A Bayesian ensemble approach to combine PM2.5 estimates from statistical models using satellite imagery and numerical model simulation

Spatiotemporal empirical analysis of particulate matter PM2.5 pollution and air quality index (AQI) trends in Africa using MERRA-2 reanalysis datasets (1980-2021)

In this study, the spatial-temporal trends of PM2.5 pollution were analyzed for subregions in Africa and the entire continent from 1980 to 2021. The distributions and trends of PM2.5 were derived from the monthly concentrations of the aerosol species from MERRA-2 reanalysis datasets comprising of sulphates (SO4), organic carbon (OC), black carbon (BC), Dust2.5 and Sea Salt (SS2.5). The resulting PM2.5 trends were compared with the climate factors, socio-economic indicators, and terrain characteristics. Using the Mann-Kendall (M-K) test, the continent and its subregions showed positive trends in PM2.5 concentrations, except for western and central Africa which exhibited marginal negative trends. The M-K trends also determined Dust2.5 as the dominant contributing aerosol factor responsible for the high PM2.5 concentrations in the northern, western and central regions of Africa, while SO4 and OC were respectively the most significant contributors to PM2.5 in the eastern and southern Africa regions. For the climate factors, the PM2.5 trends were determined to be positively correlated with the wind speed trends, while precipitation and temperature trends exhibited low and sometimes negative correlations with PM2.5. Socio-economically, highly populated, and bare/sparse vegetated areas showed higher PM2.5 concentrations, while vegetated areas tended to have lower PM2.5 concentrations. Topographically, low laying regions were observed to retain the deposited PM2.5 especially in the northern and western regions of Africa. The Air Quality Index (AQI) results showed that 94 % of the continent had an average PM2.5 of 12-35 μg/m3 hence classified as "Moderate" AQI, and the rest of the continent's PM2.5 levels was between 35 and 55 μg/m3 implying AQI classification of "Unhealthy for Sensitive People". Northern and western Africa regions had the highest AQI, while southern Africa had the lowest AQI. The approach and findings in this study can be used to complement the evaluation and management of air quality in Africa.

IoT and Satellite Sensor Data Integration for Assessment of Environmental Variables: A Case Study on NO2

This paper introduces a novel approach to increase the spatiotemporal resolution of an arbitrary environmental variable. This is achieved by utilizing machine learning algorithms to construct a satellite-like image at any given time moment, based on the measurements from IoT sensors. The target variables are calculated by an ensemble of regression models. The observed area is gridded, and partitioned into Voronoi cells based on the IoT sensors, whose measurements are available at the considered time. The pixels in each cell have a separate regression model, and take into account the measurements of the central and neighboring IoT sensors. The proposed approach was used to assess NO2 data, which were obtained from the Sentinel-5 Precursor satellite and IoT ground sensors. The approach was tested with three different machine learning algorithms: 1-nearest neighbor, linear regression and a feed-forward neural network. The highest accuracy yield was from the prediction models built with the feed-forward neural network, with an RMSE of 15.49 ×10-6 mol/m2.

An Estimation Method for PM2.5 Based on Aerosol Optical Depth Obtained from Remote Sensing Image Processing and Meteorological Factors

Understanding the spatiotemporal variations in the mass concentrations of particulate matter ≤2.5 µm (PM2.5) in size is important for controlling environmental pollution. Currently, ground measurement points of PM2.5 in China are relatively discrete, thereby limiting spatial coverage. Aerosol optical depth (AOD) data obtained from satellite remote sensing provide insights into spatiotemporal distributions for regional pollution sources. In this study, data from the Multi-Angle Implementation of Atmospheric Correction (MAIAC) AOD (1 km resolution) product from Moderate Resolution Imaging Spectroradiometer (MODIS) and hourly PM2.5 concentration ground measurements from 2015 to 2020 in Dalian, China were used. Although trends in PM2.5 and AOD were consistent over time, there were seasonal differences. Spatial distributions of AOD and PM2.5 were consistent (R2 = 0.922), with higher PM2.5 values in industrial areas. The method of cross-dividing the test set by year was adopted, with AOD and meteorological factors as the input variable and PM2.5 as the output variable. A backpropagation neural network (BPNN) model of joint cross-validation was established; the stability of the model was evaluated. The trend in the predicted values of BPNN was consistent with the monitored values; the estimation result of the BPNN with the introduction of meteorological factors is better; coefficient of determination (R2) and RMSE standard deviation (SD) between the predicted values and the monitored values in the test set were 0.663–0.752 and 0.01–0.05 μg/m3, respectively. The BPNN was simpler and the training time was shorter compared with those of a regression model and support vector regression (SVR). This study demonstrated that BPNN could be effectively applied to the MAIAC AOD data to estimate PM2.5 concentrations.

