The synthetic satellite observations were generated from high-resolution simulations conducted with the COSMO-GHG model. COSMO is a hydrostatic, limited-area model developed by the Consortium for Small-scale Modeling , for which an extension has been developed for the simulation of greenhouse gases (COSMO-GHG).

COSMO-GHG was set up to simulate CO2, CO and NOx concentration fields for nearly the complete year 2015 (1 January–25 December). The model domain extended about 750 km in the east–west and 650 km in the south–north direction. It was centered over the city of Berlin and also covered numerous power plants in Germany and neighboring countries. The spatial resolution was 1.1 km × 1.1 km horizontally with 60 vertical levels up to an altitude of 24 km. Figure  presents the model domain and marks the location of Berlin and the six largest coal-fired power plants. The detailed model setup is described by .

Initial and lateral boundary conditions (ICBCs) for meteorological variables were provided by the operational European COSMO-7 analyses of MeteoSwiss with hourly temporal and 7 km horizontal resolution. For the tracers, ICBCs were obtained from the European Centre for Medium-Range Weather Forecasts (ECMWF) through the European Earth observation program Copernicus. CO2 and CO boundary conditions were taken from a global free-running CO2 simulation with 137 levels and about 15 km horizontal resolution (T1279 spectral resolution, experiment gf39, class rd) . For NO and NO2, boundary conditions were taken from ECMWF's operational global forecasts for aerosol and chemical species with 60 vertical levels and a horizontal resolution of about 60 km (T255 resolution, experiment 0001, class mc) .

Anthropogenic emissions were obtained by combining the third version of the TNO/MACC inventory (, for version 2), which was generated by the Netherlands Organisation for Applied Scientific Research (TNO) for the Monitoring Atmospheric Composition and Climate – Phase III (MACC-3) project, with a detailed inventory provided by the city of Berlin . The inventories provide point and area sources separately for different sectors (e.g., industry, heating and road transport) using Selected Nomenclature for Air Pollution (SNAP) categories. The temporal variability of emissions was accounted for by applying diurnal, weekly and seasonal cycles according to SNAP categories. Furthermore, emissions were vertically distributed using specific vertical profiles for the different emissions categories and plume rise calculations for the six largest power plants and the major point sources in Berlin . Hourly biospheric fluxes of both photosynthesis and respiration were generated with the Vegetation Photosynthesis and Respiration Model (VPRM) at the resolution of the COSMO model .

According to the official inventory of the city of Berlin, total annual CO2 emissions of Berlin were 16.9 Mt CO2 yr-1 (in the reference year 2012 of the inventory). This is about a factor of 2 smaller than in previous studies, e.g., in the LOGOFLUX project , which relied on unrealistically high emissions as provided by the global EDGAR inventory (version 4.1). Due to the diurnal cycle of emissions, emissions were somewhat larger (about 20.0 Mt CO2 yr-1) around the time of the satellite overpasses (10:00–11:00 UTC). Table  summarizes the CO2, NOx and CO emissions of the five largest power plants in the domain.

The simulations included a total of 50 different passively transported tracers representing the three different gases further divided into different sources, release times or release altitudes. This also included background tracers constrained at the lateral boundaries by the global-scale models and two tracers for biospheric respiration and photosynthesis for CO2. Due to the reactivity of NOx, five different NOx tracers with e-folding lifetimes of 2, 4, 12, and 24 h and infinity were included, considering that the lifetime of NOx varies between about 2 and 24 h . The full list of tracers is provided in the SMARTCARB final report p. 15f.

In this study we use the following seven tracers that have been computed from the 50 tracers in the simulations.

X_BER: concentrations from time-varying emissions of Berlin;

The CO2M mission is a proposed constellation of satellites flying in a sun-synchronous low-Earth orbit with Equator crossing times around 11:30 local time. Each satellite will carry an imaging spectrometer measuring in the near-infrared (NIR) and in two shortwave infrared spectral bands (SWIR 1 and SWIR 2) for retrieving CO2 as the main payload. The NIR band is used to retrieve information on the dry air column, on surface pressure, and on aerosols and clouds. The SWIR-1 and SWIR-2 bands contain weak and strong absorption features of CO2 and provide additional information on aerosols and clouds, especially on thin cirrus clouds. A CO2 retrieval using these three bands is described for example by . CO2M is planned to carry also additional instruments for measuring NO2, aerosols and clouds. In an earlier phase, also an instrument measuring CO was considered. The preliminary system concept envisages a pixel size of 4 km2 and a swath width of 250 km or more.

