Atmosphere, Vol. 17, Pages 125: Adaptive Data-Driven Framework for Unsupervised Learning of Air Pollution in Urban Micro-Environments

Atmosphere doi: 10.3390/atmos17020125

Authors:
Abdelrahman Eid
Shehdeh Jodeh
Raghad Eid
Ghadir Hanbali
Abdelkhaleq Chakir
Estelle Roth

(1) Background: Urban traffic micro-environments show strong spatial and temporal variability. Short and intensive campaigns remain a practical approach for understanding exposure patterns in complex environments, but they need clear and interpretable summaries that are not limited to simple site or time segmentation. (2) Methods: We carried out a multi-site campaign across five traffic-affected micro-environments, where measurements covered several pollutants, gases, and meteorological variables. A machine learning framework was introduced to learn interpretable operational regimes as recurring multivariate states using clustering with stability checks, and then we evaluated their added explanatory value and cross-site transfer using a strict site hold-out design to avoid information leakage. (3) Results: Five regimes were identified, representing combinations of emission intensity and ventilation strength. Incorporating regime information increased the explanatory power of simple NO2 models and allowed the imputation of missing H2S day using regime-aware random forest with an R2 near 0.97. Regime labels remained identifiable using reduced sensor sets, while cross-site forecasting transferred well for NO2 but was limited for PM, indicating stronger local effects for particles. (4) Conclusions: Operational-regime learning can transform short multivariate campaigns into practical and interpretable summaries of urban air pollution, while supporting data recovery and cautious model transfer.

Source link

Abdelrahman Eid www.mdpi.com