Application of geostationary satellite and high-resolution meteorology data in estimating hourly PM2.5 levels during the Camp Fire episode in California

Wildland fire smoke contains large amounts of PM2.5 that can traverse tens to hundreds of kilometers, resulting in significant deterioration of air quality and excess mortality and morbidity in downwind regions. Estimating PM2.5 levels while considering the impact of wildfire smoke has been challenging due to the lack of ground monitoring coverage near the smoke plumes. We aim to estimate total PM2.5 concentration during the Camp Fire episode, the deadliest wildland fire in California history. Our random forest (RF) model combines calibrated low-cost sensor data (PurpleAir) with regulatory monitor measurements (Air Quality System, AQS) to bolster ground observations, Geostationary Operational Environmental Satellite-16 (GOES-16)'s high temporal resolution to achieve hourly predictions, and oversampling techniques (Synthetic Minority Oversampling Technique, SMOTE) to reduce model underestimation at high PM2.5 levels. In addition, meteorological fields at 3 km resolution from the High-Resolution Rapid Refresh model and land use variables were also included in the model. Our AQS-only model achieved an out of bag (OOB) R2 (RMSE) of 0.84 (12.00 μg/m3) and spatial and temporal cross-validation (CV) R2 (RMSE) of 0.74 (16.28 μg/m3) and 0.73 (16.58 μg/m3), respectively. Our AQS + Weighted PurpleAir Model achieved OOB R2 (RMSE) of 0.86 (9.52 μg/m3) and spatial and temporal CV R2 (RMSE) of 0.75 (14.93 μg/m3) and 0.79 (11.89 μg/m3), respectively. Our AQS + Weighted PurpleAir + SMOTE Model achieved OOB R2 (RMSE) of 0.92 (10.44 μg/m3) and spatial and temporal CV R2 (RMSE) of 0.84 (12.36 μg/m3) and 0.85 (14.88 μg/m3), respectively. Hourly predictions from our model may aid in epidemiological investigations of intense and acute exposure to PM2.5 during the Camp Fire episode.

Predicting spatiotemporally-resolved mean air temperature over Sweden from satellite data using an ensemble model

Mapping of air temperature (Ta) at high spatiotemporal resolution is critical to reducing exposure assessment errors in epidemiological studies on the health effects of air temperature. In this study, we applied a three-stage ensemble model to estimate daily mean Ta from satellite-based land surface temperature (Ts) over Sweden during 2001-2019 at a high spatial resolution of 1 × 1 km². The ensemble model incorporated four base models, including a generalized additive model (GAM), a generalized additive mixed model (GAMM), and two machine learning models (random forest [RF] and extreme gradient boosting [XGBoost]), and allowed the weights for each model to vary over space, with the best-performing model for each grid cell assigned the highest weight. Various spatial predictors were included as adjustment variables in all the base models, including land cover type, normalized difference vegetation index (NDVI), and elevation. The ensemble model showed high performance with an overall R² of 0.98 and a root mean square error of 1.38 °C in the ten-fold cross-validation, and outperformed each of the four base models. Although each base model performed well, the two machine learning models (RF [R² = 0.97], XGBoost [R² = 0.98]) had better performance than the two regression models (GAM [R² = 0.95], GAMM [R² = 0.96]). In the machine learning models, Ts was the dominant predictor of Ta, followed by day of year, NDVI, latitude, elevation, and longitude. The highly spatiotemporally-resolved Ta can improve temperature exposure assessment in future epidemiological studies.

