A machine learning approach to quantify meteorological drivers of ozone pollution in China from 2015 to 2019

Xiang Weng, Grant L. Forster, Peer NowackAtmospheric 2022 · DOI

TL;DR

This study employs machine learning (Random Forest Regression and Ridge Regression) to untangle the complex relationship between meteorological factors and surface ozone pollution in China from 2015-2019. The findings reveal that nonlinear relationships are crucial for accurate ozone modeling, and that incorporating larger-scale meteorological patterns significantly improves predictions, particularly in southern China.

What Problem Does This Paper Address?

Despite significant efforts to control air pollution in China, summertime surface ozone concentrations have continued to rise in many regions. While meteorology is known to play a crucial role in modulating ozone levels, its precise quantitative contribution and the nature of its influence (linear vs. nonlinear) remain uncertain. Existing methods, like Multiple Linear Regression (MLR), struggle to capture these complex interactions, especially when dealing with numerous, potentially collinear, meteorological variables. This paper aims to address this gap by developing and applying advanced machine learning techniques to better quantify the impact of meteorological drivers on ozone pollution and to identify the most influential factors across different regions of China.

Methods

The researchers utilized two primary machine learning regression algorithms:

  • Random Forest Regression (RFR): A nonlinear approach that builds an ensemble of decision trees. RFR is adept at capturing complex, non-linear relationships between predictors and the target variable.
  • Ridge Regression (RR): A linear statistical learning approach that uses L2-regularization. RR is robust in high-dimensional settings and effectively handles collinearity among predictors, making it suitable for scenarios with many meteorological variables.

These methods were compared against traditional approaches:

  • Multiple Linear Regression (MLR): A standard linear model using all available meteorological predictors.
  • Stepwise MLR: A variation of MLR that iteratively selects the most statistically significant predictors.

To assess the models' performance, the researchers employed a rigorous cross-validation strategy, splitting the data into training, validation, and testing sets. They also explored the impact of including "nonlocal" meteorological predictors (i.e., data from surrounding areas) by expanding the predictor domain for RR.

Datasets

The study utilized the following datasets:

  • Surface Ozone Data: Hourly and 8h rolling mean ozone concentrations (MDA8) from ground-based monitoring stations across China, obtained from the Ministry of Ecology and Environment (MEE) via the quotsoft.net/air repository. Data from 2015-2019 were used, with specific quality control measures applied.
  • Meteorological Data: Gridded ERA5 reanalysis data (0.25° × 0.25° spatial resolution) for 11 meteorological variables, including temperature, boundary layer height, pressure, solar radiation, humidity, precipitation, and wind components at different altitudes. This data covered the same 2015-2019 period.

No new benchmark datasets were introduced; instead, the study focused on applying advanced modeling techniques to existing, publicly available datasets.

Key Insights and Ideas

  • Nonlinearity Matters: The study strongly suggests that linear models like MLR and even RR (when using only local predictors) fail to fully capture the complex, often nonlinear, relationships between meteorological factors and ozone. RFR, by its nature, excels at this.
  • Beyond Local Influences: Incorporating meteorological data from a wider spatial domain (nonlocal predictors) significantly improves ozone prediction accuracy, especially in regions like southern China. This highlights the importance of large-scale weather systems.
  • Surface Solar Radiation is Key: Contrary to some previous studies, this research identifies surface solar radiation as a critical meteorological driver for ozone pollution, particularly in central and eastern China.
  • Regional Differences in Drivers: The dominant meteorological drivers for ozone vary significantly by region. Temperature is key in northern China, solar radiation in central/eastern regions, and humidity in southern China.
  • Meteorology's Significant Role in Trends: Approximately half of the observed ozone trends in China between 2015-2019 can be attributed to meteorological variability, underscoring its substantial influence on long-term changes.

