Review for the paper „Methane retrieval from MethaneAIR using the CO₂ Proxy Approach: A demonstration for the upcoming MethaneSAT mission” by Christopher Chan Miller et al.

General comment:

This paper describes the XCH₄ retrieval algorithm used for the MethaneAIR airborne remote sensing instrument to detect and quantify CH₄ emissions. The algorithm uses the CO₂ proxy method to eliminate light path errors due to aerosols. The quality of the retrieved XCH₄ maps is assessed, and corrections for changes in the ISRF, a destriping algorithm, and a well-described denoising algorithm are introduced. Additionally, the detection limit of the combination of MethaneAIR and the CO₂ proxy approach is estimated. Finally, the emissions of several plumes detected during measurement flights are calculated with the IME method adapted to MethaneAIR.

The paper provides valuable information about the retrieval algorithm and post-processing of the retrieved XCH₄ maps, with promising results for future flights of the instrument. I recommend publication after the following partly major comments have been addressed.

Conceptual/Broader concerns:

The background of airborne remote sensing measurements of greenhouse gases needs to be referenced appropriately, with the MAMAP instrument being the first airborne remote sensing instrument to determine emissions from point and small areal sources (e.g., Krings et al., 2011, Gerilowski et al., 2011, Krautwurst et al., 2017, Krings et al., 2018). Relevant work of the airborne remote sensing instruments AVIRIS-NG and GHGSat-Airborne should be included, too. The relevant publications for the satellite remote sensing instruments GHGSat, Prisma, ENMAP, Sentinel-2, CO2M, and Carbon Mapper might also be referenced explicitly, as often a link is made to the satellite system MethaneSat. A corresponding statement on how MethaneSAT fits in this fleet of satellites needs to be included.

The algorithm is not sufficiently described. It is unclear what mathematical basis is used, especially which regularization and optimizer were chosen (e.g., Gauss-Newton, Levenber-Marquardt, …). Similarly, the convergence criterion is unclear. As this is the first publication of the algorithm, more information should also be added for the forward model, which, to my understanding, is an absorption-only case that should easily be written down as a formula. This is especially important as you do not only use a polynomial for albedo but also for an additive offset if I interpret it correctly (it is never stated explicitly). Why is this additional polynomial included? Do you include a spectral shift fitting parameter in your retrieval? In the current state, the state vector cannot be connected to the forward model in the current state of the description. Finally, I wonder why no theoretical sensitivity study of the algorithm w.r.t. differences in other atmospheric parameters that are not part of the state vector (different albedo types, aerosols, surface elevation, and aircraft altitude changes, as an example) has been conducted. Is a theoretical sensitivity study part of another manuscript currently in preparation?

For the estimation of the detection limit, there are still a few questions open:

1. How do you treat the background? Do you overlay the simulated plume over real data? Or do you also simulate the background?

2. How do you treat measurement noise in the simulation? Is it added afterward?

3. What is the emission rate assumed for the LES simulation?

A thorough uncertainty propagation has to be included for the emission rates and detection limit considerations. The emission estimates in the later parts of the paper are given without any uncertainty range. From the numbers stated in the article, and given that you take the wind from models, there is likely significant uncertainty due to that alone, but also from retrieval precision, and maybe uncertainty in missed parts of the plume may contribute. Please add an uncertainty breakdown to the emission estimates. Also, for IME, you state that you are close to the 1%-case stated in Varon et al., 2018. However, with 1.9% precision, the uncertainty likely is larger than the 70 kg h^-1 + 5% stated in the paper. Is an according analysis for a ~2% precision instrument planned?

The paper itself is, in parts, slightly unstructured and jumps between topics. I point out some parts in the line-by-line comments but suggest introducing a few more subsections where the subject of the section covers multiple aspects.

Specific comments line-by-line:

L32-44: This paragraph might fit better in the later parts of the introduction.

L56-57: I did not find anything in Lorente et al. about the 3% success rate, as well as the causes being the robustness of the full-physics approach. While satellite retrievals usually filter out more than 90% of the data, 3% seems low to me without a source. Additionally, how do the scientific algorithms perform?