Evaluating the Utility of High-Resolution Spatiotemporal Air Pollution Data in Estimating Local PM2.5 Exposures in California from 2015–2018

Air quality management is increasingly focused not only on across-the-board reductions in ambient pollution concentrations but also on identifying and remediating elevated exposures that often occur in traditionally disadvantaged communities. Remote sensing of ambient air pollution using data derived from satellites has the potential to better inform management decisions that address environmental disparities by providing increased spatial coverage, at high-spatial resolutions, compared to air pollution exposure estimates based on ground-based monitors alone. Daily PM2.5 estimates for 2015–2018 were estimated at a 1 km2 resolution, derived from NASA’s Moderate Resolution Imaging Spectroradiometer (MODIS) satellite instrument and the Multi-Angle Implementation of Atmospheric Correction (MAIAC) algorithm in order to assess the utility of highly refined spatiotemporal air pollution data in 92 California cities and in the 13 communities included in the California Community Air Protection Program. The identification of pollution hot-spots within a city is typically not possible relying solely on the regulatory monitoring networks; however, day-to-day temporal variability was shown to be generally well represented by nearby ground-based monitoring data even in communities with strong spatial gradients in pollutant concentrations. An assessment of within-ZIP Code variability in pollution estimates indicates that high-resolution pollution estimates (i.e., 1 km2) are not always needed to identify spatial differences in exposure but become increasingly important for larger geographic areas (approximately 50 km2). Taken together, these findings can help inform strategies for use of remote sensing data for air quality management including the screening of locations with air pollution exposures that are not well represented by existing ground-based air pollution monitors.

Estimating monthly PM2.5 concentrations from satellite remote sensing data, meteorological variables, and land use data using ensemble statistical modeling and a random forest approach

Fine particulate matter (PM_2.5) is associated with various adverse health outcomes and poses serious concerns for public health. However, ground monitoring stations for PM_2.5 measurements are mostly installed in population-dense or urban areas. Thus, satellite retrieved aerosol optical depth (AOD) data, which provide spatial and temporal surrogates of exposure, have become an important tool for PM_2.5 estimates in a study area. In this study, we used AOD estimates of surface PM_2.5 together with meteorological and land use variables to estimate monthly PM_2.5 concentrations at a spatial resolution of 3 km² over Taiwan Island from 2015 to 2019. An ensemble two-stage estimation procedure was proposed, with a generalized additive model (GAM) for temporal-trend removal in the first stage and a random forest model used to assess residual spatiotemporal variations in the second stage. We obtained a model-fitting R² of 0.98 with a root mean square error (RMSE) of 1.40 μg/m³. The leave-one-out cross-validation (LOOCV) R² with seasonal stratification was 0.82, and the RMSE was 3.85 μg/m³, whereas the R² and RMSE obtained by using the pure random forest approach produced R² and RMSE values of 0.74 and 4.60 μg/m³, respectively. The results indicated that the ensemble modeling approach had a higher predictive ability than the pure machine learning method and could provide reliable PM_2.5 estimates over the entire island, which has complex terrain in terms of land use and topography.

Sub-City Scale Hourly Air Quality Forecasting by Combining Models, Satellite Observations, and Ground Measurements

While multiple information sources exist concerning surface-level air pollution, no individual source simultaneously provides large-scale spatial coverage, fine spatial and temporal resolution, and high accuracy. It is, therefore, necessary to integrate multiple data sources, using the strengths of each source to compensate for the weaknesses of others. In this study, we propose a method incorporating outputs of NASA's GEOS Composition Forecasting model system with satellite information from the TROPOMI instrument and ground measurement data on surface concentrations. Although we use ground monitoring data from the Environmental Protection Agency network in the continental United States, the model and satellite data sources used have the potential to allow for global application. This method is demonstrated using surface measurements of nitrogen dioxide as a test case in regions surrounding five major US cities. The proposed method is assessed through cross-validation against withheld ground monitoring sites. In these assessments, the proposed method demonstrates major improvements over two baseline approaches which use ground-based measurements only. Results also indicate the potential for near-term updating of forecasts based on recent ground measurements.

