the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 11 Jan 2023
Estimating Methane Emissions in the Arctic nations using surface observations from 2008 to 2019
Abstract. The Arctic is a critical region in terms of global warming. Environmental changes are already progressing steadily in high northern latitudes whereby, among other effects, a high potential of enhanced methane (CH4) emissions is induced. With CH4 being a potent greenhouse gas, additional emissions from Arctic regions may intensify global warming in the future by positive feedback. Various natural and anthropogenic sources are currently contributing to the Arctic’s CH4 budget; however the quantification of those emissions remains challenging. Assessing the amount of CH4 emissions in the Arctic and their contribution to the global budget still remains challenging. This is on the one hand due to the difficulties in carrying out accurate measurements in such remote areas. Besides, high variations in the spatial distribution of methane sources and a poor understanding of the effects of ongoing changes in carbon decomposition, vegetation and hydrology also complicate the assessment. Therefore, the aim of this work is to reduce uncertainties on current bottom-up estimates of CH4 emissions as well as soil oxidation by implementing an inverse modeling approach in order to better quantify CH4 sources and sinks for the most recent years (2008 to 2019). More precisely, the objective is to detect occurring trends in the CH4 emissions and potential changes in seasonal emission patterns. The implementation of the inversion included footprint simulations obtained with the atmospheric transport model FLEXPART, various emission estimates from inventories and land surface models as well as data of atmospheric CH4 concentrations from 42 surface observation sites in the Arctic nations. The results of the inversion showed that the majority of the CH4 sources currently present in high northern latitudes are poorly constrained by the existing observation network. Therefore, conclusions on trends and changes in the seasonal cycle could not be obtained for the corresponding CH4 sectors. Only CH4 fluxes from wetlands are adequately constrained, predominantly in North America. Within the period under study, wetland emissions show a slight negative trend in North America and a slight positive trend in East Eurasia. Overall, the estimated CH4 emissions are lower compared to the bottom-up estimates but higher than similar results from global inversions.
-
Notice on discussion status
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
-
Preprint
(11483 KB)
-
Supplement
(11213 KB)
-
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
- Preprint
(11483 KB) - Metadata XML
-
Supplement
(11213 KB) - BibTeX
- EndNote
- Final revised paper
Journal article(s) based on this preprint
Interactive discussion
Status: closed
-
RC1: 'Comment on egusphere-2022-1257', Anonymous Referee #1, 03 Feb 2023
Estimating Methane Emissions in the Arctic nations using surface observations from 2008 to 2019
Sophie Wittig et al.
GENERAL COMMENTS
The paper describes the estimation of methane emissions in the Arctic by combining model simulations with observations from the ground network. An important part of the study focusses on uncertainties, for which an ensemble approach is used. One of the conclusions is that, at least in some regions, the current observation network seems not dense enough to constrain emission fluxes. The evaluation also shows a clear difference between emission estimates by global inversion systems and systems dedicated to the Arctic regions. These are important messages that are well described and illustrated, and therefore the paper deserves to be published with some minor modifications.
The ensemble method that is used is described with sufficiently detail to understand the key aspects of the procedure, although some clarifications are needed. For example, a key aspect of the method is that from each of 500 possible realizations of the inversions an extra set of 10 possible realizations is drawn. That feels counter-intuitive, one would think that first 5000 possible inputs should be drawn, and from that a similar sized ensemble of posterior results is obtained by performing 5000 inversions. Could the authors describe in more detail the reasoning behind the chosen implementation?
It would also be useful if the authors could comment on how they see the practical application of their method in future studies. For example, a set of "most likely" configurations of the uncertainty has been found; could that be used again in future? If so, then the amount of computational resources that are now spent on running the ensemble, could these then be spent on something else, for example in extension of the particle number and time window of the trajectories?
The authors amplify in their discussion of the results that the availability of observations hampers the impact of the inversion, at least for specific regions and/or years. Could the system also be used to advise on how a future observation network should look like, if the goal would be to reduce uncertainties in particular emissions?
A novel aspect of the paper is for example the computation of sensitivity matrices, which is not often done at this scale, but provides interesting information. Could these be used to advise on extension of the observation network?
SPECIFIC COMMENTS
- p4, line 17: There is no OH sink taken into account explicitly. However, the impact of atmospheric sinks is implicitly present via the background concentrations. What is the important configuration setting that allows to neglect the explicit impact of sinks, is that the length of the time window of 10 days for the trajectories?
- 2.2.1 Pairs of $(R,B)$ are judged on their likelihood. How likely is the configuration with the highest log-likelihood actually? This might be examined by comparing a histogram of the ratios between elements of $S^j$ and std. deviations from the diagonal of $S^j_{RB}$ with the standard normal distribution. Part of this is discussed in Section 4, but it would be useful to have some insight here already.
- 3.3.2 and 3.3.3, Eq. (7) and (9): Does this actually provide a mathematically correct covariance matrix? Usually a covariance is formulated as the matrix product $DCD$, with $D$ a diagonal matrix with the std. deviations, and $C$ a correlation matrix with elements following for example the exponential decays in these equations. The leading factor would then be $\sigma_m\sigma_n$ rather than the average shown here.
- 3.3.2 and 3.3.3: The correlation lengths $d_{corr}$ and $t_{corr}$ are kept fixed. Would it in the current setup of the experiment be possible to also perturb these in the ensemble? If so, what implications would that have for the ensemble size? With longer correlation lengths the number of degrees of freedom will be come lower, and maybe more in line with the number of observations.
