Abstract

1. Introduction

The World Health Organization estimates in seven million global deaths associated with air pollution (both outdoor and household) every year, emphasising that exposure to particulate matter (PM) is among the greatest causes of concern [1]. Fine particles with an aerodynamic diameter smaller than 2.5 μm (PM_2.5) can penetrate the human circulatory system through the lungs and provoke multiple adverse health outcomes, including mortality [2], hospital admissions [3], lung dysfunction [4], cardiovascular diseases [5], and allergic reactions [6]. Usually, epidemiological studies collect air quality (AQ) data from ground monitors to quantify both short-term and long-term PM_2.5 exposure associated with acute and chronic health effects, respectively. The limitation in this health assessment approach is the lack of continuous temporal records of PM_2.5 and the limited spatial distribution of the monitors. Great Britain is an example of countries with very limited spatio-temporal coverage of PM_2.5, whereby the monitoring network is densely located only in major cities and widespread measurements of PM_2.5 only started from 2010.

Remote sensing observations of aerosol optical depth (AOD) obtained from satellites, which measures how much direct sunlight has been scattered and absorbed by aerosol particles suspended in the atmosphere, has recently been proposed as an alternative to measuring PM variability for epidemiological purposes [7]. However, while offering the advantage of global coverage and relatively high spatio-temporal resolution, the use of AOD for PM_2.5 exposure assessments presents limitations, for instance the fact that it represents the total atmospheric column concentration of the aerosol rather than surface values [8]. Unsurprisingly, early studies based only on satellite-AOD achieved very low performances in predicting PM_2.5 [9]. Recent studies have proposed more sophisticated approaches, combining AOD measures with information from other satellite products, reanalysis data, chemical transport models, and geospatial features to improve the prediction of PM_2.5. Such studies used various analytical methods, including multiple linear regression [10], land-use regression [11,12], and mixed effect models [13–17]. The last development in this research area is represented by the application of machine learning (ML) algorithms, including various architectures such as random forests [18–22], neural networks [19,21], and gradient boosting [19,23,24]. These have demonstrated higher performances, linked with an ability to model any kind of predictor(s)-response association and to deal better with the potentially complex relationship between PM_2.5, spatial, and spatio-temporal predictors [19,25].

The aim of this study is to develop and apply a multi-stage satellite-based ML model to estimate daily concentrations of PM_2.5 over a 1 km grid across Great Britain in the period 2008–2018. The analysis is based on a dataset with synchronised information from various data sources, such as several remote sensing satellite products, multiple climate and atmospheric reanalysis databases, chemical transport models, and spatial and spatio-temporal variables. The model is assessed through measures of predictive performance, error, and bias, obtained through cross-validation.

2. Materials and Methods

2.2. PM_2.5 and PM₁₀ Observed Data

Daily PM₁₀ and PM_2.5 (μg/m³) measurements in the study period were obtained from five monitoring network sources through the R package openair [29]: Automatic Urban and Rural Network (AURN), Air Quality England (AQE), King College London (KCL), Scotland Air Quality Network (SAQN), and Wales Air Quality Network (WAQN). When monitors from different sources showed the same temporal distributions (i.e., a correlation equal to 1) and were located at approximately the same coordinates, only the AURN monitor was kept. Monitors with less than 18 hours of PM_2.5 records per day as well as less than 30 days by year were removed. The final set includes 581 and 183 monitors measuring PM₁₀ and PM_2.5 along the study period, respectively. Figure 1 shows the locations of these monitors across Great Britain, illustrating how the network coverage is densely located in major cities, leaving several small cities and rural areas with only a few or no AQ records. Each monitor was indexed using the cell-ID of the 1 km grid cell that contained it.

Open in a separate window

Figure 1

Spatial distribution of 581 PM₁₀ (black star) and 183 PM_2.5 (turquoise dots) monitors across Great Britain during the study period.

