Introduction

Air pollution poses a major health risk in Europe and worldwide1,2. In 2021, the European Environment Agency (EEA) estimated over 253,000 premature deaths attributed to fine particulate matter (PM2.5), 52,000 deaths to nitrogen dioxide (NO2) and 22,000 deaths to ozone (O3) exceeding the 2021 World Health Organization (WHO) annual limits3. These exposures, both chronic and acute, also increase the risk of cardiovascular and respiratory diseases, allergic reactions, diabetes, cognitive health, and childhood development, among many others4,5. Recognizing these risks, in 2021, the WHO6 issued stricter air quality limits for each of these pollutants separately at different time scales, i.e. annual, peak season, 24 h and daily maximum 8 h, to mitigate both short-term and long-term health impacts caused by air pollutants.

To assess the threat posed by air pollution in Europe, recent compliance studies have predominantly relied on ground-based air pollutant monitoring networks7,8,9. However, these networks, concentrated primarily in urban areas, exhibit limited spatial coverage and fail to comprehensively represent the entire population. While ground-level measurements offer direct, accurate, and reliable real-world data, their spatial averaging and extrapolation introduces biases in exposure assessment. Additional limitations include frequently incomplete daily observation time series values, which can lead to biases when averaging observations from varying numbers of sites per day. Also, data availability from these networks is higher in more recent periods10, leading to inconsistencies in the prior analysis of multi-decadal concentrations changes.

Another key limitation pertains to the conventional analysis of guideline exceedances for each pollutant separately7,8,9. This approach overlooks occurrence of compound air pollution episodes, in which the WHO daily guidelines are simultaneously exceeded for two or more air pollutants. This is a noteworthy omission, as individuals may experience concentrations exceeding safe guidelines for multiple pollutants concurrently, potentially resulting in synergistic health effects that amplify overall health risks11,12. Although some have begun exploring the interactive health impacts of co-exposure to specific combinations of pollutants, such as PM2.5 and O3, further research on other combinations is imperative. Unfortunately, the unavailability of consistent daily ground-level measurements for multiple air pollutants presents a challenge in comprehending the spatio-temporal patterns of population’s co-exposure.

Using models constrained with observations represents a promising solution to these problems11. Global atmospheric composition reanalyses provide multidecadal daily estimates integrating a diverse range of satellite measurements12,13. However, due to their coarse spatial resolution and the lack of integration of surface measurements, these datasets remain affected by significant biases at ground level. The use of air pollution models constrained by surface measurements over multidecadal periods, either for Europe or globally, have mostly focused on long-term averages (annual or monthly values)14,15,16, while models predicting daily concentrations have predominantly focused on a single pollutant, primarily PM2.517,18. Consequently, consistent and accurate air pollution datasets allowing comprehensive understanding of population exposure to multiple air pollutants and its compound episodes in Europe is still lacking.

Governments worldwide are increasingly acknowledging the necessity of addressing air pollutions collectively, such as the integrated control programs in the United States19,20, due to their cost-benefit efficiency, as well as the significant benefits they offer in improving overall air quality and public health. Unfortunately, the lack of spatial-resolved daily estimates over long period for multiple air pollutants obtained from internally-consistent models impedes our understanding of how multiple pollutants and their compound episodes have evolved over time in response to air pollution policies and measures implemented in Europe. Obtaining such crucial information is vital for evaluating the effectiveness of interventions and developing targeted strategies to mitigate the health risks associated with the exposure to multiple air pollutants.

This study uses Quantile LightGBM (QLG) machine learning models21 to link ground-level station data of daily mean PM2.5, PM10, NO2, and O3 (the primary four air pollutants contributing to mortality3) concentrations with meteorological and air quality reanalysis data, aerosol optical depth (AOD) model estimations and ground-level emission data. The models estimate daily concentrations from 2003 to 2019 at a spatial resolution of 0.1°, which are then used to estimate regional population-weighted (PW) averages for 1426 NUTS3 regions in 35 European countries. These estimations were used to estimate the spatial heterogeneity and temporal evolution of (i) air pollution concentrations and the (ii) population count and (iii) cumulative time of exposure to concentrations exceeding the 2021 short-term and long-term WHO guidelines. Moreover, we analyzed the joint exceedance of WHO limits simultaneously for two or more air pollutants, providing a comprehensive assessment of compound events. This study contributes to the assessment of overall air quality in Europe within the framework of the new WHO short and long-term guidelines, and identifies spatiotemporal patterns of compound event days. This information is crucial for environmental health assessments and policymaking aimed at mitigating the health risks associated with air pollution in the European Union and the whole continent.

