The inversion algorithm DECSO (Daily Emissions Constrained by Satellite Observations) has been developed at KNMI for the purpose of deriving emissions for short-lived gases (Mijling and van der A, 2012, Fig. 1). DECSO is using a Kalman Filter implementation for assimilating emissions, which includes the following steps: the sensitivity matrix (or Jacobian) to describe the relation between gridded emissions and satellite observations is derived using an ensemble of trajectories calculations with an estimated effective lifetime of the trace gas. The satellite observations are compared to the concentrations modelled by the chemical transport model (CTM) CHIMERE v2020r3 (Menut et al., 2021). The difference between model and observations is used to adapt the emissions within the data assimilation scheme. Since these adaptations are based on additions or subtractions and no scaling is involved, emerging or disappearing emission sources are also detected. Note that the a-priori emissions are not from an external inventory but based on the emissions of the previous day which are updated based on the mismatch between observations and forecast with a spin-up period of 200 days before the emission results are used. After this period any effect of the day one a-priori emissions is gone due to the additive assimilation scheme. As prior emissions for day 1, HTAP v3 was used (Crippa et al., 2023). DECSO is currently designed to estimate NO_x and/or NH₃ emissions. Detailed information on the inversion method can be found in Mijling and van der A (2012).

For deriving NO_x emissions, we use NO₂ observations of the TROPOMI instrument on board Sentinel 5P (van Geffen et al., 2022). Only observations with a cloud radiance fraction of less than 50 % (this corresponds to a cloud fraction of about 20 %) are selected for input to DECSO, which is about 50 % of all observations. The precision of the derived NO_x emissions is estimated as 8 % for annual emissions and 25 % for monthly emissions (van der A et al., 2024).

The version we used at the start of this study is version 6.4. It is actually the same code as version 6.3, published in van der A et al. (2024), Ding et al. (2024), and Lin et al. (2024), except that the vertical emission profiles for industrial sources used in the CTM were updated following the analyses of Bieser et al. (2011). In this study we focused on the Netherlands (50–54° N, 2–9° E), which is a relatively small country, and a higher resolution emission grid of NO_x emissions was required for a detailed intercomparison. To facilitate such an intercomparison a new high-resolution version of DECSO was developed named DECSO-HR. While the original version of DECSO had a 0.1° resolution, the newly developed DECSO-HR focussed on the Netherlands at a 0.05° resolution, comparable to the resolution of the TROPOMI observations, which are 3.5×5.5 km at nadir. Besides small changes concerning the harmonization of the parameters related to spatial and temporal resolution (for example a smaller time step and a shorter distance for the trajectory calculations) the main difference between the standard DECSO and DECSO-HR is their input. The TROPOMI observations used for DECSO-HR are more numerous and have higher uncertainties than the super-observations (area-weighted means of cloud-free TROPOMI observations, see below) used in DECSO. On the other hand, the direct observations are better spatial matched with the emissions. Together with additional developments of the algorithm to improve its quality, this has led to the release of version 6.5. Some of these points of improvement were notably revealed by the comparison to the Dutch emission inventory.

The changes from 6.4 to 6.5 included some minor corrections such as including the seasonal cycle of VOCs, adapted vertical emission profiles for ship emissions and an error correction for high concentrations to avoid the assimilation to favour the majority of low emissions in the region. The main changes that will be discussed below were related to the use of a new superobservation code, a correction of NO₂ observations in winter, and a fixed lifetime specifically for the high-resolution version.

DECSO is using superobservations, which are area-weighted means of cloud-free TROPOMI observations over the model grid cells. The software to construct these superobservations has been replaced by the software of Rijsdijk et al. (2025), which provides a more realistic description of the superobservation uncertainties. At the same time, the observations of the tropospheric column till 500 hPa are calculated. However, these superobservations are not used for the HR-version of DECSO because the model grid and the TROPOMI resolution are almost equal and the advantage of superobservations disappears. For the HR version the tropospheric column of NO₂ till 500 hPa is used directly as input to DECSO. Another change for the HR-version of DECSO is that the effective lifetime of NO_x is no longer derived locally but fixed to 3 h. This significantly improves the calculation time at the cost of slightly slower convergence speed to match the emissions to the observations. This lifetime is only part of the trajectory calculation within DECSO in addition to the intrinsic NO_x lifetime of the CHIMERE model. While the former lifetime determines the convergence speed of the emission updates, the magnitudes of the emissions are mainly impacted by the intrinsic NO_x lifetime of the CHIMERE model, since the derived daily emissions are based on minimizing the difference between the model output and the satellite observations. The difference of the final monthly emissions with and without fixed lifetime in the DECSO trajectory calculation is negligible.

