We use the Weather Research and Forecasting model coupled with Chemistry (WRF-Chem) version 4.2.2 (Fast et al., 2006; Grell et al., 2005) to simulate NOx concentrations over Thailand. The model domain and configuration, including grid nudging, largely follow Thongsame et al. (2024). The MOZART-MOSAIC with aqueous phase chemistry (Emmons et al., 2010; Knote et al., 2014; Zaveri et al., 2008) scheme is used for chemical and aerosol schemes. This scheme includes interactions of the aerosols with radiation and clouds. The online MEGAN version 2.04 (Guenther et al., 2006) and offline FINN versions 2.5 (Wiedinmyer et al., 2023) are used to represent biogenic and biomass burning emissions, respectively. Anthropogenic emissions are taken from the CAMS-GLOB-ANT v5.3 inventory (Granier et al., 2019), and the initial and boundary conditions come from the Whole Atmosphere Community Climate Model (WACCM; Granier et al., 2019). Figure 1 shows the distribution of NOx emissions from CAMS-GLOB-ANT.

The simulation employs 42 vertical layers extending to approximately 63 hPa, with a horizontal resolution of 9×9km2. The model tropopause is located at approximately 100 hPa. To compare with GEMS, WRF-Chem NO2 concentrations are interpolated to GEMS vertical pressure levels. Tropospheric NO2 VCD from WRF-Chem is then calculated by integrating NO2 from the surface to the GEMS-defined tropopause at 230 hPa. The GEMS averaging kernel is applied to the model-simulated NO2 profiles to account for the satellite's retrieval sensitivity, following previous approaches (Cooper et al., 2020; Park et al., 2024). The averaging kernel accounts for the sensitivity of the satellite and is applied to the model to mitigate uncertainties in the retrieval algorithm. Further details regarding the averaging kernels and the applied equations are provided in Sect. S6 in the Supplement.

GEMS is an ultraviolet-visible instrument that measures back-scattered solar spectra in the 300–500 nm wavelength range (Kim et al., 2020). It is a geostationary orbit satellite, which was launched in February 2020, with a resolution of 3.5×8km2 (Kim et al., 2023). The GEMS data span from November 2022 to the present, covering a geographical domain from 75 to 145° E and 5° S to 45° N (Park et al., 2024). Tropospheric NO2 VCDs are retrieved using a Differential Optical Absorption Spectroscopy (DOAS) algorithm, applied over the 432–450 nm spectral window (Kim et al., 2023). Several studies have demonstrated strong correlations between GEMS NO2 and ground-based measurements in regions such as China and South Korea (Kim et al., 2023; Li et al., 2023b; Seo et al., 2025).

For this research, we utilize version 3 of the GEMS Level 2 tropospheric NO2 VCD (https://nesc.nier.go.kr/en/html/index.do, last access: 27 December 2024) in September 2023 for the Thailand area. For September 2023, the dataset provides up to 10 consecutive hourly snapshots of tropospheric NO2 VCD during daylight, from 00:45 to 07:45 UTC, with additional observations at 22:45 and 23:45 UTC.

For this analysis, we select data between 00:45–04:45 and 06:45–07:45 UTC (07:45–11:45 and 13:45–14:45 LT), which collectively cover over 50 % of Thailand. As the GEMS scan at 05:45 UTC does not extend to Thailand, the 05:45 UTC data are linearly interpolated from adjacent hours (04:45 and 06:45 UTC) for updating the emission with the mass balance method. To ensure data quality, we retain only pixels with a cloud fraction below 0.3 (Park et al., 2024). We also filter data based on the final algorithm quality flag, a metric representing the reliability of each pixel's retrieval. Although a flag of 0 indicates an ideal measurement from the perspective of algorithmic issues, a systematic surface reflectance issue in GEMS version 3 prevents most pixels from achieving this score. Consequently, we use pixels with a quality flag of 0 or 1, ensuring sufficient data coverage while excluding severely degraded retrievals. Additionally, we have found striping artifacts present in the GEMS data (Fig. 2a). To ensure data quality and improve the reliability of the emission adjustment process, we manually remove these from our analysis. The data coverage in North Thailand is limited, and retrieval errors are relatively high in both North Thailand and the Bangkok Metropolitan Region (Fig. 2b and c). In GEMS version 3, the tropopause pressure level is assumed to be fixed at 230 hPa across all pixels. As GEMS does not publicly release full a priori model profiles required to change the tropopause in the algorithm, we are constrained to adopt the 230 hPa definition to ensure consistency between the satellite retrieval and our WRF-Chem model. Our sensitivity test using WRF-Chem reveals that a fixed tropopause at 230 hPa and the model's varying tropopause have a negligible impact during the daytime. This is because the NO2 in Thailand during September is overwhelmingly dominated by near-surface emissions, and the NO2 beyond 230 hPa is almost zero everywhere in Thailand.

TROPOMI (TROPOspheric Monitoring Instrument) is a nadir-viewing UV–Visible spectrometer aboard Sentinel-5P that measures solar radiation backscattered from the Earth to retrieve atmospheric trace gases such as NO2, O3, SO2, HCHO, CH4, and CO (Judd et al., 2020; Veefkind et al., 2012). In this study, we use the Level 2 tropospheric NO2 VCD product from TROPOMI version 2, which represents the NO2 amount between the surface and tropopause. The TROPOMI tropospheric NO2 retrieval uses a DOAS technique in the 405–465 nm spectral window, and its horizontal resolution is approximately 3.5×5.5km2 (Judd et al., 2020; Zhao et al., 2020).