Results

Data validation

Our models demonstrate robust spatial cross-validation performance (see Fig. 1, Figs. S2, S3 in Supplementary Information) in estimating the European ground-level concentrations of PM2.5, PM10, NO2 and the maximum daily 8h average of O3 (here referred to as MDA8 O3 for simplicity), with a NRMSE (Normalized Root Mean Square Error) of 1.85%, 2.71%, 8.99% and 3.20%, and a Pearson correlation of 0.80, 0.79, 0.79 and 0.90, respectively. PM2.5 and O3 estimations were nearly unbiased (NMB (Normalized Mean Bias) = −0.9% and 0.24%, respectively), while PM10 and NO2 were slightly underestimated (NMB = −3.81% and −2.00%, respectively). Table S4 (Supplementary Information) shows that the model estimations clearly outperform reanalysis data from CAMSRA12 and MERRA-213, and Table S5 (Supplementary Information) that the temporal cross-validation of the model estimates is consistent over the whole period. The comparison between whole-period averages of observed and estimated daily values is shown in Fig. S1 (Supplementary Information), and the mean, standard deviation, median, inter-quartile range, trend, and Pearson correlation in Northern, Southern, Western and Eastern Europe in Table S6 (Supplementary Information).

Fig. 1: Validation of estimated pollutants.
figure 1

Comparison between observed and model-estimated PM2.5 (a), PM10 (b), NO2 (c), MDA8 (maximum daily 8 h average) O3 (d) concentrations from 2003 to 2019 under spatial cross-validation.

Full size image

Population-weighted concentrations and trends

Figure 2 depicts the long-term averages (left panels) and trends (right) in PW concentrations. Whole-period continental averages for PM2.5, PM10, NO2 and MDA8 O3 were 14.34, 22.01, 13.46, and 74.51 μg/m³, respectively. PM2.5 and PM10 was higher in Northern Italy and Eastern Europe, with high PM10 additionally in Southern Europe. High NO2 was mainly observed in Northern Italy and some areas of Western Europe, such as in the south of the United Kingdom, Belgium and the Netherlands. MDA8 O3 was latitudinally orientated, with the highest concentrations in the Mediterranean. PM2.5, PM10 and NO2 generally decreased in most of Europe, with an average annual rate of −1.72%, −2.72% and −2.45%, respectively. The most important reductions in PM2.5 and PM10 were observed in Central Europe, while for NO2 they were found in mostly urban areas of Western Europe, which correspond to the areas with the highest concentrations. In contrast, MDA8 O3 increased by 0.58% in Southern Europe, while it decreased or showed nonsignificant trend in the rest of the continent.

Fig. 2: Spatial Distribution and Trends of air pollution in Europe.
figure 2

17-year Mean Population-Weighted Concentrations of estimated PM2.5 (a), PM10 (c), NO2 (e), and MDA8 (maximum daily 8h average) O3 (g) (Unit: μg/m³), and its Average Annual percentages Changes (in %, calculated using Theil-Sen slope dividing mean estimates) (b, d, f, h).

Full size image

Cumulative time of exposure

Figure 3 depicts the year-to-year time-series and whole-period average maps of the short-term unclean air exposure time (orange bars and maps) and the percentage of population living in short-term clear air areas (blue curves). Here, the annual unclean air exposure time represents the PW average annual number of days in which the WHO daily limit for an air pollutant is exceeded, while population in clean air areas represents percentage of people living in areas where air quality meets recommended standards. Detailed definitions are provided in the methodology section. Overall, we observed a consistent decrease in unclean air exposure time for PM2.5, PM10 and NO2 throughout Europe, with approximately 60, 46, and 48 fewer unclean air days in 2019 compared to 2003, respectively. For MDA8 O3, it generally increased, with the exception of the extreme year of 2003. In 2003, under the record-breaking summer temperatures22, O3 levels were similar to those registered within the period 2015–2019. Higher values of unclean air exposure time were found in Eastern Europe and Northern Italy for PM2.5 and PM10, in mostly urban regions (particularly in Western and Central Europe and Northern Italy) for NO2, and in Southern and Eastern Europe for MDA8 O3.