3. Results

3.2. Stage-2 Results

The Stage-2 procedure is illustrated in Figure 2, showing missing satellite-AOD, imputed through a combination of multiple modelled-AOD wavelengths and sub-day times. Table 2 shows the performance of Stage-2 models validated using the OOB CV method. The results presented consistently high R² (ranging from 0.963 to 0.988), low RMSE (varying between 0.007 to 0.010), and almost no bias (intercept zero and slope close to 1 in almost all years and wavelengths).

Open in a separate window

Figure 2

Satellite-AOD 0.47 μm values are represented in Figure 2a (mean of all Terra- and Aqua-MODIS passing times). The missing values in grey were imputed through a combination of multiple modelled-AOD wavelengths and sub-day times, represented in Figure 2b by modelled-AOD 0.47 μm at 12:00. The Stage-2 output is shown in Figure 2c, illustrating the full coverage of AOD 0.47 μm based on the combination of measurements and estimations across Great Britain. The maps correspond to values measured or reconstructed on 6 July 2018.

Table 2

Predicted-AOD 0.47 μm and 0.55 .55 obtained from Stage-2 RF models were regressed against measured Satellite-AOD 0.47 μm and 0.55 μm in a linear regression model. The performance was evaluated using OOB CV together with RMSE (μg /m³), intercept (μg/m³), and slope (μg /m³).

Stage-2 OOB CV
	Predicted-AOD 0.47 μm				Predicted-AOD 0.55 μm
	R2	RMSE	Inter.	Slope	R2	RMSE	Inter.	Slope
2008	0.977	0.010	–0.001	1.009	0.977	0.007	–0.001	1.009
2009	0.976	0.010	–0.001	1.010	0.976	0.007	–0.001	1.010
2010	0.968	0.009	–0.001	1.013	0.968	0.007	–0.001	1.013
2011	0.988	0.010	–0.001	1.005	0.988	0.007	0.000	1.005
2012	0.980	0.010	–0.001	1.008	0.981	0.007	–0.001	1.008
2013	0.984	0.010	–0.001	1.007	0.984	0.007	–0.001	1 006
2014	0.970	0.009	–0.001	1.012	0.970	0.007	–0.001	1.012
2015	0.972	0.009	–0.001	1.011	0.973	0.007	–0.001	1.011
2016	0.975	0.009	–0.001	1.010	0.975	0.007	–0.001	1.010
2017	0.963	0.009	–0.001	1.015	0.963	0.007	–0.001	1.014
2018	0.969	0.010	–0.001	1.01113	0.969	0.007	–0.001	1.013
Mean	0.978	0.010	–0.001	1.009	0.9778	0.007	–0.001	1.009

Open in a separate window

3.4. Stage-4 Results

Stage-4 provides the prediction of daily PM_2.5 concentrations for each of the 234.429 1 km grids covering Great Britain. Results indicate an annual PM_2.5 average of 9.41 μg/m³ for 2008, 10.17 μg/m³ for 2013, 8.05 μg/m³ for 2018, and 8.84 μg/m³ for 2008–2018 but with a strong spatial and temporal variation. The spatial distribution of annual average PM_2.5 concentrations for 2008, 2013, and 2018 are shown in Figure 3, revealing a decrease of pollution levels in recent years across the whole territory, although slightly stronger in England. Table S5 in Supplementary Materials provides the same figures for all the years, confirming the decreasing trend. The spatial comparison suggests that PM_2.5 concentrations are lower in Scotland and Wales compared to the more populated southern regions of England, with hotspots located in urban areas such as Liverpool, Manchester, Birmingham, and Greater London. At the bottom of Figure 3, the maps display the corresponding annual average of PM_2.5 levels in London, demonstrating the precision of the multi-stage ML model in reconstructing PM_2.5 concentrations in 1 km grid cells within urban areas. The maps show local hotspots of high pollution, with a spatial distribution that, however, changes along the study period.

Open in a separate window

Figure 3

Stage-4 predicted PM_2.5 concentrations across Great Britain (Top) and London (Bottom) for 2008, 2013, and 2018 aggregated by annual means. All plots were built under the same colour scale.

