Surface NO₂ measurements with hourly resolution from 2005 to 2023 were collected from approximately 7021 ground-based monitoring stations worldwide, including China National Environmental Monitoring Centre (CNEMC), Environmental Protection Agency (EPA), European Environment Agency (EEA), Tropospheric Ozone Assessment Report database (TOAR), and Open Air Quality (OpenAQ). To create model labels, duplicate data were removed by retaining the highest priority measurements according to the specified order (CNEMC/EPA/EEA > TOAR > OpenAQ). The raw data underwent additional quality control measures, including removal of outliers such as negative values, duplicate values (occurring for more than three consecutive hours), and extreme values (exceeding the 99.9th percentile). Days with at least 18 valid hourly NO₂ measurements were identified to calculate the daily average for each monitoring site. To ensure consistency with auxiliary variables, all daily NO₂ measurements were adjusted to the Universal Time Coordinated (UTC) zone and then used for independent training and validation of the machine learning model. Finally, daily observations were spatially aggregated to unique grid locations based on geographic coordinates, and the distribution of sample counts and NO₂ descriptive statistics is summarized in Table S2 in the Supplement.

To accurately predict NO₂ concentrations, we incorporated a total of 17 independent features derived from multiple data sources. Meteorological reanalysis data were obtained from the ECMWF Reanalysis v5 (ERA5) dataset, including boundary layer height (BLH), temperature (TEMP), wind speed (WS), dewpoint temperature (DEW), pressure (PRE), precipitation (PRECIP), evaporation (EVA), and long-wave solar radiation (LWSR), all at a spatial resolution of 0.25° × 0.25° for the period 2005–2023. These meteorological variables are known to play key roles in controlling pollutant dispersion, vertical mixing, chemical reaction rates, and surface–atmosphere exchanges, and therefore provide essential physical constraints for NO₂ prediction. Satellite data of tropospheric column NO₂, formaldehyde (HCHO), and O₃ were retrieved from the L3 level products in OMI instrument at a resolution of 0.25° × 0.25° for 2005–2018, and from the L2 level products in TROPOMI instrument at a finer resolution of 5.5 km × 3.5 km for 2019–2023, providing complementary chemical information related to NO_x emissions, photochemical activity, and oxidant levels.

To enable consistent long-term analysis and minimize discrepancies arising from differences in sensor characteristics, retrieval algorithms, horizontal resolution, and overpass time between OMI and TROPOMI, we applied a seasonal, grid-specific adjustment to OMI products using TROPOMI as a reference during their overlap period (2019–2022). This harmonization step is particularly important for long-term trend analyses, as uncorrected inter-sensor differences could introduce artificial discontinuities into the time series. Specifically, OMI observations for years without TROPOMI coverage were corrected according to 1ΩOMIadji,yr,m=ΩOMIi,yr,m+ΔΩweightsi,m, where 2ΔΩweightsi,m=14∑yr=2019yr=2022(ΩTROPOMIi,yr,m-ΩOMIi,yr,m), where ΩOMIadji,yr,m, ΩOMIi,yr,m, ΩTROPOMIi,yr,m is adjusted OMI, original OMI and original TROPOMI products for grid i at year yr season m, respectively. ΔΩweightsi,m represents the difference between OMI and TROPOMI at target resolution. This correction strategy preserves the spatial patterns and seasonal variability of the original satellite products while ensuring temporal continuity and consistency of satellite-derived chemical variables across the entire study period.

Aerosol Optical Depth (AOD) data, an important proxy for particulate matter, were sourced from MODIS product (MCD19A2CMG) at a resolution of 0.05° × 0.05° for 2005–2023. The frequent occurrence of clouds in optical remote sensing images hampers the detection of trace gases beneath the cloud layer, leading to gaps in satellite data. To address this challenge, we implemented the eXtreme Gradient Boosting (XGBoost) machine learning algorithm, known for its efficiency, to fill these gaps and obtain spatially comprehensive retrieval data. This gap-filling procedure improves the representativeness of aerosol-related information and reduces potential biases associated with uneven data availability.

