Introduction

Exposure to ambient particulate matter is recognized as a leading global health concern1. Over the past decades, extensive epidemiological studies have investigated the relationship between mortality risk and long-term exposure to mass-based metrics, e.g. ambient PM2.5 and PM102,3,4,5. However, toxicological studies suggest that the relevant dose metric may not only be related to particle mass6,7, but also dependent on the size, shape, and surface properties of particles8,9,10,11,12.

Ultrafine particles (UFPs), defined as particles with diameters less than 100 nm, pose significant health risks due to their small size and high specific surface area, which allow them to easily penetrate and contact with human tissues and organs13,14. Despite their substantial contribution to the total number of particles, UFPs contribute minimally to overall mass15,16,17, rendering mass-based metrics inadequate for accurately characterizing their influence. Limited studies indicate a potential association between increased UFP exposure and adverse effects in vulnerable populations, e.g. elderly and children, particularly those with respiratory and cardiovascular diseases7,18,19,20. However, definitive health effects of UFPs remain inconclusive, primarily due to insufficient exposure data on spatial variations in UFP concentrations2,21,22.

The World Health Organization (WHO) has recommended in its recent Global Air Quality Guidelines (AQG)23 that ambient UFPs should be quantified by particle number concentration (PNC). The WHO also advocates using emerging science and technology to enhance methods for assessing UFP exposure, which is crucial for epidemiological studies and effective UFP management. Monitoring PNC is inherently more challenging than measuring mass-based metrics due to the predominance of nanoscale particles and the need for high resolution and sensitivity. The transient nature of UFPs also leads to highly variable results depending on time and location. This process requires expensive, research-grade instruments24, resulting in limited monitoring data. Consequently, the development of exposure assessment methodologies for PNC is still in its early stages25,26.

To determine the spatial variability of ambient concentrations, atmospheric dispersion models have been employed to evaluate the UFP exposure in major European cities15, with a particular focus on urban areas27,28. However, the uncertainty associated with particle number emission inventories is significantly higher than that of commonly regulated pollutants15,29,30,31,32,33,34. Additionally, the complex microphysics of number-based aerosol dynamics introduces substantial uncertainties and high computational costs35,36. As an alternative, data-driven methods, e.g. land-use regression (LUR) models, have been utilized to assess PNC exposure. In Amsterdam, a measurement campaign collected PNC data from 50 homes over a one-week period each, marking the initial efforts to evaluate PNC exposure using LUR models37. Inspired by the Measurements of UFP and Soot in two Cities (MUSiC) and the European Study of Cohorts for Air Pollution Effects (ESCAPE) projects, LUR models have increasingly been applied for small-scale PNC exposure assessment in European cities38,39,40, as well as in North American cities4,41,42 and in China43.

Recently, machine learning (ML) methods have been proposed for assessing UFP exposure. Random forest (RF) machine learning models have been integrated with LUR to improve UFP exposure assessment24,44. Additionally, LUR has been combined with Deep Convolutional Neural Network (CNN) models to generate high-resolution models of annual median ambient PNC using aerial images4. These studies have demonstrated that ML methods outperform traditional LUR models44.

Most existing models primarily rely on geographic information (e.g. land use, road and traffic, and population density). However, they do not fully utilize the extensive data available on regulated pollutants (e.g. NOx, Ozone, PM2.5, and PM10) from both monitoring and air quality modeling sources. By leveraging this abundant data, emerging ML methods can more effectively capture the complex relationships between UFP and these regulated pollutants, potentially leading to significant improvements in the accuracy and performance of UFP exposure assessments.

This study presents the development of an innovative stacking-based ensemble machine learning framework that integrates data-driven and physical-chemical models to conduct a national-scale assessment of UFP exposure at high spatial (1 km) and temporal (1 h) resolutions, encompassing urban centers, urban clusters, and rural areas. The globally unique long-term standardized UFP measurements in Switzerland and the superior design of the model significantly enhanced its generalization capabilities. High spatial and temporal resolutions were achieved by leveraging the Copernicus Atmosphere Monitoring Service (CAMS) validated air quality reanalysis dataset, in conjunction with meteorological data from the ECMWF Reanalysis v5 (ERA5) and the traffic volume with 100 m resolution from the Open Transport Map (OTM)42. The high-resolution PNC obtained from the model was employed to comprehensively assess the national UFP exposure across different area types. It reveals a relatively high exposure to UFP with the national average of (9.3 ± 4.7) × 103 particles‧cm−3. We highlighted a nonlinear relationship between the annual cumulative durations of high UFP exposure respectively based on the 24-h mean and the 1-h reference levels recommended by the WHO AQG, and disparities among various types of areas. This study promotes the adoption of innovative ML models as a cost-effective method for evaluating UFP exposure, utilizing available measurements of regulated pollutants. Furthermore, high-resolution UFP data allowed detailed exposure analyses, revealing crucial patterns and informing public health. It enhances our understanding of UFP exposure on a national scale and provides valuable insights for future UFP standards.

Results

Data-driven model for particle number concentration

A data-driven Stacking Technique for Ensemble Modeling of Particle Number Concentration (Stem-PNC) has been proposed to estimate PNC based on regulated pollutants (i.e. NOx, PM10, PM2.5, and Ozone), in conjunction with meteorological parameters (i.e. wind speed, temperature, radiation, relative humidity, and precipitation) and traffic information. Stem-PNC was trained using the globally unique, long-term standardized measurements of PNC concentrations, from the Swiss National Air Pollution Monitoring Network (NABEL). Since 2005, NABEL has measured the number concentration of particles with diameters from 5 nm to 3 μm45. Stem-PNC was trained on hourly PNC measurements collected over four years (2016–2019, covering 78% of the data) and tested on data from 2020 (22%). 5% of the training dataset was allocated as a validation set for hyperparameter tuning, ensuring that the test dataset remained completely independent for the final evaluation. The model utilized pollutant datasets from NABEL, meteorological data (MeteoSwiss), traffic data, and temporal features. Both training and predictions were conducted on data collected from monitoring stations. Additional details on Stem-PNC and the training dataset are provided in the “Methods” section.

