Review of AMT manuscript egusphere-2024-1551, by Wells et al. (2024)

The manuscript presents global retrievals of four VOCs - methanol, ethyne, ethene, and HCN - from the polar-orbiting CrIS satellite sounder. First, their spectral signatures are identified in CrIS measurements taken in a fire plume from the 2019/2020 Australian bushfires. The retrieval methodology, which is based on a neural network approach, is then explained. Subsequently, the global distributions of the total columns of these VOCs are presented and discussed. CrIS data of methanol, ethyne, and HCN are compared with total columns derived from ground-based FTIR measurements at various NDACC sites, revealing notable correlations between spaceborne and ground-based measurements. Finally, the CrIS-derived distributions of these VOCs are compared with GEOS-Chem model simulations, and the discrepancies are analysed.

By presenting new satellite products of VOCs and conducting inter-comparisons with independent data and model simulations, the manuscript aligns perfectly with the scope of AMT. Although these four species have already been retrieved by other nadir-viewing or limb-sounding satellite instruments, these new CrIS products are of particular interest to the research field, given the increasing need for satellite observations of VOCs. With its low instrumental noise and early afternoon overpass time, CrIS shows promising capabilities for measuring VOCs. Overall, the paper is well-written and clearly structured, with results that are well-discussed and effectively supported by figures.

The retrieval framework is not new and builds upon the heritage of past work using other satellite instruments, which should be clearly recognized in the manuscript. The main new element is the introduction of a predictor for the vertical atmospheric distribution of VOCs, based on the relationship between the CrIS measured signal and GEOS-Chem vertical profiles. However, I have major concerns about this, detailed further below. The primary issue is that it might fall short of representing the full range of atmospheric variability of VOCs, as it constrains the CrIS retrievals within the range of VOC profiles from GEOS-Chem. As a global model, GEOS-Chem has a limited representation of the spatial and vertical variability of VOCs. For the same reasons, I do not think this predictor significantly improves the comparison with ground-based FTIR measurements, as would accounting for the a priori profile and vertical sensitivity respective to each instrument. I also have concerns about the feasibility of accurately retrieving these VOCs under all conditions, particularly at background concentrations, which might be within the instrumental noise. The example of VOC detections during the Australian bushfires presented in the manuscript is not convincing, as this event provided very favourable conditions. I have additional comments detailed below. I believe these comments need to be carefully addressed before proceeding further with the manuscript.

