Estimation of pan-European, daily total, fine-mode and coarse-mode Aerosol Optical Depth at 0.1° resolution to facilitate air quality assessments

Science of the Total Environment 918 (2024) 170593 Available online 1 February 2024 0048-9697/© 2024 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/). Estimation of pan-European, daily total, fine-mode and coarse-mode Aerosol Optical Depth at 0.1◦resolution to facilitate air quality assessments Zhao-Yue Chen a , b , * , Raúl Fernando M´ endez Turrubiates a , Herv´ e Petetin c , Aleksander Lacima c , Carlos P´ erez García-Pando c , d , Joan Ballester a , * a ISGLOBAL, Barcelona, Spain b Universitat Pompeu Fabra (UPF), Barcelona, Spain c Barcelona Supercomputing Center, Barcelona, Spain d ICREA, Catalan Institution for Research and Advanced Studies, Barcelona, Spain HIGHLIGHTS GRAPHICAL ABSTRACT •New fAOD significantly improves weak AOD-PM associations in Europe. •They addressed coverage gaps in satellite aerosol data. •Different-sized PMs’ estimation required different AOD fractions. •A significant decline in AOD over Europe, driven by fAOD ARTICLE INFO Editor: Anastasia Paschalidou Keywords: Aerosol Optical Depth Satellite Aerosol Particulate matter fAOD cAOD ABSTRACT Aerosol Optical Depth (AOD) data derived from satellites is crucial for estimating spatially-resolved PM concentrations, but existing AOD data over land remain affected by several limitations (e.g., data gaps, coarser resolution, higher uncertainty or lack of size fraction data), which weakens the AOD-PM relationship. We developed a 0.1◦resolution daily AOD data set over Europe over the period 2003–2020, based on two-stage Quantile Machine Learning (QML) frameworks. Our approach first fills gaps in satellite AOD data and then constructs three components’ models to obtain reliable full-coverage AOD along with Fine-mode AOD (fAOD) and Coarse-mode AOD (cAOD). These models are based on AERONET (AErosol RObotic NETwork) observations, Gap-filled satellite AOD, climate and atmospheric composition reanalyses. Our QML AOD products exhibit better quality with an out-of-sample R 2 equal to 0.68 for AOD, 0.66 for fAOD and 0.65 for cAOD, which is 23–92 %, 11–13 % and 115–132 % higher than the corresponding satellite or reanalysis products, respectively. Over 91.6 %, 81.6 %, and 88.9 % of QML AOD, fAOD and cAOD predictions fall within ±20 % Expected Error (EE) envelopes, respectively. Previous studies reported that a weak satellite AOD-PM correlation across Europe (Pearson correlation coefficient (PCC) around 0.1). Our QML products exhibit higher correlations with ground-level PMs, particularly when broadly matched by size: AOD with PM10, fAOD with PM2.5, cAOD with PM coarse (R =0.41, * Corresponding authors at: ISGLOBAL, Barcelona, Spain. E-mail addresses: [email protected] (Z.-Y. Chen), [email protected] (J. Ballester). Contents lists available at ScienceDirect Science of the Total Environment journal homepage: www.elsevier.com/locate/scitotenv https://doi.org/10.1016/j.scitotenv.2024.170593 Received 16 October 2023; Received in revised form 12 January 2024; Accepted 29 January 2024 Science of the Total Environment 918 (2024) 170593 2 0.45 and 0.26, respectively). Different AOD fractions more effectively distinct PM size fractions, than total AOD. Our QML aerosol dataset and models pioneer full-coverage, daily high-resolution monitoring of fine-mode and coarse-mode aerosols, effectively addressing existing AOD challenges for further PMs exposures’ estimations. This dataset opens avenues for more in-depth exploration of the impacts of aerosols on human health, climate, visibility, and biogeochemical processes, offering valuable insights for air quality management and environmental health risk assessment. 1. Introduction In light of the alarming increase in ambient particulate matter (PM)- related deaths, from an estimated 2.1 million in 1990 to 4.14 million in 2019 (Institute for Health Metrics, 2020), it is crucial to address this pressing global health issue. In Europe alone, fine PMs (PM2.5) account for an estimated 307,000 premature deaths annually, surpassing other major pollutants (European Environment Agency, 2021). Accurate spatiotemporal air pollutant distribution, especially PM, is crucial. However, establishing and maintaining extensive ground-level monitoring stations, though essential, often faces economically unfeasible or space limits problem (Maag et al., 2018), limiting accurate spatial estimations for air pollutants in epidemiological modelling or pollutant control (Zhang et al., 2021). To address these challenges, satellite observation of atmospheric aerosols offers broader coverage compared to surface-level PM monitoring (Griffin, 2013). These atmospheric aerosols are measured as Aerosol Optical Depth (AOD), by detecting how much sunlight is absorbed and scattered by suspended particles. In general, AOD serves as a crucial indicator for surface PM estimates, especially in areas with limited ground-based monitoring data. The application of existing AOD data (satellite-derived AOD or reanalysis AOD data) to estimate PM in Europe encounters some obstacles, which might undermine its effectiveness and feasibility. Guo et al. (2023) reported that the regions north of 45◦N, where most European countries located, generally have much higher missing rate in satellite-derived AOD than other areas. Thicker clouds, low sun angle at high-latitudes, more frequent snow cover and the occurrences of polar day and polar night contribute to the substantial absence of satellite AOD measurements, emphasizing the critical necessity for comprehensive gap-filling strategies (Z.-Y. Chen et al., 2019; Gupta et al., 2016; Wei et al., 2018). Furthermore, the accuracy of satellite AOD is still subject to various factors, like instrument calibration, cloud contamination, and climate or geographic conditions (He et al., 2021), resulting in inconsistent AOD-PM correlations over many regions. Notably, Europe and South America exhibit weaker Pearson correlation coefficients (PCC) ranging from 0.1 to 0.12, which contrast with stronger the correlations observed in Northern America or East Asia (Christopher and Gupta, 2020), which range from 0.45 to 0.70. Compounded by limited scans from polar satellites, temporal disparities between satellite observations (1–2 times on daytime) and ground-level PM monitoring data (24-hour) weakens the association between AOD and PM. Overall, applying existing satellite-derived AOD on precise PM estimations in Europe remains a challenge. While aerosol reanalyses aim to mitigate some of these satellite-AOD limitations by offering complete AOD assimilation, they still suffer from coarse resolutions (~50–100 km) due to computational constraints (Bouttier, 2009) and inherent biases from uncertainties in emission inventories (Huang et al., 2021), necessitating urgent efforts to downscale these coarse data for enhanced accuracy. Moreover, the type and size distribution of aerosols play a crucial role in influencing the amount of light that is scattered or absorbed by the aerosols, impacting the AOD-PM relationship (Yan et al., 2017; Zang et al., 2021). However, many studies still rely on total AOD as the main predictor for PM2.5 or PM10, due to the lack of reliable size-segmented AOD data, such as fine-mode aerosols (fAOD) or Coarse-mode AOD (cAOD) (Ferrero et al., 2019; You et al., 2015). This lack can lead to inaccurate PM predictions, particularly in regions with diverse aerosol compositions. For instance, in