This study applies machine learning to estimate hourly PM₂.₅ and PM₁₀ concentrations across eastern China using Himawari-8 satellite data, analyzing trends, influencing factors (2015–2023), and urban–rural disparities. The results are well presented. Below are comments and suggestions for improving the manuscript:

The particulate matter designations “PM2.5” and “PM10” should consistently use subscript formatting (i.e., PM₂.₅ and PM₁₀) throughout the manuscript for scientific precision.

Numerous previous studies have derived hourly surface PM concentrations from Himawari-8 observations in China (doi:10.5194/acp-21-7863-2021). These should be briefly summarized in the Introduction.

Similarly, the Extreme Trees model has been previously applied successfully for satellite-based PM₂.₅ (doi:10.1038/s41467-023-43862-3; doi:10.1016/j.rse.2020.112136) and PM₁₀ (doi:10.1016/j.envint.2020.106290) estimation. A concise summary of these efforts should be added. In addition, a clear justification for selecting this particular model over other machine learning approaches is needed.

Line 91: The acronym “TOAR” appears before it is defined. All acronyms should be spelled out at first mention for clarity (e.g., “Tropospheric Ozone Assessment Report (TOAR)”).

Lines 97–101: The authors should clarify whether only Himawari-8 data were used, or whether Himawari-9 (which became operational in December 2022) was included in the 2022–2023 period. This is important for ensuring temporal consistency.

Lines 119–121: The data sources and preprocessing steps for elevation (HEIGHT), land use and land cover (LUCC), and population density (RK) should be explicitly described.

Equation 1: The model uses only top-of-atmosphere (TOA) reflectance, without accounting for viewing or solar illumination angles, which are known to influence aerosol retrievals. The authors should provide justification for their exclusion.

Figure 1: The methodology used to simultaneously estimate PM₂.₅ and PM₁₀ via a multi-output model is unclear. A brief explanation or schematic would improve reader understanding.

Equation 2: The terms SS_res and SS_tot should be formally defined in the text or figure caption.

Line 163: “SHAP” should be spelled out as “SHapley Additive exPlanations (SHAP)” upon first use.

Line 168: The selection of “20 times” for permutation testing appears arbitrary. A statistical or methodological justification is necessary.

Line 172: The purpose of the provided URL is unclear. The authors should clarify what resource it links to and its relevance.

Lines 212–213: The reported temporal cross-validation R² values (0.41 for PM₁₀, 0.51 for PM₂.₅) seem inconsistent with the claim of “robust stability.” The authors should address this discrepancy or revise the description accordingly.

Figure 2: The placement of accuracy labels is too close to the subplot boundaries, potentially affecting readability. Adjust the positions to improve visual clarity.

Line 243: The manuscript does not evaluate relative reduction trends (i.e., trends normalized by baseline concentrations), which are crucial for comparing changes across regions with differing pollution levels. Consider incorporating this analysis.

Figure 3: Clearly define the boundaries (e.g., interquartile range, whiskers) of the box plots in the caption. Additionally, the color bar ranges in panels C–F are too broad, masking regional differences. Narrowing the ranges would better highlight spatial variability.

Lines 263–269: The number of decimal places reported is inconsistent. Standardize numerical precision across the section, preferably to two decimal places.

Lines 277–278: The inclusion of temporal variables (year and month) as proxies for anthropogenic drivers requires further explanation. Clarify their interpretability in the context of human activity patterns.

Figure 8A: The x-axis range is too narrow, truncating some boxplot distributions. Expanding the axis limits would allow for clearer visualization of data variability.

This Manuscript uses an Extreme Trees based machine learning model to identify the drivers in changes of urban and rural PM in China. This is a good effort, and the authors demonstrate reasonable applicability of their approach.

Following are key points that need to be addressed:

1. Since the authors use spatial - temporal datasets, why was LSTM and Convolutional Neural Networks not applied?

2. Figure 7: Explain the physical justification for why SHAP values for temperature are negative in summer and positive in winter? Would these change between urban versus rural areas? For example, biogenic emissions might increase in summer at higher temperatures increasing secondary organic aerosol formation. In winter, reducing temperatures might increase demand for residential heating. Further discussions are needed here.

3. Figure 8 and related discussions: The figure is not clear. Discussions on lines 352-356 suggest different results for how interannual variations change between urban and rural areas using the 2 approaches: Relative contributions versus SHAP. Why are these different? Are SHAP values more reliable? The authors seem to just combine results from relative contribution and SHAP in their Abstract and Discussions. However, physical justification is needed to figure out what causes these differences.

4. What about role of photochemistry? The authors include solar radiation, however, it does not show up as a key variable in SHAP interpretability analyses.

5. Conclusions: Line 402-405: The authors rightfully acknowledge that anthropogenic influences are just represented by a time variable in their analyses. This is clearly insufficient. If possible, the authors should consider emissions, photochemistry (ozone, OH radicals, NOx, VOCs) etc. in their analyses.