A machine learning model to estimate ambient PM_2.5 concentrations in industrialized highveld region of South Africa

Highlights

• We developed a random forest model to estimate daily PM_2.5 concentrations at 1 km² resolution in South Africa.

• Our model captured seasonal trends and spatial patterns of PM_2.5 with relatively high accuracy.

• High PM_2.5 levels were identified in low-income settlements and industrial areas in western Mpumalanga.

• PM2.5 levels decreased in north of Gauteng province after the implementation of new air quality standard.

Abstract

Exposure to fine particulate matter (PM_2.5) has been linked to a substantial disease burden globally, yet little has been done to estimate the population health risks of PM_2.5 in South Africa due to the lack of high-resolution PM_2.5 exposure estimates. We developed a random forest model to estimate daily PM_2.5 concentrations at 1 km² resolution in and around industrialized Gauteng Province, South Africa, by combining satellite aerosol optical depth (AOD), meteorology, land use, and socioeconomic data. We then compared PM_2.5 concentrations in the study domain before and after the implementation of the new national air quality standards. We aimed to test whether machine learning models are suitable for regions with sparse ground observations such as South Africa and which predictors played important roles in PM_2.5 modeling. The cross-validation R² and Root Mean Square Error of our model was 0.80 and 9.40 μg/m³, respectively. Satellite AOD, seasonal indicator, total precipitation, and population were among the most important predictors. Model-estimated PM_2.5 levels successfully captured the temporal pattern recorded by ground observations. Spatially, the highest annual PM_2.5 concentration appeared in central and northern Gauteng, including northern Johannesburg and the city of Tshwane. Since the 2016 changes in national PM_2.5 standards, PM_2.5 concentrations have decreased in most of our study region, although levels in Johannesburg and its surrounding areas have remained relatively constant. This is anadvanced PM_2.5 model for South Africa with high prediction accuracy at the daily level and at a relatively high spatial resolution. Our study provided a reference for predictor selection, and our results can be used for a variety of purposes, including epidemiological research, burden of disease assessments, and policy evaluation.

Introduction

Fine particulate matter (PM_2.5, airborne particles with an aerodynamic diameter of less than 2.5 μm) is a ubiquitous air pollutant that harms human health and wellbeing. Numerous epidemiological studies of both short- and long-term exposure have reported strong associations with adverse health outcomes, including premature mortality and morbidity from a range of diseases (Atkinson et al., 2014; Burnett et al., 2018; Liu et al., 2019). Populations living in developing regions, in particular, are often exposed to levels of particulate matter greatly exceeding WHO standards, and an estimated 4.2 million premature death worldwide are attributable to ambient PM_2.5 (Cohen et al., 2017; World Health Organization, 2016).

In addition to natural sources such as biomass burning, dust storms, and ocean spray, PM_2.5 and its precursors are emitted from several anthropogenic sources, including industrial activities, power generation, vehicle traffic, agriculture burning, and household fuel use (Tucker, 2000). The diverse emission sources and secondary production that occurs in the atmosphere results in complex distributions of PM_2.5 in space and time (Seinfeld and Pandis, 2016). Current air quality monitoring networks are often insufficient to quantify PM_2.5 exposure and health risk at the local level even in high-income countries. Routine monitoring is even more sparse or nonexistent in many low- and middle-income countries (Brauer et al., 2016), however satellite data can be used to fill this gap.

The past decade has seen the increasing application of satellite remote sensing products such as aerosol optical depth (AOD) to estimate surface PM_2.5 concentrations. AOD measures the light extinction of aerosol particles at a given wavelength as it passes through the atmospheric column. Although AOD is often strongly correlated with surface PM_2.5 concentration, this relationship is nonlinear and modified by various factors such as meteorology, particle vertical distribution, and particle chemical composition (Hoff and Christopher, 2009; Liu et al., 2005; Sorek-Hamer et al., 2020). Over the past two decades, a number of statistical models have been proposed to capture these relationships at different spatial and temporal scales in order to improve prediction accuracy and robustness, including linear mixed-effects models (Ma et al., 2016b), geographically weighted regression (GWR) (Ma et al., 2014), generalized additive models (Strawa et al., 2013), Bayesian downscaler (Chang et al., 2013), and multi-stage models (Kloog et al., 2014). Most recently, machine learning models such as random forests (Brokamp et al., 2018; Hu et al., 2017) and neural network (Li et al., 2020) have shown high prediction accuracy. These advanced satellite-driven models are useful tools to fill the data gaps left by sparse ground monitors networks and enable more comprehensive assessments of PM_2.5 exposure and its associated health effects. Random forest model is a good choice for areas where advanced models have never been built to explore the spatiotemporal patterns of PM_2.5 because it not only has high prediction accuracy but also provides guidance for predictor selection which is very helpful for future research.