The greatest contribution of this study is in the ability of the satellite-based ML models to reconstruct daily levels of PM_2.5 over a 1 km grid across a wide geographical domain. Figure 4 displays the time series plots of observed and predicted PM_2.5 concentrations at three monitoring sites and their locations within the geographical domain of Great Britain. While not necessarily representative, the results show that the error varies depending on location, type of monitor, and period. However, generally the plots indicate a very good performance of the ML algorithms in recovering the observed temporal variation in PM_2.5 levels, capturing peaks and periods of stable low concentrations.

Open in a separate window

Figure 4

Stage-4 daily series of observed and predicted PM_2.5 values across Great Britain for 2008, 2013, and 2018: (1) Traffic monitor—located in Dundee (Scotland); (2) Urban monitor—located in London (England); and (3) Rural monitor—located in Sutton Bridge (England).

Figure 5 complements the analysis of daily variations by displaying the spatial distribution of PM_2.5 estimations across Great Britain (Top) and London (Bottom) for specific days within the study period. It is interesting to note the wide variation in PM_2.5 concentrations between days in the same area and between areas on different days. For instance, the maps in the left panels represent a day with almost no variation and generally low concentrations; the maps in the mid panels display a strong north–south split likely linked to weather conditions; the right panels show a more complex pattern with a wider range of PM_2.5 values, a large area with very high pollution concentrations located in east London, and hotspots in highly populated urban areas across England.

Open in a separate window

Figure 5

Stage-4 day-specific PM_2.5 estimations across Great Britain (Top) and London (Bottom).

4. Discussion

This study presents the first application of satellite-based spatiotemporal ML methods to reconstruct levels of pollution across Great Britain, providing estimates of daily PM_2.5 concentrations over a 1 km grid during 2008–2018. The multi-stage ML framework provided significant advantages, allowing the combination of information from multiple data sources, such as air quality monitoring networks, remote sensing satellite products, chemical dispersion models, reanalysis databases, and administrative census data, among others.

The beginning of the study period (2008 and 2009) had a lower quantity of monitors measuring PM_2.5 across Great Britain, but this number increased considerably from 2010 after a new measuring network had been established in 2009 [46]. Therefore, Stage-1 was an extremely important step to extend the number of PM_2.5 measurements. The implementation of Stage-2 was also relevant to fill the gaps in MAIAC AOD retrievals from satellites, thus maximizing the available information. Stage-3 produced an ML prediction model based on a long list of informative predictors, accounting for potentially complex inter-relationships and functional forms. The Stage-3 ML algorithms offered excellent performance, showing an average cross-validated R² of 0.767 across the period, with an increased predictive ability in the last years. Stage-4 provided a single PM_2.5 estimation for each of the 234.429 1 km grid cells in each of the 4018 days, totalling around 950 million data points. The methodology provides complete spatial coverage, high resolution, and a relative small error of the predictions, and the ability to capture variations in PM_2.5 concentrations across both spatial and temporal domains. The model offers a prediction accuracy that makes the output suitable for application in epidemiological studies on the short- and long-term health effects of air pollution. As demonstrated in Figure 4, the methodological approach presented in this study was able to capture the daily PM_2.5 variability across different years, locations, and monitor types.

The output from the empirical ML model developed in this study complements existing databases of modelled PM_2.5 in the United Kingdom, with some advantages for applications in epidemiological studies. Country-wide maps generated by emission-dispersion models are usually available at a coarser spatial or temporal resolution [32,47,48], and they generally show lower small-scale accuracy when tested against observed monitoring data [49,50]. The spatiotemporal ML models presented here demonstrated comparable predictive performance to similar methods applied in other countries, based either on single-learner ML models [13,20], ensemble ML models [18,19,21], or generalised additive models (GAM) [15]. Ref [21] developed an ensemble ML model (composed by RF, Deep Neural Networks, GAM, Gradient Boosting, K-nearest Neighbour) to estimate PM_2.5 for Greater London, reaching a mean 2005–2013 CV spatial-R2 of 0.396. Using the 2008–2013 period, this study reached a CV spatial-R2 of 0.637 for the whole of Great Britain. Modelling a larger area might have provided more information and a higher spatial variability, improving substantially the CV spatial-R2. The‘application of a single learner to model air pollution in both spatial and temporal domains for Great Britain achieved a satisfactory performance. Nonetheless, ensemble model formats can be used in future applications using the same area of study and variables to assess how much performance gain is reached compared to the RF-only learner.

