The manuscript addresses the issue of accurately and efficiently calculating plume injection heights, aiming to enhance the accuracy of estimating smoke exposure during Australia's biomass-burning season. For this purpose, the authors employed two distinct methodologies to quantify the fractions of fire emissions  injected above the planetary boundary layer: a climatological approach with the use of climatological monthly mean injection profiles from IS4FIRES and daily injection heights taken from the GFAS emission inventory and a novel machine-learning approach using random forest models trained using plume heights from MISR satellite instrument. The effect of wildfires on the daily smoke PM2.5 concentrations in Australian cities is estimated using the STILT Lagrangian particle dispersion model.   

The results presented in the manuscript reveal that the machine-learning approach generally outperformed the climatological in predicting  the daily plume injection fractions. Furthermore, considering the injection fractions predicted by the random forest model leads generally improves the agreement between PM2.5 concentrations (estimated by STILT) in Australian cities and surface measurements. Lastly, based on the findings from the Random Forest and STILT models, smoke-related PM2.5 in Australian cities accounts for up to 52% of total PM2.5 during the biomass-burning season.

Overall the manuscript is well-written, and the results are interesting.

However, in my opinion, there are some issues to be addressed and some clarifications to be provided before its publication in ACP.

Specific comments

ll 61-62: You described the drawbacks of using MISR and CALIOP to estimate the plume injection heights.

Are there any drawbacks regarding the use of TROPOMI?

Why did you choose PBL heights from MERRA-2? There are differences between the PBL heights provided by different datasets (Ding et al., 2021; Guo et al., 2021). How the uncertainties on the MERRA-2 PBL heights can affect your results? 

l 122:  A brief definition of the  "top and bottom heights of plumes", "MHMI", and "injection height"

would be helpful for readers less familiar with these terms.

Please provide a short discussion on the role of the plume injection height in surface air pollution based referring to new studies such as Li et al., 2023.

l 260: "Smoke PM2.5 is typically defined as the sum of the fire-related BC and organic matter (OM)" Please provide a reference. The non-biomass burning OM (e.g. biogenic organic aerosols do not affect the concentrations?

ll 283-284: "We assume that the fire emissions injected above the PBL have

no impact on the surface PM2.5." Is it possible that this may introduce a low bias to the results?

Please add some references regarding the synoptic conditions favoring the occurrence of wildfires. i.e. :

ll 324-325 (During this period, ... northern Australia), ll 326-327 (During the monsoon periods ... of wildfires) and

ll 331-332 (when low-pressure systems ... coastal areas). 

ll 345-348: According to Fig. 2, there are many cases where we observe an overestimation of injection fractions by INJ-CLIM compared to MISR observations, mainly in cases with relatively low fire emissions. Can you elaborate on this? 

Also, please show the normalized mean bias in Fig. 2 (and also Fig. 3, at least in Figs. 2b,3b)

ll 359-361: What about the low bias for large injection fractions? Please elaborate more on the potential causes of the biases.

ll 364: How does the meridional wind speed at 10 m determine the atmospheric stability? 

I understand that this is probably true, but maybe it's better to add a short explanation or reference.

l 371: "incomplete knowledge of the local temperature profile" Can you elaborate more on this?

sec 3.1 I believe that it would be both interesting and useful for the readers to provide some connection between the results presented in Fig. 4 and those discussed in the previous sections of the manuscript.

E.g., I can see that over coastal SE Australia (a region where the fire emissions are very pronounced according to Fig. 1)

the mean fractions of total OC fire emissions injected above the PBL are larger when calculated using the random forest method. Is this connected to the fact that RF capture better strong fire events compared to the climatological method?

ll 424-426 and 4.1, generally: Can you elaborate more on why there are many cases with no improvement when using the INJ-RF, based also on the results presented in Table S1? 

Also, you state (ll 444-446) that in Melbourne the model has problems in reproducing successfully the 

daily variations of PM2.5 during low fire years. However even during 2019 (Fig. 7f) the performance of INJ-RF is not good in this city (it's worse than CTL and INJ-CLIM)

l458: 2016 or 2015?

l470: "2 to 36%" according to which method?

ll 489-495: In Brisbane I see that the INJ-CLIM method performs slightly better than INJ-RF even during years with relatively large fires (e.g. 2019). Generally, I can see a similar performance for INJ-CLIM and INJ-RF in this city. Also according to Table S1, the NMB is larger for INJ-RF.

sect 4.3 The discussion on the contribution of long-term smoke PM2.5 to regional air quality is very interesting (results shown in Fig. S4). Consider moving Fig. S4 to the main manuscript. 

ll 571-573: The NMBs do not agree with those in Fig. 7.

sect 5. ll 561-562 "The random forest model predicts plume behavior that best agrees with observed surface PM2.5."

I agree that the performance of the RF model approach is generally very satisfactory (and in most cases yields better results compared to the climatological approach). However, I suggest that the authors moderate somewhat their expressions regarding the performance of this method.

