Predicting Air Pollution from sparse and movable observation points using machine learning techniques

Exposure to pollutants is closely linked to respiratory illness, cardiovascular disease, and premature mortality. Accurate full-field prediction of air pollutant concentrations is essential for assessing exposure to pollution and guide sustainable urban planning. However, the intrinsic interaction among pollutants, hard-to-predict weather patterns, along with limited and randomly placed monitoring stations make this a complex task. While the domain has shifted from traditional interpolation methods towards machine learning algorithms, generation of high-resolution maps remain challenging. In this study, we use hourly available sparse data and apply data-driven techniques to provide faster and accurate reconstruction of four key pollutants - NO₂, O₃, PM_2.5 and PM₁₀. Models are trained on full-field simulation data and evaluated on real-world observations collected from 20-25 monitoring stations in the city of Paris. We propose multi-pollutant modelling using both discriminative and ensemble-based generative approaches, investigate the impact of incorporating historical data into discriminative models, and introduce stochastic modelling via diffusion techniques to capture the variability in spatial distribution. Despite observing anomalies in spatial map and recording noisy observations, the proposed ML models achieve high structural similarity (SSIM) in field reconstruction. By utilizing noise-based augmentation strategy, we facilitate prediction of real-world data without model retraining. The models exhibit superior generalization ability on real-data by predicting realistic pollution patterns on time periods that lie outside the training period. These findings highlight the potential of ML-models for reliable real-world deployment in reconstruction tasks.