation was based on a limited building-height assessment of only 519 buildings in Shenyang, with a reported mean height deviation of approximately 1 m and an accuracy of 86.8%. This raises the question of whether the same level of accuracy can be assumed for nationwide and multi-temporal building-height mapping. In addition, Section 4.1 would benefit from a clearer explanation of how the 2019 reference heights are transferred backward to 1990. In particular, if a pixel experienced in-situ height changes over the past three decades, such as demolition followed by reconstruction of a taller building, it would be helpful to clarify how such cases are handled during annual reference sample generation.
3. On Page 9, Lines 195–198, the authors state that applying the 1999 CCDC mask uniformly to 1990–1999 has “negligible effects” on reference data accuracy. However, Page 13, Lines 271–274 show that the validation accuracy of annual reference samples remains only around 65%–70% during 1990–2002. Therefore, I suggest replacing “negligible effects” with a more cautious statement, such as “limited but non-negligible uncertainty may remain for the pre-2000 period.” The authors may also clarify in Section 4.1 or Section 6 that early-year estimates are more suitable for regional-scale or trend-level analyses, rather than for strong interpretation of pixel-level inter-annual changes.
4. On Page 10–11, Lines 225–228, the manuscript states that the annual reference height samples are randomly split into 90% training and 10% testing sets. Since both subsets are derived from the same CCDC-filtered and 2019-reference-based sample generation procedure, this evaluation is better interpreted as testing the consistency of model fitting against CCDC-derived reference labels, rather than as a fully independent validation of spatial or temporal generalization. I suggest that the authors make this distinction clearer when presenting the accuracy results, so that the reported RMSE, MAE, and R² are not over-interpreted as completely independent validation metrics.
5. Page 16, Lines 335–342 show that latitude and longitude are the most important predictors, even exceeding most optical and SAR variables. Meanwhile, Page 15–16, Lines 318–326 report clear regional discrepancies in model performance, especially poorer performance in parts of southern China and the Pearl River Delta. This suggests that the model may partly rely on regional urban-form priors rather than only pixel-level remote-sensing signals. I recommend that the authors add a short discussion on whether the strong importance of geographic coordinates helps capture spatial heterogeneity, or whether it may introduce risks when extrapolating to regions or cities with limited representation in the training samples.
6. Several numerical values should be checked for consistency. The Abstract reports an RMSE range of 5.96–6.69 m, whereas Page 13, Lines 283–284 report 5.94–6.69 m, and Page 29, Lines 490–491 again report 5.96–6.69 m. Similarly, Page 27, Line 460 reports the 2019 national building volume as 884.69 km³, while Page 27, Lines 463–464 report 884.49 km³. I suggest that the authors harmonize these values throughout the manuscript to avoid giving the impression of inconsistent data versions or calculation procedures.