We applied Stem-PNC to estimate the particle number concentrations at the five NABEL stations for the year 2020. Despite the significant changes in PNC distributions due to the COVID-19 pandemic, which differed markedly from the 2016–2019 training dataset, Stem-PNC successfully predicted the weekly temporal trends of PNC at all five sites, as illustrated in Fig. 1a. Quantitative evaluation of the model performance is shown in Fig. 1b–e. The model accuracy improved with longer averaging periods, as indicated by the increase in the coefficient of determination (R2) from 0.85 for hourly averages to 0.92 for monthly averages. This suggests that Stem-PNC is well-suited for long-term exposure assessment. The strong agreement between Stem-PNC predictions and actual measurements highlights the robust generalization capabilities of the model.

Fig. 1: Performance evaluation of the Stacking Technique for Ensemble Modeling of Particle Number Concentration (Stem-PNC).
Fig. 1: Performance evaluation of the Stacking Technique for Ensemble Modeling of Particle Number Concentration (Stem-PNC).
Full size image

a Comparison of measured and Stem-PNC predicted weekly mean particle number concentrations in 2020 at the five stations: Basel-Biningen (BAS), Bern-Bollwerk (BER), Harkingen-A1 (HAE), Lugano-University (LUG), Rigi-Seebodenalp (RIG). The shaded area represents the standard deviation. The daily and monthly results in 2020 can be found in Supplementary Note 3. The quantitative comparisons are presented for different averaging periods: b hourly results; c daily average results; d weekly average results; e monthly average results. Dashed lines represent 1:2 (or 2:1) ratio lines. f Comparisons among different data-driven machine learning models for particle number concentration, including Linear Models, Support Vector Machines, Neighbor-based Models, Tree-based Models and Deep Learning Models and Stem-PNC. The error bars represent the standard deviation. Statistical metrics: Coefficient of Determination (R2), Mean Bias (M-Bias), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE). More details about metrics formulas, stacking structure, model categories, hyper-parameterization, and more prediction results in each category are in Supplementary Note 2. Stem-PNC was trained on hourly PNC measurements from 2016–2019 (78% of the data, with 5% used for hyperparameter tuning) and tested on 2020 data (22%), using pollutant data from National Air Pollution Monitoring Network (NABEL), meteorological data from MeteoSwiss, traffic data, and temporal features.

These capabilities are attributed to three synergistic factors: the extensive training datasets, the inclusion of regulated pollutants in the model, and the stacking technique. The large training dataset enhances the generalization ability by providing a diverse set of PNC distributions, reducing the likelihood of overfitting, and enabling the model to capture accurate underlying relationships between PNC and regulated pollutants. The stacking technique further boosts generalization by integrating the predictive strengths of multiple models, resulting in more accurate and reliable predictions of unseen data.

Figure 1f demonstrates that Stem-PNC outperforms a range of models, from simple linear models to more advanced Deep Learning models. Notably, despite its relatively lower complexity compared to the Deep Learning framework, the performance of the Stem-PNC stacking model is comparable with the Deep Learning model. This is evidenced by a higher R2 (0.845), a lower Root Mean Square Error (RMSE = 4594), and a significantly reduced Mean Bias (124). However, deep learning models have higher hardware requirements, such as the need for A6000 GPUs, compared to the Stem-PNC model, which can operate effectively on less specialized hardware. These findings highlight the cost-efficiency and practical utility of the Stem-PNC stacking model, especially in resource-constrained environments. Supplementary Notes 2.2 and 2.7 provide detailed information about the hardware requirements and computational costs for each model.

As discussed in Supplementary Note 2.10, among the deep learning models, the CNN with sliding time windows achieved an R2 of 0.848, slightly outperforming the Stem-PNC model (0.845). However, Stem-PNC maintained the lowest mean bias (123.5), likely due to the robustness of its ensemble approach. While these results suggest that deep learning models, e.g. CNN, did not significantly outperform the Stem-PNC model, this should be interpreted with caution. The current findings do not necessarily indicate that deep learning methods are inherently less effective; rather, the relatively small and spatially sparse PNC dataset may limit their full potential. With a larger and more spatially comprehensive dataset, the performance of deep learning models, particularly CNNs, may improve substantially, allowing them to better exploit complex spatial and temporal patterns in the future.

Generalization ability of Stem-PNC

In this study, we tried to improve the spatial performance mainly through three aspects. Firstly, we implemented a data pooling strategy by combining measurements from different stations during model training. This approach enhances the model’s ability to identify broader spatial patterns, thereby improving its generalizability across diverse locations. Secondly, we applied a random time sequence for training by shuffling the temporal order of the data. This allows the model to focus on capturing the relationship between PNC and the immediate input variables without being influenced by temporal patterns or trends. This approach reduces the risk of overfitting to temporal variations. Thirdly, we integrated data-driven and physical-chemical models to strengthen temporal and spatial performance by leveraging the strengths of both approaches. Physical-chemical models provide temporally and spatially consistent inputs that support robust generalization across different times and areas, while the data-driven model captures complex relationships between PNC and regulated pollutants, e.g. NOx, PM10, PM2.5, and Ozone.

To evaluate the spatial and temporal generalization capabilities of Stem-PNC, a five-fold station-holdout cross-validation method was employed. In this approach, data from each of the five monitoring stations were sequentially excluded from the training dataset, ensuring that each station’s data served as an independent test set in one of the folds. Consequently, the model was trained on the data from the remaining four stations with the data between 2016 and 2019 and then evaluated on the withheld station’s data from 2020. The detailed configurations of the cross-validations are shown in Table S10 in Supplementary Note 5.

By excluding one station during each training iteration, this method provides a rigorous test of the model’s spatial generalization capabilities, as the excluded station’s data has never been used in training and is reserved solely for evaluation purposes. This approach simulates real-world conditions by predicting PNC levels at an unseen location, ensuring that the model learns generalizable spatial patterns rather than overfitting to specific stations. This approach provides a robust evaluation of the model’s performance across various spatial domains. Further details are provided in Supplementary Note 4. This procedure facilitated a comprehensive assessment of the model’s generalization performance over both spatial and temporal dimensions. The process was repeated iteratively for each station, providing a thorough analysis of the Stem-PNC model’s capacity to generalize across varied spatial domains and temporal intervals.

