the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 15 Jul 2024
An Investigation of the Impact of Canadian wildfires on US Air Quality using Satellite, Model and Ground Measurements
Abstract. Canadian wildfires transport large concentrations of particulate matter into the United States, leading to various impacts on surface temperature, radiation balance, visibility, and exacerbating pollution-related respiratory conditions. Using a combination of surface, satellite and numerical models, this study quantifies the increase in surface fine particulate matter (PM2.5) in the Continental United States due to long-range transported smoke from Canadian wildfires during a wildfire episode from August 9th to 25th, 2018. As a widely used indicator of surface pollution levels, satellite-retrieved AOD can provide crucial information on columnar pollution mass. However, the daily spatial coverage of satellite AOD is restricted due to cloud cover. In order to quantify the daily changes of surface pollution, we fill in the AOD gaps by utilizing simulated 10-m spatial resolution AOD from a chemistry transport model (CTM). In addition, different processes affecting smoke vertical and horizontal transport are examined using two CTM simulations. Meteorological variables associated with these processes are then selected along with the gap-filled AOD product to assess the surface pollution using geographically weighted regression (GWR) and random forest (RF) models. It is found that synoptic pressure systems dominate the horizontal transport of Canadian smoke, and the ascending air and robust winds from a low-pressure system contributes to the long distances of the transport path. Other processes affecting the vertical distribution of pollutants, including boundary layer entrainment, precipitation and terrain-induced vertical mixing are also analyzed in this paper. Our results show that Canadian fires caused a substantial increase in total PM2.5, reaching up to 13 % (62 μgm−3) across different US EPA regions during 2018 August wildfire event.
- Preprint
(6539 KB) - Metadata XML
-
Supplement
(3725 KB) - BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2024-1781', Anonymous Referee #1, 22 Nov 2024
This manuscript demonstrates the application of a method that combines information from surface monitors, satellites, and a 3D model to estimate wildfire smoke impacts over a 17-day time period. There are significant issues with the writing and some aspects of the science, so I feel the manuscript requires major revisions before it can be published.
Major commentsIn general, the writing is imprecise and there are quite a few typos. I’ve pointed out many examples in the specific comments. These issues caused the review to take about twice as long as it would have if the authors had edited their manuscript better. It should not be the reviewers’ responsibility to edit writing mistakes and imprecision, and it adds to the stress on the peer review system.
Section 4.3 and equation 9. The method will fail when a significant fraction of the smoke impact is out of the PBL. For example, if the AOD without smoke is 0.1 and the AOD with smoke is 1.0, but 100% of the smoke is in the free troposphere, the real impact of removing smoke should have no effect on surface PM2.5. However, equation 9 would attribute 90% of the surface PM2.5 to smoke (as opposed to the correct value, 0%). This erroneous calculation could have a large impact on the results in Table 3, which has some of the main results of the paper that show up in the abstract.
Sections 5.1-5.4 seem very scattered, and I am not currently seeing how they contribute to the main points of the paper. I recommend removing some of these sections and better integrating the remaining parts with the following sections.
How is the modeled AOD defined in a cloudy column? Aerosol scattering blows up as f(RH) approaches 100%. How is this handled in a cloudy gridbox and how do the assumptions affect the results?
Specific commentsTitle: Is “measurements” also applying to “satellite”? If not, then it should be “satellites”. If yes, then it’s confusing to have “model” in the middle.
L8: 10-m grid spacing is impressive.
L16: unclear what these numbers mean. 62 ug m-3 would be far more than a 13% increase considering typical non-smoke concentrations in the US are <10 ug m-3.
L22: 92 mortalities over what time period?
L23: I’d say “has been found to be 3-4 times greater” given that this is just one study and smoke and non-smoke composition varies.
L25-28: Is this sentence about wildfire smoke PM2.5 or PM2.5 in general?
L30-31: Using past tense in a sentence about the future. I’d say “have been projected to increase from…”.
L32 and in general: Focus on making the topic sentence be an intro to the full paragraph. This paragraph moves a long way from FRP and injection heights.
L33-36: Boreal injections have been found to be higher on average than tropics in the MISR plume height climatology (https://www.mdpi.com/2072-4292/10/10/1609), so the contrasting of limited cases seems to give the wrong impression.
L47: Volatile organic compounds, by definition of being volatile, do not condense. Some of the oxidation products of VOCs can condense. Also, the citations here are not dedicated investigations of the specific processes listed in this sentence, though these studies do exist.
L54-59: Of what data sources?
L59: Which of the approaches above are the traditional ones?
L63: Strange to say that AOD is an *essential* indicator of surface pollution since it is a column measurement. There are much more direct ways to get surface PM (e.g., an in situ measurement), so AOD is not essential. The PM2.5:AOD ratio should be discussed in this paragraph.
