Exposure to fine particulate matter (PM2.5) poses a significant global health risk, yet extreme concentration patterns remain underexplored. This study estimates daily PM2.5 concentrations from 1980-2023, validated against the WHO ambient air quality database. An ensemble of deep learning models (CNN, LSTM, DNN) incorporating meteorological inputs achieved robust predictive accuracy (RMSE < 17.85 µg/m³, R² > 0.894). Global and regional variations in population-weighted PM2.5 extremes [average annual, annual maximum, 99th percentile, days exceeding the USEPA standard of 35.5 μg/m³ (AQI > 100) weighted by population density] were analysed. Results reveal persistently high PM2.5 extremes in China, India, and Pakistan, contrasted with declining levels in Europe and North America. Significant variability in African nations like Rwanda and Benin was also observed. 79.7% of the global population and 66.3% of land areas exceeded the USEPA annual standards (9 μg/m³). Seasonal disparities underscore region-specific pollution trends. These findings advocate for phased, locally adaptive air quality strategies, especially in low-income and emerging economies.

Exposure to fine particulate matter (PM) is a major environmental and public health concern, linked to increased risks of disease and early mortality. Outdoor air pollution, which includes ambient PM, contributed to approximately 6.67 million (95% UI: 5.90-7.49) premature deaths worldwide in 2019. PM is a composite mixture of various aerosols and chemical compounds with varying toxicity and mass fractions. Consequently, the actual health impact of ambient PM could be greater than what is estimated by looking at the total PM mass alone. Exposure to ambient PM, whether for a short duration or over an extended period, poses significant health risks, even at low concentrations. Similarly, the World Health Organisation (WHO) has updated its air quality guidelines (AQG) to set new standards for outdoor PM exposure. The revised air quality standards have lowered the permissible annual mean PM exposure threshold from 10 to 5 μg/m³, while the 24 h mean exposure criterion has been adjusted from 25 to 15 μg/m³, reflecting a more stringent regulatory framework for public health protection. However, still over 99% of the global population remained subjected to ambient PM concentrations surpassing the WHO's annual AQG threshold, indicating widespread exceedance of recommended exposure limits.

Estimating global PM concentrations is essential for assessing population exposure and evaluating the associated epidemiological and health risks of air pollution. Although ground-based PM monitoring networks have expanded globally, including those integrated into the WHO ambient air pollution database, their spatial heterogeneity and insufficient coverage pose significant challenges in accurately quantifying global PM exposure. While early studies relied on interpolation methods and chemical transport models (CTMs) to quantify PM concentrations. However, the incorporation of satellite-derived aerosol optical depth (AOD) has enhanced estimation precision by providing higher spatial and temporal resolution. Ji et al. utilised Moderate Resolution Imaging Spectroradiometer (MODIS) AOD datasets, validated by AERONET, integrating multi-source remote sensing and meteorological parameters to predict PM concentrations in the Beijing-Tianjin-Hebei region with high precision. However, Zhang et al. introduced an alternative approach using multi-angle polarised top of atmosphere reflectance observations acquired from the GaoFen-5B satellite, demonstrating improved retrieval accuracy over traditional AOD-based methods. Similarly, integrating machine learning and deep learning (ML/DL) models with satellite data, reanalysis datasets, ground-based monitoring stations and model simulations, offer promising opportunities to reduce the uncertainties in global PM assessments, which further improves the accuracy. These technologies are increasingly employed for PM concentration forecasting due to their high predictive accuracy and capability to handle complex, nonlinear relationships. Wang et al. integrated meteorological variables from the ERA5 reanalysis dataset with machine learning algorithms to mitigate AOD discontinuities and enhance the spatiotemporal resolution of PM estimations in North China. Similarly, Li et al. employed a two-stage regression model with bias correction to compute PM concentrations using AOD, further assessing the mortality risk associated with pollution exposure in Iran. However, DL models, especially convolutional neural networks (CNNs) and long short-term memory (LSTM) networks, have shown higher accuracy in spatiotemporal PM estimation than conventional regression and ML models. Similarly, Koo et al. developed a ConvLSTM-DNN hybrid model tailored for Seoul, demonstrating superior performance over traditional Community Multiscale Air Quality (CMAQ) models by better capturing spatiotemporal patterns. Liu et al. further advanced short-term PM forecasting by incorporating a complex decomposition-based CNN-LSTM hybrid framework, optimising feature extraction and temporal sequence prediction through the combined advantages of convolutional and recurrent architectures.

