The NOx point source catalog is based on NO2 TVCDs from TROPOMI for the period from May 2018 to November 2021, using the consistently reprocessed data product provided via the S5-P Products Algorithm Laboratory (PAL) based on NO2 processor version v2.3.1. The main improvement of the PAL product compared to the product versions (≤ v1.3) used in is the change in the cloud product due to an updated FRESCO algorithm, generally leading to higher cloud altitudes and thus lower AMFs and higher TVCDs. In addition, “for cloud-free scenes a surface albedo correction is introduced based on the observed reflectance, which also leads to a general increase in the tropospheric NO2 columns over polluted scenes of order 15 %” . Both changes lead to an overall increase of NO2 TVCDs of about 10 %–40 %.

Meteorological data are taken from ERA5 reanalysis   provided by the European Centre for Medium-Range Weather Forecasts (ECMWF). ERA5 data are used with a truncation at T639, corresponding to ≈ 0.3∘ resolution.

In order to reduce the data amount, we created an intermediate meteorological dataset in which the original model output, containing horizontal wind fields (needed for the calculation of horizontal fluxes) and temperature and pressure (needed for estimating the NOx/NO2 ratio), was interpolated on a regular horizontal grid with a resolution of 1∘ and stored in intervals of 6 h.

In the analysis below, horizontal wind fields are vertically interpolated to 500 m (default)/300 m (sensitivity analysis) above ground level (a.g.l.) (see Sect. ).

As in v1, “ozone mixing ratios, used for the scaling of NO2 to NOx, were taken from the Earth System Chemistry integrated Modelling (ESCiMo) project , using the RC1SD-base-10a simulation for the years 2000–2010. The monthly mean climatology was calculated from the model fields sampled online along the overpass time of OMI aboard Aura (which is close to the TROPOMI overpass time) using the MESSy SORBIT submodel . As the divergence is sensitive for the added NOx at the source, the relevant NOx/NO2 ratio is that close to ground. We thus took O3 concentrations from the lowest model layer” .

TROPOMI's directionally dependent surface Lambertian-equivalent reflectivity (DLER) v1.0 , based on the algorithm described in , is taken from https://www.temis.nl/surface/albedo/tropomi_ler.php (last access: 27 June 2023). The minimum LER for clear conditions, averaged over all months, at 440 nm is used for identifying regions with good observation conditions (i.e., deserts).

We use the Global Power Plant Database (GPPD) , in order to automatically identify NOx point sources corresponding to power plants. The GPPD lists about 35 000 power plants of all kinds, including solar, nuclear, and hydro power. For our purpose, we created a subset of those power plants using coal, gas, oil, petroleum coke, biomass, and waste as primary fuel and skip power plants with capacities below 100 MW.

We make use of the latest release (v1.3) of GPPD. However, this update does not include power plants that have been shut down recently but were still active during the time period investigated in this study. For instance, the Navajo power plant was one of the top emitters of NOx in the USA in 2019, as reported by EPA and also listed in v1 of the catalog. This power plant was shut down at the end of 2019 and is consequently not listed in GPPD v1.3.

Thus, we extend v1.3 of the GPPD by all power plants from v1.2 which are not included in v1.3. This adds 45 power plants, which we labeled as “(v1.2)” in the combined GPPD database.

The resulting GPPD database comprises 4741 power plants, of which 2291, 1995, and 378 use gas, coal, and oil as primary fuel, respectively.

In order to automatically identify cities close to the detected NOx point sources, the basic version of the World Cities Database (WCD), provided at https://simplemaps.com/data/world-cities (last access: 27 June 2023), is used. Only cities with more than 100 000 inhabitants are considered.

For validation purpose, we compare the NOx point source emission catalog to the following regional emission databases:

The German Environment Agency (Umweltbundesamt, UBA) provides the Pollutant Release and Transfer Register (PRTR) for Germany . The PRTR contains annual NOx emissions of all facilities, covering energy sector as well as metal, chemical, mineral, and other industries. In this study, we have included PRTR data for the years 2018–2020.

