A Decadal Study of PM2.5 Concentrations over Delhi using MERRA-2 and Ground Measurements: Predictive Insights via Machine Learning

Abstract

This study investigates the spatial and temporal variations of PM_2.5 concentrations in Delhi from 2014 to 2023, utilizing ground-based measurements from the Central Pollution Control Board (CPCB) and MERRA-2 reanalysis data. The analysis reveals strong positive correlations (r > 0.90) across all districts, highlighting city-wide factors influencing PM_2.5 levels, such as vehicular emissions, industrial activities, and regional weather patterns. Seasonal patterns show PM_2.5 concentrations peaking during winter, attributed to lower temperatures, reduced wind speeds, and increased emissions from heating sources.To enhance the accuracy of PM_2.5 predictions, various machine learning (ML) models were employed, including Extra Trees Regressor, Random Forest Regressor, Light Gradient Boosting Machine (LGBM) Regressor, and a Stacking Regressor. These models utilized MERRA-2 sub-parameters like Dust, Organic Carbon, Black Carbon, Sea Salt, and Sulfate. The Stacking Regressor demonstrated the best performance, achieving an R² value of 0.67 and a significant improvement in correlation with CPCB measurements (r = 0.86). The ML models significantly improved the prediction accuracy of PM_2.5 concentrations compared to the original MERRA-2 data, reducing the Mean Bias from -39.4 µg/m³ to around 10.4µg/m³ and the Root Mean Squared Error (RMSE) from 71.1 µg/m³ to below 40 µg/m³. Additionally, the Fraction of predictions within a factor of 2 increased from 0.61 for MERRA-2 to over 0.89 for all ML models.These findings underscore the effectiveness of integrating machine learning models with MERRA-2 sub-parameters to accurately estimate PM_2.5 concentrations. This approach provides more reliable predictions of air quality, essential for developing targeted and effective air quality management strategies in Delhi.