For comparison with WRF-Chem simulations, we filter the TROPOMI data to retain only retrievals with a quality assurance value (qa_value) greater than 0.5, thereby excluding pixels affected by clouds, snow/ice, or retrieval artifacts. To ensure consistency in vertical sensitivity between the satellite and model data, we apply the TROPOMI-provided averaging kernels to the model-simulated NO2 profiles. The TROPOMI satellite provides data around 10:30 am–01:30 pm LT for Thailand. The model outputs at the closest time to TROPOMI are used for comparison.

For ground-based monitoring data, we use the data from the Pollution Control Department (PCD) of Thailand air quality stations (http://air4thai.pcd.go.th/webV3/, last access: 20 December 2024) to evaluate surface NOx concentrations from WRF-Chem. The PCD stations' measurements are based on the chemiluminescence technique, which detects NOy species (Pollution Control Department, 2022; Zhang et al., 2009). Hence, we compare WRF-Chem NOy (which includes NOX and its oxidation products) with NOx from PCD. The PCD stations (20) are classified into roadside (14) and non-roadside stations (6) as shown in Fig. S14 in the Supplement. On average, the simulated NOy concentrations from WRF-Chem is higher than NOx by 5 % at the roadside stations and 6 % at the non-roadside stations.

Although IFDMB assumes negligible horizontal transport due to the short atmospheric lifetime of NO2, our high-resolution simulations (9×9km2) violate this assumption, as there may be significant horizontal transport of NO2 to surrounding grid cells. To mitigate the smearing error, we regrid both GEMS and WRF-Chem NO2 VCDs to a coarser 45×45km2 resolution by spatially averaging over surrounding grid cells. To update NOx emissions, we apply IFDMB in three different configurations: Fixed diurnal, Daylight diurnal, and Temporal average (Table 1). Each approach is described below. For each method, we use some form of a monthly average of the GEMS data to update emissions. This helps reduce the impact of noise and missing data in satellite observations. Additionally, the smearing error is minimized when considering monthly average NO2 VCDs.

Three metrics, namely, normalized mean bias (NMB), normalized root mean square error (NRMSE), and correlation coefficient (R), are used to evaluate the performance of the IFDMB solution in comparison to our simulation using the prior emissions. The observation error (σo) provided by GEMS and TROPOMI data is used to normalize mean bias (MB) and root mean square error (RMSE). We compare our results with tropospheric NO2 VCD from GEMS and TROPOMI, and with in situ NOx measurements from PCD stations. For GEMS and TROPOMI, the monthly average tropospheric NO2 VCDs in the Thailand grid cells are compared with the monthly average WRF-Chem results. This is because only emissions in Thailand are updated, while emissions outside Thailand remain the same. For the satellite data, our metrics are normalized with the observation error. However, for PCD stations, we evaluate our results with non-normalized MB, non-normalized RMSE, and R. This is because the PCD stations have not reported the observation error. Additionally, we compare the monthly average diurnal NOx profile with PCD data.

This section shows the performance of the WRF-Chem simulation using the prior emissions (the CAMS-GLOB-ANT inventory), aka the baseline simulation, compared to GEMS NO2 VCDs before applying any mass balance methods. Overall, the baseline simulation significantly overestimates the monthly average NO2 VCDs in the Bangkok Metropolitan Region and East Thailand (Fig. 4). The baseline simulation shows that the NO2 VCD is dominated by large point sources from many sectors. In North Thailand, the prior emissions identify the Mae Moh coal power plant (energy sector) as the main source of NO2, with a signal that significantly overwhelms the urban emissions from Chiang Mai and Lampang (Fig. 1). In the Bangkok Metropolitan Region, the model suggests a mixed contribution from the dense residential and transportation sectors of the urban area. Additionally, industrial estates in Saraburi and Rayong produce high NO2 – as strong as urban emissions in Bangkok. GEMS presents a contrasting picture that emphasizes the role of urban sources. In North Thailand, GEMS has the highest NO2 VCDs centered over the urban areas of Chiang Mai and Lampang, which are dominated by residential and transportation emissions. The Mae Moh power plant appears as only a moderate signal in the GEMS observations. This pattern holds in the Bangkok Metropolitan Region, where GEMS again highlights the urban area, while the signals from major industrial estates in Rayong and Saraburi are notably weaker than the urban emissions. Additionally, the GEMS NO2 VCDs in North Thailand are as large as in the Bangkok Metropolitan Region, while the baseline simulation shows significantly larger NO2 VCDs in the Bangkok Metropolitan Region. This direct contradiction between the baseline simulation and GEMS highlights that the CAMS-GLOB-ANT is likely to overestimate energy and industrial emissions relative to the residential and transportation emissions in Thailand. The baseline simulation shows high NMB and NRMSE in most regions, with comparatively better performance in North and West Thailand (Figs. 5 and S15). However, GEMS data availability is high in Central Thailand and the Bangkok Metropolitan Region, moderate in the North and East, and very limited in West Thailand (Fig. 2). When averaged across all grid cells in Thailand, the baseline simulation yields an NMB of 83.52 and an NRMSE of 151.94 relative to GEMS. These metrics quantitatively support the widespread overestimation of modeled NO2, though the correlation coefficient across the domain remains moderate (R=0.65). In the Bangkok Metropolitan Region, the baseline simulation shows an even more pronounced overestimation, with an NMB of 278.67 and an NRMSE of 336.69. Despite these high biases, the correlation coefficient in this area is exceptionally strong (R=0.96). This suggests that while the baseline NO2 spatial distribution strongly agrees with the satellite, the absolute VCDs of the prior emissions in this region are severely overestimated.