During this study we noticed the TROPOMI NO₂ observations were unrealistically low or even negative during days in the winter with strong wind coming from northerly to westerly directions. This was corrected by using a reference sector above a relatively clean area over the North Sea to determine the bias between the CTM of DECSO and the NO₂ retrievals. The effect of this correction can also be seen in Fig. 4b.

Although DECSO provides additional information on anthropogenic emission and soil emissions (Lin et al., 2024), in this paper only the total NO_x emissions are analysed.

The national Dutch emission inventory (in Dutch Emissieregistratie) reports emissions for 375 different substances towards air and water, including the most important air pollutants like NO_x, SO_x and particulate matter and all greenhouse gases. It also compiles the Informative Inventory Report about air pollutant emissions (Staats et al., 2025) which is annually published to fulfil obligations under CLRTAP and NECD. The methods used to calculate emissions are detailed in the yearly method reports of each sector (Honig et al., 2025; van der Zee et al., 2025; Visschedijk et al., 2025; Witt et al., 2025). Most data from the emission inventory, including spatial gridded data and emissions per province and municipality for this analysis were downloaded from https://www.emissieregistratie.nl/data (last access: 19 March 2026). Some additional data like the large point source data were produced by the emission inventory on special request.

To calculate national, provincial and municipal Dutch emissions from the satellite data grid cells within the Netherlands, excluding the North Sea, were selected using the exactextractr package in R (Baston, 2025). For grid cells partially in an area, the emissions were distributed based on the share of the grid cell within the Netherlands. Regridding of inventory data to the DECSO resolution was carried out using the resample function of the terra package in R by calculating mean fluxes (Hijmans, 2025). Monthly inventory emissions were calculated by mapping the emissions from the emission inventory to GNFR sectors in order to use temporal emission profiles from the CAMS-TEMPO v4.1 data set (Guevara et al., 2021). The majority of NO_x emissions in L_AgriOther in the Netherlands is from application of manure to agricultural soils. However, the profile for NO_x emissions from L_AgriOther in CAMS-TEMPO is based on burning of agricultural waste, a practice not legal in the Netherlands. Therefore, the profile for NH₃ was used instead.

For additional emission data outside the Netherlands CAMS-REG-ANT v8.1 was used (Kuenen et al., 2022), which utilizes the reported national emissions and uses a consistent method for gridding of the emissions. Monthly emissions were estimated using CAMS-TEMPO v4.1 (Guevara et al., 2021). Official national, provincial and municipal borders of the Netherlands were downloaded from PDOK (2024). Information on the Dutch road network were downloaded from RWS (2024).

The uncertainty of the NO_x emissions quantified with DECSO at the scale of the Netherlands was previously quantified and estimated as 8 % (95 % confidence interval) (van der A et al., 2024). Additional uncertainties arise from the applied country mask, which have so far not been quantified. The uncertainty of the total Dutch NO_x emissions in the Dutch National Inventory was previously calculated with a Monte-Carlo analysis based on the estimated uncertainties of individual emission sources and was found to be 19 % (95 % confidence interval) (Staats et al., 2025). Uncertainties in the spatial or temporal distribution of the Dutch National Inventory have so far not been analysed.

Figure 2 shows a comparison of emissions maps for the year 2019 as reported by the Dutch national emission inventory and as derived from satellite observations analysed with the DECSO algorithm (total emissions) using the original DECSO 6.4.1 at 0.1° resolution and the newly introduced DECSO-HR 6.5 with a resolution of 0.05°. As the Dutch national emission inventory covers only the Netherlands, emissions for neighbouring countries were added from CAMS-REG-ANT (Kuenen et al., 2022). To allow for an easier visual comparison, the inventory dataset is also shown interpolated to the DECSO-HR resolution of 0.05°.

