During the TRACER-AQ campaign, a total of seven Pandora instruments operated across three sites in Houston (Table 1). Pandora instruments are ground-based UV–VIS spectrometers that measure spectrally resolved radiances, and this work only utilizes those collected via direct-sun observations (Herman et al., 2009). Trace gas spectral fitting routines are employed similarly to remote sensing and aircraft observations (Judd et al., 2020) to characterize column concentrations of gases (e.g., NO2). Details on the Pandora instruments and their fitting routines are discussed in detail in past studies (Cede, 2021; Herman et al., 2009). The study was designed to have two Pandoras operating coincidently at each site during the campaign; however, due to instrument failures, an uneven number of observations were obtained at each site. In order to evenly weigh the observations between the three sites, we selected data from a single Pandora instrument at each site. Pan no. 58 at La Porte, no. 61 at Aldine, and no. 25 at the University of Houston were chosen for the following reasons: as indicated in Table 1, Pan no. 61 and Pan no. 58 clearly have the largest temporal coverage during the TRACER-AQ time period. While Pan no. 188 measured more frequently at the University of Houston than Pan no. 25, Pan no. 188 was operated on a tower about 70 m above the surface, which results in missing portions of the tropospheric column when operated in direct-sun mode.

Locations of the three sites are presented in Fig. 2f. These three chosen instruments are bolded in Table 1​​​​​​​. Pandora direct-sun retrievals represent the “total vertical column” of NO2 which differs from the aircraft measurements that only measure the tropospheric column. We directly compare these disparate sources by adding a “stratospheric NO2 column component” derived from TROPOMI estimates to the aircraft measurements (see Sects. 2.2 and 2.3) to compare total column amounts.

The GCAS instrument was installed on the NASA G-V aircraft. The GCAS instrument employs charge-coupled device array detectors to observe backscattered light. These data can be used to retrieve column densities of gases like NO2 below the aircraft using DOAS computing software (Danckaert et al., 2017). During TRACER-AQ, GCAS collected data over the Houston metropolitan area across 12 d during late August and throughout September 2021. The flight strategy of the aircraft included flying the plane in a “lawnmower” fashion with flight lines spaced 6.3 km apart, ensuring overlap at flight altitude (FL280) with the instrument field of view of 45° creating one gapless map of NO2 up to three times per flight day with an average differential slant column pixel size of 250 m × 250 m. NO2 observations from GCAS are publicly available at the NASA Atmospheric Sciences Data Center (NASA/LARC/SD/ASDC, 2022a). Observations from 2 of the flight days – a test flight (30 August) and a flight over the Gulf of Mexico (27 September) – are excluded from this study because they provided no meaningful data over Houston. Given the relatively short time frame of flight data collection, meteorological conditions have an influence on the fine-scale patterns in NO2 columns observations. Owing to this, we summarize some basic conditions and information of the 10 flight days that focused on Houston (Table 2). Wind and meteorological conditions were determined by review of historical weather archives taken at Houston Hobby Airport (EOSDIS Worldview, 2023; Weather Underground, 2023).

The publicly available GCAS measurements (version R2) include a version of the dataset with reprocessed AMFs to include NO2 vertical profile estimates from the fine-scale (444 × 444 m2) WRF–CAMx simulation used in this analysis (Sect. 2.4). Air mass factors use this vertical profile information to account for altitude-dependent sensitivities in remote sensing observations. The original vertical profiles in the dataset were derived from a global model, GEOS-CF (Keller et al., 2021), which had a coarser spatial resolution (0.25° × 0.25°). Lastly, to directly compare GCAS measurements to other NO2 column concentrations, we regrid them to a common grid; in this study, we chose the fine-scale WRF–CAMx grid. Only cloud-free GCAS data are considered in this analysis.

To characterize the accuracy and precision of GCAS measurements we compare them to observations from the Pandora instruments (Sect. 3.1). This comparison requires both spatial and temporal screening. Spatially, we restrict our comparison to only the GCAS pixels that contain Pandora instruments. Temporally, we screen out all Pandora measurements that are more than 15 min removed from a GCAS overpass and then identify the Pandora measurement time within this 30 min window that most closely matches the GCAS overpass time. While we choose this 30 min window as an upper-bound cut-off, 96 % and 90 % of all Pandora closest matches occur within a 20 and 15 min window of GCAS overpasses, respectively, indicating that this choice of window will have a minimal impact on our results. After screening the data, we also account for the fact that GCAS only measures the tropospheric component of the NO2 column. There is a substantial but predictable “above-aircraft” column that is not reflected in the GCAS measurements. This is primarily associated with stratospheric NO2. To account for this, we approximate the above-aircraft component of the GCAS NO2 columns using the stratospheric NO2 column component of TROPOMI measurements (Sect. 2.3) and add this to GCAS observations. Additionally, we add an “above-aircraft” but below-troposphere partial column amount based on the CAMx simulation, and then we calculate the column. The three highest levels of CAMx amount to 0.57 × 1015 molec. cm-2, and we add this amount to GCAS.