L85-87: This sentence makes it seem like there are 2 different WRF-Chem domains, implying 2 different simulations. In the later WRF-Chem description, it seems like this is not the case.
L116: T2M, U10M, and SP are not directly related to the AOD:PM2.5 ratio. I would guess that including them could lead to overfitting.
L128: Should “in-line” be “online”?
L130: Both aerosols and clouds have microphysical processes (condensation, coagulation, etc.), so please say “aerosol-cloud interaction” here. Similarly, since aerosols have chemical and physical processes, what does “interactions between chemistry, aerosols, and physics” mean here?
L140: 2-moment
L146: “All the key species… are included” is a bold statement. Are you sure?
Table 1 or accompanying texts: What are the lat-lon limits of the domain?
L176: Double “the”
Section 4.1.2: How is AOD defined in the CTM in a cloudy column (see major comment)?
L180-182: This sentence is confusing.
Section 4.2: Please explain how variability in plume height is captured in the 2 methods in this section.
Section 4.3: See major comment.
L264-265: Is this other study investigating the exact same time period? If not, is the comparison relevant?
L268-272: These several sentences were all confusing.
Figure 3. It’s very hard to read the text on this figure. Can we see well enough what we should be taking from the maps? Would something like 500 hPa geopotential heights be better for showing the steering of smoke? Should the maps include Canada?
L281: Is there an estimate of dry dep? Normally dry dep is fairly slow for accumulation-mode sizes.
Section 5.3 and 5.4. These just show special cases for specific locations while the rest of the paper shows smoke affects more broadly. I don’t think these sections add much, and the manuscript may benefit from cutting them.
Figure 6 needs axis labels. Also, neither the caption nor the text says which day it’s for.
Figure 10 gets discussed before Figure 9.
L365: “AOD is crucial” same comment about the use of “essential” earlier.
L372-374: It’s not clear to me that these are actually good-enough performance values, particularly the RMSEs. Can you give more context to these values?
Figure 11: Please make missing data white or gray rather than blue, which is also the color of low AODs. It’s currently very hard to tell what is missing data vs. low AOD.
L418: Units of RMSE.
Figure 14 and Tables 2 and 3: Many (most?) readers won’t know the EPA regions (at least not precisely). There needs to be a map showing the regions.
Figure 14: For RF and GWR, are the averages calculated only from the locations of the ground stations (to be consistent with the average of the ground stations)? Or are the averages from RF and GWR calculated from the entire region (in which case, differences with the monitors could be due to differences in the locations used in the averaging)?
L454-455: Please explain the “FRP < 10000 MW” threshold. MW is an energy rate. How is a rate summed over a year (should the integral of a rate over time be in Joules, not Watts)? Or is it an average rate over the season?
L458-459: I’m not sure how a 17-day investigation alone can give insight into what is happening in the long-term multi-year trend. There is no investigation of how Canadian smoke has changed during the multi-year time period.
L467-481: A very important thing missing from these limitations is uncertainties in the plume heights. (1) It affects horizontal transport speed and direction in WRF-Chem, causing errors in the gap filling. (2) It affects the PM:AOD ratio, and if there’s variability in plume height between monitors, this will cause errors in the inferred surface concentrations. (3) It will certainly have issues within equation 9 in getting the smoke fractional contribution to PM, as discussed earlier.
L476: Figure 4 seems like more monitors than I would expect from FRM alone. Are you sure you aren’t using FRM+FEM monitors?
L506: particulate *matter*
L522: I don’t see how a 17-day study confirms fires becoming the dominant source.
Citation: https://doi.org/10.5194/egusphere-2024-1781-RC1 -
RC2: 'Comment on egusphere-2024-1781', Anonymous Referee #2, 25 Nov 2024
This study combined satellite observations, model simulations, and ground measurements to assess the impact of long-term transport smoke from Canadian wildfires on U.S. PM2.5 levels during a 17-day study period. This topic is attractive but the manuscript is not well structed and the presentation of results/conclusion is confusing. The text is hard to read and the figures (too many figures) are in low quality. Detailed comments are listed below
- The structure of this manuscript needs to be carefully re-design. For example, the first paragraph of introduction talked about PM2.5, fire-related PM2.5, fire-related PM2.5 health effects and again PM2.5 health effects without a logical sequence.
- The citation format needs to be double check throughout the text.
- When averaging 1-km resolution AOD to 0.1-degree grid, how to deal with missing AOD data?
- Description of the AOD gap-filling method is not clear. In the third step, “calculating the ratio of pixels with valid MAIAC AOD over the total pixels”, are the total pixels satisfied the pixel selection conditions defined above? If the ratio is less than or equal to 80%, the AOD is filled by geo-weighted method including all the pixels within the pre-selected box or including only the pixels satisfied the pixel selection conditions defined above?