- Section 3.3.2 How is the "near-surface residence time" defined? That is, at which heights above the surface are the particles receiving emissions? I guess that a layer structure is defined, with a model for how emissions are distributed over these layers; could this be clarified?
- p 16, line 23: Isn't "EAC4" is simply the ECMWF experiment ID, rather than an abbreviation?
SPELL AND GRAMMAR
- line 23: should be "for" 2005?
- p 13, table 1: not "RANDERSON" but "Randerson"
- p 15, line 10: Think this should be: ".., the input to H consists of the monthly CH4 mixing ratios sectioned into sub-regions and sectors, as well as the monthly averages .."
- p 18, line 19: not "H ere" but "Here"
- p 23, line 15: not "into to" but 'into"
- p 34, line 1: "lower than the prior"
- p 34, line 6: "Minor shifts"
Citation: https://doi.org/10.5194/egusphere-2022-1257-RC1 -
AC1: 'Reply on RC1', Sophie Wittig, 11 Apr 2023
We thank the reviewers for their time reviewing our manuscript. We are thankful for their insights and fruitful comments helping to improve our work. We reply below to individual comments. Reviewers comments are reported in bold, our replies in regular font, while modifications in the manuscript are written in italics.
p4, line 17: There is no OH sink taken into account explicitly. However, the impact of atmospheric sinks is implicitly present via the background concentrations. What is the important configuration setting that allows to neglect the explicit impact of sinks, is that the length of the time window of 10 days for the trajectories?
The justification is mentioned in section 3.3, page 15, line 9 to 13: “Chemical reactions with free radicals in the atmosphere are neglected since the air masses in the studied domain change rather quickly (up to 2 months) compared to the lifetime of CH4 molecules (≈9 years).”
We simulate backward transport over a 10-day period, which means that a peak of 100 ppb caused by emissions, would be oxidized by a rate of less than 0.5 ppb after 10 days, hence negligible for our computations. It should be noted that the OH sink is accounted for in the global simulations used to compute the background. We clarify the latter aspect in the manuscript.
In Sect. 2.1 we added:
We neglect chemical oxidation of CH4 emitted in our regional domain by OH as further explained in Section 3.3, although oxidation by OH is still accounted for in the global model used for the background (Section 3.3.3).
The header of Sect. 3.3 was modified to:
The oxidation of methane by hydroxyl radicals (OH) is neglected inside the domain of simulation since the lifetime of CH4 is ≈9 years (Prather et al., 2012) and the air masses remain in the domain up to 2 months (Berchet et al., 2020). The methane sink from the hydroxyl radical is however accounted for in the global simulations used to compute the background mixing ratios (Section 3.3.3).
2.2.1 Pairs of $(R,B)$ are judged on their likelihood. How likely is the configuration with the highest log-likelihood actually? This might be examined by comparing a histogram of the ratios between elements of $S^j$ and std. deviations from the diagonal of $S^j_{RB}$ with the standard normal distribution. Part of this is discussed in Section 4, but it would be useful to have some insight here already.
We agree that the log-likelihood method tries to attribute diagonal terms of $S^j_{RB}$ as close as possible to the diagonal terms of $S^j$, while keeping the ln | SjRji,Bji| as small as possible, depending on the degrees of freedom available. Still, we find it difficult to comment directly on the diagonal terms of these matrices as their ratio is very sensitive to the definition of the problem, to the observations, to the parameters optimized by the log-likelihood, etc.
Nevertheless, the evaluation of the use of the log-likelihood method in atmospheric inverse modeling approaches has been subject of other studies and is not a specific aim of this work. The selection of this method to evaluate the error estimates of matrices of the ensemble was therefore based on other studies concerning this topic.
We added the following references analysing further the log-likelihood method: (Winiarek et al., 2011; Berchet et al., 2013, 2015a)
3.3.2 and 3.3.3, Eq. (7) and (9): Does this actually provide a mathematically correct covariance matrix? Usually a covariance is formulated as the matrix product $DCD$, with $D$ a diagonal matrix with the std. deviations, and $C$ a correlation matrix with elements following for example the exponential decays in these equations. The leading factor would then be $\sigma_m\sigma_n$ rather than the average shown here.
The equation was indeed incorrectly written and has been corrected. As mentioned below, the accounting of background covariances should anyway be improved in a future work to include the 4D background fields in the control vector instead of the background contribution time series by station.
3.3.2 and 3.3.3: The correlation lengths $d_{corr}$ and $t_{corr}$ are kept fixed. Would it in the current setup of the experiment be possible to also perturb these in the ensemble? If so, what implications would that have for the ensemble size? With longer correlation lengths the number of degrees of freedom will be come lower, and maybe more in line with the number of observations.
Drawing Monte-Carlo members by varying d_corr and t_corr is perfectly possible. We agree that larger d_corr and t_corr reduce the number of degrees of freedom relative to the background. Inquiring the likelihood of configurations relative to d_corr and t_corr would be an interesting new axis of work; however, we limited our present work to fixed d_corr and t_corr to keep the message clearer and also because in our set-up the background takes the shape of time series by station; too strong correlations between stations may be counter-productive in terms of inversions as nearby stations can be influenced by different background air masses.
