Li et al. presented an new machine learning scheme and applied it to OMI SO2. It is an interesting preliminary study and the work should be published after minor revision. The paper is well written and structured. Overall, the figures are not of sufficient resolution.  

Main comment: although it is an interesting study, I am concerned by the fact that the NN function is heavily weighted by SRR. The SRR, as defined by the author, actually contains the desired result. Therefore, it is not a surprise that this works. The only real task of the NN is to resolve the RMS dependence. Also, the fact that the noise gets reduced is in a way artificial, as it is the direct result of the constrain on the clean pixels that should be SCD = zero.  The reduction of the bias poses also the problem of a possible overcorrection. On long-term averages, are the weak emissions sources detected by OMI PCA still present in the NN data set?

In my opinion, what would be much stronger is to train the NN directly with the corresponding radiances. This is done to some extend (through a PCA transformation) but it is coming at the end of the paper. It is  pity it is not put more in front.

Minor comments

L55: I agree with the necessity to improve the retrievals but is a 10% noise increase a real problem for addressing long-term trend monitoring? I do not think so. The appearance of instrument issue like row anomaly is more of a problem.

Figure 2b is confusing. Is the SRR unit less?  The SCD is DU and RMS has no unit, thus SRR should be expressed in DU (?).

Section 2.2. The classification of pixels is quite complex. As explained in the text, the parameters used (a1,a2) have been determined by testing. However, it would be good to illustrate the impact of the (a1,a2) settings on the final results. Currently, it is hard to judge if the complexity is worth, compared to a more simple classification.

Section 2.3: the processing is not performed separately for each row. Why not? Would it improve/degrade the results?

Section 3.1: the SCDs over SAA are much better in the NN analysis, which is surprising. Any idea why?

Figure 7: the figure quality is not sufficient. When zooming over the subplots it is hard to see the emission patterns described in the text.

Figure 10b : the PCA-NN results seem to show striping features, although the analysis is performed separately for each row. Why?

Review of manuscript "A New Machine Learning based Analysis for Improving Satellite Retrieved Atmospheric Composition Data: OMI SO2 as an Example" by Can Li et al. 2022



In the manuscript a new and interesting method to minimize noise and artifacts using machine learning has been presented and applied to OMI SO2 data. The paper is well written and only minor revisions are needed.

My main comment is that an optimization (or sensitivity analysis) of the applied NN architecture is missing. A simple NN architecture from a completely different scientific field ("for reconstruction of RGB images from hyperspectral radiances") was used. In general a simple architecure is fine as a starting point, but what I am clearly missing is a sensitivity analysis of much the NN architecture is affecting the results.

Furthermore I suggest some restructuring of the paper - I think it would be good to start with the simple approach of using the linear interpolation model and then begin the NN model.

Althoug the comparison and different maps are nice to see, I suggest to also add line plots as a function e.g. latitude such that is easier to see biases and differences.

Detailed comments:

•  Sect.1, L49: Please provide numbers or references for the background noise of OMI SO2 SCDs.
• Sect 1. L80ff: The 20 DU limit of the FP-ILM retrieval only applies to the SO2 LH retrievals. It was not yet applied to SO2 VCD retrievals
• Sect 2.1, L109: How are pixels with enhanced SO2 after volcanic eruptions detected/filtered? Do you apply a VCD threshold? Please describe
• Sect 2.1, L128: By how much do the monthly medians/stddev change every month? Do you see jumps in the results from one month to the other?
• Sect. 2.2, L146: I am a bit concerned that the a1/a2 factors are based on trial and error and there is no robust criterion to determine them. This makes appying the whole method to other sensors (or even to OMI for a different time span) difficult - I assume the factors will vary over time, especially related to instrument degradation. Can you perhaps show how sensitive your results are to (small) changes in a1/a2?
• Sect 2.3, L180ff: As already mentioned in my main comment, I am a bit concerned about the choice of the NN architecture. Although it is a good starting point to use a simple archictecture, it definitely needs to be optimized for the specific problem, especially independently for your NN1 and NN2.
  
  Is there a reason for the choice of activation functions? I.e. using soft-sign and then sigmoid is not really common - I suggest to use ReLu or something related for both hidden layers.
• Sect. 4.3: I suggest to put this in front before you apply the NNs, to show that a simple linear interpolation is not sufficient and therefore you apply NNs.
• Figure 4&5: Suggest to add line plot of SCD as a function of latitude. With this plot you probably better see biases and the differences
• Figure 5: Suggest to use different color scale from -0.1 - 0.1
• Figure6b: Relative differences are always problematic for SO2 plot since in clean areas the SCD is close to zero and hence the relative difference becomes extremly high, as can be seen in the figure. Suggest to use absolute differences here.
• Figure 7: The maps are rather confusing and do not provide addtional information. I sgueest to rather repleace them by line plots as a function of e.g. latitude (see my comment above)
• Figure 9&10: Suggest to add line plots of SCD as a function of e.g. latitude. With this plot you probably better see biases and the differences