As part of the 2004 National Environmental Management: Air Quality Act (Act No. 39 of 2004), approximately 130 ground ambient air quality stations have been established or, for those already in existence, incorporated into the national reporting of air quality levels on the South African Air Quality Information System (SAAQIS, http://saaqis.environment.gov.za/). These stations monitor criteria pollutants and precursors including PM₁₀, PM_2.5, carbon monoxide (CO), nitrogen oxides (NOx, NO, and NO₂), ozone (O₃), and sulfur dioxide (SO₂), as well as meteorological factors (Gwaze and Mashele, 2018; South African Air Quality Information System, 2010). Most of these ground stations are in areas with poor air quality, such as low-income settlements that use solid fuel for cooking, heating or lighting (i.e., domestic burning activities), urban areas, areas near large roads, and industrial areas. These stations are situated within communities to assist with the assessment of population exposure to air pollution. The ambient monitoring stations in South Africa are limited in spatial coverage, data availability, and data quality. Only 20 ground stations had available PM_2.5 data in our study domain. Hourly raw measurements were only available for 47% of the modeling days in our study period. After quality control, this percentage decreased to 40%.

The South African government promulgated National Ambient Air Quality Standards (NAAQS) for many criteria pollutants in 2009 (Department of Environmental Affairs, 2009), and in 2012 (Department of Environmental Affairs, 2012a), promulgated PM_2.5 standards. The PM₁₀ and PM_2.5 standards were designed to become more stringent over time. Table S1 displays the PM_2.5 NAAQS with compliance date; ultimately (i.e., starting 1 January 2030), the 24-h standard of 25 μg/m³ will be equivalent to the WHO Air Quality Guideline, and the annual standard of 15 μg/m³ would be equivalent to the WHO Interim Target 3 (World Health Organization, 2006).

Ambient air quality in South Africa is impacted by a range of sources including natural sources such as biomass burning, dust, lightning, and biogenic sources, as well as anthropogenic sources such as industry, vehicles, domestic burning, and waste burning (Wright et al., 2017). In South Africa, coal is the dominant energy resource, providing 69% of primary energy needs and more than 90% of electricity (The World Bank, 2015; Zulu et al., 2019). Emissions from other human activities, including industry, mining, mobile vehicle, domestic burning, waste burning, also contribute to PM_2.5 pollution, resulting in a significant public health burden (Katoto et al., 2019; Pacella et al., 2007; Wright et al., 2017).

Using monitoring data and interpolation with the Benefits Mapping and Analysis Program (BenMAP) model, Altieri and Keen (2019) estimated that if the WHO Guidelines for annual average PM_2.5 concentrations were met across South Africa, that 28,000 premature deaths (95th percentile CI 15000–52,000) in South Africa could be avoided, with economic costs of over $29 billion (~4.5% of South African's GDP). A key source of uncertainty in this estimate is from the limited coverage of PM measurements across South Africa.

The South African government declared three national air quality priority areas where ambient air quality does (or is projected to) exceed NAAQS to focus concerted and specific air quality measures to address the poor air quality (Department of Environmental Affairs, 2005). These Priority Areas are the Vaal Triangle Airshed Priority Area (VTAPA) (Department of Environmental Affairs and Tourism, 2006), the Highveld Priority Area (HPA) (Department of Environmental Affairs and Tourism, 2007), and the Waterberg-Bojanala Priority Area (WBPA) (Department of Environmental Affairs, 2012b) (Fig. 1). The Priority Areas contain most of South Africa's large industrial hubs and coal-fired power plants. The Priority Areas also contain portions of the Gauteng Province, where the large mega-city conurbation of Johannesburg-Tshwane-Ekurhuleni is located, the domain of this study.

Ambient air quality levels in the study domain often exceed South African National Ambient Air Quality Standards, with exceedances driven by high PM_2.5, PM_10, and ozone concentrations (Gregor et al., 2019; Kogieluxmie and Venkataraman, 2019; uMoya-NILU, 2017; Venter et al., 2012). A decreasing trend in PM_2.5 concentrations has been found at some monitoring sites in the HPA and the VTAPA, though at the current rate, compliance with current standards will take years at some sites and decades at others. PM concentrations generally peak in the dry winter season, driven by increases in both emissions (e.g., domestic burning, wind-blown dust, biomass burning) and stagnant meteorological conditions (Hersey et al., 2015; Tyson et al., 1988; Xulu et al., 2020). In general, PM concentrations are higher in low-income settlements compared to monitoring sites in industrial and middle-class residential areas (Hersey et al., 2015). Large-scale regional biomass burning impacts this region in the late winter and spring, leading to peak Aerosol Optical Depth (AOD) levels (Archibald et al., 2010; Duncan et al., 2003; Hersey et al., 2015; Queface et al., 2011; Tesfaye et al., 2011).

The ability to accurately estimate exposure to ground-level PM_2.5 is essential to study its adverse health effects and assess the effectiveness of air pollution control measures. In this study, we built a 1 km² resolution daily PM_2.5 concentration model in Gauteng Province and the surrounding areas from 2014 to 2018, based on satellite AOD, meteorological fields, land use data, and socioeconomic variables. This is an advanced model rarely developed in South Africa or elsewhere in Africa. Our goal is to explore the feasibility of developing machine learning models in this region with sparse and incomplete ground measurements and gain insight on important predictors of PM2.5 levels. We also assessed the change in regional PM_2.5 levels before and after the implementation of new national air quality standards.