| 19 May 2026
Satellite Retrieval of Tropospheric NO2 under Fire Conditions
Abstract. Open fires, including wildfires and planned fires, emit large amounts of nitrogen oxides (NOx = NO + NO2) into the atmosphere, and their environmental impacts are becoming increasingly severe under climate change. Satellite-based retrievals of tropospheric nitrogen dioxide (NO2) vertical column densities (VCDs) provide broad spatial coverage and continuous monitoring capabilities for assessing NOx pollution under fire conditions. However, due to the lack of explicit fire-related a priori information in current satellite NO2 retrieval algorithms, the resulting data products exhibit large uncertainties under fire conditions. Here, we use the Peking University OMI NO2 (POMINO) algorithm to investigate the impact of including fire-related a priori information on the tropospheric NO2 retrieval for the TROPOspheric Monitoring Instrument (TROPOMI) sensor. We conduct sensitivity experiments by including and excluding fire-related a priori information in the NO2 retrieval process over the western United States in September 2020, a period of intense wildfire activities. The a priori information is taken from GEOS-Chem simulations with and without fire emissions, as well as with different fire emission injection heights. In addition, NO2 retrieval based on a priori information from the Global Earth Observing System Composition Forecast (GEOS-CF) data is also conducted for comparison. Our results show that including fire-related a priori information in the retrieval significantly increases tropospheric NO2 VCDs, primarily due to enhanced NO2 concentrations in the lower layers of the a priori NO2 profile. Retrieved tropospheric NO2 VCDs increase by up to 100 % at locations greatly impacted by fires and by about 80 % in surrounding areas. Differences in fire emission injection height lead to up to ~30 % variations in the retrieved VCDs in fire regions, indicating a secondary but non-negligible effect. Validation against EPA surface NO2 measurements shows improved agreement when fire-related a priori information is included, particularly with lower biases (-12 % versus -44 %) over fire-affected regions. These results highlight the importance of incorporating fire-related a priori information in satellite NO2 retrievals to obtain more accurate data for air quality assessments under fire conditions.
In their paper, Wang et al. study the impact of various a-priori profiles - based on actual fire emission information - on the NO2 tropospheric column retrieval. The paper is relatively short (which I like) and has a clear message. It is well written, the approach and data sets used are well explained, and it contains a good set of references. The complexity and uncertainties of the retrievals over fires is an important message. I found the comparison with the ground based observations compelling, even though uncertainties are large in transforming columns to surface concentrations. The sensitivity experiments, in particular the injection height, are interesting. Therefore the paper deserves to be published to my opinion.
However, I have a set of general and specific comments which I would like the authors to answer before the paper is published.
First I would argue that the authors do not make a fair comparison with the TROPOMI-DOMINO product. This official product is based on coarse 1 degree model output from TM5, a system that does not contain daily information on fires and can not resolve emissions at the scale of the actual TROPOMI footprint (roughly 0.05 degree resolution). This is well known, e.g. see Douros et al., https://doi.org/10.5194/gmd-16-509-2023, who discuss the two optimal ways of comparing with satellite NO2 retrievals (not influenced by the a-priori profile shape), and describe a European TROPOMI product based on CAMS-Europe a-priori profiles with a resolution of 0.1 degree. In the TROPOMI Product User Manual users are encouraged to either make use of the kernels to compare with model output, or to replace the TM5 a-priori (using kernels and AMFs from the product) to obtain a more quantitative retrieval product and comparison on the regional scale.
I would like to encourage the authors to show the result of a "DOMINO-GC-Fire" product, where the NO2 a-priori of the official product is replaced by the GC-Fire GEOS-Chem tropospheric profiles, and to be compared to POMINO-GC-Fire.. This additional product does allow a more in depth discussion of the differences between POMINO (with explicit aerosol treatment) and the DOMINO approach (with an implicit aerosol-cloud approach).
