Global Accuracy Comparison from Multi-Source NO₂ Products Based on Pandora Observations

Highlights

What are the main findings?

• TROPOMI exhibits the highest accuracy among the four global TOTNO₂ products (R = 0.88, RMSE = 4.83 Pmolec·cm⁻²), while CAMS reanalysis shows the largest bias and error (MB = −4.91 Pmolec·cm⁻², RMSE = 9.15 Pmolec·cm⁻²).
• All four TOTNO₂ products (OMI, TROPOMI, GOME-2, CAMS) systematically underestimate Pandora ground-based observations and follow a clear “winter-high, summer-low” seasonal cycle.

What are the implications of the main findings?

• Linear bias correction reduces systematic errors by more than 79% and lowers RMSE by 4–28%, significantly improving the consistency between satellite/CAMS data and ground-based measurements.
• The superior accuracy and stability of TROPOMI make it a suitable reference for calibrating coarser-resolution products like CAMS, facilitating multi-source NO₂ data fusion and air quality applications.

Abstract

Effective evaluation of the accuracy of multi-source NO₂ products from different satellites and reanalysis is of great significance for data fusion and application. Based on NO₂ observation data from Pandora stations worldwide, we verify and compare the accuracy of the total column density of NO₂ (TOTNO₂) from OMI, TROPOMI, GOME-2 satellites and CAMS reanalysis. The mean biases of the four TOTNO₂ datasets relative to the Pandora station observation data are all negative, indicating that all four TOTNO₂ products show systematic underestimation with respect to Pandora. Overall, TROPOMI has the highest correlation (R = 0.88) and the smallest root mean square error (RMSE = 4.83 Pmolec·cm⁻²), suggesting that among the four TOTNO₂ products, the accuracy of TROPOMI TOTNO₂ is higher compared with the other TOTNO₂ products. The accuracies of OMI and GOME-2 are in the middle, while the performance of CAMS is the poorest. The TOTNO₂ values and accuracies from the four TOTNO₂ products both show a seasonal characteristic. Among the four TOTNO₂ products, the accuracy is higher in summer, and the error increases in autumn and winter. After performing linear fitting correction on the four NO₂ products, the mean biases of each data are reduced by more than 79%, and the RMSE decreases by 4–28%. The consistency of the four TOTNO₂ products with the ground-based observation data is significantly improved.

1. Introduction

NO₂ is not only an important trace gas pollutant in the atmosphere [1], participating in the control process of ozone (O₃) and hydroxyl radicals in the troposphere, but also a key primary air pollutant and a key precursor for ozone formation [2,3]. The sources of NO₂ can be traced back to natural processes and human activities. Natural sources of NO₂ mainly include soil emissions, lightning, wildfires and volcanic activities. Human emission sources are highly concentrated in fossil fuel consumption in the fields of traffic and industrial processes [4,5]. When NO₂ concentration is too high, it not only has a serious impact on air quality [6] but also poses a threat to the human respiratory system [7]. Therefore, accurately monitoring and analyzing the spatio-temporal distribution and the concentration changes of NO₂ in the atmosphere is of vital importance for pollution control and policy-making for human health [5].

There are various NO₂ data sources, such as ground-based station data, satellite observation data, and reanalysis data [3,8], but there are differences among them. Ground-based station data can provide high-precision and high-time-resolution column gas concentration data and are widely used for satellite data verification [9]. However, due to the uneven distribution and limited number of observation stations, large-scale detection cannot be achieved and the cost is relatively high. Satellite NO₂ monitoring, with its advantages of global coverage, continuous observation and providing long-term series data, has become an important technology for monitoring global atmospheric NO₂ concentration. At present, many atmospheric composition monitoring satellites have been successfully launched internationally. For example, the European Remote-Sensing Satellite-2 (ERS-2) was launched by the European Space Agency (ESA) in 1995. ERS-2 is equipped with the Global Ozone Monitoring Experiment (GOME), and this Satellite ceased operation in 2003 [10]. The Environmental Satellite (ENVISAT), launched in 2002, is equipped with a Scanning Imaging Absorption SpectroMeter for Atmospheric CHartographY (SCIAMACHY), which aims to provide data on the concentration of trace gases in the troposphere and stratosphere [11]. The Aura satellite launched by National Aeronautics and Space Administration (NASA) is equipped with an Ozone Monitoring Instrument (OMI) sensor [12]. Subsequently, the Global Ozone Monitoring Experiment-2 (GOME-2) was carried by the Meteorological Operational Satellite-A/B (MetOp-A/B) Satellite [13]. The Tropospheric Monitoring Instrument (TROPOMI), carried by the Sentinel-5 Precursor (Sentinel-5P) satellite, was launched in 2017 with the mission of atmospheric monitoring, marking a significant improvement in the spatial resolution of satellite monitoring of NO₂ [14]. At present, OMI, GOME-2 and TROPOMI are in normal operation. Furthermore, the reanalysis dataset of the Copernicus Atmosphere Monitoring Service (CAMS) generates a continuous spatiotemporal data field with high spatial resolution and high temporal resolution [15]. Ground-based NO₂ observation data have high accuracy and thus are often used as the benchmark for satellite accuracy verification.