To determine the optimal number of iterations for the IFDMB inversion, we calculate the total cost function J, Jprediction, and Jparameter based on comparisons between NO2 VCDs from WRF-Chem and GEMS at each iteration (Fig. 6). Our results show that J decreases substantially during iterations 1 and 2 across all update schemes, but it begins to increase beyond iteration 3 (Fig. 6a). This increase in the total cost function is primarily driven by a rise in Jparameter (Fig. 6c). The Jprediction term slightly decreases after iteration 2. However, Jprediction does not decrease monotonically due to overadjustment; it may increase in some iterations before decreasing again in subsequent steps (Fig. 6b). There are several reasons supporting the rapid convergence within 2–4 iterations. The iteration in IFDMB accounts for horizontal transport and the nonlinear chemistry of NO2. Regarding transport, we reduce the smearing error by regridding both GEMS and WRF-Chem NO2 VCDs to a coarser resolution. We also use monthly averages of GEMS and WRF-Chem, which reduce the impact of noise and missing data in the satellite observations and further minimize the smearing error. Together, these steps largely confine NO2 transported out of a source grid cell to the same coarse grid cell. The transport effect that iterations would otherwise need to resolve is therefore absorbed before the inversion begins, which is why only few iterations are required. Additionally, the GEMS observation error is on the order of 1013 (Fig. 2c), whereas the GEMS NO2 VCD data is on the order of 1015 (Fig. 2a). This large difference in magnitude results in a dominant Jprediction term in Eq. (1), which drives a substantial emission adjustment during the very first iteration, followed by much smaller adjustments in subsequent iterations. Furthermore, our regularization parameters prevent over-adjustment and avoid oscillatory fluctuations during convergence. In this study, we assume a 50 % relative uncertainty for the prior emissions error. However, the true uncertainty for specific sectors like power plants and industry may be lower due to direct emissions reporting. This large assumed emission uncertainty yields a relatively small Jparameter penalty term, giving the IFDMB algorithm the flexibility to make large initial emission adjustments to reconcile the model with the GEMS observations. Our analysis shows that the daylight diurnal scheme exhibits significant overadjustment because it assumes an immediate response of NO2 VCDs to emission changes. For the fixed diurnal and daylight diurnal update schemes, iteration 2 provides the lowest J and is thus selected as optimal. For the temporal average update scheme, iteration 4 yields the optimal performance.

Additionally, we performed a sensitivity analysis of the regularization parameters in the IFDMB framework, using a fixed diurnal update, to evaluate the impacts of C1 and C2. We found that varying these parameters alters the results significantly – yielding differences of ∼30 % to 60 % – with the model showing particular sensitivity to C1 (Table S1 in the Supplement). Increasing C1 from 1×1014 to 2×1015 molec.cm-2 limits the influence of satellite constraints, resulting in higher NMB and NRMSE values, particularly in Central and East Thailand (Figs. S1 and S2). A low value for C1 permits large emission adjustments across all of Thailand, including regions with high GEMS retrieval errors, such as North Thailand (Fig. S3). Conversely, a higher C1 value restricts these adjustments. In North Thailand, where retrieval uncertainty is high, increasing C1 benefits the IFDMB by preventing the model from overfitting to highly uncertain GEMS data. However, this same increase has a detrimental impact on the Bangkok Metropolitan Region. Because uncertainty is comparatively low here, a high C1 limits emission improvements. Regarding C2, increasing its value from 1 to 10 molkm-2h-1 has minimal impact when C1 is low.

This is because a low C1 permits aggressive emission adjustments nationwide; variations in C2 become negligible. The impact of C2 is much more pronounced when C1 is high. C2 primarily governs adjustments for smaller emission sources in rural areas. Consequently, a high C2 allows for greater adjustments in these low-emission zones, which leads to overfitting to the GEMS data in the rural North. A low C2 leads to underfitting in Central Thailand and the Bangkok Metropolitan Region. Overall, the adjustable parameters in the IFDMB offer valuable flexibility to manage the mismatch between GEMS and TROPOMI by moderating emission updates in high-error regions and balancing prior and posterior estimates. In this study, we applied fixed regularization parameters across the entirety of Thailand. However, our findings indicate that a large C1 benefits high-uncertainty areas like North Thailand, whereas a low C1 benefits the Bangkok Metropolitan Region and Central Thailand. This highlights the potential of using spatially varying parameters based on regional uncertainty rather than fixed domain-wide values. Nevertheless, because we aim to treat these parameters as fixed statistical values rather than tuned parameters, an exhaustive fine-tuning exploration is beyond the scope of this paper.

