Review of “Unleashing the Potential of Geostationary Satellite Observations in Air Quality Forecasting Through Artificial Intelligence Techniques” by Zhang et al.

Major Comments

This study by Zhang et al. entitled “Unleashing the Potential of Geostationary Satellite Observations in Air Quality Forecasting Through Artificial Intelligence Techniques” presents a new machine-learning framework – GeoNet – that synthesizes geostationary observations of columnar NO₂ from the Geostationary Environment Monitoring Spectrometer (GEMS) with meteorological parameters to forecast surface-level NO₂in East China. Overall, this study represents a significant advancement in surface-level pollution forecasting given its use of the unprecedented hourly data provided by GEMS. I believe that this manuscript is well-written and consistent; however, I have a few comments below.

First, if possible, it would be useful to validate the GEMS observations using ground-based spectrometers (e.g., PGN) specifically for the study region and time period. Additionally, unless I missed it, I don’t believe the time periods for model training and validation were ever stated; if this is the case they should be added to the main text. Second, when investigating feature importance, it would be useful to also identify variability in the feature importance to uncover whether some components are more stable than others in GeoNet and to identify if the significance of geostationary observations is consistent across different days and seasons. Lastly, I suggest that the authors update their analysis in Figure 4 to include the GeoNet predictions regridded to the CAMS grid to identify how much of the improvement in predictions is attributable specifically to enhancements in spatial resolution.

I have included line-specific comments below:

Minor Comments

L53-54: While I agree with this statement, it should be mentioned that for air pollution forecasting to facilitate health benefits, infrastructure needs to be created that communicate risks and appropriate responses to risks to the public.

L55: I think you can drop the second limited in this line.

L75: Maybe it would be useful to give an example or two here (i.e., TROPOMI + OMI).

L78-81: Another limitation of the polar orbiting satellites that is worth mentioning is that typically (at least in the case of TROPOMI) the satellite observes at roughly the same time of day (early afternoon) which makes it difficult to predict concentrations at other times of the day with different meteorological (boundary layer height) and photochemical conditions.

L92: It would be better to describe GEMS as having “unprecedented temporal and spatial resolution andcoverage” as ground-level monitors can observe hourly NO₂ but are limited in time and aircraft remote-sensing can observe NO₂ at sub hourly resolution but over a limited temporal coverage (usually a few days or weeks). The resolution alone isn’t necessarily unique but rather than combined spatial + temporal resolution with extended spatial and temporal coverage.

L117-120: Were you able to validate these data for the study time period / domain? If possible, it may be useful to compare GEMS to ground-based spectrometers in the study domain to get an idea of performance.

L207-208: I don’t think you need this sentence as it is already mentioned in the methods section.

Figure 3: It would be interesting to present the variance of these different components as well in a). Are these importance values pretty consistent regardless of season and day, or do they vary substantially day to day?

Figure 4: Have you assessed how much of the reductions in performance are attributable to resolution? If not, I suggest regridding the GeoNet prediction to the resolution of CAMS and comparing this “GeoNET_coarse” product to the observations to characterize how much of the improved performance is attributable to enhanced spatial resolution.

Figure 5: The colorbar in a is not labeled, and throughout the font is small (especially in the yaxis of c and d), I suggest updating to improve readability.

L338-339: I don’t believe the timeframe of this study was mentioned at all in the main text. What months / years was this prediction trained on and for what period was it validated?

The authors attempted to improve the short-term prediction of surface NO2 at a high spatial and temporal resolution by taking advantage of the GEMS NO2 prodcuts and a neural network model. They successfully forecasted full-coverage surface NO2 for the next 24 hours and identifed the critical role of GEMS NO2. Their results demonstrate the potential application of the GEMS products in air quality prediction.

Overall, this is an important study and the results presented here will be useful for future applications of GEMS products as well as the geostationary satellite observations. I am happy to see its publication in due course. However, before that, I still have a few concerns or suggestions for the authors.

I would sugget to move the model configuration and optimization into the main text. This will be very helpful for readers to understand the model.

In the handling of missing data, the authors tried to set them to a fill value of zero. Is it reasonable? It looks reasonable to fill values of diurnal NO₂ climatology  (e.g., the seasonal mean diurnal NO2). In addition, as shown in Fig.2, it looks the three methods of handling missing data perform similarly in term of R2 and RMSE. So I don’t think it is necessary to highlight the “weakest” or “strongest” configuration.

I would also suggest to move Fig.S12 in the main text which shows the advantage of GEMS measurements.

The authors also show that the performance of GeoNet model degrades notably after t+16h. Is there any possible solution to overcome this short predicability?

As mentioned in my last comment, the authors also highlight the possible applications to other air pollutants. However, the chemistry and lifetime of other air pollutants might be very different from NO2. For example, if the GEMS tropospheric ozone measurement is useful for the prediction of surface ozone? Some more detailed discussions on the possible application to other pollutants would be very insightful.

BTW, I am not sure whether it is necessary to include NO2 in the Title of this manuscript since the authors didn’t talk too much about major air pollutants (i.e., PM2.5 and ozone)