At present, satellite observations can provide long-term and wide-range NO₂ data, which are widely used for the analysis of spatiotemporal variations in NO₂ [16,17,18,19], but their accuracy and reliability are unknown. Therefore, it is necessary to verify the accuracy of the multiple NO₂ data. Judd et al. took the NO₂ data of the Pandora station as the reference value to comparatively analyze the accuracy of the Sentinel-5 Precursor (S-5P) NO₂ data in the New York area. The results showed that there is a significant correlation between the Pandora and TROPOMI NO₂ data [20]. Herman et al. analyzed the accuracy of OMI NO₂ data based on Pandora station NO₂ data, and the results show that OMI systematically underestimates the amount of NO₂ in the atmosphere [21]. Celarier et al. comprehensively verified the OMI NO₂ product based on Systeme d’Analyse par Observations Zenithales (SAOZ) and DOAS ground observation data. The results indicate that OMI NO₂ has a good correlation with ground measurements, but there is a systematic underestimation bias [22]. Peters et al. conducted a comparative analysis of the NO₂ data of GOME-2 and SCIAMACHY based on MAX-DOAS data and found that there is only about 1% difference between SCIAMACHY and GOME-2, indicating a good consistency between the two NO₂ data [23]. Wang et al. verified the accuracy of GOME-2A/B using MAX-DOAS tropospheric NO₂ data from Wuxi, China, from 2011 to 2014, and found that the estimated values of GOME-2A/B products are systematically higher, approximately 30% higher [24]. Pinardi et al. verified the NO₂ accuracy of the GOME-2 satellite based on MAX-DOAS ground-based data. The results show that the correlation between GOME-2 NO₂ data and ground observation NO₂ data in clean areas is above 0.8, and the consistency is good [25]. Gruzdev et al. compared the NO₂ data from Russian ground stations with the NO₂ column concentration of the OMI satellite. The results show that the correlation coefficient between the NO₂ column concentration of OMI and that of the ground observation is 0.92, indicating a good consistency between the two types of data [26]. Douros et al. compared the NO₂ concentrations in Europe using CAMS and TROPOMI and found that there is a high consistency between the tropospheric NO₂ column concentrations of the two datasets in summer, with the relative differences in most cities not exceeding 15%. In winter, there are significant differences in column concentrations in most parts of Europe, with the maximum relative difference reaching up to 50% [8].

At present, on the one hand, the accuracy verification of NO₂ products from remote sensing or reanalysis mainly focuses on the evaluation and comparison of one or two NO₂ products. The differences in temporal matching principles of the remote sensing NO₂ data and ground reference NO₂ data with different time coverage ranges in different papers make it impossible to directly verify and compare different NO₂ products. On the other hand, the accuracy verification and comparative analysis of NO₂ data from different satellites are of great significance for the application of NO₂ data and data fusion. At present, there are relatively few studies on the accuracy verification and analysis of different satellite data under the same conditions, including the same station and the same temporal matching principle, and there are also few studies on the accuracy verification and comparison of CAMS reanalysis data and different satellites under the same conditions. In addition, there are certain differences in the accuracy and applicability of different satellites and reanalysis NO₂ products in different regions and at different times around the world. Therefore, in this paper, the total column density of NO₂ (TOTNO₂) products of OMI, TROPOMI, GOME-2 and CAMS are verified and compared from August 2021 to July 2023 using global Pandora station data as reference values based on the same temporal matching principle, and bias correction models for four TOTNO₂ products have been established.

2. Data and Methods

2.1. Data

2.1.1. Pandora Data

The Pandonia Global Network (PGN) is a collaborative project between NASA (Washington, DC, USA) and ESA (Paris, France), which uses ground-based Pandora spectrometers to monitor trace gases, such as nitrogen dioxide, ozone, sulfur dioxide and formaldehyde [27]. The main purpose of PGN is to provide long-term mass observations of the total column density and vertical resolution concentration of a series of trace gases [28]. The Pandora instrument is a solar spectrophotometer. Its observation conditions are that the Solar Zenith Angle (SZA) is less than 80°. The wavelength range is 280–525 nm, and it can obtain high signal-to-noise ratio spectral data in the ultraviolet band and the visible light-ultraviolet mixed band separately. The minimum integration time is 4 ms, the direct solar field of view angle is 1.6° [29], and the inversion is performed approximately every 2 min [30]. The data used in this study were sourced from https://data.hetzner.pandonia-global-network.org/ (accessed on 21 June 2026) and filtered based on the following criteria: Normalized Root Mean Square > 0.05, quality flags selected as Q0, Q1, Q10, and Q11, SolarZenithAngle < 75, and NO₂ value > 0.

