Spatial prediction of on-road air pollution using long-term mobile monitoring: Insights from Delhi

A rapid increase in the density of urban activities nudges for a dense air quality monitoring network. Mobile monitoring using low-cost air quality devices provides a valuable method for capturing spatiotemporal variations of pollutants in the absence of a dense air quality monitoring network. Few studies advocated using linear and nonlinear models, whereas others have utilized machine learning (ML) models for spatial prediction. However, the application in the existing studies is limited to a shorter period and smaller area. Additionally, an understanding of the selection of these models is absent. This study uses ${\mathrm{PM}}_{2.5}$ concentrations from 15 low-cost air quality devices deployed in buses in Delhi for over eight months to compare the performance of different model categories. ${\mathrm{PM}}_{2.5}$ data is aggregated at the midpoint of the 1110 road segments. Various predictor variables, which exhibit spatiotemporal variations, are used in the prediction models. Among the linear models, Backward Stepwise Regression achieved the highest R² (0.61) for the training dataset, and among ML models, Extreme Gradient Boosting exhibits the highest R² (0.98). Temperature, humidity, built-up area, building height, road length, and traffic signals are the main influencing predictor variables. ML models perform better among all model categories, whereas linear models have a smaller divergence between training and validation R². Additionally, linear models have better prediction consistency than nonlinear and ML models. These results confirm the high performance of ML models and exhibit the potential for improving prediction accuracy by splitting the data into smaller time bins and including more road segments.