The TROPOMI instrument, on board the Sentinel-5P satellite, has measured total slant columns of NO2 daily at approximately 13:30 local time (LT) globally from 30 April 2018 to the present (Copernicus Sentinel-5P, 2021). The slant column measurements were converted into tropospheric vertical column amounts by subtracting off a stratospheric NO2 component and transforming the remaining tropospheric slant column to vertical column using an air mass factor. We downloaded the publicly available data​​​​​​​ (https://dataspace.copernicus.eu/explore-data/data-collections/sentinel-data/sentinel-5p​​​​​​​, last access: May 2023; https://dataspace.copernicus.eu/, last access: September 2024) coincident with the TRACER-AQ campaign in September 2021 for overpasses of Houston. In this study, we primarily consider measurements from the latest version (2.4.0) (Eskes et al., 2023); however, we additionally consider measurements processed using the version 2.3.1 algorithm (van Geffen et al., 2022) and the NASA Multi-Decadal Nitrogen Dioxide and Derived Products from Satellites (MINDS) product (Lamsal et al., 2022) and intercompare these different versions (Fig. 3). All product versions stem from the same slant column retrieval but differ in the calculation of the air mass factor for slant to vertical column conversions and, in the case of NASA MINDS, separation of the stratosphere and troposphere (Bucsela et al., 2013). The main difference between versions 2.3.1 and 2.4.0 is the use of the 0.125° × 0.125° directional Lambertian equivalent reflectivity (DLER) climatology derived from TROPOMI observations which replaces an old 0.5° × 0.5° Lambertian equivalent reflectivity (LER) dataset used in v2.3.1 (Eskes et al., 2023). NASA MINDS uses a geometry-dependent surface Lambertian equivalent reflectivity (GLER) product for its surface reflectivity input into the AMF calculation based on MODIS observations. The other main difference in these products includes use of different a priori NO2 profiles (1° × 1° TM5-MP for v2.3.1 and v2.4.0 vs. 0.25° × 0.25° GMI simulation for NASA MINDS). A comparison between TROPOMI version 2.4.0 and a MAX-DOAS network found that in moderately polluted locations TROPOMI had a median bias of -35 %. A comparison between TROPOMI version 2.4.0 and PGN found a median bias of -18 % over polluted stations (Lambert et al., 2023).

These publicly available TROPOMI data are further processed for this study. We screen TROPOMI measurements to consider cloud coverage and erroneous data using the recommended qa_value filter (> 0.75). We regrid the TROPOMI NO2 observations (resolution of 3.5 × 5.5 km2 at nadir) onto the WRF–CAMx grid (444 × 444 m2). When comparing TROPOMI observations to Pandora instruments we follow the same spatial and temporal screening approach as discussed for GCAS. Spatially, we identify the CAMx grid cell in which each Pandora instrument is located and only consider TROPOMI measurements that were regridded to these grid cells. We intercompare GCAS, TROPOMI, and CAMx at this resolution but also compare the three datasets at a coarser resolution (Sect. 3.4) to account for resolution-dependent errors. Temporally, we screen out all Pandora measurements that are more than 15 min removed from a TROPOMI overpass time and then identify the Pandora measurement time within this 30 min window that most closely matches the TROPOMI overpass time. While we choose this 30 min window as an upper-bound cut-off, 100 % and 97 % of all Pandora closest matches occur within a 20 and 15 min window of TROPOMI overpasses, respectively, indicating that this choice of window will have little impact on our results. Using WRF–CAMx vertical profile information we calculate both a total and tropospheric NO2 column from TROPOMI v2.4.0 measurements using new AMF derived from the WRF simulation, and we difference the total and tropospheric values to calculate a stratospheric NO2 column component from TROPOMI. We take the spatial and temporal average of this stratospheric component in Houston during the TRACER-AQ campaign to calculate a constant bias correction to convert tropospheric NO2 columns – from GCAS and WRF–CAMx – to quasi-total NO2 columns when comparing them to total NO2 column measurements from Pandora instruments. This stratospheric vertical column NO2 amount of 3.0 × 1015 is typical for Houston during summer (Geddes et al., 2018). Boersma et al. (2018) suggest that 0.5 × 1015 molec. cm-2 is the upper limit of structural uncertainty in the stratospheric estimate. This uncertainty should be considered when reviewing results that compare total column amounts (i.e., results comparing GCAS and CAMx to Pandora). We additionally account for diurnal variation in the stratospheric column by applying the results from work by Li et al. (2021); they calculate a daytime stratospheric NO2 column increase rate of 1.34 × 1014 molec. cm-2 starting at 07:00 LT. We apply this increase rate by calculating the difference in hours between the dataset times – either the GCAS overpass times or CAMx simulation hours – and 13:30 LT,​​​​​​​ i.e., the approximate TROPOMI overpass time, and then multiply this difference by the increase rate. In doing so, total column values before the TROPOMI overpass are decreased and total column values after the overpass are increased.

For this study, a set of simulations were conducted employing version 4.3.3 of the Advanced Research Weather Research and Forecasting (WRF) model (Skamarock et al., 2021) jointly with the Comprehensive Air Quality Model with Extensions (CAMx) v7.20 with the CB6r5 chemical mechanism for a simulation period that matched the September 2021 TRACER-AQ time frame. A new high-resolution modeling platform was designed specifically for this study that adopted prior approaches used in Texas Commission on Environmental Quality (TCEQ) state implementation plan (SIP) modeling (TCEQ, 2021) to update emissions.

The WRF model is a mesoscale numerical weather prediction system designed to serve both operational forecasting and atmospheric research needs (Skamarock et al., 2005, 2008). We define the WRF modeling domains as slightly larger than the corresponding CAMx domains (Fig. 1) to avoid possible numerical artifacts near domain boundaries when transferring WRF meteorology to CAMx. The 36 km CAMx domain (red) includes the continental United States, Mexico, and parts of Central America and Canada. The 36 km, 12 km (blue), and East Texas 4 km (green) domains are also used by the TCEQ for State Implementation Plan (SIP) modeling. The higher-resolution domains (1.333 km (orange) and 0.444 km (cyan)) were selected to include the most relevant GCAS flight tracks while considering computational expense.

