Supriya Mantri, Laura Horton, Adam Povey

National Centre for Earth Observation, University of Leicester, Space Park, Leicester, LE4 5SP, UK

The paper describes the importance of high resolution PM2.5 data for air quality monitoring and health studies. It suggests a new machine learning technique for correcting the AOD-to-PM2.5 ratio using inputs from Sentinel-3, MERRA-2 reanalysis, and high-resolution geographical indicators. That post-processed corrected AOD-to-PM2.5 ratio was used to estimate the PM2.5, which was then compared to OPEN-AQ ground stations across Europe (specifically, Paris and Madrid) for 2019. The product was shown to work best for low PM2.5 values.

As researchers of the interface between satellite aerosol observations and air quality, we found this an interesting and inspirational publication which we enjoyed discussing and would like to see published. We share some minor corrections and comments for the authors to consider:

• The introduction was clear with good motivations and references but, in line 25, the reference to the WHO air quality guideline does not specify the time frame. Is it an annual average?
• Is it possible to give any additional reasoning behind the use of hourly downscaling of daily PM averages? This strikes us as one of the more consequential choices in this methodology and neglects complexities such as the diurnal cycle. (We aren’t questioning the authors’ judgement, nor disagreeing; merely curious as we wouldn’t have thought of this.)
• In section 2.3 the authors mention they use MERRA-2 reanalysis variables as an input for the model and provide a lengthy list in the appendix. Have all of the available variables been used as an input or are there some variables not included? It may be helpful for those designing similar algorithms if you also mention which variables were not included and, if so, why not?
• Could you provide a reference for the RH equation on line 93? Or is it at standard temperature/pressure?
• In the NASA Black Marble Night Light section (2.6.2), the authors could include a reference to the data set used, especially a DOI so we can distinguish which of the four available datasets was used.
• We found Sections 3.3 and 3.4 quite opaque. They could be improved by adding more specific details about the process, such as the equations used as an input to the correction model. A flow chart of the steps used in the approach would make it more clear for an audience which is less familiar with machine learning and neural networks.
• Those of us with little experience with neural networks did not understand what Figure 2 wished to convey and those of us used to neural networks felt Figure 2 adds little to the text; it could be removed.
• At line 183, it would be helpful to have an understanding of what is meant by “slightly better”?
• Figure 3 could be improved as it is hard to distinguish between lines, perhaps using a filled histogram of stacked bars.
• We are curious if the authors considered any methods to amplify the availability of high PM2.5 observations, such as data augmentation? Our understanding was that the balancing of training data is an important step in constructing a neural network to recognise rare events and we would value the authors’ opinion.
• From line 223, what do you mean by “the fully learned approach were less accurate than with the post correction approach”? What metric of accuracy was used and how significant was the difference? This would help guide our own efforts in neural network generation.
• In Figure 4, are the authors certain about the plotting of panel B? Compared to panel A there appears to be some duplication of points up the y-axis. For example, there are three points at the extreme right of (A) but over ten in (B). This effect is not exhibited in panels D-F.
• In Figure 5 (comparison of the uncorrected and corrected methods at the ground stations), do the authors have any understanding of why there is a large discrepancy between OpenAQ and both satellite estimates for most of the sites (i.e. the blue dots do not overlap the red line in 5/9 cases shown). Is it because of some local source (e.g. roads or small industrial buildings) in close proximity of the stations that isn’t present in Madrid?
• In Figures 5 and 6, could the dots representing sites and arrows indicating one be made substantially larger and outlined with a colour not in the plot (such as green or blue)? Our older member had failed to see them on his own.
• In Figure 7 bottom left (Hull Freetown) why is the uncertainty so large April? What are the possible reasons for high PM2.5 in February?
• If practical, it would be interesting to add an appendix highlighting a few ensemble members for Fig 5 and/or 6. This would demonstrate if the smoothness of the fields shown at the top right of those figures is due to the action of the neural network or due to the median filter over the ensemble.
• It would be interesting to hear if training the model over an area with higher PM2.5 levels, such as South Africa or India, and then testing over central Europe improves the model’s performance, particularly for higher PM2.5 values.

Some typographic corrections,

L3: we consider a satellite-based

L4: satellite retrievals of Aerosol

L5: we employ a machine learning

L33: retrievals and an AOD-to-PM

L45: PM2.5 at a spatial resolution

L54: estimation of the AOD-to-PM

L62: The Sentinel-3 POPCORN AOD product is based on the post-process corrected

L76: by station; 1-hour

L154: due to the huge dimensionality

L181: A linear activation

L190: was divided into training

L248: is obtained by post-correcting

L249: to obtain similar post-process [this is actually a meaningful change; as written you say ‘products corrected in a similar manner’ and our revision says ‘similar products that have been corrected']