Additionally, we also included NDVI (2005–2023, MCD19A3CMG) to account for vegetation cover, population density data (2005–2020, WorldPop, https://www.worldpop.org/, last access: 29 April 2026) to capture human activity, land use data (2005–2022, MCD12C1, providing sub-pixel proportions of nine class in forests, shrublands, savannas, wetlands, croplands, urban, snow and ice, barren, water) to understand land cover changes, and road network density and Digital Elevation Model (DEM) to incorporate infrastructural and topographical influences.

To ensure consistency across all input variables, the higher spatial resolution variables were aggregated to a homogeneous 0.1° × 0.1° grid using averaging, while the lower spatial resolution variables were resampled to the same grid using bilinear interpolation. Bilinear interpolation estimates values at target grid cells by linearly weighting the four nearest neighboring grid points in both latitude and longitude directions, thereby providing a smooth spatial transition while preserving large-scale gradients and avoiding abrupt discontinuities. This approach is widely adopted in atmospheric and remote sensing applications for spatial resampling, as it balances computational efficiency with the need to retain spatial coherence. This harmonization process enables the seamless integration of multi-source datasets within the machine learning framework. These processed variables were then used for subsequent air pollution modeling. This diverse set of features enabled a comprehensive machine learning approach to modeling NO₂ concentrations, providing robust predictions across various spatial and temporal scales.

This study applied an innovative deep learning model, the Air Transformer (AiT), to improve the accuracy of atmospheric pollutant concentration prediction. The AiT model is based on the Transformer architecture and is specifically designed to capture the spatiotemporal dependencies of atmospheric pollutants. The AiT model structure builds on the successful application of the Transformer in natural language processing, incorporating cross-channel self-attention alongside pixel self-attention and utilizing a decoder for enhanced prediction capabilities. The multi-dimensional attention mechanism enables the model to capture dependencies not only across temporal and spatial dimensions but also among data channels, such as different pollutant concentrations and meteorological parameters. Transformers have the advantage of having no recurrent units, and therefore require less training time than earlier recurrent neural architectures (RNNs) such as long short-term memory (LSTM) (Zhao et al., 2019; Liu et al., 2024).

The self-attention mechanism is the core component of the AiT model. For each input sequence, attention weights are generated by computing the dot product between the query, key, and value vectors. These weights are used to perform a weighted sum of the input sequence values, producing new feature representations. The self-attention mechanism further enhances the model's representational power by allowing it to simultaneously focus on different parts of the input data, capturing richer spatiotemporal dependencies. In this study, the AiT model establishes a robust global surface–column relationship by integrating variables related to pollutant transport, chemical transformation, and column density within approximately 20 km of each target grid, enabling the reconstruction of gapless daily surface NO₂ concentrations at a spatial resolution of 10 km.

Multi-source data were input into AiT as model features and labels. These data were pre-processed through feature extraction layers and then fed into the Transformer's encoder. The encoder comprises multiple self-attention layers and feedforward neural network layers. Utilizing the multi-head attention mechanism, it processes input data in parallel, capturing features at various temporal and spatial scales. The AiT model is trained using supervised learning, leveraging a large amount of historical data. The model's loss function comprises prediction errors (such as mean squared error (MSE), Root Mean Square Error (RMSE)) and regularization terms to ensure prediction accuracy and stability. To improve the model's generalization ability, we introduced a multi-task learning strategy which was detailed discussed in our previous work (Tao et al., 2024). Specifically, the model not only predicts pollutant concentrations but also simultaneously predicts related meteorological parameters and remote sensing observations. This multi-task learning strategy helps the model better understand the interrelationships between different variables, thereby enhancing prediction performance.

