Review of

Combining POLDER-3 satellite observations and WRF-Chem numerical simulations to derive biomass burning aerosol properties over the Southeast Atlantic region

By Alexandre Siméon et al.

This manuscript explores aerosol absorption in the South East Atlantic Ocean (SEA), both in cloud free conditions and above clouds, using WRF-Chem constrained by POLDER-3 observations. While the time frame is limited (half of one month), this manuscript is a valuable investigations of that snapshot in time, and illustration of the strengths and weaknesses of WRF-Chem for these conditions. I would rate the scientific significance of this (timely) manuscript to be excellent. There’s a few things I would do differently regarding scientific approach and presentation, so I rate those as ‘good’ but am willing to rate higher if convinced otherwise. Overall, I believe this is a paper that should be published after addressing a few issues I’ll describe below.

1. In this review, I’ll reveal myself as a remote sensing specialist with less expertise in models such as WRF-Chem. With that in mind, I’d like to question the two step method of constraining WRF-Chem with POLDER and other data. Is this common practice? What is the benefit compared to constraining in one step, in other words adjusting WRF-Chem all at once based on POLDER and met data? I could understand differentiating between constraints that come from the POLDER retrievals and other data. But that is not the case – AOD from POLDER is used in step 1, microphysical properties in step 2. The problem is that the POLDER algorithms work by simultaneously retrieving all of those parameters, which I would think should be considered when using those data to adjust WRF-Chem. For example, if one were to impose an incorrect AOD in a POLDER retrieval, there would potentially be an incorrect retrieval of aerosol absorption (and vice versa). In a similar fashion, by initially modifying emissions to match AOD, and then later constraining aerosol optical properties might mean AOD no longer matches. I wonder if the POLDER/WRF-Chem match for AOD in Figure 16 and the ACAOD in Figure 17 could be further improved by constraining emissions a second time at this point. Which begs the question: why not just do it all at the same time. Am I missing something here?

1b. As an aside, I find the nomenclature for S1 “reference configuration” to be confusing. “Reference,” to me, implies some external (to this effort) source of data, while instead S1 represents the first stage of constrained WRF-Chem results. Alternatives that may be better:

“partially constrained configuration” or “emissions adjusted configuration” or something else more informative. Furthermore, Figure 3 could be made more useful by also listing what is constrained in WRF-Chem at each step.

2. Continuing the discussion on constraint methodology: why is it so approximate? In section 3.2.1 you adjust the APIFLAMEv1 BB emissions inventory by a factor of 1.5, while the ratio of WRF-Chem to POLDER-3/GRASP AOD(565) is 0.49/1.10 = 2.24. I imagine one reason could be that AOD does not directly relate to emissions estimates, but a discussion of how you go from that ratio value to 1.5 is needed. The same could be said for the choice of the Williams et al. (2007) BC refractive index, BCx2, OC/2.5 and 2.5% BrOC: is this the best match you can make? Could you do some more fine tuning to get a better constraint, or are the computational needs too much?

2b. By the way, I think table 4 would be more informative if the ‘score’ you used in section 3.3 were also included. And regarding how that score is calculated: why not use statistically appropriate parameters such as the Mean Absolute Error (MAE) or Root-Mean Squared Error (RMSE)?

3. Another issue relates to how you visualize and assess the differences between POLDER and the various versions of WRF-Chem. I like the images in, for example, Fig. 7, 8, 11, etc. since they give a good intuitive understanding of the spatial context in comparing the two. However, it is qualitative. Since you’ve aggregated the POLDER data to the coarser WRF-Chem model grid, you have the ability to do a pixel by pixel comparison – in which case you can calculate and visualize more statistics. In most cases, you present a mean difference between POLDER and WRF-Chem, when instead you could show a histogram of differences, and make a mean-bias (“Bland-Altman”) or some other sort of plot. Please just don’t use the (unfortunately common) scatterplot and calculate correlation coefficients, as that would be statistically inappropriate (see for example, Altman and Bland, 1983, https://doi.org/10.2307/2987937). One of the reasons this matters is because AOD (and presumably ACAOD) is not normally distributed (Sayer and Knobelspiesse, 2019, https://doi.org/10.5194/acp-19-15023-2019).
4. Essentially, you’re only assessing half of one month of one year. I realize that many of your choices are driven by computational resources, but how do you expect these results to fit within the annual cycle of BB in southern Africa, and how relevant is it considering inter-annual variability? I was expecting some discussion on this. These results need to be expressed in terms of their general usefulness. Perhaps a few words on computation resources needed for this analysis would be useful too, should somebody want to scale up to a full year or years.
5. Is the RETRO anthropogenic emissions inventory, which doesn’t cover the year of your analysis, the best source for that information?
6. When considering POLDER measurement uncertainty, please recall that this often refers to per pixel estimates. In that sense, are the error bars in Figure 14 appropriate, since they are applied to geographically averaged values?