Fig. 3: Population exposure to short-term air pollutants in Europe.
figure 3

Annual unclean air exposure time (Unit: Days, presented by bar plots) exceeding WHO Daily Limits, and the population (%) in short-term clean air areas (depicted by the blue line) for PM2.5 (a), PM10 (c), NO2 (e), and MDA8 (maximum daily 8 h average) O3 (g) in Europe. The spatial distribution of 17-years average annual unclean air exposure time (b, d, f, h). Short-term clean air areas here indicate those regions with 17-Year average annual unclean air exposure time Less Than 4 Days (WHO standard).

Full size image

Population in short-term or long-term exposure to clean air areas

As expected, in general terms, the trend of the annual percentage of population in short-term clean air areas was found to be negatively correlated with the evolution of the unclean air exposure time. Among the four pollutants, the population in short-term clean air areas for PM10 exhibited the largest increase, rising from 8% in 2003 to 77% in 2019, an increase that represents 367.9 million people. The population living in clean air areas for NO2 and PM2.5 also increased from 21% to 49% and from 3 to 11%, respectively, corresponding to an increase of 141.2 and 41.8 million people compared to 2003, respectively. Changes in MDA8 O3 exhibited significant annual variation. By 2019, the percentage of population in short-term clean air areas dropped from 62% to 26%, equivalent to around 219 million fewer people compared to 2004 (note: we here exclude 2003, which was an exceptional year in terms of O3 concentrations). Most short-term clean air areas were mainly located in Northern Europe, Scotland, the island of Ireland and Northern Spain.

As a consequence of the general decline in air pollution levels, we found an increasing trend in the population residing in long-term clean air areas of PM2.5, PM10, and NO2 (Fig. 4a–c), reaching 10.1, 102.8 and 75.9 million people in 2019, respectively. These numbers constitute about 1.90%, 19.85%, and 14.66% of the total population, compared to around 0.05%, 2.26%, and 2.88% corresponding to 2003. The distribution of the population living in long-term clean air areas is spatially heterogeneous. For example, the long-term clean air population for PM10 and NO2 grew faster in western Europe compared to other areas, while the increasing trend for PM2.5’s clean air population was faster in northern Europe. The year-to-year changes in population living in long-term clean air areas was found to be sometimes abrupt, as soon as densely populated areas started to comply with the regulation limits. Regarding MDA8 O3, almost no areas meet the WHO standard of 60 μg/m3. Therefore, we adopted the WHO interim target 2 (70 μg/m3) as the reference threshold, which showed no clear trend over the years (Fig. 4d). Most long-term clean air population for MDA8 O3 is found in western Europe.

Fig. 4: Population exposure to long-term air pollutants in Europe.
figure 4

Annual Population in long-term clean air areas (million, depicted by bar plots) and annual concentration (μg/m³, represented by blue line) in Europe for PM2.5 (a), PM10 (b), NO2 (c), and MDA8 (maximum daily 8 h average) O3 (d).

Full size image

Population exposure to compound episodes

In Fig. 5, we analyzed the annual unclean air exposure time for the joint exceedance of multiple air pollutants, or compound event days. Generally, the annual unclean air exposure time for various types of compound event days decreased significantly in Europe, dropping from 78.5 days to 17.4 days. During 2012–2019, around 86.26% of the European population experienced at least one day per year with compound event days, which is approximately 10% lower than the figures from 2003 to 2011 (see Table S7 in Supplementary Information). We identified four primary types of compound event days in Europe, namely PM2.5-NO2, PM2.5-PM10, PM2.5-O3 and PM2.5-PM10-NO2 days, which collectively accounted for over 94.6% of all compound event days during the entire study period (see Fig. 5). Notably, compound event days also played a particularly important role in contributing to unclean air days, accounting for over 87% and 88% of unclean air days for PM10 and NO2, respectively (see Figure S4 in Supplementary Information).