Figure 2 illustrates that Stem-PNC demonstrated commendable performance on the independent test sets, corresponding to the data from stations sequentially excluded in each fold of the validation process. For the BER (urban, roadside) and LUG (urban, background) stations, the predictions generated by Stem-PNC exhibited strong concordance with the actual measurements, as evidenced by the scattering of data points along the 1:1 reference line. This alignment suggested that Stem-PNC possessed robust spatial and temporal generalization capabilities within urban environments, which are critical for assessing exposure to PNCs. Conversely, at the HAE (rural, next to motorway) and RIG (rural, altitude > 1000 m) stations, discrepancies were observed wherein Stem-PNC respectively underestimated and overestimated the PNCs when the data of these stations were excluded from the training dataset. These outcomes were anticipated due to distinct environmental characteristics: the proximity of HAE to a major motorway results in higher emission levels, while RIG, being in a rural and elevated locale, typically registers at the lower end of the PNC spectrum. Such extremities in PNCs at HAE and RIG represent the upper and lower bounds of particle number distributions observed across the network, likely mirroring the broader national context of Switzerland.

Fig. 2: The cross-validation results at National Air Pollution Monitoring Network (NABEL) stations.
Fig. 2: The cross-validation results at National Air Pollution Monitoring Network (NABEL) stations.
Full size image

a–e display cross-validation predictions for each of the five different sites, with dashed lines marking the 1:2 (or 2:1) ratio lines. Each plot is labeled in bold with the station omitted during the training process. Although each model was trained by excluding one station, predictions were generated for all five stations during testing. f presents a comparative radar plot showing results with and without the exclusion of specific sites, along with a scenario where all five sites were included in training (indicated by the red dashed line). The site name in brackets identifies the station excluded from training. To ensure comparability, hourly metrics were normalized across test stations, with further adjustments applied so values closer to 1 indicate lower actual values, reflecting improved predictive accuracy and generalization. Notably, f displays Stem-PNC metrics based on hourly normalized results, in contrast to the monthly time scales in (a)–(e). The metrics include Coefficient of Determination (R2), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Square Error (MSE), Explained Variance Score (EVS), and Mean Bias (M-Bias). Additional daily and weekly results are provided in Supplementary Note 4.1.

In Supplementary Note 4, we provide scatter plots of weekly and daily measurements versus model predictions across various stations, allowing for a detailed assessment of the model’s performance at shorter time scales. We observe that discrepancies, including both underestimation and overestimation, are more pronounced at the HAE and RIG stations, with some predictions deviating beyond the 1:2 and 2:1 reference lines. These deviations are notably reduced when data is averaged over longer periods, such as monthly averages. This pattern suggests that spatially divergent locations, which may not be fully represented within the training dataset, could yield larger deviations in short-term predictions. Therefore, we recommend interpreting short-term results with caution in such areas, as longer-term averaged predictions provide more robust estimates. The limitations of spatial generalizability and their implications for divergent locations are further documented and quantified in Supplementary Note 4, where we evaluate prediction accuracy in relation to station-specific characteristics, highlighting how the model’s generalizability may be affected in areas with extreme features.

Further analysis revealed that Stem-PNC underestimated PNCs at BAS (suburban, background), located in a park within a suburban region. Figure 1 illustrates that PNC levels at BAS were low, nearing the lower boundary of the spectrum. Excluding data at BAS caused Stem-PNC to approximate from lower concentration levels, predominantly influenced by the rural station RIG, which led to the underestimation in suburban areas. Therefore, integrating data from BAS can substantially improve the spatial and temporal generalization capabilities of Stem-PNC, particularly for suburban and rural settings.

Overall, the monthly mean predictions generated by the Stem-PNC model consistently fell within a factor of two of the actual measurements, with R2 values ranging from 0.76 to 0.91, indicating a notable degree of accuracy. These statistical metrics underscore the model’s efficacy in capturing the underlying patterns and dynamics governing UFP across diverse geographical locations and temporal periods.

Distribution of particle number concentration in Switzerland

For the implementation of the developed Stem-PNC model in UFP exposure assessment across Switzerland, it is crucial to obtain data on regulated pollutants (NOx, PM10, PM2.5, and Ozone) with high spatial and temporal resolutions. Such comprehensive data are typically produced by sophisticated air quality model systems. We utilized the Copernicus Atmosphere Monitoring Service (CAMS) validated air quality reanalysis dataset46, which provides data at a spatial resolution of 0.1 degrees (approximately 10 km) and a temporal resolution of one hour. This dataset integrates extensive validated observational data from satellites, ground-based, and airborne measurement systems with advanced atmospheric models. The integration of observational data and model results has substantially enhanced the accuracy and reliability of the calculated spatial-temporal distributions of regulated pollutants.

To improve the spatial resolution of CAMS data from 10 km to 1 km, we developed a machine learning-based downscaling technique using Light Gradient Boosting Machine (lightGBM)47,48 and Gradient Boosting Machine49. Details about the downscaling technique can be found in “Methods” section. The Stem-PNC leveraged the downscaling outputs from CAMS as well as the meteorological data from ECMWF Reanalysis v5 (ERA5)50 to estimate particle number concentrations across Switzerland. These estimates were achieved with high temporal (hourly) and spatial (1 km) resolutions.

In Fig. 3, the workflow outlines the four main components of training and spatial extrapolation with Stem-PNC. First, feature selection and target outputs are established using data from multiple sources. Stem-PNC’s input format is structured as Ntime × MFeatures, including meteorological, traffic, and temporal data, with PNC as the output. The model utilizes a two-layer stacking approach with a base model and a meta-model, employing cross-validation to assess generalization. For spatial extrapolation, the trained Stem-PNC model, using pre-trained weights, is combined with the CAMS Downscale Model to incorporate high-resolution CAMS data. By integrating this data with ERA5 meteorological, traffic, and temporal features as 2D inputs, Stem-PNC can estimate 2D PNC distributions across broader regions, capturing localized variations in particle concentrations.

Fig. 3: Workflow for Training and Estimating.
Fig. 3: Workflow for Training and Estimating.
Full size image

a Feature selection and target outputs from National Air Pollution Monitoring Network (NABEL) data, including the training dataset (2016–2019, with 5% allocated for hyperparameter tuning) and the test dataset (2020) on Stacking Technique for Ensemble Modeling of Particle Number Concentration (Stem-PNC). b Stacking model structure with base and meta models, using cross-validation to assess generalization. c Integration of Copernicus Atmosphere Monitoring Service (CAMS) and European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis v5 (ERA5) data along with meteorological, traffic, and temporal features. d Spatial extrapolation to estimate 2D PNC distribution, leveraging high-resolution CAMS data and the trained Stem-PNC. NABEL monitoring stations include Basel-Binningen (BAS), Bern-Bollwerk (BER), Beromünster (BRM), Chaumont (CHA), Davos-Seehornwald (DAV), Dübendorf-Empa (DUE), Härkingen-A1 (HAE), Jungfraujoch (JUN), Lausanne-César-Roux (LAU), Lugano-Università (LUG), Magadino-Cadenazzo (MAG), Payerne (PAY), Rigi-Seebodenalp (RIG), Sion-Aéroport-A9 (SIO), Tänikon (TAE), and Zürich-Kaserne (ZUE).