As in v1, TROPOMI data are restricted to qa values (quality indicator of NO2 TVCDs provided in the operational products) above 0.75, as recommended in , removing cloudy pixels (cloud radiance fractions above 50 %) as well as anomalies (like solar eclipses) in the TROPOMI NO2 dataset. Also, the selection of solar zenith angles (SZAs) below 65∘ is the same as in v1, restricting the TROPOMI data to favorable observation conditions for tropospheric NO2 from space. In particular, over midlatitudes, measurements in winter are skipped by this selection (Fig. ), also avoiding complications due to potential snow cover.

In v2, viewing zenith angles (VZAs) are restricted additionally to values below 56∘, avoiding less favorable viewing conditions at the swath edges. In addition, this selection limits the maximum pixel width (across track) to 11 km. In contrast to v1, no regional preselection of potentially polluted regions was made in v2. Only high latitudes (north from 72∘ N or south from 50∘ S) are skipped directly.

In this study, horizontal transport is described by horizontal wind fields at a fixed “plume height”. This is a simplifying assumption, as the emissions take place at a stack height of about 200 m but are uplifted and vertically mixed within the boundary layer during downwind transport.

For the quantification of point source emissions, the focus of this study is set to the horizontal transport close to the point source, where spatial gradients are largest. As shown in , power plant emissions at 200 m stack height quickly rise to about 500 m within the first hundred meters. investigated the effective height of CO2 emissions for atmospheric transport simulations. This is closely related to the question of which altitude has to be considered in order to describe horizontal transport of a fresh power plant plume appropriately. For summer around noon, they report mean effective heights of about 450 m (with a long tail towards larger values).

In this study, we assume an effective plume height of 500 m a.g.l. For individual stations and specific meteorological situations, systematic deviations might occur. In order to quantify the impact of this assumption, we thus also performed the analysis for a plume height of 300 m (see Sect. ).

ERA5 wind fields are vertically interpolated to the assumed plume height (Sect. ). In addition, the AMF correction is applied consistently for the same height (Sect. ).

The AMF and the averaging kernel (AK), reflecting the total and height-dependent sensitivity of satellite measurements for atmospheric trace gases, are key concepts for the interpretation and quantification of trace gas column densities . The operational NO2 TVCD is based on AMFs calculated for an a priori vertical profiles of NO2 taken from a global chemistry model. With the AK provided in the TROPOMI data, i.e., the ratio of height-dependent (box-)AMF to the total AMF, the AMF can be adjusted to a different a posteriori vertical profile .

In polluted regions, vertical profiles of NO2 are generally highly complex, and a local NOx source can typically not be adequately represented by global chemistry models with comparably coarse spatial resolution. In the case of NOx point sources, however, the horizontal gradient in NO2 is basically sensitive to the NO2 excess added by the point source. Any “background” NO2 (which might be considerably polluted in densely populated regions) is intrinsically corrected for by the spatial derivative, i.e., the difference between NO2 levels upwind and downwind of the point source.

Thus, for the quantification of point source emissions, the AMF has to be corrected with respect to the NO2 excess added by the point source. Hence, we apply an AMF scaling factor 1cAMF=AMFplumeAMFPAL, where AMFPAL is the tropospheric AMF applied in the PAL product, and AMFplume is calculated from the AK based on a delta-peak profile at plume height (default 500 m); i.e., cAMF reflects how much higher the plume AMF is compared to the a priori value. For the detected point sources, the AMF correction is about 1.61 ± 0.32. Figure a displays the respective frequency distribution for the point sources listed in the v2 catalog.

Note that in v1, no AMF correction was applied, as the AK provided in the TROPOMI data used in v1 was based on a cloud height that was reported to be biased low , as discussed in , while no reprocessed dataset was available at that time.