L61/62: The instruments should be named, and MAMAP and GHG-Sat Airborne included and referenced. Additionally, a quick categorization apart from satellite/airborne would be helpful.

L66/67: The loosening of the retrieval accuracy is true. However, lower retrieval accuracy comes at the price of higher noise, false positives, and more substantial systematic errors.

L92-94: As the accuracy depends on the SNR (among others) and for smaller ground scenes at higher spectral resolution, the number of photons per detector pixel gets low, a source supporting that statement (e.g., simulation studies for MethaneSAT) would be helpful.

L114ff: Have the polarization sensitivity measurements been evaluated already in a paper? Or is this planned? This could have a significant impact on the accuracy and precision.

L131f: The CO₂ proxy method for methane for airborne measurements was first applied by Krings et al., 2011, which I suggest adding to the citation.

L146: The retrieval window for CO₂ contains hardly any background information, which would be available in the longer wavelength range for both MethaneAIR and MethaneSAT. Why did you exclude those wavelength ranges?

L148f: Why do you assume that the higher spectral resolution will allow you to only measure from 1597 nm upwards? Did you do any simulations supporting this statement? Otherwise, my previous question is more urgent: Why did you choose not to include the background, especially in the longer wavelength range above 1618 nm?

L178-L183: While this information is important, it breaks the retrieval description here. I suggest moving the paragraph to a more general “advantages of the retrieval” paragraph at the end of the section. Also, do I assume correctly that the forward model is absorption only? Or does it use aerosol assumptions without optimizing them?

L186: Does this mean you only optimize the troposphere in the retrieval, i.e., only below the aircraft? Would this change if you fly lower? And do you optimize the 13 levels separately? If yes, how does this work with only 1 – 1.5 DOF for CH₄? And how do you optimize/treat water vapor?

L195: “can lead to significant impact on the light focus on the detector” – could you please clarify what you mean by this and why the F-number is important here?

Table 1: The Albedo a priori is stated to be the MethaneAIR Radiance – averaged over which wavelength range? Also, an albedo uncertainty of 100% is very small for low albedos, e.g. 0.05, while it is pretty large for typical albedos of 0.2-0.3. Is this on purpose?

Finally, why do you include a radiance offset in the state vector, which is an additive offset, I assume?

Fig 3: Interestingly, the RMS after correction is lower at regions on the chip where the defocus is more prominent (as indicated by the larger Squeeze factor). While it aligns with the L1 prediction (how is it calculated from the radiance uncertainty?), it is still interesting that a lower resolution seems to perform better than a higher resolution (provocatively speaking).

L213ff: Why do you use the OCO-2 algorithm a priori uncertainties? Especially for H₂O, I assume some difference in uncertainty between a satellite covering the whole earth with the natural variability of water vapor globally and the local variability of water vapor during an airborne remote sensing flight of 2 hours. The same holds for the surface pressure uncertainty. And why use the UoL GOSAT Proxy retrieval covariance matrices for the CH₄ and CO₂ profiles and not covariances derived from the GGG2020 Priori Profile Software results?

L244: “temperature changes can defocus light at the FPA” – what do you mean by that? A defocus of the image in the FPA plane? How does this happen mechanically?

Figure 5: Interestingly, the largest biases and bias changes do not appear in the regions with the largest ISRF change. Or is this the bias, including the ISRF fit? Then, my question would be why the bias is largest in the area where the lowest correction was necessary.

Figure 6: First, I am curious about the dips/jumps in temperature. Do they correlate with some in-flight operations?

Second, I suggest adding a shifted orange (dashed?) line to visualize the correlation better so that the steepest gradient overlaps with the temperature curve.

L250: 10 s corresponds to approximately how many spectra? I.e., how robust is this median over, e.g., plumes? Do you do that as a running median, or each 10 s step is condensed to one median value?

L254: “The relative XCH₄ cross track bias is then derived by subtracting the mean of all the cross track background values.” I do not understand this. Can you provide a formula for the bias in cross track pixel i at time t?

L257: “is a common feature of other 2D grating spectrometers.” – please cite according examples.