Figure 4a–d present the seasonal average distributions of particle number concentrations in Switzerland for 2020, estimated by the Stem-PNC model in conjunction with the CAMS and ERA5 dataset. The results generally indicate elevated PNCs in the northern regions and along major roadways, attributable to significant emissions of fine particles from vehicular traffic. Particle number concentrations were higher in January to March (JFB) and October to December (OND) as compared to April to June (AMJ) and July to September (JAS). The observed seasonal fluctuations in PNC align with those of other regulated pollutants, e.g. NOx. These variations are primarily attributed to reduced atmospheric mixing and the low height of the atmospheric boundary layer during colder seasons, which facilitates the accumulation of PNCs. Additionally, summer rainfall might be also a factor that influences the variation of PNC14.

Fig. 4: Estimated seasonal distributions of particle number concentration (PNC) in Switzerland for 2020.
Fig. 4: Estimated seasonal distributions of particle number concentration (PNC) in Switzerland for 2020.
Full size image

The data are segmented into four periods: a JFM (January, February, March), b AMJ (April, May, June), c JAS (July, August, September), and d OND (October, November, December). e The comparison of the average annual data from National Air Pollution Monitoring Network (NABEL) sites, site-specific prediction results, and recalculated grid-average prediction data for PNC. The shadings represent the standard deviation of particle number concentrations. NABEL monitoring stations include Basel-Binningen (BAS), Bern-Bollwerk (BER), Härkingen-A1 (HAE), Lugano-Università (LUG), and Rigi-Seebodenalp (RIG). The boxplot comparisons for the different annual data are provided in Supplementary Note 6.6, Fig. S20.

Figure 4e compares the annual average PNC measurements from the five NABEL stations with the results from the Stem-PNC model. It is important to note that the CAMS and ERA5 datasets, which have relatively low resolutions, were used as input data. Therefore, the results reflect the performance of the entire estimation system. The site-specific and grid-average estimations were respectively derived using sub-grid (at the location of the measurement site) and grid-averaged (1 km by 1 km) traffic information. Both the site-specific and grid-average estimations closely align with the measurements. The largest deviations were observed at the HAE station. The difference between the site-specific estimation and the measurement reached a maximum of 1.7 × 103 particles‧cm−3, representing an approximately 8% deviation from the measurements at the HAE station. The difference between the grid-average estimation and the measurement peaked at 4.4 × 103 particles‧cm−3, corresponding to about a 20% deviation from the measurement. While the site-specific estimations outperformed the grid-average estimations, both of them underestimated the PNC, likely due to the localized high emissions from the adjacent motorways. The grid-average estimations reflected the averaged PNC within the grid, where most of the area was far from the motorways, thus resulting in lower PNCs. Overall, the Stem-PNC model, in conjunction with the CAMS and ERA5 datasets, provides PNC estimations with sufficient accuracy for exposure assessment.

Exposure assessment of ultrafine particles

Figure 5a, b represents the statistics of the gridded (1 km by 1 km) annual mean exposure assessment of UFPs in Switzerland. The population count generally exhibits a downward trend as PNC increases, as shown in Fig. 5a. However, it should be noted that approximately 20% (1.7 million) of the Swiss population is exposed to high UFP levels exceeding an annual mean of 104 particles‧cm−3, which is considered as high exposure (24-h mean) suggested by WHO AQG23. Three distinct peaks are observed below 2 × 104 particles‧cm−3. The first peak, at the lower end of the spectrum close to 3000 particles‧cm−3, is primarily due to contributions from rural areas. The second peak, around 104 particles‧cm−3, is attributed to urban centers and urban clusters. Notably, there is a minor peak in PNC between 1.5 × 104 and 1.7 × 104 particles‧cm−3, corresponding to regions encompassing major transportation hubs and road networks (in Supplementary Note 7.1), indicating that local increases in UFP are driven by these specific conditions.

Fig. 5: Annual mean exposure assessment of ultrafine particles (UFP) in Switzerland.
Fig. 5: Annual mean exposure assessment of ultrafine particles (UFP) in Switzerland.
Full size image

a Distribution of the gridded (1 km by 1 km) UFP exposure for the total population. b Distributions of the gridded UFP exposure across different area types: urban centers, urban clusters, and rural areas. c Population-weighted community-level exposure across 2682 communities in Switzerland. The pie chart illustrates the population fractions across the three annual mean exposure levels, below 104, 104 to 2 × 104, and above 2 × 104 particles‧cm−3, and the height of the bars indicates the community-level exposure of the corresponding communities. d Distributions of the community-level UFP exposure across the three different area types. The shaded regions indicate the probability density of the data. e Comparison of coefficient of variations within communities for particle number concentrations and PM2.5 across the three different area types. The whiskers extend to the smallest and largest values within 1.5 times the interquartile range (IQR = Q3–Q1) from the lower (bottom edge, Q1, 25%) and upper (top edge, Q3, 75%) quartiles, respectively.

Figure 5b summarizes the UFP exposure in urban centers, urban clusters, and rural areas. Approximately 2.12 million residents (24%) live in urban centers, 4.16 million (47%) in urban clusters, and 2.56 million (29%) in rural areas, as indicated by the cumulative population. Among those exposed to UFP levels exceeding 2 × 104 particles‧cm−3, 55.8% are from urban centers, 38.3% from urban clusters, and 5.9% from rural areas. While there is a pronounced focus on UFP exposure risks in urban centers, attention should also be extended to certain high-UFP areas within urban clusters.

Considering that most activities of individuals occur within a specific spatial extent, evaluating community-level exposure may be more reasonable than focusing on a single grid. Community-level exposure was assessed using population-weighted PNCs of the grids (1 km by 1 km) within the corresponding community, as detailed in Supplementary Note 7.3. Grids with higher population densities exert greater influence on the community-level exposure assessment. As shown in Fig. 5c, approximately 4.6% of the total population lives in communities with annual mean UFP exposure levels above 2 × 104 particles‧cm−3, and 31.2% experience community-level UFP exposure between 104 and 2 × 104 particles‧cm−3.