Major comments

• Section 2.1: The cloud screening appears to be based on a simple temperature difference between the surface temperature and the brightness temperature from a single channel. The source of the surface temperature is unclear. If provided by CrIS, it is uncertain how the surface temperature can be obtained if the scene is cloudy. Additionally, when the surface temperature is low, clouds might have a similar temperature, making them difficult or impossible to detect. Using a single channel is also risky because it might not indicate residual clouds, while such clouds can affect the HRI and significantly bias the retrieved columns.
• Lines 108-110 and Section 2: The entire retrieval methodology employed here, including the use of HRI and neural-network-based inversion, draws from approaches previously developed for the retrieval of ammonia and other species (including methanol) from IASI satellite observations, as described in published works (e.g., Whitburn et al., 2016; Franco et al., 2018). While it is standard practice to adopt already established methodologies, it is necessary to acknowledge this influence and explicitly reference the seminal studies.
• Section 2.2: Detecting the spectral signatures of VOCs in CrIS radiances from a concentrated plume during the 2019/2020 Australian bushfires does not imply that these species can be detected elsewhere, in other conditions. These Australian fires represent by far the most favourable conditions in the last decade for detecting the spectral signatures of any weak absorbers in satellite data (i.e., exceptional gas concentrations and very high injection height). The literature is full of such examples. Hence, this event cannot be used to demonstrate that VOCs can be detected and retrieved worldwide.
• Following the previous comment, I have concerns regarding the global distributions of these VOCs, particularly in remote environments. What is the typical detection threshold of single CrIS pixel for these species? Can they truly be measured from individual CrIS spectra under all conditions? For instance, ethene columns below 1x10¹⁵ molec cm^-2, derived from ground-based FTIR measurements at clean-air sites, are considered undetectable (Toon et al., 2018). The detection threshold with spaceborne instruments should be significantly higher. Therefore, it is important to consider whether robust retrievals are possible in remote areas, such as over the oceans, and what the satellite is actually measuring. Is it genuinely background abundance, or is it merely instrumental noise?
• Lines 192-204: It is unclear why 10 different neural networks are trained, only to average their outputs to obtain a final prediction. Can the spread of predictions between these 10 networks truly be used to assess the performance of the training? If you use 50 or 100 different networks, the spread might be reduced, but this would be due to increased statistics, not necessarily because the training is better.
• The same remark applies to Lines 219-232: a filter is built based on the spread between the 10-ANN predictions to discard measurements with too low sensitivity. However, such sensitivity should depend primarily on the uncertainties of the input parameters (HRI, and other predictors), not on the spread between different networks. Indeed, the neural network is built with noise-free, synthetic data, while the retrievals are performed with actual inputs that have their own noise and uncertainties.
• Section 2.3: Input VOC profiles from GEOS-Chem are used to produce the dataset for the training of the neural network, and a P90 predictor, also derived from GEOS-Chem, is used to capture the variations in the spatial and temporal vertical distributions of the VOCs. However, the methodology for employing these profiles and P90 assumptions in the actual column retrievals is unclear. Is a single CrIS pixel retrieved multiple times for different P90 values? If P90 is available from external sources, is the gas column that is ultimately retrieved an interpolation between two columns previously retrieved using the two closest P90 assumptions? Clarification on this process is needed.
• Throughout the study, the P90 is claimed to capture the vertical dependence of the VOCs and enable consistent vertical profile assumptions in comparisons between CrIS and models or other measurements. However, I disagree for several reasons:
  • Using VOC profiles and P90 values from GEOS-Chem constrains the CrIS retrievals within the range of variability of the model. This makes the retrievals dependent on a model, assuming that the model represents the true state of the VOCs. However, global models often fail to accurately represent the actual spatial and vertical variability of these species. This issue is exacerbated with fire plumes, which global models typically struggle to simulate accurately. Consequently, one can question how CrIS retrievals, trained and driven within the range of variability of the model, can produce more realistic representations of the VOCs.
  • While this approach facilitates comparisons between CrIS and GEOS-Chem, it does not enable internally consistent comparisons with other models, and even less so with other measurements. For a given P90 value, any vertical distributions can potentially occur in other models, and even more so with other independent measurements, which may significantly diverge from the vertical assumptions of GEOS-Chem. For example, in the CrIS-FTIR comparison, the range of FTIR retrieved profile shapes of VOCs below the P90 value likely differs significantly from those of GEOS-Chem. The neural network has not been trained to account for such profile variability, and the P90 value alone does not indicate the underlying profile.
  • Additionally, the P90 predictor does not account for the inhomogeneous vertical sensitivity of CrIS, which is likely degraded in the lower layers where the bulk of VOCs is typically found, affecting comparisons with models. Moreover, this sensitivity varies by instrument (the sensitivity of FTIR is supposed to be better).
• A consequence of the CrIS retrieval dependency on GEOS-Chem can be observed in the limited extent of the training set in Figs 4 and 5. The number of synthetic predictions for high gas columns is low and quite limited to weak thermal contrasts, where, by definition, the sensitivity of the sounder is reduced. This might explain why, in Fig. 5, the prediction precision of HCN, ethyne, and ethene is poor for high columns, while this precision is better for lower columns with equivalent thermal contrast. The range of thermal contrast is also relatively narrow. It is not unusual to observe thermal contrasts beyond 15-20K. Therefore, it is questionable whether the neural network can generalize for observational conditions outside the conditions with which it has been trained.
• Section 4: This section lacks a comprehensive explanation of how the FTIR-CrIS comparison is performed. In particular, are the FTIR and CrIS measurements co-located in space and time? It is mentioned that CrIS retrievals are interpolated to the local P90 value from the FTIR retrieved profile. Are the FTIR and CrIS measurements compared in pairs? However, the comparisons are shown for daily and monthly averages. For days where FTIR measurements are available for part of the day (e.g., in the early morning only), is the subsequent daily average really comparable with the CrIS overpass, given the variability of species like methanol and ethene?