References

Ding, F., Iredell, L., Theobald, M., Wei,J., and Meyer, D. : PBL height from AIRS, GPS RO, and MERRA-2 products

in NASA GES DISC and their 10-year seasonal mean intercomparison. Earth and Space Science, 8, e2021EA001859. https://doi.org/10.1029/2021EA001859, 2021

Guo, J., Zhang, J., Yang, K., Liao, H., Zhang, S., Huang, K., Lv, Y., Shao, J., Yu, T., Tong, B., Li, J., Su, T., Yim, S. H. L., Stoffelen, A., Zhai, P., and Xu, X.: Investigation of near-global daytime boundary layer height using high-resolution radiosondes: first results and comparison with ERA5, MERRA-2, JRA-55, and NCEP-2 reanalyses, Atmos. Chem. Phys., 21, 17079–17097, https://doi.org/10.5194/acp-21-17079-2021, 2021.

Li, Y., Tong, D., Ma, S., Freitas, S. R., Ahmadov, R., Sofiev, M., Zhang, X., Kondragunta, S., Kahn, R., Tang, Y., Baker, B., Campbell, P.,

Saylor, R., Grell, G., and Li, F.: Impacts of estimated plume rise on PM2.5 exceedance prediction during extreme wildfire events: a comparison of  three schemes (Briggs, Freitas, and Sofiev), Atmos. Chem. Phys., 23, 3083–3101, https://doi.org/10.5194/acp-23-3083-2023, 2023.

First off, authors, apologies for double posting and keeping the discussion period open. I couldn't find this in my account to review because apparently a whole new account was created to review this paper. I posted the review previously, but I was asked to resubmit it under this account.

Review as previously posted: In this paper, the authors compare three different methods for estimating plume injection heights/emission fractions above and below PBL: (1) a control with all emissions in the boundary layer, (2) using climatological injection heights from MISR and (3) using machine learning to estimate the injection height. These injection heights are used with the STILT model to estimate and compare surface concentrations in different cities across Australia.

The paper is generally well-written, and the methods seem sound. However, I felt like the methods section was very, very long and then was followed by sparse actual results and a lacking discussion section. I was left with many questions about their results. In general, I would have liked a greater discussion of where there were biases and hypotheses about what might be the cause, along with more discussion about when and where their machine learning injection height estimates did not improve the comparisons (line 475-476 seems like something that could be further explored). The Discussion and Conclusion section is mainly just a rehashing of results rather than a synthesis, comparison with previous studies, and necessary future work (rather than just apply it elsewhere and use more data). The abstract mentions that their model has the best agreement in several key cities, but it was not the best in other cities and other time periods. I’m not overall convinced that it is better than using the climatological injection heights. Before it can be published, it needs to provide more insight and analysis of the results.

Specific comments

Abstract, line 29-31: State what this 63% is in comparison to in the sentence

Line 72: PRM is only used two more time in the manuscript, I’d just write plume rise model each time to cut down on abbreviations

Line 76: MHMI needs more explanation

Line 90-92: But may not include chemistry, other PM sources, or background smoke

Line 100: It also requires that meteorology, emissions, and the injection heights are correct

Line 105: remove “improved” as it hasn’t been shown to be improved yet.

Line 130: remove “need to”

Line 157-8: Remove “for a limited set of months” as this is discussed later with the actual months listed.

Line 181: what is the spatial distribution of these MISR records over Australia?

Line 283-284: They can’t mix down?

Line 321-335 seems like intro section information

Line 363-373 Discussion of Figure 3c should come before the evaluations (before 3.2)

Line 426: change to “on modeling smoke concentration for exposure estimates in Australia.

Line 429: I’m not really sure this is a good idea for discussing exposure. I would have also liked the statistics from the daily observations included in at least the supplement (rather than just the 10-day averages). I’m also not sure how long smoke events normally last in these different areas.

Line 443: “the 10-day moving average of daily PM2.5”

Table 1. Put in the shorthand names used in Figure 3 Feature Importance Plot

Figure 3a and 3b should be combined into Figure 2 and Figure 3c should just be it’s own plot as it seems out of place here.

Figure 5 Change to “and southeastern Australia (Total, blue bars in a and b) The y-axis feels confusing for a and b with one in linear and one in log scale. In b, there needs to be more labels on the y-axis. There seems to be lots more year-to-year variability in the INJ-Clim for southeastern Australia compared to Northern Australia

Figure 6 (and Figure S5) isn’t really useful. The pie charts don’t provide much information and aren’t in line geographically, so I’d either just put the numbers on the map, in a table, or just use Figure S4 as it is a more useful/less confusing plot with similar information. If you are attached to the pie charts, then they should represent the total PM2.5 concentrations and be closer to the actual location.

Figure 7. What are the statistics for the same day as opposed to the 10-day average? Do these smoke events often last a week or longer or are the peaks on shorter timescales?

Figure 9. The color bar levels are odd here. I can really only see two levels. Also, these suggest that for all of these locations, most of the air reaching the site comes from over the ocean. There’s not discussion of how that might be impacting the results. Does this pattern look different on just fire/smoke days compared to the rest of the season?  

Figures S1 and S2 should be combined.

I find Figure S6 really interesting and would like more discussion about it. These fire seasons are mentioned in the introduction, but perhaps some more discussion about how the different fire seasons might impact the model performance would be useful.