Impact of transport model resolution and a-priori assumptions on inverse modeling of Swiss F-gases emissions
Abstract. Inverse modeling is a widely used top-down method to infer greenhouse gas (GHG) emissions and their spatial distribution based on atmospheric observations. The errors associated with inverse modeling have multiple sources, such as observations and a-priori emission estimates, but they are often dominated by the transport model error. Here, we utilize the Lagrangian Particle Dispersion Model (LPDM) FLEXPART, driven by the meteorological fields of the regional numerical weather prediction model COSMO. The main source of errors in LPDMs is the turbulence diffusion parameterization and the meteorological fields. The latter are outputs of an Eulerian model. Recently, we introduced an improved parameterization scheme of the turbulence diffusion in FLEXPART, which significantly improves FLEXPART-COSMO simulations at 1 km resolution. We exploit F-gases measurements from two extended field campaigns on the Swiss Plateau (in Beromünster and Sottens) and we conduct both high- (1 km) and low-resolution (7 km) FLEXPART transport simulations that are then used in a Bayesian analytical inversion to estimate spatial emission distributions. Our results for four F-gases (HFC-134a, HFC-125, HFC-32, SF6) indicate that both high-resolution inversions and a dense measurement network significantly improve the ability to estimate the spatial distribution of emissions. Furthermore, the total emission estimates from the high-resolution inversions (351±44 Mg yr−1 for HFC-134a, 101±21 Mg yr−1 for HFC-125, 50±8 Mg yr−1 for HFC-32, 9.0±1.1 Mg yr−1 for SF6) are significantly higher compared to the low-resolution inversions (20–40 % increase) and result in total a-posteriori emission estimates that are closer to national inventory values as reported to the UNFCCC (10–20 % difference between high-resolution inversion estimates and inventory values compared to 30–40 % difference between the low-resolution inversion estimates and inventory values). Specifically, we attribute these improvements to a better representation of the atmospheric flow in complex terrain in the high-resolution model, partly induced by the more realistic topography. We further conduct numerous sensitivity inversions, varying different parameters and variables of our Bayesian inversion framework to explore the whole range of uncertainty in the inversion errors (e.g., inversion grid, spatial distribution of a-priori emissions, covariance parameters like baseline uncertainty and spatial correlation length, temporal resolution of the assimilated observations, observation network, seasonality of emissions). From the above-mentioned parameters, we find that the uncertainty of the mole fraction baseline and the spatial distribution of the a-priori emissions have the largest impact on the a-posteriori total emission estimates and their spatial distribution. This study is a step towards mitigating the errors associated with the transport models and better characterizing the uncertainty inherent in the inversion error. Improvements in the latter will facilitate the validation and standardization of the national GHG emission inventories and support policymakers.
Ioannis Katharopoulos et al.
Status: final response (author comments only)
- RC1: 'Comment on acp-2022-723', Anonymous Referee #1, 06 Feb 2023
- RC2: 'Comment on acp-2022-723', Anonymous Referee #2, 21 Mar 2023
Ioannis Katharopoulos et al.
Atmospheric Halocarbon Observations at Beromünster, Switzerland, and Bayesian Inverse Modeling to assess Emissions https://zenodo.org/record/5843548#.Y1FKv3ZBw2w
Ioannis Katharopoulos et al.
Viewed (geographical distribution)
The paper presents inverse modelling results using atmospheric transport models at varying spatial resolution. It is certainly of interest to the scientific community. In general the paper is well written, and I recommend publication after the following minor concerns have been addressed.
Regarding the sensitivity of inversion results to the assumed a priori emission distribution, it should be discussed a bit more why the inversion is not able to adjust and correct the spatial pattern. Which part is related to station density (coverage of the combined sensitivity) and which part is related to the lack of flexibility via the a priori uncertainty? It is a good suggestion to use different prior estimates with different spatial patterns, but then it needs to be ensured that those different estimates actually cover the range of possible distributions.
Lines 51-54: It was actually Lin et al (2003) who showed this for the first time.
Lines 58-66: Errors in inversions and in transport have been discussed in the literature prior to 2018, please cite earlier studies.
Line 89: reference Bergamaschi et al., change year from 2017 to 2022
Line 331: change “a-posterior” to “a posteriori”.
Line 345: ad a comma between the indices in the subscript (as in Eq. 12)
Line 349: a temporal correlation length of 0.01 days would only have an impact if the observations would be at a higher rate that about 1/hour. Is this the case? May be this should be mentioned clearly. It is not so clear that representation and model errors are uncorrelated between e.g. subsequent days or even hours, so this might need additional explanation or discussion.
Table 2: SEAS1 and SEAS2 have identical entries in the table, may be mention in a footnote to table 2 what the difference is.
Line 458: please briefly explain the iterative approach here, what is iterated? Are simply posterior residuals used to inform on model-data mismatch error?
Lines 484-486: What are typical scales for near- vs. far-field? This needs to be elaborated a bit more. To me it is unclear why in the far-field the diffusion should be depending on the size of the eddies, certainly at some distance the main cause of “diffusion” is the loss of correlation (or coherence) in the mean wind fields.
Line 494: remove comma after “overfitting”
Line 519: add “.” at the end of the line sentence
Line 524: When I calculate the relative uncertainties, I get 12.5% and 18.8% for Base1 and Base7 respectively. Please correct
Figure 6 caption: e) and f) are missing
Line 542: Is it not expected that the reduced chi-square values are always close to 1 given that uncertainty covariance parameters are estimated using maximum likelihood?
Line 553: “rational state” not clear why that would not be rational, given the (wrong) prior, that is the solution one retrieves using a completely rational approach.
Line 569: why is the uncertainty for the SEAS2 inversion results for JJA not given? It would also be interesting to discuss if the seasonality in retrieved fluxes is significant.
Line 577: It would be interesting to see if taking into account seasonality improves the performance (e.g. using statistics as shown in Table 3). Not allowing for seasonal variations in combination with seasonal changes in transport patterns can be expected to result in larger model-data mismatch error.
Line 597 / Table 3: it would be helpful to also have the posterior estimates and uncertainties included.
Lines 681-685: It needs to be mentioned that the case of Switzerland is special given the orography. The manuscript has not shown similar impacts of resolution increase in domains with more benign orography.
Lines 694-695 “While … uncertainty,” the sensitivity is the other way around (emission estimates being sensitive to parameters), please correct.
Lines 736-737: May be reformulate “… which is not true for the low-resolution inversions, in which when the uniform distribution is employed the low-resolution inversion fails and produces unrealistic results.” e.g. to “… which is not true for the low-resolution inversions, which fails and produces unrealistic results when using the uniform distribution.”