Cluster-enhanced Ensemble Learning for Mapping Surface Ozone in China

As global air quality deteriorates, ground-level ozone (O3) has gained significant attention as a major secondary pollutant. Its simulation and prediction have become key research areas. Existing studies on ozone concentration simulation and prediction primarily focus on individual urban scales, confirming the importance of wind direction in ozone simulation. However, its impact has not been sufficiently addressed in nationwide studies. This research integrates multisource observational data from 2003 to 2019 across China, constructing a comprehensive meteorological-environmental dataset that includes ERA5 meteorological reanalysis data, CAMS and MERRA2 chemical reanalysis data, MODIS land use and leaf area index data, MODIS aerosol optical depth, OMI/Aura satellite NO2 and O3 data, BS total column ozone data, and the CEDS emission inventory. The study emphasizes the spatiotemporal variation of wind direction and incorporates wind direction as a factor to enhance the accuracy of ozone concentration simulations. Furthermore, the research explores the impact of meteorological factors on ozone concentration in China and its spatiotemporal variation characteristics. This research employs cluster-enhanced ensemble learning methodologies. First, the K-means clustering algorithm categorizes ozone-related data, providing a basis for further analysis. The optimal cluster number is determined using the elbow method. Next, various ensemble learning models, including penalized linear models (LASSO), multi-layer perceptron neural networks (MLP), random forests (RF), LightGBM, XGBoost, and CatBoost, are applied to simulate ozone concentrations. Hyperparameters are optimized via randomized grid search to enhance model performance. To improve spatial prediction accuracy, geographically weighted generalized additive models (GWGAM) are incorporated to better capture local ozone variations. Additionally, the study quantifies wind direction influence using the 24-hour local recirculation index (R_24). This index measures pollutant transport by comparing straight-line and total trajectory distances, reflecting outward transport and recirculation effects. The study found that meteorological factors play a key role in the distribution of ozone concentration. The 24-hour local circulation index is negatively correlated with ozone concentration in non-Beijing-Tianjin-Hebei regions, with a correlation index of approximately -0.178. The test results show that lightGBM and XGBoost models performed best among the six sub-models, with an RMSE of approximately 5.70 ppbv and an R2 exceeding 0.800, ranked as LightGBM > XGBoost > CatBoost > RF = MLP > LASSO. The ensemble model outperformed all individual sub-models and achieved the best performance. Additionally, the ensemble model was validated for its generalization ability in predicting both annual and monthly average concentrations, and the importance of the R_24 feature in each sub-model was computed. The simulation results show that ozone concentrations are higher in eastern and central China (particularly in East, Central, and South China), and these areas have seen a steady increase in ozone levels over time. In contrast, the southwest, northeast, Taiwan, and Hainan regions have relatively lower ozone concentrations. The spatial distribution of ozone pollution aligns with actual observational data. From 2014 to 2019, ozone pollution showed a steady upward trend, consistent with observed data. In conclusion, this study improves the accuracy of ozone concentration simulation and provides a theoretical foundation for air pollution control and policy development. The study quantifies the role of R_24 in ozone modeling, providing scientific support for optimizing ozone simulation models and improving air quality management policies in China.

Grantor dc:publisher

Helsingin yliopisto

Year dc:date.issued

2025

Author dc:creator

• Qin, Tian

• ozone
• Ensemble Learning Models
• Geographically Weighted Generalized Additive Model
• 24-hour local recirculation index

Statement dc:rights

• In Copyright 1.0

Language dc:language.iso

eng

Handle dc:identifier.uri

http://hdl.handle.net/10138/598123

OAI identifier oai:identifier

oai:helda.helsinki.fi:10138/598123