Supplementary material to "Tropospheric ozone trends and attributions over East and Southeast Asia in 1995–2019: An integrated assessment using statistical methods, machine learning models, and multiple chemical transport models"

Xiao Lu, Yiming Liu, Jiayin Su, Xiang Weng, Tabish Ansari, Yuqiang Zhang, Guowen He, Yuqi Zhu, Haolin Wang, Ganquan Zeng, Jingyu Li, Cheng He, Shuai Li, Teerachai Amnuaylojaroen, Tim Butler, Fan Qi, Shaojia Fan, Grant L. Forster, Meng Gao, Jianlin Hu, Yugo Kanaya, Mohd Talib Latif, Keding Lu, Philippe Nédélec, Peer Nowack, Bastien Sauvage, Xiaobin Xu, Lin Zhang, Ke Li, Ja‐Ho Koo, Tatsuya Nagashima2024 · DOI

Introduction

Tropospheric ozone ($O_3$) is a critical environmental challenge, acting as both a potent greenhouse gas and a toxic pollutant that impacts human health and agricultural productivity. In East and Southeast Asia, rapid industrialization and evolving environmental policies over the last quarter-century have led to complex and highly variable ozone trends. Attributing these trends to either meteorological shifts or changes in anthropogenic emissions is difficult due to the non-linear interactions between weather and atmospheric chemistry. This study employs an integrated ensemble approach—combining statistical regressions, machine learning, and multiple chemical transport models (CTMs)—to disentangle these drivers over the period 1995–2019.

Methodology

The researchers utilized a hierarchical modeling framework to ensure robust attribution. This involved comparing Multiple Linear Regression (MLR), Ridge Regression (RR), and Random Forest Regression (RFR) to "de-weather" ground-based Maximum Daily 8-hour Average (MDA8) $O_3$ measurements. By stripping away meteorological variability, the underlying emission-driven trends could be isolated.

Methodological framework for de-weathering ozone data using statistical and machine learning models.

In addition to statistical methods, three major CTMs—GEOS-Chem, WRF-CMAQ, and CAM4-chem—were used to simulate atmospheric chemistry. Source tagging techniques ($NO_x$ and VOC tagging) were applied within these models to track the geographical origins of observed ozone in specific Asian city-clusters. Meteorological data was sourced from high-resolution ERA5 and MERRA-2 re-analyses.

A primary finding is the superior performance of non-linear models. Random Forest Regression consistently outperformed linear methods, achieving $R^2$ values up to 0.64 in regions like the Pearl River Delta (PRD). This highlights the complexity of weather-ozone interactions.

Analysis of meteorological trends from 1995 to 2019 reveals significant increases in surface solar radiation and temperature across much of East Asia. These shifts likely exacerbated ozone pollution independently of emission changes. Temperature (at 2m) and solar radiation emerged as the most powerful predictors of ozone levels across nearly all studied Asian cities.

Trends in solar radiation, temperature, specific humidity, and wind speed over Asia (1995–2019).

Regional Insights and Chemical Shifts

  • Chemical Sensitivity: A decline in the $H_2O_2/HNO_3$ ratio in the North China Plain indicates a shift in chemical regimes, suggesting that ozone production has become more sensitive to specific emission controls over time.
  • Source Attribution: While local East Asian VOC emissions remain dominant, "Other" regions, including Southeast Asia and global background transport, contribute significantly to the total tropospheric column ozone.
  • Vertical Consistency: Despite discrepancies in absolute concentrations in the upper atmosphere, the CTMs showed consistent vertical profiles in the lower troposphere, reinforcing confidence in surface-level results.

Limitations and Future Directions

The study identifies lower ML model performance in topographically complex areas like the Sichuan Basin, suggesting a need for better representation of local sources and terrain. Furthermore, data gaps in Southeast Asia necessitate more refined regional tagging to fully understand transboundary transport. Finally, while trends are consistent across models, absolute ozone magnitudes remain sensitive to uncertainties in emission inventories in these rapidly developing regions.

Summary updated

BibTeX
@misc{lu2024supplementary,
  author = {Xiao Lu and Yiming Liu and Jiayin Su and Xiang Weng and Tabish Ansari and Yuqiang Zhang and Guowen He and Yuqi Zhu and Haolin Wang and Ganquan Zeng and Jingyu Li and Cheng He and Shuai Li and Teerachai Amnuaylojaroen and Tim Butler and Fan Qi and Shaojia Fan and Grant L. Forster and Meng Gao and Jianlin Hu and Yugo Kanaya and Mohd Talib Latif and Keding Lu and Philippe Nédélec and Peer Nowack and Bastien Sauvage and Xiaobin Xu and Lin Zhang and Ke Li and Ja‐Ho Koo and Tatsuya Nagashima},
  title = {{Supplementary material to "Tropospheric ozone trends and attributions over East and Southeast Asia in 1995–2019: An integrated assessment using statistical methods, machine learning models, and multiple chemical transport models"}},
  year = {2024},
  doi = {10.5194/egusphere-2024-3702-supplement},
}
Figure