The violin plots in Fig. 5d depict the distributions of community-level annual mean UFP exposure in urban centers, urban clusters, and rural areas, highlighting the range and frequency of UFP exposure within each area type. There are 2686 communities: 93 in urban centers (3.5%), 803 in urban clusters (29.9%), and 1790 in rural areas (66.6%). The community-level UFP exposure in urban centers exhibits the widest range, with values from 6.1 × 103 to 2.7 × 104 particles‧cm−3, and a median of 1.2 × 104 particles‧cm−3. Urban clusters display a mid-range distribution, with values from 1.3 × 103 to 2.3 × 104 particles‧cm−3, and a median of 7.8 × 103 particles‧cm−3. Conversely, rural areas show a more compact distribution, with values from 1.5 × 103 to 1.7 × 104 particles‧cm−3, and a median of 4.4 × 103 particles‧cm−3. Overall, community-level exposure ranges from (5.5 ± 2.3) × 103 particles‧cm−3 in rural areas, (8.5 ± 3.1) × 103 particles‧cm−3 in urban clusters, to (1.4 ± 0.5) × 104 particles‧cm−3 in urban centers, with a national average of (9.3 ± 4.7) × 103 particles‧cm−3.

We adopted the Coefficient of Variation (CV) to investigate the spatial heterogeneity of UFP and PM2.5 within communities across different area types, as shown in Fig. 5e. The median CVs for UFP and PM2.5 at the community level in Switzerland are 0.38 and 0.032, respectively, indicating that the spatial distribution of UFP is significantly more heterogeneous than that of PM2.5. The median CVs for UFP are relatively consistent across different area types, ranging from 0.38 in urban centers and rural areas to 0.41 in urban clusters. In contrast, the heterogeneity of PM2.5 varies more widely across area types, with median CVs of 0.026 in rural areas, 0.087 in urban clusters, and 0.12 in urban centers. This indicates that the spatial distribution of PM2.5 is much more homogeneous in rural areas compared to urban clusters and urban centers. The average ratio between the CVs of UFP and PM2.5 is 10.1 ± 12.2 among all the communities, ranging from 4.7 ± 4.2 in urban centers to 10.0 ± 11.5 in urban clusters, and 13.8 ± 15.1 in rural areas. These substantial differences in spatial heterogeneity suggest that monitoring strategies for UFP may need to differ from those currently used for mass-based measurements.

Evaluation of the high UFP exposure levels suggested by WHO

Community-level exposure to UFP was evaluated based on the 24-h mean and 1-h reference levels for high exposure as suggested by the recent WHO AQG23, where high UFP exposure is considered to be more than 104 particles‧cm−3 for a 24-h mean or 2 × 104 particles‧cm−3 by hourly PNC.

We calculated the annual cumulative hours and days for each community that exceeded the WHO-recommended 1-h and 24-h mean reference levels for high PNC, as shown in Fig. 6a, b. Most communities fall below the 1-h high exposure level, with those exceeding it primarily located in urban centers, as depicted in Fig. 6a. The average duration exceeding the 1-h high exposure level is about 597 ± 1050 h per year. The results indicate that the temporal duration and spatial extent of communities exceeding the 24-h mean high exposure level (Fig. 6b) are significantly larger than those exceeding the hourly high exposure level (Fig. 6a), suggesting that it is more common to surpass the 24-h mean high exposure level. The average duration exceeding the 24-h mean high exposure level is 122 ± 118 days per year.

Fig. 6: Distributions of the annual duration of communities exceeding high ultrafine particle (UFP) exposure levels recommended by the World Health Organization Air Quality Guidelines (WHO AQG).
Fig. 6: Distributions of the annual duration of communities exceeding high ultrafine particle (UFP) exposure levels recommended by the World Health Organization Air Quality Guidelines (WHO AQG).
Full size image

a Distribution of the annual cumulative hours at the community level exceeding the hourly high exposure level (2 × 104 particles‧cm−3). b Distribution of the annual cumulative days at the community level exceeding the 24-h mean high exposure level (104 particles‧cm−3). c Correlation between the annual durations of high UFP exposure respectively identified based on the 24-h mean and 1-h levels suggested by the WHO AQG. The size and color of the dots represent the population of each community, while the bars indicate the population experiencing various annual durations of high exposure. The populations without any high UFP exposure are not shown in the figure; these are 1.20 million based on the 24-h mean level and 2.87 million based on the hourly level.

Figure 6c clearly illustrates the significant nonlinear relationship between the number of days when the PNC exceeds the WHO-recommended daily average and the number of hours when the PNC exceeds the hourly average. This indicates that the two WHO-recommended reference levels might not consistently reflect the same high UFP exposure patterns. The population with various durations of high exposure remains relatively stable based on the hourly level, as shown in the bottom panel of Fig. 6c. However, the durations display an increasing trend with longer durations based on the 24-h mean level, as indicated by the left panel of Fig. 6c. The nonlinear relationship between the two reference levels should be considered in future recommendations for high UFP exposure.

Figure 7a presents the distributions of the annual cumulative hours exceeding the hourly high exposure level of UFPs in the communities within urban centers, urban clusters, and rural areas. Regarding the lower end of the distribution, approximately 47.9% of the total population experiences hourly high UFP exposure for less than 24 h per year, similar to the fraction in urban clusters (49%). This fraction is significantly higher in rural areas, reaching 86.3% of the residents, due to the lack of emission sources such as road traffic. In contrast, only 13.4% of the residents in urban centers experience less than 24 h of high exposure annually. The average annual durations exceeding the 1-h high exposure level are about (1.59 ± 1.45) × 103 h in urban centers, 318 ± 577 h in urban clusters, and only 59 ± 226 h in rural areas.

Fig. 7: Distributions of the exposed population fractions categorized by annual cumulative durations exceeding the 1-h and 24-h mean high ultrafine particle (UFP) exposure levels recommended by World Health Organization Air Quality Guidelines (WHO AQG).
Fig. 7: Distributions of the exposed population fractions categorized by annual cumulative durations exceeding the 1-h and 24-h mean high ultrafine particle (UFP) exposure levels recommended by World Health Organization Air Quality Guidelines (WHO AQG).
Full size image

a Exposed population fraction based on the 1-h high UFP level (2 × 104 particles‧cm−3), including four categories of total population, the population in urban centers, urban clusters, and rural areas. b Correlation between the fractions of the population exposed to the 1-h high UFP level in urban centers and urban clusters. c Exposed population fraction based on the 24-h mean high UFP level (1 × 104 particles‧cm−3). d Correlation between the fractions of the population exposed to the 24-h mean high UFP level in urban centers and urban clusters.