As in v1, the TROPOMI NO2 TVCD is upscaled to NOx by the scaling factor cNOx, which is calculated based on the photo-stationary state (PSS) according to 2cNOx:=[NOx][NO2]=1+[NO][NO2]=1+Jk[O3], where

the photolysis frequency of NO2 J is parameterized as 0.0167×exp⁡(-0.575/cos(SZA)) s-1, as proposed by , with SZA taken from TROPOMI;

In , the divergence of the horizontal flux F=wV, with horizontal wind fields w and TVCD V, was calculated: 3D=∇⋅F. According to product rule, this equals 4D=∇⋅(wV)=w⋅∇V+V∇⋅w.

The first term is the scalar product of the (horizontal) wind vector and the spatial gradient of the TVCD; in meteorology, this is denoted as “advection” in the sense of “the rate of change of the value of the advected property” . Below, we use the term advection in this sense for the quantity A: 5A:=w⋅∇V.

The second term of Eq. () reflects the divergence of wind fields scaled by the TVCD. However, as we are interested in flux changes caused by local NOx emissions rather than by non-vanishing divergence of the wind fields, we now directly calculate the advection according to Eq. (), as also proposed recently by . Therefore, in v2, the impact of non-vanishing divergence of the wind field is explicitly skipped. However, the resulting mean maps of A and D are very similar, as the temporal mean divergence of wind fields (at 500 m above ground) is negligibly small. Thus, the switch from “divergence” to “advection” is rather a change in terminology but appropriately describes the retrieval steps that have actually been implemented in the processing of the v2 catalog.

As in v1 of the catalog, only observations with wind speeds above 2 m s-1 are considered in the further processing.

In , spatial derivatives were calculated after gridding the TROPOMI data on a regular latitude–longitude grid. In contrast, proposed to calculate spatial derivatives directly on the native TROPOMI grid (along-track and across-track).

The main advantage of taking derivatives directly on the TROPOMI grid is the handling of gaps (e.g., due to cloud masking): on the TROPOMI grid, a gap in TVCD just results in a gap in the respective gradient; in contrast, if the derivative is calculated for a temporal mean on a regular latitude–longitude grid, as in v1, gaps on individual days cause steps in the mean distribution, resulting in spikes in the spatial derivatives.

In order to calculate the advection on the TROPOMI grid, the following steps are performed:

For each TROPOMI pixel, horizontal wind fields from ERA5 are interpolated linearly to the assumed plume altitude (default: 500 m above ground) and to the observation time and latitude and longitude of the TROPOMI pixel center.

In v1, systematic artifacts of the divergence map were reported over mountains with high tropospheric TVCDs, in particular over parts of China, which hinders the identification and quantification of NOx point sources. These artifacts were explained by inaccurate wind fields over mountains in . However, a recent study by shows that these patterns are rather caused by 3D transport effects which have been ignored so far in the simplified 2D divergence approach.

derives a “topography-wind” term in order to correct for this effect: 6Ctopo:=V/Hsh⋅w0⋅∇z0, with NOx TVCD V (without AMF correction), NOx scale height Hsh, surface wind speed w0, and surface elevation z0.

We include this correction term in order to account for topographic effects: Ctopo is calculated for each TROPOMI pixel based on the surface elevation and surface wind speed (10 m) provided in the PAL NO2 data and assuming an a priori NOx scale height of 1 km.

The topography-corrected advection is then derived as 7A*:=A+f⋅Ctopo, where the scaling factor f is derived empirically as 1.5 (corresponding to a net NOx scale height of 1/1.5 km = 667 m) in order to minimize topography effects, as shown in Appendix . From now on, we denote the topography-corrected advection as A* in labels and equations but still refer to it as “advection” in the text for the sake of simplicity; i.e., the application of the topographic correction is implied below.

For each TROPOMI orbit, the NOx advection derived on the TROPOMI grid, as well as all other relevant variables like NOx and AMF scaling factors, wind speed, or topographic correction, is re-gridded on a regular lat–long grid with 0.025∘ resolution, considering latitudes from 50∘ S to 72∘ N. Afterwards, temporal averages are calculated for daily, monthly, and annual periods as well as for the complete time series covered by the PAL NO2 product (May 2018–November 2021).

