Constructing a spatiotemporally coherent long-term PM_2.5 concentration dataset over China during 1980–2019 using a machine learning approach

Highlights

• Long-term PM_2.5 is essential due to the coverage deficiency of surface observations.

• A machine learning model with visibility and many auxiliary data inputs is applied.

• A 1-degree gridded daily PM_2.5 dataset over China for 1980–2019 is constructed.

• The model performs well with a high coefficient of determination and low bias.

• It will be a promising tool for assessing related impacts on environment and climate.

Abstract

The lack of long-term observations and satellite retrievals of health-damaging fine particulate matter in China has demanded the estimates of historical PM_2.5 (particulate matter less than 2.5 μm in diameter) concentrations. This study constructs a gridded near-surface PM_2.5 concentration dataset across China covering 1980–2019 using the space-time random forest model with atmospheric visibility observations and other auxiliary data. The modeled daily PM_2.5 concentrations are in excellent agreement with ground measurements, with a coefficient of determination of 0.95 and mean relative error of 12%. Besides the atmospheric visibility which explains 30% of total importance of variables in the model, emissions and meteorological conditions are also key factors affecting PM_2.5 predictions. From 1980 to 2014, the model-predicted PM_2.5 concentrations increased constantly with the maximum growth rate of 5–10 μg/m³/decade over eastern China. Due to the clean air actions, PM_2.5 concentrations have decreased effectively at a rate over 50 μg/m³/decade in the North China Plain and 20–50 μg/m³/decade over many regions of China during 2014–2019. The newly generated dataset of 1-degree gridded PM_2.5 concentrations for the past 40 years across China provides a useful means for investigating interannual and decadal environmental and climate impacts related to aerosols.

Introduction

Particulate matter is one of the major health-damaging components in the atmosphere, especially those with the aerodynamic diameters smaller than 2.5 μm (PM_2.5). Long-term exposure to PM_2.5 can increase risks of many adverse health issues, including respiratory and cardiovascular diseases, lung cancer and premature death (Crouse et al., 2012; Pope et al., 2002; Xing et al., 2016; Zhang et al., 2017). The presence of high concentrations of PM_2.5 also reduces atmospheric visibility, influences the public transportation, and thus adversely affects social and economic activities (Zhang et al., 2014). These aerosol particles also influence climate via aerosol-radiation and aerosol-cloud interactions (Boucher et al., 2013; Yang et al., 2020). Through long-range transport, the local environmental and climatic impacts of aerosols near major source regions can be extended globally (Ren et al., 2020; Wang et al., 2014).

Aerosol concentrations in China are experiencing great changes in recent decades. The rapid industrial development and urbanization were primarily responsible for the increasing tendency of PM_2.5 concentrations before 2010 (Yang et al., 2016; Cohen et al., 2017). From 2013 to 2017, the PM pollution was alleviated, with the national averaged concentration reduced by one third, primarily owing to the implementation of clean air actions in China (Huang et al., 2018). Following the growing public health concern, many air quality monitoring stations have been established to measure real-time PM_2.5 concentrations since 2013. However, the measurements are limited to a short temporal coverage and have uneven spatial distributions (Wang et al., 2019; Zhao et al., 2020), so they are insufficient to describe the long-term characteristics of PM_2.5 in China. Because the spatiotemporal variation of PM_2.5 and its relationship with changes in emissions, weather and climate can be used to improve the current understanding of pollution formation and provide the scientific basis of air quality improvement to policy makers, it is essential to produce a long-term dataset of gridded surface PM_2.5 concentrations based on real observed data in China.

To overcome the spatiotemporal coverage deficiency of surface PM_2.5 observations, satellite remote sensing data have been widely used to estimate surface PM_2.5 concentrations recently (Fang et al., 2016; Wei et al., 2019). In general, aerosol optical depth (AOD) derived from satellite has a positive correlation with near-surface PM_2.5 concentrations. Based on this, a variety of statistical models, including multiple linear regression (Chelani, 2019), geographically weighted regression (Ma et al., 2014; Guo et al., 2017), linear mixed-effect model (Zheng et al., 2016), and two-stage model (Ma et al., 2016; Yao et al., 2019), have been applied to assess PM_2.5. In addition, machine learning has become a modern tool for a regression task nowadays due to its computational efficiency and state-of-the-art performance (Stafoggia et al., 2019). Wei et al. (2019) produced PM_2.5 concentrations at 1-km resolution in China for 2016 based on satellite AOD using the space-time random forest (STRF) model, with a cross-validation (CV) coefficient of determination (R²) of 0.85. Li et al. (2017) estimated PM_2.5 in 2015 over China using a geo-intelligent deep learning model together with satellite AOD data, in which the CV R² increases from 0.42 to 0.88 relative to the traditional neural network method. However, these estimated PM_2.5 data still have some limitations in certain aspects. First of all, the Moderate Resolution Imaging Spectroradiometer (MODIS) data were not available until 1999 and Suomi National Polar-orbiting Partnership (S-NPP) satellite was launched in 2011. Most of the studies mentioned above used AOD derived from these two satellites to predict PM_2.5, and consequently the PM_2.5 data are not available before 2000 (van Donkelaar et al., 2015; Xue et al., 2019). Additionally, AOD represents aerosol loading in the entire atmospheric column and its relationship with near-surface PM_2.5 concentrations is largely influenced by planetary boundary layer height, relative humidity, temperature, and other factors (Liu et al., 2009). Moreover, algorithm bias, signal uncertainty, and cloud contamination induce biases to the PM_2.5 estimation from AOD (Stafoggia et al., 2019; Xiao et al., 2017).

Atmospheric visibility measurements, which have been available for several decades in China, were demonstrated to be a promising alternative for estimating near-surface PM_2.5 concentrations (Shen et al., 2016). Li et al. (2020) derived PM_2.5 concentrations over North China in 2014 using the combination of visibility observations and GEOS-Chem model simulations and reported that the estimated PM_2.5 was highly correlated to surface observations in time and space, with a correlation coefficient of 0.96 and 0.79, respectively. Liu et al. (2017) estimated historical (1957–1964 and 1973–2014) PM_2.5 in China using visibility measurements and a statistical approach, and found that the model can accurately estimate PM_2.5 concentrations with the CV R² of 0.71. They also reported an increasing trend of 1.9 μg/m³/decade averaged over China during 1957–2014. Due to the better abilities in dealing with non-linear and complex relationships between variables than traditional statistical approaches, machine learning methods can also be used in the visibility-PM_2.5 prediction. By using a machine learning model (the Extreme Gradient Boosting), Gui et al. (2020) constructed surface PM_2.5 concentrations in 2018 over China based on visibility and meteorological data, which offered the potential in reconstructing long-term PM_2.5 data in China with a machine learning method. Furthermore, in addition to visibility and meteorology, other factors such as emissions, topography, population and land use data, should be considered in the machine learning model to simulate PM_2.5 concentrations and spatiotemporal distributions.

In this study, we construct a gridded dataset of near-surface PM_2.5 concentrations across China covering 1980–2019 using the STRF model along with atmospheric visibility and other auxiliary data (e.g., meteorology, anthropogenic emissions, land use, topography, population density and spatiotemporal information), which have a longer time coverage and are more representative of the near-surface aerosols than the data based on satellite AOD. The performance of the STRF model in estimating PM_2.5 in China is evaluated and the long-term variations of PM_2.5 are characterized.