Fig. 5: Composition changes of multi-pollutant compound episodes in Europe.
figure 5

The composition of annual compound unclean-air exposure days for multiple air pollutants across 17 years (a), with further focus on three pollutants combinations (b).

Full size image

The decreasing trends in compound event days are consistent across different subcontinental domain, although their contributions of compound event days vary (Fig. 5). Eastern Europe was found to be dominated by PM2.5-PM10 days, while PM2.5-NO2 days were more frequent in Western Europe. Southern Europe experiences a wider variety of types of compound event days, mainly PM2.5-NO2, PM2.5-O3 and PM2.5-PM10-NO2. Although most compound event days are decreasing over decades, PM2.5-O3 days is the only one that increased (see Table s8 in Supplementary Information), going from 4.62 to 5.30 days per year when comparing the periods from 2012–2019 to 2003–2011. And PM2.5-NO2 declined more slowly than other major types of compound event days. Consequently, PM2.5-NO2 and PM2.5-O3 days became the two predominant types of combinations in Europe during 2012–2019. Compound event days also exhibited a clear seasonal pattern in Fig. S5 (Supplementary Information). Compound event days with unclean levels of O3 were more common from March to October, while those compounds involving PM2.5, PM10, or NO2 tend to occur during colder seasons. Additionally, Figs. S6, S7 (Supplementary Information) illustrates a noticeable year-to-year decreasing trend in compound event days involving PM2.5, PM10, or NO2.

Discussion

In general, our study provides a comprehensive assessment of spatial and temporal inequities in population exposure to air pollutants in 1426 regions across 35 European countries, representing 543 million people. Our findings reveal a substantial reduction in European population exposure to most air pollutants. However, PM2.5 and O3 levels continue to surpass WHO guidelines in numerous regions, resulting in a relatively higher number of people exposed to unclean air levels. Moreover, our assessment of compound event days showed annual occurrence of compound event days decreased from 78.5–17.4 days over 2003–2019, but over 86.3% of the European population still experienced at least one compound event days per year in 2012–2019. PM2.5-O3 was the only compound event days that increased and became the second most frequent type of compounds in Europe during 2012–2019. Overall, our findings present comprehensive evidences of both short and long-term exposure to the main pollutants with largest impact on human health and mortality, by performing an exhaustive continental-wide regional analysis not restricted to urban settings only. Additionally, it introduces valuable insights into compound event days involving these pollutants, significantly enriching our understanding of multi-hazard exposure, and potentially guiding air pollution management policies.

Our research fills a critical gap in the literature by offering daily estimations of multiple air pollutants in Europe over the period 2003–2019, including PM2·5, PM10, NO2, and MDA8 O3. Unlike prior studies providing annual or monthly estimations over multidecadal periods14,15,16, our daily air pollution estimations fill a crucial need for detailed data (either short-term or long-term) essential for conducting health impact studies and environmental monitoring. While prior studies focused mainly on single-pollutant estimations, mostly on PM2·512,13, our study simultaneously provides estimations for multiple pollutants with enhanced predictive accuracy, achieving a strong correlation coefficient of approximately 0.79 to 0.90 for spatial cross-validation and 0.81 to 0.91 for temporal cross-validation. For example, Lary et al. estimated daily PM2·5 concentrations globally over the period 1997–2014 by using remote sensing and meteorological data17, with a correlation coefficient of 0·52–0·75. Moreover, Yu et al. used deep ensemble machine learning to estimate global daily PM2·5 concentrations in 2001–201918, with Spearman correlation of around 0.76 with ground-level observations throughout Europe. Furthermore, our estimates also outperform CAMSRA and MERRA-2 reanalysis (see Table S4 in Supplementary Information).