2.1.2. OMI NO₂ Product

OMI is a hyperspectral imaging spectrometer operating in a sun-synchronous orbit [31], providing global observational data of ozone and major atmospheric pollutant gases such as NO₂ and SO₂, with a spatial resolution of 13 × 24 km [32], a transit time of close to 13:45 local time [33], and a time range from 1 October 2004 to 1 January 2025. In this paper we use OMI NO₂ vertical column data from August 2021 to July 2023, the data can be downloaded at NASA’s Goddard space flight center (Greenbelt, MD, USA) for free access (https://disc.gsfc.nasa.gov/datasets/OMI_MINDS_NO2_1.1/summary?keywords=NO2 (accessed on 21 June 2026)), by filtering and retaining data with cloud cover less than 0.5, solar zenith angle less than 70°, positive NO₂ column amounts, and quality flags (VcdQualityFlags and XTrackQualityFlags) set to zero.

2.1.3. TROPOMI NO₂ Product

TROPOMI is an instrument carried on the Sentinel-5 Precursor satellite launched by the European Space Agency (ESA) with a sun-synchronous orbit, which was successfully launched on 13 October 2017 [34,35]. TROPOMI is mainly used to obtain the vertical column data of key components in the atmosphere, including NO₂, ozone (O₃), sulfur dioxide (SO₂), formaldehyde (HCHO), methane (CH₄), and carbon monoxide (CO), etc. TROPOMI has a scanning width of 2600 km, enabling daily global coverage, and its transit time is approximately 13:00 local time [33]. Its spatial resolution is relatively high, at 3.5 km × 7 km, and has been improved to 3.5 km × 5.5 km since 6 August 2019 [36]. TROPOMI can provide both near-real-time data and offline data. TROPOMI NO₂ data can be downloaded from the ESA’s official website (https://dataspace.copernicus.eu/ (accessed on 21 June 2026)), filtering and retaining data with cloud cover less than 0.5, solar zenith angle less than 70°, a qa_value of at least 0.75, and a positive nitrogendioxide_total_column.

2.1.4. GOME-2 NO₂ Product

The Metop series satellites (METOP-A, METOP-B and METOP-C), as an important component of the Global meteorological observation network, are equipped with a variety of high-precision instruments, including the GOME-2, Infrared Atmospheric Sounding Interferometer (IASI) and Advanced Microwave Sounding Unit-A (AMSU-A) [37]. On 17 September 2012, the Metop-B satellite was successfully launched, with an orbital altitude of 817 km [38,39]. The Metop-B satellite was developed by the European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT, Darmstadt, Germany). The meteorological satellite system managed by EUMETSAT is designed to provide high-quality atmospheric monitoring and meteorological data, supporting global climate change research and meteorological forecasting efforts.

The GOME-2 carried by the Metop-B satellite plays a significant role in the monitoring of nitrogen dioxide and is one of the main instruments for atmospheric composition monitoring. It can measure multiple atmospheric pollutants such as O₃, NO₂, and sulfate aerosols, and provide high-precision data covering the world [35]. This paper screens the vertical column NO₂ concentration data that meet the criteria of cloud cover < 50% and solar zenith angle < 70°, and NO₂ concentration > 0.

2.1.5. CAMS NO₂ Product

Copernicus Atmosphere Monitoring Service Reanalysis (CAMSRA) is a global atmospheric composition reanalysis dataset produced by the European Centre for Medium-Range Weather Forecasts (ECMWF, Reading, UK) [39], also known as ECMWF’s Fourth-generation Atmospheric Composition Reanalysis (EAC4). This data reanalysis process is not subject to the time limit of real-time forecasting, thus enabling the collection of more observational data and the application of an improved inversion algorithm to historical observations, which enhances the quality of the final product [38]. The advantage of this dataset lies in that even in regions with sparse observations, the model can provide reasonable estimated values and offer long-term, globally covered data records in a unified format, which greatly facilitates the spatio-temporal analysis and trend research of atmospheric components. This study uses the global NO₂ total column density data provided by CAMS, with a temporal resolution of 3 h and a spatial resolution of 0.75° × 0.75° [36], which can be obtained from the official website (https://ads.atmosphere.copernicus.eu/ (accessed on 21 June 2026)).