For the CO2, CO and NO2 satellite observations, different instrument scenarios were prescribed by ESA for this study in terms of orbit, spatial resolution, and spatial and temporal coverage of the CO2M instrument. In addition, the Sentinel-5 instrument on board the Meteorological Operational Satellite – Second Generation A (MetOp-SG-A) was studied as an alternative platform for CO and NO2 measurements. Sentinel-5 will be an imaging spectrometer measuring, among others, NO2 and CO columns with a spatial resolution of 7 km× 7 km and a 2650 km swath. MetOp-SG-A will be also on a sun-synchronous orbit but with different Equator crossing times and repeat cycles than the CO2M mission.

In addition to a single satellite, the potential of a constellation of multiple CO2M satellites was also studied. The basic assumption for a constellation is that the individual satellites are spaced with equal angular distance in the same orbit with the same orbit parameters, for instance separated by 180, 120 and 90∘ on a full circle in the case of 2, 3 and 4 satellites. The individual satellites can be distinguished by their starting longitude at the Equator of the first orbit in the repeat cycle. Here, we analyze constellations between one and six satellites.

For the computation of orbits, we adopted the orbit simulator of the Netherlands Institute for Space Research (SRON). Since this simulator makes a few simplifying assumptions, such as circular orbits and tiled ground pixels, satellite and instrument parameters were slightly modified to preserve essential parameters. In particular, orbit periods were calculated to match a given cycle duration and length. The period then determines the altitude and inclination of a circular, sun-synchronous orbit. The altitude of the circular orbits is slightly larger than the typically used mean altitude for elliptic orbits. Since the altitude affects the size of the ground pixels and the width of the swath, field of view and along-track sampling time were set to match exactly the prescribed pixel size at subsatellite point as well as the prescribed swath width. As a result, the number of across-track pixels for the simulated Sentinel-5 did not match exactly the number of pixels for the real Sentinel-5 instrument.

Tables  and summarize the orbits and viewing geometries of the two satellites. The CO2M satellite is assumed to have a 250 km wide swath and an 11 d repeat cycle. Within the 11 d cycle, the instrument provides nearly global spatial coverage (Fig. ). For locations in the SMARTCARB model domain, either one or two overpasses occur during the 11 d repeat cycle depending on the Equator starting longitude of the satellite. The wider swath of Sentinel-5 results in near-daily global coverage (not shown).

XCO2, CO and NO2 column densities were sampled along the satellite swath for 1 year using the tracers from the COSMO-GHG simulations. For CO2M, XCO2, CO and NO2 columns were mapped onto the 2 km × 2 km size pixels along the 250 km wide swath.

For Sentinel-5, CO and NO2 columns were sampled with up to 7.5 km × 7.5 km resolution along the 2670 km wide swath of the Sentinel-5 instrument. Due to the wide swath, the pixel sizes grow towards the edge of the swath. In this study, only the spatial overlap between Sentinel-5 and CO2M were of interest because Sentinel-5 was used for detecting the CO2 emission plumes inside the swath of CO2M.

The error characteristics of the CO2, NO2 and CO instruments were specified in collaboration with ESA based on previous studies for CarbonSat and on performance requirements for Sentinel-5 . For the instruments on the CO2M satellites, two or three scenarios were included in order to cover a realistic range between more or less demanding instruments. Table  summarizes the single sounding precision of the different instruments.

For XCO2, three different uncertainty scenarios were considered which relate back to the performance estimates derived for the CarbonSat mission concept. CarbonSat was a CO2 imaging spectrometer proposed for ESA's eighth Earth Explorer mission, with specifications similar to those of the CO2M mission. A detailed error budget for CarbonSat was presented in the CarbonSat report for mission selection . In the LOGOFLUX study, error parameterization formulas (EPFs) for random and systematic errors were developed, which account for errors introduced by solar zenith angle (SZA), surface reflectance in the near-infrared (NIR) and shortwave infrared (SWIR 1), cirrus clouds and aerosol optical depth .