areas with abundant coarse aerosols, higher AOD levels may primarily reflect the presence of larger particles, failing to capture the contribution of smaller, health-threatening particles like PM2.5. Satellite-derived products also have some challenges in this issue. For instance, Moderate Resolution Imaging Spectroradiometer (MODIS) fine mode fraction products are only available over oceanic areas, due to high uncertainties in its products over land (Levy et al., 2013); Polarization and Directionality of the Earth’s Reflectance (POLDER) products, although informative, are accessible only for short time spans (Dubovik et al., 2019). To address this concern, a few experimental studies (Chen et al., 2020; Yan et al., 2022) suggested applying machine learning methods to enhance satellite data quality and to calibrate stronger biases from physical methods during the cold season. However, these approaches are still constrained by satellite data limitations like missing gaps and limited scans per day from polar satellites. Moreover, these fine-segment products often have coarser spatial resolutions (0.5–1◦) than AOD-only data, insufficient for precise PM exposure estimation. Despite these challenges, fAOD data are essential for accurate PM predictions and for understanding the role of aerosols in the global radiative budget, carbon, nitrogen, and sulfur cycles, and climate change (Alahmad et al., 2023; Chung et al., 2005). These aerosols are closely linked to human activities (Bellouin et al., 2005), underscoring the pivotal role of fAOD distribution in identifying potential anthropogenic contributions to air pollution and global climate change (Yan et al., 2021). Overall, the estimations of air pollution exposure and subsequent health assessments in Europe encounter significant constraints due to limitations inherent in existing AOD and their size-segmented products. There exists a necessity for a full-coverage, higher resolution, sizesegmented AOD dataset covering the entirety of Europe over an extended decade to rectify these limitations and facilitate robust assessments of air quality and health implications. Thus, we developed a new set of AOD models to generate the first full-coverage daily sizesegmented AOD dataset at a 0.1-degree resolution for Europe spanning 2003 to 2020. This dataset significantly enhances our comprehension of aerosol distribution across time and space within the continent. To overcome satellite data limitations, we initially trained the MAIAC AOD gaps-filled model to handle the missingness of MAIAC AOD. Next, we employed Quantile Machine Learning (QML) models with ground-level AERONET data, Gap-filled MAIAC, climate and aerosol reanalyses, to derive a high-resolution daily dataset of AOD, fAOD and cAOD covering Europe over the period 2003–2020. Subsequently, we assessed whether these improved AOD products exhibited a more robust correlation with surface PM. These improved estimates exhibit stronger correlations with PM10, PM2.5, and PMcoarse compared to previous products. This enhancement lays a critical groundwork for estimating the exposures of air pollutants in instances where ground-level monitoring stations are absent. Unlike prior studies, our method addresses limitations addresses limitations like data quality, resolution and comprehensive coverage, bridging satellite aerosol data gaps and provides reliable fine-mode and coarse-mode estimates, essential for epidemiological studies and environmental monitoring. 2. Data 2.1. AERONET data We collected cloud-screened, ground-based AOD data from AEROZ.-Y. Chen et al. Science of the Total Environment 918 (2024) 170593 3 NET v3.0 (Holben et al., 2006). This data source also provides the decomposition of total AOD into fAOD and cAOD, based on a Spectral Deconvolution Algorithm validated by (O’neill et al., 2003). The uncertainty (median RMSE) in the AOD, fAOD and cAOD calculation is reported as 0.005, 0.016 and 0.013. As it is commonly done in the literature, AERONET data are here considered as the “ground truth” to validate other aerosol data (Bright and Gueymard, 2019; Gueymard and Yang, 2020). Our data are from 257 sites located in the research domain here considered: 27-72N ×25W-45E. To be comparable with the satellite and reanalysis data, the AERONET AOD data at 550 nm (AOD550) was interpolated from the AOD500 (Duarte and Duarte, 2020; Gupta et al., 2020). The Eq. (1) used for this interpolation is as follows: AOD550 =AOD500*(550 500)−∝t (1) where ∝tis the is the AERONET AOD Ångstr¨ om exponent at 500 nm, which is obtained from AERONET SDA output. Before obtaining the fAOD550 and cAOD550, we first transformed the Fine mode fraction at 550 nm (FMF550) from the 500 nm (FMF500) using the Eq. (2): FMF550 = fAOD500*(550 500)−∝f AOD500*(550 500)−∝t=FMF500*(550 500)∝t−∝f (2) where ∝fis the AERONET fAOD Ångstr¨ om exponent at 500 nm. All of these parameters are available from AERONET SDA products. Finally, we obtained fAOD550 and cAOD550 by following formula: fAOD550 =AOD550*FMF550 (3) cAOD550 =AOD550*(1−FMF550)(4) 2.2. Satellite data We collected daily AOD data for the period 2003–2020 from MODIS Collection 6 (C6) (NASA, 2023). The data are based on the Multiangle Implementation of Atmospheric Correction (MAIAC) algorithm, with time series analysis and an image-based processing algorithm to provide accurate, fixed-grid aerosol estimations (Lyapustin et al., 2011). The data were filtered for the quality assurance flags of the NASA guidelines (i.e. Quality Assurance of cloud mask =clear sky) (Lyapustin et al., 2018), and provided 1 km ×1 km resolution over Europe. 2.3. MERRA-2 reanalysis We also retrieved data from MERRA-2, the atmospheric composition reanalysis developed by NASA. The Goddard Earth Observing System Model (GEOS-5) data assimilation system ingest various sources (e.g. data from AERONET sites, MODIS, MISR and AVHRR sensors) to simulate 3-hourly aerosol data (1980 onwards) at 0.625◦×0.5◦resolution (Randles et al., 2017). Previous studies (Che et al., 2019) showed that MERRA-2 reproduces the general trends in annual and seasonal AOD, both at regional and global scales, but with significant biases in some locations. 2.4. CAMS reanalysis CAMSRA, an ECMWF air quality reanalysis, used hourly AOD data from Envisat’s AATSR, MODIS Aqua and Terra sensors, and in-situ measurements (Inness et al., 2019). Although coarser in resolution (0.75◦×0.75◦), it assimilated wide range of Aerosol components since 2003 (Bozzo et al., 2017; Flemming et al., 2015), with slightly underperforming MERRA-2 in AOD estimates (Gueymard and Yang, 2020). 2.5. ERA-5 reanalysis for atmospheric meteorological data Previous studies (Gui et al., 2019; Huang et al., 2007; Tai et al., 2010; Yan et al., 2022; Zhou and Savij¨ arvi, 2014) have analysed the associations between weather conditions and the concentration of fineand coarse-mode aerosols. For example, high-pressure events, characterised by atmospheric stability and low winds, retain the smaller particles, which is seen with higher-than-normal fine-mode aerosol levels (Gui et al., 2019; Tai et al., 2010). Moreover, rainfall washes out the particles from the lower part of the troposphere, especially the largest particles. There are other pathways by which aerosols can also affect weather conditions, for example by reflecting and absorbing the incoming UV radiation (Zhou and Savij¨ arvi, 2014), or by changing the conditions for the condensation of water in the cloud (Huang et al., 2007). To account for the impact of meteorological factors on aerosols, we collected atmospheric variables from ECMWF ERA-5 reanalysis, including boundary layer height, downward UV radiation, cloud cover, and precipitation. More information about the resolution and data source for each meteorological variable can be found in Table S2. 