Recent independent validation studies of the operational GEMS v3 product over Bangkok and South Korea report low biases in NO2 columns relative to ground-based sun-photometer and DOAS measurements (Bae et al., 2025; Jung et al., 2025). Bae et al. (2025) shows that GEMS v3 increasingly underestimates NO2 relative to Pandora under high-NO2 conditions (>1×1016 molecules cm⁻²) as is the case for Bangkok pollution levels. In Jung et al. (2025), validation results over Bangkok indicate a pronounced low bias in GEMS tropospheric NO2 columns relative to Pandora, with regression slopes of ∼ 0.35 for v2.0 and ∼0.28 for v3.0, indicating increasing underestimation at higher NO2 levels. While moderate correlations (r≈ 0.6–0.7) suggest that GEMS captures temporal variability, column magnitudes are substantially underestimated, particularly under polluted conditions. The persistence of this behavior in the v3.0 product indicates that the low bias is not fully corrected by recent algorithm updates and is consistent with retrieval sensitivity limitations in highly polluted urban environments (Jung et al., 2025).

To address the low bias in GEMS prior to the NOx satellite emission inversion, we first quantify the GEMS bias relative to Pandora measurements over our period of interest (14–27 March 2024). We compare total column GEMS NO2 to Pandora Level 2 direct-sun total column retrieval, filtering for high quality measurements (quality flag = 10) and averaging to hourly means. GEMS columns are sampled at the nearest grid cell to each Pandora site and temporally collocated. Results are shown in Fig. S1 in the Supplement. GEMS NO2 columns are generally about a factor of two lower than Pandora measurements (mean bias ≈-9.2×105 molecules cm⁻²). This factor-of-two difference persists throughout March 2024 (Fig. S1).

To assess the sensitivity of the top-down NOx inversion framework to this bias, we apply a simple correction factor to GEMS prior to inversion. Specifically, GEMS NO2 columns over the BMR are scaled by a factor of 1.67 derived from the Pandora comparison. The inversion is then performed with and without the bias-corrected columns.

Previous studies have used satellite data (OMI, TROPOMI, SCIAMACHY, GOME (-2), OMPS) to estimate top-down NOx emissions over urban areas, but were limited to once-a-day, mid-afternoon measurements, leaving temporal variability in emissions largely unaddressed (Beirle et al., 2011; Goldberg et al., 2017, 2019). Ground-based and aircraft measurements of emissions can be challenging to constrain given boundary layer dynamics, such as changes in boundary layer height, stability, and vertical mixing, which can strongly influence observed concentrations of trace gases (Goldberg et al., 2017, 2019). Additionally, temporal allocation in bottom-up emission inventories remain a significant uncertainty, particularly at hourly and daily timescales because default temporal profiles (diurnal, weekly, seasonal) often fail to capture real activity patterns, meteorological influences, and sector-specific variability (Goldberg et al., 2017, 2019; Mues et al., 2014). GEMS observations, however, offer a unique opportunity to provide hourly emissions rates over Bangkok, complementing existing inventories that are usually provided at monthly scales. To estimate NOx emissions over Bangkok in March 2024, we apply the methodology of Kuhlmann et al. (2024), implemented through the openly available data-driven emission quantification (ddeq v1) Python library (see step 1, Fig. 2). The standard ddeq library implements computationally inexpensive methods (e.g., Gaussian Plume Inversion (GP), Cross-Sectional Flux (CSF) method, Integrated Mass Enhancement (IME) method) to estimate emissions from Sentinel5P TROPOMI images (Graziosi and Manca, 2025; Meier et al., 2024). One study found limited temporal sampling from polar-orbiting satellites (i.e., TROPOMI) can lead to systematic underestimation of NOx emissions, particularly because wintertime conditions with higher emissions are often poorly observed due to cloud cover (Meier et al., 2024). Geostationary observations, including GEMS, address this limitation by providing hourly measurements that increase the likelihood of usable data on partially cloudy days and enable direct resolution of daytime emission variability, which is not accessible from once-daily LEO observations. For our application, we estimate emissions for daylight hours from GEMS tropospheric column NO2 data over Bangkok during the ASIA-AQ campaign. A brief description of inversion methods and specifics is provided here. A complete description of the inversion methods and algorithm can be found in Kuhlmann et al. (2024).

