Review of Goudar et al. Plume detection and estimate emissions for biomass burning plumes from TROPOMI Carbon monoxide observations using APE v1.0" submitted to GMD.

Goudar et al. presented an automated way to determine CO emissions from fires using a combination of observational data and modeling. This is an interesting first study and the authors demonstrated well the potential of this approach although there are several limitations to this approach. The paper is sometimes hard to follow and could be improved for its structure and writing (especially in the introduced). It should be published in GMD, after few (mostly minor) corrections.

Comments

One main comment is that the derived emissions are not compared to any other CO emissions estimates, so it is difficult to judge how good these are.  I understand that this could perhaps be the subject of a separate paper but this is not mentioned by the authors. I would suggest that the authors prepare supplementary material with the estimated CO emissions e.g., in the form of a spreadsheet so that the values could be compared with other emission values by others.

Introduction

-p3: You list 4 methods but discuss only three. Please discuss the missing one: IME.

Section 2

-section 2.1.1: the choice of r_max=4km is puzzling. Will it not automatically discard the megafires from the analysis?  Maybe this should be emphasized in the text?

-Figure2. I don't understand this figure. What is the message? Is this supposed to be good? Many fire-counts are not considered by DBSCAN. Why? Because there are less fire counts than n_min? The selection of fires does not consider any criterion on the fire intensity (FRP). Why (not)?  Also noticeable are the fire counts over sea. To what these pixels correspond?

-Figure 4 shows a relatively isolated CO plume but how is the plume detection working for the other plumes close to each other?

-p7, l 142: What is a 'connected region'. What is the CO VCD criterion related to this?

-section 2.3.1: Is the re-centering needed? Or is to facilitate the Gaussian fit? Please clarify

-p11 l219: what is the name of the model used for the simulations? Is it defined somewhere?

-The authors attempt to account for wind variability in the horizontal and vertical dimensions. However, there is an additional flux term due to the partial derivative of the wind which is not accounted for (see the divergence method of Beirle et al., Sc. Adv, 2019). Can you quantify this?

Section 3

-P12: Going from 622 to 196 plumes is in a way disappointing. Does that mean that only ~1/3 of the fires made a meaningful CO signal in the TROPOMI data? Please elaborate.

-The discussion on errors should be expanded. The error characterization based on standard error (Eq.5) does not account for any systematic error and mixes random errors and real CO flux variability, so it is not a very good metric. I would propose including a table summarizing all error sources and estimating them. 

-Section 3.1.1: Generally, zlag seems higher than zc which is in contradiction with Fig 7b. It is confusing. Perhaps it is due to an unfortunate choice in the illustration?

-P13, l273: the author states: “a relation between plume height rise and these two variables can be expected as higher FRP means higher temperature which heats up the air, leading rise of the warm air.” However, this process of self-heating is likely not accounted for in the Lagrangian modeling. In fact, the approach presented here is in fact limited to a certain range of fires not too low (because of the limit of detection of the satellites) and not too big (because self-heating and other non-linear processes are not well represented). Therefore, Fig 9b is misleading. The differences are very small, but it does not mean Ec is good because the Elag is not representing all the physics.

Typos/text suggestions

-acronyms are sometimes defined multiple times. Please define acronyms only once.

-Both acronyms ‘Tropomi’ and ‘TROPOMI’ are used in the text. Please use one or the other throughout the text.

-several suplots /maps have no units. Please define the units for all figures.

-several figures or subplots would be better placed in the supplementary material: Figs 7c-e, Fig 8.

-P2, l27: ‘CO in atmosphere and Shi et al.’ -> ‘CO in the atmosphere. Shi et al.’

-P2, l30: ‘has been on increase’ -> ‘has been increasing’

-P2, l37: ‘between two measurements’ ->‘between the two measurements’

-P2, l50: refer to the use of VIIRS for methane cloud masking doesn’t help the clarity of the text.