- Regarding the hyperparameter of RF model, only the number of tree and number of parameters are considered and the model is over-fitted. The hyperparameter number of leaves (nodes) needs to be optimized to deal with the over fitting issue.
- In section 4.3, the control PM2.5 is calculated from simulated AOD, filled AOD and predicted PM2.5. I am not sure what is the contribution of AOD here. The control PM2.5 can be directly assessed from the estimated PM2.5 by calculating the Canadian fire contribution ratio as the simulated control PM2.5 divided by the simulated true PM2.5. Due to the non-linear relationship between AOD and PM2.5, the AOD-based calculation may introduce larger uncertainty. What’s advantage of including AOD in this calculation?
- The result reported in line 295 directly compared AOD maps in Figure 2 and PM2.5 maps in Figure 4 and discussed the contribution of Canadian wildfires vs. US local fires to surface pollution. However, AOD and PM2.5 are not linearly correlated. This discussion here is with large uncertainty.
- Section 5.3 discusses the PBL changes at hourly scale with diurnal patterns that is not consistent with other results at daily scale. This section can be removed.
- Line 365-370, figure 10 appear early than figure 9 in the manuscript.
- The different temporal trends reported by local stations, GWR and RF indicating that the model predictions are nor solid. What is the performance of the temporal trends of PM2.5 from EPA stations and from RF/GWR estimations at the same location? Please adjust the model prediction before further analyses.
- In Table 3, region 1 on Aug-22 with >12.2% Canadian wildfires’ PM2.5 effects but only 3.3 µgm-35 concentrations? Please double check the number in table 2 and 3.
- Figure 1 What is “r”? Please label it in the figure or in the caption.
- Figure 3 There are too many numbers in the figure.
- Figure 4 Are these purple dots showing missing measurements or showing measurements of 0 μg/m3?
- Figure 6 What date is this figure about?
- Figure 11 (a), missing AOD should be shown as missing not as 0 in a continuous color scheme.
- Figure 12 and Figure 13, the caption should be “100-fold CV” rather than “k-fold CV”. Some key parameters, e.g. R and RMSE, should be added in the figures.
- Figure 14 Why are some data points not on the line?
Citation: https://doi.org/10.5194/egusphere-2024-1781-RC2 -
RC3: 'Comment on egusphere-2024-1781', Anonymous Referee #3, 27 Nov 2024
In this paper, the authors used a different approach than what we usually use to estimate the impact of Canadian wildfires on PM2.5 concentrations in the United States. The traditional approach is WRF-CHEM sensitivity simulation. Two simulations are performed with and without Canadian wildfire emissions. The simulation differences are used to estimate the contribution of Canadian wildfires to the increase in ground PM2.5 concentrations in the United States. In this approach, observational data (ground stations or satellites) are only used to compare with WRF-CHEM results to demonstrate the reliability of model simulations. Therefore, the observational data have little impact on the final results and conclusions.
The highlight of this paper is that the authors incorporate observational data into the estimation, which includes not only ground PM2.5 observations but also satellite AOD observations. The authors first fill the AOD observation gaps caused by satellite limitations through Kriging interpolation and CTM interpolation, and then use two machine learning methods, GWR and RF, combined with PM2.5 ground observations to estimate the daily PM2.5 concentrations at each point in the model simulation area.
I don't have too many questions about the treatment of AOD, but based on Figures 12 and 13, I doubt whether the PM25 surface concentration can be correctly inverted from AOD even with machine learning methods.
As can be seen in Figures 12 and 13, the probability that the 2D PM25 ground concentrations predicted using the 3D columnar data (AOD) match the observed values is low. For example, if the PM25 observations are in the range 0-10, the PM25 predicted by GWR and RF are in the range 0-50 (k-fold plot); the colors representing higher probabilities are overwritten by the colors representing lower probabilities. This may be due to the large markers used by the authors when plotting the graphs. But overall, Figures 12 and 13 are not sufficient to prove the most critical assumptions used in this paper and should be revisited.
Another big problem related to the above discussion is Equation 9. In this equation, the authors claim that the AOD/PM25 relationship derived from WRF-CHEM simulations that include Canadian wildfire emissions can be used for WRF-CHEM simulations that do not include Canadian wildfire emissions. Ground station PM25 observations must be used to derive the AOD/PM25 relationship, but the ground stations only match the model simulation scenario with Canadian wildfire emissions. Since the control runs have no matching observations, the AOD/PM25 relationship in the control is theoretically unsolvable, and using other relationships instead is incorrect.
Apart from these issues, the rest of the paper is well organized and logically clear.
Citation: https://doi.org/10.5194/egusphere-2024-1781-RC3
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 382 | 128 | 18 | 528 | 28 | 10 | 10 |
- HTML: 382
- PDF: 128
- XML: 18
- Total: 528
- Supplement: 28
- BibTeX: 10
- EndNote: 10
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1