ANNOUNCEMENTS
Delhi faces a persistent and severe air quality crisis, with particulate matter (PM2.5 and PM10) concentrations consistently exceeding permissible limits established by both national and international standards. Despite the implementation of regulatory frameworks and mitigation efforts, the fine-scale spatio-temporal variability of air pollution and its underlying drivers remain inadequately characterized. This study adopts an interdisciplinary framework that integrates environmental science, geospatial analysis, advanced statistical modeling, and machine learning to enhance the understanding of air quality dynamics and support the development of effective mitigation strategies.
Leveraging a comprehensive dataset from 39 air quality monitoring stations spanning 2019 to 2023, this research employs trend analysis and time series modeling to quantify temporal variations at individual monitoring sites. Furthermore, advanced machine learning and deep learning models, including Random Forest, Support Vector Machines, Artificial Neural Networks, and Bidirectional Long short-term memory (Bi-LSTM) models were integrated into the time series framework to improve predictive accuracy and provide insights into non-linear and complex interactions influencing pollutant concentrations. The framework uniquely utilized station-specific hyperparameter optimization, comprehensive exogenous variables (co-pollutants: NO₂, SO₂, O₃, CO; and meteorological parameters), and Fourier-transformed functions which explicitly capture the multi-scale seasonal variations. Land Use Regression (LUR) models were developed to assess the spatial variability of particulate pollution, incorporating a diverse array of geographic, meteorological, and socio-economic predictors. The integration of machine learning techniques, including Random Forest and Deep Neural Networks (DNN), further augmented the LUR framework, enabling a deeper understanding of the non-linear relationships governing pollutant distribution.
Trend analysis conducted utilizing the robust monitoring network revealed significant seasonal variations, with PM concentrations peaking during winter months due to adverse meteorological conditions and anthropogenic activities, while the monsoon season showed notable reductions. Furthermore, we observed a broadly improving profile for PM2.5 levels, marked by statistically significant decline across most monitoring stations, while, PM10 displayed pronounced spatial heterogeneity, with significant rise at several key monitoring stations. This underscores the need for region-specific air quality management strategies, informed by unique pollution trends at each monitoring station, to effectively address local air quality challenges.
Time series modeling using traditional models like SARIMAX, machine learning models including RF and SVM, and deep learning models such as ANN and Bi-LSTM provided reliable predictions. We demonstrated the superior predictive capability of ANN and Bi-LSTM models, achieving testing R squared values of 0.81–0.98 and root mean square error (RMSE) as low as 10.75 μg/m³, capturing both seasonal and non-linear variations.
LUR models utilizing an extensive set of geospatial predictors demonstrated a strong relation between particulate concentrations and urban features, accounting for 36–47% of the variability in PM2.5 levels and 31–52% in PM10 levels. Moreover, the LUR-RF model, delivered the best performance, especially notable in the annual and post-monsoon periods, reflecting higher accuracy and the lowest RMSE scores.
These findings underscore the complex interplay of local and regional factors influencing air quality and highlight the potential of advanced modeling techniques to inform targeted policy interventions. This study highlights the spatio-temporal complexity of air pollution in Delhi and the critical role of predictive models in shaping effective interventions. Furthermore, it offers a replicable decision–support framework for sustainable urban air management, with the potential to mitigate the health and economic burdens of air pollution in densely populated regions.