A next step will be to transform our observation operator to include global 4D fields as the background instead of their extraction/convolution at given sites accounting for backward transport. This would allow to realistically link background from different stations by accounting by the non-isotropic transport.
We modified the end of Sect. 3.3.3 to comment on that aspect:
Drawing Monte-Carlo members by varying dcorr and tcorr is perfectly possible and would impact the number of degrees of freedom relative to the background (the higher dcorr and tcorr, the smaller the number of degrees of freedom). In the present work, we limit our analysis to a single set of dcorr and tcorr. The chosen configuration is a compromise to account for the consistent influence of the background between neighbouring stations and successive time steps while avoiding forcing unrealistic isotropic correlations when close sites are influenced by different background air masses. A future work will include the four-dimensional background fields in the control vector as the background.
Section 3.3.2 How is the "near-surface residence time" defined? That is, at which heights above the surface are the particles receiving emissions? I guess that a layer structure is defined, with a model for how emissions are distributed over these layers; could this be clarified?
We defined a height threshold below which particles are considered as influenced by surface fluxes. We do not distribute surface fluxes into different layers. This is a known caveat for biomass burning in particular that should account for an injection height.
We add in Sect. 3.3.2:
The near-surface hereby corresponds to particles below 500 m instead of, for instance, the planetary boundary height (PBL) since simulated PBL in the Arctic can be unrealistically small, especially during the winter months.
p 16, line 23: Isn't "EAC4" is simply the ECMWF experiment ID, rather than an abbreviation?
“EAC4” is in fact an abbreviation for “ECMWF Atmospheric Composition Reanalysis 4” as also described in the corresponding reference (Innes et al. 2019, section 3.3.3, page 16, line 24).
Citation: https://doi.org/10.5194/egusphere-2022-1257-AC1
-
RC2: 'Comment on egusphere-2022-1257', Anonymous Referee #2, 10 Feb 2023
The manuscript by Wittig et al deals with Arctic methane emissions, their distribution, trends, seasonality and uncertainties, by applying atmospheric inversion modeling methods. The manuscript is thorough and well written. Research on the topic is needed, since the Arctic regional methane budgets are subject of large uncertainties, and the future temperature increases are anticipated to strongly affect the Arctic regions. The manuscript pays specific attention to the uncertainties and sensitivities of the estimated fluxes, which is especially valuable, as such thorough analysis is not too often seen. Overall, the manuscript could be accepted with minor modifications, however, some choices were made that might deserve further analysis or at minimum some more discussion.
p.4 l 17. Neglecting the OH sink seems a little unexpected, because of its importance, trends and seasonality. What where the challenges in applying it ? How would it affect the results ?
p.12 l.17 Do the wetland emissions include freshwater emissions, and are they optimized?
p.14 l.20 What would be the effect of changing these uncertainty assumptions (50%, 60%,100%), as they usually are not very well defined?
p.15 l.10 Observations are averaged over quite a long time, one month. How would the results change if a shorter time period was used? Would it affect the share of the observed signal used for optimization of background vs regional emissions?
p.16 l.6 Did you use boundary layer height for defining near-surface conditions?
p.16 l.25 The implementation used for the inversion? Maybe some more details here?
p.18 l.12 Presume the climatological priors are still realistic in their own framework. Maybe rephrase ‘not good enough as year-to-year changes were present’ or equivalent
p.25 l.15 Satellite measurements (e.g. TROPOMI) do have plenty of data points above 30N, especially during summer, extending to high Arctic. Please rephrase.
p-28 l.3 Are the negative fluxes only removed from the figure, or also from the calculation of the trend?
The text has some typos, please use a spell checker.
Citation: https://doi.org/10.5194/egusphere-2022-1257-RC2 -
AC2: 'Reply on RC2', Sophie Wittig, 11 Apr 2023
We thank the reviewers for their time reviewing our manuscript. We are thankful for their insights and fruitful comments helping to improve our work. We reply below to individual comments. Reviewers comments are reported in bold, our replies in regular font, while modifications in the manuscript are written in italics.
p.4 l 17. Neglecting the OH sink seems a little unexpected, because of its importance, trends and seasonality. What where the challenges in applying it ? How would it affect the results ?
The justification is mentioned in section 3.3, page 15, line 9 to 13: “Chemical reactions with free radicals in the atmosphere are neglected since the air masses in the studied domain change rather quickly (up to 2 months) compared to the lifetime of CH4 molecules (≈9 years)."
We simulate backward transport over a 10-day period, which means that a peak of 100 ppb caused by emissions, would be oxidized by a rate of less than 0.5ppb after 10 days, hence negligible for our computations. It should be noted that the OH sink is accounted for in the global simulations used to compute the background.
We clarify the latter aspect in the manuscrip:
In Sect. 2.1 we added: We neglect chemical oxidation of CH4 emitted in our regional domain by OH as further explained in Section 3.3, although oxidation by OH is still accounted for in the global model used for the background (Section 3.3.3).
The header of Sect. 3.3 was modified to:
The oxidation of methane by hydroxyl radicals (OH) is neglected inside the domain of simulation since the lifetime of CH4 is ≈9 years (Prather et al., 2012) and the air masses remain in the domain up to 2 months (Berchet et al., 2020). The methane sink from the hydroxyl radical is however accounted for in the global simulations used to compute the background mixing ratios (Section 3.3.3).
(see also 1st comment by reviewer #1)
p.12 l.17 Do the wetland emissions include freshwater emissions, and are they optimized?
