General comments:

The authors try to refine the ground meteorological stations surrounding the Beijing-Tianjin-Hebei region to achieve an improved forecast for particulate matter. This topic is interesting and has practical implications since right now more and more stations are constructed but few studies have studied how they can help improve numerical forecasts in reality. The refining approach introduced in this paper by considering the sensitive areas is reasonable and logical. Overall, the manuscript is well written and clearly structured.

However, there are still several issues that should be addressed before acceptance.

Major comments:

1. Based on the sensitive areas identified by the CNOP associated with the 48 forecasts, the authors first identify the essential observation network, and then scatter the remaining station according to the comprehensive sensitivity. So the accurate calculate of CNOP is the basis of the study. In section 2.3, the authors presented a detailed description on the definition of CNOP-type error. However, the descriptions on how to calculate the CNOP is brief and insufficient. I suggest the authors add more details of the algorithm on Line 195.

1. In section 5. The authors select two forecasts which possess large forecast errors in the control run as examples to show that the cost-effective stations provide observations of equivalent efficiency of the whole constructed stations. However, to better demonstrate the effectiveness of the cost-effective observations, I recommend the authors to have a look at the improvements when the cost-effective observations are removed from the whole station observations. If the remained observations (after the removal of the cost-effective observations) contribute to a slight improvement of the PM2.5 forecasts but with a larger number, then it will be more convincing that the cost-effective observations are necessary for the PM2.5 forecasts in BTH.

1. Also, in section 5, the authors only take two examples to present the detailed interpretations, which is not enough to me. Even if the authors have explained on Line 584 that the assimilations of the cost-effective station observations and all the constructed station observations correct the meteorological conditions for the PM2.5 forecasts in a similar way, it is suggested to add more examples or discuss the overall corrected meteorological conditions in more detail. For example, the authors may use the atmospheric stability to quantify the meteorological conditions for the accumulation or dissipation of PM2.5 concentrations.

Minor comments:

1. Line 104. For the application of CNOP in field campaigns, Feng et al., (2022) demonstrated its validity on identifying sensitive areas for typhoon forecasting.

Feng, J., Qin, X., Wu, C., and coauthors. Improving typhoon predictions by assimilating the retrieval of atmospheric temperature profiles from the FengYun-4A's Geostationary Interferometric Infrared Sounder (GIIRS). Atmospheric Research, 280(15), 106391.

1. Line 288. The authors use “target observation” here, but in the introduction part they used “targeted observation”. Please unify the usage.
2. Line 290. When the “cost-effective” first appeared in the manuscript, I did not quite understand what it means. More explanations should be added here.
3. Line 323, when determining the sensitive areas, the authors should clarify here that CNOP-type initial errors are superimposed on the ground meteorological fields in the “truth run”.
4. Line 420, the 110E~120E should be 110^oE~120^o Also the 34N~36N.
5. Line 705, it is recommended to mention in the section 6 that the improvements are based on the OSSEs, which means the simulated observations from ERA5 are assimilated to the control run to show the effectiveness of the newly refined station observations. However, how the improvements will be when the real observations from the refined station network are assimilated still needs further studies.
6. The boundary layer height is also an important meteorological variable for PM2.5 forecasts. Why do not the authors consider the perturbation of this variable in the study?

The manuscript entitled “An approach to refining the ground meteorological observation stations for improving PM2.5 forecasts in Beijing-Tianjin-Hebei region” introduced an approach to refine the ground stations by identifying the sensitive areas for targeted observations. The study is highly related to the studies of predictability, target observation and data assimilation. And it provides a scientific guidance on optimizing the ground stations. I believe the approach is not only useful for air quality forecasts, but can also be used to the forecasts of extreme weather events. Nevertheless, there is a gap between publication and the manuscript in current version. I hope the following comments will help authors improve the manuscript.

Specific comments:

1. Line 42. There are a great many publications addressing the meteorological conditions on PM2.5 variations, but the authors only cite one, which is not enough. More references are needed here.
2. Line 68. “assimilating more observations may not necessarily lead to much higher forecast benefits.” References are needed here.
3. Line 75. How are the worse forecast skills possible when the sensitivities are low? Please provide a detailed explanation here.
4. Line 195-202. The descriptions are insufficient and confuse me. Please add more details and make it clear.
5. Line 313. Please clarify that the real “meteorological” observations are not in public archive, because in section 3.1, the authors have compared the simulations with the observed PM2.5 concentrations.
6. Line 322. Is the CNOP-type initial error that what has been described in section 2.3? It is suggested to add a detailed description on what variables the CNOP-type errors have contained here.
7. Line 349. “the area with larger values of TME can be regarded as the sensitive areas”. It is ambiguous. Is there a threshold for the definition of sensitive area or just determined subjectively?
8. Line 453. “the essential stations can indeed provide additional observations that help increase the skill of the PM2.5 forecasts, in comparison to other constructed stations but not in the sensitive grids”. The authors did not do any comparison experiments to show the improvements are higher than assimilating the station observations which are not in the sensitive grids. How can they get such conclusions?
9. As shown in Figure 7 (a1, a2), assimilating the observations will lead to worse forecasts since the AEv and AEM are negative. It is hard to understand. Why will assimilating the observations will lead to worse forecasts?
10. Line 703. “It is clear that assimilating the fewer observations can lead to higher PM2.5 forecast skills”. It is inaccurate. It is suggested to rephrase it more carefully.
11. It is suggested to mention the limitation of the study in section6 that the results are based on OSSEs. If the real observations are available, how the refined station observations help improve the air quality forecasts deserve deeper studies.