Model training was conducted for 300 epochs using the LAMB optimizer, which supports large batch sizes without sacrificing convergence performance, thereby accelerating training on large-scale datasets. A warmup learning rate schedule was applied, with the learning rate peaking at 0.001, and a batch size of 6144 was used. An early stopping strategy was implemented and triggered after 30 consecutive epochs without improvement. To mitigate overfitting, dropout was applied within linear layers and self-attention modules, and a Gaussian Error Linear Unit (GeLU) activation function was employed throughout the network.

The AiT model was implemented and trained using the PyTorch framework. Prior to training, all input features were normalized across the entire dataset. Model performance was evaluated using both sample-based and site-based cross-validation approaches. In these evaluations, all samples and monitoring sites were randomly divided into five folds, with four folds used for training and one fold reserved for independent validation, allowing for a rigorous assessment of the model's spatiotemporal generalization capability.

The performance of our models was evaluated using two five-fold cross-validation approaches: random-based and spatial-based. The random-based approach involved 16 841 418 samples, with 3 368 283 samples used for validation in each fold. This approach ensures a comprehensive assessment of the model's predictive performance across diverse temporal and spatial conditions represented in the full dataset, effectively testing the model's ability to reproduce observed NO₂ variability under a wide range of atmospheric and environmental settings. By randomly partitioning samples, this strategy primarily evaluates the overall fitting accuracy and stability of the model when predicting NO₂ concentrations at locations and times similar to those seen during training.

In contrast, the spatial-based approach was designed to explicitly assess the model's spatial transferability. This evaluation included data from 7021 monitoring sites, of which 1404 sites were held out for validation in each fold. By withholding entire monitoring locations from the training process, this approach tests the model's ability to generalize to previously unseen locations, which is critical for producing reliable spatially continuous NO₂ fields, particularly in regions with sparse or no ground observations. This site-based validation provides a more stringent and realistic assessment of model performance for global-scale mapping applications.

In both cross-validation approaches, validation data were strictly excluded from model training, ensuring an unbiased evaluation of predictive performance. Together, the random-based and spatial-based cross-validation strategies provide complementary perspectives on model accuracy and generalization capability, with the former primarily reflecting interpolation performance within the sampled space–time domain and the latter evaluating extrapolation ability to unmonitored locations. This combined evaluation framework is particularly well suited for global NO₂ mapping applications, where reliable prediction in regions with limited observational constraints is essential.

To accurately assess regional pollution levels and the potential health impacts, we employ two complementary weighted averaging methods: geographically area-weighted averaging (Eq. 3) and population-weighted averaging (Eq. 4). These equations incorporate spatial area and population data to produce weighted averages that more accurately reflect both geographical and demographic factors in regional NO₂ concentration assessments. 3Carea-weighted=∑i=1nCi×Ai∑i=1nAi, where Ci is the NO₂ concentration in the ith region, Ai is the area of the ith region, and n is the total number of regions. 4Cpopulation-weighted=∑i=1nCi×Pi∑i=1nPi, where Ci is the NO₂ concentration in the ith region, Pi is the population of the ith region, and n is the total number of regions. We used WorldPOP population datasets for this study. Population-weighted NO₂ concentration is a crucial indicator for assessing the public health impact of air pollution, as it integrates both pollutant concentration levels and the distribution of human populations across different regions. This metric is particularly useful for capturing the actual exposure of populations to NO₂ pollution, which is not adequately represented by area-weighted concentrations alone. By incorporating population density, the population-weighted NO₂ concentration metric provides a more accurate reflection of the exposure risk faced by people living and working in polluted areas.