The wetland emissions dataset by Poulter et al. explicitly aims at excluding CH4 emissions from freshwater systems such as lakes and rivers. In high northern latitude regions the distinction is however challenging, since ponds in permafrost peatlands and thermokarst lakes are usually shallow (less than 1 m depth), which falls of the definition commonly used for wetlands (standing water up to 2 – 2.5 m depth). CH4 emissions from freshwater systems are therefore not optimized taken into account as a separate source.
We propose to comment on that aspect in Sect. 3.2.1:
Terrestrial freshwater systems other than wetlands are hereby not taken into account as a separate source since ponds in permafrost peatlands and thermokarst lakes are usually shallow (less than 1 m depth), which falls under the usual definition for wetlands with standing water up to a depth of 2 to 2.5 m (e.g. Tineret al., 2015).
p.14 l.20 What would be the effect of changing these uncertainty assumptions (50%, 60%,100%), as they usually are not very well defined?
The given uncertainty assumptions only serve as reference values to compute 500 different sets of error estimates, as described in section 3.2.2 on page 14 and 15. The chosen log-normal random distribution of uncertainties leads to a higher number of uncertainty configurations centered around these reference values, yet a large number of different uncertainty estimates are included for each CH4 emission source.
Moreover, we attribute a different weight in our posterior analysis of inversion results depending on the likelihood of the matrices R and B. Hence our posterior estimates account for the most likely uncertainties.
We propose to clarify this aspect in Sect. 3.2.2:
We expect little sensitivity of our results to these prescribed reference values as we weight every sample according to its log-likelihood (see Sect. 2.2.1).
p.15 l.10 Observations are averaged over quite a long time, one month. How would the results change if a shorter time period was used? Would it affect the share of the observed signal used for optimization of background vs regional emissions?
The share of information used to optimize background versus emissions is presented in Figure 8, page 21. Only less than 10% of available pieces of information is used to optimize emissions, while 25-40% is lost in noise, the rest is used for the background.
Using daily observations instead of monthly data could reduce the noise share, and also move part of the background contribution to the emissions. However, there are still obstacles to such ideal outcome:
1) the used background mixing ratio estimates were predominantly too high and therefore a large share of observations was used by the inversion to constrain the background anyway.
2) Observations and simulations at the daily scale are predominantly constituted of alternating peaks due to emission plumes; however, simulated and observed plumes are not necessary synchronized. Using daily observation with temporal mismatches between the numerical realm and the real world would actual worsen usable information. This is particularly critical as we have diagonal observation error covariance matrices R, as is usually done in the inversion community.
3) increasing the duration of backward transport (from 10 days to several weeks) could mechanically reduce the share background information; however, this would critically increase the computation time of footprints and was not affordable for our exploratory study; future work may try inquiring this issue.
We added a comment on that aspect in Sect. 4.2.2:
Another option to increase the ratio of information used to constrain emissions instead of the background would be to use higher temporal resolution for the observations. A compromise would be to use daily afternoon averages instead of monthly averages. However, as illustrated in Belikov et al. (2019), Arctic observations, even at the daily scale, can see strong daily peaks due to emissions in combination with meteorological conditions. Such peaks may be ill-reproduced by the model and could be shifted in time and magnitude, posing a double-penalty effect to the inversion (Vanderbecken et al.). This can have a critical impact on the inversion conclusion, especially with diagonal observation covariance matrices R, as done in the present work, consistent with to the general practice of the inversion community.
We also added a comment in the conclusion:
Increased observation resolution instead of the monthly means used in the present work should be explored as well to potentially improve the constraints on fluxes, although at higher temporal resolution, missed emission peaks in simulations compared to observations could lead to depreciated emission estimates.
p.16 l.6 Did you use boundary layer height for defining near-surface conditions?
Near-surface corresponds to particles below 500 m. We use such a minimal threshold for the surface influence instead of the planetary layer because simulated PBL heights in the Arctic can be unrealistically small, especially in winter, which would cause some periods to virtually see no influence from the surface.
It can be proven that if assuming that the PBL is perfectly mixed, using 500m instead of the PBL height is equivalent when the PBL is higher than 500m. As mentioned above for PBL lower than 500m, this is not mathematically equivalent, but is preferred to avoid numerical artifacts due to erroneously represented winter stable PBL, and also to insufficient number of virtual particles.
We add in Sect. 3.3.2:
The near-surface hereby corresponds to particles below 500 m instead of, for instance, the planetary boundary height (PBL) since simulated PBL in the Arctic can10be unrealistically small, especially during the winter months.
p.16 l.25 The implementation used for the inversion? Maybe some more details here?
We propose to clarify to:
The practical computation of background time series from background fields and backward trajectories is carried out using the Community Inversion Framework, …
p.18 l.12 Presume the climatological priors are still realistic in their own framework. Maybe rephrase ‘not good enough as year-to-year changes were present’ or equivalent
The proposed change has been implemented as following:
This confirms that the climatological priors are not good enough as year-to-year changes were present and the inversion can realistically improve the flux trends.
p.25 l.15 Satellite measurements (e.g. TROPOMI) do have plenty of data points above 30N, especially during summer, extending to high Arctic. Please rephrase.