The temporal mean advection map of A* is the basis for the identification and quantification of NOx point sources. A high-resolution map is provided in the Supplement. Note that regions with less than 10 % temporal coverage have been skipped. This criterion removes regions with poor statistics of the filtered data, caused by frequent cloud cover, snow and ice cover, and/or low wind speeds.

As in , point sources are identified in an automated iterative process in which local maxima of the temporal mean (May 2018–November 2021) advection map are successively checked for being point sources. The criteria for classifying potential point source candidates have been extended and modified, as explained in detail below.

In v2, a default radius 15 km is considered for the quantification of NOx emissions, in contrast to 22 km in v1. This reduces the skipping of point sources due to interfering sources nearby. A new quantity used during the categorization procedure is the “peak area fraction” which is just defined as the percentage of grid pixels (on 0.025∘ grid) within a given radius around the candidate that have advection values above a threshold (here: 30 % of the local maximum). This quantity helps to identify single spikes (with very low peak area fraction) as artifacts, as well as area sources (with large peak area fraction).

Point sources are identified by a fully automated iterative procedure, in which a “candidate” is identified and classified in each iteration step:

The candidate location is defined by the absolute maximum of the advection map. Figure a displays the maximum advection value as a function of the iteration step.

For the v2 catalog, 50 000 candidates have been processed. In the beginning, a high fraction of candidates is classified as a point source (Figs. b and a). For the first 330 iterations, where maximum advection is >2 µg m-2 s-1, 52 % of all candidates are found to be point sources and another 32 % as area sources. In later iteration steps, while maximum advection decreases by almost 2 orders of magnitude, the majority of candidates is classified as gap (initially due to actual gaps in the input data, later due to the removal of prior candidates nearby) or negative (due to artificial dipolar patterns and due to maxima close to the local advection noise level). Within iterations 40 000–50 000, only 10 significant point sources have been found, and further iterations are not meaningful.

For the candidates identified as point source, the respective NOx emissions are quantified by spatial integration of the advection map (Sect. ), corrected for NOx lifetime (Sect. ).

The iterative classification algorithm yields 2475 point source candidates. For v2 of the catalog, we select significant and reliable point sources by different criteria, i.e., the detection limit, the integration error, the contribution from topographic correction, and the temporal persistence of the derived emissions.

Version 2 of the NOx point source catalog can be found at https://doi.org/10.26050/WDCC/No_xPointEmissionsV2 .

Based on consistently reprocessed TROPOMI NO2 data for the time period May 2018 to November 2021 (PAL product), combined with wind fields from ERA5, we compiled an updated catalog (v2) of NOx emissions from point sources worldwide.

Compared to v1 of the catalog , several improvements were implemented; the most important ones are

the usage of the PAL product ,

Steps 2 and 3 reduce noise and systematic artifacts, respectively, in the temporal mean advection maps, allowing for the automated detection of NOx point sources (1139 compared to 451 in v1). Steps 1 and 4–6 result in far higher (factor of ≈ 3) and more realistic NOx emissions, where the strongest contribution comes from the AMF correction (≈ 1.6) and the lifetime correction (≈ 1.4). Comparisons to PRTR emissions for Germany and eGRID emissions for the USA agree well, with a remaining low bias of about 20 % of the updated catalog for the detected point sources (excluding those close to cities).

Due to step 6, shorter time periods like annual means can also be considered, and the annual emissions included in the v2 catalog do reflect for instance the reduction of power plant emissions from Niederaußem and Neurath between 2018 and 2020 or the shutdown of the Navajo power plant at the end of 2019 well.

Future updates will focus on including wind fields with improved spatial and temporal resolution. In addition, the impact of 3D radiative transfer effects on AMFs for power plant plumes will be investigated in more detail.