Improved constraints on ammonia emissions and deposition from co-assimilating NH<sub>3</sub> and NO<sub>2</sub> satellite observations over the Netherlands

Wizenberg, Tyler; Dammers, Enrico; Segers, Arjo; Shephard, Mark W.; Coheur, Pierre; Clarisse, Lieven; Van Damme, Martin; Eskes, Henk; Wichink Kruit, Roy; van der Graaf, Shelley; Schaap, Martijn

doi:10.5194/egusphere-2025-5926

Preprints

https://doi.org/10.5194/egusphere-2025-5926

Preprints

11 Jan 2026

| 11 Jan 2026

Status: this preprint is open for discussion and under review for Atmospheric Chemistry and Physics (ACP).

Improved constraints on ammonia emissions and deposition from co-assimilating NH₃ and NO₂ satellite observations over the Netherlands

Tyler Wizenberg, Enrico Dammers, Arjo Segers, Mark W. Shephard, Pierre Coheur, Lieven Clarisse, Martin Van Damme, Henk Eskes, Roy Wichink Kruit, Shelley van der Graaf, and Martijn Schaap

Abstract. Ammonia (NH₃) and nitrogen dioxide (NO₂) are key components of reactive nitrogen, strongly affecting air quality and ecosystem health. However, long-term constraints on ammonia emissions and deposition remain uncertain due to sparse in situ measurements and limitations of individual satellite products. We jointly assimilate five years (2018–2022) of NH₃ and NO₂ satellite observations over the Netherlands to improve constraints on reactive nitrogen concentrations, emissions, and deposition. NH₃ retrievals from the Infrared Atmospheric Sounding Interferometer (IASI) and the Cross-Track Infrared Sounder (CrIS) are combined with NO₂ observations from the TROPOspheric Monitoring Instrument (TROPOMI) within the LOTOS-EUROS chemical transport model using a Local Ensemble Transform Kalman Filter. The co-assimilation produces coherent year-to-year adjustments in modeled NH₃ concentration, emission, and deposition fields. Validation against measurements from the Dutch National Air Quality Monitoring Network (LML) shows reduced biases, clearer diurnal cycles, and improved correlations. Sensitivity experiments demonstrate that including TROPOMI NO₂ alongside IASI and CrIS NH₃ yields the lowest NH₃ surface bias versus LML, highlighting the added value of coupling chemically related satellite observations. Comparisons with monthly Measurements of Ammonia in Nature (MAN) observations showed improved correlations but persistent spatial biases due to representativeness differences, while MAN sensors co-located with LML stations exhibited consistent improvements. These results demonstrate that co-assimilating complementary satellite observations can substantially improve constraints on ammonia emissions and deposition, with direct relevance for air-quality assessment and nitrogen policy applications.

Received: 28 Nov 2025 – Discussion started: 11 Jan 2026

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Tyler Wizenberg, Enrico Dammers, Arjo Segers, Mark W. Shephard, Pierre Coheur, Lieven Clarisse, Martin Van Damme, Henk Eskes, Roy Wichink Kruit, Shelley van der Graaf, and Martijn Schaap

Status: open (until 22 Feb 2026)

Post a comment Subscribe to comment alert

RC1:
'Comment on egusphere-2025-5926', Anonymous Referee #1, 29 Jan 2026 reply
The manuscript “Improved constraints on ammonia emissions and deposition from co-assimilating NH3 and NO2 satellite observations over the Netherlands” investigates the impact of assimilating ammonia and nitrogen dioxide retrievals from satellite-borne instruments on ammonia emissions in the Netherlands using the local ensemble transform Kalman filter (LETKF) with the LOTOS-EUROS chemical transport model. The paper discusses the impact of assimilating different combinations of retrievals of NH3 from IASI, NH3 from CrIS, and NO2 from TROPOMI. Following describing the impact on the ammonia emissions, the paper then examines the impact of using the posterior NH3 emission on the atmospheric concentration of ammonia and on the deposition of NHx as well as comparisons to surface observations.

The authors show that the posterior ammonia emissions generally improve the agreement between the LOTOS-EUROS model and observations from the LML network, but degrades the comparison to observations from the MAN network. The authors attribute the degradation of the comparisons to the MAN observations to differences in representativeness of the MAN observations, the satellite retrievals, and the horizontal grid of the LOTOS-EUROS model.

As ammonia plays a significant role in particulate matter formation and nitrogen deposition, thereby important for human and ecosystem health, but bottom-up emission inventories of ammonia emissions are often poorly constrained, this paper provides relevant information. While a number of previous studies have examined using ammonia observations from satellites to improve ammonia emissions, few (if any) have expanded this to include observations of other chemical species (in this case NO2), and so provides a novel aspect to the work as well.