We propose to clarify to:
(ii) global inversions constrained by satellite measurements (GOSAT, IASI) provide fewer data points above 30°N compared to regions in lower latitudes because of factors such as the solar zenith angle, the surface albedo and the limited light during the polar night; more recent inversions start using TROPOMI data (Tsuruta et al., 2023), with possibly a higher density of data points at high latitudes, but these data were not publicly available at the time of the present work,...
p-28 l.3 Are the negative fluxes only removed from the figure, or also from the calculation of the trend?
The trend and mean annual emissions from biomass burning were calculated for each inversion set-up individually. Configurations with negative mean annual emissions of CH4 were subsequently removed from the ensemble, therefore the trends of configurations with negative mean annual CH4 fluxes are not shown in the figures.
Citation: https://doi.org/10.5194/egusphere-2022-1257-AC2
-
AC2: 'Reply on RC2', Sophie Wittig, 11 Apr 2023
Interactive discussion
Status: closed
-
RC1: 'Comment on egusphere-2022-1257', Anonymous Referee #1, 03 Feb 2023
Estimating Methane Emissions in the Arctic nations using surface observations from 2008 to 2019
Sophie Wittig et al.
GENERAL COMMENTS
The paper describes the estimation of methane emissions in the Arctic by combining model simulations with observations from the ground network. An important part of the study focusses on uncertainties, for which an ensemble approach is used. One of the conclusions is that, at least in some regions, the current observation network seems not dense enough to constrain emission fluxes. The evaluation also shows a clear difference between emission estimates by global inversion systems and systems dedicated to the Arctic regions. These are important messages that are well described and illustrated, and therefore the paper deserves to be published with some minor modifications.
The ensemble method that is used is described with sufficiently detail to understand the key aspects of the procedure, although some clarifications are needed. For example, a key aspect of the method is that from each of 500 possible realizations of the inversions an extra set of 10 possible realizations is drawn. That feels counter-intuitive, one would think that first 5000 possible inputs should be drawn, and from that a similar sized ensemble of posterior results is obtained by performing 5000 inversions. Could the authors describe in more detail the reasoning behind the chosen implementation?
It would also be useful if the authors could comment on how they see the practical application of their method in future studies. For example, a set of "most likely" configurations of the uncertainty has been found; could that be used again in future? If so, then the amount of computational resources that are now spent on running the ensemble, could these then be spent on something else, for example in extension of the particle number and time window of the trajectories?
The authors amplify in their discussion of the results that the availability of observations hampers the impact of the inversion, at least for specific regions and/or years. Could the system also be used to advise on how a future observation network should look like, if the goal would be to reduce uncertainties in particular emissions?
A novel aspect of the paper is for example the computation of sensitivity matrices, which is not often done at this scale, but provides interesting information. Could these be used to advise on extension of the observation network?
SPECIFIC COMMENTS
- p4, line 17: There is no OH sink taken into account explicitly. However, the impact of atmospheric sinks is implicitly present via the background concentrations. What is the important configuration setting that allows to neglect the explicit impact of sinks, is that the length of the time window of 10 days for the trajectories?
- 2.2.1 Pairs of $(R,B)$ are judged on their likelihood. How likely is the configuration with the highest log-likelihood actually? This might be examined by comparing a histogram of the ratios between elements of $S^j$ and std. deviations from the diagonal of $S^j_{RB}$ with the standard normal distribution. Part of this is discussed in Section 4, but it would be useful to have some insight here already.
- 3.3.2 and 3.3.3, Eq. (7) and (9): Does this actually provide a mathematically correct covariance matrix? Usually a covariance is formulated as the matrix product $DCD$, with $D$ a diagonal matrix with the std. deviations, and $C$ a correlation matrix with elements following for example the exponential decays in these equations. The leading factor would then be $\sigma_m\sigma_n$ rather than the average shown here.
- 3.3.2 and 3.3.3: The correlation lengths $d_{corr}$ and $t_{corr}$ are kept fixed. Would it in the current setup of the experiment be possible to also perturb these in the ensemble? If so, what implications would that have for the ensemble size? With longer correlation lengths the number of degrees of freedom will be come lower, and maybe more in line with the number of observations.
- Section 3.3.2 How is the "near-surface residence time" defined? That is, at which heights above the surface are the particles receiving emissions? I guess that a layer structure is defined, with a model for how emissions are distributed over these layers; could this be clarified?
- p 16, line 23: Isn't "EAC4" is simply the ECMWF experiment ID, rather than an abbreviation?
SPELL AND GRAMMAR
- line 23: should be "for" 2005?
- p 13, table 1: not "RANDERSON" but "Randerson"
- p 15, line 10: Think this should be: ".., the input to H consists of the monthly CH4 mixing ratios sectioned into sub-regions and sectors, as well as the monthly averages .."
- p 18, line 19: not "H ere" but "Here"
- p 23, line 15: not "into to" but 'into"
- p 34, line 1: "lower than the prior"
- p 34, line 6: "Minor shifts"
Citation: https://doi.org/10.5194/egusphere-2022-1257-RC1 -
AC1: 'Reply on RC1', Sophie Wittig, 11 Apr 2023
We thank the reviewers for their time reviewing our manuscript. We are thankful for their insights and fruitful comments helping to improve our work. We reply below to individual comments. Reviewers comments are reported in bold, our replies in regular font, while modifications in the manuscript are written in italics.
p4, line 17: There is no OH sink taken into account explicitly. However, the impact of atmospheric sinks is implicitly present via the background concentrations. What is the important configuration setting that allows to neglect the explicit impact of sinks, is that the length of the time window of 10 days for the trajectories?