-P3, l74: ‘deliberated’-> ‘discussed’

-P4, l91: ‘Mostly, an emission plume created by a burning’->’Essentially, a plume emitted by a fire’. The sentence states that a fire in a single VIIRS pixel cannot be detected by TROPOMI. Why not? On what is based such statement?

-p4, l107: ‘constrained’ -> ‘restricted’

-p6, l129: l129: Gaussian filter : is this a 2D convolution?

-Fig5d is not appearing in the manuscript.

-p18, l362: doesn’t -> does not

-p18, l370: ‘reliable’ is subjective. You don't have any way to assess whether it is more reliable or not.

P18, l384-387: What about overlapping plumes from different fires? Isn't there a way to improve on this?

“Plume detection and estimate emissions for biomass burning plumes from TROPOMI Carbon monoxide observations using APE v1.0” by Goudar, Anema, Kumar, Borsdorff, and Landgraf

The manuscript describes a method for the automatic detection of CO fire plumes and the quantification of associated CO emissions. The method is applied to two regions: Australia (October 2019) and USA (September 2020); the exact extent of these regions is unclear. Fire locations were obtained from VIIRS active fire data. CO plumes from these fires were automatically identified in TROPOMI data using several filters (a Gaussian filter, presumably for noise reduction; a Sobel filter for gradient detection) and publicly available tools (the “marker-based watershed” algorithm, the “label” algorithm). CO emission rates were then estimated using the cross-section flux method. Some uncertainties in the emission results are discussed. A 97.9 % success rate in plume detection and emissions calculations is quoted.

The goal (automatic detection of CO fire plumes and emissions quantification) is relevant. However, the manuscript fails at demonstrating to what extent does the method described reach the goal. Main issues include:

- unclear how “fully automated” is this method. Numerous thresholds are applied but their values are not justified. Also: how universal are the numerous threshold values required by the method? How reliable are the results when applied to other locations and times?

- unclear what percent of all VIIRS fires are detected and their emissions successfully quantified with this method. The quoted 97.9 % success rate seems too high (considering that many plumes seem to be rejected due to different reasons) and is not properly justified.

- unclear how valid the plume height values and the emission values quoted in the manuscript are, since no attempt was made to validate those with respect to in situ measurements.

More specific issues include:

- details needed for analysis replication are missing, e.g., filter size, filter formulation.

- tools are not described and, thus, become black boxes to the reader.

line 5: please explain “APE”

7: please explain “VIIRS”

10: “IS4FIRES”?

15: should month names be spelled without abbreviation?

16: “The emissions were severely under-predicted”. With respect to what? Were there any in situ measurements used to validate the emissions calculated here?

22: please clarify “idealized cases”

25-26: please provide reference for “it is a weak greenhouse gas”

30: “CO emissions due fossil fuel burning has been on increase”, consider rewording to “CO emissions due to fossil fuel burning have increased”

32-33: please consider rewording “Thus, it becomes essential to understand the effect of CO on air-quality and climate by measuring it accurately on a global and local scales which helps to quantify CO emissions” for readability

65: please correct “using the the wind”

74: since the present tense was consistently used before (lines 71-74), please consider rewording to “The study results are deliberated in Section 3”. Consider using “discussed” instead of “deliberated”.

80: “extracts TROPOMI CO data”

87-88: please explain acronyms as they are introduced

90: Should “Furthermore” be “From now on” or similar?

91-92: Could missing fire counts (due to missing VIIRS pixels because of, for example, clouds/smoke) result in fires and, thus, in plumes not being identified by the automated plume detector? Some CO plumes may only be detectable downwind from the fire, if clouds/smoke mask the fire, which is not uncommon.

92: also 5.5 x 7 km^2, since all but one of the cases analyzed here postdate August 2019.

95: “low density areas”?

98: Please correct to “The minimum number […] has been set to n_min=10” Also, why 10?

104: “for part of data granule S5P_OFFL_L2__CO_____??? over Australia”

105: a granule would be much larger than 41 x 41 pixels; please consider using “subset” instead of “granule” here as well as in lines 107, 108, 109, 110, 111, 114, 128, and Fig. 3 caption. (A granule would have whatever size is covered by a whole TROPOMI file.)