Additional information on the WRF–CAMx modeling is included in the Supplement including the WRF physics options (Table S1 in the Supplement), vertical layer mapping from WRF to CAMx (Table S2), and CAMx science options (Table S3). We used 0.25° Global Forecasting System (GFS) data assimilation system (GDAS) analysis data (DOC/NOAA/NWS/NCEP/EMC, 2023) as initial conditions for the WRF meteorological model; these GDAS data are also used for boundary conditions and data assimilation. We configured the output time steps of WRF to 15 min for the higher-resolution domains. Conducting WRF simulations at fine spatial resolutions (i.e., 4, 1.333, and 0.444 km) requires careful consideration of physical schemes that are sensitive to grid spacing. We turn off the convective cumulus parameterization scheme for the fine grids because WRF can explicitly simulate convection for them. For coarser grids we turn on the cumulus parameterization to account for sub-grid-scale convection. The other physics options (Table S1) are kept consistent across the different resolutions. The CAMx simulation was first performed over the coarser domains (36, 12, and 4 km) from which initial and boundary conditions were extracted for the higher-resolution domains. TCEQ developed the 2019 modeling emissions inventory for the Dallas–Fort Worth (DFW) and Houston–Galveston–Brazoria (HGB) attainment demonstration (AD) SIP revisions (TCEQ, 2021). Starting with this inventory we implement further changes as discussed in the next paragraph.

First, we update the CAMx modeling emissions inventory from the TCEQ platform to incorporate 2021 hourly continuous emissions monitoring systems (CEMS) (EPA, 2023) data for the 11 major electric generating units (EGUs) listed in Table S4. We download hourly data from Clean Air Markets Program Data (CAMPD) for the 11 EGUs for the 30 August to 27 September 2021 period, and stack parameters were based on the TCEQ 2019 emissions platform (TCEQ, 2021). Second, we update shipping emissions to incorporate MARINER v2 (Ramboll, 2022a) emissions built with 2021 automatic identification system (AIS) data for the higher-resolution domains. Third, we reprocess link-based on-road mobile emissions for the higher-resolution domains. Fourth, we update biogenic emissions and lightning NOx (LNOx) based on WRF meteorology. Specifically, we use the Model of Emissions of Gases and Aerosols from Nature (MEGAN) version 3.2 (Guenther et al., 2012) for biogenic emissions, the Fire Inventory of NCAR (FINN) version 2.2 (Wiedinmyer et al., 2011) for fire emissions, and lightning NOx emissions derived by applying the CAMx LNOx processor to the 2021 meteorological data from the WRF simulation. Considering that wildfires in the Houston area are rare and that LNOx emissions are associated with convective clouds that obscure remote sensing column observations, we excluded these two emission sources from the finer resolution domains (the 1.333 and 0.444 km domains) but included them in the larger domains. These two sources represent a small fraction of emissions in the local Houston area that are the primary focus of the finer resolution simulations. Lastly, we regrid all other gridded emissions from the coarser domains to the high-resolution domains without refining their spatial resolution. Specifically, all point sources are geolocated to the grid cell containing the source. On-road mobile source emissions and shipping emissions were provided for individual links which we allocated to 444 m grid cells and are based on known roadway networks, ship tracks, and traffic patterns. Airport and railyard emissions were allocated to 444 m grid cells within the property boundary. Other sources retained the 4 km grid resolution provided by the TCEQ. Daily emissions of NOx and volatile organic compounds (VOCs) in metric tonnes per day (t d-1) for a September weekday are presented in Table 3.

We evaluate the WRF simulation meteorology by comparing surface-level wind speed, direction, temperature, and water vapor mixing ratio to observations from 16 ground-level monitors (Figs. S2–S11; Tables S5–S8) and calculate the mean bias error (MBE), mean absolute error (MAE), and Pearson R-squared (R2) statistics as defined in Table S2. Circular statistics are calculated using the Astropy circular statistics package for Python (The Astropy Collaboration et al., 2022). We obtain integrated surface data from NCDC in the DS3505 format (10.7289/V5M32STR, Vose et al., 2014). These data consist mainly of airport locations and have good meteorological siting and quality assurance procedures.

