IDENTIFICATION OF THE DRIVING FACTORS OF NO2 CONCENTRATIONS IN PAKISTAN USING MACHINE LEARNING

Abstract

Nitrogen dioxide (NO₂) is a short-lived trace gas but highly toxic in nature that causes serious health issues in a country like Pakistan where ground monitoring infrastructure is very poor to support targeted mitigation. Modelling the concentration of tropospheric NO₂ is very important for public health and environmental concerns. This study proposes the use of multi-satellite datasets and machine learning approach to identify the driving factors of elevated NO₂ concentration in Pakistan.  High resolution (0.05° × 0.05°) tropospheric NO₂ dataset was collected from Sentinel-5P TROPOMI sensor and combined with 19 other predictor variables selected from various auxiliary domains such as meteorology, land-use/land-cover, anthropogenic emissions and topography. Pearson correlation coefficient ‘r’ was computed and highly correlated coexisting features were dropped from the feature set using a fair threshold of |r|<0.85. To model non-linear and complex interactions between tropospheric NO₂ pollution and its drivers, machine learning algorithms, Random Forest (RF) and Extreme Gradient Boosting (XGBoost), were trained on 80% of the dataset and evaluated on the remaining 20% of the dataset. The XGBoost model outperformed the RF with a R²=0.928 (92.8% accuracy in explaining the dataset), RMSE = 3.70×10⁻⁶ mol/m², and MAE = 2.40× 10⁻⁶ mol/m².  SHapley Additive exPlanations (SHAP) values were evaluated to find the relative importance and marginal contribution of each predictor variable in the final prediction of the model. Visuals like geospatial maps, scatter plots and beeswarm plots were drawn using Python libraries to understand the underlying dynamics. The results show that road density, surface temperature, surface elevation and industrial emissions are the main driving factors that control the tropospheric NO₂ variation across Pakistan.  Geographical differences were observed across the four canonical domains of spatial attribution: across 65.5 % of the grid cells in Pakistan, the  first domain of Demographics & Topography remained dominant, providing baseline corridors of accumulation of natural processes in low-lying river plains and high density transportation corridors; the second domain of Meteorology governed the natural processes across 29.8 % of the grid cells in Pakistan, with natural dispersion and the creation of thermal inversion driven winter peaks in pristine northern highlands and coastal belts; the third domain of Direct Anthropogenic Emissions, with extreme degrees of indicators of concentration, governed acute and severe hot spots across 4.7 % of the grid cells across Pakistan, concentrated over the Lahore – Faisalabad industrial belt and Karachi. These results reveal that boundary layer compression in the presence of dense transport and industrial corridors, and not a surge in emissions alone, trigger recurrent winter smog crises. This research offers an easily understood, airshed-specific methodology for moving from blanket restrictions to emission controls that are coupled to meteorology.