Overall, the paper is relevant and generally well written. The examination of the additional information provided by the NO2 retrievals from TROPOMI are particularly interesting. There are a few places where the clarity of the manuscript needs improvement. Also, many of the comparison statistics in the paper are presented without uncertainties and thus the reader cannot determine if differences in these statistics are statistically significant or not.
Major Comments:
I think Section 2.1.1 needs to be revised to make the description of the assimilation system clearer.

For instance, on line 98, is only the beta parameters set via the LETKF or 3D gas concentrations are set with the LETKF as well? Please rewrite this sentence to more clearly differentiate between the inversion variables set by the LETKF and all other model variables.

Related to this, when the NO2 retrievals are assimilated and compared against the model output, are all differences attributed to mismatches in NH3 emissions or to other factors as well? If it is the case that the LETKF only adjusts the NH3 emissions, then is it a reasonable assumption to attribute all observed NO2 differences to the NH3 emissions instead of other factors?

What is the frequency of the analyses produced (i.e. t_k - t_{k-1})? How exactly are different times handled with respect to the temporal correlation in Eq. (1) and what is the assimilation window used? Is the assimilation window for time t_k set as from t_k – delta_t/2 to t_k + delta_t/2, where delta_t is the analyses frequency, and then there is some sort of formula that combines {x_k} for different times using Eq. (1) or does the assimilation window encompass multiple times {t_k} and Eq. (1) is used within a single analysis (or some other method)? Please clarify in the text.

Lines 102-104 describe the ensemble initialization, but at this point I don't think a definition of beta has been given, so I'm not certain what the stated standard deviation (of one) is the standard deviation of. At the start of the paragraph, it is stated that beta is a “perturbation factor”. Please give a precise definition of this. e.g. is this a multiplicative factor or some other type of perturbation. On line 104, it is stated that “The values of the mean and standard deviation for the ensemble can be adjusted if desired.” I don’t know what this means, please clarify.

I’m less familiar with the specifics of the LETKF as compared to other ensemble methods, in which localization is done quite differently, so I was confused by some of the details in Section 2.1.2. I’m aware that for the ‘R localization’ that is typically used in the LETKF, the optimal localization length scale is often quite a bit smaller than that used in the ‘B localization’ used in many other ensemble systems (where the Schur product of the ensemble-based B matrix is taken with the localization matrix). However, I wasn’t aware that ‘R localization’ had anything to do with any of the properties of the observations like the instrument’s horizontal resolution (for ‘B localization’ these are unrelated). Could you add some references and maybe some more text explaining this, especially for readers less familiar with ‘R localization’.

At the beginning of this section, Shin et al. (2016) is referenced, in which they state that they choose a horizontal localization scale of 2*(10/3)^0.5 * 500 km ~ 1825 km. Although, the Shin et al. (2016) paper is for NWP instead of a relatively short-lived atmospheric gas, I’m surprised that the localization length scales between the two different applications would be so different (compared to the 15 km and 5 km used in for this work).

The ensemble size used in this work (12) is quite small compared to most (non-LETKF) ensemble systems. On line 100, the paper cites Van Der Graaf et al. (2022) as determining this number to be sufficiently large. Van Der Graaf et al. (2022) states “A limited ensemble size of N=12 was found to be sufficient to describe the imposed model uncertainty, which is not too complicated due to short lifetime of NH₃ and therefore strong relation between concentrations and nearby emissions.” There doesn’t seem to be a more quantitative analysis of the dependency of the LETKF on ensemble size. If ‘B localization’ was used (in a non-LETKF system), the localization length scale would be much larger than 15 km or 5 km, and so N=12 would be a very small ensemble. N=12 would be sufficient to specify the error covariances within a 15 km x 15 km (or 5 km x 5 km) region, but I don’t have an intuitive understanding on why the ensemble size would need to be of very different sizes depending on whether B or R localization was used. Could you also add either some references or reasoning on why this is the case.

On line 278, it is stated that “cases where the measured and modeled precipitation exceeded a mean absolute deviation of 1-sigma were excluded” from the evaluation statistics. I would have imagined that a 1-sigma cutoff would exclude quite a lot of observations and would have thought setting this at something like >= 2-sigma would be more appropriate. Could you either add a justification of the 1-sigma cutoff or change this to a higher cutoff.

I don’t think Section 3.4 is necessary in the main text as I’m not sure it adds much to the conclusions of the paper. I think it would be best to move this section to the Supplement or an appendix and the authors can add a sentence or two in the main text to referencing the supplement/appendix. Also, I think Eq. (11) might be missing a trace on the right-hand side of the equation.