Generally, meteorological conditions simulated by WRF agree with ground-level observations especially on the more data-rich non-cloudy days, which are the most important for our intercomparison; however, performance depends on the specific measure of meteorology considered. Across all days, the WRF wind direction was well correlated (R2 = 0.76) and had minimal bias (MBE = 8°) but some unsystematic errors (MAE = 26°) compared with observations. This indicates that the model generally captures variability in wind direction without a notable bias; however, considering any individual observation the simulated direction may differ by 20–30°. For non-cloudy days – which are more relevant for our intercomparisons due to more data – correlation for wind direction was similar (R2 = 0.73), and the bias and error were reduced (MBE = -5° and MAE = 21°). Simulations of wind speed were more poorly correlated (R2 = 0.26) and had some unsystematic error (MAE = 1.20 m s-1); however, there was very little systematic bias in the wind speed simulation (MBE = -0.02 m s-1). Correlation and unsystematic errors improve on the non-cloudy days (R2 = 0.37 and MAE = 1.08), while there is still no notable systematic bias (MBE = -0.13 m s-1). Considering wind speeds at 09:00 and 13:00 LT​​​​​​​ (Figs. S2–S11), it appears that observations in the afternoon degrade correlation compared with the morning and that, generally, simulated wind speeds are better correlated with observations in downtown Houston than in the southeastern part of the domain near Galveston Bay. Comparisons between GCAS observations and WRF–CAMx simulations show that the model represents the dominant direction and dispersion of identifiable plumes from known sources. The wind speed bias is sufficiently low that model uncertainty will not lead to systematic errors in plume advection. Across the 8 non-cloudy days, hourly and site-specific – across the 16 monitors – WRF wind direction (R2 = 0.3–0.8; MAE = 14–32°), wind speed (R2 = 0.1–0.5; MAE = 0.94–1.35 m s-1), temperature (R2 = 0.69–0.81; MAE = 0.93–1.18 K), and water vapor mixing ratio (R2 = 0.28–0.78; MAE = 0.87–3.11 g kg-1) performed moderately compared with observations given the fine spatial and temporal resolution. Additionally, we compared simulated hourly NO2 (Fig. S12 in the Supplement) and maximum daily 8 h average or “MDA8” O3 (Fig. S13) to observations from 17 TCEQ continuous air monitoring stations (CAMS) operating in Houston. We found poor performance and a strong negative bias in the simulated surface-level NO2 (normalized mean bias (NMB) = -59 %), while simulated surface-level MDA8 O3 had a much weaker bias (NMB = -2.5 %) compared with observations. Comparisons to ozonesondes (Figs. S14–S18) suggest that WRF simulates more aggressive vertical mixing than what is observed. This is consistent with our findings of a stronger negative bias at the surface level than for the columns, as emitted NO2 at the surface is advected vertically quicker in WRF–CAMx than in reality.

We further intercompare these data by grouping them at locations and then calculating their average diurnal profiles during the TRACER-AQ campaign for both column- and surface-level NO2. Specifically we compare GCAS, CAMx, Pandora, and GEOS-CF (Keller et al., 2021) NO2 columns at the three Pandora sites during TRACER-AQ flight days. We include NO2 data from GEOS-CF – that will be used for processing NO2 remote sensing observations from the NASA Tropospheric Emissions: Monitoring of Pollution (TEMPO) mission – to characterize differences between a global simulation and our regional WRF–CAMx modeling. Simulated surface and NO2 columns from GEOS-CF are obtained through the GMAO OPeNDAP interface (https://opendap.nccs.nasa.gov/dods/) for all of 2021 and filtered to the specific Pandora instrument locations and during TRACER-AQ flight days. We apply both spatial and temporal screening. Spatially, we identify the CAMx grid cell for GCAS and CAMx, as well as the GEOS-CF grid cell in which the Pandora instrument is located. Temporally, for GCAS, we round all overpass times to the nearest hour and calculate the median value for each hour across all overpasses and days. For CAMx, GEOS-CF, and Pandora we identify the simulated and observed NO2 column concentration closest to the hour and calculate the median value across all flight days and locations.

For diurnal comparisons at the surface, we use surface-level NO2 concentrations from CAMx and GEOS-CF and apply the same temporal screening. Spatially, for the surface level we consider concentrations at a point in between the three Pandora instruments that is representative of downtown concentrations (29.7° N, 95.3° W). We choose this point to represent the temporal behavior of the wider regions rather than individual sites. Additionally, we download hourly NO2 concentrations from the US Environmental Protection Agency (EPA) Air Quality System (AQS) (https://aqs.epa.gov/aqsweb/airdata/download_files.html, last access: July 2023). We download all hourly data for 2021 for the United States and filter the TRACER-AQ flight days and for monitors in Harris County. We identify the median hourly concentrations across these monitors and the TRACER-AQ flight days.

The observations from ground-based Pandora instruments are considered the most accurate of all observational platforms measuring column NO2 presented in this project due to low uncertainties in their air mass factors (Herman et al., 2009) when operating in direct-sun mode. The air mass factor in this mode is calculated from simple solar geometry – unlike TROPOMI and GCAS, which rely on a priori assumptions like the vertical NO2 profile and surface reflectivity. Pandora AMFs are not reliant on an a priori profile as the data we are using are only in direct-sun mode in cloud-free scenarios. AMF for Pandora is analogous to pathlength through the atmosphere relative to the vertical path. Since all of the signal is from a direct-sun path (with extremely minimal scattering), this is purely geometric. Given this, we use Pandora observations as our reference dataset to characterize the performance of the two observational datasets – GCAS and TROPOMI – along with the WRF–CAMx simulation across three sites (Table 1). These three sites (Aldine, La Porte, and the University of Houston) are located in the heavily polluted inner region of Houston that we denote as “urban Houston” (Fig. 2f). Background observations from Pandora instruments in less polluted sites were unavailable during the TRACER-AQ campaign, so there is less certainty about the performance of GCAS, TROPOMI, and CAMx outside of urban Houston. We consider the performance of GCAS processed with a CAMx-based AMF (Fig. 2a) and TROPOMI processed with a CAMx-based AMF (Fig. 2b) and CAMx (Fig. 2c), and we also consider the performance of GCAS and TROPOMI with the operational AMFs (Fig. 2d and e) individually and then intercompare the three datasets across the 10 GCAS observation days (Fig. 2g).