2.6. ERA5-land reanalysis for surface data Apart from atmospheric meteorological data, surface data, including vegetation indices and soil types, impact aerosols. For example, forests contribute to a large extent to particle removal, and previous studies found the deposition velocity of ultrafine particles is generally more sensitive to leaf area index than leaf area density (Huang et al., 2015; Lin et al., 2018). Also, the dry deposition of particles is affected by properties of the vegetation elements (such as leaves and branches) and soil types (Gr¨ onholm et al., 2009). In this study, we also found the significant contributions of leaf area index high vegetation, leaf area index low vegetation and soil types in aerosol modelling. Higher leaf area index high vegetation means more evergreen trees, deciduous trees or forest, while higher leaf area index low vegetation represents more crops and mixed farming, grass or shrubs. For bare ground or places with no leaves, both of them will be close to zero. The soil types describe how coarse the soil is, representing the water holding ability of soil. Coarser soil generally has lower water holding ability. ERA5-Land provides relevant surface variables (Table S2), including some near-ground meteorological data or land cover data (e.g., wind speed, relatively humidity, vegetation index), at a resolution of 0.1◦×0.1◦, influencing satellite data quality in reanalysis. 2.7. Ground-level particulate matter data To investigate the potential application of QML AOD, fAOD and cAOD predictions, PM 10 , PM 2.5 and PM coarse (PM 2.5–10 ) data from 2003 to 2020 in Europe were obtained from EEA’s AirBase and Air Quality eReporting (AQER). AirBase is a repository of air quality data that contains measurements from 1969 to 2012, while AQER is a newer system that was introduced in 2013 to replace the AirBase system. We collected data from both sources to ensure a wide coverage of the time period. For AQER, we only collected the E1a data, fully validated after an official delivery. 3. Methodology 3.1. Overall model In this study, we used quantile lightGBM (Light Gradient Boosting Machine) models to separately obtain predictions of AOD, fAOD and cAOD. This tree-based framework is known for its high performance and computational efficiency, particularly suited for handling complex relationships within large-scale research areas (Ke et al., 2017). Its flexible framework allows for customizable APIs, enabling the customization of a distance weighted loss function to improve model performance and the Z.-Y. Chen et al. Science of the Total Environment 918 (2024) 170593 4 generation of quantile predictions, aiding in estimating prediction uncertainties. Given the high missing rate (75.79 %) of Satellite AOD from 2003 to 2020, we employed a two-step approach. Initially, we employed a gap-filling model based on selected variables (listed in Table S1) to impute missing MAIAC AOD values. Subsequently, we established three separate models to estimate total AOD, Fine-mode AOD (fAOD) and Coarse-mode AOD (cAOD), using gap-filled MAIAC AOD, AERONET AOD, and selected factors listed in Table S1, following the methodological steps outlined in Fig. 1. 3.2. Variable selection and model developments To ensure consistent resolution across datasets for subsequent analysis, we employed bilinear resampling to standardize all gridded data to a horizontal resolution of 0.1◦×0.1◦(equivalent to approximately 9 km at mid-latitudes). Despite the higher native resolution of the 1 km MAIAC data, persistent high missing rates may introduce uncertainty to the predictions. The subsequent available dataset with complete coverage operated at 0.1-degree resolution. To ensure robustness in our methodology, we’ve opted for the 0.1-degree resolution as the final resolution. For temporal resolution, we computed daily averages from various datasets, including hourly data from MERRA-2, ERA5, ERA5land, and 3-hourly from CAMSRA data. In the first gap-filling step, we randomly subset 2000 available 0.1degree regrided MAIAC grid cells each day (or all available grids if <2000) to create training datasets for subsequent gap-filled models. In the calibration step, we extracted the corresponding grid at AERONET site locations to train the calibration/components models. Variable selection employed the Boruta procedure (Kursa and Rudnicki, 2010), executed twice: firstly, for gap-filling MAIAC AOD, and secondly, for AOD, fAOD, and cAOD prediction. The robustness of Boruta method is well validated in previous simulation studies (Degenhardt et al., 2019). This algorithm identifies significant variables by comparing original feature importance scores to shadow feature scores, with importance marked if significantly higher. Boruta was executed five times with 20 % random subsamples of sites, resulting in statistically significant variables (p-value <0.05) listed in Table S1, and top 20 important features illustrated in Fig. S1. According to training loss gain in the models, we find that dust and sea salt AOD contribute more to MAIAC AOD gap-filled process, likely due to their relatively simple spatial pattern and distinctive spectral characteristics, making it easier for the model to detect and estimate their contributions when satellite AOD data is unknown. Additionally, combining AOD data from CAMSRA, MERRA, and Gap-filled MAIAC are key contributors to total AOD estimates. When focusing on fAOD estimates, organic matter and sulphate AOD play a prominent role. For cAOD estimates, dust AOD and windspeed emerge as the primary influencing factors. Moreover, meteorological factors such as temperature, boundary layer dissipation, humidity, windspeed, and air pressure play critical roles in our models. These findings illuminate the intricate interplay between aerosol components, meteorological data and their impact on AOD predictions, enhancing our understanding of AOD variations. To train the gap-filled and calibration/components models, we began by randomly allocating 70 % of the sites as training data for the quantile lightGBM models. An additional 20 % of the sites were designated for optimizing model hyperparameters or validating subsequent performance enhancement techniques (as detailed in the Section 3.3). The final 10 % of sites served as entirely independent test data for unbiased evaluation. 3.3. Calibration/components models’ improvements To address limited AERONET sites, in the second step, we applied three techniques to improve our calibration/components models: - Distance weighted loss function: given the heterogeneous spatial distribution of the 257 AERONET sites, we introduced distance weights in the loss function to avoid the overfitting in regions with higher AERONET station density. This mitigates selective overfitting, giving less weight to denser station areas. The mathematical formulation is Fig. 1. The Quantile Machine Learning (QML) Workflow for AOD, fAOD and cAOD, Comprising Gap-filled (Green part), Calibration (Blue part, same process applied to three components models for AOD, fAOD and cAOD) and independent PMs exploration stage (Red part). Z.