Here, the same EPFs were adopted but only applied to compute random errors. These were calculated based on SZA and surface reflectance in the NIR and SWIR-1 band. Surface reflectances were taken from the MODIS MCD43A3 product (version 006) at 1 km spatial resolution . A detailed consideration of cirrus clouds and aerosols and their impact on systematic errors was outside the scope of the study as it would have required the collection and processing of a large amount of additional data. The possible impact of not considering systematic errors will briefly be discussed in Sect. .

The random error calculated with the EPFs for the so-called vegetation-50 scenario (VEG50, i.e., vegetation albedo and SZA of 50∘) is about 1.5 ppm. In the model domain, mean random errors are slightly smaller at 1.3 ppm. To obtain random errors for the three instrument scenarios with σVEG50 of 0.5, 0.7 and 1.0 ppm, the computed errors were divided by 3.0, 2.14 and 1.5, respectively.

For NO2 VCDs, the overall uncertainties are due to (a) measurement noise and spectral fitting affecting the slant column densities; (b) uncertainties related to the separation of the stratospheric and tropospheric column; and (c) uncertainties in the auxiliary parameters used for air mass factor (AMF) calculations such as clouds, surface reflectance, a priori profile shapes and aerosols . The total uncertainties are dominated by uncertainties from spectral fitting for background pixels and by uncertainties in AMF calculations for polluted pixels, respectively. Typical spectral fitting uncertainties of previous instruments such as the Ozone Monitoring Instrument (OMI) were of the order of 1–2 × 1015 molecules cm-2 and AMF uncertainties of the order of 15 %–20 %. These ranges were used to define two different scenarios for a possible CO2M NO2 instrument (Table ). For the Sentinel-5 UVNS instrument, we assumed a relative uncertainty of 20 % and a minimum uncertainty of 1.3 × 1015 molecules cm-2. In the presence of clouds, the reference noise was increased using the empirical formula developed by . For a cloud fraction of 30 %, random noise is approximately doubled.

For CO VCDs, the total uncertainty depends on the (a) fitting noise; (b) a priori CO and CH4 profiles; and (c) surface reflectance, aerosols and clouds. We assumed a single sounding precision of 4.0 × 1017 molecules cm-2 and a relative precision of 10 % and 20 % for both Sentinel-5 and the CO2M mission.

Satellite observations require filtering for clouds, which significantly reduces the number of observations available for plume detection. For the CO2 product, we removed all CO2 pixels with cloud fractions larger than 1 % because the CO2 requires rigorous cloud filtering . The NO2 retrieval can tolerate larger errors and is therefore less sensitive to clouds. For the NO2 product, we used a cloud threshold of 30 % as often applied in satellite NO2 studies e.g.,. For CO, a cloud threshold of 5 % was used, which is motivated by the cloud threshold used for the MOPITT CO product .

Previous studies used the MODIS cloud mask product available at 1 km resolution for masking cloudy CO2 observations . Since CO and NO2 observations can tolerate larger cloud fractions, a cloud fraction product would be needed for masking pixels with different thresholds, but the MODIS cloud product is only available at 5 km resolution . Therefore, we used total cloud fractions computed by COSMO-GHG, i.e., the same model as used for the tracer transport simulations, that are available at model resolution. COSMO-GHG computes total cloud fraction from layer cloud fractions assuming minimum overlap. The differences between cloud masks derived from COSMO-GHG and MODIS products and their effects on data yield are discussed in Sect. .

We developed a new but simple plume detection algorithm that uses a statistical test to detect signal enhancements which are significant with respect to instrument noise and variability in background levels. The plume is then identified as a coherent structure of significant pixels. The algorithm involves three processing steps as laid out in Fig. .

The first step of the plume detection algorithm finds satellite pixels with CO2, CO or NO2 values significantly larger than the background field using a statistical Z test, for which the distribution of the test statistics can be approximated by a normal distribution e.g.,. The Z test computes a z value given by 1z=x-μσ, where x is an observation from a population with mean value μ and standard deviation σ. The value x is considered significantly larger than μ when the z value is greater that or equal to a critical value. The critical value is calculated from the inverse cumulative distribution function of the normal distribution for a probability q. The probability that x is not significantly larger than μ is the p value (p=1-q).