This section presents the results from WRF-Chem simulations using the IFDMB method evaluated with GEMS. The IFDMB with all three update schemes reduces both NMB and NRMSE in GEMS compared to the baseline simulation (Fig. 5). The IFDMB with fixed diurnal update demonstrates the highest improvement among the three update schemes. The NMB and NRMSE reduce by about 60 % for the fixed diurnal update averaged across all grid cells in Thailand. Reductions are most pronounced in Central Thailand and the Bangkok Metropolitan Region (Figs. 5 and S15). In Bangkok Metropolitan Region, the IFDMB with fixed diurnal update reduces NMB to 29.46 and NRMSE to 51.93. The correlation coefficient remains the same as 0.96. Overestimated NO2 VCDs from transportation and residential sectors are substantially corrected. However, improvements in North and West Thailand are more modest. The improvement in those areas is limited due to the low data density and high errors in those areas. The correlation coefficient (R) improves from 0.65 (baseline simulation) to 0.84.

The IFDMB with daylight diurnal and temporal average updates produces similar results. Both methods improve the performance in urban areas, particularly in Bangkok and big cities in North Thailand, but are less effective in other areas (Figs. 5 and S15). The NMB is reduced by 14 % for the IFDMB with daylight diurnal update and 9 % for the temporal average update averaged across all grid cells in Thailand. However, within the Bangkok Metropolitan Region, these two schemes show lower performance. Specifically in the Bangkok Metropolitan Region, the NMB decreases to 160.26 under the daylight diurnal update and to 110.52 under the temporal average update. Similarly, the NRMSE in this region drops to 214.43 and 158.45 for the daylight diurnal and temporal average updates, respectively. Despite these remaining absolute biases, the correlation coefficient R remains exceptionally strong at approximately 0.96 for both methods. Notably, both updates exhibit small improvements during early morning hours (e.g., 08:00–09:00 am LT) compared to the IFDMB with fixed diurnal update (Fig. S16). Since these methods only update emissions during the daylight, uncorrected nighttime emissions likely contribute to the error of NO2 VCD in the early morning. In contrast, the fixed diurnal update, which adjusted emissions at all time steps, shows lower NMB and NRMSE throughout the day. The NRMSE reduction for these two methods is around 20 % and correlation coefficients improve to 0.74 for the daylight diurnal update and 0.78 for the temporal average update.

The use of GEMS satellite data to update NOx emissions with the IFDMB method provides an increase in emissions in the North and West Thailand, while emission reductions are obtained in other areas (Fig. 7). The fixed diurnal update produces the largest emission adjustments, as it applies the scaling factor uniformly across all time steps, including periods with and without GEMS observations (Fig. 8). In this case, daily NOx emissions increase approximately 20 molekm-2h-1 in North and West Thailand, while they decrease by about 30 molekm-2h-1 in other regions (Fig. 7). The IFDMB with daylight diurnal update, which only modifies emissions during hours with available GEMS data, results in the smallest overall changes. Increases and decreases are generally below 20 molekm-2h-1 across all regions (Fig. 7). The IFDMB with temporal average update produces similar results to the daylight diurnal update. However, the emission reduction in the Bangkok Metropolitan Region is stronger than the daylight diurnal update, with the reduction of about 30 molekm-2h-1. A notable feature across all update schemes is the significant amplification of NOx emissions in Lampang. This enhancement arises because the CAMS-GLOB-ANT inventory lacks representation of large point sources in that area, whereas GEMS indicates elevated NO2 VCDs. Consequently, the IFDMB compensates for the discrepancy between simulated and observed NO2 by intensifying the small prior emissions in Lampang.

For the diurnal variability, prior emissions follow almost the same diurnal profile across all regions in Thailand, depending on the dominant sectors. However, GEMS-informed updates reveal differences in diurnal variability across Thailand (Fig. 8). In Bangkok and North Thailand, the prior profile has a sharp morning peak between 08:00–09:00 LT, followed by a flat midday period extending to 16:00 LT. In contrast, both the IFDMB daylight diurnal and temporal average updates exhibit more dynamic variability, with pronounced peaks between 09:00–11:00 LT and a subsequent decline throughout the GEMS observation window in Bangkok. In North Thailand, the IFDMB daylight diurnal update peaks around 12:00–14:00 LT, while the temporal average update peaks earlier, around 09:00–11:00 LT. Our study reveals that the diurnal variability from Thongsame et al. (2024), which shows peaks aligned with morning and evening rush hours, is different from the GEMS diurnal variability. GEMS provides the potential to refine the diurnal variability during the daylight in Thailand. However, GEMS tends to underestimate NO2 VCDs in the early morning and late afternoon, due to unfavorable viewing geometries during those times (Park et al., 2024). Future research is required to analyze this uncertainty and its impacts on diurnal variability.