The justification is mentioned in section 3.3, page 15, line 9 to 13: “Chemical reactions with free radicals in the atmosphere are neglected since the air masses in the studied domain change rather quickly (up to 2 months) compared to the lifetime of CH4 molecules (≈9 years).”
We simulate backward transport over a 10-day period, which means that a peak of 100 ppb caused by emissions, would be oxidized by a rate of less than 0.5 ppb after 10 days, hence negligible for our computations. It should be noted that the OH sink is accounted for in the global simulations used to compute the background. We clarify the latter aspect in the manuscript.
In Sect. 2.1 we added:
We neglect chemical oxidation of CH4 emitted in our regional domain by OH as further explained in Section 3.3, although oxidation by OH is still accounted for in the global model used for the background (Section 3.3.3).
The header of Sect. 3.3 was modified to:
The oxidation of methane by hydroxyl radicals (OH) is neglected inside the domain of simulation since the lifetime of CH4 is ≈9 years (Prather et al., 2012) and the air masses remain in the domain up to 2 months (Berchet et al., 2020). The methane sink from the hydroxyl radical is however accounted for in the global simulations used to compute the background mixing ratios (Section 3.3.3).
2.2.1 Pairs of $(R,B)$ are judged on their likelihood. How likely is the configuration with the highest log-likelihood actually? This might be examined by comparing a histogram of the ratios between elements of $S^j$ and std. deviations from the diagonal of $S^j_{RB}$ with the standard normal distribution. Part of this is discussed in Section 4, but it would be useful to have some insight here already.
We agree that the log-likelihood method tries to attribute diagonal terms of $S^j_{RB}$ as close as possible to the diagonal terms of $S^j$, while keeping the ln | SjRji,Bji| as small as possible, depending on the degrees of freedom available. Still, we find it difficult to comment directly on the diagonal terms of these matrices as their ratio is very sensitive to the definition of the problem, to the observations, to the parameters optimized by the log-likelihood, etc.
Nevertheless, the evaluation of the use of the log-likelihood method in atmospheric inverse modeling approaches has been subject of other studies and is not a specific aim of this work. The selection of this method to evaluate the error estimates of matrices of the ensemble was therefore based on other studies concerning this topic.
We added the following references analysing further the log-likelihood method: (Winiarek et al., 2011; Berchet et al., 2013, 2015a)
3.3.2 and 3.3.3, Eq. (7) and (9): Does this actually provide a mathematically correct covariance matrix? Usually a covariance is formulated as the matrix product $DCD$, with $D$ a diagonal matrix with the std. deviations, and $C$ a correlation matrix with elements following for example the exponential decays in these equations. The leading factor would then be $\sigma_m\sigma_n$ rather than the average shown here.
The equation was indeed incorrectly written and has been corrected. As mentioned below, the accounting of background covariances should anyway be improved in a future work to include the 4D background fields in the control vector instead of the background contribution time series by station.
3.3.2 and 3.3.3: The correlation lengths $d_{corr}$ and $t_{corr}$ are kept fixed. Would it in the current setup of the experiment be possible to also perturb these in the ensemble? If so, what implications would that have for the ensemble size? With longer correlation lengths the number of degrees of freedom will be come lower, and maybe more in line with the number of observations.
Drawing Monte-Carlo members by varying d_corr and t_corr is perfectly possible. We agree that larger d_corr and t_corr reduce the number of degrees of freedom relative to the background. Inquiring the likelihood of configurations relative to d_corr and t_corr would be an interesting new axis of work; however, we limited our present work to fixed d_corr and t_corr to keep the message clearer and also because in our set-up the background takes the shape of time series by station; too strong correlations between stations may be counter-productive in terms of inversions as nearby stations can be influenced by different background air masses.
A next step will be to transform our observation operator to include global 4D fields as the background instead of their extraction/convolution at given sites accounting for backward transport. This would allow to realistically link background from different stations by accounting by the non-isotropic transport.
We modified the end of Sect. 3.3.3 to comment on that aspect:
Drawing Monte-Carlo members by varying dcorr and tcorr is perfectly possible and would impact the number of degrees of freedom relative to the background (the higher dcorr and tcorr, the smaller the number of degrees of freedom). In the present work, we limit our analysis to a single set of dcorr and tcorr. The chosen configuration is a compromise to account for the consistent influence of the background between neighbouring stations and successive time steps while avoiding forcing unrealistic isotropic correlations when close sites are influenced by different background air masses. A future work will include the four-dimensional background fields in the control vector as the background.
Section 3.3.2 How is the "near-surface residence time" defined? That is, at which heights above the surface are the particles receiving emissions? I guess that a layer structure is defined, with a model for how emissions are distributed over these layers; could this be clarified?
We defined a height threshold below which particles are considered as influenced by surface fluxes. We do not distribute surface fluxes into different layers. This is a known caveat for biomass burning in particular that should account for an injection height.
We add in Sect. 3.3.2:
The near-surface hereby corresponds to particles below 500 m instead of, for instance, the planetary boundary height (PBL) since simulated PBL in the Arctic can be unrealistically small, especially during the winter months.
p 16, line 23: Isn't "EAC4" is simply the ECMWF experiment ID, rather than an abbreviation?