Regarding the concentrations of the pollutants, we observe the largest declines in PM10 in most of Europe, with an approximate annual decrease of 2.72%, followed by NO2 (2.45%) and PM2.5 (1.72%). Instead, we find that MDA8 O3 rose by about 0.5% per year if the outlier representing year 2003 is excluded. These trends align with previous studies, which reported annual declines of around 1.7–2.2% in NO223,24 and 1–2% in PM2.514,24,25, as well as undefined trend between (−0.3 and +0.5%) in MDA8 O324,26, over last two decades in Europe. These trends were also in agreement with10 reanalysis products (CAMSRA and Merra-2) and ground-level observations11, with an annual average decrease of 2.1–3.3% in PM10, 2.3–2.5% in NO2, 0.9–1.7% in PM2.5 and a 0.1–0.9% annual increase in MDA8 O3. Overall, our estimations provide further evidence of the slight upward trend of MDA8 O3 in Europe over the last decades, when other pollutants decreased under the European Union’s (EU) efforts to implement air quality control measures. Notably, this upward trend of MDA8 O3 is latitudinally oriented, and largely related to temperatures and sunlight. These conditions promote the formation of O3 from precursor pollutants like nitrogen oxides (NOx) and volatile organic compounds (VOCs). Previous studies27,28,29 suggested that the reduction of NOx may have alleviated O3 depletion in and around cities, particularly at night, due to lower titration of O3 by NOx. Moreover, these studies underscore the necessity of prioritizing stronger control measures on VOCs over NOx for effective urban O3 mitigation27,28.

Previous analyses with WHO 2021 guideline7,8,9,30 have primarily focused on urban areas, constrained by limited monitoring stations. These studies often faced inconsistencies in assessing different air pollutants against WHO guidelines, discrepancies in the availability of daily observations across different pollutants over space and time. Our study overcome these limits by providing full-coverage estimations covering European population exposure and time, enhancing a more thorough understanding of spatial and temporal disparities related to WHO guidelines. It highlighted a notable decrease in European population exposure to PM2.5, PM10 and NO2, contrasting with the rise in MDA8 O3 exposure. Additionally, the average exposure time and population exposed to unclean air areas for PM2.5 and O3 is much higher than for the other two pollutants, highlighting the urgency for greater control for these pollutants. Furthermore, recent studies31,32 have also linked MDA8 O3 exposure exceeding WHO daily limits to substantial rises in hospital admissions for heart attacks, heart failure, and strokes. This further emphasizes the importance of addressing the increasing role of O3 exposure, especially in the context of rapidly increasing threats from climate change in Europe.

Despite significant progress in reducing air pollution, our assessment found that over 85% of Europeans still experienced at least one day per year with compound event days, notably prevalent in Eastern and Western Europe (see Table S7 in Supplementary Information). It highlights the persistent need for heightened attention to exposure of compound event days. We also found that PM2.5-O3 days have become the second most prevalent category of compound event days in Europe, with their contribution increasing from 4.43% in 2004 to 35.23% in 2019. Recent increases in PM2.5-O3 days, especially in lower latitudes during warm seasons, are likely linked to climate change and complex interplay between PM2.5 and O3. Emission sources such as vehicle exhaust and industrial processes release both PM2.5 and O3 precursors like VOCs and NOx. Global warming intensifies sunlight and raises temperature, particularly in summer, accelerating O3 formation through photochemical reactions33. Subsequently, higher levels of O3 will oxidize volatile organic gases or secondary organic aerosols in the atmosphere34, leading to the condensation of certain oxidized compounds, ultimately forming secondary PM2.5 particles. Also, climate change increases the likelihood of wildfires, contributing to elevated O3 and PM levels35. Lastly, biogenic VOCs (BVOCs) have been identified as second largest sources of the O3 production in summer36. The emission rate of BVOCs also rises with increasing temperatures, reaching peak levels at around 38–40 °C37, due to heightened metabolic activity in vegetation.