In this section, we compare tropospheric NO2 VCDs from the WRF-Chem simulations and GEMS with those from TROPOMI (Fig. 9). Because GEMS and TROPOMI differ in their averaging kernels and overpass times, a direct comparison is not feasible. Therefore, we focus on assessing the relative spatial distributions of NO2. To compare the NO2 VCD between GEMS and TROPOMI, WRF-Chem is used as the intercomparison platform, following the methodology of Zhang et al. (2010). This approach applies the AKs from each satellite to the same WRF-Chem outputs to assess how the distinct vertical sensitivities of the instruments differ relative to the model (Fig. 9b, d, and f). To ensure consistency, the GEMS and WRF-Chem VCDs closest to the TROPOMI overpass time are selected for comparison (Fig. 9c and d). The differences between the GEMS and TROPOMI overpass times are about 15 min. The monthly average GEMS NO2 VCDs at the TROPOMI overpass times are slightly higher than the monthly average NO2 derived from all GEMS observations (Fig. 9a and c). Moreover, they exhibit increased noise, particularly over rural areas. Overall, both satellites show similar spatial patterns of NO2, except in North Thailand. High NO2 is consistently observed in the Bangkok Metropolitan Region, Saraburi, and Rayong from both GEMS and TROPOMI. In North Thailand, both satellites detect high NO2 associated with the Mae Moh power plant; however, GEMS shows anomalously high NO2 over Lampang that is not observed by TROPOMI (Fig. 9c and e). The NO2 levels in Lampang from GEMS exceed those in Bangkok, whereas TROPOMI indicates the highest NO2 over the Bangkok Metropolitan Region. Additionally, TROPOMI generally has higher background NO2 levels compared to source regions. To further examine the magnitude differences, we apply the respective averaging kernels of GEMS and TROPOMI to the baseline simulation. The results indicate that TROPOMI retrieves higher NO2 VCDs than GEMS across most regions, except in Lampang, where GEMS values remain substantially larger (Fig. 9d and f).

Our baseline simulation overestimates tropospheric NO2 VCDs across Thailand relative to TROPOMI satellite observations (Fig. 9e and f). The model predicts strong NO2 VCD from energy, industry, transportation, and residential sectors almost everywhere, whereas TROPOMI detects high NO2 only in specific areas. In particular, TROPOMI shows high NO2 only over the Bangkok Metropolitan Region (attributed largely to residential emissions) and at certain major point sources (e.g., the Mae Moh power plant in the North and large industrial zones in Central Thailand). Most other areas, including city centers in North and Northeast Thailand, exhibit only moderate NO2 VCDs in the TROPOMI data. Additionally, TROPOMI shows moderate to low NO2 along the major roads in Thailand, whereas the baseline simulation shows clear NO2 enhancements. The monthly NMB from the baseline simulation is 0.8 (Figs. 10 and S17). This emphasizes the overestimation of TROPOMI NO2 VCDs with the baseline simulation. The monthly NRMSE is 1.21 (Fig. 10). We have found that the baseline simulation is moderately correlated with the monthly average TROPOMI (R=0.79).

The IFDMB with fixed diurnal update significantly reduces NMB and NRMSE for TROPOMI, except in North Thailand. The NMB drops by roughly 25 % relative to the baseline simulation, and NRMSE also declines by about 29 % (Fig. 10). Because the baseline simulation overpredicts NO2 in nearly all locations, the emission reductions from the IFDMB with fixed diurnal update lead to a closer match with TROPOMI across most of Thailand. For example, in the Bangkok Metropolitan Region and East Thailand, the IFDMB with fixed diurnal update significantly improves the NMB and NRMSE, while moderately reducing NMB and NRMSE in Central and Northeast Thailand (Figs. 10 and S17). In these areas, the NO2 VCD from transportation and residential sectors is decreased, aligning better with the TROPOMI levels.

In contrast, the IFDMB with fixed diurnal update worsens the agreement with TROPOMI in North Thailand by increasing the NMB and NRMSE. This degradation is primarily due to the discrepancy between GEMS and TROPOMI NO2 observations (Fig. 9c and e). While TROPOMI shows relatively low NO2 VCDs in North Thailand compared to the Bangkok Metropolitan Region, GEMS has high NO2 VCDs in the North. Moreover, the baseline simulation already overestimates NO2 VCDs in North Thailand relative to TROPOMI (Fig. 10a). As a result, the IFDMB – driven by GEMS – further increases NOx emissions in the North, exacerbating the overestimation when compared to TROPOMI (Fig. 10b). This leads to a noticeable decline in the overall WRF-Chem–TROPOMI correlation, dropping to R≈0.61 (Fig. 10b). However, in the Bangkok Metropolitan Region, the update slightly improves the agreement, with monthly correlation coefficients increasing from 0.90 to 0.93 (Table S4).

The daylight diurnal and temporal average updates exhibit spatial trends similar to the fixed diurnal update, with improvements in most regions but degradation in North Thailand (Figs. 10 and S17). Across all grid cells, these two schemes do not yield net improvements in overall NMB and NRMSE (Fig. 10). The daylight diurnal update slightly increases NMB by 5 % and reduces NRMSE by 2 % relative to the baseline simulation. In contrast, the temporal average update increases NMB and NRMSE by 31 % and 17 %, respectively. The apparent deterioration in national-scale performance primarily stems from the mismatch between GEMS and TROPOMI in North Thailand.

To isolate this effect, we evaluate model performance in the Bangkok Metropolitan Region, where GEMS and TROPOMI show strong consistency (Table S4). In this region, the daylight diurnal update reduces NMB and NRMSE by 54 % and 39 %, respectively, compared with the baseline simulation (Table S4). The temporal average update yields even larger improvements, reducing NMB by 81 % and NRMSE by 61 % (Table S4). However, this update scheme performs the poorest in North Thailand. The temporal average update shows stronger emission adjustment near TROPOMI overpass time (09:00–11:00 LT) compared to the daylight diurnal update. This leads to more impact on the model performance with TROPOMI.