First, within the ddeq framework, the source location (Bangkok; 13.7563° N, 100.5018° E), GEMS v3 NO2 column data, and ERA5 wind fields are read. We chose to incorporate ERA5 wind fields as opposed to those provided by our high resolution WRF-Chem simulation in this analysis given they presented significantly lower bias when compared to surface observations (Fig. S2; Sect. 6.1). The ERA5 wind fields were downloaded and prepared by the ddeq library. Cloudy pixels (CF > 0.3) were removed. Next, a plume detection algorithm is implemented to identify the Bangkok urban plume within the GEMS image. The Bangkok plume subregion is estimated based on the source location and ERA5 wind field. Generally, a wind vector is taken at the source location, and the plume is assumed to be located downwind, and a rectangular polygon can be drawn with the along- and across-wind direction (Kuhlmann et al., 2024). An example of these plume detections and associated rectangular polygons is shown in Fig. S3. To aid the algorithm's plume detection process, we identify two additional sources north of Bangkok in the Saraburi province to avoid overlapping plumes from other NOx sources. These correspond to the Khao Wong (14.692856° N, 100.817204° E), and Thap Kwang (14.645313° N, 101.077650° E) regions, which we identified through NO2 patterns visible in the GEMS data in Fig. S3. These areas are not representative of Bangkok urban emissions as they are a source of industrial activity related to limestone quarries and mining operations outside the inversion domain (Makkwao and Prueksasit, 2021). Next, a center curve is fitted to the data, and natural coordinates are calculated for the detected plume associated to Bangkok to prep for the inversion algorithm. The coordinates are computed as the distance along and perpendicular to the wind vector and for curved plumes this distance is computed as the arc length. Lastly, the data is prepared for the satellite-derived emissions estimation by calculating and removing the background field and converting the NO2 columns to kg m⁻². We estimate satellite-derived NOx emissions using three inversion techniques included in ddeq: CSF, GP, and IME methods. Below is a summary of the CSF method defined in Kuhlmann et al. (2024). GP and IME method descriptions are available in the Supplement (e.g., S1) and Kuhlmann et al. (2024). In the following subsections and in Sect. 4, we refer to the WRF-Chem input inventory based on EDGAR as the prior emissions, the CSF-based estimates from GEMS as the satellite-derived emissions, the CSF-based estimates from the model-simulated columns as the model-derived emissions, and the final emissions after applying the optimization framework, as the posterior emissions.

The CSF approach applies mass conservation to quantify the NOx flux transported downwind of Bangkok using (i) the wind speed perpendicular to the plume and (ii) the NO2 enhancement integrated along the plume (line density). The NOx flux, F (kg s-1), is defined as: 1F=u⋅q where u (m s-1) is the effective transport wind speed and q (kg m-1) is the NO2 line density. The emission rate, Q (kg s-1), is then inferred by correcting for chemical decay along the plume using: 2Q=F(x)D(x,τ)

Here, D(xτ) represents the along-plume decay for lifetime τ. Here, τ represents an effective plume lifetime that accounts for both chemical decay and plume dispersion (Kuhlmann et al., 2024).