Minor comments / typos

• Title: Typically, "long-term" refers to climatological series spanning at least 20-30 years. Perhaps "decadal" would be more appropriate.
• Lines 95-98: There is also a more recent, neural-network-based HCN product from IASI published by Rosanka et al. (2021).
• Lines 155-158: The 2019/2020 Australian bushfires are known for their heavy smoke aerosol content. Are these aerosols accounted for during the fit? Can the residuals of an in-plume spectrum, likely affected by broadband absorptions from smoke aerosols, be compared with those from an out-of-plume spectrum that is free of such aerosols?
• Lines 161-162: “confirming their importance in fire plumes and the underlying spectral signals driving the HRI values.” The spectral fits in Section 2.2 are performed to demonstrate that the spectral signatures of the VOCs are present in CrIS radiances. However, from what I understood, the HRI is not a fit but is calculated over an entire spectral range that is larger than just the spectral features highlighted by the residuals of the fits. Can it be confirmed that the signal contributing to the HRI comes solely from the targeted VOC species and that no other absorbers contribute significantly?
• Lines 190-191: Ozone is a major interference in the absorption band of methanol, and HCN and ethyne (and even ethene) absorb in a spectral range with significant absorption features of CO2. Are these potential interferences accounted for in the retrieval? If not, what could be the impact on the retrieved columns? In addition, the HRI is set up based on CrIS observations from 2019, but then applied to the entire 2012-2023 time series. During this time span, CO2 concentrations have increased from <400 ppm to >420 ppm. Could this bias the retrieved VOC columns before and after 2019?
• Lines 190-191: Are three levels enough to describe the atmospheric temperature profile for the retrieval? There might be significant variability between these levels. Additionally, in Lines 210-211, what do you mean by “dispersed among multiple input variables that individually have only minor impacts”? What are these variables? An accurate representation of the temperature profile is usually important, and misrepresentation can lead to biased retrieved concentrations.
• Lines 206-207: The vertical location of the VOC in the atmosphere influences the thermal contrast. Isn't this the same as saying that the measured signal is largely dependent on the temperature profile and surface temperature?
• Line 209, typo: “yield”
• Lines 171-177: I don’t understand why each simulation is replicated 25 times with random noise, only to take the mean of all 25 replications to obtain the final HRI value. This process cancels out the noise that was initially added.
• Figure 6: What is responsible for the high HCN and ethene columns observed in winter over the North Atlantic Ocean and along the eastern coast of Siberia? It seems unlikely that these are due to outflows from continents. Additionally, concerning ethyne, we do not observe strong latitudinal gradients between the Northern and Southern Hemispheres, which contrasts with what might be expected for an anthropogenic hydrocarbon.
• Figure 7: The seasonal cycle of ethene in India contrasts sharply with that of the other species. Why?
• Figure 11: The daily CrIS data show negative columns of methanol and HCN in the scatter plots, but this is not the case with the CrIS ethyne data (Fig. 11c). Were those filtered out?
• Section 4: CrIS column offsets are derived from the comparison with FTIR columns, based on the intercept of the linear regression between the two datasets. However, there is considerable disparity around this regression, which seems to be partly driven by the FTIR sites included in the comparison. How do these intercepts, and hence the column offsets, vary if more or fewer FTIR sites are included in the comparison?
• Line 311: “VOC spatial distributions”. Do you mean “VOC vertical distributions”?
• Lines 330-333: Few measurements of ethene columns exist, but total columns have been retrieved in various environments up until 2016 with the MkIV FTIR instrument (Toon et al., 2018), including urban sites with high ethene concentrations. Additional ground-based FTIR retrievals of ethene were conducted at the sub-polar Eureka NDACC site in Canada during the overpass of a fire plume (Wizenberg et al., 2023). These measurements could be valuable for comparison with CrIS data.
• Lines 376-380: I believe that the DOFS is not a valid argument to explain the weaker CrIS/NDACC agreement for HCN and ethyne. The DOFS values for methanol, HCN, and ethyne retrieved from ground-based FTIR measurements typically range between 1 and 2, rarely exceeding 2. Therefore, the DOFS values are similar for these three species at most FTIR sites. Instead, I lean towards the view that the weaker CrIS/NDACC agreement for HCN and ethyne may be attributed to the GEOS-Chem-driven training of CrIS retrievals, which may be less robust in capturing the variability seen in retrieved FTIR profiles for these two species.

References:

• Franco, B. et al.: A General Framework for Global Retrievals of Trace Gases from IASI: Application to Methanol, Formic Acid, and PAN, J. Geophys. Res.-Atmos., 123, 13963–13984, https://doi.org/10.1029/2018jd029633, 2018.
• Rosanka, S. et al.: The impact of organic pollutants from Indonesian peatland fires on the tropospheric and lower stratospheric composition, Atmos. Chem. Phys., 21, 11257–11288, https://doi.org/10.5194/acp-21-11257-2021, 2021.
• Toon, G. C. et al.: Measurements of atmospheric ethene by solar absorption FTIR spectrometry, Atmos. Chem. Phys., 18, 5075–5088, https://doi.org/10.5194/acp-18-5075-2018, 2018.
• Wizenberg, T. et al.: Exceptional wildfire enhancements of PAN, C2H4, CH3OH, and HCOOH over the Canadian high Arctic during August 2017. Journal of Geophysical Research: Atmospheres, 128, e2022JD038052. https://doi.org/10.1029/2022JD038052, 2023.
• Whitburn, S. et al.: A flexible and robust neural network IASI NH3 retrieval algorithm, J. Geophys. Res.-Atmos., 121, 6581–6599, https://doi.org/10.1002/2016jd024828, 2016.

Comments:

• Line 185-187. This line needs more evidence, as without an averaging kernel this retrieval seems to be too heavily reliant on strong vertical mixing to be able to detect these species. These species are likely not most sensitive at the surface so any location with a high P90 value is probably missing a significant amount of  information.
• Following the previous comment, it would be helpful to explain how the P90 is applied from the GEOS-Chem grid to match the gridded HRI values. This would likely create problems in regions with strong elevation gradients or strong emissions sources.
• Section 4 could explain more how the NDACC measurements are synced to the CrIS overpass time. Also, are the concentrations being compared via the 1 grid cell that contains each site? If so is there any meteorological screening applied to avoid abnormal conditions? Needs to be explained more thoroughly.