Hyperlocal Air Pollution Mapping: Innovations in Mobile Sensing with AI

Abstract

This doctoral research investigates novel methodologies for high-resolution urban environmental exposure assessment, with particular emphasis on air pollution mapping through mobile monitoring platforms. The study addresses critical limitations inherent in mobile monitoring approaches that leverage machine learning, transfer learning and deep learning techniques to map urban hyperlocal air pollution across Amsterdam, Copenhagen, and Rotterdam. The research makes four principal contributions to the field of environmental exposure science: First, it formally characterizes and mitigates the exposure disparities problem arising from discrepancies between mobile on-road measurements and residential exposure estimates. Through transfer learning, specifically implemented TrAdaBoost and relative unconstrained least-squares importance fitting (RuLSIF) algorithms within land use regression (LUR) modeling architectures, the study demonstrates significant improvements (20-30% enhancement in predictive accuracy) in estimating long-term residential exposures from transient mobile measurements. Second, the work advances spatiotemporal modeling techniques for mobile monitoring data, developing methodologies to generate hourly resolved pollution maps despite the inherent temporal sparsity of mobile measurements. The proposed approach integrates spatial autocorrelation structures with temporal variation patterns, enabling creation of dynamic exposure estimates that better account for mobility patterns. Third, the research investigates and overcomes limitations in spatial generalizability of LUR models through development of the CORrelation ALignment (CORAL) ensemble method. This unsupervised domain adaptation technique successfully transfers model knowledge between geographically distinct urban areas while maintaining predictive performance, addressing a longstanding challenge in the field. Fourth, the study introduces the Visual Land Use Regression (VLUR) framework, which incorporates deep learning-derived features from street-view imagery to enhance traditional LUR models. This innovation improvements in model accuracy while identifying previously unrecognized built environment features (e.g., chimney density, traffic infrastructure) as significant predictors of hyperlocal air pollution variation. The research further demonstrates the extensibility of mobile monitoring platforms to additional environmental stressors through pilot studies of noise and urban heat mapping. These investigations inform the conceptual development of an Exposure Fundamental Model, proposing a unified framework for multimodal environmental exposure assessment through mobile sensing and artificial intelligence. These methodological advances carry significant implications for environmental epidemiology and urban planning. The enhanced exposure assessment capabilities support more precise health effects estimation, while the high-resolution mapping outputs enable targeted intervention strategies. The technical approaches developed in this work offer scalable solutions for cities worldwide to overcome traditional monitoring limitations and advance evidence-based urban environmental management.