First, comparing DECSO 6.4.1 and DECSO-HR 6.5, the increased resolution clearly provides additional detail. While large emission hotspots within and outside the Netherlands, like Amsterdam, Rotterdam, Antwerp and the Rhine–Ruhr area can be recognized in both data as reported earlier (van der A et al., 2024), the improved resolution of DECSO-HR 6.5 reveals additional valuable details, like distinguishing the emissions from Amsterdam and the industrial complexes in IJmuiden (in the North-West of the Netherlands, see also Fig. 3) or resolving individual emission hotspots within the German Rhine–Ruhr area (in the South-East of the domain).

Next, the comparison between DECSO-HR 6.5 (Fig. 2b) and the national emission inventory (Fig. 2c and d) directly shows the good spatial agreement between these two independent emission datasets. Both emission maps clearly show the urban emission hotspots of Amsterdam, Rotterdam and Utrecht in the Western part of the Netherlands. On the inventory emission maps one can also clearly recognize the network of high-ways and rivers (used for inland navigation) in the Netherlands. These features are also indicated in the DECSO-HR 6.5 emission map (Fig. 2b), while not being recognizable in DECSO 6.4.1. Interestingly, while not all shipping routes on the North Sea in the inventory dataset are resolved in DECSO-HR 6.5, the main shipping routes towards Rotterdam and Antwerp are still captured.

Zooming closer on the biggest cities Rotterdam and Den Haag (Fig. 3, left panel) and Amsterdam (Fig. 3, right panel) and their surrounding industrial areas in DECSO-HR 6.5, one can clearly see the good correlation between registered point sources (blue circles) and the satellite emissions estimates. Also Schiphol airport (runways marked with black diamonds) is clearly visible in the satellite data set. Interestingly, the highway network around these cities (indicated in rainbow colors) is not visible in the DECSO-HR emission maps despite road traffic being a major source of NO_x as will be discussed further below.

Next, for a quantitative comparison of the emission inventory with satellite-based emission estimates, we compare the timeseries of total national emissions from 2019 to 2023 for DECSO 6.4.1 and DECSO-HR 6.5 as shown in Fig. 4. To calculate national total emissions from the DECSO datasets, grid cells within the Netherlands, excluding the Dutch part of the North Sea, were added up. For the time span investigated here, the national inventory reported monotonously decreasing emissions from 76 kt (N) in 2019 to 61 kt (N) in 2023 (Staats et al., 2025). In comparison, DECSO 6.4.1, the latest version of DECSO when this work started, reported only 44 kt (N) in 2019. By contrast, DECSO-HR 6.5 was much closer to the inventory with 68 kt (N) in 2019 and agrees with the inventory within the uncertainty range for all years except for 2021.

There is, however, a notably difference in the inter-annual variation between the different emission data: While the inventory emissions decline monotonously from 2019 to 2023, DECSO-HR 6.5 reports a minimum in 2021 and an increase from 2021 to 2022. This might be due to the effects of the COVID pandemic not being captured accurately by the emission inventory. While changes in the activity data, i.e. the economic activity, were presumably accurately captured by national statistics, there are also possible, less obvious effects on the emission factors, for example lower emission factors from road traffic as traffic was less congested. While there were attempts to also include these effects, these certainly lead to higher uncertainties in the inventory emissions that are so far not well understood. These data thus might imply that emission reduction during the pandemic was underestimated by the national inventory, but other explanations are also feasible, for example meteorological conditions.

Despite these differences, fitting a trendline to the emissions from both DECSO-HR 6.5 and the inventory (dashed line in Fig. 4a) led to similar emission trends with a yearly reduction of 3.1 kt (N) for DECSO-HR 6.5 and 3.3 kt (N) for the inventory, showing that emission trends on intermediate timescales of several years agree between the two data sets.