A key feature of the algorithm is that the trace gas observation of a single pixel is replaced by a spatial average of pixels in a defined neighborhood. This allows identifying weak plumes with signals of individual pixels well below instrument noise but also bares the risk of diluting the signal at the plume edges. Figure  shows examples of neighborhoods with different sizes ns that have been used for testing the algorithm. A large neighborhood results in a stronger smoothing and may therefore produce a larger number of false positives, i.e., pixels outside of the plume that are wrongly assigned to the plume. An ideal neighborhood size balances the need for sensitive plume detection with the requirement of a low fraction of false positives.

Whether a signal enhancement is detectable is primarily determined by the z value (Eq. ), which can also be interpreted as the signal-to-noise ratio (SNR) at a spatially smoothed satellite pixel. Thereby, the signal is the enhancement above background due to the plume, and the noise is composed of both instrument noise and spatial variability in the background. Since the background, i.e., the trace gas field in the absence of the plume, cannot be directly observed, it needs to be estimated either from the observable trace gas field surrounding the plume or from a climatology or a model. Instrument noise and background variability can have both random and systematic components. The random component would reduce with the inverse square root of the number of valid pixels n in the neighborhood, whereas the systematic error would be approximately independent of n. The number of valid pixels n can be smaller than the size of the neighborhood ns when pixels are missing, e.g., due to clouds or at the boundary of the satellite swath. Therefore, the z value or the SNR can be calculated as follows: 2SNR=Xobs-Xbgσrand2n+σsys2, where Xobs represents the spatially averaged satellite observations, Xbg is the estimated background value, and σrand and σsys are the random and systematic errors, respectively. Equation () can be calculated for each satellite pixel and compared with the critical value to determine which Xobs's are significantly larger than Xbg.

To compute the SNRs from the satellite image, we computed observed values Xobs as the local means of a neighborhood of size ns (Fig. ). For sake of simplicity, the background Xbg in this study was estimated from the X_BER_BG and X_PP_BG model tracers for a 200 km× 200 km square centered on the city of Berlin and a 100 km× 100 km square centered on each power plants. The implications of this assumption are discussed in Sect. .

For random and systematic errors we assumed that the instrument uncertainty is purely random and that the background variability is purely systematic. For the instrument uncertainty, random errors as listed in Table  were used, which reduce with the number of valid pixels in the local neighborhood n due to the inverse scaling by the number of pixels. The background uncertainty σbg2 was computed from the spatial variance of the background model tracer in a fixed domain surrounding the source as described above. Since it is assumed to be systematic, it does not decrease with the size of the neighborhood.

The result of the Z test is a binary image with “true” values where pixels are significantly enhanced above the background and “false” values where they are not (Fig. b). Since the local mean can still be computed for missing center pixels – using neighboring pixels – missing pixels can also be detected as enhanced above the background.

In the second step, pixels that are enhanced (true) and connected are assumed to belong to the same plume. We label regions of connected pixels using a standard labeling algorithm. Neighboring pixels are identified using a Moore neighborhood, where each pixel has eight potential neighbors. Each region is assigned a unique integer value (Fig. c).

Finally, in the third step, all connected regions that do not intersect with the source region are removed, leaving only regions that overlap with the source of the plume (Fig. d). For cities, the source region is defined by a circle with a radius of 15 km and for point sources by a circle with a radius of 5 km. The last step may remove regions that are part of the real plume but separated from the source by weak signals or missing values (e.g., the region labeled “2” in Fig. c).

The plume detection algorithm was applied to the CO2 plumes of Berlin and Jänschwalde. The detectability of CO2 plumes was evaluated for the different instrument scenarios by comparing the detected plume with the “true plume” defined by the field of the CO2 tracer released by the respective source (CO2_BER in the case of Berlin, CO2_PP for Jänschwalde) above a low threshold. For Jänschwalde, the performance had to be additionally evaluated by visual inspection due to frequent overlaps with the plumes of other power plants also contained in the CO2_PP tracer.

To evaluate the performance of Sentinel-5 with respect to detecting plumes as observed by CO2M, detected pixels had to be projected onto the pixels of the CO2M instrument. Sentinel-5 pixels were thus only used over the swath of the CO2M satellite rather than over the whole swath of Sentinel-5.

When using NO2 or CO for plume detection, the performance was assessed by comparing the detected pixels with the true CO2 plume rather than the true plume of the auxiliary gas. In this way, the degree of congruence between the CO2 and auxiliary trace gas plumes was considered as well.