In 2008, a new Air Quality Directive came into force, bringing considerable changes to the following UK annual air quality assessments in 2010, setting an annual mean target of 25 μg/m³ [51]. Several policy controls and emission reductions had been put in place aiming to reduce from 2010 the road traffic PM emission by 83%, off-road mobile machinery by 54%, and energy production by 32% until 2020 [52]. Independently of the temporal aggregation (daily or annual), all maps shown in this study detected this considerable drop in the PM_2.5 concentrations from 2010 across Great Britain.

Some limitations of this study must be acknowledged. First, the multi-stage model relies on the extension of the observed series of PM_2.5 by predicting values based on co-located PM₁₀ in Stage-1. This step was necessary for the application of the method in the early years characterized by sparse PM_2.5 monitoring, which is likely to have contributed to the lower predictive performance in this period. Morover, the cross-validation procedures revealed the presence of some bias in the Stage-1 predictions, particularly in the spatial domain, which was probably linked to limitations in modelling the relative distribution of the two PM components. Second, while the model displayed a good performance throughout the period, the accuracy is worse in the temporal domain in the summer (Table S4), suggesting limitations in capturing the higher temporal variation in this season. Third, the generalization of the prediction model was dependent on the selected locations of the monitors, which may be not representative of the study domain. This can result in an underestimation of the error and potential biases in the predictions in more remote and less represented areas if structural spatial differences are not entirely captured by the model covariates.

The findings of this study also highlight additional limitations not yet discussed in previous studies related to the use of remote sensing satellite products for reconstructing air pollution exposure for epidemiological purposes. The RF importance ranking order showed that Stage-3 was mainly informed by EMEP4UK PM_2.5, regional and local interpolated PM_2.5 estimations, and meteorological variables, while the contribution from Stage-2 outputs (i.e., predicted-AOD 0.47 μm and 0.55 μm) was very limited. Therefore, direct satellite observations (i.e., AOD, NDVI, elevation, land cover, impervious surfaces, and night-time light) offered a relatively minor contribution to represent ground PM_2.5 exposures (Stage-3), while satellite-based products (i.e., indirect satellite observations) such as EMEP4UK and climate reanalysis products played a much more relevant role to derive reliable PM_2.5 estimates across Great Britain.

Future directions in remote sensing are pointing to new satellite instruments developed for air pollution monitoring which are likely to provide better resolution and reliability, thereby improving the predictive performance. As examples, the recently launched Copernicus Sentinel-5 Precursor [53] and the two future earth observation missions from European Space Agency (e.g., Copernicus Sentinel-4 [54] and Sentinel-5 [55]) were developed to provide information on atmospheric variables, such as air quality parameters (nitrogen dioxide, ozone, and aerosols). Finally, future research developments related to advanced statistical methodology, as demonstrated by the ML framework proposed here, include the application of geostatistical techniques, the use of alternative single-learner or ensemble ML algorithms, and statistical downscaling methods to increase further the resolution of the predictions.

5. Conclusions

This study developed and applied a multi-stage ML model, combining data for multiple sources, including remote sensing satellite products, climate and atmospheric reanalysis models, chemical transport models, and geospatial features, to generate a complete map of daily PM_2.5 concentrations in a 1 km grid across Great Britain between 2008 and 2018. The model showed good performance overall and in both spatial and temporal domains, with an accuracy that is compatible with the use of such reconstructed values as a proxy for PM_2.5 exposures in epidemiological studies. In particular, the availability of high-resolution measures that can be linked as such or aggregated at different spatial and/or temporal scales makes the output suitable for investigations on both transient and chronic health risks associated to short and long-term exposures to PM_2.5, respectively.

Supplementary Material

Acknowledgments

Footnotes

References