-Y. Chen et al. Science of the Total Environment 918 (2024) 170593 5 Wi=(Di−Dmin) (D−Dmin)(5) L(Y,Y*) = ∑n i=1Wi*L(yi,y* i)(6) where Di is the distance of station i to its nearest site; Dmin and D denote the minimum and average distance to their nearest site; and L(Y,Y*) and L(yi,y* i)are the overall loss function (Y and Y* are the observation and prediction vectors) and the loss functions for station i (yi and y* i are the observations and predictions at station i). - minimum directional distance: We used this method to assess model predictions’ sensitivity to site distribution along the longitudinal and latitudinal axes. Fig. S2 outlines how we constructed spatial calibration variables, obtaining four quadrants by drawing the longitudinal and latitudinal axes. A distance of 2000 km was defined as an upper threshold in each quadrant. Distances to closest stations were included as inputs, accounting for minimum directional distance. - white-noise data augment: this technique reduces overfitting by adding white noise to predictors. Gaussian-distributed white noise (zero mean, predictor’s variance) was introduced to duplicated training datasets. These techniques enhance model stability and accuracy for locations with fewer monitoring stations. Two validation approaches were employed: randomly selecting 20 % of sites for training (keeping 80 % for validation) and using the top 20 % farthest sites (with distances at least over 130.5 km) for training, evaluating against remaining sites. and trained the model using the remaining sites. These techniques improved the overall performance of AOD, fAOD, and cAOD predictions, especially in sparsely monitored regions (Table S3). In sensitive analysis, we also compared the predictive accuracy of LightGBM, Random Forest, XGBoost, and Support Vector Machines (SVM) using 5-fold cross-validation. LightGBM had the highest predictive accuracy, as shown in Table s8. 3.4. AOD model validation To further assess our models’ out-of-sample predictive capabilities, we also performed nested 5-fold cross-validation, yielding both spatial and temporal out-of-sample predictions. The 257 AERONET sites were divided into five equal subsamples, where four were for training and tuning, and one for prediction. Models were trained on 80 % of the four subsamples, with 20 % used for hyperparameter optimization. This process was repeated for each of the five subsamples, generating five models and out-of-sample predictions for all the sites. For the temporal out-of-sample predictions an analogous strategy was used, but with six subperiods of three consecutive years, covering our 18-year period. Similar structures, hyperparameters, and results were found for spatial and temporal models (Table S1), ensuring model strategy stability. Final configurations were used to retrain models with all observations, predicting daily AOD, fAOD, and cAOD spatiotemporal estimates. To interpret these out-of-sample predictions, comparing our predictions with MERRA-2, CAMSRA, and MAIAC data involved evaluating against AERONET AOD. For the fAOD and cAOD, the reanalysis data did not provide fine-mode and coarse-mode information. Hence, we used Phy-DL (PDL) Fine-mode Fraction (FMF) global products as reference for fAOD and cAOD, considering their proven accuracy in global comparisons ((Yan et al., 2022). They found PDL FMF outperforms existing FMF products: Polarization and Directionality of the Earth’s Reflectance (POLDER) FMF, Multi-angle Imaging Spectro Radiometer (MISR) FMF and MODIS FMF. Their corresponding correlation with AERONET AOD is 0.78, 0.48, 0.42 and 0.37. Besides, the general correlation between PDL fAOD (=PDL FMF * MAIAC AOD) and AERONET AOD (PCC = 0.781) in our domain is similar with the result reported by (Yan et al., 2022). The Evaluation metrics (R-squared, NMB (Normalized Mean Bias), NRMSE (Normalized Root Mean Square Error), 90 % PI coverage, and percentage of predictions within 20 % expected error envelopes) are detailed in Table S3. 3.5. Exploratory correlation analysis with surface PMs One crucial application of QML AOD, fAOD, and cAOD predictions is their role in improving PM10, PM2.5, and PMcoarse predictions. To establish this connection, we investigate the relationship between aerosol size distribution and surface PM levels. Reliable AOD, fAOD, and cAOD data enable us to assess their potential as enhanced PM indicators. Since PM monitoring stations are distinct from AERONET stations, we first predict QML AOD and its components for PM monitoring locations. We then compare Spearman correlation coefficients (SCC) across various AOD products with PM10, PM2.5, and PMcoarse (Red part in Fig. 1). SCC is suitable for evaluating the often monotonic but not strictly linear AOD-PM relationship. Furthermore, PMs data are independent dataset compared to AERONET data, so this enhances the independent validation of AOD product quality and its practical application value. 4. Results and discussion 4.1. Description of AERONET data Fig. 2 shows the spatial and temporal distribution of AERONET station data. On average, each site has 662 observations during the 2003–2020 period, with approximately 28 % of sites having over 1000 observations. Early-built sites (constructed before 2010) are distributed relatively evenly throughout Europe, allowing models in the early period to be trained with data that represents the continent as a whole. Annual numbers of observations have steadily risen over the last two decades, peaking in 2018. In relative terms, numbers of daily observations per site gradually increased until 2011, when it stabilised at around 150 annual days per site. This plateau period (2011–2020) indicates that many sites average approximately 150 observable days due to clear skies required by sun photometer or intermittent offline periods for calibration (GLOBE, 2010; Holben et al., 2006)). The earlier rising trend hints that some AERONET sites might not have consistently operated, potentially impacting QML AOD model performance before the plateau phase. Fig. 3 depicts the spatiotemporal distribution of AERONET AOD, fAOD and cAOD observations. Central and eastern Europe exhibit higher AOD and fAOD, reflecting significant anthropogenic pollution sources primarily linked to anthropogenic (Bellouin et al., 2005) and secondary aerosols (e.g., sulphate, nitrate, ammonium) (Seinfeld and Pandis, 1998; Zhao et al., 2018), In contrast, southern Europe experiences elevated cAOD due to dust intrusions from the Sahara desert, as well as Mediterranean sea salt advection (Meloni et al., 2008; Prospero et al., 2014). The long-term trend in AOD reveals 28.5 % decrease per decade for AOD and 27.2 % decrease for fAOD. Given that fAOD is largely associated with anthropogenic emissions, its decreasing trend reflects the reduction in anthropogenic pollution resulting from air quality plans implemented in Europe. Notably, cAOD remains stable over time. 4.2. Spatial and temporal evaluation To further validate the AOD models, we compared in time and space the AOD predictions with the estimates from MERRA-2, CAMSRA and MAIAC. To account for the large fraction of missing values in satellite MAIAC AOD among AERONET sites (64 %), we divided the validation results into two subgroups based on the availability of satellite MAIAC AOD data. The first subgroup, referred to as the “Sat scenario”, included validation dates and sites where satellite MAIAC AOD data were available. The second subgroup, referred to as the “Non-Sat scenario”, Z.