The information of OMI, TROPOMI, GOME-2 satellite data, Pandora ground-based station and CAMS reanalysis data is shown in

Table 1. Information on a multi-source satellite, CAMS and Pandora ground station.

Type	Instrument	Data Version	Spatial Resolution/km	Time Resolution	Time Range
Ground-based	Pandora	-	-	2 min	2010–present
Satellite observation	OMI	MINDS v1.1	13 × 24	98.8 min	1 October 2004–1 January 2025
	TROPOMI	HiR v2	3.5 × 5.5	1 d	13 October 2017–present
	GOME-2	TM4-NO2A v 2.3	80 × 40	1 d	17 September 2012–10 January 2024
CAMS reanalysis	-	EAC4 v2	80 × 80	3 h	1 December 2003–31 December 2024

.

2.2. Research Methods

This study aims to evaluate and compare the accuracies of the TOTNO₂ products from OMI, TROPOMI, GOME-2 satellite data and CAMS reanalysis data based on the Pandora station, explore their spatio-temporal distribution characteristics, and perform linear correction. We only use the Pandora TOTNO₂ data to independently verify the accuracy of the TOTNO₂ products from OMI, TROPOMI, GOME-2 and CAMS, and do not compare the TOTNO₂ among OMI, TROPOMI, GOME-2 and CAMS. When one kind of satellite TOTNO₂ data is verified based on Pandora, it is completely independent of the other two kinds of satellite TOTNO₂ data. This study is mainly divided into three steps. Firstly, qualified data are selected, and the satellite data or CAMS data are matched with the ground station data according to certain temporal and spatial matching principles. Secondly, the temporal and spatial distribution characteristics of TOTNO₂ are analyzed, and the accuracy of the matching results is evaluated and analyzed. Finally, through linear correction of the data, the systematic errors of the data are processed to obtain more accurate results. The research process is shown in Figure 1.

2.2.1. Data Matching Method

In terms of time, the mean value of the TOTNO₂ from Pandora station observation data within 5 min before or after the satellite transit time or the CAMS data time is selected. Spatially, with the ground station as the center, the satellite and CAMS data within the distance range of native resolution around the station are selected to obtain the mean value. Finally, better matching data pairs of the stations and satellite or CAMS data are obtained. Based on the matching results, a comprehensive comparative analysis of different data sources is conducted.

2.2.2. Screening Method

Based on the data from 148 Pandora stations worldwide, we conduct a systematic accuracy verification of four TOTNO₂ products from TROPOMI, OMI, GOME-2 and CAMS. The low- and medium-quality Pandora observation data contain large uncertain errors that compromise data accuracy evaluation. Stations with too few valid data pairs cannot provide statistically meaningful results, while a high requirement on the number of matching data pairs would leave too few usable Pandora stations. Therefore, we selected only Pandora stations with high-quality observations and more than 10 satellite-matched data pairs for accuracy assessment, and eventually, 46 Pandora stations with better matching effects were obtained for in-depth analysis, thus ensuring the reliability and representativeness of the verification results. The global distribution of the 46 stations is shown in Figure 2.

2.2.3. Statistical Indicators

To quantitatively evaluate the consistency and differences among different data sources, the following statistical indicators are adopted in this study: Mean Bias (MB), Root Mean Square Error (RMSE), Standard Deviation (STD), and Correlation Coefficient (R).

$$MB={\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n\left(TOTNO{2}_{sate,i}-TOTNO{2}_{site,i}\right)}$$

(1)

$$RMSE=\sqrt{{\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n{(TOTNO{2}_{sate,i}-TOTNO{2}_{site,i})}^2}}$$

(2)

$$STD=\sqrt{{\displaystyle \frac{1}{n}}{\displaystyle \sum _{i-1}^n(TOTNO{2}_{sate,i}-TOTNO{2}_{site,i}-\overline{TOTNO{2}_{sate}-TOTNO{2}_{site}{)}^2}}}$$

(3)

$$R={\displaystyle \frac{{\displaystyle \sum _{i=1}^n(TOTNO{2}_{sate,i}-\overline{TOTNO{2}_{sate}})(TOTNO{2}_{site,i}-\overline{TOTNO{2}_{site}})}}{\sqrt{{\displaystyle \sum _{i=1}^n{(TOTNO{2}_{sate,i}-\overline{TOTNO{2}_{site}})}^2{\displaystyle \sum _{i=1}^n{(TOTNO{2}_{sate,i}-\overline{TOTNO{2}_{site}})}^2}}}}}$$

(4)

where $TOTNO{2}_{sate}$ represents the total column density of nitrogen dioxide for satellites/CAMS reanalysis, and $TOTNO{2}_{site}$ represents the total column density of nitrogen dioxide for the Pandora station.

