Abstract
Exposure to fine particulate matter (PM_2.5) is the leading environmental risk factor for mortality worldwide. However, sparse and unevenly distributed ground monitors limit comprehensive assessment of its spatial heterogeneity, chemical composition, and temporal variability. Satellite retrievals and chemical transport models can provide insights into PM_2.5 concentrations for extending coverage to unmonitored regions, and integrating this information through data-driven methods can further enhance the estimation of ambient PM_2.5. This dissertation develops a progression of deep learning frameworks that incorporate geophysical process-based information, paired with rigorous spatial validation and uncertainty quantification, to estimate PM_2.5 at increasingly fine spatial, compositional, and temporal scales. We first develop a convolutional neural network (CNN) that corrects the local bias in monthly geophysical PM_2.5 concentrations globally from 1998 – 2019 with a fine spatial resolution of 0.01°×0.01°. A loss function anchored to the geophysical a priori constrains the unrealistic behavior of classic mean-squared-error training in monitor-sparse regions. We also introduce novel spatial cross-validation for air quality to examine the importance of considering spatial autocorrelation. Building upon this foundation, we extend this framework across North America (2000–2023) to estimate total PM_2.5 and its chemical components, integrating geophysical a priori information that markedly improves performance for components such as nitrate and ammonium. We introduce Buffered Leave Isolated Sites and Clusters Out (BLISCO) cross-validation, which reveals that traditional spatial cross-validation overstates accuracy and understates uncertainty because of spatial autocorrelation. We represent spatial uncertainty for PM_2.5 and its components based on the statistical results of BLISCO cross-validation by integrating information from both the spatial distribution of ground observations and the variability in the spatial representation of predictors. In the third section, we develop a Mixture-of-Experts Ensemble 3D-CNN to generate gapless daily PM_2.5 estimates at 1-km resolution across North America (2019–2023), paired with a Mahalanobis-distance uncertainty framework. The resulting record, a five-year population-weighted mean of 8.2 µg/m³, with wildfire-driven summer peaks and elevated uncertainty in 2023, provides a spatially continuous, uncertainty-quantified basis for air quality management and epidemiological research. Together, these contributions show that embedding geophysical knowledge within deep learning, with rigorous spatial validation and uncertainty quantification, yields PM_2.5 estimates that are simultaneously more accurate and more trustworthy across scales relevant to exposure assessment.
Committee Chair
Randall Martin
Committee Members
Jay Turner; Jian Wang; Nathan Jacobs; Rajan Chakrabarty
Degree
Doctor of Philosophy (PhD)
Author's Department
Energy, Environmental & Chemical Engineering
Document Type
Dissertation
Date of Award
7-28-2026
Language
English (en)
DOI
https://doi.org/10.7936/v3qy-ys35
Recommended Citation
Shen, Siyuan, "Advancing Geophysically Informed Data-Driven Estimation of Ambient Fine Particulate Matter with Uncertainty Quantification" (2026). McKelvey School of Engineering Graduate Student Theses & Dissertations. 1414.
The definitive version is available at https://doi.org/10.7936/v3qy-ys35