Overall, the IFDMB with all three update schemes reduce the model bias and error in areas such as the Bangkok Metropolitan Region, Saraburi, and Rayong industrial estates. This is because TROPOMI's spatial pattern resembles GEMS, with high NO2 VCDs over the Bangkok Metropolitan Region and moderate VCDs over the Saraburi and Rayong industrial estates. However, improving model performance in North Thailand remains a challenge because GEMS and TROPOMI exhibit substantial differences in NO2 VCDs over this region. TROPOMI NO2 VCDs in North Thailand align more closely with the baseline simulation, having a strong signal from the Mae Moh power plant that is higher than the surrounding urban areas. In contrast, GEMS presents the highest NO2 VCDs over the Lampang urban area instead. The inconsistency between GEMS and TROPOMI in North Thailand is a crucial finding.

Our analysis indicates that the GEMS–TROPOMI discrepancy primarily arises from uncertainties in the prior model profile used by the GEMS retrieval algorithm. TROPOMI version 2 uses the TM5-MP global model with the MACCity emission inventory for its retrieval algorithm (Douros et al., 2023; Williams et al., 2017). GEMS version 3 employs the GEOS-Chem model with the ASIA-AQv3 emission inventory as its prior input for NO2 retrievals. Our analysis shows that the ASIA-AQv3 inventory reports anomalously high NOx emissions in the Lampang region compared to other inventories (Fig. S6). This high NOx emission likely causes GEMS to overestimate NO2 VCDs over Lampang which is not observed by TROPOMI and the baseline simulation. The high NOx emissions from ASIA-AQv3 also result in significantly larger retrieval errors in the Lampang area (Fig. 2). Supporting this interpretation, GEMS NO2 using the Direct Vertical Column Fitting (DVCF) algorithm from Yang et al. (2024) yields much lower NO2 VCDs over Lampang, in line with the TROPOMI observations. Notably, the GEMS version 3 data used in this study do agree well with TROPOMI over the Bangkok Metropolitan Region, Saraburi, and Rayong. Detailed comparisons are provided in Sects. S2 and S3.

These differences between satellite products highlight the importance of bias correction, calibration, and cross-validation for satellite-based NO2 measurements. For example, Kim et al. (2023) also find that both GEMS and TROPOMI underestimate NO2 VCDs compared to Pandora measurements in Seosan, South Korea, with GEMS exhibiting a greater degree of underestimation than TROPOMI. The fact that GEMS and TROPOMI can yield markedly different NO2 (as seen in Thailand) underscores the need to calibrate and validate these satellite retrievals using independent observations in Thailand. However, achieving robust calibration in Thailand is challenging due to the scarcity of ground-based measurements. In 2023, only a single Pandora spectrometer was operational in Bangkok. Data from this instrument exhibit substantial uncertainty during the Thai winter season (October–January) (Kanchanachat, 2026), and there is no cross-validation by other instruments. Furthermore, compared to Korea, China, and Japan, the Bangkok Pandora site reports significantly lower NO2 VCDs relative to GEMS (Jung et al., 2025). Consequently, we exclude its observations from our analysis. We recommend leveraging data from the ASIA-AQ field campaign to improve the accuracy and consistency of satellite NO2 assessments in Thailand. Even though this field campaign is limited to the Bangkok and Chiang Mai areas and spans only a short time period, it provides useful information to resolve the biases in the satellite products.

Hourly observations from PCD stations are used to evaluate how different emission update schemes affect the simulated diurnal variability of NOx. As the GEMS hourly scanning times occur between 08:00 and 15:00 Thailand local time, model–observation comparisons are limited to daylight hours. We restrict the comparison to 08:00–15:00 LT period to ensure a fair evaluation, as only the fixed diurnal scheme adjusts emissions outside this period, while the others revert to the prior emissions. The PCD stations are mostly located in Central and Northeast Thailand, and the Bangkok Metropolitan Region (Fig. S14). The baseline simulation systematically underestimates monthly average NOx mixing ratios by 10–15 ppbv at 14 roadside PCD sites throughout the day and underestimates daytime NOx by 3 ppbv or less at six non-roadside PCD sites (Fig. S18). The IFDMB with GEMS-based top-down emission adjustment exacerbates the underestimation of NOx mixing ratios at roadside PCD stations (Fig. S18), although it provides the same magnitude of concentration for non-roadside PCD stations. All inversion schemes lead to an increase in bias and error at roadside stations. These discrepancies arise primarily from differences in spatial resolution between the GEMS (∼3.5×8km2), WRF-Chem (9×9km2), and PCD stations (point-scale), as well as strong local contributions of transportation emissions at the PCD sites. Moreover, both GEMS observations and WRF-Chem simulations are regridded to 45×45km2 for the mass balance calculation in this study to reduce smearing errors associated with horizontal transport. However, this introduces resolution errors by averaging over spatially heterogeneous emission sources. This grid-averaged value is heavily influenced by regional NO2 pollution and is therefore much lower than the NO2 mixing ratio from local hotspots, such as a busy roadway, that may occupy only a small fraction of that grid cell. Consequently, we do not focus on reproducing absolute NOx mixing ratios in this study but instead assess changes and improvements in diurnal patterns as shown in NOx concentrations normalized by their monthly average value (Fig. 11).