110: how were the 80% and 85% thresholds selected? Also, please explain the meaning of QA>0.5

Fig. 2: “fire-counts that were not clustered” was it because there were less than 10 fire counts within a 4 km radius? Please explain. Also: why 41x41 pixels? Why 7x7 pixels?

116: please explain “gold standard data”

121: “Thus, the watershed algorithm segments the regions into valleys and mountains (CO enhancements) based on a given marker” Valleys suggest low CO regions and mountains high CO regions, i.e., plumes. However, line 125 seems to say that what the algorithm does is to look for low/high boundary zones, i.e., zones of maximum slope change. Please clarify.

How does this method perform compared to simply calculating the background value in a TROPOMI scene and then selecting clusters of pixels above that value?

Fig. 4: the map shown in panel (b) seems to display zones of maximum slope change or, as the text states, zones of “gradient” change. However, panel (b) is labeled “Elevation map”, which does not seem very appropriate. Please reword.

Also, it looks like only one of several plumes in this 41x41 TROPOMI subset is detected, even though several fires are shown in the same 41x41 subset in Fig. 3; please clarify what happened to the other plumes in this subset, including the largest of them all, in both size and CO value: did the algorithm identify all of them or just one of them?

4c-4d: plumes detected are much shorter than the actual plumes.

125: Ielev does not show a continuous variable (like elevation, or CO value) but it rather shows where the maximum change in that variable occurs. Consider renaming it to Iedge or similar.

128-162: Please clarify if this example illustrates the process followed to either 1) identify a single plume in the 41x41 TROPOMI subset; the process is then repeated for each of the remaining plumes in the subset or 2) all plumes in the 41x41 TROPOMI subset at once. If 1) is true: please clarify text. If 2) is true: most plumes are missed, please discuss.

129: “First, high frequency components of the CO-image are reduced by a Gaussian filter” Please explain, is that to remove noise? What is the size of the filter, is the size constant for all plumes, how was it selected?

130: “the elevation map Ielev is computed using a Sobel operator” Describe with an equation what the Sobel filter does. Also, “elevation” seems incorrect here, since the Sobel filter would highlight zones of maximum change in slope in the input. Consider changing to Isobel or similar.

134: stating that Imark is initialized with zeroes would suffice, no need for an equation.

136-140: for clarity and simplicity, consider rewording to “[…] clear CO enhancement. Pixel Imark(i,j) is considered CO enhanced (i.e., Imark(i,j)=2) if Is(i,j) is either above the median of Is or above the mean of the 15x15 pixels centered at Is(i,j). Otherwise, Imark(i,j)=1. For our example in Figure 4 [...]” (no equations needed). Also: why 15x15 pixels? Is this size fixed, or does it change from plume to plume?

140-141: the meaning of the last sentence in the paragraph is unclear. It looks like the result of the step that was just described (where Imark is populated with either ones or twos) is illustrated by panel 4d, not 4c. What is panel 4c? How is it relevant? Please comment on the plumes present in 4a and 4b but absent in 4d; one of the absent plumes was the largest of them all, in both size and CO values.

142: how does this new tool work?

149: Where does “14” come from? Eq. 3 is not needed, since it does not add to what’s already in the text.

How is this plume detection algorithm better than a simpler approach, such as identifying groups of CO pixels with values above that of the background? It looks like the latter would have sufficed to identify all the plumes in this 41x41 TROPOMI subset, while this plume detection algorithm (at least according to Fig. 4) missed most of them.

164. Please provide reference for the cross-sectional flux method.

169. (Here and elsewhere in the manuscript) wind velocity from ERA5 data is expressed in the manuscript as “u”. Usually (and that includes the ERA5 dataset) u represents the E-W component of wind; v would represent the N-S component. Is u in the manuscript really the E-W component of the wind? Shouldn’t the wind velocity be calculated according to the plume’s direction? Please clarify

175. Please clarify “The plume line results from a second order curve fit through the pixel centers of the identified pixels”

176. Why 2.5 km?

177. Why 500 m?

Fig. 5 caption: please correct typo, “transaction” should probably be “transect”. c): why missing value at 0 km from the source?