GEMS NOx emissions for Bangkok were estimated for daytime hours (00:45–06:45 UTC) during the ASIA-AQ deployment, 14–27 March 2024. The bias-corrected satellite inversion results are denoted with the subscript “BC”. Days with significant cloud contamination were removed from the analyses on a case-by-case basis, leaving satellite-derived emissions estimates for 14, 15, 17, 18, 22, 24, and 25 March 2024. Satellite-derived emissions were computed for both weekends and weekdays, however, due to fewer weekends in the study period, weekday estimates are estimated to be more robust. We aggregate satellite-derived emissions hourly across weekdays to produce a summary daytime profile as illustrated in Fig. 3a and b. Satellite-derived emissions generally range between 0.5 and 4 kg s⁻¹ dependent on the inversion method. Bias-corrected results (Fig. 3b) illustrate a larger NOx range, between 0.4 and 5 kg s⁻¹. Differences between CSF, GP, and IME satellite-derived emission estimates are expected, as each method relies on distinct assumptions regarding transport, plume geometry, and chemical loss. Similar spreads between methods have been reported in previous satellite-based emission studies (Hakkarainen et al., 2024; Santaren et al., 2025), particularly in urban environments, and can be interpreted as a measure of structural uncertainty rather than inconsistency between methodologies. Nevertheless, in this application, the CSF method illustrates a distinct daytime pattern with satellite-derived emissions peaking between 08:00 and 09:00 LT (bias-corrected) coinciding with morning rush hour traffic, decreasing by approximately 65 % by 14:00 LT. The GP and IME results illustrate a relatively stable daytime pattern and overall lower (< 50 %) morning satellite-derived emission compared to the CSF method.

These results reflect both methodological differences and uncertainties in the underlying NO2 observations. While the satellite-derived emissions might provide strong constraints on daytime variability, uncertainties in the GEMS NO2 retrieval, particularly under high NO2 conditions may influence the magnitude and timing of the inferred emissions as depicted in Fig. 3a and b. The bias-corrected emissions are approximately 70 % higher on average, indicating a substantial sensitivity of the inversion to the assumed retrieval bias.

To identify the optimal inversion method to be used for model correction, we test which method can successfully recover WRF-Chem's prior daytime NOx emission profile. To do so, we apply the ddeq inversion algorithms to WRF_Base, to produce “top-down” estimates from the model, which we will refer to as model-derived emissions. To replicate the satellite output, we re-gridded the model output from its native resolution to the satellite swath grid using the Universal Regridder for Geospatial Data (xESMF) Python package (Zhuang et al., 2025). A nearest-neighbor interpolation scheme was used to preserve spatial gradients and align the model resolution to that of the satellite observations. This spatial re-gridding was performed on an hourly basis, corresponding to the nearest observation time (00:45–06:45 UTC). Following re-gridding, model NO2 mixing ratios were vertically interpolated to the satellite pressure levels using a one-dimensional linear interpolation to ensure consistency between the model and observed data for averaging kernel application. After interpolating model output to GEMS's spatial and pressure grid, we computed the model's tropospheric column NO2 (molecules cm⁻²). This is done by converting the model's standard output of NO2 volume mixing ratio (ppmV) to a number density with the ideal gas law: 9NO2NumberDensity=P⋅χNO2⋅NAR⋅T where χNO2 is the NO2 mixing ratio in ppmV, P is pressure in Pa, T is temperature in K, R is the gas constant (8.314 J mol⁻¹ K⁻¹), and NA is Avogadro's number (6.022×1023 mol-1). The number density is then multiplied by the thickness of each vertical model layer Δz to obtain the NO2 amount per unit area. Summing across all layers yields the total column NO2: 10NO2WRF-GEMS=1104×∑kAk⋅NO2NumberDensity,k⋅Δzk where Ak is the averaging kernel value at layer k, Δzk is the thickness of the model layer k, in meters, and 1104 converts the units to molecules cm⁻². As WRF-Chem has no stratosphere, we regard this as the tropospheric column NO2. The satellite averaging kernel weights each model layer based on how sensitive GEMS is to that part of the atmosphere. This allows for a direct comparison with GEMS retrievals.