The detailed discussion about the Transformer can be seen in the Methods. Here we mainly focus on the performance evaluation from the AiT-NO₂ products. We first evaluate the model using Random cross-validation and Spatial cross-validation. In the Random cross-validation, the NO₂ model exhibited outstanding predictive accuracy, with an R2 of 0.91, an RMSE of 2.32 ppbv, and an MAE of 1.56 ppbv for daily predictions (Fig. 1a). The performance further improved for monthly and annual predictions, achieving R2 values of 0.98 and 0.99, RMSE values of 0.97 and 0.59 ppbv, and MAE values of 0.64 ppbv and 0.33 ppbv, respectively (Fig. 1c, e). This improvement reflects the reduction of short-term noise and random variability through temporal aggregation, indicating that the model effectively captures both high-frequency variability and long-term NO₂ signals.

Similarly, the Spatial cross-validation results were also impressive. For daily predictions, the NO₂ model achieved an R2 of 0.62, an RMSE of 4.77 ppbv, and an MAE of 3.31 ppbv (Fig. 1b). For monthly and annual predictions, the model showed R2 values of 0.65 and 0.63, RMSE values of 3.92 and 3.64 ppbv, and MAE values of 2.71 and 2.51 ppbv, respectively (Fig. 1d, f). As expected, predictive performance under Spatial cross-validation is lower than that under Random cross-validation, reflecting the greater difficulty of extrapolating to previously unseen locations. Nevertheless, the relatively stable performance across temporal scales indicates that the model maintains robust spatial transferability and is capable of reproducing NO₂ variability in regions with limited or no monitoring coverage.

In addition to the global evaluation, continental-scale performance is summarized in Table S2. Under random cross-validation, the model maintains high predictive accuracy across Europe, Asia, North America, and South America (R2= 0.89–0.92). Spatial cross-validation results show greater variability across continents, with R2 values ranging from 0.54 to 0.68 in regions with dense monitoring coverage. Performance in Africa and Australia is substantially lower under spatial cross-validation, which is attributable to the limited number of available monitoring grids in this region.

Overall, the performance metrics obtained from both validation strategies demonstrate that the AiT-NO₂ model performs consistently well across multiple temporal scales and validation settings. Compared with previous studies, our framework not only provides improvements in spatial resolution (0.1°) and temporal coverage (daily estimates from 2005–2023), but also achieves higher or comparable predictive performance. For example, reported random cross-validation R2 values in previous regional NO₂ modeling studies typically range from 0.70 to 0.88, with RMSE values between approximately 3–6 ppbv depending on region and temporal aggregation (Chan et al., 2021; Shao et al., 2023; Wong et al., 2021). In contrast, our model achieves an R2 of 0.91 and an RMSE of 2.32 ppbv for daily predictions under random cross-validation at the global scale. Notably, many earlier studies were conducted at regional or national scales, often benefiting from dense monitoring networks and region-specific tuning, whereas globally consistent daily surface NO₂ datasets with comparable resolution remain relatively limited. Furthermore, the Spatial cross-validation results are generally comparable to, or exceed, those reported in previous global and regional NO₂ modeling studies (Wei et al., 2019, 2023), underscoring the robustness and reliability of the AiT-based framework for constructing spatially continuous, long-term NO₂ datasets.