Imputing Satellite-Derived Aerosol Optical Depth Using a Multi-Resolution Spatial Model and Random Forest for PM2.5 Prediction

A task for environmental health research is to produce complete pollution exposure maps despite limited monitoring data. Satellite-derived aerosol optical depth (AOD) is frequently used as a predictor in various models to improve PM2.5 estimation, despite significant gaps in coverage. We analyze PM2.5 and AOD from July 2011 in the contiguous United States. We examine two methods to aid in gap-filling AOD: (1) lattice kriging, a spatial statistical method adapted to handle large amounts data, and (2) random forest, a tree-based machine learning method. First, we evaluate each model’s performance in the spatial prediction of AOD, and we additionally consider ensemble methods for combining the predictors. In order to accurately assess the predictive performance of these methods, we construct spatially clustered holdouts to mimic the observed patterns of missing data. Finally, we assess whether gap-filling AOD through one of the proposed ensemble methods can improve prediction of PM2.5 in a random forest model. Our results suggest that ensemble methods of combining lattice kriging and random forest can improve AOD gap-filling. Based on summary metrics of performance, PM2.5 predictions based on random forest models were largely similar regardless of the inclusion of gap-filled AOD, but there was some variability in daily model predictions.

The long-term trend of PM2.5-related mortality in China: The effects of source data selection

Quantification of PM_2.5 exposure and associated mortality is critical to inform policy making. Previous studies estimated varying PM_2.5-related mortality in China due to the usage of different source data, but rarely justify the data selection. To quantify the sensitivity of mortality assessment to source data, we first constructed state-of-the-art PM_2.5 predictions during 2000-2018 at a 1-km resolution with an ensemble machine learning model that filled missing data explicitly. We also calibrated and fused various gridded population data with a geostatistical method. Then we assessed the PM_2.5-related mortality with various PM_2.5 predictions, population distributions, exposure-response functions, and baseline mortalities. We found that in addition to the well documented uncertainties in the exposure-response functions, missingness in PM_2.5 prediction, PM_2.5 prediction error, and prediction error in population distribution resulted to a 40.5%, 25.2% and 15.9% lower mortality assessment compared to the mortality assessed with the best-performed source data, respectively. With the best-performed source data, we estimated a total of approximately 25 million PM_2.5-related mortality during 2001-2017 in China. From 2001 to 2017, The PM_2.5 variations, growth and aging of population, decrease in baseline mortality led to a 7.8% increase, a 42.0% increase and a 24.6% decrease in PM_2.5-related mortality, separately. We showed that with the strict clean air policies implemented in 2013, the population-weighted PM_2.5 concentration decreased remarkably at an annual rate of 4.5 μg/m³, leading to a decrease of 179 thousand PM_2.5-related deaths nationwide during 2013-2017. The mortality decrease due to PM_2.5 reduction was offset by the population growth and aging population.

Development of Europe-Wide Models for Particle Elemental Composition Using Supervised Linear Regression and Random Forest

We developed Europe-wide models of long-term exposure to eight elements (copper, iron, potassium, nickel, sulfur, silicon, vanadium, and zinc) in particulate matter with diameter <2.5 μm (PM_2.5) using standardized measurements for one-year periods between October 2008 and April 2011 in 19 study areas across Europe, with supervised linear regression (SLR) and random forest (RF) algorithms. Potential predictor variables were obtained from satellites, chemical transport models, land-use, traffic, and industrial point source databases to represent different sources. Overall model performance across Europe was moderate to good for all elements with hold-out-validation R-squared ranging from 0.41 to 0.90. RF consistently outperformed SLR. Models explained within-area variation much less than the overall variation, with similar performance for RF and SLR. Maps proved a useful additional model evaluation tool. Models differed substantially between elements regarding major predictor variables, broadly reflecting known sources. Agreement between the two algorithm predictions was generally high at the overall European level and varied substantially at the national level. Applying the two models in epidemiological studies could lead to different associations with health. If both between- and within-area exposure variability are exploited, RF may be preferred. If only within-area variability is used, both methods should be interpreted equally.

Application of Bayesian Additive Regression Trees for Estimating Daily Concentrations of PM2.5 Components

Bayesian additive regression tree (BART) is a recent statistical method that combines ensemble learning and nonparametric regression. BART is constructed under a probabilistic framework that also allows for model-based prediction uncertainty quantification. We evaluated the application of BART in predicting daily concentrations of four fine particulate matter (PM_2.5) components (elemental carbon, organic carbon, nitrate, and sulfate) in California during the period 2005 to 2014. We demonstrate in this paper how BART can be tuned to optimize prediction performance and how to evaluate variable importance. Our BART models included, as predictors, a large suite of land-use variables, meteorological conditions, satellite-derived aerosol optical depth parameters, and simulations from a chemical transport model. In cross-validation experiments, BART demonstrated good out-of-sample prediction performance at monitoring locations (R² from 0.62 to 0.73). More importantly, prediction intervals associated with concentration estimates from BART showed good coverage probability at locations with and without monitoring data. In our case study, major PM_2.5 components could be estimated with good accuracy, especially when collocated PM_2.5 total mass observations were available. In conclusion, BART is an attractive approach for modeling ambient air pollution levels, especially for its ability to provide uncertainty in estimates that may be useful for subsequent health impact and health effect analyses.