-Y. Chen et al. Science of the Total Environment 918 (2024) 170593 6 included validation dates and sites where satellite MAIAC AOD data were not available. This division was made in order to ensure comparability among the different products and to provide insights into the factors that limit the performance of the models. Unlike MERRA-2, CAMSRA and MAIAC, we used our models to generate out-of-sample predictions in space and time in order to shed more light into the validation. 4.2.1. Total AOD product Fig. 4 validates the different AOD products against the AERONET observations, by comparing QML’s spatial and temporal out-of-sample AOD predictions with MERRA-2, CAMSRA and MAIAC AOD. The NMB (Normalized Mean Bias) maps show that QML aligns closely with AERONET data but slightly underestimates both spatial and temporal out-of-sample predictions. Additionally, QML exhibits the lowest NRMSE, about 33 % smaller than other products. Conversely, CAMSRA overestimates, especially in Northern Europe, the UK, and certain Mediterranean coastal areas, while satellite AOD tends to underestimate inland and overestimate coastal regions. In Fig. 4 (a1–e1), the box plots portray bias (estimation minus AERONET) across AERONET AOD deciles (Q1–Q10), with the range in each decile additionally shown in Table S4. The orange and green lines represent the expected error (EE) envelopes of ±0.05 ±20 %/40 % for each decile, used in previous studies (Levy et al., 2010; Xiao et al., 2016; Yan et al., 2022). The ±0.05 accounts for measurement errors, like instrument calibration uncertainties or atmospheric condition uncertainties. For fAOD and cAOD, this threshold is adjusted to ±0.025 due to their lower variability compared to total AOD. This balance ensures dependable assessment for small percentiles and sufficient evaluation for larger AOD values. Overall, QML demonstrates high agreement, with ~94 % (temporal) and ~92 % (spatial) predictions within 20 % EE envelopes, surpassing other products’ 20 % EE ranging from 76 % to 78 %. QML exhibits the highest R-squared: 0.71 (temporal) and 0.68 (spatial). In the Sat scenario, all products overestimate lower AOD deciles, underestimating higher ones. However, in this scene, QML has >96 % predictions within 20 % EE, less bias in lower deciles compared to others. Non-Sat scenario sees significant improvements in QML estimates than other estimates; QML narrows bias in Q1–Q8, achieving 88 %–92 % within 20 % EE. Overall, both scenarios show that QML AOD mainly fixed the overestimations problem of reanalysis data in lower quantiles, especially in Non-Sat scenario. Lastly, as QML AOD is less likely to overestimate values in the lower range, but like other products, it may still exhibit some overestimates in higher ranges, leading to a slightly negative NMB. Fig. 5 presents annual AOD product performance via R-square, NMB, and NRMSE. QML consistently outperforms reanalysis and satellite estimates in R-squared. NMB increases slightly for all, shifting from Fig. 2. Spatial and temporal distribution of AERONET sites: (a) number of daily values, (b) built-year and (c) number of observations for whole domain (left axis, green bars) and average of observable days per site (right axis, blue line). Z.-Y. Chen et al. Science of the Total Environment 918 (2024) 170593 7 negative to positive values for MERRA-2 and MAIAC estimates. This trend could be blamed for the decrease in AERONET AOD over time, coupled with the overestimating tendency of most products in the lower deciles. In contrast, QML’s NMB remains relatively constant. NRMSE slightly decreases over time for all, with QML consistently lowest. After categorizing the data into two scenarios, it was evident that most products exhibited slightly worse in the non-Sat scenario, probably due to the missing information from their satellite inputs. Nevertheless, QML consistently maintained superior performance in both scenarios. Further validation statistics for individual months (Fig. S3) and weekdays (Fig. S4) are available in the Supplementary, confirming QML’s temporal and spatial predictions’ superiority over reanalysis and satellite AOD estimates. All products demonstrated better fit with AERONET data during the warm season (April–September), while predicting AOD in the cold season (October to March) proved more challenging due to cloud and rain hindrances to sun photometer observations (GLOBE, 2010). CAMSRA AOD tended to overestimate AOD in warm seasons and underestimate it in cold seasons, unlike QML’s relatively stable performance across seasons. Weekly cycles exhibited no significant variations among days, likely due to human activities’ limited impact on columnintegrated AOD. QML assigns 90 % prediction intervals (PI) to estimate prediction uncertainty (Table S4). QML’s temporal and spatial PI achieved average coverages of 84.2 % and 83.5 %, respectively, slightly below the anticipated 90 %. Q1 and Q10 intervals showed difficulty, with coverage at 55 % and 68 % respectively, suggesting slight underestimation of intervals by our quantile models. 4.2.2. Fine-mode AOD product Fig. 6 validates the QML and PDL fAOD products against the AERONET observations. PDL fAOD displays uneven predictions, overestimating in Southwest Europe and underestimating in Central and Eastern Europe. This pattern is probably due to the preference of PDL to overestimate small fAOD values and to underestimate high fAOD values (as seen in Fig. 6 (a1)), coupling with lower fAOD values in Southwest Europe than Central and Eastern Europe (Fig. 3 (b)). In contrast, In contrast, QML fAOD narrows bias in smaller deciles (Q1–Q5), avoiding similar patterns, with over 89 % (spatial) and 90 % (temporal) predictions within 20 % EE in Sat scenarios. Underestimation of high fAOD values persists in QML estimates for some locations. In Non-Sat scenarios, QML maintains 77–79 % of the predictions within 20 % EE. QML’s overall R-squared is 0.66 (spatial) and 0.68 (temporal). As the PDL fAOD is only available in the satellite scenario, Fig. 7 (left columns) compares PDL and QML fAOD year-to-year in Sat scenario. PDL fAOD’s performance varies over years more than QML, with NMB (%) underestimating pre-2008 due to its uneven spatial bias in early period. Prior to 2008, more AERONET sites were located in western Europe (Fig. S5), where PDL fAOD was underestimated (Fig. 6A). Afterward, more sites were set in areas in which the product tends to overestimate the predictions (e.g., Southwest Europe), offsetting the Fig. 3. Spatial distribution of the median value of AERONET (a) AOD, (b) fAOD and (c) cAOD data. (d) Temporal evolution of AOD (red +blue), fAOD (blue) and cAOD (red). Z.-Y. Chen et al. Science of the Total Environment 918 (2024) 170593 8 Fig. 4. Spatial distribution maps of Normalized Mean Bias (NMB) (a–e) and Box Plots depicting Bias in AOD Estimations (Estimation minus AERONet Data) across Various Quantiles of AERONet AOD (a1–e1) for the Period 2003–2020. The datasets considered comprise: MERRA-2 AOD (a-a1), CAMSRA AOD (b-b1), MAIAC AOD (c-c1), Out-of-Sample QML Spatial Prediction (QML AOD Spat) (d-d1), and Out-of-Sample QML Temporal Prediction (QML AOD Temp) (e-e1). The ‘N(%)’ Represents the Sample Size Proportion. The 20 % and 40 % Expected Error (EE) Envelopes Indicate ±20 % and ±40 % Deviations from Observations. Within Each Box Plot, the Upper, Middle, and Lower Lines Correspond to the 75th, Median, and 25th Percentiles, Respectively. Z.