For the evaluation, we computed true positives (TPs), false positives (FPs) and the positive predictive value (PPV=TP/(TP+FP)) . A good algorithm should have a much smaller number of FPs than TPs and, therefore, a PPV close to 1. To remove the impact of different cloud thresholds, TPs, FPs and PPVs were only computed for cloud-free pixels using a cloud threshold of 1 % because we are primarily interested in valid CO2 observations that can be used for estimating CO2 emissions.

Figure  shows the CO2 and NO2 plumes of Berlin on 21 April 2015 observed by CO2M and Sentinel-5 for different instrument scenarios. The outlines of the real plumes are overlaid as solid and dashed lines for CO2 and NO2, respectively. Since the CO2 instruments have a lower cloud threshold, a band of cirrus clouds is obscuring the plume in the CO2 observations but not in NO2. Successfully detected pixels are shown as black crosses, and the number of detected pixels (median of all 20 noise realizations) are presented in the legend. On average, a CO2 instrument detects 116±46, 48±40 and 24±30 pixels with noise scenarios σVEG50 of 0.5, 0.7 and 1.0 ppm, respectively (Fig. a and b). The number of true positives is slightly smaller, having on average 2 false-positive pixels. Consequently, the PPV is high, ranging between 0.85 and 0.99 for high- and low-noise, respectively.

For the NO2 measurement the band of thin cirrus clouds is not an issue. The NO2 instrument can therefore detect a much larger number of pixels, i.e., 1242±99 and 1203±155 in the case of the low-noise and the high-noise scenarios, respectively (Fig. c). On average, the fraction of FPs is relatively large, and the PPV is only 0.80±0.05 and 0.77±0.10 for the low- and high-noise scenarios, respectively. The small PPV is caused by interference with the plume of Jänschwalde, which is just south of the plume of Berlin. For cases where no neighboring plumes have been detected falsely with the NO2 instrument, the spatial match between CO2 and NO2 plumes is generally high, suggesting a high degree of spatial overlap between the CO2 and NO2 plumes.

The Sentinel-5 NO2 instrument is also able to detect the CO2 plume with 879±114 CO2M pixels, but since it is measured 2 h earlier, the NO2 plume seen by Sentinel-5 (dashed line in Fig. d) is slightly shifted with respect to the CO2 plume (solid line). As a consequence, the PPVs is low (0.60±0.04).

Figure  presents two examples where the CO2 instrument fails to detect the CO2 plume. In Fig. a, the CO2 field has a pronounced spatial gradient resulting in a high variance of the background. This gradient is not present in the much-shorter-lived trace gas NO2, making it possible to detect the plume using an NO2 instrument (Fig. b). Similar situations occur in roughly 20 % of cloud-free swaths. Figure c and d show a second example where the CO2 instrument cannot detect the plume because the signal is very weak due to strong winds. Owing to its better SNR, the NO2 instrument is able to detect the plume also in this situation.

Figure  presents two examples comparing the NO2 plume observed by Sentinel-5 to the CO2 plume observed 2 h later by the CO2M satellite. In the first example (panels a and b), Sentinel-5 fails to detect any plume due to clouds, which have largely disappeared by the time of the CO2M overpass. In the second example, both Sentinel-5 and the CO2M satellite detect a plume of similar size, but the Sentinel-5 plume is significantly displaced due to changes in the prevailing winds between the two overpasses.

To count the number of plumes detectable under the different instrument scenarios, we analyze the 50 plumes observed by a constellation of six satellites, which we had classified in Sect.  as being potentially useful based on the idealized tracer XBER having more than 100 pixels above a threshold of 0.05 ppm. Table  summarizes the results in terms of number and size of the detected plumes. A plume was only counted as detected when at least 100 CO2 pixels were correctly detected (true positives) and when at least 80 % of the detected pixels were true positives (PPV ≥ 0.80). The PPV threshold was found useful for removing plumes interfering with others or plumes shifted due to the earlier overpass time of Sentinel-5.

Table  shows that the CO2 instruments detect significantly fewer plumes than the NO2 instruments. Depending on the instrument noise scenario, the CO2 instruments detect plumes with more than 100 pixels with a success rate of only 12 % to 32 %, while for the NO2 instruments the success rates are 68 % to 70 %. Surprisingly, the NO2 instrument with low noise performs slightly worse than the high-noise instrument. This is an artifact of the algorithm often detecting small plumes not related to emissions from Berlin in the case of a low-noise instrument. The Sentinel-5 NO2 instrument detects 20 % of the plumes, thus only half the success rate of the NO2 instrument on the CO2M satellite. The main reason for this low success rate is the spatial mismatch of the plumes due to the 2 h difference in overpass times.