We compute model-derived NOx emissions from re-gridded WRF_Base (which we denote as WRF-GEMS) with ddeq and the resulting average daytime profile for Bangkok is illustrated in Fig. 3c. The shape and magnitude of model-derived NOx emissions differ substantially from the satellite-derived emission data. Model-derived emissions are initially >50 % larger than GEMS-derived emissions. Additionally, there are minimal changes in the daytime pattern present in any of the inversion methods. In fact, WRF-GEMS D02 sees an increase in emission between 08:00 and 10:00 LT, a pattern that disagrees with the observed satellite-derived emissions. The results here further illustrate the need for an hourly correction to the model's emission profile, rather than applying a single scaling factor across all hours. The WRF_Base prior emission profile is shown by the black dashed line. For this comparison, we smoothed the prior emission profile using a 3 h backward-looking average to account for the accumulation and chemical evolution represented implicitly in the inversion results. Overlaying the prior emissions shows that the model-derived CSF method most accurately recovers them, likely due to its use of downwind fluxes and multiple cross-sections across the plume, which provide additional constraints on plume evolution and reduce sensitivity to local errors in wind fields, plume definition, and missing data compared to the IME and GP methods. As a result, we select the CSF inversion method moving forward to correct and scale WRF-Chem's daytime NOx emission profile.

Using satellite-derived NOx emissions and the posterior results above, we correct the WRF-Chem 24 h input emissions over the BMR. Since GEMS provides estimates over 01:00–07:00 UTC, we only have posterior model emissions over this window. To apply these updates to the full gridded 24 h WRF-Chem anthropogenic emission input (00:00–23:00 UTC), while preserving the model's spatial distribution, we compute hourly scaling factors which we use to adjust the gridded prior NOx emissions. Because GEMS provides constraints only during the daytime overpass window, the derived scaling below primarily reflects daytime emission adjustments. We therefore apply the mean scaling factor uniformly during the nighttime hours to preserve the prior temporal structure and avoid introducing artificial discontinuities between the constrained and unconstrained hours, particularly given the large daytime emission reductions. This approach is consistent with previous satellite-constrained emission studies (e.g., Goldberg et al., 2019) which constrain emission magnitude based on a single daytime value derived from OMI. Where available, locally derived temporal profiles should be used to better represent regional emission patterns; in this study, our prior profile reflects the best available regional representation for Asia at the time of study.

To compute the factors, hourly WRF-Chem NO and NO2 prior emission fields were summed across the vertical levels and masked to the Bangkok region bounded by N–S: 13.5–15.0°, W–E: 100.2–100.9°, using the WRF-Chem grid. In WRF-Chem, NOx emissions are stored as NO2-equivalent mass (“NOx as NO2”), so all conversions use the molecular weight of NO2. The total emissions were converted to mol h⁻¹ using a molecular weight of 46.01 g mol⁻¹ and an area of 16 km² per grid cell (following the resolution of D02). Each posterior emission, x^h was converted from kg s⁻¹ to mol h⁻¹ and divided by the corresponding original WRF-Chem total to compute a scale factor, f for each hour h: 16f=x^h×100046.01×3600EHWRF where EHWRF is the original domain-integrated prior NOx emissions for hour h. Since the optimization covered hours with GEMS measurements, the scaling factors for the remaining model hours were computed using the ratio of average daytime posterior emissions where averages are taken over the constrained hours. Finally, we apply the scaling factors spatially over the BMR. The scaling factor was applied to the hourly WRF-Chem anthropogenic emissions prior files and only to NO and NO2 grid cells within the BMR mask. This preserves the original spatial distribution but adjusts the total prior emissions to match the posterior values. The updated emission files were then used to initialize an updated model simulation, ran for the ASIA-AQ deployment period, which we will refer to as WRF_Updated (see step 3, Fig. 2).

Figure 4e compares the average diurnal prior profile over BMR with the updated profile derived from the optimization and scaling procedure. The optimized diurnal profile based on the bias-corrected GEMS constraint is illustrated by the light pink line. The updated profiles show a notable increase in NOx between 07:00 and 10:00 UTC (14:00 to 17:00 LT), which may reflect the beginning of a rush-hour signal over the city. Nighttime emissions retain the overall shape of the prior profile but are scaled downward based on the daytime average. Figure S6 maps the spatial distribution of these updates across BMR, showing the largest differences in the late afternoon and early evening, with localized reductions of up to 400 mol km⁻² h⁻¹ in the updated input emissions.