-Y. Chen et al. Science of the Total Environment 918 (2024) 170593 9 bias and bringing the NMB of PDL fAOD closer to 0 after 2008. The NRMSE results also suggest that PDL fAOD has a larger error than QML fAOD. In contrast, the QML fits better with stable errors over years. We note that the availability of PDL AOD in our dataset is lower than the satellite MAIAC AOD (21 % VS 36 %), because PDL FMF is not always available when satellite AOD is available. Therefore, it is necessary to obtain fAOD estimates in non-Sat scenarios. Fig. S6 shows that QML’s Non-Sat R-squared is around 0.58–0.70, around 10 % lower than the sat scene. In the seasonal analysis (Fig. S7), fAOD patterns mirror total AOD, performing better in summer. QML fAOD stable seasonal performance surpasses PDL. At the weekly level, there is no difference in performance among different days of the week (Fig. S8). QML fAOD’s PI coverage (Table S5) averages 83.3 % (temporal) and 84.4 % (spatial), akin to total AOD results, Fig. 5. Performance comparison (R-squared, NMB, NRMSE) of Various Data Sources for AOD Prediction: QML spatial out-of-sample prediction (QML AOD-Spat), temporal out-of-sample prediction (QML AOD-Temp), CAMS AOD, MAIACAOD, MERRA-2 AOD. Z.-Y. Chen et al. Science of the Total Environment 918 (2024) 170593 16 period (Papayannis et al., 2008). 4.5. Summary of evaluation of the AOD products Fig. 13 summarizes AOD product comparisons across eight dimensions: Accuracy, Stability, Bias within 20 % EE, Correlation with PMs, TI Index, Coverage, Resolution, and Product Period. Comparison is based on common 2003–2020 period. For AOD, QML product performs best in accuracy (R 2 : 0.69 (QML AOD) vs 0.36–0.56 (others)) and stability (variation coefficients of R 2 : 0.24 (QML AOD) vs 0.35–0.53 (others)), and also as an optimal indicator of ground-level PM10 (Correlation with PM10: 0.41). Correlation with AERONET AOD is generally linked to PM correlation, but exceptions exist. CAMSRA, well-fitted with AERONET AOD, shows the lowest PM10 correlation among products. Factors like potential over-fitting in location or time might influence these behaviours. For example, the location of ground-level PM observations generally does not coincide with the AERONET sites, but it is still unknown how CAMSRA AOD perform in those locations without AERONET ground-truth measurements. Any potential over-fitting may worsen the AOD-PM relationship. Furthermore, PM is measured almost every day, but AERONET AOD is only available with clear sky condition. Thus, it is impossible to validate the performance of CAMSRA in those cloudy or rainy days. These unknown biases also additionally deteriorate the AOD-PM relationships. Among trends, satellite MAIAC AOD is least consistent with AERONET AOD due to missing values, so it requires additional caution in their trend analysis. Additionally, the satellite AOD is the only product without full temporal coverage, with only 24.21 % coverage. MERRA-2 provides the longest record (from 1980), while our period starts in 2003 due to CAMSRA input. Fig. 13 (b) highlights QML fAOD’s superiority over PDL fAOD. QML fAOD is more accurate, stable, and correlates better with PM2.5. QML improves PDL’s shortcomings, enhancing overall spatial coverage (14.96 % to 100 %) and resolution (1◦to 0.1◦). In trend comparison, QML fAOD aligns better with AERONET data due to improved coverage. QML greatly improves the performance for cAOD (Fig. 13(c)): more accurate, robust performance and more highly correlated with Coarse PM. In the trend comparison, the QML cAOD agrees well with AERONET data, while PDL cAOD shows the opposite trend against AERONET data (infinite TI value). Overall, the QML cAOD provides a valuable tool for coarse aerosol time series analysis, with its full coverage and higher resolution aiding future Coarse PM estimation. 5. Conclusion Europe is the one of regions with the poorest association between satellite AOD and ambient PM2.5 (Christopher and Gupta, 2020), so it is urgent to establish a reliable link between satellite AOD and ambient PM2.5, enhancing accurate PM estimation. To address this gap, we’ve developed an innovative 18-year aerosol dataset (AOD, fAOD, and cAOD) at a 0.1◦resolution, offering valuable insight into the distribution of particles of varying sizes across Europe. The out-of-sample validation rigorously assesses QML AOD, fAOD, and cAOD in both spatial and temporal dimensions. These products outperform other products, achieving 21 %–55 % lower NRMSE (reached: 20.63 %, 23.12 %, 30.86 %) and 11–132 % higher R 2 (0.68, 0.66, 0.65). Over 91.6 %, 81.6 %, and 88.9 % of biases respectively fall within a ±20 % EE envelope. Notably, QML fAOD significantly improves PM2.5 association, doubling spearman correlation to 0.45 from 0.10 to 0.29. Additionally, our findings also reveal QML AOD, fAOD, and cAOD serve as better indicators for PM10, PM2.5, and Coarse PM, respectively. This newly developed aerosol dataset and the associated models not only overcome the limitations posed by missing satellite aerosol information, such as lower coverage and temporal discontinuity, but also fulfil the pressing demand for reliable fine-mode and coarse-mode AOD data. This dataset serves as a valuable tool for monitoring and analysing aerosols at varying spatial and temporal scales in Europe, offering profound insights into their multifaceted impacts on human health, the environment, and the climate. Additionally, it equips researchers and policymakers with a powerful resource to enhance air quality monitoring and analysis, thereby contributing to our collective efforts to safeguard human health and the planet’s well-being. Financial support In this initial version of the geodatabase, the authors from ISGlobal would like to express their gratitude for the support they received from various organizations. The Spanish Ministry of Science and Innovation’s “Centro de Excelencia Severo Ochoa 2019–2023” Program (CEX2018000806-S-20-1), the Ministry of Research and Universities of the Government of Catalonia (2021 SGR 01563), and the Generalitat de Catalunya through the CERCA Program all provided support. ZC acknowledges support from the grant PRE2020-091985 funded by MCIN/AEI/10.13039/501100011033 and by European Social Fund invests in your future. CP acknowledges funding from the AXA Research Fund through the AXA Chair on Sand and Dust Storms at BSC, the European Research Council (ERC) under the Horizon 2020 research and innovation program through the ERC Consolidator Grant FRAGMENT (grant agreement no. 773051), H2020 ACTRIS IMP (#871115), and the Department of Research and Universities of the Government of Catalonia through the Atmospheric Composition Research Group (code 2021 SGR 01550). HP has received funding from the Ministerio de Ciencia e Innovaci´ on through the MITIGATE project (grant no. PID2020113840RA-I00 funded by MCIN/AEI/10.13039/501100011033) and the Ramon y Cajal grant (RYC2021-034511-I) and the European Union’s NextGeneration EU/PRTR (PID2020-116324RA695). JB gratefully acknowledges funding from the European Union’s Horizon 2020 and Horizon Europe research and innovation programs under grant agreements No 865564 (European Research Council Consolidator Grant EARLY-ADAPT) and 101069213 (European Research Council Proof-ofConcept HHS-EWS), as well as from the Spanish Ministry of Science and Innovation under grant agreement No RYC2018-025446-I (programme Ram´ on y Cajal). Research data The pan-European high-resolution aerosol optical depth (AOD) daily estimations and its fraction products developed by this study are available at https://zenodo.org/record/8315721 (Chen et al., 2023). The QML AOD data are in the Geotiff format on a daily scale. CRediT authorship contribution statement Zhao-Yue Chen: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Software, Validation, Visualization, Writing – original draft, Writing – review & editing. Raúl Fernando M´ endez Turrubiates: Data curation, Formal analysis, Resources, Software. Herv´ e Petetin: Conceptualization, Methodology, Supervision, Validation, Writing – review & editing. Aleksander Lacima: Methodology, Validation. Carlos P´ erez García-Pando: Conceptualization, Funding acquisition, Methodology, Supervision, Validation, Writing – review & editing. Joan Ballester: Conceptualization, Funding acquisition, Investigation, Methodology, Project administration, Supervision, Validation, Writing – review & editing. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Z.-Y. Chen et al. Science of the Total Environment 918 (2024) 170593 17 Data availability Data will be made available on request. Acknowledgements The authors gratefully acknowledge the European Centre for Medium Range Weather Forecasts, MERRA-2, ERA-5 and AERONET teams for their effort in making the data available. Appendix A. Supplementary data Supplementary data to this article can be found online at https://doi. org/10.1016/j.scitotenv.2024.170593. References Alahmad, B., Khraisha, H., Althalji, K., Borchert, W., Al-Mulla, F., Koutrakis, P., 2023. Connections between air pollution, climate change & cardiovascular health. Can. J. Cardiol. https://doi.org/10.1016/j.cjca.2023.03.025. Bellouin, N., Boucher, O., Haywood, J., Reddy, M.S., 2005. Global estimate of aerosol direct radiative forcing from satellite measurements. Nature 438, 1138–1141. https://doi.org/10.1038/nature04348. Bouttier, F., 2009. Fine scale versus large scale data assimilation—a discussion. In: Fifth WMO Symposium on Data Assimilation. WMO World Weather Research Programme Melbourne, VIC, Australia, p. 8. Bozzo, A., Remy, S., Benedetti, A., Flemming, J., Bechtold, P., Rodwell, M.J., Morcrette, J.-J., 2017. Implementation of a CAMS-based Aerosol Climatology in the IFS. European Centre for Medium-Range Weather Forecasts Reading, UK. Bright, J.M., Gueymard, C.A., 2019. Climate-specific and global validation of MODIS Aqua and Terra aerosol optical depth at 452 AERONET stations. Sol. Energy 183, 594–605. Che, H., Gui, K., Xia, X., Wang, Y., Holben, B.N., Goloub, P., Cuevas-Agull´ o, E., Wang, H., Zheng, Y., Zhao, H., 2019. Large contribution of meteorological factors to interdecadal changes in regional aerosol optical depth. Atmos. Chem. Phys. 19, 10497–10523. Chen, X., de Leeuw, G., Arola, A., Liu, S., Liu, Y., Li, Z., Zhang, K., 2020. Joint retrieval of the aerosol fine mode fraction and optical depth using MODIS spectral reflectance over northern and eastern China: artificial neural network method. Remote Sens. Environ. 249, 112006. Chen, Y., Zhao, C., Ming, Y., 2019b. Potential impacts of Arctic warming on Northern Hemisphere mid-latitude aerosol optical depth. Clim. Dyn. 53, 1637–1651. https:// doi.org/10.1007/s00382-019-04706-3. Chen, Z.-Y., Zhang, T.-H., Zhang, R., Zhu, Z.-M., Yang, J., Chen, P.-Y., Ou, C.-Q., Guo, Y., 2019a. Extreme gradient boosting model to estimate PM2. 5 concentrations with missing-filled satellite data in China. Atmos. Environ. https://doi.org/10.1016/j. atmosenv.2019.01.027. Chen, Zhaoyue, M´ endez, R., Petetin, H., Lacima, A., García-Pando, C.P., Ballester, J., 2023. A Pan-European, Quantile Machine Learning (QML) Based, Total, Fine-mode and Coarse-mode Aerosol Optical Depth Dataset (QML AOD). https://doi.org/ 10.5281/ZENODO.7756570. Christopher, S., Gupta, P., 2020. Global distribution of column satellite aerosol optical depth to surface PM2.5 relationships. Remote Sens. 12, 1–13. https://doi.org/ 10.3390/rs12121985. Chung, C.E., Ramanathan, V., Kim, D., Podgorny, I.A., 2005. Global anthropogenic aerosol direct forcing derived from satellite and ground-based observations. J. Geophys. Res. Atmos. 110. Crippa, M., Janssens-Maenhout, G., Dentener, F., Guizzardi, D., Sindelarova, K., Muntean, M., Van Dingenen, R., Granier, C., 2016. Forty years of improvements in European air quality: regional policy-industry interactions with global impacts. Atmos. Chem. Phys. 16, 3825–3841. https://doi.org/10.5194/acp-16-3825-2016. Degenhardt, F., Seifert, S., Szymczak, S., 2019. Evaluation of variable selection methods for random forests and omics data sets. Brief. Bioinform. 20, 492–503. Duarte, R.M.B.O., Duarte, A.C., 2020. Urban atmospheric aerosols: sources, analysis, and effects. Atmosphere (Basel) 11, 1221. Dubovik, O., Li, Z., Mishchenko, M.I., Tanr´ e, D., Karol, Y., Bojkov, B., Cairns, B., Diner, D.J., Espinosa, W.R., Goloub, P., 2019. Polarimetric remote sensing of atmospheric aerosols: instruments, methodologies, results, and perspectives. J. Quant. Spectrosc. Radiat. Transf. 224, 474–511. European Environment Agency, 2021. Health Impacts of Air Pollution in Europe, , 19/ 2021. ed.2021 https://doi.org/10.2800/08097 Ferrero, L., Riccio, A., Ferrini, B.S., D’Angelo, L., Rovelli, G., Casati, M., Angelini, F., Barnaba, F., Gobbi, G.P., Cataldi, M., 2019. Satellite AOD conversion into ground PM10, PM2. 5 and PM1 over the Po valley (Milan, Italy) exploiting information on aerosol vertical profiles, chemistry, hygroscopicity and meteorology. Atmos. Pollut. Res. 10, 1895–1912. Flemming, J., Huijnen, V., Arteta, J., Bechtold, P., Beljaars, A., Blechschmidt, A.-M., Diamantakis, M., Engelen, R.J., Gaudel, A., Inness, A., 2015. Tropospheric chemistry in the Integrated Forecasting System of ECMWF. Geosci. Model Dev. 8, 975–1003. GLOBE, 2010. Summary of Aerosols Protocol [WWW Document]. URL. https://instesre. org/Aerosols/Aerosols_HTML.htm. Griffin, R.J., 2013. The sources and impacts of tropospheric particulate matter. Nat. Educ. Knowl. 4, 1. Gr¨ onholm, T., Launiainen, S., Ahlm, L., Mårtensson, E.M., Kulmala, M., Vesala, T., Nilsson, E.D., 2009. Aerosol particle dry deposition to canopy and forest floor measured by two-layer eddy covariance system. J. Geophys. Res. Atmos. 114. Gueymard, C.A., Yang, D., 2020. Worldwide validation of CAMS and MERRA-2 reanalysis aerosol optical depth products using 15 years of AERONET observations. Atmos. Environ. 225, 117216. Gui, K., Che, H., Wang, Y., Wang, H., Zhang, L., Zhao, H., Zheng, Y., Sun, T., Zhang, X., 2019. Satellite-derived PM2. 5 concentration trends over Eastern China from 1998 to 2016: relationships to emissions and meteorological parameters. Environ. Pollut. 247, 1125–1133. Guo, B., Wang, Z., Pei, L., Zhu, X., Chen, Q., Wu, H., Zhang, W., Zhang, D., 2023. Reconstructing MODIS aerosol optical depth and exploring dynamic and influential factors of AOD via random forest at the global scale. Atmos. Environ. 315, 120159 https://doi.org/10.1016/j.atmosenv.2023.120159. Gupta, P., Levy, R.C., Mattoo, S., Remer, L.A., Munchak, L.A., 2016. A surface reflectance scheme for retrieving aerosol optical depth over urban surfaces in MODIS Dark Target retrieval algorithm. Atmos. Meas. Tech. 9, 3293–3308. Gupta, P., Remer, L.A., Patadia, F., Levy, R.C., Christopher, S.A., 2020. High-resolution gridded level 3 aerosol optical depth data from MODIS. Remote Sens. 12, 2847. He, Q., Wang, M., Hung, S., Yim, L., 2021. The Spatiotemporal Relationship Between PM 2. 