As we mentioned earlier, the SoGA2024 for the first time included NO₂ in its report, as well as the annual NO₂ concentration datasets they generated using a global LUR model (Anenberg et al., 2022; Larkin et al., 2023). The LUR-based dataset was selected for comparison due to its adoption in the State of Global Air (SoGA) assessment for global NO₂ health burden estimation, where it functions as an authoritative and policy-relevant exposure benchmark. Here, we mainly focus on the performance of the two datasets in simulating NO₂ concentration in major urban observation sites in several observation-dense countries, such as Austria, Belgium, China, France, Germany, India, Poland, Spain, and US. Overall, the AiTNO₂ model shows much better alignment with observational values, outperforming the LUR model. In most countries, except the US, predictions from the AiTNO₂ model tend to be slightly higher or almost comparable to the observational data, whereas the LUR model consistently underestimates NO₂ concentrations. For countries like China and India, where pollution levels are particularly high, the AiTNO₂ model provides significantly more accurate predictions, closely reflecting the actual pollution patterns observed in these regions. This highlights the efficacy of the AiTNO₂ model. We estimated that the LUR method could significantly underestimate the regional NO₂ concentration by at least 50 % for most of the regions, except for HI (Fig. S3; Table S2). When examining different regions, our results has much higher NO₂ estimation compared to LUR NO₂, particularly in the Northern Hemisphere (1.6 to 2.1 times) and tropical regions (2.5 to 3.0 times), where there are more monitoring stations and higher pollution levels (Fig. S4). The primary reason for this discrepancy is that the AiTNO₂ model relies on urban stations data for training, which are typically located in city centres with high pollution levels, enabling it to more effectively capture elevated NO₂ concentrations in these localized urban areas. In contrast, the LUR model relies on a set of spatial predictor variables derived from land-use and environmental datasets, including population density, impervious surface area, road networks, vegetation indices (e.g., NDVI and tree cover), elevation, and satellite-based NO₂ observations from OMI. As a result, the LUR framework is less capable of fully reflecting high NO₂ concentration areas, especially in urban settings. In addition, we further compared our product with another independent NO₂ dataset (Copper et al., 2022) to provide a broader evaluation of model performance (Fig. S10). Furthermore, the AiTNO₂ model was featured in higher sensitivity in capturing the temporal and spatial variations of atmospheric pollutants, accurately predicting these changes by learning from time series data (Tao et al., 2024).

In conclusion, both cross-validation approaches confirm the strong predictive capabilities of our model. The Random validation highlights the model's accuracy across diverse data points, while the Spatial validation underscores its effectiveness in generalizing to new urban locations in regions with strong monitoring infrastructure. These comprehensive evaluations attest to the high quality and reliability of our predictive models, making them valuable tools for accurate NO₂ concentration predictions. The AiTNO₂ model shows higher accuracy in predicting NO₂ concentrations across the selected countries, especially in regions with severe pollution and dense monitoring networks. While the LUR model captures the general trends in NO₂ variation, it may provide relatively conservative estimates in urban areas and may underestimate concentrations in rural environments in some countries. The AiTNO₂ model, by leveraging deep learning and observational data, compensates for these limitations and demonstrates its potential as a valuable tool for air pollution monitoring and forecasting. These advantages highlight the value of the AiT-based approach for constructing high-resolution global NO₂ datasets and for improving the representation of surface NO₂ concentrations in data-driven air quality analyses.

Based on the geographically area-weighted average AiTNO₂, we have a systematic understanding of the global spatial and temporal trends in NO₂. The global annual average concentration of NO₂ exhibited distinct trends among regions from 2005 to 2023 (Fig. S5a). Specifically, from 2005 to 2010, NO₂ concentrations remained relatively stable, increasing slightly from 3.32 to 3.44 ppbv. From 2010 to 2015, concentrations continued to rise, peaking at 4.05 ppbv in 2015. This upward phase coincides with rapid industrial expansion, increasing fossil fuel consumption, and accelerated urbanization in several emerging economies, which contributed to sustained growth in anthropogenic NO_x emissions during this period. Subsequently, despite a decrease to 3.73 ppbv by 2019, concentrations remained relatively high. The post-2015 decline is broadly consistent with the implementation of stricter emission standards, large-scale installation of pollution control technologies, and structural shifts in energy systems in major emitting regions. These policy-driven reductions partially offset emission growth from ongoing economic activities. In 2020, due to the significant reduction in global economic activities and transportation caused by the COVID-19 pandemic, NO₂ concentrations notably dropped to 3.42 ppbv. This abrupt decline highlights the strong sensitivity of surface NO₂ to short-term changes in transportation intensity and industrial output, underscoring the dominant contribution of anthropogenic combustion sources. From 2021 to 2023, with the gradual recovery of economic activities, NO₂ concentrations rebounded slightly in 2021 but then declined again in 2022 and 2023, reaching 3.38 ppbv in 2023. The post-pandemic evolution suggests that while mobility and economic activity resumed, ongoing structural adjustments in emission sources and continued regulatory efforts may have moderated the rebound in NO₂ concentrations.