2.2.4. Bias Correction Method

Firstly, in the matching TOTNO₂ dataset, a random sampling strategy is employed to divide the training set and the validation set. A total of 80% of the samples are used for linear fitting model construction, while 20% are reserved for independent validation to ensure the generalization ability of the corrected model and the objectivity of the evaluation. Subsequently, on the training set, the linear regression is conducted using the observation values from the Pandora station as the dependent variable and the satellite observations or reanalysis data as the independent variable, which is shown in Equation (5).

$$TOTNO{2}_{site}=a\times TOTNO{2}_{sate}+b$$

(5)

3. Results

3.1. Comparison of the Temporal Characteristics of NO₂ Products Three Kinds of Satellite and CAMS Reanalysis with Those of the Pandora Station

In order to more accurately analyze the time variation patterns of the four TOTNO₂ products and the TOTNO₂ from ground-based observation stations, the differences in time variation patterns between the Northern and Southern Hemispheres are taken into account. The temporal characteristics of TOTNO₂ products from OMI, TROPOMI, GOME-2 and CAMS, as well as Pandora from August 2021 to July 2023 in the Southern Hemisphere and the Northern Hemisphere are shown in Figure 3.

The temporal characteristics of TOTNO₂ products from OMI, TROPOMI, GOME-2 and CAMS, as well as Pandora from August 2021 to July 2023, are shown in Figure 3. As shown in Figure 3, overall, the four TOTNO₂ products exhibit a consistent trend of change with the data from the Pandora station, all showing distinctly seasonal characteristics. The TOTNO₂ concentration is relatively low in summer, gradually increases in autumn and winter, and reaches its peak in winter. In the Northern Hemisphere, ground-based TOTNO₂ is higher than the three satellite products and CAMS in most cases. There are too few ground-based observation stations in the Southern Hemisphere, so this phenomenon is not obvious.

Among the TOTNO₂ datasets from the three satellites, the TOTNO₂ of TROPOMI and OMI are the closest, and they also have the same temporal characteristics as Pandora. However, the consistency between GOME-2 and Pandora’s NO₂ products lies between those of other satellite data and CAMS.

3.2. Accuracy Assessment of Multi-Source NO₂ Products Based on Pandora Ground-Based Observations

3.2.1. Overall Accuracy Assessment of NO₂ Products

The scatter plots and box plots comparing the four TOTNO₂ data products (OMI, TROPOMI, GOME-2 and CAMS) with the ground-based observations at the global Pandora stations are shown in Figure 4 and Figure 5, respectively.

Figure 4 shows that the correlation between TROPOMI and Pandora TOTNO₂ observations is the highest (R = 0.88), and the RMSE value of 4.83 Pmolec·cm⁻² is the smallest among the four products, indicating the best consistency between TROPOMI and ground-based observations. The mean bias between TROPOMI and Pandora is −1.48 Pmolec·cm⁻², which indicates that there is a systematic deviation in TROPOMI TOTNO₂, which is significantly underestimated compared to the Pandora station. The RMSE and STD of OMI are similar to those of TROPOMI, indicating that the accuracies of the two TOTNO₂ products are comparable. The mean absolute bias of OMI is slightly larger than that of TROPOMI, indicating that the systematic bias of the OMI TOTNO₂ product is smaller compared to that of TROPOMI.

Figure 5 shows that all four TOTNO₂ products exhibit a systematic negative bias. The median biases of the four TOTNO₂ products are all less than 0. The results indicate that the overall values of these four TOTNO₂ products are underestimated compared to the ground-based TOTNO₂ products. The upper quartile of the CAMS bias is less than 0, indicating that the CAMS TOTNO₂ product is significantly underestimated by more than 75% compared to ground-based observations. The median bias between OMI and GOME-2 is similar, but the dispersion of OMI is smaller, indicating that the OMI NO₂ product has better accuracy and stability. The maximum and minimum biases of TROPOMI are both relatively small, and the bias range of TROPOMI is more concentrated, indicating that the accuracy and stability of the TROPOMI NO₂ product are better compared to the other three products.

3.2.2. The Accuracy of the Four NO₂ Products Changes over Time

The monthly root mean square error results of the four NO₂ products (OMI, TROPOMI, GOME-2 and CAMS) relative to the observations at the Pandora station are shown in Figure 6. Due to the scarcity of stations in the Southern Hemisphere, this analysis is conducted based on the stations in the Northern Hemisphere.

