He et al. developed a machine learning model to estimate gridded ozone data using meteorological parameters, pollutant variables, geographical covariates, temporal dummy variables, and ground station observation data. They conducted the modeling based on data after 2014 and extended the ozone estimates to the period 2000-2013, ultimately generating seamless ozone data for 2000-2020. Detailed exposure risk analyses were also conducted. Given the persistently high ozone concentrations in China in recent years, the data and analysis results presented in this paper are of value for decision-making. Overall, the paper generates long-term ozone data, the methods and framework employed are reasonable, and the results and analysis offer some insightful takeaways, but there are still several issues that need to be addressed:

1. The authors emphasize the importance and contribution of surface temperature as a proxy. However, two aspects warrant further exploration:
   
   a) The conclusion that surface temperature is important is derived from the model's variable importance analysis. What would happen if surface temperature were removed from the model? Would the overall modeling accuracy decrease, and by how much? Additionally, how would the spatial mapping results be affected in the absence of high-resolution surface temperature data? A visual example to illustrate this would be helpful.
   
   b) We are aware of global warming. If surface temperature plays a decisive role in ozone estimation, this may be suitable for the integration between ozone and surface temperature observation, but is it equally applicable for hindcasting? Further discussion on this point could enhance the scientific value of the paper.
2. The paper estimates historical ozone data over an extended period, which is difficult to validate. While observation data from Hong Kong are scarce, they provide valuable validation. However, this is not clearly explained in the paper. How many observation sites in Hong Kong were used, and what time periods do the data cover? This information is crucial for assessing the accuracy of the historical ozone estimates. A more in-depth analysis would be beneficial. For example, if a few years of data are available, how does the accuracy vary year by year?
3. The time periods are divided into 2001-2007, 2008-2015, and 2016-2020 for trend analysis. Were these periods defined based on trend identification, or was the segmentation arbitrary? This needs to be clarified in the paper.
4. The paper presents results across multiple scales (e.g., pixel, county, region, national), offering a comprehensive view of the analysis from different perspectives. Figure 4 presents fine-grained county-level analysis, but it is unclear what new insights this scale of analysis provides. It seems neither as detailed as pixel-level analysis nor as regionally distinctive as the regional-scale analysis. It would be helpful to clearly state the key conclusions derived from this level of analysis.
5. The exposure levels for the period 2016-2020 show a significant increase in the NCP region, especially compared to 2001-2007. What is the underlying cause of this increase? Additionally, the PRD has long been considered a high-concentration ozone area, but this does not seem to be reflected in the results.
6. What are the physical or chemical mechanisms through which aerosol optical depth is used to estimate near-surface ozone? Please provide further explanation.
7. Lines 155-167, does "province level" refer to statistical analysis by province, rather than validation by province?
8. Figure 2 compares the results obtained in June 2018 with those from previous studies. Why was this particular time chosen for comparison?
9. Line 231, a writing mistake, "NCP)" should be "NCP."
10. The discussion section summarizes many of the paper's findings and analyses, a more discussion of uncertainties and insights regarding ozone pollution control would be beneficial.

The study analyzed the long-term and short-term characteristics of surface ozone spatialtemporal variation and exposure in 2000-2020, which provides important support for environmental management and health research. This study used a machine learning method based on satellite retrieved surface temperature, combined with multi-scale and multi-temporal dimensions analysis, to reveal the spatial and temporal distribution of ozone and its potential health risks. This detailed and comprehensive study have high scientific value, and provide a reference basis for policy making in related fields. However, there are still some aspects that need to be further improved.

Major:

1、The background of the article is not well researched, and the introduction only briefly mentions some O₃ data at high spatial resolution for O₃. However, there are already studies (Shang et al., 2024. https://doi.org/10.1021/acs.estlett.4c00106) where hourly O₃ data are available, and the advantages of your research over these studies should be fully explained.

2、There is too much restatement in the discussion section and the conclusion section, I suggest that the language be refined and that the two sections, or some of them, be considered to be merged together.

3、Why aerosol optical depth used as an index for O₃ retrieval? As far as I know, solar radiation is also an important factor in the photochemical reaction of O₃. Will it be improved if it is included as a feature in the model for training?

Minor:

• Figure 3 lacks the label of (b)
• Line 231 has an extra set of parentheses after the NCP.