When comparing the observational and simulated datasets with Pandora observations, we consider the total column NO2 and we add a stratospheric component from TROPOMI to the tropospheric column NO2 of GCAS and CAMx to total column as discussed in the methodology. For TROPOMI, we use an AMF derived from the CAMx simulation to calculate a tropospheric NO2 column from TROPOMI: following the TROPOMI user guide, we multiply the total averaging kernel by the ratio of the total air mass factor to the tropospheric air mass factor. We difference the total column NO2 from TROPOMI with the tropospheric column to estimate a constant stratospheric NO2 column amount that we add to GCAS and CAMx when we compare them with Pandora. This corresponds to a mean value of 3.0 × 1015.For GCAS, we apply an additional amount, above the aircraft and below the tropopause, to account for the NO2 column in the upper troposphere. We calculate the column of levels 27–29 which correspond to 9400–18 100 m above sea level – that extends roughly from the height at which GCAS flies of around 9100 m to the tropopause – and apply this to the GCAS results; this corresponds to a value of 0.57 × 1015 molec. cm-2. For all results that include comparison with Pandora we present total NO2 columns, and for results where we only intercompare GCAS, TROPOMI, and CAMx, we compare the tropospheric column. All statistical measures (e.g., R2) are defined in the Supplement.

In Fig. 2a–c we characterize the performance of the observational and simulated datasets of NO2 column concentrations across the three sites in Houston. For each of the GCAS flight days, we compare GCAS and TROPOMI observations against Pandora measurements for every overpass that was not obstructed by cloud coverage; for CAMx we compare simulated columns for every daytime hour of each GCAS flight day.

Observations from GCAS were both well correlated (r2 = 0.79) and slightly high biased (NMB = +3.4 %) when compared with measurements from Pandora. Use of the CAMx AMF in place of the operational AMF had a minimal impact on comparisons with Pandora (from r2 = 0.78 and NMB = +6.5 %). Observations from TROPOMI on GCAS flight days were also well correlated with Pandora measurements (r2 = 0.73), but there was a negative bias (NMB = -22.8 %) in v2.4.0. This bias was worse for more NO2 polluted scenarios. This negative bias may be attributable to the coarser resolution of TROPOMI compared with GCAS that weakens its ability to capture fine-scale plumes (Wagner et al., 2023) of NO2 associated with road systems, airports, power stations, and industrial facilities. Similarly to GCAS, use of the CAMx AMF in place of the operational AMF for TROPOMI had a minimal impact on comparisons with Pandora (from r2 = 0.76 and NMB = -23.1 %).

We calculate the ratios of the TROPOMI v2.4.0 product with the CAMx AMF compared with the operational AMF (Fig. S1) in September 2021 throughout the domain and note that tropospheric column NO2 increases in the urban core and decreases in the city outskirts. The areas with Pandora instruments, in suburban Houston, have roughly equivalent values. Given that Pandora instruments were not located at either the most or least polluted areas of the metropolitan area, the benefit of the CAMx AMF may be underrepresented by our findings at the Pandora sites.

We compare simulated NO2 columns from CAMx with Pandora measurements; however, in this comparison there are more points to intercompare as columns were simulated for each hour of every flight day by CAMx and observed multiple times per hour from Pandora. The CAMx-simulated columns were less correlated with Pandora measurements (r2 = 0.34) than TROPOMI and GCAS, and they had a consistent negative bias (NMB = -21.2 %). This poor correlation could partially be explained by differences in WRF-simulated meteorology and observed meteorology specifically from differences in wind speed and direction and an inability to fully capture the bay breeze in Houston. We find that the WRF-simulated wind direction (R2 = 0.76 and MBE = 8°), temperature (R2 = 0.71 and MBE = 0.39 K), and water vapor mixing ratio (R2 = 0.86 and MBE = -1.45 g kg-1) (Figs. S2–S11; Tables S5–S8) are generally well correlated and minimally biased compared with observations; however, there are some unsystematic errors in wind direction (MAE = 26°) and poor correlation in wind speed (R2 = 0.26) that would likely degrade correlation between observed and simulated NO2 columns. While there are errors in the meteorological conditions, the biases at the surface are all small, including minimal bias in the wind speed (MBE = -0.02 m s-1), indicating that the negative biases in NO2 columns are likely attributable to an underestimate of NOx emissions; however, the WRF meteorological performance could partially explain the poor correlation and absolute errors in simulated NO2 columns. We also note that generally, the model performance is stronger on windier days, when speeds exceed 4 m s-1 (R2 = 0.5 and 0.32), than on calmer days, when speeds are below 3 m s-1 (R2 = 0.07, 0.1, and 0.25). Additionally, there can be substantial differences in vertical mixing coefficients in different schemes in the models, and these can impact the biases in column concentrations (de Foy et al., 2007; Riess et al., 2023). We briefly compare meteorology and the ozone mixing ratio in the WRF–CAMx simulation with ozonesonde data (https://www-air.larc.nasa.gov/cgi-bin/ArcView/traceraq.2021, last access: February 2024) and find that while temperature and pressure are captured well, there is variable performance in the vertical structure for the ozone mixing ratio, wind speed, and wind direction (Figs. S14–S18).

