Machine learning estimation of surface ozone using Sentinel-5P data and meteorological and ground observations

Abstract

Surface ozone is a critical air pollutant with serious implications for human health and ecosystem balance. This study utilized Sentinel-5P satellite data, meteorological parameters, and ground observations from 2018 to 2021 to estimate surface ozone concentrations in Tehran, capital of Iran. Four machine learning algorithms, including Random Forest (RF), Support Vector Regression (SVR), Multilayer Perceptron (MLP), and Extreme Gradient Boosting (XGBoost), were employed for this purpose. The RF algorithm outperformed the others, achieving average Mean Bias Error (MBE) of 0.14 ($\mu g/m^3$), Root Mean Square Error (RMSE) of 13.54 ($\mu g/m^3$), and Normalized RMSE (NRMSE) of 9.48%. Statistical analysis using the Wilcoxon test confirmed that the RF algorithm’s performance was significantly superior. Permutation analysis was then applied to interpret the RF behavior in surface ozone estimation, highlighting the contribution of total column nitrogen dioxide, temperature, relative humidity, and v-component wind speed. Further analysis revealed that incorporating meteorological parameters at various pressure levels. Finally, an Ozone Exposure Risk (OER) index was introduced to relatively vulnerable areas in Tehran by integrating surface ozone estimates with population density data. Based on the results, the southern and eastern parts of the study region exhibited the highest OER compared to the other regions. The results of this study can serve as input for policymakers and urban planners to inform the establishment of legislation that promotes sustainable urban development and improves society’s public health.