L259-263: The second ISRF PCA score has a totally different time evolution than the first ISRF PCA score. The same holds for the XCH₄ PCA scores. Therefore, if the second ISRF PCA score has a similar correlation with both XCH₄ PCA scores, I would be hesitant to call the correlation between the first components high. For completeness, what is the correlation between the first ISRF PCA score and the second XCH4 PCA score?

Equation 5: Please expand this to include the complete formula from the pixel segments to the pixel bias.

L258ff: It might be worth putting the across-track bias correction part (which you later also call “destriping”) into its own subsection and separating it from the general ISRF fit result analysis. Additionally, state clearly at the beginning that the bias correction is to remove across-track biases, i.e., stripes. Do you have to calculate each flight's correction separately, or is the correction found for one RF sufficient?

Figure 8: While the correction largely removes the striping, there is still some residual track-to-track bias (and, in some cases, some residual stripes) left over. Did you quantify this?

L305f: How do you rescale the MethaneAir retrievals with the EM27/SUN XCO₂? Are you using the EM27 XCO₂ instead of the model XCO₂ in the proxy formula?

Figure 9, right side: are the lines and dots shifted such that the value is 0 for both dot and line at the first dot?

L309: Please provide threshold values for the filters and how and why you chose these values.

L313f: “This could be partially due to different EM27/SUN spectrometers used…”: Should that not at least partly be mitigated by rescaling the retrieval with the EM27 XCO₂?

L339: Why did you choose the standard V1 product, assuming you mean the operational product and not a newer version? As far as I know, significant improvements have been made since this version, and they might affect the comparison. Also, while the uncertainty due to the proxy normalization is sufficient to explain the bias, it might be that in the TROPOMI results, there is still a bias, which could either improve or worsen the comparison.

L358: Why use the a priori estimate and not the fitted albedo?

P21, second paragraph: Likely, this paragraph makes much more sense once the optimal estimator and regularization are named. However, as of now, it is a bit confusing, as it is not fully clear what the goal of the paragraph is.

Heading to 6.1: In this paragraph, additionally, IME is partly introduced, which should somehow be either reflected in the heading or, preferably, the introduction to IME is moved to a separate subsection.

L400: It sounds like the filter was already implemented by Frankenberg et al., 2016 and Varon et al., 2018, which, to my knowledge, is not the case. Is there a source where the advantages of this filter to IME are studied? If not, please provide some context for why it is advantageous here.

L406: U_eff is, to my knowledge, determined from the u₁₀ 10-meter wind from models. This is done via LES, but not for each target scene, at least in the papers you cited. Also, the MethaneAIR inversion paper (Chulakadabba et al., 2023, disc) states that u_eff is a function of log(U₁₀) + 0.6. Please also cite Chulakadabba at the first mention of the IME approach applied here.

L411f: What is the “3-sigma iteratively clipped mean”?

P23 last line: Please include the curves for the native resolution, maybe in the supplement.

L435-439: While this is a real consideration, the longer plume detection raises additionally the question of how to treat plumes cut off by the swath edge in the IME approach like the plumes in Fig. 19. This gets additionally important for highly variable plumes or puff releases.

L458: This is a very strong statement, especially as no MethaneSAT data yet exists. I suggest rephrasing that such detached plumes should be detectable by MethaneSAT.

After Line 460, I suggest the beginning of a separate subsection, as this is a significant point of the paper.

Figure 17: Please enlarge the red cross, and maybe color it differently (possibly black), as it is hardly visible and due to the blue in the figure also slightly hurting the eyes. Similarly, the red triangles in the Google Earth overlay are hardly visible. Maybe mark the whole region of compressors with a box?

L464: You state that the lower resolution of the simulation will lead to an overly pessimistic estimate of the detection limit. What is the reason for that?

L466: Why do you assume such a low wind speed? The detection limit is linked to wind speed, too, and wind speeds of ~ 5 m/s and higher are not unusual, especially in cloud-free conditions over sand or shrubland, not uncommon for O & G production sites.

L467: How exactly do you scale the plume sample? Do you multiply the base emission and all XCH₄ values by the same constant? Is this realistic?