Figure 6 shows that the monthly RMSE values of the four TOTNO₂ products show seasonal fluctuations. The RMSE values of the four TOTNO₂ are generally higher in winter. In December 2021, the RMSE of GOME-2, OMI, CAMS and TROPOMI reached 9.20 Pmolec·cm⁻², 6.92 Pmolec·cm⁻², 15.69 Pmolec·cm⁻² and 9.46 Pmolec·cm⁻², respectively, while they significantly decreased in summer. In July 2022, the RMSEs of GOME-2, OMI, CAMS and TROPOMI drop to 5.18 Pmolec·cm⁻², 2.94 Pmolec·cm⁻², 6.59 Pmolec·cm⁻² and 3.32 Pmolec·cm⁻², respectively.

Table 2. The number of collocation pairs of OMI, TROPOMI, GOME-2 and CAMS relative to Pandora.

Month	GOME-2	OMI	CAMS	TROPOMI
2021-08	219	23	1190	155
2021-09	258	16	1141	179
2021-10	246	19	909	138
2021-11	304	25	817	169
2021-12	204	16	553	153
2022-01	111	26	580	133
2022-02	207	28	709	112
2022-03	195	29	812	115
2022-04	203	19	833	96
2022-05	164	37	1040	93
2022-06	139	43	1077	80
2022-07	168	29	1110	90
2022-08	174	25	1087	133
2022-09	207	23	1002	165
2022-10	231	36	895	180
2022-11	192	20	710	157
2022-12	106	50	517	121
2023-01	80	9	293	48
2023-02	40	13	338	0
2023-03	94	24	373	0
2023-04	81	32	443	67
2023-05	129	63	662	68
2023-06	90	29	597	0
2023-07	89	42	657	0

indicates that the number of collocation pairs shows an uneven distribution over the period from August 2021 to July 2023. Due to the fact that the number of collocation pairs of the four datasets in the period from August 2022 to July 2023 is generally smaller than that of the four datasets in the period from August 2021 to July 2022, the seasonal variation in RMSE from August 2022 to July 2023 is not as obvious as that from August 2021 to July 2022.

3.2.3. Station-Level Accuracy Evaluation from NO₂ Products

Figure 7, Figure 8, Figure 9 and Figure 10 present the station-level accuracy evaluation statistics of four TOTNO₂ products from OMI, TROPOMI, GOME-2 and CAMS relative to 46 Pandora stations worldwide. At most observation stations, Figure 8 indicates that the RMSE of the CAMS TOTNO₂ product is significantly higher than that of the other three satellite products. Moreover, the correlation coefficient of the CAMS TOTNO₂ product at most observation stations is significantly lower than that of the other three satellite products. The results show that the accuracy of CAMS TOTNO₂ is lower than that of other satellite products. TROPOMI TOTNO₂ consistently demonstrates the highest and most stable accuracy at most stations. The accuracy of OMI TOTNO₂ is generally stable, with relatively high accuracy at many stations (such as Beijing-RADI), similar to that of TROPOMI. However, as shown in Figure 11, the error of OMI significantly increases in some highly polluted cities (such as Incheon-ESC, Yokosuka). The station-level accuracy of GOME-2 varies greatly, and the errors are the highest at some stations, especially at high-pollution stations such as Incheon-ESC.

From the perspective of spatial distribution patterns, at most of the stations, four NO₂ products exhibit good consistency, and the root mean square error values of these four NO₂ products at most stations are either relatively large or relatively small. The findings indicate that the RMSE values of the four products are highly correlated with the quality and concentration of the station observations. The errors of the four TOTNO₂ products are relatively small at low NO₂ stations (such as Dalanzadgad), and the consistency with respect to Pandora stations is good. However, in regions with high NO₂ concentrations, especially in urban clusters in East Asia and North America, the errors of the four TOTNO₂ products increase. The differences between CAMS and satellite data are particularly prominent among them.

3.3. Analysis of Bias Correction for TOTNO₂ Products from OMI, TROPOMI, GOME-2 and CAMS

The biases of four TOTNO₂ products from OMI, TROPOMI, GOME-2 and CAMS are corrected based on a linear fitting model. A total of 80% of matching results from four TOTNO₂ products are randomly selected for linear fitting to obtain the linear fitting equation. Then, the remaining 20% of the four TOTNO₂ products is used for verification. The accuracy comparison of the four TOTNO₂ products before and after bias correction is shown in

Table 3. Comparison of accuracy before and after correction of NO₂ products.

