Suitability of Machine Learning Models and their Performance for PM2.5 Estimation using high-resolution satellite-driven datasets over Northwest India

Abstract

The estimation of particulate matter (PM2.5) levels using high-resolution satellite-driven datasets and machine learning algorithms represents a potential advancement in air quality monitoring. The traditional ground-based PM2.5 measurements, while accurate, suffer from limited spatial coverage, prompting the need for satellite-based retrieval methods. The machine learning (ML) algorithms convert high-resolution satellite-derived Aerosol Optical Depth (AOD) into PM2.5 and enhance the accuracy estimation. Hereby, this study used satellite-driven PM2.5 estimation framework based on Multi-Angle Implementation of Atmospheric Correction (MAIAC) AOD at 1 km and meteorological data over Northwest India (NW). This study used Random Forest (RF), Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), and Adaptive Boosting (AdaBoost) models to integrate the MAIAC AOD with meteorological variables. The datasets have been pre-processed and optimized for better accuracy from 2022 to 2023, aligning with ground observations. RF and XGBoost (R2 = 0.91 and 0.91, RMSE = 10.36 µg/m3 and 10.39 µg/m3, Bias = 0.30 µg/m3 and 0.48 µg/m3, respectively) outperform AdaBoost and SVM in this study. The estimated PM2.5 values exceed the prescribed National Ambient Air Quality Standards (NAAQS), with mean 24-h and annual average concentrations of 74.05 µg/m3 and 70.53 µg/m3, underlining severe air pollution in the region. This study presented a region-specific, multi-source PM2.5 estimation framework that integrates satellite AOD, meteorological drivers, and optimized machine learning model to improve spatial accuracy and robustness. These approaches provide actionable insights for air quality monitoring and policymaking, enhance the ability to capture the complexity of PM2.5 variability, and facilitate predictive models that could help in efficient air quality management.