Observed NOx mixing ratios typically peak near 08:00 LT, during the morning rush hour, whereas the baseline simulation and WRF-Chem simulations using posterior emissions have an earlier peak between 06:00 and 07:00 LT. Furthermore, the model has higher NOx levels at night, while observations have high NOx during the day. This discrepancy indicates that the diurnal profile from Thongsame et al. (2024) used in the baseline simulation does not fully capture the real-world traffic patterns for NOx. In particular, the default diurnal profile fails to reproduce the morning traffic peak and instead shows a high level of NOx during nighttime. The fixed diurnal update scheme does not improve the temporal correlation between modeled and observed NOx mixing ratios at roadside or non-roadside stations (Fig. 11). As this update scheme does not alter the shape of the diurnal profile, the simulated NOx peak continues to occur at night, leaving the morning underestimation unresolved. The daylight diurnal and temporal average updates also show limited improvement in diurnal variability (Fig. 11). At roadside stations, both produce afternoon NOx diurnal variability closer to that of the PCD data but fail to reproduce the morning peak. At non-roadside stations, both schemes capture similar morning diurnal variability yet diverge in the afternoon. The temporal average update yields a slight mid-day dip, whereas the daylight diurnal update produces a modest increase (Fig. 11). The PCD data exhibit a small increase around 14:00 LT and a drop by 15:00 LT, which is more consistent with the daylight diurnal update. Overall, the GEMS-based diurnal variability does not capture the observed morning rush-hour but slightly improves afternoon diurnal variability. However, using the GEMS satellite to only update the diurnal variability introduces an unrealistic diurnal profile in the model – an abrupt increase and drop in the NOx emissions profile (Fig. 8). We recommend implementing a smoothing technique in future work to mitigate such discontinuities. In this study, we applied a 5-point moving average to the posterior emissions from the IFDMB with the daylight diurnal update (Sect. S4). While this smoothing technique has a small impact on absolute NOx concentrations, it yields a more physically realistic NOx concentration profile by mitigating the abrupt changes in the transition zones bordering the GEMS observation window. In summary, while the fixed diurnal update does not improve model–observation agreement due to its unchanged emission profile, the daylight diurnal and temporal average updates offer some potential for spatially resolved diurnal refinements. However, the lack of satellite observations during critical rush-hour periods (early morning and evening) limits their effectiveness. Although GEMS-based updates alone do not substantially improve NOx simulations at PCD stations due to resolution and timing mismatches, the satellite provides valuable information for refining regionally specific diurnal patterns, which could enhance WRF-Chem performance in future studies.

Our findings also underscore the tradeoffs in the application of mass balance methods. The IFDMB approach, when applied at the 45×45km2 resolution, offers computational efficiency and is well-suited for adjusting regional-scale emissions over long temporal periods. However, its coarse spatial resolution limits its ability to capture fine-scale urban variability and accurately constrain emissions from local sources such as traffic. Moreover, reducing or removing the aggregation for IFDMB provides no significant difference in roadside stations (Sect. S5). For non-roadside stations, the 45×45km2 resolution performs better than 27×27km2 resolution and 9×9km2 resolution (Fig. S11). In contrast, other inversion techniques – such as 4D-Var or ensemble Kalman filter (EnKF) methods – coupled with high-resolution models, would be more effective for resolving local emissions. However, these approaches are computationally expensive and may not be feasible for long-term simulation. Therefore, the IFDMB with GEMS data provides practical and efficient adjustments to regional NOx emissions in Thailand.

This section provides an evaluation of the performance and inherent trade-offs of the different update schemes used in our study. The IFDMB method is a practical advancement of the mass balance approach for emission adjustment by integrating satellite and prior information. In our study, regularization parameters C1 and C2 are introduced to balance constraints between the prior emissions and GEMS. These parameters control the trade-off between overfitting and underfitting to the satellite data. For example, the discrepancy between GEMS and TROPOMI in North Thailand indicates high uncertainty in the satellite observations. In this region, a higher C1 value, which affects the prediction error term, allows IFDMB to place less weight on the GEMS data, reducing the risk of over-adjustment in areas with more uncertain observations. Although our study does not fully evaluate the sensitivity of IFDMB to C1 and C2, preliminary tests with varying parameter values are presented in Sect. S1.

We find that with C1=1×1014, IFDMB can partially accommodate the uncertainty in the GEMS data due to ASIA-AQv3 inventory used for the GEMS retrieval. Our IFDMB still leads to increased bias and error for WRF-Chem in North Thailand when compared with TROPOMI, particularly for the daylight diurnal and temporal average update schemes. These schemes are more uncertain than the fixed diurnal update because of fewer available data points in each update. The C2 parameter, which affects the emissions error term, helps prevent IFDMB from excessively adjusting small emissions, but our results indicate that C1 has a stronger influence on IFDMB performance than C2. The overall performance of IFDMB also depends on the relative uncertainties from GEMS and the prior emissions. In this study, we apply a fixed uncertainty of 50 % to the prior emissions, which may not reflect the true variability in real-world conditions. Additionally, the GEMS error is lower than the error reported in TROPOMI. Therefore, future work should focus on developing more accurate, region-specific estimates of both satellite and prior emission uncertainties to further improve IFDMB performance and reliability.