Approximately 25.5% of the total population exceeds the hourly high exposure level for more than 720 h (1 month) annually, with this fraction rising to 62.4% in urban centers. The exposed total population, as well as those in urban clusters and rural areas, generally decreases as the cumulative duration of high exposure increases. Conversely, in urban centers, the exposed population displays an upward trend, reaching 21.9% for cumulative durations between 2880 and 5760 h (4–8 months) per year. No population experiences the hourly high exposure level for more than 5760 h annually. Figure 7b reveals a negative correlation between the exposure populations in urban centers and urban clusters, indicating that the hourly high exposure periods exceeding 2 × 104 particles‧cm−3 are common in urban centers but rare in urban clusters.

As shown in Fig. 7c, when the 24-h mean high UFP exposure reference level is applied, the exposed total population rises as the cumulative duration of high exposure increases, peaking at 21.1% of the total population with more than 240 days of high exposure, which is opposite to the downward trend observed based on the hourly reference level (Fig. 7a). About 56% of the total population experiences high UFP exposure for over 60 days based on the 24-h mean reference level. Particularly in urban centers, the majority (81.7%) of the population experiences high exposure for over 120 days intervals, and the fraction in urban clusters is 36.9%. It indicates that about 80% of the residents in rural areas exceed the 24-h mean reference level for less than 30 days. The average annual durations exceeding the 24-h mean high exposure level are about 239 ± 104 days in urban centers, 102 ± 94 days in urban clusters, and only 27 ± 52 days in rural areas. A positive correlation between the exposure populations in urban centers and urban clusters has been observed as shown in Fig. 7d.

Discussion

The innovative stacking-based ensemble machine learning model, Stem-PNC, integrates data-driven and physical-chemical models to achieve comprehensive UFP exposure assessments with high accuracy and resolution. The Stem-PNC model, utilizing the stacking method, demonstrates significant advantages in air quality modeling due to its strong generalization capabilities and flexibility in model selection. Leveraging the unparalleled long-term standardized PNC measurements in Switzerland, along with the extensive dataset of the regulated pollutants and model results from CAMS, the generalization abilities of the model are significantly enhanced. The Stem-PNC model has achieved a high R2 of 0.85 and a low mean bias of 124 when compared to measurements. The largest annual mean deviations were at the HAE station near motorways, with a maximum site-specific estimation difference of 1.7 × 103 particles‧cm−3 (approximately 8% of the mean PNC). Notably, Stem-PNC demands only 4% of the computational time and energy consumption needed by deep learning models. These highlight its cost-efficiency and practical utility, underscoring its superior predictive and generalization abilities compared to other models.

By integrating machine learning downscaling techniques with CAMS grid data and ERA5 meteorological data, we achieved high-quality UFP estimations at a 1 km spatial resolution and hourly temporal resolution. The national-scale UFP exposure assessment has provided the first comprehensive picture of the exposure across various area types, i.e. urban centers, urban clusters, and rural areas. The analysis has revealed that approximately 20% (1.7 million) of the Swiss population is exposed to high UFP levels exceeding an annual mean of 104 particles‧cm−3. The annual mean community-level exposure is as high as (1.4 ± 0.5) × 104 particles‧cm−3 in urban centers, (8.5 ± 3.1) × 103 particles‧cm−3 in urban clusters, and decreases to (5.5 ± 2.3) × 103 particles‧cm−3 in rural areas. Notably, about 10% and 5.6% of the population in urban centers and urban clusters, respectively, about half a million residents in total, face an annual mean UFP exposure above 2 × 104 particles‧cm−3. Therefore, UFP exposure risk should be considered for both urban centers and clusters.

A significant nonlinear relationship was identified between the number of days when the PNC exceeds the WHO-recommended daily average and the number of hours when the PNC exceeds the hourly average, following an exponential relation with an exponent of 0.35. The average durations exceeding the 1-h and 24-h mean high exposure levels are about 597 ± 1050 h and 122 ± 118 days per year. Additionally, disparities in the trends for high exposure durations among different types of areas were observed. There is a negative correlation between the exposure populations in urban centers and urban clusters based on the 1-h reference level (2 × 104 particles‧cm−3), indicating that the hourly high exposure periods are common in urban centers but rare in urban clusters. Conversely, a positive correlation is observed between the exposure populations in urban centers and urban clusters when the 24-h mean reference level is applied, suggesting that the two reference levels might not consistently reflect the same high UFP exposure patterns.

This nonlinear relationship between the two WHO-recommended reference levels should be considered in future recommendations for high UFP exposure. The current WHO AQG suggested that the two reference levels for high PNC are interchangeable. However, the results here indicate that they are not equivalent, and both are necessary to manage UFP exposure effectively. One level aims to control daily exposure to prevent it from being excessively high, while the other focuses on controlling hourly exposure to prevent extraordinarily high peaks. The spatial heterogeneity of UFP is about 10 times larger than that of PM2.5 on average, ranging from 4.7 ± 4.2 in urban centers to 10.0 ± 11.5 in urban clusters, and 13.8 ± 15.1 in rural areas, indicating that monitoring strategies for UFP should differ from those currently used for mass-based measurements.

The Stem-PNC model, trained using the long-term standardized PNC measurements in Switzerland, could be fine-tuned with a small dataset of UFP measurements from another region, and utilized for large-scale UFP exposure assessment in other countries, where similar emission standards are applied. The combination of temporal and spatial generalization, along with computational efficiency, positions Stem-PNC as a valuable tool in UFP modeling and epidemiological studies. Furthermore, the exposure conditions in Switzerland investigated in this study could serve as a reference, providing crucial insights for the development of future UFP standards.

