A fine-scale spatiotemporal air quality modeling framework by combining big but noisy data
Author(s)
Fan, Hongwei
Type
Thesis or dissertation
Abstract
Ambient air pollution is a major public health crisis. Traditional air quality monitoring relies on sparse reference-grade stations, which are costly and unable to capture fine-scale spatial variability. Although low-cost sensor networks provide extensive spatial coverage, their measurements often contain substantial noise and uncertainty. This dissertation addresses the challenge of transforming “big but noisy” spatiotemporal datasets—including low-cost sensor observations, land-use, traffic, and meteorological data—into accurate, high-resolution hourly maps of PM2.5 and NO2 concentrations.
Firstly, a scalable calibration methodology was developed to significantly enhance data reliability for networks encompassing hundreds of low-cost sensors. A novel transformer-based calibration approach was introduced, uniquely leveraging temporal dependencies (e.g., diurnal variations) in sensor data, thereby reducing the root mean squared error (RMSE) by 10–20% compared to traditional methods.
Secondly, an innovative deep learning framework based on the Vision Transformer (ViT) architecture was introduced, capitalising on its attention mechanism to model intricate spatial patterns across the urban landscape. Applied to calibrated sensor data and auxiliary inputs, the ViT-based model generates PM2.5 and NO2 predictions at 100 m × 100 m hourly resolution, consistently outperforming three established baselines.
Finally, comprehensive uncertainty quantification is implemented. Uncertainty arising from calibrated low-cost sensors and proposed ViT-based model predictions are systematically estimated. To effectively integrate these uncertainties, a Bayesian Maximum Entropy (BME) data fusion framework was developed, providing robust probabilistic predictions and rigorous uncertainty quantification for fine-scale urban air quality estimations.
Collectively, this research addresses critical challenges in large-scale calibration of low-cost sensors, development of fine-scale air quality models, and rigorous uncertainty quantification through data fusion. The proposed framework enables reliable urban air quality estimation at fine spatiotemporal scales, supporting evidence-based environmental management and public health policy.
Keywords: Air pollution, low-cost sensors, calibration, deep learning, Vision Transformer, Bayesian Maximum Entropy, spatiotemporal mapping, uncertainty quantification.
Firstly, a scalable calibration methodology was developed to significantly enhance data reliability for networks encompassing hundreds of low-cost sensors. A novel transformer-based calibration approach was introduced, uniquely leveraging temporal dependencies (e.g., diurnal variations) in sensor data, thereby reducing the root mean squared error (RMSE) by 10–20% compared to traditional methods.
Secondly, an innovative deep learning framework based on the Vision Transformer (ViT) architecture was introduced, capitalising on its attention mechanism to model intricate spatial patterns across the urban landscape. Applied to calibrated sensor data and auxiliary inputs, the ViT-based model generates PM2.5 and NO2 predictions at 100 m × 100 m hourly resolution, consistently outperforming three established baselines.
Finally, comprehensive uncertainty quantification is implemented. Uncertainty arising from calibrated low-cost sensors and proposed ViT-based model predictions are systematically estimated. To effectively integrate these uncertainties, a Bayesian Maximum Entropy (BME) data fusion framework was developed, providing robust probabilistic predictions and rigorous uncertainty quantification for fine-scale urban air quality estimations.
Collectively, this research addresses critical challenges in large-scale calibration of low-cost sensors, development of fine-scale air quality models, and rigorous uncertainty quantification through data fusion. The proposed framework enables reliable urban air quality estimation at fine spatiotemporal scales, supporting evidence-based environmental management and public health policy.
Keywords: Air pollution, low-cost sensors, calibration, deep learning, Vision Transformer, Bayesian Maximum Entropy, spatiotemporal mapping, uncertainty quantification.
Version
Open Access
Date Issued
2025-07-30
Date Awarded
2026-06-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
de Nazelle, Audrey
Arcucci, Rossella
Publisher Department
Centre for Environmental Policy
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)