Type	Bias Correction	MD /Pmolec·cm−2	RMSE /Pmolec·cm−2	STD /Pmolec·cm−2	R
OMI	Before correction	−1.06	5.05	4.77	0.74
	After correction	0.22	4.87	4.70	0.74
TROPOMI	Before correction	−1.57	4.70	4.43	0.91
	After correction	−0.16	3.38	3.38	0.91
GOME-2	Before correction	−2.40	7.87	7.50	0.51
	After correction	0.16	7.50	7.50	0.51
CAMS	Before correction	−4.77	8.88	7.49	0.31
	After correction	0.14	7.36	7.35	0.31

. Table 3 shows that the MB values of the four NO₂ products are all less than 0.25 Pmolec.cm⁻², indicating that after correction, the mean biases of OMI, TROPOMI, GOME-2 and CAMS have significantly decreased by more than 79%. After correction, the RMSE values of OMI, TROPOMI, GOME-2 and CAMS decreased by 4%, 28%, 5% and 17%, respectively. The standard deviation (STD) and correlation coefficient (R) of the corrected data changed relatively little.

4. Discussion

4.1. Discussion on the Change in Accuracy Before and After the Downgrade of OMI

Due to the severe degradation of detector response and polarization sensitivity of OMI since 2019, the OMI TOTNO₂ product has been corrected for the period after 2019. To compare the accuracy of the corrected OMI TOTNO₂ products (post-2019) with that of the products before 2019, we use observation data from the same Pandora stations over the same length of time (two years) to evaluate the data from 2017 to 2018 and August 2021–July 2023, respectively, as shown in Figure 12.

Figure 12 indicates that the MB, RMSE, and STD of the OMI TOTNO₂ products from 2017 to 2018 are smaller than those of the OMI products from August 2021 to July 2023. The findings indicate that the accuracy of the OMI TOTNO₂ products from 2017 to 2018 is slightly superior to that of the OMI products from August 2021 to July 2023. The main reasons are as follows: Firstly, the OMI TOTNO₂ data products after 2019 have undergone bias correction [40]. Secondly, we select high-quality OMI NO₂ products and exclude abnormal row data, so the accuracy of the OMI TOTNO₂ product does not change much.

4.2. Discussion on Temporal Characteristics and Systematic Underestimation

The seasonal variation in NO₂ is mainly influenced by both meteorological conditions and human emissions. In summer, the intense solar radiation accelerates the photolysis reaction of NO₂, while the higher atmospheric boundary layer facilitates the vertical dispersion of pollutants. In contrast, during winter, the increase in coal burning for heating leads to higher NO₂ emissions, combined with a lower atmospheric boundary layer and frequent inversion phenomena, making it difficult for pollutants to disperse, thereby resulting in a cumulative increase in NO₂ concentration.

The differences among satellite products mainly result from various factors such as inversion algorithms, instrument resolution, and observation quality. The main reasons for the general underestimation of actual TOTNO₂ concentration by satellite NO₂ products may be as follows. On the one hand, NO₂ is mainly concentrated near the ground surface, and the reduced sensitivity of satellites to signals near the ground surface leads to an underestimation of the NO₂ concentration near the ground surface. On the other hand, high-pollution data are systematically excluded during the cloud screening process. Similarly, the underestimated satellite TOTNO₂ product is assimilated into the CAMS reanalysis product, resulting in a relatively low accuracy of the CAMS TOTNO₂ product.

The better accuracy and stability of the TROPOMI TOTNO₂ product compared to the other three products may be attributed to the superior spatial resolution that enables it to capture the spatial heterogeneity of urban NO₂ more precisely. The differences between satellites may stem from factors such as their spatial resolution and inversion algorithms.

4.3. Discussion on Seasonal Accuracy Variation

The seasonal variation in RMSE—higher in winter and lower in summer—arises from multiple physical mechanisms that affect satellite retrievals and reanalysis data assimilation. In winter, several factors jointly increase retrieval uncertainties. First, aerosol interference is more pronounced: higher aerosol optical thickness (AOT) due to anthropogenic emissions (e.g., heating, stagnant conditions) enhances photon scattering and absorption, altering the effective optical path and biasing NO₂ slant column density estimates. Second, air mass factor (AMF) uncertainties grow under low solar elevation angles (winter overpass times) and high aerosol loading, as the AMF depends on the assumed vertical profile of NO₂ and aerosols—deviations from the a priori profile introduce systematic errors. Third, boundary layer dynamics in winter often produce strong temperature inversions, trapping pollutants near the surface and creating sharp vertical gradients that are poorly captured by the coarse vertical resolution of model priors. Fourth, cloud screening effects are more challenging: while clouds can reduce data availability, partial cloud cover and sub-pixel clouds are more frequent in winter mid-latitudes, leading to misclassification and residual cloud contamination that degrades retrieval accuracy. Finally, differences in satellite overpass times relative to solar zenith angle (SZA) matter—higher SZA in winter increases the optical path length, reducing sensitivity to near-surface NO₂ and amplifying the impact of all the above uncertainties. In contrast, summer offers cleaner atmospheric conditions (lower AOT), smaller SZA, more stable boundary layers, and fewer cloud interference issues, all of which improve the accuracy of TOTNO₂ products. Thus, the observed RMSE seasonality is a direct consequence of these physically intertwined factors, rather than a simple correlation with pollutant concentration alone.