The ensemble ML/DL methods, have shown great promise in geospatial air pollution prediction. The ensemble learning algorithms combine predictions from multiple models to achieve optimal results. Previous studies have shown that ensemble ML consistently achieve higher accuracy than individual models in evaluating environmental exposures. Di et al. developed an ensemble modelling framework by integrating multiple ML algorithms with diverse predictor variables to quantify daily PM concentrations at a 1 km × 1 km resolution across the U.S., achieving cross-validated R² values of 0.86 for daily estimates and 0.89 for annual averages. Similarly, Mohammadi et al. used meteorological data and four machine learning models (ANN, RF, SVM, KNN) and ensembles of classification trees RF to predict PM levels in Isfahan, finding ANN to be the most accurate (90.1%) for air pollution forecasting. Nowadays, a more advanced approach called deep ensemble machine learning model (DEMLM), which integrates various bases and other parameters to enhance prediction performance is used to predict atmospheric variables.

Exposure and risk assessment of mean PM and its extremes is crucial, as its concentration is expected to rise significantly in many regions worldwide. Extreme PM events can be categorised into short-term spikes, which occur due to sudden emissions or meteorological conditions lasting a few days, and long-term exposure, which results from persistently high levels over months or years. Recent global studies on PM concentrations have emphasised long-term exposure, typically using annual or monthly data. However, research on short-term PM exposure, spanning hours to days, remains relatively limited globally. More studies are now estimating daily PM concentrations at regional scales in countries like the United States, China, India, and Europe. These studies focus on short-term exposure and its global spatio-temporal variations. Kelly et al. developed Nine PM-Exposure models, including CTMs, interpolation, satellite-derived, Bayesian, and machine learning methods, estimated a 1 µg/m³ average PM decrease from 2011 to 2028, with greater reductions in high-emission areas and reduced exposure inequality among racial/ethnic groups. Similarly, Aunan et al. developed an Integrated Population-Weighted Exposure (IPWE) method combining ambient and household PM exposure, estimating 1.15 million annual premature deaths in China, with rural exposure nearly twice that of urban areas. However, variability in global training datasets and methodological inconsistencies in PM estimation hinder the comparability of regional assessments, complicating efforts to derive a consistent global exposure profile. Therefore, it is essential for studies to estimate extreme ambient global daily PM concentrations and track their variations over time and space using consistent study designs, modelling methods, and data source. Additionally, advancements in data collection and integration techniques will be essential in overcoming current limitations and providing more reliable and actionable insights into air quality management and policymaking.

Several challenges remain in global PM estimation and exposure assessment. The lack of ground-based observations in developing regions limits model accuracy, necessitating improved data integration techniques. Existing models exhibit limitations in accurately predicting extreme pollution events due to their low frequency and high variability, underscoring the necessity for more resilient forecasting frameworks. Inconsistencies in methodological approaches across studies hinder the comparability of global PM estimates, underscoring the importance of standardised modelling frameworks. While deep learning models have shown promise, their interpretability remains a challenge, requiring the development of ensemble ML/DL approaches for air quality prediction. This research introduces key innovations in extreme air pollution monitoring and analysis to address these gaps. The primary methodological advancement lies in the deployment of a DEMLM (hybrid deep leaning CNN-LSTM-DNN model) model for global PM quantification and prediction at an unprecedented spatiotemporal resolution of 0.5° × 0.625° and 1 day, respectively, using the meteorological and geographic features, spanning 1980 to 2023 and biases in the predicted data has been removed using WHO ambient air quality database. This long-term and high-resolution approach significantly improves existing models, enabling more precise tracking of pollution patterns. Furthermore, national and regional population-weighted mean PM concentrations were examined, including the number of days populations were exposed to PM levels that exceeded the US EPA daily limit of 35.5 μg/m³ (Air Quality Index > 100 ie;unhealthy for sensitive groups), maximum one-day pollution and 99 percentile of PM, respectively. In this context, population-weighted pollution extremes (PWPE) encompasses four distinct metrics that collectively characterise severe air pollution exposure: (1) exceedances of the USEPA standard of 35.5 μg/m³ for 24-hour PM concentrations, representing regulatory threshold violations; (2) annual maximum daily PM concentrations, capturing the most severe pollution episodes; (3) 99 percentile PM values, indicating recurrent extreme pollution events; and (4) population-weighted mean concentrations that exceed health-based guidelines, where all metrics are weighted by population density to prioritise human exposure assessment over simple spatial averaging. This analysis aimed to reveal the spatiotemporal distribution of exposure to PM over the four decades, providing insights into patterns and trends that could inform public health strategies and policy decisions. The DEMLM framework represents a significant leap forward in environmental health research, offering enhanced capabilities for capturing subtle spatiotemporal variations in PM distribution. This study advances existing knowledge by providing a more granular, comprehensive view of global air pollution patterns, enabling better-informed public health strategies and policy decisions. However, the practical significance extends to identifying regions experiencing extreme pollution events and temporal fluctuations, thereby facilitating targeted intervention strategies for air quality management.

Unequal spatio-temporal distribution of population-weighted pollution extremes through deep learning - npj Climate and Atmospheric Science