5 and AOD in China : Influencing Factors and Implications for Satellite PM 2.5 Estimations by MAIAC AOD.. Holben, B.N., Eck, T.F., Slutsker, I., Smirnov, A., Sinyuk, A., Schafer, J., Giles, D., Dubovik, O., 2006. AERONET’s version 2.0 quality assurance criteria. In: Remote Sensing of the Atmosphere and Clouds, 6408, pp. 134–147. https://doi.org/ 10.1117/12.706524. Huang, C.W., Lin, M.Y., Khlystov, A., Katul, G.G., 2015. The effects of leaf size and microroughness on the branch-scale collection efficiency of ultrafine particles. J. Geophys. Res. Atmos. 120, 3370–3385. Huang, L., Liu, S., Yang, Z., Xing, J., Zhang, J., Bian, J., Li, S., Sahu, S.K., Wang, S., Liu, T.-Y., 2021. Exploring deep learning for air pollutant emission estimation. Geosci. Model Dev. 14, 4641–4654. Huang, Y., Chameides, W.L., Dickinson, R.E., 2007. Direct and indirect effects of anthropogenic aerosols on regional precipitation over east Asia. J. Geophys. Res. Atmos. 112. Inness, A., Ades, M., Agustí-Panareda, A., Barr´ e, J., Benedictow, A., Blechschmidt, A.-M., Dominguez, J.J., Engelen, R., Eskes, H., Flemming, J., 2019. The CAMS reanalysis of atmospheric composition. Atmos. Chem. Phys. 19, 3515–3556. Institute for Health Metrics, 2020. State of Global Air 2020 (Special Report). Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W., Ye, Q., Liu, T.-Y., 2017. Lightgbm: a highly efficient gradient boosting decision tree. Adv. Neural Inf. Proces. Syst. 30. Kulmala, M., Aurela, M., Carbone, S., Saarnio, K., Frey, A., Saarikoski, S., Teinil¨ a, K., Hillamo, R., 2014. Seasonal and diurnal changes in inorganic ions, carbonaceous matter and mass in ambient aerosol particles in an urban, background area. Boreal Environ. Res. 19, 71–86. Kursa, M.B., Rudnicki, W.R., 2010. Feature selection with the Boruta package. J. Stat. Softw. 36, 1–13. Levy, R.C., Remer, L.A., Kleidman, R.G., Mattoo, S., Ichoku, C., Kahn, R., Eck, T.F., 2010. Global evaluation of the Collection 5 MODIS dark-target aerosol products over land. Atmos. Chem. Phys. 10, 10399–10420. Levy, R.C., Mattoo, S., Munchak, L.A., Remer, L.A., Sayer, A.M., Patadia, F., Hsu, N.C., 2013. The Collection 6 MODIS aerosol products over land and ocean. Atmos. Meas. Tech. 6, 2989–3034. https://doi.org/10.5194/amt-6-2989-2013. Lin, C., Lau, A.K.H., Fung, J.C.H., Lao, X.Q., Li, Y., Li, C., 2019. Assessing the effect of the long-term variations in aerosol characteristics on satellite remote sensing of PM2. 5 using an observation-based model. Environ. Sci. Technol. 53, 2990–3000. Lin, X., Chamecki, M., Katul, G., Yu, X., 2018. Effects of leaf area index and density on ultrafine particle deposition onto forest canopies: a LES study. Atmos. Environ. 189, 153–163. Lyapustin, A., Wang, Y., Laszlo, I., Kahn, R., Korkin, S., Remer, L., Levy, R., Reid, J.S., 2011. Multiangle implementation of atmospheric correction (MAIAC): 2. Aerosol algorithm. J. Geophys. Res. 116, D3. Lyapustin, A., Wang, Y., Korkin, S., Huang, D., 2018. MODIS Collection 6 MAIAC algorithm. Atmos. Meas. Tech. 11. Maag, B., Zhou, Z., Thiele, L., 2018. A survey on sensor calibration in air pollution monitoring deployments. IEEE Internet Things J. 5, 4857–4870. Martins, N.R., da Graça, G.C., 2017. Simulation of the effect of fine particle pollution on the potential for natural ventilation of non-domestic buildings in European cities. Build. Environ. 115, 236–250. Meloni, D., Di Sarra, A., Monteleone, F., Pace, G., Piacentino, S., Sferlazzo, D.M., 2008. Seasonal transport patterns of intense Saharan dust events at the Mediterranean island of Lampedusa. Atmos. Res. 88, 134–148. NASA, 2023. [WWW document]. https://ladsweb.modaps.eosdis.nasa.gov/. O’neill, N.T., Eck, T.F., Smirnov, A., Holben, B.N., Thulasiraman, S., 2003. Spectral discrimination of coarse and fine mode optical depth. J. Geophys. Res. Atmos. 108. Papayannis, A., Amiridis, V., Mona, L., Tsaknakis, G., Balis, D., B¨ osenberg, J., Chaikovski, A., De Tomasi, F., Grigorov, I., Mattis, I., 2008. Systematic lidar observations of Saharan dust over Europe in the frame of EARLINET (2000–2002). J. Geophys. Res. Atmos. 113. Prospero, J.M., Collard, F., Molini´ e, J., Jeannot, A., 2014. Characterizing the annual cycle of African dust transport to the Caribbean Basin and South America and its impact on the environment and air quality. Glob. Biogeochem. Cycles 28, 757–773. Z.-Y. Chen et al. Science of the Total Environment 918 (2024) 170593 18 Randles, C.A., Da Silva, A.M., Buchard, V., Colarco, P.R., Darmenov, A., Govindaraju, R., Smirnov, A., Holben, B., Ferrare, R., Hair, J., 2017. The MERRA-2 aerosol reanalysis, 1980 onward. Part I: system description and data assimilation evaluation. J. Clim. 30, 6823–6850. Seinfeld, J.H., Pandis, S.N., 1998. From air pollution to climate change. Atmos. Chem. Phys. 1326. Tai, A.P.K., Mickley, L.J., Jacob, D.J., 2010. Correlations between fine particulate matter (PM2. 5) and meteorological variables in the United States: implications for the sensitivity of PM2. 5 to climate change. Atmos. Environ. 44, 3976–3984. Wei, J., Sun, L., Huang, B., Bilal, M., Zhang, Z., Wang, L., 2018. Verification, improvement and application of aerosol optical depths in China part 1: intercomparison of NPP-VIIRS and Aqua-MODIS. Atmos. Environ. 175, 221–233. Xiao, Q., Zhang, H., Choi, M., Li, S., Kondragunta, S., Kim, J., Holben, B., Levy, R.C., Liu, Y., 2016. Evaluation of VIIRS, GOCI, and MODIS Collection 6 AOD retrievals against ground sunphotometer observations over East Asia. Atmos. Chem. Phys. 16, 1255–1269. Yan, X., Shi, W., Li, Zhanqing, Li, Zhengqiang, Luo, N., Zhao, W., Wang, H., Yu, X., 2017. Satellite-based PM2. 5 estimation using fine-mode aerosol optical thickness over China. Atmos. Environ. 170, 290–302. Yan, X., Zang, Z., Zhao, C., Husi, L., 2021. Understanding global changes in fine-mode aerosols during 2008–2017 using statistical methods and deep learning approach. Environ. Int. 149, 106392 https://doi.org/10.1016/j.envint.2021.106392. Yan, X., Zang, Z., Li, Z., Luo, N., Zuo, C., Jiang, Y., Li, D., Guo, Y., Zhao, W., Shi, W., Cribb, M., 2022. A global land aerosol fine-mode fraction dataset (2001-2020) retrieved from MODIS using hybrid physical and deep learning approaches. Earth Syst. Sci. Data 14, 1193–1213. https://doi.org/10.5194/essd-14-1193-2022. You, W., Zang, Z., Zhang, L., Li, Z., Chen, D., Zhang, G., 2015. Estimating ground-level PM10 concentration in northwestern China using geographically weighted regression based on satellite AOD combined with CALIPSO and MODIS fire count. Remote Sens. Environ. 168, 276–285. Zang, Z., Li, D., Guo, Y., Shi, W., Yan, X., 2021. Superior PM2. 5 estimation by integrating aerosol fine mode data from the Himawari-8 satellite in deep and classical machine learning models. Remote Sens. 13, 2779. Zhang, Ying, Li, Z., Bai, K., Wei, Y., Xie, Y., Zhang, Yuanxun, Ou, Y., Cohen, J., Zhang, Yuhuan, Peng, Z., 2021. Satellite remote sensing of atmospheric particulate matter mass concentration: advances, challenges, and perspectives. Fundam. Res. 1, 240–258. Zhao, B., Jiang, J.H., Diner, D.J., Su, H., Gu, Y., Liou, K.N., Jiang, Z., Huang, L., Takano, Y., Fan, X., Omar, A.H., 2018. Intra-annual variations of regional aerosol optical depth, vertical distribution, and particle types from multiple satellite and ground-based observational datasets. Atmos. Chem. Phys. 18, 11247–11260. https://doi.org/10.5194/acp-18-11247-2018. Zheng, C., Zhao, C., Zhu, Y., Wang, Y., Shi, X., Wu, X., Chen, T., Wu, F., Qiu, Y., 2017. Analysis of influential factors for the relationship between PM$_{2.5}$ and AOD in Beijing. Atmos. Chem. Phys. 17, 13473–13489. https://doi.org/10.5194/acp-1713473-2017. Zhou, Y., Savij¨ arvi, H., 2014. The effect of aerosols on long wave radiation and global warming. Atmos. Res. 135, 102–111. Z.-Y. Chen et al.