182. Please explain what are the terms H0, H1, and A0

203. Fig. 6 shows two distinct plumes approximately 100 km long each, resulting from two fires 50 km apart. According to the text, both plumes were rejected by this algorithm because they were too close to each other. How close is too close? What’s the minimus plume size detectable with this method? These and other limitations of the method presented in this manuscript should be discussed both in the abstract and in the conclusions sections.

What are the dashed lines in Fig. 6?

197, 198. “remove overlapping fires”. Fires like those shown in Fig, 6 are not overlapping, both the fire sources and the plumes appear quite distinct. Consider rewording “overlapping” by “closer than … km”.

207. Please explain briefly why is the uncertainty in injection height about 500 m.

220. does “on the right-hand side” refer to equation 8? If yes, then consider rewording to “The velocity v” or similar. If not, please explain what does it refer to.

224. Why 6 hours?

Fig. 7 b) values at distance=0, 2.5, and 5 km from the source (i.e., at the fire source and close to it) are missing; please explain. The text states elsewhere (e.g., l. 366) than plumes go higher away from the source but the opposite behavior is shown here.

234-245. Please quantify what proportion of plumes are rejected due to: lack of GFAS injection height, disagreement between Lagrangian particles flow direction and actual plume direction, wind velocity below 2 m/s.

244. Please clarify if “the velocity at the TROPOMI measurement time” refers to wind velocity.

260: “To conclude, presented automated algorithm can successfully detect plumes and compute emissions for ≈ 97.9% of the cases.” It looks like the percentage of plumes detected is much lower that that. How was this figure calculated? A few sentences earlier the text says “the plume detection algorithm [...] identified 196 plumes among 622 cases” and lists numerous cases which were not successfully processed due to a number of reasons. The detection rate quoted does not seem feasible, unless relevant qualifiers are missing.

264. How is it decided what is the number of transects along the downwind direction to be considered? Does this number change from plume to plume, or is it universal?

265. Is “w.r.t” acceptable in a manuscript?

Fig. 9. The blue crosses and blue dots are too similar to tell them apart. Consider using other symbols or separate colors instead.

3.1.1. As expected, emissions calculated using plume height Zlag and Zc differ when the heights themselves differ. Unclear if results vary from Australia to USA; thus, please consider using the same symbol/color for data points from both locations. Unclear if all 4 panels are relevant; some seem redundant. Consider showing one panel with height difference (between zlag and zc) versus emissions difference and another panel with height difference (between zlag and fix z=100 m) versus emissions difference.

277-287. Unclear where the discussion is going until the last sentence “although the overall effect of the Lagrangian estimate of the plume height on the emission estimate is minor, we could identify several cases where the emissions estimate become more reliable.” Consider starting the paragraph with this sentence and add a very brief description of relevant data.

295. 10% change in emissions seems to be much smaller than some of the emission uncertainties discussed later on (e.g., l. 344-345). Also, a 10% variation in emissions was qualified as “minor” elsewhere (l. 19). Please discuss.

325. “It should be noted that this uncertainty has been reduced to 3.4ppb in the newer versions of L2 product”. How much is that in percent value?

328. Please clarify “as the pixel size of TROPOMI is high”.

332. “TROPOMI”

339. Table 2: please include percent differences.

351. How universal is this method? It seems to have many steps requiring thresholds which seem to have been selected based on specific examples. Would the same thresholds result in the desired results if the method was applied to other regions, other time periods?

355. Please clarify: 97.9% of what? Many plumes were rejected based on proximity to other plumes, lack of injection height data, … Such high percentage seems off.

363. 22 out of how many cases? Alternatively, please provide a percent value. Otherwise, “22” alone is not informative.