In Fig. 2g, we intercompare the daily variability in biases across the 10 GCAS flight days. There were no TROPOMI data for the first 2 flight days because cloud coverage blocked TROPOMI observations at the Pandora sites during its overpass time. The daily average biases of GCAS observations were consistently small throughout the entire period: they ranged from -2.1 to +1.2 × 1015 molec. cm-2 on 10 September and 3 and 24 September, respectively. TROPOMI observations were consistently biased systematically low: they ranged from -4.8 to -0.5 × 1015 molec. cm-2; however, on all days except 26 September, the daily averaged TROPOMI biases were more negatively biased than -1.3 × 1015 molec. cm-2 compared with Pandora measurements. Unlike the two observational datasets, the bias in simulated CAMx NO2 columns had much higher daily variability. On some days, such as 3 September, there was little bias in simulated columns compared with Pandora measurements, and on other days, such as 26 September, there was a minor high bias (+1.2 × 1015 molec. cm-2); however, on most days there was a negative bias that was the strongest on 23 September when NO2 columns were biased as low as -7.5 × 1015 molec. cm-2. Generally, simulated CAMx columns perform better on weekend days (11, 25, and 26 September), which is investigated in greater detail in Sect. 3.4.

We intercompare TROPOMI observations to Pandora measurements across three different algorithms: version 2.3.1 (Fig. 3a, d), version 2.4.0 (Fig. 3b, e), and the NASA MINDS product (Fig. 3c, f) using both the CAMx AMF (top row) and the operational AMF (bottom row) for the same Pandora instruments in Houston during the TRACER-AQ campaign. Overall, the choice of algorithm and AMF does affect the performance of TROPOMI compared with Pandora, albeit slightly. Regardless of AMF, version 2.4.0 appears to have the worst normalized mean bias in Houston during TRACER-AQ (r2 = 0.73 and NMB = -22.8 %), version 2.3.1 is improved (r2 = 0.72 and NMB = -18.3 %), and the NASA MINDS product performs comparably (r2 = 0.69 and NMB = -18.2 %) to version 2.3.1. Notably, NASA MINDS data for 11 September are missing, so these data are excluded from Fig. 3c and f. For version 2.3.1 and version 2.4.0 the CAMx AMF slightly improves the bias; however, for the MINDS product the CAMx AMF slightly worsens the bias compared with the operational AMF. The correlation is generally unaffected by the choice of AMF. We choose TROPOMI version 2.4.0 for the intercomparison in the following sections as it is the most recent version.

The comparisons between Pandora measurements and the datasets indicate that GCAS observations are in best agreement with Pandora. While TROPOMI performs worse than GCAS, it still decisively outperforms simulated NO2 columns from CAMx at Pandora sites in both correlation and bias despite its coarser resolution. With the above in mind, in this section we present NO2 columns observed from GCAS and TROPOMI and simulated from CAMx at the 444 × 444 m2 resolution of the CAMx grid. We extend the prior comparison beyond focusing on three discrete points in urban Houston to the entire CAMx domain to get a more complete picture of the spatial components of these datasets. For each dataset we consider observations across all 10 GCAS flight days. We begin by comparing GCAS observations only with CAMx-simulated columns across all GCAS overpasses as these data are less limited temporally than TROPOMI observations (Fig. 4).

When considering data from all GCAS overpasses (Fig. 4a, b, c) we observe a consistent negative bias in the CAMx product compared with GCAS observations throughout the domain that worsens in the downtown (DT) area (-2.7 × 1015 molec. cm-2) compared with background levels in the rural East Galveston Bay (RB) (-1.2 × 1015 molec. cm-2). Near the W. A. Parish power station in the southwestern area of the domain there is a mixture of positive and negative biases in the CAMx-simulated columns that are likely indicative of errors in wind speeds or directions in the CAMx simulation. Overall, the CAMx-simulated columns were well correlated with GCAS observations (r2 = 0.78), but the negative bias was substantial (NMB = -30.6 %) (Fig. 4d).

We continue this comparison in Fig. 5 where we limit the GCAS and CAMx values temporally around TROPOMI overpasses. For Fig. 5, we screen out all observations that are ±90 min from TROPOMI overpass for each day and then temporally average the observations across the GCAS flight days (Fig. 5a, b, c). We difference, both absolutely (Fig. 5d, e, f) and relatively (Fig. 5g, h, i), the three pairs of datasets and present them in scatter density plots (Fig. 5j, k, l). We focus on three regions: downtown Houston (DT) (red), the rural East Galveston Bay (RB) (blue), and all other areas (OA) (green), and we calculate the mean values and differences for these areas in the top left of each of the plots. The results presented in Fig. 5 are the temporal average across all flight days; however, similar figures for individual flight days are presented in Figs. S19–S28.

First, we consider the spatial distribution of NO2 columns from GCAS (Fig. 5a), TROPOMI (Fig. 5b), and CAMx (Fig. 5c) independently. For all three datasets, NO2 columns are higher in downtown Houston than in the rural East Galveston Bay; generally, they are between 3 and 5 times as large. The two finer-resolution datasets, GCAS and CAMx, also capture NO2 peaks associated with point sources like those from W. A. Parish, Texas City, as well as Baytown and the ship channel. A map of the major point sources discussed in this work is included in Fig. S29. The coarser resolution of TROPOMI leads to fewer identifiable peaks associated with point sources; however, there are slightly elevated observed values near the W. A. Parish and Texas City power plants as well as the ship channel. Observations from GCAS and TROPOMI reveal a more diffuse peak in NO2 columns in and around downtown Houston that includes elevated levels of NO2 in the western part of the city. Simulated columns from CAMx, on the other hand, primarily estimate higher NO2 values in the eastern area of downtown Houston and have lower NO2 values in the western area of the city.

