Spatiotemporal assessment and machine learning-based prediction of PM2.5 Emissions from biomass combustion in Rural India

Background: Biomass combustion remains a dominant cooking practice in rural India, contributing significantly to indoor and ambient air pollution. Exposure to particulate matter (PM), particularly PM2.5, poses severe health risks, especially in poorly ventilated households. This study aims to analyze PM emission patterns from traditional Chulha use and assess the influence of meteorological factors using machine learning techniques. Methods: Particulate matter concentrations (PM₁, PM2.5, PM₄, PM₁₀) were monitored over a one-year period in two rural sites- Jhunjhunu (JJN) and Mahendragarh (Mgarh). Data were collected across three daily cooking intervals and combined with meteorological variables (temperature, humidity, wind speed, rainfall). Regression models (Linear, Random Forest, XGBoost), classification algorithms, and unsupervised learning (K-Means, Isolation Forest) were applied to predict, classify, and analyze pollution patterns. Results: PM levels peaked during winter and evening cooking hours, often exceeding WHO air quality standards 2021. Meteorological variables, particularly temperature and wind speed, showed strong seasonal influence on pollutant dispersion. Random Forest Regression achieved the best predictive performance (R² = 0.87, RMSE = 18.3 µg/m³ in Jhunjhunu), while classification accuracy reached 98%. SHAP analysis identified PM2.5 lag, humidity, and wind chill as key predictors. Clustering revealed distinct pollution regimes, and anomaly detection successfully flagged episodic high-pollution events. Conclusion: The integration of temporal, meteorological, and machine learning analysis offers a robust framework for understanding rural air pollution. The findings underscore the need for clean cooking interventions, targeted health risk communication, and the application of predictive tools in rural air quality management.