Besides the inter-annual emission trend, we also compared the intra-annual variation of emissions. Therefore, mean monthly emissions from satellite observations were compared to monthly inventory emissions (Fig. 4b). As the Dutch national inventory only produces annual emissions, monthly inventory emissions were estimated by distributing the inventory emissions per GNFR sector using the CAMS-TEMPO temporal profiles (Guevara et al., 2021). This comparison revealed large differences in the temporal profiles. While the inventory emissions were roughly constant throughout the year with higher emissions from stationary combustion and road transport in winter being compensated by higher emissions from agriculture and non-road mobile machinery in summer, DECSO 6.4.1 shows a strong minimum of emissions in winter, with emissions in January being just 20 % of emissions in June (Fig. 4b). This deviation in winter also explains the large deviation in yearly emissions between the inventory and DECSO 6.4.1 in Fig. 4a. The low DECSO emissions in winter were caused by a bias in the TROPOMI NO₂ retrievals for observations with a strong north-western wind in the winter when NO₂ observations were close to zero or even negative. Correction of this bias in DECSO-HR 6.5 (see Sect. 2.1) therefore remedied this winterly mismatch. Monthly emissions according to DECSO-HR 6.5 agree within the uncertainty range with inventory emissions.

Taken together, the improvements introduced with DECSO-HR 6.5 not only provided more spatial information but also improved the match to inventory emissions.

After analysing the temporal match of the emission datasets, we next considered their spatial correlation as shown in Fig. 5. Figure 5a shows a difference map for 2019 between the national inventory and DECSO-HR 6.5. In general, this map confirms the good agreement between the two data sets. There are no systematic deviations or biases apparent on a national scale and deviations could only be observed in the Amsterdam/IJmuiden and Rotterdam areas (Fig. 5b). In these areas, DECSO yields overall lower emissions than reported in the inventory. Interestingly, it appears that in these areas there might be an alternating pattern of positive and negative deviations, which will be further discussed below.

In the zoom-in views of urban areas, also highways and Schiphol airport are indicated as the transport sector is the main source of NO_x emissions in the Netherlands. Schiphol airport (with black diamond's indicating the runways) is clearly visible in the DECSO emission maps (Fig. 3). Considering also the difference maps (Fig. 5b), DECSO-HR 6.5 reports higher emissions than the national inventory over Schiphol airport. This might be caused by the national inventory only reporting emissions for the landing and take-off phase of the flights (LTO) up to 3000 ft (∼914 m) according to international agreements, thereby missing emissions above 3000 ft, which have been estimated at 4 kt (N) over the entire Netherlands (Witt et al., 2024) of which about 2/3 are related to air traffic at Amsterdam Schiphol airport. These emissions will in parts still contribute to satellite observations of column densities. The highway network on the other hand was not very apparent in the emission maps (Fig. 3) or the difference maps (Fig. 5b), indicating that as a diffuse source road traffic will be difficult to spatially isolate in the DECSO emissions.

As these two areas around Rotterdam and Amsterdam have large industrial sides and harbours and consequently high emissions, we hypothesized that the deviation between the two datasets was depending on the intensity of the fluxes. To further explore these deviations, emissions according to DECSO-HR 6.5 were plotted per grid cell as a function of the inventory emissions in that same grid cell (Fig. 6a), which revealed two different behaviours at low and high emissions. At low emissions (below roughly 0.25 kt (N), yellow shading) emissions per grid cells showed on average good correlation (dark blue circles), albeit with some scatter. At larger emissions (blue shading), however, the two datasets started to deviate, with the mean DECSO emissions moving further and further from the unity line. This can be seen even more clearly when considering the difference between the two datasets as a function of inventory emissions (Fig. 6b). The mean deviation was close to 0 at small emissions below 0.25 kt (N) (yellow shading) and started to increase at higher emissions (blue shading). It reached 0.5 kt (N) for grid cells with emissions around 1 kt (N), a deviation of almost 50 % for these highly-emitting grid cells. These deviations at high emissions had a large impact on the total estimated national emissions: Grid cells with an emission larger than 0.5 kt (N) in the emission inventory were responsible for 53 % of the observed deviation between inventory and satellite observations on a national scale in 2019, despite representing only 0.4 % of grid cells.