Figures S5b and 5c show that the NO₂ concentrations in the Northern Hemisphere and tropical regions exhibited similar trends. This similarity reflects the dominance of anthropogenic NO_x emissions in the Northern Hemisphere, which accounts for the majority of global industrial production, transportation activity, and energy consumption. As a result, large-scale emission control policies and economic transitions in this hemisphere exert a disproportionate influence on global NO₂ variability. In contrast to global trends, the Southern Hemisphere exhibited unique characteristics in the annual average NO₂ concentrations from 2005 to 2023 (Fig. S5d). Between 2005 and 2009, NO₂ levels decreased from 2.83 to 2.66 ppbv, likely due to the initial implementation of air pollution control measures in certain regions. From 2010 to 2015, NO₂ concentrations fluctuated, overall declining from 2.68 to 2.65 ppbv. This period's fluctuations may reflect the dynamic balance between economic activities and air quality management policies. Between 2016 and 2019, NO₂ levels remained relatively stable, ranging from 2.63 ppbv to 2.69 ppbv, indicating a balance between ongoing mitigation efforts and economic growth. In 2020, the COVID-19 pandemic led to a significant reduction in economic activities and transportation, causing NO₂ concentrations to drop sharply to 2.45 ppbv, highlighting the substantial impact of anthropogenic activities on NO₂ emissions. From 2021 to 2023, as economic activities gradually resumed, NO₂ levels increased markedly, reaching 3.15 ppbv in 2022 before slightly declining to 3.11 ppbv in 2023. This upward trend is likely associated with economic recovery, increased traffic, and the relaxation of emission control measures in some areas. Compared with the Northern Hemisphere, the Southern Hemisphere is characterized by lower anthropogenic emission densities, different energy consumption patterns, and a relatively larger influence of natural sources such as biomass burning and lightning-produced NO_x. In particular, interannual variability associated with fire activity and meteorological conditions may play a more prominent role in shaping regional NO₂ patterns. Moreover, the smaller extent of highly industrialized urban clusters limits sustained long-term anthropogenic-driven growth in NO₂ concentrations.

Overall, these results highlight pronounced regional differences in long-term NO₂ evolution and underscore the combined influence of anthropogenic activity, emission control measures, and large-scale external perturbations on global and hemispheric NO₂ concentrations.

We examine the spatial distribution and regional evolution of NO₂ concentrations from 2005 to 2023 to characterize geographic heterogeneity and regional hotspots. Significant spatial heterogeneous in NO₂ concentration across regions have been identified from 2005 to 2023 (Fig. S7), with hotspots in eastern China and northern India. In 2005, regions including the Beijing-Tianjin-Hebei area and the Yangtze River Delta exhibited very high NO₂ concentrations due to their status as major industrial and economic centres. However, after 2013, these regions experienced significant reductions in NO₂ levels, largely attributed to the Chinese government's stringent air quality management policies, including the Air Pollution Prevention and Control Action Plan, which effectively reduced industrial emissions and vehicle exhaust (Feng et al., 2019; Liu et al., 2024; Gao et al., 2022; Geng et al., 2019). In India, NO₂ hotspots are concentrated in major cities located in the Indo-Gangetic Plain such as Delhi and Mumbai. Between 2005 and 2015, NO₂ concentrations in these cities gradually increased due to rapid urbanization and industrialization. However, recent years have seen a decrease in NO₂ levels as the Indian government has intensified efforts to control air pollution (Xie et al., 2024). Conversely, Western European countries like Germany, France, and the United Kingdom had high NO₂ concentrations in 2005, reflecting their industrialized cities and busy transportation networks, and then significantly decreased due to the European Union's strict emission standards and air quality regulations (Guerreiro et al., 2014; Sicard et al., 2021; Georgoulias et al., 2019; Crippa et al., 2016). In North America, the northeastern industrial corridor of the US and Canada, along with major cities like Los Angeles and New York, shared a similar pattern as Western European countries. While overall NO₂ concentrations are lower in Africa and South America, industrial centres and major cities like Johannesburg in South Africa and Sao Paulo in Brazil have shown increased NO₂ levels in certain years. These areas require further pollution control measures to improve air quality.