A multi-city air pollution population exposure study: Combined use of chemical-transport and random-Forest models with dynamic population data

Cities are severely affected by air pollution. Local emissions and urban structures can produce large spatial heterogeneities. We aim to improve the estimation of NO₂, O₃, PM_2.5 and PM₁₀ concentrations in 6 Italian metropolitan areas, using chemical-transport and machine learning models, and to assess the effect on population exposure by using information on urban population mobility. Three years (2013-2015) of simulations were performed by the Chemical-Transport Model (CTM) FARM, at 1 km resolution, fed by boundary conditions provided by national-scale simulations, local emission inventories and meteorological fields. A downscaling of daily air pollutants at higher resolution (200 m) was then carried out by means of a machine learning Random-Forest (RF) model, considering CTM and spatial-temporal predictors, such as population, land-use, surface greenness and vehicular traffic, as input. RF achieved mean cross-validation (CV) R² of 0.59, 0.72, 0.76 and 0.75 for NO₂, PM₁₀, PM_2.5 and O₃, respectively, improving results from CTM alone. Mean concentration fields exhibited clear geographical gradients caused by climate conditions, local emission sources and photochemical processes. Time series of population weighted exposure (PWE) were estimated for two months of the year 2015 and for five cities, by combining population mobility data (derived from mobile phone traffic volumes data), and concentration levels from the RF model. PWE_RF metric better approximated the observed concentrations compared with the predictions from either CTM alone or CTM and RF combined, especially for pollutants exhibiting strong spatial gradients, such as NO₂. 50% of the population was estimated to be exposed to NO₂ concentrations between 12 and 38 μg/m³ and PM₁₀ between 20 and 35 μg/m³. This work supports the potential of machine learning methods in predicting air pollutant levels in urban areas at high spatial and temporal resolutions.

Fine particulate matter pollution in North China: Seasonal-spatial variations, source apportionment, sector and regional transport contributions

Large areas of mainland China have been suffering frequently from heavy haze pollution during the past years, which feature high concentrations of fine particulate matter (PM_2.5, particulate matters with aerodynamic diameters less than 2.5 μm) and low visibility. Moreover, these areas manifested strong regional complex pollution characteristics, particularly in North China including Beijing and the five surrounding provinces (BSFP). In this study, by using the localized comprehensive emission inventory of BSFP region in 2012 as an input, the Comprehensive Air Quality Model with Extensions-Particulate Matter Source Apportionment Technology (CAMx/PSAT) was used to assess the seasonal variations and source apportionment of PM_2.5 in the highly polluted BSFP region, with a specific focus on the sectoral and sub-regional contributions to PM_2.5 in Beijing during winter and summer. Results showed that the PM_2.5 concentrations of BSFP region was higher in winter than that in summer. And the heavily polluted area in BSFP region shrinked noticeably in summer, compared with winter. As for source apportionment of PM_2.5, residential and remaining industrial sectors constituted the top two sources of PM_2.5 mass concentrations in Beijing. In addition, agricultural source represented a major contributor to ammonium, whereas transportation and power sectors constituted major sources to nitrates. In terms of contributions from sub-regions, the local sources ranked as the dominant contributors to PM_2.5 in Beijing, while the main external contributions originated from the surrounding areas, such as Hebei and Shandong. Results of daily source apportionment to PM_2.5 in Beijing showed that sub-regional long-distance transport became stronger when haze pollution was severe, in which contribution from remaining industrial sector would be higher than that of other periods. The results will allow for an improved understanding of the causes and origins of heavy regional PM_2.5 pollution, and thus will benefit the development of effective joint control policies and identification of key polluting emission categories in North China and ultimately serve as references for other highly polluted megacities in the world.