Methods

Stacking technique for ensemble modeling of particle number concentration

The Stacking Technique51,52 for Ensemble Modeling of Particle Number Concentration (Stem-PNC) employs the stacking technique, also known as stacked generalization, an ensemble learning approach that combines multiple prediction models to enhance the overall predictive performance. Stem-PNC consists of two model levels. The first level comprises four base learners: K-Nearest Neighbors (KNN), Decision Tree (Tree), Random Forest (RF), and Light Gradient Boosting Machine (LGB). These models are chosen for their ability to capture the nonlinear relationships between PNC, regulated pollutants, meteorological data, and traffic information. The second layer, known as the meta-learner, aggregates the predictions of the base learners by training a sub-learner. Stem-PNC utilizes a multilayer perceptron (MLP) neural network as the meta-learner.

$${{{\rm{P}}}}{{{{\rm{NC}}}}}_{{pred}}=f({{{\rm{NABEL}}}},{{{\rm{Meteo}}}},{{{\rm{Traffic}}}},{{{\rm{Time}}}};\Theta ),$$
(1)

where PNCpred is the predicted PNC, derived from a machine learning model f such as Random Forest or LightGBM, where Θ represents the parameters of the model. The data used as features in this model include NABEL (NABEL data), Meteo (Meteorological Data), Traffic (Traffic Data), and Time (Time Data).

The stacked model offers enhanced generalization capabilities, enabling broad scenario application and the integration of diverse predictor types, thereby increasing the model’s robustness and stability. The stacking model used in this study is as follows:

$$\hat{y}={f}_{{MLP}}{\left\{{\left[{f}_{{KNN}}\left({{{\bf{X}}}}\right){f}_{{RF}}\left({{{\bf{X}}}}\right){f}_{{LGB}}\left({{{\bf{X}}}}\right){f}_{{Tree}}\left({{{\bf{X}}}}\right)\right]}_{{{{\rm{First\; Layer}}}}}\right\}}_{{{{\rm{Second\; Layer}}}}},$$
(2)

where X represents the input feature set and \(\hat{y}\) denotes the target variable.

The PNC measurements are taken continuously with a high temporal resolution of 1 h at five different stations across Switzerland, Basel-Biningen (BAS, suburban, background), Bern-Bollwerk (BER, urban, roadside), Harkingen-A1 (HAE, rural, next to motorway), Lugano-University (LUG, urban, background), Rigi-Seebodenalp (RIG, rural, altitude > 1000 m). In addition to PNC, regulated pollutants (NOx, PM10, PM2.5, and Ozone) are also monitored at these NABEL sites. Stem-PNC was trained using hourly PNC measurements collected over a four-year period (2016–2019) at these five stations. Metrological data were obtained from MeteoSwiss. Details about the study area and the dataset are provided in Supplementary Note 1.

Stem-PNC demonstrates superior predictive and generalized performance compared to machine learning models and greater time efficiency than deep learning methods. By combining multiple base models into a meta-model, Stem-PNC enhances prediction accuracy and reduces overfitting. This ensemble approach integrates various strengths of different models, leading to more robust and reliable predictions. Additionally, the simpler, faster-to-train base models in Stem-PNC require less hyperparameter tuning and computational power than deep learning models, resulting in quicker predictions and efficient performance.

The performance of Stem-PNC was evaluated against five different types of data-driven machine learning methods: Linear Models, Support Vector Machines, Neighbor-based Models, Tree-based Models, and Deep Learning (DL) Models. Statistical metrics such as the R2, M-Bias, RMSE, and Mean Absolute Error (MAE) were used for these comparisons. A smaller absolute value of M-Bias indicates better generalization capability compared to other models. Mean prediction metrics are calculated within each category, with standard deviation represented as error bars in Fig. 1f. Among the categories, Tree-based Models exhibit the highest standard deviation. To enhance clarity, some metrics with similar evaluation functions were omitted. Additional details and definitions of these statistical metrics, the structure of stacking, the criteria for categorizing models, the rationale for their division, hyper-parameterization, and more prediction results in each category are available in Supplementary Note 2. To further analyze the predictive model’s results, model interpretability was visualized using SHapley Additive exPlanations (SHAP53,54) for Random Forests, as detailed in Supplementary Note 2.8.

We conducted additional experiments focusing on the most correlated features: NOX, PM10, and traffic. To evaluate their individual and combined predictive power, we trained models using each feature separately, in pairs, and with all three features together, comparing their performance to that of the comprehensive, full-feature Stem-PNC model. Our results indicate that, while models using only the most correlated features (NOX, PM10, and traffic) achieved reasonable accuracy, their predictive performance was notably lower than that of the full-feature Stem-PNC model.

Specifically, R2 values ranged from 0.679 with NOx alone to 0.794 when combining NOx, PM10, and traffic. Correspondingly, root mean square error (RMSE) increased considerably across the reduced-feature models, with RMSE values ranging from 5308 to 6623. Incorporating additional features (e.g. ozone) further improved model performance, increasing R2 to 0.813 and reducing RMSE errors (RMSE = 5053). The full-feature Stem-PNC model achieved the highest accuracy, with R2 reaching 0.845 and significantly lower RMSE (RMSE = 4594). This variation is primarily due to the unique characteristics of the RIG station, where the rural setting and high altitude (>1000 m) introduce additional complexities for models with fewer features. A detailed analysis is provided in Supplementary Note 2.9.

These findings illustrate the limitations of models trained solely on highly correlated features and underscore the advantage of incorporating a broader range of input variables. Each additional feature contributes incrementally to the model’s predictive accuracy.

Score-metrics to generalization ability of Stem-PNC

For evaluating prediction results, certain metrics such as the Coefficient of Determination (R2) and Explained Variance Score (EVS) naturally range from 0 to 1. However, metrics like Mean Bias (M-Bias), Root Mean Square Error (RMSE), Mean Square Error (MSE), and Mean Absolute Error (MAE) vary over a broader range and are more challenging to compare directly due to their disparate magnitudes. To facilitate comparisons across these metrics, the Minmax scalar is employed for normalization. For each metric x (such as MAE, MSE, RMSE, M-Bias), normalization is calculated using the formula:

$${{{\rm{Score}}}}{-}{{{\rm{Metrics}}}}=1-\frac{x-{x}_{\min }}{{x}_{\max }-{x}_{\min }},$$
(3)

Here, xmin and xmax are the minimum and maximum values of the metric in the dataset, respectively. This formula ensures that a lower original metric value translates into a higher normalized score, reflecting better performance.

Downscaling CAMS data: methodology and formulas

Although the spatial resolution of CAMS data is adequate for broad-scale assessments, it necessitates refinement to fulfill the requirements of detailed PNC exposure analysis in Switzerland, particularly given the significant spatial variability of PNC. The formulas for calculating prediction features and the rationale for selecting models for high-resolution CAMS predictions are detailed in Supplementary Notes 6.1 and 6.2, respectively.