The focus of the paper is on the impact of a more realistic a-priori profile on the column retrieved over fires. However, this is not the only factor impacting the retrieval. Fire plumes are very complicated and diverse, depending on the fuel type, amount of water, pyroconvection, chemical reactions and aerosol formation. The optical properties, impacting the retrieval and air mass factors are also complex. These complications of the fire plumes should should be discussed in more detail. What would be an order of magnitude uncertanty estimate linked to such plume aspects? Please distinguish profile aspects from other AMF related uncertainties.
A direct comparison between the GC-Fire modelled columns and the POMINO retrieval columns would be a useful extra result to show. Is the model able to produce realistic column amounts over/near fires?
The authors mention differences in quality filtering between the POMINO and operational TROPOMI products, related to cloud fraction estimates. It would be useful to show some more results and extend the discussion on this, with maybe an extra figure to show the differences in coverage. The authors present a cloud height sensitivity experiment, but more analysis on the clouds would be useful. Are cloud heights realistic for these fire cases?
The comparisons with the surface observations are interesting and relevant. But these depends on many modelling, retrieval, and profile details and the uncertainty in the factor 2. Quantitative conclusions beyond an order of magnitude are difficult to make it seems. Pandora is a good validation source, but, as shown by the authors, they do not sample the fire regions. It would be good to explore other possible datasets. In particular, are there aircraft campaign datasets that may be exploited, e.g. FIREX-AQ? Profiles from such campaigns will provide additional and more quantitative evaluation material.
It would be important to elaborate on the overall uncertainty estimate, and contributions to the uncertainties in the retrieval product.
More minor comments:
l 49, l 199: The later paper of van Geffen, 2022, is more relevant:
https://doi.org/10.5194/amt-15-2037-2022
l 116: The ATBD contains the detailed description of the algorithm. Could you please refer to the latest update of the ATBD and to van Geffen, 2022, which described v2.2, which is similar to v2.4. The ATBD, PUM and ReadMe are available from the S-5P Sentwiki products page, https://sentiwiki.copernicus.eu/web/s5p-products. Please include the ATBD or this site as reference.
Table 1: TROPOMI RPRO v2.4: Cloud fraction is not from FRESCO, but is computed in the NO2 fitting window. Please add the version of the DLER (v1.0).
l 227, p10: Correction factor of 2. Could you include a rough estimate of the uncertainty of this factor? Translating columns to surface concentrations has significant uncertainties, and discussing these in more detail would be useful.
Figure S2: Could you add the methematical equation for the generalised normal distribution (e.g. in the supplement)? Is it a mix of exponential and Gaussian? Is it symmetric in positive and negative? I did not undestand the plot S2. Since it is the difference between fire and nofire, I would expect it to be asymmetrical, with a tail on the positive side. This is not so clear from the figure. Why is the 95% applied on both the positive and negative sides?
l 283: "providing a robust basis ". Enhancements are clearly detected, but "robust" for me would mean that realistic error estimates are available. The paper provides sensitivity experiments, which can serve to quantify some components of the uncertainty, but are not full uncertainty estimates combining all major contributions to the VCD uncertainty.
l 288: "This sampling approach leads to a smaller number of valid data points in TM5-MP-DOMINO than in POMINO-CF by 11% in our study domain." This is interesting and it would be good to see more details, e.g. maps comparing the filtering (or qa_value) for both products.
l 351, 363: The injection height sensitivity test 100%, 65% and 5% emissions injected in the BL. Is this a realistic range of uncertainties related to injection heights? In other words: is the range of column results indicative of the uncertainty in the VCD linked to injection height?
l 370: Similar question: is the +- 50hPa a realistic perturbation, representing uncertainties in cloud top pressure? It would be useful to include a histogram of cloud pressure values over the fire regions to compare with the profiles in Fig. 4.
Figure 6: Please mention the type of linear fit that is used. Is it ordinary least-square fitting? It is good to use a fitting method that is x-y axis symmetric.
Conclusion: Can the authors provide an estimated overall uncertainties of the NO2 column amount for fire-affected pixels? The injection height gives variations of 30%, even though the impact on the profile shape is not dramatic (Fig.4). Other aspects, e.g. the optical properties and chemical/aerosol composition of fire plumes, may lead to significant additional uncertainties. Please discuss this.
Henk Eskes