To place the updated NOx emissions in context of existing emission inventories, we compared total monthly emissions over the BMR against several widely used bottom-up inventories, our top-down emissions (including bias-corrected results – This StudyBC), and emissions from a chemical reanalysis product (Fig. 5). These include HTAP v3.2 (base 2019), a local Thailand Inventory (THAI-KMUTT) (base 2019), MIX v2.3 (base 2017), EDGAR v5 (base 2015), CEDS (base 2024), our study (2024), and the Tropospheric Chemistry Reanalysis (TCR-3) top-down emissions TCR-3 (2024) (Miyazaki et al., 2020). Across inventories, monthly totals for March can vary nearly by an order of magnitude, highlighting large uncertainty in regional NOx sources and emissions processing methodology. For example, although both HTAP v3.2 and MIX v2.3 rely on the REAS framework (Kurokawa and Ohara, 2020) for Asian anthropogenic emissions, they are derived from different REAS versions, temporal coverage, and processing assumptions, which can lead to substantial differences in absolute emission magnitudes over Bangkok. HTAP v3.2 is based on REAS v3.2.1, which provides monthly emissions for 2000–2015 at a spatial resolution of 0.25°. In the HTAP product, these emissions are re-gridded to 0.1° by assuming a uniform spatial distribution within each grid cell (Guizzardi et al., 2025). In contrast, the MIX inventory is based on REAS v2, which covers 2000–2008 and reflects earlier emission estimates and activity data (Li et al., 2024). REAS v2 was subsequently extended to 2010 following the scaling approach described by Kurokawa and Ohara (2020).

Overall, HTAP v3.2 depicts the highest NOx totals at 31 kt month⁻¹. THAI-KMUTT follows with emissions at ∼ 20 kt month⁻¹, reflecting its incorporation of detailed regional activity from local emissions data. The coarser-resolution MIX v2.3, and EDGAR v5, and CEDS inventories reflect lower totals, at ∼ 13, 11, 10 kt month⁻¹ respectively, consistent with the use of different emission estimation methodology and older base years for MIX and EDGAR which correspond to lower macro-economic (e.g., GDP) indicators in Thailand compared to 2019. Lastly, GEMS top-down NOx estimates (including bias-corrected results) and TCR-3 estimates based on the chemical reanalysis using TROPOMI NO2 indicate substantially lower emissions (2–4 kt month-1). These reduced magnitudes may result from both recent emission declines and structural differences in how emissions are computed (i.e., top-down perspective).

It is important to note uncertainties are present and differ amongst emission methods. For example, bottom-up inventories depend on activity data, emission factors, spatial approximations, and assumptions that may not capture rapid socio-economic changes or region-specific behavior. Top-down estimates incorporate uncertainties from their respective satellite retrievals, reanalysis data (e.g., ERA5), averaging kernels, chemical lifetime assumptions, and forecast-model transport and chemistry (e.g., TCR-3). Together, these differences highlight the value of combining observational constraints with updated bottom-up information to define NOx emission estimates.

To evaluate model performance of surface trace gases, modeled surface mixing ratios of NO2, NOx, and O₃ were evaluated against observations across BMR during 14–27 March 2024. Surface observations of these constituents (in ppbV) were obtained from Thailand PCD ground monitor network data. Like the evaluation of wind speeds and direction, data was co-located by extracting model information from the nearest grid cell and matched to the observation timesteps. Station data and model data were averaged across stations located within the Bangkok urban plume, bounded by N–S (13.5–14.6° N), W–E (100.2–101° E). Figure S8b illustrates the station locations and averaging domain in Bangkok. For the air quality evaluation, we restrict the station averaging domain to a smaller area over the Bangkok urban plume. This smaller domain allows us to assess how the inversion and optimization specifically correct the local plume structure as opposed to a broader regional background.

We additionally evaluate modeled NO2 columns against Pandora measurements for the Bangkok site. Pandora NO2 evaluation used Level 2 direct-sun total column retrievals, filtered for high-quality measurements (quality flag =10), and averaged to hourly means. Tropospheric columns from Pandora were estimated by subtracting coincident or closest GEMS stratospheric NO2 at the nearest satellite pixel. Model columns were sampled at the nearest grid cell to the Pandora site and temporally matched to observations.