“EAC4” is in fact an abbreviation for “ECMWF Atmospheric Composition Reanalysis 4” as also described in the corresponding reference (Innes et al. 2019, section 3.3.3, page 16, line 24).
Citation: https://doi.org/10.5194/egusphere-2022-1257-AC1
-
RC2: 'Comment on egusphere-2022-1257', Anonymous Referee #2, 10 Feb 2023
The manuscript by Wittig et al deals with Arctic methane emissions, their distribution, trends, seasonality and uncertainties, by applying atmospheric inversion modeling methods. The manuscript is thorough and well written. Research on the topic is needed, since the Arctic regional methane budgets are subject of large uncertainties, and the future temperature increases are anticipated to strongly affect the Arctic regions. The manuscript pays specific attention to the uncertainties and sensitivities of the estimated fluxes, which is especially valuable, as such thorough analysis is not too often seen. Overall, the manuscript could be accepted with minor modifications, however, some choices were made that might deserve further analysis or at minimum some more discussion.
p.4 l 17. Neglecting the OH sink seems a little unexpected, because of its importance, trends and seasonality. What where the challenges in applying it ? How would it affect the results ?
p.12 l.17 Do the wetland emissions include freshwater emissions, and are they optimized?
p.14 l.20 What would be the effect of changing these uncertainty assumptions (50%, 60%,100%), as they usually are not very well defined?
p.15 l.10 Observations are averaged over quite a long time, one month. How would the results change if a shorter time period was used? Would it affect the share of the observed signal used for optimization of background vs regional emissions?
p.16 l.6 Did you use boundary layer height for defining near-surface conditions?
p.16 l.25 The implementation used for the inversion? Maybe some more details here?
p.18 l.12 Presume the climatological priors are still realistic in their own framework. Maybe rephrase ‘not good enough as year-to-year changes were present’ or equivalent
p.25 l.15 Satellite measurements (e.g. TROPOMI) do have plenty of data points above 30N, especially during summer, extending to high Arctic. Please rephrase.
p-28 l.3 Are the negative fluxes only removed from the figure, or also from the calculation of the trend?
The text has some typos, please use a spell checker.
Citation: https://doi.org/10.5194/egusphere-2022-1257-RC2 -
AC2: 'Reply on RC2', Sophie Wittig, 11 Apr 2023
We thank the reviewers for their time reviewing our manuscript. We are thankful for their insights and fruitful comments helping to improve our work. We reply below to individual comments. Reviewers comments are reported in bold, our replies in regular font, while modifications in the manuscript are written in italics.
p.4 l 17. Neglecting the OH sink seems a little unexpected, because of its importance, trends and seasonality. What where the challenges in applying it ? How would it affect the results ?
The justification is mentioned in section 3.3, page 15, line 9 to 13: “Chemical reactions with free radicals in the atmosphere are neglected since the air masses in the studied domain change rather quickly (up to 2 months) compared to the lifetime of CH4 molecules (≈9 years)."
We simulate backward transport over a 10-day period, which means that a peak of 100 ppb caused by emissions, would be oxidized by a rate of less than 0.5ppb after 10 days, hence negligible for our computations. It should be noted that the OH sink is accounted for in the global simulations used to compute the background.
We clarify the latter aspect in the manuscrip:
In Sect. 2.1 we added: We neglect chemical oxidation of CH4 emitted in our regional domain by OH as further explained in Section 3.3, although oxidation by OH is still accounted for in the global model used for the background (Section 3.3.3).
The header of Sect. 3.3 was modified to:
The oxidation of methane by hydroxyl radicals (OH) is neglected inside the domain of simulation since the lifetime of CH4 is ≈9 years (Prather et al., 2012) and the air masses remain in the domain up to 2 months (Berchet et al., 2020). The methane sink from the hydroxyl radical is however accounted for in the global simulations used to compute the background mixing ratios (Section 3.3.3).
(see also 1st comment by reviewer #1)
p.12 l.17 Do the wetland emissions include freshwater emissions, and are they optimized?
The wetland emissions dataset by Poulter et al. explicitly aims at excluding CH4 emissions from freshwater systems such as lakes and rivers. In high northern latitude regions the distinction is however challenging, since ponds in permafrost peatlands and thermokarst lakes are usually shallow (less than 1 m depth), which falls of the definition commonly used for wetlands (standing water up to 2 – 2.5 m depth). CH4 emissions from freshwater systems are therefore not optimized taken into account as a separate source.
We propose to comment on that aspect in Sect. 3.2.1:
Terrestrial freshwater systems other than wetlands are hereby not taken into account as a separate source since ponds in permafrost peatlands and thermokarst lakes are usually shallow (less than 1 m depth), which falls under the usual definition for wetlands with standing water up to a depth of 2 to 2.5 m (e.g. Tineret al., 2015).
p.14 l.20 What would be the effect of changing these uncertainty assumptions (50%, 60%,100%), as they usually are not very well defined?
The given uncertainty assumptions only serve as reference values to compute 500 different sets of error estimates, as described in section 3.2.2 on page 14 and 15. The chosen log-normal random distribution of uncertainties leads to a higher number of uncertainty configurations centered around these reference values, yet a large number of different uncertainty estimates are included for each CH4 emission source.
Moreover, we attribute a different weight in our posterior analysis of inversion results depending on the likelihood of the matrices R and B. Hence our posterior estimates account for the most likely uncertainties.