Ozone management presents a complex challenge due to its secondary formation pathway. Conventional air pollution control strategies, which focus on reducing primary pollutant emissions, may not be sufficient to effectively mitigate O3 exceedances and associated compound event days. However, addressing climate change, which influences ozone formation through increased sunlight and rising temperatures, is crucial for long-term ozone management and protection of public health. This approach not only slows global warming but also curtails the rise of O3 formation triggered by photochemical reactions in warmer seasons. Moreover, surface or tropospheric O3, beyond impacting air quality, acts as a greenhouse gas. Its ability to absorb infrared radiation contributes to the trapping of heat in the lower atmosphere. By reducing tropospheric O3 levels, we can help mitigate its role in the greenhouse effect, potentially breaking the cycle that leads to further O3 generation. Implementing policies to prevent and manage wildfires can help in controlling the release of these compounds into the atmosphere, thereby reducing O3 formation. Lastly, vehicles stand as the most prominent contributor to anthropogenic VOC emissions36. Implementing rigorous policies to control and diminish VOC emissions from vehicles can notably impact O3 formation, particularly in urban areas characterized by dense vehicular traffic. Additionally, choosing low-BVOCs emission plants for urban green spaces also aids in mitigating BVOCs emissions, further improving air quality and reducing O3 precursors.

while EEA8 and WHO7 provided varying compliance estimates for long-term unclean air populations in urban settings (see Table S9 in Supplementary Information), possibly due to differences in the analyzed urban areas, countries, periods or calculation methods, our analysis suggests around 98.10%, 80.15% and 86.34% of the population in the 35 European countries lived in 2019 in unsafe air areas for PM2.5, PM10 and NO2, respectively. These results align closely with EEA’s urban estimates of 97%, 81% and 94% for the 27 countries of the European Union (EU-27), respectively. However, our NO2 estimates are more consisted with WHO’s estimates in boarder human settlements settings, possibly due to wider inclusion of population beyond urban areas in EEA’s analysis. Notably, urban areas are more susceptible to experience higher NO2, primarily driven by emissions from vehicles and residential sources30,38. Spatially, Northern Europe exhibits a significantly higher population proportion living in long-term clean air areas for PM2.5, PM10, and NO2 compared to the rest of the continent (see Fig. S7 in the Supplementary Information). Additionally, with the introduction of the new long-term guideline for O3 in 2021, the EEA8,30 started to report non-compliance in all countries with the peak season O3 standard in 2021 and 2022, which concurs with our findings from 2003 to 2019. These results underscore the significant improvements made in European air quality control for PM10 and NO2, while challenges in controlling O3 levels underscore the need for a policy shift. Addressing global warming and air quality together with more comprehensive solutions is essential, requiring a macro perspective to collaborate with policymakers for effective action.

This study has several limitations worth acknowledging. Although we have conducted spatial and temporal cross-validations to assess the quality of our air pollution estimates, biases might persist due to the uneven distribution of ground-level stations and the limited number of observations in earlier periods. Also, the population exposure in this study does not include population changes within a year. Despite these limitations, our study serves as a solid foundation for future research and policy development addressing air quality management and public health concerns throughout Europe.

Methods

Exposure estimation and validation

In this study, we trained four separated Quantile LightGBM (QLG) models21 for daily PM2.5, PM10, NO2 and MDA8 O3, with ground-level measurements from European environment information and observation network (Eionet). These four individual models for each pollutant are developed separately to maximize the information gathered from the varying numbers of background monitoring sites for each pollutant. Specifically, we used 1310, 2438, 1867 and 2021 sites for PM2.5, PM10, NO2 and MDA8 O3. To train these models, we gathered data from multiple sources, which are further described in the Supplementary Information (pp 3–6). The datasets encompassed various atmospheric aerosol data, like model predictions of size-resolved aerosol optical depth (AOD)39, Atmospheric composition reanalysis data from CAMSRA12 and MERRA-213. Additionally, we incorporated reanalysis data from ERA5-land and ERA540, Gridded climate observations from E-OBS41, Land use data, including road density data from GRIP global roads database42, vegetation-related data from ERA5-land, Köppen-Geiger climate classification and local climate zone data from world urban dataset43, Emission data from CAMSRA global emission inventories. These datasets spanned from January 1, 2003 to December 31, 2019, with different spatial and temporal resolutions provided in Table S1(Supplementary Information).