There were insufficient details given about the uncertainties of the statistical values presented in Section 3.5. In Fig.7, the slope of the line of best fit has uncertainties associated with it, but the statistics in the lower-right corner (R, mu, sigma) do not. To properly compare these statistics between the base and optimized cases, uncertainties need to be added so the reader can determine whether the differences between these cases is statistically significant or not.

For instance, on lines 456-457, it states “The correlation in the temporal means also improved slightly from R = 0.84 to R = 0.85, and the slope of the regression also improved from 0.79 to 0.91”, but without uncertainties on these numbers the reader cannot tell if these changes are statistically significant or not (as with other subsequent comparisons done later in the paper).

Uncertainties for R, mu, sigma should be added to all scatter plots (Figs. 7, 8, 11-14, 16), Table 2 (also add uncertainties for the slope here), as well to the main text. So line 457 should read '...the slope of the regression also improved from 0.79 +/- 0.03 to 0.91 +/- 0.03' and similarly for whenever R, mu, or sigma values are stated in the main text. Also, in Fig. B2, the difference plot should show the uncertainties (i.e. the standard error) of the mean differences.

For a lot of the comparisons between the model and surface observations in Section 3.5, scatter plots and their associated statistics were compiled for particular spatial and temporal average separately, for instance in Fig. 7 & 8, Figs. 11 & 12, and Figs. 13 & 14. I think there are more informative ways of computing and presenting these statistics to convey both spatial and temporal information, as well as the statistics overall. It is difficult for the reader to translate the information in the scatter plots of the temporal or spatial mean values into meaningful temporal or spatial information.

If scatter plots are used, I think they should only display the ‘raw’ (unaveraged) data. But its often hard for the reader to compare two different scatter plots that have more than ~10 data points, so I would redo all the scatter plots using the unaveraged data and put them in the supplement/appendix (for completeness) and copy the statistical values for R, mu, sigma, and the slope into a Table in the main text (something similar to Table 2, again make sure the uncertainties on each value are stated in the table).

For temporal information, I would plot the data on a time series. For spatial information, I would plot the data on a map, or for LML comparisons since there are only 6 stations, you could also do something like a box and whiskers plot (or something similar) with each station being at a different place on the x-axis. Plotting as a time series or on a map directly conveys the temporal or spatial information.

On line 593, it is stated “whereas most MAN sites are located within Natura2000 areas” How many and what percentage are in these conditions? Can you recompute the comparison statistics excluding all of these stations? Or is doing this basically the same at the set of 6 stations colocated with the LML stations? Would it be possible to provide some sort of map showing the land use, maybe like the dominant land use in each grid point, to add support to the statement that LML stations are more representative of the model domain?

Minor Comments:
Line 35: “increased oxidation of NO2 consumes HNO3” NO2 is a precursor to HNO3, so why does this consume HNO3? Please add more details here.

Line 92: Is a ‘)’ missing somewhere on this line?

Lines 99-100: “These are represented” What does ‘these’ refer to here? Later in the sentence it states that the ensemble “capture the uncertainties in both the model and observations”, why does the ensemble capture uncertainties in the observations? This sentence is a bit confusing, please consider rewriting.

Line 107: Put the equation defining X^f inline with the text after “the state vector” in line 105. The lines 105 to 112 could be rearranged to make these sentences more clear.

Line 111: “includes the application of the emissions perturbation factors,” Give more details about this.

Lines 114-116: Would increase clarity if \bar{x} was renamed to \bar{x}^f.

Line 128: Change “limit” to “approximation”.

Line 132: Change “representing” to “represents”.

Line 143: Looking at page 2558 of Shin et al. (2016), they say that they use the fifth order polynomial of Gaspari and Cohn 1999. Is there another part of Shin et al. (2016) where they use this Gaussian function? The Gaspari and Cohn is very similar to a Gaussian, so there would presumably be little difference between the two but was just confused about the reference here.

Section 2.1.3: I think the paper would flow better if the information in this section was moved to Section 2.1.

Section 2.2: I think the total column is being used from all three retrievals (IASI, CrIS, TROPOMI), but this is not clearly stated in the text. References are made to the total column in a few places in Section 2.2, but somewhere it should be clearly stated that the total columns are the observations being used in the LETKF.

Figures 2, 4, 5: It's difficult to see the differences between rows (a) and (b), and while row (c) show the relative differences, there are several places where the base values are low, so that large percentage differences may correspond to small absolute changes. Can plots of the absolute differences be added as well, or some other way of presenting the results to show where substantial changes occur? The figure might also benefit from changing the units and making the panels larger (they are on the small size and are a bit hard to read).