This leads to the question where these large deviations are coming from, an underestimation by the satellite-derived emissions or an overestimation by the inventory. On the inventory side, these high-emission grid cells relate to areas of high industrial activity whose emissions are registered as large point sources. The quantity of these emissions is being reported by the companies themselves, according to standardized methods and controlled by local environmental authorities (Staats et al., 2025). That means that the location of these emission sources is precisely known leading to neglectable spatial uncertainty. On the other hand, while there are certainly uncertainties in the amount of emissions, there are no records of systematic overreporting of industrial emissions. Looking at satellite-derived emissions, a part of this deviation might be explained by the emissions from large point sources being wrongly allocated to neighbouring grid cells, in other words appearing smeared out over a larger area which might be caused by uncertainties in the wind fields. This would lead to an alternating pattern of over- and underestimated grid cells, as can indeed be seen in Fig. 5b. However, while this fits qualitatively, these alternating pixels do not compensate each other quantitively, meaning that even when adding these grid cells up DECSO reports lower emissions for highly emitting areas than the inventory. In absence of a known ground truth, more research and further independent assessments will be needed to fully understand these deviations.

Figure 6c shows the same data as Fig. 6a but plotted on a double-logarithmic scale in order to better assess grid cells with low emissions, which are dominated by diffuse sources. Here, again, different behaviours can be distinguished. At very low inventory emissions below 0.01 kt (N), the DECSO emissions reached a plateau at around 0.01 kt (N), which is presumably caused by NO_x emissions from soils. The overall contribution of these grid cells to the total emissions is very small with only 3 %, despite representing 22 % of all grid cells. Above 0.01 kt (N) the two data show a good correlation, however with a slight offset towards inventory emissions. This means that also in these low-emitting grid cells the inventory reports on average higher emissions than DECSO, albeit with a very small difference.

Taken together, a good spatial correlation between the two emission data sets was observed, with deviations at both ends of the emission range. Of these two deviations, however, the one at low emissions hat a negligible impact on overall emission levels, while the deviation at high emissions had a major impact.

The high resolution of the DECSO-HR algorithm of 0.05° (about 5 km in the Netherlands) becomes comparable to the size of cities, municipalities and provinces, which might allow comparison and verification of inventory emissions even on these smaller, local scales. This could potentially be very useful, not only for a better understanding of emission sources, but also because emissions at the local scales have a high relevance for policies.

Here, we will evaluate these possibilities for the use-case of the Netherlands beginning on the scale of provinces (Fig. 7a), of which there exist 12 in the Netherlands, with populations ranging from 390 000 to 3 800 000 and land areas between 1400 and 5400 km². Comparing the emissions per province according to the inventory and according to DECSO-HR 6.5, we see good agreement between the two data sets. A linear fit revealed a slope of 0.89 with R2 of 0.93. That means that a small bias remains, where also on a provincial level the emission inventory reports slightly higher emissions than DECSO-HR 6.5. However, the good correlation supports the possibilities of emission verification by satellite observations also on the regional scale.

Next, we zoom-in closer, looking at the scale of municipalities (in Dutch gemeenten), which can have a size as large as 523 km² and as small as 7 km² with a population of up to 930 000 for the bigger cities but also as small as 1000 (Fig. 7b and c). This analysis reveals even on this finer scale a good correlation, which can visually be better appreciated on a log-log scale (Fig. 7c). Fitting a trendline returned a slope of 0.67 with R2 of 0.84. While most of the data points in general follow the unity line with some added scatter, systematic deviations can be observed especially for municipalities with very large emissions, most prominently Rotterdam, Amsterdam, Velsen and Vlissingen. This can be explained by the deviation between DECSO-HR 6.5 and the inventory at high emissions discussed above (Fig. 6), which subsequently leads to a deviation for municipalities with strong industrial sources. Furthermore, emissions in grid cells that are divided between two municipalities are split between these two municipalities based on area, which effectively spreads out the emissions and leads to an underestimation of highly emitting municipalities. To test the impact of the improved resolution on this analysis, it was repeated after aggregating the DECSO data to a cell size of 0.1°. This led to a much worse correlation between the two datasets with a slope of 0.566 and R2 of 0.75. Taken together, these analyses showed that while the higher resolution did improve the agreement between inventory and DECSO data on a municipal level, there are still some discrepancies on these spatial scales.