General comments:

Currently, many studies are using spatially sparse fixed-site measurements to map air pollution on a large scale, ignoring the local spatial heterogeneities such as the intra-city variations. This article evaluated the performance of various algorithms across different scales and validated the accuracy separately in subsets categorized by road density and population density. They found that the model performance varied significantly at different spatial locations. The pattern was found to be different in “global” and local models. The comparison between “global” and local models in terms of intra-city distribution patterns is valuable. However, in its present form, I cannot recommend the article for publication. With substantial revision and restructuring, this article could be a useful addition to the existing literature.

The writing needs further improvement. The current version is not easy to read. First, this is too long. I appreciate the solid work of the authors. But please simplify the main text and consider moving some descriptions/figures to the Appendix. Keep only the core story in the main text and make sure the primary findings and the most important messages stand out. Second, consider restructuring the method/data and the result section. Third, the caption of figures and tables needs more details, including the unit of NO₂. Fourth, Clarify definitions like “Far from road” vs “Rural”. Last, please pay attention to the tense usage.

Specific comments:

1. The mobile measurements from Kerckhoffs et. al., 2019 were measured on the road. How can they validate the accuracy for the “far from road” group? Did you perform any adjustments?
2. Table 1 describes the predictor features. Why not include land use proportions? Land Use Regression models are efficient and well-accepted methods in air pollution modeling.
3. Figure 4. It would be better to plot the map of differences between the model tested and the benchmark (i.e., NO2 estimations from Kerckhoffs et. al., 2019). I would be curious about the difference in spatial distributions between the “global” and local models.
4. A restructuring of the data/method section is recommended. Begin with the introduction of the data source, ensuring clarity on the source of the population information and road class when discussing spatial groups. Consider adding a table summarizing model input/algorithms for ease of understanding. Move some algorithm introductions to the appendix.
5. Please explain why 20-fold cross-validation?

Technical corrections:

I have listed some specific points. But not limited to them.

Abstract:

The abstract attempts to encompass numerous findings but allocates insufficient space to elucidate the methodology and experimental setting. A substantial rephrasing of the abstract is needed.

Line 1-5, toing and froing, can be simplified.

Line 6, please provide more details about the meaning of “spatial heterogeneity” in this context.

Line 9-10 what is the local and global model? Define first, before using it.

Methodology:

Line 100-105, not clear. How do you divide the area? Purpose? What is the time frame of these national measurements? Frequency of measuring? Any preprocessing? More details are needed here. How do you define the less densely populated area? What is the source of the population density data?

Line 121, “rural”= “Far from roads”? Please keep the terminology consistent.

Line 123,  the label of models should be provided as the legend in the figure instead of in the caption.

Line 130-135, unit of NO₂ is missing. This paragraph is not informative. The values can be integrated into the figure 1.

Line 145, More details about kriging and accuracy are needed.

Line 160-165, is the traffic volume used as the annual average? Table 1. it would help readers to understand the data distribution by adding columns such as numbers and some statistics like mean, median, quantiles etc.

Line 168, the section title should begin with a capital letter, and further refinement is necessary in terms of formatting.

Line 190, not clear. Please do not refer to the citation but to the dataset you have described in section 2.1.

Line 195, rephrase please instead of a direct quote.

Line 196, details of the tuning strategy are missing.  

Result and discussion:

Line 465, how do you compare the influence of predictors between cities? The feature importance is a relative value. The magnitude is not meaningful when compared to the other models.

Line 515, which is opposite to the common knowledge (see Hoek et. al., 2008). Can you explain why non-linear model predictions were smoother?

Dear authors,

please, find attached an extended report about your paper.

Best regards