For the update schemes, all three schemes for IFDMB offer improvements over the prior simulation when evaluated against GEMS data. However, each scheme involves trade-offs in how it treats diurnal variability, and these trade-offs determine which scheme is best suited to a given application. For applications focused on monthly-average NO2 improvement or on updating emissions at the national or regional scale where the diurnal cycle matters less than spatial accuracy, we recommend the fixed diurnal update. This scheme applies a uniform monthly-averaged correction factor to emissions at all hours and produces the most substantial improvements against GEMS observations. However, this method could not correct the diurnal variability of emissions. The ground-based PCD data clearly show that the baseline simulation misrepresents the diurnal cycle, with the model peaking at night instead of during morning and evening rush hours. The high NOx concentration during the nighttime from the model is likely due to the low planetary boundary layer at that time (Thongsame et al., 2024). The fixed diurnal update simply scales down the incorrect diurnal profile, without improving the diurnal variability. We find that it lowers the correlation at 08:00 LT from 0.70 to 0.60 at roadside stations and from 0.70 to 0.66 at non-roadside stations. Even so, the fixed diurnal update remains a simple and effective method, as it is the only scheme able to update emissions during both daytime and nighttime using monthly-average data; because GEMS only provides information during daylight hours, the other two schemes must leave nighttime emissions unchanged from the prior.

The daylight diurnal and temporal average updates are designed to address diurnal issues by incorporating hourly GEMS data. These schemes only adjust emissions during daylight hours, resulting in more modest improvements in GEMS agreement. Their key weakness, however, is the introduction of an unrealistic diurnal profile. Because the algorithms have no observational constraints outside the GEMS timing window, they revert to the prior profile at night. This is due to the limitations of our update schemes. For studies focused on daytime hourly NOx concentrations and diurnal variability, we recommend the daylight diurnal update scheme. Because the daylight diurnal and temporal average schemes yield no significant difference in performance, we favor the daylight diurnal update for its simpler implementation. It improves the correlation at 08:00 LT from 0.70 to 0.75 at non-roadside PCD stations, while the correlation at roadside stations remains unchanged at 0.70. By applying the smoothing technique, the impact of the unrealistic diurnal profile would be reduced. Unlike the fixed diurnal update, it can adjust the diurnal profile region by region rather than applying a single uniform profile across the domain, though this benefit remains limited by GEMS's spatial resolution. The challenge of this scheme is how to refine emissions outside the GEMS coverage period, as no UV/Vis satellite observations are available at night. Future research should consider integrating additional datasets – such as from the recent ASIA-AQ field campaign data, Pandora spectrometer measurements, and PCD stations – with satellite observations to develop accurate diurnal profiles. Nevertheless, GEMS provides valuable information for constraining daylight diurnal variability across different regions of Thailand.

We compare our results with another study using GEMS to adjust regional emissions. Our baseline simulation shows consistent results with Park et al. (2024) when comparing monthly average NO2 VCDs with GEMS data. Specifically, in our September 2023 simulation, the baseline simulation overestimates NO2 VCDs across most of Thailand, except in the North. Similarly, using the CMAQ model for March–April 2022, Park et al. (2024) report an overestimation of NO2 VCD compared to GEMS across all regions of Thailand, including the North. The key difference in the North is due to the dominant emission sources during the two study periods. In Park et al. (2024), the simulation period is during the haze season, when biomass burning emissions mainly contribute to NOx in North Thailand during March–April (Lalitaporn, 2017; Oo et al., 2021). In contrast, our study period in September 2023 corresponds to the non-haze season, when biomass burning activity is minimal. Therefore, the underestimation in North Thailand, in our case, is likely attributable to anthropogenic emission inventories rather than biomass burning. Additionally, Park et al. (2024) use GEMS version 2 for emission adjustment, while our study utilizes GEMS version 3, which may also contribute to differences.

Regarding top-down emission updates, our IFDMB produces similar regional trends to Park et al. (2024) across all update schemes. In both studies, posterior NOx emissions decrease in the Bangkok Metropolitan Region, Central, and Northeast Thailand. However, our study shows an increase in posterior NOx emissions in North Thailand, while Park et al. (2024) report a decrease in this region. This contrast again likely reflects seasonal differences in emission sources – anthropogenic in our case, biomass burning in theirs.

Our analysis also shows that the daylight diurnal and temporal average update schemes produce higher simulation errors in the morning hours when evaluated against GEMS, compared to midday periods. This pattern arises because these update schemes modify emissions only during daylight hours while leaving nighttime emissions unchanged. Although NO2 is a short-lived species, it still can have a multi-hour lifetime, meaning that emissions from previous hours influence concentrations at any given time. Consequently, unchanged nighttime emissions can impact early morning NO2 concentrations. Furthermore, as noted in recent work (Lange et al., 2024; Park et al., 2024), GEMS observations in the early morning and late afternoon are subject to greater uncertainty due to unfavorable viewing geometries, often leading to underestimation of NO2 columns. This observational bias further complicates the evaluation of morning model performance when using GEMS data alone. In the future, we plan to integrate other observations, which provide nighttime data, for these update schemes. As our study is one of the first applications of GEMS geostationary data to update local emissions specifically focused on Thailand, we have not yet applied other instruments to infer nighttime anthropogenic activities such as thermal infrared (TIR) instruments or VIIRS night light products.