We next consider the three products compared with one another through three methods: absolute difference (Fig. 5d, e, f), relative difference (Fig. 5g, h, i), and scatter density plots (Fig. 5j, k, l). We intercompare these three products by isolating three sets of pairs: GCAS and TROPOMI, GCAS and CAMx, and TROPOMI and CAMx.

First, considering GCAS and TROPOMI, there appears to be a systematic low bias in TROPOMI observations throughout nearly the entire domain. Regardless of the spatial subset, the low bias in TROPOMI was consistent and ranged from -27 % in downtown to -32 % in the rural bay (Fig. 5g). In an absolute sense, on average TROPOMI was between 2.1 and 0.7 × 1015 molec. cm-2 lower than GCAS (Fig. 5d) across the three locations. Throughout the entire domain, observations from GCAS and TROPOMI were well correlated (r2 = 0.85), but TROPOMI had an overall negative normalized mean bias of -31.6 % (Fig. 5j). We note that this low bias is slightly greater than what we would expect from considering the biases of these products relative to Pandora measurements as we do in Sect. 3.1; doing this we would expect TROPOMI to be low biased relative to GCAS by around 23 %. This slight additional negative bias indicates that either the three Pandora sites are unable to capture the full extent of the negative TROPOMI bias, and that TROPOMI may be lower biased outside of these sites (e.g., areas outside of downtown Houston), or that GCAS observations may be biased additionally high outside of these sites. Notably, there are a few areas surrounding point sources in the eastern area of downtown and around the W. A. Parish plant in which TROPOMI observes higher NO2 columns than GCAS. This is likely attributable to the coarser resolution of TROPOMI that results in peaks of NO2 spreading into surrounding areas that are in the same TROPOMI grid cell.

Second, comparing GCAS with CAMx we again find a low bias relative to GCAS, albeit one with a higher degree of spatial variability. In the remote bay, CAMx-simulated columns are lower than GCAS compared with elsewhere in the domain (-50 %) (Fig. 5h), while downtown and background levels are similarly biased at 32 % and 39 %, respectively. This lower bias in the low emissions rural East Galveston Bay is indicative of an underestimation of background NO2 columns in the CAMx simulation. Across these three regions the mean absolute differences range from -2.4 to -1.2 × 1015 molec. cm-2 (Fig. 5e). Visually, the negative bias in CAMx appears to be stronger in downtown and to the west, east, and northwest of downtown, and less to the south and southwest of downtown. Overall, GCAS and CAMx are well correlated (r2 = 0.74) (Fig. 5k); however, simulated columns from CAMx have a worse negative bias (NMB = -38.1 %) against GCAS than what is captured at the Pandora sites of approximately -21 %. Around some point sources CAMx columns are positively biased against GCAS observations. This high bias in CAMx is likely attributable to differences in wind speed and direction in the WRF simulation than in reality. These differences could contribute to NO2 plumes being advected in incorrect directions.

Lastly, when comparing observed columns from TROPOMI to simulated columns from CAMx, biases have a great degree of spatial variability; however, in general, CAMx is negatively biased. In a relative sense (Fig. 5i), the CAMx-simulated columns are lowest compared with TROPOMI in the rural bay (-26 %) and similar in downtown (-7 %) and in other areas (-11 %). There are a few areas where this pattern does not hold: both in the area southwest of downtown Houston and near point sources, CAMx is biased high compared with TROPOMI. These results indicate that simulated columns from CAMx are underestimated in downtown Houston and that this underestimation could potentially be attributable to an incorrect advection of NO2 from some downtown source to the southwest perhaps in conjunction with an underestimate of emissions in this downtown area. Overall, TROPOMI and CAMx are well correlated (r2 = 0.73), and there is a spatially heterogeneous low bias when considering the two products throughout the domain (NMB = -9.7 %) (Fig. 5l).

The comparisons presented in the prior section are done at the high resolution of the CAMx grid (444 × 444 m2). Here, we characterize the effect of the coarser resolution of TROPOMI by performing an additional comparison of the three datasets at the 0.05° × 0.05° resolution (approximately 5.5 × 5.5 km2) (Fig. 6). We average all of the NO2 columns from this finer resolution to the coarser resolution based on the centroid of the fine resolution grid cells. This new coarser resolution is comparable to that of the TROPOMI observations at nadir (on average 3.5 × 5.5 km2). We additionally present comparisons at two further coarser resolutions in the Supplement: 0.25° × 0.25° (Fig. S30) and 0.1° × 0.1° (Fig. S31).

Generally, this change in resolution has only a minor effect on the trends discussed in the prior section. Observed NO2 columns from GCAS and TROPOMI have a collocated peak in downtown Houston and NO2 columns from TROPOMI are still systematically biased lower compared with GCAS. Simulated NO2 columns from CAMx are clearly lower than GCAS in the area directly west of downtown and slightly higher southwest of downtown compared with TROPOMI (Fig. 6a, b, c). Considering the spatial distribution of absolute (Fig. 6d, e, f) and relative (Fig. 6g, h, i) differences between the three products, the low bias in TROPOMI compared with GCAS is generally homogenous throughout the domain. On the other hand, there are clear peaks in negative biases in downtown and western Houston when comparing CAMx to GCAS, and in some areas southwest of downtown biases are small and positive. Averaging observations to this coarser resolution improved the correlation for all three pairs (r2 = 0.93, 0.82, and 0.83 for GCAS and TROPOMI, GCAS and CAMx, and TROPOMI and CAMx, respectively), while the biases remained comparable to what was found in the comparison at a finer resolution (Fig. 5j, k, l).