Figure 7d depicts a comparison of WRF_Base (light purple triangles), WRF_Updated (dark purple squares), WRF_Updated+BC (magenta dots) and Pandora (pink stars), tropospheric NO2 column measurements for days with high quality data. Throughout this period, Pandora measurements generally ranged between approximately 7×1015 to 3.5×1016 molecules cm⁻² and fit between WRF_Base which places columns higher (e.g., up to 5×1016 molecules cm-2), WRF_Updated+BC and WRF_Updated which places columns lower (e.g. 2×1015 molecules cm⁻²). Average statistics between model cases for this analysis are shown in Table 2. Biases are generally similar between model cases. WRF_Base overestimates column NO2 as illustrated by the mean bias (+7.9×1015 molecules cm-2) whereas emissions updates contribute to an underestimation (-6.6×1015 molecules cm-2). However, there are some improvements in absolute error metrics with ME and RMSE decreasing by roughly 20 % and 12 % respectively, and NME dropping from 61 % to 48 %. The greatest bias improvements overall are seen in WRF_Updated+BC (-3.8×1015 molecules cm⁻²). Figure S13 illustrates the spatial distribution of mean tropospheric model NO2 bias for GEMS and Pandora during the time reflected in Fig. 7d. In the WRF_Base simulation, a strong positive bias is seen within and north of the Pandora site, indicating the overestimation of NO2 columns (up to 2×1016 molecules cm⁻²) in the urban plume relative to GEMS. This pattern reiterates the point that the prior anthropogenic emissions (based on EDGAR v5) were too large with pollutant accumulation occurring downwind of Bangkok as a result. After applying the emission updates, the WRF_Updated simulation essentially eliminates this bias. The overall bias near the Pandora site becomes close to neutral or slightly negative, demonstrating the optimization effectively corrected the spatial overprediction. WRF_Updated+BC further reduces the residual negative bias relative to WRF_Updated with NMB improving from -46 % to -26 % and RMSE decreasing from 8.2×1015 to 6.4×1015 molecules cm⁻². This indicates that accounting for the GEMS retrieval bias in the initial optimization process partially corrects the remaining underestimation, although some spatial discrepancies persist dude to transport and plume placement error. Examples of daily biases for WRF_Base D01, D02, and WRF_Updated D01, D02 can be found in Figs. S14 and S15.

The Airborne and Satellite Investigation of Asian Air Quality (ASIA-AQ) was a NASA field campaign conducted in February–March 2024 to advance the understanding of urban and regional air quality across East and Southeast Asia. Targeting several megacities (e.g., Manila, Seoul, Bangkok, Chiang Mai), the campaign combined satellite observations, aircraft measurements, ground-based monitoring, and modeling approaches to characterize pollution sources and validate satellite retrievals. A main objective of ASIA-AQ was to evaluate the data from GEMS. Airborne observations were collected using the NASA DC-8 in situ and LaRC G-III remote sensing aircraft, with coordinated support from ground-based networks such as Pandora and AERONET. Additionally, chemical transport models (e.g., GEOS-Chem, GEOS-FP, MUSICA, WRF-Chem, WRF-CMAQ) played a key role in real-time flight planning and post-campaign interpretation.

In our study, the comparisons with independent aircraft and ground-based observations indicate the satellite-constrained emissions without bias-corrections presented here may represent a low estimate. In particular, the systematic low bias in GEMS NO2 is consistent with the negative biases often seen in the WRF_Updated simulations. Importantly, this bias could not be diagnosed using the satellite data alone.

To further evaluate the relative behavior of GEMS and airborne GCAS NO2 columns, we compare their tropospheric columns using WRF_Updated+BC as a common transfer framework (Fig. S21). Figure S21 compares the observed GCAS/GEMS NO2 column ratios with ratios calculated after both datasets are mapped through WRF_Updated+BC. In the observations, GCAS columns are consistently higher than the raw GEMS columns as illustrated by the orange line, with GCAS/GEMS ratios of ∼ 2–3 on most days and values reaching ∼ 6–7 on 21 March. In contrast, the corresponding WRF-GCAS_Updated+BC/WRF-GEMS_Updated+BC ratios are closer to 1:1 (∼ 1.1–1.7), even on 21 March. This behavior is consistent with a low bias in GEMS NO2 columns relative to GCAS. After applying the bias correction, the GCAS/GEMS_BC ratios approach 1:1 on most days, indicating improved consistency between the datasets, and supporting the effectiveness of the applied GEMS bias correction.