This refinement was achieved by utilizing long-term observational data of regulated pollutants collected from 16 NABEL stations across Switzerland. These stations include Basel-Binningen (BAS [Suburban]), Bern-Bollwerk (BER [Urban, Traffic]), Beromünster (BRM [Rural, Altitude < 1000 m]), Chaumont (CHA [Rural, Altitude > 1000 m]), Davos-Seehornwald (DAV [Rural, Altitude > 1000 m]), Dübendorf-Empa (DUE [Suburban]), Härkingen-A1 (HAE [Rural, Motorway]), Jungfraujoch (JUN [High Alpine]), Lausanne-César-Roux (LAU [Urban, Traffic]), Lugano-Università (LUG [Urban]), Magadino-Cadenazzo (MAG [Rural, Altitude < 1000 m]), Payerne (PAY [Rural, Altitude < 1000 m]), Rigi-Seebodenalp (RIG [Rural, Altitude > 1000 m]), Sion-Aéroport-A9 (SIO [Rural, Motorway]), Tänikon (TAE [Rural, Altitude < 1000 m]), and Zürich-Kaserne (ZUE [Urban]). Similar to the approach in Stem-PNC, meteorological data, date, and traffic information were also integrated into the downscaling method. The high-resolution CAMS prediction model was trained using data from 2016 to 2019 and used to predict the data for 2020.

By downscaling CAMS data, higher-resolution pollutant distributions can be obtained. These distributions are influenced by various factors including meteorological conditions, human activities, and topography55, all of which are crucial for improving the reliability and accuracy of model predictions. To enhance data richness and correct discrepancies, the ratio of pollutant data from NABEL to CAMS, referred to as Measurements in Eq. (4), is utilized initially for data correction.

To further enrich the feature set, meteorological, traffic, and time-related data are incorporated. Fusing multiple datasets mitigates the limitations inherent in relying on a single data source, thereby ensuring more continuous and comprehensive coverage. Machine learning methods are employed to discern complex patterns and relationships among these factors, enhancing our understanding of how they collectively influence pollutant distribution.

Moreover, machine learning enhances the flexibility and adaptability of model predictions, particularly when applied to different datasets. This adaptability ensures that once a model is trained and validated, it can effectively predict air quality in different regions, even those without extensive monitoring networks. This capability is vital for maintaining good generalization performance across varied geographical locations.

$${{{\rm{Measurements}}}}=\frac{{{{\rm{pollutio}}}}{{{{\rm{n}}}}}_{{NABEL}}}{{{{\rm{pollutio}}}}{{{{\rm{n}}}}}_{{CAMS}}}\propto f({{{\rm{pollutio}}}}{{{{\rm{n}}}}}_{{CAMS}},{{{\rm{Meteo}}}},{{{\rm{Traffic}}}},{{{\rm{Time}}}}),$$
(4)

The predicted Measurements were then multiplied by the linearly interpolated pollution data to obtain high-resolution pollutant distributions in Eq. (5). This process enhances the accuracy of the scaled-up data, ensuring that the finer details of pollution patterns are captured effectively.

$${{{\rm{pollutio}}}}{{{{\rm{n}}}}}_{{High}{-}{res\; pollution}}={{{\rm{pollutio}}}}{{{{\rm{n}}}}}_{{CAMS}}\times {{{{\rm{Measurements}}}}}_{{pred}}$$
(5)

PNC estimation: methodology and formulas

The Stem-PNC model, integrated with machine learning-based downscaling of CAMS data, is utilized to estimate PNC. Initially, the Stem-PNC model is trained specifically for time-series predictions. Following this, the CAMS downscaling process produces distributions of regulated pollutants at a high spatial resolution of 1 km\(\times\)1 km. These high-resolution outputs are then merged with meteorological, traffic, and time data to derive a comprehensive high-resolution PNC using Eq. (6).

This multi-source integration significantly improves the model’s accuracy in estimating spatial and temporal variations in PNC. Additionally, this approach allows for the application of the trained Stem-PNC model to accurately estimate the spatial and temporal distributions of PNC. As a result, it enables precise predictions of PNC across various regions and times, thereby enhancing the model’s utility and effectiveness in real-world scenarios.

$${{{\rm{P}}}}{{{{\rm{NC}}}}}_{{High}-{res}}\propto f({{{\rm{pollutio}}}}{{{{\rm{n}}}}}_{{High}-{res}},{{{\rm{Meteo}}}},{{{\rm{Traffic}}}},{{{\rm{Time}}}})$$
(6)

In this study, we focus on exposure assessment, selecting a spatial resolution of 1 km to align with the available population distribution data. The estimated PNC values represent only the average PNC within each grid cell. As illustrated by the sub-grid and grid-average results in Fig. 4e, the 1 km resolution may not capture fine-scale variations near high-emission sources, e.g. the motorways. Supplementary Fig. S20 in Supplementary Note 6.6 indicates that while grid averages effectively represent the median values of PNC distributions, closely matching sub-grid results and measurements, they may miss high hourly PNC values. Future studies would benefit from improved spatial resolution of PNC estimates. However, finer spatial scales can increase measurement errors and noise56,57, necessitating denser monitoring networks and larger PNC observation datasets for model training, which are currently unavailable. Additionally, more detailed population distribution data, or movement data, would be required for accurate exposure assessment at finer spatial estimations.

Population data and area types

The population data utilized in this analysis is sourced from the Gridded Population of the World (GPW), v458, specifically from the year 2020 and at a resolution of 2.5 arc-minutes (~1 km at the equator). The categorization of area types is based on the Eurostat methodology—Applying the Degree of Urbanisation59. According to this classification, areas are divided into three types: Urban Centers, characterized by a population density of at least 1500 inhabitants per km2; Urban Clusters, with a population density of at least 300 inhabitants per km2; and Rural Areas, where the population density falls below 300 inhabitants per km2. Further details regarding the estimated PNC and population can be found in Supplementary Notes 7 and 8.

In this study, we mainly focused on long-term exposure, and we used annual population data due to minimal changes in Switzerland60,61. In contrast, developing countries experience more dynamic shifts, including immigration. Regarding short-term exposure, future work could consider combining dynamic population mapping62,63,64,65 with hourly PNC estimations to improve exposure assessments at finer temporal resolutions. Although this study uses static population density for time-averaged exposure, the model’s temporal accuracy suggests that integrating dynamic data could better capture exposure during peak hours, allowing for more precise evaluations in regions with significant population changes. This integration would require additional data sources, such as cell phone-based position data, which may become available in the future. For more detailed explanations, see Supplementary Note 7.5.