We computed daily averages if the data were originally available at hourly or 3-hourly resolution. To ensure consistent spatial resolutions, all continuous gridded data were bilinearly resampled to a horizontal resolution of 0.1° × 0.1°. For the Köppen–Geiger climate classification (approximately 0.08° × 0.08°), nearest neighbor interpolation was applied during resampling. Regarding the local climate zone data (1 km), resampling involved the use of the most frequent category. Subsequently, to align with the stations’ observation, we extracted modeling data within a 0.05-degree buffer around each station’s location. This extraction process employed the area-weighted average for continuous variables and the dominant category for categorical data.

To select the most relevant features for each air pollutant model, we employed the Boruta feature selection procedure44. This method considers interactions and nonlinear relationships during the selection of variables, making it robust and efficient for removing noise45. The selected variables for each air pollutant model are listed in Table S2(Supplementary Information), and the 20 most important variables are listed in Figure S8(Supplementary Information).

As ground-level monitoring stations tend to be in and around urban areas30,46, this may weaken the capacity of the model to estimate the concentrations in regions farther away from these sites, while causing overfitting in areas with higher station density. We employed a distance-weighted loss function (Supplementary Information pp9) during model training to address this issue and ensure that the model properly represents the areas with fewer monitoring sites. This approach involved assigning weights to the loss function based on the normalized distances between each site and its nearest neighboring sites. By doing so, we aimed to mitigate the potential biases associated with the non-uniform distribution of monitoring stations across the study area.

To evaluate the out-of-sample predictive capacity of the models, we used two different approaches. First, we randomly selected 10% of the ground-level sites as test sites to validate the model performance. Second, we conducted nested 5-fold cross-validation to obtain spatial and temporal out-of-sample predictions separately. For spatial out-of-sample predictions, we randomly divided the monitoring sites into five equal-sized subsamples. In each loop of predictions, four subsamples were used for model training and tuning, while the remaining subsample was used to obtain the out-of-sample predictions. For temporal out-of-sample predictions, we split the 17-year period into six subperiods consisting of three or two consecutive years. After obtaining these out-of-sample predictions, we calculated validation metrics (Supplementary Information pp9) such as the Pearson Correlation, NMB, and NRMSE to assess the performance of the models.

Indicators calculation

We developed indicators describing three main aspects: (i) air pollution concentrations and the (ii) population count and (iii) cumulative time of exposure for individuals to air pollution values exceeding the guidelines (formulas and threshold of WHO guidelines (Table S3) are provided in the Supplementary Information, pp9–11):

We used the Re-Gridded Population of the World Version 4 (GPWv4) to calculate the daily PW regional concentrations for PM2.5, PM10, NO2 and O3 for all the grid-cells included in each of the 1426 NUTS3 regions. Further details are provided in the Supplementary Information (pp9). Changes over time were compared to annual PW concentrations with the Theil-Sen slope10,24, which estimates the annual rate of change by taking the median of all possible pairwise slopes between data points. The Theil-Sen slope is less sensitive to outliers and is suitable for analyzing time series data with potential fluctuations and irregularities. Additionally, we used the Mann-Kendall test to evaluate the significance level of the trends24.

We developed an indicator to represent the cumulative time of exposure for individuals to air pollution values exceeding the guidelines. Thus, the annual unclean air exposure time represents the PW average annual number of days in which the WHO daily limit for an air pollutant is exceeded. The indicator involves calculating person-days surpassing daily limits for each grid-cell annually, aggregated them over a region or set of regions (unit: person*day), and finally dividing by the total population of the region(s) (unit: day). The annual unclean air exposure time can be calculated for one individual air pollutant, or for two or more air pollutants simultaneously exceeding the WHO limits the same day and grid-cell (for compound event days). More details are provided in Supplementary Information (pp 10, 11).

To assess the population exposed to air pollution values in relation to established guidelines, we defined indicators called “Population in clean air areas” and “Population in unclean air areas”, as the percentage of people living in areas where air quality either meets or exceeds recommended standards. Notably, ‘clean air’ aligned with WHO standards widespread adopted by many governments, balancing current expenditure and practicality against health benefits, but not imply entirely safe air level. For short-term guideline, we imposed that daily or 8h maximum values are met 99% of days in a given year, while for long-term limits, annual or peak season limits are not exceeded. Complementary criteria were used for unclean air areas. The mathematical formulation is described in the Supplementary Information (p10, 11).