L483: What is the uncertainty range on the 75% number of detectable sources? As this is quite a high number, and the wind speed uncertainty alone accounts for 15-50%, plus 70 kg h^-1 emission uncertainty, which for sources near the detection limit is also nearing 50%, I assume this number should have quite a substantial uncertainty.

L500: Is there a reason why Frankenberg et al 2016 is not on the list, as you cited it already in other places?

L503f: This is a very interesting thought. How many overpasses would you think it needs to assess that the emissions are stable enough to reduce revisiting times?

L509: Which of the nine overpasses do you consider covering the plume fully? In Fig. 19, I see six plumes I would consider “cut off” and only three plumes fully sampled.

L530: While true, such large persistent leaks have also been detected by TROPOMI, GHGSat, and EMIT, e.g., in Turkmenistan. Therefore, it enhances the capabilities to inform operators but is not the only instrument for that, which is implied by the statement.

Figure 20: Was this plume recorded in one go, or has it been sampled over multiple tracks? If the latter is the case, over which time frame was the plume recorded, and how do you treat the wind in this case for the emission estimate?

L533: “the operational MethaneSAT” – please add “designated” as MethaneSAT is not flying yet.

L545f: Does this assume purely random noise? What about systematic effects that cannot be reduced by averaging?

Technical corrections/Typos:

L220: “clouds with high-optical”: here a word seems to be missing

L237: “… is the mean …”: “The” missing

L264: mechanistically sounds strange here – do you mean “directly”?

L267: please replace “(top right”) with the according subfigure number

L299: “2-3 hours of MethaneAIR observations” – please add at which altitude, as this determines the swath width.

Figure 12: Typo in the caption: “wiht” -> with

Figure 14: Label and ticks of the y-axis on the right plot are covered

L511: There is no reference in the bibliography for the US EPA GHG reporting program, or it is not listed under this name.

L512: The study's measurements were taken ten years ago, and the paper was published eight years ago, so please remove the “recent”.

L525: Please give the 2.7 Tg a^-1 also in kg h^-1 for better comparison with the other values in the paragraph.

L543/544: two times “strong”

List of citations:

Gerilowski, K., Tretner, A., Krings, T., Buchwitz, M., Bertagnolio, P. P., Belemezov, F., Erzinger, J., Burrows, J. P., and Bovensmann, H. MAMAP – a new spectrometer system for column-averaged methane and carbon dioxide observations from aircraft: instrument description and performance analysis. Atmospheric Measurement Techniques, 4, no. 2:pp. 215–243. doi:10.5194/amt-4-215-2011. URL http://www.atmos-meas-tech.net/4/215/2011, 2011.

Krautwurst, S., Gerilowski, K., Jonsson, H. H., Thompson, D. R., Kolyer, R. W., Iraci, L. T., Thorpe, A. K., Horstjann, M., Eastwood, M., Leifer, I., Vigil, S. A., Krings, T., Borchardt, J., Buchwitz, M., Fladeland, M. M., Burrows, J. P., and Bovensmann, H. Methane emissions from a Californian landfill, determined from airborne remote sensing and in situ measurements. Atmospheric Measurement Techniques, 10, no. 9:pp. 3429–3452. doi:10.5194/amt-10-3429-2017, 2017.

Krings, T., Gerilowski, K., Buchwitz, M., Reuter, M., Tretner, A., Erzinger, J., Heinze, D., Pflüger, U., Burrows, J. P., and Bovensmann, H. MAMAP – a new spectrometer system for column-averaged methane and carbon dioxide observations from aircraft: retrieval algorithm and first inversions for point source emission rates. Atmospheric Measurement Techniques, 4, no. 9:pp. 1735–1758. doi:10.5194/amt-4-1735-2011. URL http://www.atmos-meas-tech.net/4/1735/2011, 2011.

Krings, T., Neininger, B., Gerilowski, K., Krautwurst, S., Buchwitz, M., Burrows, J. P., Lindemann, C., Ruhtz, T., Sch¨uttemeyer, D., and Bovensmann, H. Airborne remote sensing and in situ measurements of atmospheric CO2 to quantify point source emissions. Atmospheric Measurement Techniques, 11, no. 2:pp. 721–739. doi:10.5194/amt-11-721-2018, 2018.

See supplement for the full review.