Key Results

  • Model Performance: RFR consistently outperformed linear methods (MLR, stepwise MLR) in predicting ozone using local meteorological data, achieving higher R² values (0.5–0.6) compared to linear models (0.4–0.5).
  • Nonlocal Predictors Boost Accuracy: Including nonlocal meteorological data with RR (RR-2D) further improved model skill, particularly in southern China, where R² increased from 0.47 to 0.6.
  • Dominant Drivers: Temperature, surface solar radiation, and relative humidity were identified as the most important meteorological drivers across China, with their regional dominance varying.
  • Trend Attribution: An estimated 45% of the overall April-October ozone trend in China from 2015-2019 was attributed to meteorological factors, with RR-2D providing this estimate.
  • Regional Trend Differences: BTH and PRD showed significant positive observed ozone trends, while YRD and Sichuan had no significant trends. Meteorological factors played a more central role in driving recent positive trends in PRD.

Limitations and Future Work

  • Data Period: The study period of 2015-2019 is relatively short for robust long-term trend analysis, and the authors acknowledge that trend analyses are explorative.
  • Anthropogenic Factors: While the study focuses on meteorological drivers, it acknowledges that residual trends (after accounting for meteorology) likely reflect anthropogenic influences. Future work could further disentangle these contributions.
  • Nonlinearity in RR-2D: While RR-2D showed strong performance, it doesn't inherently capture nonlinearity. The authors suggest that combining nonlinear capabilities with nonlocal predictors could further enhance modeling.
  • Extrapolation: The paper notes a potential trade-off in RR's ability to extrapolate, which might lead to biases in predicting extreme ozone anomalies.
  • Future Work: The authors recommend considering surface solar radiation more broadly in future controlling factor analyses and highlight the importance of incorporating nonlocal meteorological context for more accurate ozone pollution modeling and trend attribution. They also suggest that the machine learning framework can be extended to investigate the impact of climate change on future ozone pollution.

Summary updated

Abstract

Surface ozone concentrations increased in many regions of China from 2015 to 2019. While the central role of meteorology in modulating ozone pollution is widely acknowledged, its quantitative contribution remains highly uncertain. Here, we use a data-driven machine learning approach to assess the impacts of meteorology on surface ozone variations in China for the period 2015–2019, considering the months of highest ozone pollution from April to October. To quantify the importance of various meteorological driver variables, we apply nonlinear random forest regression (RFR) and linear ridge regression (RR) to learn about the relationship between meteorological variability and surface ozone in China, and contrast the results to those obtained with the widely used multiple linear regression (MLR) and stepwise MLR. We show that RFR outperforms the three linear methods when predicting ozone using local meteorological predictor variables, as evident from its higher coefficients of determination (R2) with observations (0.5–0.6 across China) when compared to the linear methods (typically R2 = 0.4–0.5). This refers to the importance of nonlinear relationships between local meteorological factors and ozone, which are not captured by linear regression algorithms. In addition, we find that including nonlocal meteorological predictors can further improve the modelling skill of RR, particularly for southern China where the averaged R2 increases from 0.47 to 0.6. Moreover, this improved RR shows a higher averaged meteorological contribution to the increased trend of ozone pollution in that region, pointing towards an elevated importance of large-scale meteorological phenomena for ozone pollution in southern China. Overall, RFR and RR are in close agreement concerning the leading meteorological drivers behind regional ozone pollution. In line with expectations, our analysis underlines that hot and dry weather conditions with high sunlight intensity are strongly related to high ozone pollution across China, thus further validating our novel approach. In contrast to previous studies, we also highlight surface solar radiation as a key meteorological variable to be considered in future analyses. By comparing our meteorology based predictions with observed ozone values between 2015 and 2019, we estimate that almost half of the 2015–2019 ozone trends across China might have been caused by meteorological variability. These insights are of particular importance given possible increases in the frequency and intensity of weather extremes such as heatwaves under climate change.
BibTeX
@inproceedings{weng2022machine,
  author = {Xiang Weng and Grant L. Forster and Peer Nowack},
  title = {{A machine learning approach to quantify meteorological drivers of ozone pollution in China from 2015 to 2019}},
  booktitle = {Atmospheric},
  year = {2022},
  doi = {10.5194/acp-22-8385-2022},
}
Figure