When incorporating bias-corrected retrievals into the top-down inversions and NOx emission optimization process, the resulting WRF_Updated+BC run reduces the negative biases in the independent validation, in some cases bringing the model results even closer to observations as seen in the surface air quality analysis. However, this improvement is not uniform. In certain comparisons (e.g., GCAS and DC-8), WRF_Updated+BC introduces a tendency toward overprediction in morning hours, highlighting tradeoffs associated with the bias correction. These results suggest that while accounting for the retrieval bias can improve mean model performance, additional uncertainties such as overestimated wind speeds and associated transport errors continue to influence the representation of modeled NO2.

Here, the integration of ground-based, airborne, satellite, and model data provides a powerful framework not only for improving emissions but also for identifying limitations within individual observing systems. While WRF_Updated clearly outperforms the baseline model, and WRF_Updated+BC offers targeted improvements in reducing systematic bias, the combined observational evidence highlights the necessity of a multi-platform validation to fully interpret the satellite-based emission estimates.

Despite this low bias in the GEMS retrieval, the high-frequency daytime sampling provided by geostationary observations offers critical constraints on daytime variability and plume evolution that are particularly valuable for emission inversion and air quality modelling. For example, the GEMS-constrained emission adjustments presented here were critical for improving the temporal evolution of NOx in WRF-Chem, resulting in substantial and robust improvements in model performance across independent evaluations. Future work may further benefit from continued refinement of bias-corrected GEMS products and upcoming algorithm improvements (e.g., v4), alongside improved representation of spatial representation of emissions within the model.

The current model emission update framework optimizes the temporal evolution of emissions but does not explicitly resolve regional variability, as a single set of hourly scaling factors is applied uniformly across all grid cells within the BMR. As a result, the inversion preserves the spatial structure of the prior inventory, and any inaccuracies in the spatial distribution of emissions are not corrected.

Extending this approach to resolve emissions at finer scales would require allowing emissions to vary across grid cells or subregions, supported by additional constraints. These constraints could include regional higher-resolution regional prior inventories when available, as well as higher-resolution observations (e.g., GCAS) that can better resolve the urban variability in NO2 as shown in this analysis.

Accurate urban NOx emissions remain a major challenge for air quality modeling efforts, but geostationary satellites now offer a path forward. This work takes a novel approach in improving urban NOx emissions using daytime GEO satellite observations. We quantify and reduce biases in modeled NO2 over the Bangkok Metropolitan Region (BMR), by integrating GEMS constraints into a high-resolution model's daytime prior emission profile. We first used top-down inversions of hourly GEMS NO2 columns with the Cross-Sectional Flux (CSF) method to develop an average daytime NOx profile for the BMR in March 2024. Using this information, we developed and applied an optimization framework that incorporates physical constraints (i.e., with regards to emissions accumulation and lifetime) to reshape the WRF-Chem's daytime emission pattern and magnitude to better reflect observed emissions variability in the BMR.

This represents one of the first applications of the ddeq framework to estimate hourly urban NOx emissions with geostationary observations in Southeast Asia. In contrast to a full chemical reanalysis, the ddeq-based framework provides an efficient and scalable alternative that does not require repeated model reinitialization of extensive chemical state optimization. This makes the approach particularly suitable for regional applications and for broader application across different modeling systems and urban environments within the domain of GEO sensors.