Lines 172-175: This paragraph should be moved to the beginning of Section 3.5, right before Section 3.5.1.

Lines 407-408: Remove the sentence “Independent evaluation against ground-based observations … atmospheric conditions.”

Lines 460-461: “This indicates that the spatial pattern ... without assimilation.” Is this sentence referring to Fig. 7 or Fig. 8? Its placement in the paragraph suggests that it is referring to Fig. 8, but would maybe make more sentence inferencing Fig.7? Please reword.

10: It is a bit hard to compare the three different sets (base, assimilation, observations) in these plots. Might be clearer with just 3 curves of the mean, maybe with a shaded standard error region around each curve, instead of the box and whiskers plot.

Lines 472-474: “…even though only morning and afternoon satellite overpasses were used” I assume this is the case because the impact from the assimilation remains in the system for at least 6-12 hours. Could you add a sentence about this?

9: For the legend and the caption, I think 'base' and 'optimized' or a similar wording is better here than 'before' and 'after'. Also, why is the title of the legend 'Sites & Mean', i.e. why 'mean'?

Line 492: “whereas the Kalman filter adjusts emissions uniformly without differentiating among source types.” But as long as emissions are adjusted on a per grid cell basis, does this matter? For whatever mix of emission source types within a grid cell, there will be a total diurnal pattern that will be a mix of the diurnal patterns of each source. But from the perspective of the assimilation, I would have thought that whatever sort of mix of diurnal profiles wouldn't matter since it is just trying to match the overall diurnal profile of that particular grid cell. Maybe some diurnal profiles are more difficult to capture than others, so maybe the assimilation will be more successful at matching the diurnal pattern from some sources over others, but I would not have thought that this is a feature of the Kalman filter not being able to differentiate between source types.

Line 512: Remove “individually”.

Fig 12: Is it a coincidence that the values for mu in Fig. 12 are identical (at least to 1 decimal place) to those in Fig. 11?

Line 454: As with the comments above for Fig. 9, change ‘before’ and ‘after’ to something like ‘base’ and ‘optimized’ or something similar.

15: What does 'absolute' mean here (second row)? If referring to the absolute value, I would assume that there wouldn't be any negative values in panels (d) to (f). Please clarify what exactly is being plotted in the middle and bottom rows. Is panel (g) supposed to tell use which run (base or opt) is better depending on if the point is blue or red? Please clarify and make a reference to this in the main text.

Lines 571-572: “This suggests that the assimilation enhances large-scale spatial and seasonal variability”. I don't follow where this comes from, please clarify.

Line 575: “show similar behavior” Does this mean similar to the comparison of MAN sites in Fig. 15 or to the LML observations mentioned in the previous sentence?

Lines 578-590: Starting with the sentence “Additionally, to support … LETKF simulations.” seems a bit out of order. Combine these two paragraphs and move this sentence to the end of the paragraph as a conclusion sentence.

Table 2: Maybe add horizontal lines to separate each base/optimized pair so that it is easier for the reader to compare the numbers between the base and optimized statistics for each case.

Reply
Citation: https://doi.org/10.5194/egusphere-2025-5926-RC1

Tyler Wizenberg, Enrico Dammers, Arjo Segers, Mark W. Shephard, Pierre Coheur, Lieven Clarisse, Martin Van Damme, Henk Eskes, Roy Wichink Kruit, Shelley van der Graaf, and Martijn Schaap

Viewed

Total article views: 250 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
149	87	14	250	12	16

HTML: 149
PDF: 87
XML: 14
Total: 250
BibTeX: 12
EndNote: 16

Views and downloads (calculated since 11 Jan 2026)

Month	HTML	PDF	XML	Total
Jan 2026	149	87	14	250

Cumulative views and downloads (calculated since 11 Jan 2026)

Month	HTML	PDF	XML	Total
Jan 2026	149	87	14	250

Viewed (geographical distribution)

Total article views: 242 (including HTML, PDF, and XML) Thereof 242 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 30 Jan 2026

Short summary

We combined satellite data on ammonia and nitrogen dioxide concentrations with a regional model to better estimate nitrogen emissions and deposition in the Netherlands. This improved the accuracy of pollution maps and trends, especially in agricultural regions. Our approach helps identify when and where emissions are highest, supporting better air quality management and environmental protection by showing the benefits of using multiple satellite datasets together.


Total:	0
HTML:	0
PDF:	0
XML:	0

Improved constraints on ammonia emissions and deposition from co-assimilating NH3 and NO2 satellite observations over the Netherlands

Viewed

Viewed (geographical distribution)

Improved constraints on ammonia emissions and deposition from co-assimilating NH₃ and NO₂ satellite observations over the Netherlands