Figure  shows the number of plumes per month with at least 100 detected pixels for different constellations between one and six satellites for the CO2 low-noise and the NO2 high-noise scenario. The number of plume detections per month is small and therefore highly sensitive to the specific orbit configuration. For example, two satellites seem to detect more plumes than three, but this result is caused by an unfavorable orbit for observing Berlin for the constellation with three satellites and unfavorable cloud cover as already discussed earlier. The standard deviation was estimated from the number of detectable plumes using satellites with different starting longitudes (i.e., east–west displacements of all orbits). The figure shows that the number of observed plumes generally increases with the number of satellites as expected, but statistical noise can mask the increase from one constellation to the next. The figure confirms the much lower success rates of the CO2 instruments as compared to the NO2 instruments as expected from the computed signal-to-noise ratios.

There are six major power plants in the model domain: Jänschwalde, Boxberg, Schwarze Pumpe, Lippendorf, Turów and Pątnów. Because no model tracer was defined for individual power plants but instead only for the sum of all of them, the true plume of an individual power plant is not known. Therefore, we applied a visual inspection to identify those plume detections which erroneously included neighboring plumes. Furthermore, we limit the analysis to Jänschwalde.

As an example, Fig.  shows the successful detection of the CO2 plume of Jänschwalde on 2 November 2015 by different instruments. Since CO2 emissions of Jänschwalde are high with 33.3 Mt CO2 yr-1, the XCO2 signal is very strong and can be detected well even with a high-noise instrument (σVEG50=1.0 ppm). With low noise (σVEG50=0.5 ppm) the weaker plumes of Schwarze Pumpe and Boxberg are visible as well. The NO2 instrument detects the four plumes in the region well. On this day also the Sentinel-5 NO2 instrument successfully detects the plume of Jänschwalde and other point sources. Figure  presents a second, more challenging example for 17 February 2015. The CO2 instrument successfully detects the plume with 0.5 ppm uncertainty, but with 1.0 ppm uncertainty, the number of detected pixels is likely too small to be useful for emission estimation. The reason for the low number of detected pixels in this case is the strong horizontal gradient in the CO2 background. The NO2 instrument detects the plume, but because the NO2 plume of Jänschwalde overlaps with neighboring plumes, these plumes are erroneously assigned to Jänschwalde as well. At the coarser resolution of Sentinel-5 the plumes of the individual power plants can hardly be separated and, moreover, the time difference of 2 h results in a plume location that is shifted with respect to the plume seen by the CO2M satellite.

Table  summarizes the results of the plume detection for Jänschwalde under the different instrument scenarios. It shows the number of detected plumes with at least 10 pixels and, in addition, the number of plumes with at least 10 valid CO2 pixels (cloud cover <1 %). Note that, in the case of the much narrower plumes from power plants, fewer pixels are required to form a useful plume. We classified detections that include large parts of the background as failed but still counted detections that include neighboring plumes as successful (e.g., Fig. c) because they successfully identified the location of the plume. Since the classification is not always unambiguous, we assigned an uncertainty of about ±5 plumes at most.

In the year 2015, the number of detectable plumes with more than 10 pixels for a constellation of six satellites was between 40 and 45 for a CO2 instrument with σVEG50 of 0.5, 0.7 and 1.0 ppm. At the same time, the NO2 instrument detected about 90 plumes for the low- and high-noise scenario. When only plumes with more than 10 cloud-free CO2 observations were considered, the number was reduced to about 70 plumes. For a smaller number of satellites, the number of detectable plumes would be correspondingly smaller. The NO2 instrument detects more plumes because of its lower sensitivity to clouds, which makes it possible to trace the plume to the source even for partly cloudy scenes. On the other hand, the NO2 instrument often detects overlapping plumes (e.g., Boxberg and Schwarze Pumpe) because the instrument is much more sensitive to small signals further away from the origin than the CO2 instrument. The mean plume size was about 100 pixels for the low-noise CO2 instrument. The plumes detected with the high-noise CO2 instrument were about half the size. The NO2 instruments detected a similar number of CO2 pixels to the low-noise CO2 instrument, but when all detected pixels are counted the number of pixels doubles.