Re-running WRF-Chem (D01 – 20 km; D02 – 4 km) with an updated NOx emission profile (WRF_Updated; WRF_Updated+BC) led to substantial improvements across multiple independent datasets. Model evaluation with ground monitors, Pandora, GCAS, and DC-8 observations consistently showed reduced biases and errors relative to the baseline simulation based on EDGAR v5 emissions. For example, surface-level comparisons with Thailand PCD ground network data confirmed some of these trends in WRF_Updated, with mean biases in NO2 and NOx decreasing from +12 to +14 ppbV in WRF_Base to roughly -3 ppbV in WRF_Updated, and 0.02 to 0.1 ppbV in WRF_Updated+BC. At the Bangkok Pandora site, mean bias shifted from strongly positive (+7.9×1015 molecules cm⁻²) to negative, and normalized mean errors decreased by ∼ 20 %–4 % in WRF_Updated and WRF_Updated+BC. Evaluation against GCAS airborne column retrievals further showed that the emissions updates improved spatial variability and magnitude of NO2 across the Bangkok urban plume, reducing mean and root mean square errors by ∼ 55 %–63 % in WRF_Updated and WRF_Updated+BC compared to WRF_Base. The DC-8 in situ vertical profiles further illustrated that WRF_Updated could substantially reduce near-surface overestimation in morning hours and preserve the observed vertical structure of NO2 mixing ratios.

Overall, these results demonstrate that incorporating geostationary satellite constraints into regional, high-resolution (4 km) chemical transport models can substantially improve the representation of urban air quality over Bangkok. Large biases in the baseline simulation are likely driven by reliance on outdated global emissions inventories (e.g., EDGAR v5), which do not reflect recent changes in regional anthropogenic activity, as well as uncertainties in bottom-up methodologies where updated local estimates appear systematically high. The satellite-constrained emission estimates derived here are consistent with independent top-down approaches (e.g., TCR-3) based on different satellite platforms and methodologies, increasing confidence in the inferred reductions.

Comparisons with independent aircraft and ground-based observations further indicate that the GEMS-constrained emissions presented here may represent an underestimate. In particular, the systematic low bias reported for the GEMS v3 NO2 product is consistent with the remaining negative biases observed in WRF_Updated. By accounting for this bias through the application of bias-corrected retrievals, WRF_Updated+BC reduces these negative biases, although in some cases introduces a tendency toward overprediction, highlighting tradeoffs associated with the correction. While model output (e.g., WRF_Base) can serve as a transfer standard for comparing different observing systems, these results highlight the importance of integrating ground-based, airborne, satellite, and model information to robustly identify biases and improve emission estimates.

A key strength of this framework is the use of hourly daytime constraints uniquely provided by geostationary observations, which enable direct characterization of daytime emission variability and plume evolution that cannot be captured by once-daily low-Earth-orbit measurements. These daytime constraints are particularly important for urban environments, where emissions, chemistry, and boundary-layer dynamics vary rapidly and strongly influence air quality impacts.

Coarse (∼ 12 km) simulations have been shown to inadequately represent circulations in coastal environments, and nonlinear NOx chemistry, leading to systematic biases in simulations of NO2 (Hsu et al., 2026; Valin et al., 2011; Verreyken et al., 2025; Yu et al., 2023). These limitations in model resolution are expected in coastal megacities such as Bangkok with complex local topography and land-sea contrasts. Consistent with recent TEMPO-based emission studies indicating that model resolution can limit the robustness of GEO-based top-down NOx constraints (Hsu et al., 2026), these considerations motivate the use of 4 km WRF-Chem simulations in this work as a necessary framework for accurately interpreting geostationary satellite observations and constraining urban NOx emissions.

While remaining discrepancies, including the negative model biases, are likely influenced in part by overpredicted model wind speeds and associated transport errors, future work could benefit from finer-resolution simulations (< 4 km) and more advanced urban parameterizations (e.g., multi-layer urban canopy models; Liu et al., 2025) to better represent urban flow, drag, and mixing. Continued validation and further development of bias-corrected GEMS retrievals will also strengthen the use of geostationary NO2 products. Nevertheless, the high-frequency daytime sampling provided by GEMS already offers critical information for emission inversion, and the integration of GEMS-derived constraints into WRF-Chem represents a scalable pathway toward near-real-time, satellite-informed emissions estimation and improved air quality forecasting for rapidly developing megacities.