More minor points:

1. Why are the reference wavelengths for SSA and AOD different for the POLDER/GRASP and above cloud products different? I agree that they’re close enough to not matter, but it should be explained.
2. Page 2, line 45: Clarify if “BBA direct radiative forcing” refers or global or SEAO.
3. Page 14, line 125: I think you mean “specialty” not “specificity”
4. Table 2 and discussion in text. I’m a little confused what you mean by “range of uncertainties” here. Does this really mean range of the values reported in the literature?
5. Page 21, lines 507-8: Is the really the primary source of the differences, or could it also be due to the nature of your WRF-Chem constraint as well?
6. Page 25, line 577: Is the score calculated for AOD, SSA, AND the top of atmosphere spectral observations? If so, the latter are used to derive the former, so the logic seems a bit circular.
7. Page 29, line 657: You mention here and elsewhere a “spatial correlation coefficient” but give no description of what you mean by this or how it is calculated. Similarly on Page 30, line 687. Are those error metrics what I’m calling for in 3? Regardless, they need to be explained.

Review of: Combining POLDER-3 satellite observations and WRF-Chem numerical simulations to derive biomass-burins aerosol properties over the Southeast Atlantic region. Authors: A. Simeon, F. Waquet, J-C Pere, F. Docos, F. Thieuleux, F. Peers, S. Turquety, and I. Chiapello.

Recommendation: Accept with Minor Revisions

In this study, the authors focus on biomass burning (BB) particle plumes transported above clouds over the Southeast Atlantic (SEA) region, off the southwest coast of Africa. They employ simulations from a regional model (WRF-Chem) coupled with meteorological reanalyses data and aerosol retrievals from POLDER in clear sky (POLDER/GRASP) and cloudy scenes (POLDER-3/AERO-AC), to better characterize the physico-chemical and absorption properties of aerosols.

Other reviewers have asked relevant questions, so I will just add some additional comments below:

Page 7, line 155: What is the temporal resolution of the WRF-Chem configuration?

                            Are the vertical levels evenly distributed from surface to 50hPa or the vertical resolution is finer near the surface? Could you clarify in the text?

             Line 157: Whilst I agree that the first half of July is representative for the whole month, this short period is not representative for the entire biomass-burning season. As Adebiyi et al, 2015 and Deaconu et al, 2019 showed, September and October months are characterized by different meteorological conditions and larger amounts of BBA transported over the SEAO. Also, the BBA are lifted at higher altitudes, limiting the contact aerosol-clouds. How do you justify choosing this period to study?

                            What was the computational cost of the 30 days (with 15 days spin-up) simulation and could you apply the optimized model configuration over September/October 2008 to check the consistency over this period?

Figures 1 and 2: From these figures looks like you could have chosen a different study area for the clear-sky cases, say between 5^o and 15^o N, that would have covered more of the biomass-burning emissions (correlated with the PM2.5 in Fig.1) and also more clear-sky days (Fig.2). Why did you choose this particular box?

Page 10, Fig2b: You are showing number of observation days of clear and cloudy scenes. It would be useful to plot also the data with coincident aerosol retrievals (e.g. number of observation days used in the study) for clear and cloudy skies.

Page 16, line 395: What is the scale of the biases between the AOD in clear-sky POLDER/GRASP compared to WRF-Chem reference configuration? From Fig. 7, it looks like the model is still strongly underestimating the AOD over land. Since the range of uncertainties for BB emission inventories reported in literature in 2 to 4, why did you choose to scale the APlFLAMEv1 with a factor of 1.5?

Page 17, line 408: Deaconu el al., 2019 showed that oxygen pressure method underestimates the cloud height compared to the CALIOP retrieval, by about 2-300 m for low clouds. Therefore, the WRF-Chem underestimation of cloud top could not be as high as 500 m. Could you have used the CTH from CALIOP instead of POLDER for the model optimization (ore scale up the POLDER CTH using CALIOP retrievals)?

In the beginning of the paper, you mention having simulated only half of July as representative for the entire month. In Fig.2 you mention ‘01-15 July period’ and everywhere else you mention only ‘July 2008’, which leaves the wrong impression the data (satellite and/or model) are averaged over the entire month. Please clarify in the captions and in the text where necessary.

Minor corrections:

Page 10, line 279: ‘…to evaluate the aerosol extinction simulated with WRF-Chem at 550, as well as…’

Page 16, line 360: ‘…This bias could be due to…’

Page 17, line 399: ‘…besides the simulated aerosol loads…’

Page 18, line 431: ‘…WRF-Chem configuration simulates well the …’

               line 447: ‘…depolarization ratio method…’

Sometimes the phrases are too long, and there are missing commas or linking words that could make reading more elegant and easier.