We further discussed the annual NO₂ changes in seven super-regions (Fig. 2a, b) following the definitions by Global Burden of Disease (GBD) (Fig. S2): South-East Asia, East Asia & Oceania (SEAO), Central Europe, Eastern Europe & Central Asia (CEECA), High-income (HI), Latin America & Caribbean (LAC), North Africa & Middle East (NAME), South Asia (SA), and Sub-Saharan Africa (SSA). SA region has consistently the highest area-weighted NO₂ levels among all regions, followed by SEAO and NAME. NO₂ concentrations in these regions, as well as in CEECA, showed an increasing trend until 2015. In SA, SEAO, and NAME, the rise was primarily driven by rapid industrialization, urbanization, and increased transportation demand, while in CEECA, the increase was also linked to a reliance on fossil fuels and slower implementation of emission control measures. After 2015, NO₂ levels began to stabilize or decline due to local emission reduction efforts and global events like the COVID-19 pandemic.

In contrast, AiTNO₂ concentrations in other super-regions (LAC, HI, and SSA) are relatively low, with annual averages of less than 4 ppbv for the past two decades (Fig. 2a). The LAC region benefits from a combination of lower vehicle density in rural areas and successful urban air quality initiatives in major cities (Riojas-Rodríguez et al., 2016; Huneeus et al., 2020). The lower NO₂ levels in the HI region are largely due to stringent environmental regulations, advanced pollution control technologies, and a significant shift towards cleaner energy sources. SSA presents a unique case where low industrialization and vehicle ownership contribute to lower NO₂ levels. However, the region is still impacted by biomass burning for cooking and heating, which can lead to localized spikes in NO₂ (Fig. S7).

When considering the population exposure, we found that high levels of NO₂ were also observed in some developed regions, such as HI (including North America, Western Europe, Southern Latin America, Asia Pacific, and Australasia), which ranked first among the seven super-regions from 2005 to 2007 (Figs. 2c and S9). The higher population-weighted average of NO₂ in HI was mainly attributed to several population dense countries, including South Korea, Japan, Singapore, Andorra, Germany and Italy (Figs. 2d and S9). Figure 3 illustrates the proportion of super regions that include the top ten countries with the highest population-weighted NO₂ exposure over the past two decades. During the early years of the study period, the HI region accounted for 60 % of the top ten countries in terms of NO₂ exposure. However, this proportion gradually decreased, leaving only one country from the HI region among the top ten in recent years. In contrast, the NAME region experienced a sharp increase in its share of countries with high NO₂ exposure, eventually becoming the predominant area for such countries. This shift underscores the dynamic nature of NO₂ exposure distribution, influenced by both changes in NO₂ emissions and population dynamics.

We estimated that in recent years, SEAO, NAME, and SA are among the top regions with the highest population-weighted NO₂ concentration, with multiple year average of 10.78 ± 1.02, 10.54 ± 0.55, 9.81 ± 0.70 ppbv from 2005 to 2023, respectively. NO₂ exposure was increasing by 0.19 and 0.72 ppbv per decade in NAME and SA, while decreasing by 0.68, 0.40, 2.53, 0.85, 0.12 ppbv per decade in SEAO, CEECA, HI, LAC, and SSA, respectively (Fig. 2c). These contrasting trends highlight substantial temporal changes in population-weighted NO₂ distributions across super-regions, reflecting the combined influence of evolving emission patterns and regional population dynamics.