4.4. Discussion on Bias Correction

The results show that the linear fitting model has a significant effect on correcting the systematic errors of TOTNO₂ products. After correction, the RMSE of the four TOTNO₂ products has significantly decreased, suggesting a significant improvement in the overall accuracy. Among the four products, the TOTNO₂ product of the corrected CAMS has the most significant improvement in accuracy. The standard deviation (STD) and correlation coefficient (R) of the corrected data changed relatively little, indicating that the correction process mainly eliminates the systematic errors while still being influenced by random errors. This is because random errors are affected by various uncertain factors, including the meteorological conditions, surface reflectivity, and aerosol interference, which do not have a stable systematic change pattern, and are difficult to predict and eliminate through simple linear models. However, bias correction still improved the accuracy of satellite TOTNO₂ products, making them more consistent with ground-based observations.

However, the limitations of the linear correction approach should be acknowledged. Satellite NO₂ retrieval errors are often nonlinear and strongly modulated by season, aerosol loading, surface reflectivity, boundary layer conditions, and pollution levels. The linear regression model employed here is overly simplistic and does not account for these complex factors. Consequently, the physical meaning and generalizability of the linear correction remain limited. Future work should explore more sophisticated, physically based or machine-learning methods to better capture the nonlinear nature of retrieval errors across varying environmental conditions.

5. Conclusions

NO₂ participates in the photochemical cycle and serves as an important precursor for ozone (O₃) and secondary aerosols (such as nitrate particles), directly influencing regional air quality and atmospheric oxidation capacity. Based on the TOTNO₂ observation data from the global Pandora stations as the reference value, the temporal variation characteristics of TOTNO₂ products from OMI, TROPOMI, GOME-2 and CAMS reanalysis data and their consistency with Pandora are investigated. We systematically compare and analyze the consistency of four TOTNO₂ products with Pandora and use the linear fitting method to correct the bias of the four kinds of TOTNO₂ products. The main conclusions are as follows:

(1)

The four TOTNO₂ data sources relative to the Pandora observations generally show systematic underestimation, with negative mean bias values. Among them, TROPOMI TOTNO₂ products perform best, with the highest correlation coefficient (average R = 0.88) and the smallest RMSE value (RMSE = 4.83 Pmolec·cm⁻²), reflecting its advantage in capturing the distribution of urban NO₂ with high spatial resolution. The accuracies of the OMI and GOME-2 TOTNO₂ products are comparable and intermediate. The TOTNO₂ data from CAMS reanalysis shows the largest bias and RMSE (MB = −4.91 Pmolec·cm⁻², RMSE = 9.15 Pmolec·cm⁻²), and the results are unstable in different regions, indicating a significant discrepancy between the model simulation and the ground-based observation values.

(2)

In terms of the temporal variation characteristics, an obvious seasonality, characterized by higher TOTNO₂ values in winter and lower TOTNO₂ values in summer, is evident in all four TOTNO₂ products, and this variation agrees well with the Pandora observation. The increase in TOTNO₂ concentration in winter is primarily attributed to increased anthropogenic emissions (particularly from heating combustion), reduced photochemical loss rates, and a marked decrease in atmospheric dispersion capacity caused by adverse meteorological conditions such as temperature inversions. In summer, due to enhanced photolysis and the elevation of the boundary layer, the concentration drops to the lowest level. From the perspective of seasonal variation in TOTNO₂ product accuracy, each data source has the smallest bias and the highest correlation in summer, while the RMSE and STD generally increase in autumn and winter, which may be related to the decrease in the solar elevation angle in winter and the aggravation of aerosol influence.

(3)

In terms of bias correction, the correction method based on linear fitting effectively reduces the systematic bias of the four types of TOTNO₂ data. After correction, the mean bias values of OMI, TROPOMI, GOME-2 and CAMS decreased by 79%, 90%, 93% and 97%, respectively, and the root mean square error values decreased by 4%, 28%, 5% and 17%, respectively. The results indicate that the systematic bias values of the four kinds of TOTNO₂ products have been significantly reduced and the accuracies have significantly improved. However, the random error (STD) and correlation of the four kinds of TOTNO₂ products do not change significantly. The four kinds of TOTNO₂ products after correction are overall closer to the ground-based observations, providing a reliable foundation for the fusion and application of multi-source NO₂ data.