We propose to clarify this aspect in Sect. 3.2.2:
We expect little sensitivity of our results to these prescribed reference values as we weight every sample according to its log-likelihood (see Sect. 2.2.1).
p.15 l.10 Observations are averaged over quite a long time, one month. How would the results change if a shorter time period was used? Would it affect the share of the observed signal used for optimization of background vs regional emissions?
The share of information used to optimize background versus emissions is presented in Figure 8, page 21. Only less than 10% of available pieces of information is used to optimize emissions, while 25-40% is lost in noise, the rest is used for the background.
Using daily observations instead of monthly data could reduce the noise share, and also move part of the background contribution to the emissions. However, there are still obstacles to such ideal outcome:
1) the used background mixing ratio estimates were predominantly too high and therefore a large share of observations was used by the inversion to constrain the background anyway.
2) Observations and simulations at the daily scale are predominantly constituted of alternating peaks due to emission plumes; however, simulated and observed plumes are not necessary synchronized. Using daily observation with temporal mismatches between the numerical realm and the real world would actual worsen usable information. This is particularly critical as we have diagonal observation error covariance matrices R, as is usually done in the inversion community.
3) increasing the duration of backward transport (from 10 days to several weeks) could mechanically reduce the share background information; however, this would critically increase the computation time of footprints and was not affordable for our exploratory study; future work may try inquiring this issue.
We added a comment on that aspect in Sect. 4.2.2:
Another option to increase the ratio of information used to constrain emissions instead of the background would be to use higher temporal resolution for the observations. A compromise would be to use daily afternoon averages instead of monthly averages. However, as illustrated in Belikov et al. (2019), Arctic observations, even at the daily scale, can see strong daily peaks due to emissions in combination with meteorological conditions. Such peaks may be ill-reproduced by the model and could be shifted in time and magnitude, posing a double-penalty effect to the inversion (Vanderbecken et al.). This can have a critical impact on the inversion conclusion, especially with diagonal observation covariance matrices R, as done in the present work, consistent with to the general practice of the inversion community.
We also added a comment in the conclusion:
Increased observation resolution instead of the monthly means used in the present work should be explored as well to potentially improve the constraints on fluxes, although at higher temporal resolution, missed emission peaks in simulations compared to observations could lead to depreciated emission estimates.
p.16 l.6 Did you use boundary layer height for defining near-surface conditions?
Near-surface corresponds to particles below 500 m. We use such a minimal threshold for the surface influence instead of the planetary layer because simulated PBL heights in the Arctic can be unrealistically small, especially in winter, which would cause some periods to virtually see no influence from the surface.
It can be proven that if assuming that the PBL is perfectly mixed, using 500m instead of the PBL height is equivalent when the PBL is higher than 500m. As mentioned above for PBL lower than 500m, this is not mathematically equivalent, but is preferred to avoid numerical artifacts due to erroneously represented winter stable PBL, and also to insufficient number of virtual particles.
We add in Sect. 3.3.2:
The near-surface hereby corresponds to particles below 500 m instead of, for instance, the planetary boundary height (PBL) since simulated PBL in the Arctic can10be unrealistically small, especially during the winter months.
p.16 l.25 The implementation used for the inversion? Maybe some more details here?
We propose to clarify to:
The practical computation of background time series from background fields and backward trajectories is carried out using the Community Inversion Framework, …
p.18 l.12 Presume the climatological priors are still realistic in their own framework. Maybe rephrase ‘not good enough as year-to-year changes were present’ or equivalent
The proposed change has been implemented as following:
This confirms that the climatological priors are not good enough as year-to-year changes were present and the inversion can realistically improve the flux trends.
p.25 l.15 Satellite measurements (e.g. TROPOMI) do have plenty of data points above 30N, especially during summer, extending to high Arctic. Please rephrase.
We propose to clarify to:
(ii) global inversions constrained by satellite measurements (GOSAT, IASI) provide fewer data points above 30°N compared to regions in lower latitudes because of factors such as the solar zenith angle, the surface albedo and the limited light during the polar night; more recent inversions start using TROPOMI data (Tsuruta et al., 2023), with possibly a higher density of data points at high latitudes, but these data were not publicly available at the time of the present work,...
p-28 l.3 Are the negative fluxes only removed from the figure, or also from the calculation of the trend?
The trend and mean annual emissions from biomass burning were calculated for each inversion set-up individually. Configurations with negative mean annual emissions of CH4 were subsequently removed from the ensemble, therefore the trends of configurations with negative mean annual CH4 fluxes are not shown in the figures.
Citation: https://doi.org/10.5194/egusphere-2022-1257-AC2
-
AC2: 'Reply on RC2', Sophie Wittig, 11 Apr 2023
Peer review completion
Journal article(s) based on this preprint
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 414 | 231 | 23 | 668 | 51 | 10 | 15 |
- HTML: 414
- PDF: 231
- XML: 23
- Total: 668
- Supplement: 51
- BibTeX: 10
- EndNote: 15
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Sophie Wittig
Isabelle Pison
Marielle Saunois
Joël Thanwerdas
Adrien Martinez
Jean-Daniel Paris
Tochinobu Machida
Motoki Sasakawa
Douglas E. J. Worthy
Rona L. Thompson
Espen Sollum
Michael Arshinov
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
- Preprint
(11483 KB) - Metadata XML
-
Supplement
(11213 KB) - BibTeX
- EndNote
- Final revised paper