Three of the 10 GCAS flight days occurred on weekends (11, 25, and 26 September), and observations from GCAS and TROPOMI – along with simulated NO2 columns from CAMx – exhibited different patterns on weekends vs. weekdays (1, 3, 8–10, 23, and 24 September). This difference in observed and simulated patterns is explored in greater detail in this section, first through comparisons with Pandora measurements (Fig. 7) and then through spatial comparisons of the products on weekdays vs. weekends (Fig. 8). When interpreting these results, it should be considered that weekend data are limited to only 3 d. This data sparsity introduces a high degree of uncertainty in conclusions derived from this analysis. Day-to-day changes in meteorological conditions are likely responsible for some of the exhibited differences, so they cannot solely be attributed to differences in emission patterns.

First, we consider how comparisons of the observational datasets – GCAS and TROPOMI – with Pandora change on weekends compared with weekdays. Biases for both GCAS and TROPOMI become more positive on weekends, NMB = 10.2 % and NMB = -15.7 %, respectively, than on weekdays, NMB = 1.5 % and NMB = -25.2 %. GCAS observations are slightly better correlated to Pandora measurements on weekends (r2 = 0.89 vs. r2 = 0.76); however, TROPOMI observations are worse correlated (r2 = 0.42 vs. r2 = 0.69), which is likely attributable to a limited number of observations that are at lower NO2 column levels with limited dynamic range. Overall, biases are slightly worse for GCAS and better for TROPOMI on weekends; however, given the small number of measurements it is unclear whether this pattern is attributable to meteorological conditions or if it is attributable to some systematic bias in the instruments.

Simulated NO2 columns from CAMx exhibit clearer weekday vs. weekend patterns, and since these simulated columns are available for every hour of the day there is a greater number of measurements to support these findings than for the two observational datasets. While the correlation is slightly degraded on weekends (r2 = 0.30 vs. r2 = 0.37), the negative bias in simulated columns compared with Pandora measurements is reduced on weekends (NMB = -25.5 % vs. NMB = -9.5 %).

GCAS and TROPOMI observations of NO2 column concentrations are higher on weekdays (Fig. 8a, b) than on weekends (Fig. 8d, e). This is true in downtown Houston and the rural bay where weekday GCAS observations are 2.1 × 1015 molec. cm-2 (24 %) and 0.6 × 1015 molec. cm-2 (24 %) higher, respectively, on weekdays than on weekends. In other areas of Houston, GCAS observations on weekdays are higher than on weekends but not to the same degree (20 %). A similar pattern occurs for TROPOMI: in downtown Houston TROPOMI columns are 20 % higher on weekdays than on weekends but comparable in other areas and -5 % lower in the rural bay. This comparison again implicates some underestimated weekday source of NO2 in CAMx that is of great importance in the western area of Houston; however, due to the lack of data on weekends – that is apparent in the discontinuities in the weekend NO2 column concentrations of TROPOMI – it is difficult to examine this quantitatively.

Comparing weekday columns simulated from CAMx with weekend columns, we find that the mean concentrations for the three defined areas are nearly identical (Fig. 8c and f), although columns on weekdays are higher south and southwest of downtown, while columns on weekends are higher within downtown. These spatial patterns are further revealed in the difference plots (Fig. 8i and l) where the difference in weekday vs. weekend values appears to be split right along interstate highway I-10: north of I-10 weekday values are much lower than weekend values, while south of I-10 the opposite is true. This difference is likely attributable to different meteorological conditions on these days. Overall, simulated CAMx columns are substantially lower than GCAS and TROPOMI on weekdays but more similar on weekends, implying that weekday emissions may be underestimated in the TCEQ inventory.

Lastly, we characterize the diurnal profiles of simulated and observed NO2 columns during the TRACER-AQ campaign in downtown Houston (Fig. 9). First, considering column concentrations, we find generally good agreement during the early morning (08:00–10:00 LT) across the two simulated datasets (CAMx and GEOS-CF) and two observational datasets (GCAS and Pandora). Interestingly, between 09:00 and 11:00 LT, Pandora column measurements show a slight increase, while model simulations show a slight decrease during the same time interval. During midday and the afternoon (11:00–16:00 LT) – that corresponds to the period with the most GCAS observations – GEOS-CF columns generally agree well with Pandora observations. In the evening (17:00–19:00 LT), GEOS-CF columns have a substantial high bias across these flight days. The GEOS-CF mismatch in the evening has implications for TEMPO NO2 evening retrievals if this is a persistent bias in other urban areas since satellite instruments are especially sensitive to a priori assumptions at low sun angles.

Second, considering surface concentrations, we see a similar trend. Generally, there is great agreement across the three datasets (CAMx, AQS observations, and GEOS-CF) in the early morning (06:00–09:00 LT) before they begin to diverge with the two simulated produces maintaining comparable magnitudes with low biases compared with surface monitors. At around midday to the afternoon (12:00–17:00 LT) both simulated products have a low bias compared with observed surface-level NO2; however, the bias in CAMx concentrations is worse. Some of the apparent low bias may be related to an artificial high bias in NO2 chemiluminescence surface monitors (Dunlea et al., 2007; Lamsal et al., 2008). In the evening (18:00–19:00 LT), surface-level NO2 from GEOS-CF climbs rapidly; however, observed NO2 from the AQS and simulated NO2 from CAMx increase only slightly. The large increase in NO2 in GEOS-CF in the evening appears at both the surface and in the column, potentially indicating issues capturing boundary layer dynamics.