Dissertation Defense - Anusha Srirenganathanmalarvizhi - Department of Geography and Geoinformation Science

PhD Candidate: Anusha Srirenganathan Malarvizhi
PhD of Science, Earth Systems and Geoinformation Sciences
Department of Geography and Geoinformation Science 

Dissertation Title: A Geospatial and Deep Learning Framework for Fine-Grained Spatiotemporal PM2.5 Estimation based on Multi-Sourced Air Quality and Meteorological Data 

Dissertation Chair: Chaowei Yang, PhD, Department of Geography and Geoinformation Science

Committee Members: 

Dr. Matthew Rice (GMU, Geography and Geoinformation Science)
Dr. Donglian Sun (GMU, Geography and Geoinformation Science
Dr. Qian Liu (GMU, Geography and Geoinformation Science)
Dr. Daniel Tong (GMU, Atmospheric, Oceanic & Earth Sciences)

Abstract: Particulate matter (PM), particularly fine particles with aerodynamic diameters below 2.5 μm (PM₂.₅), poses serious health and environmental threats worldwide. While ground-based regulatory stations provide accurate surface PM₂.₅ measurements, their limited spatial coverage restricts continuous monitoring. Satellite observations complement these networks by offering broad spatial coverage; however, satellites do not directly measure PM₂.₅ but instead retrieve Aerosol Optical Depth (AOD), a column-integrated measure of atmospheric aerosols that serves as a proxy for surface PM₂.₅ concentrations. Persistent gaps in AOD retrievals due to cloud cover, along with the complex, nonlinear, and heterogeneous relationship between AOD and surface PM₂.₅, pose major challenges for high-spatiotemporal retrieval. Artificial Intelligence/Machine Learning (AI/ML) models have demonstrated superior performance in capturing these nonlinear dependencies; however, they often exhibit overconfidence—particularly during extreme events, which limits their reliability for decision support. This dissertation addresses these challenges by proposing novel AI/ML frameworks that integrate generative modeling for AOD imputation, transformer-based architectures for high-resolution PM₂.₅ forecasting, and probabilistic approaches for uncertainty quantification. The proposed models are evaluated through two region-specific use cases representing high pollution events: the June 2023 Canadian wildfire in the Northeastern United States, used to assess AOD imputation and fusion performance, and the January 2025 Los Angeles wildfire in California, used to evaluate fine-scale PM₂.₅ retrieval and uncertainty quantification (UQ) under extreme episodic pollution.  

The first component focuses on addressing missing AOD retrievals using generative modeling. A Generative Adversarial Imputation Network (GAIN) is extended to reconstruct missing Moderate Resolution Imaging Spectroradiometer (MODIS) Multi-Angle Implementation of Atmospheric Correction (MAIAC) AOD observations and fuse them with reanalysis data for improved spatial continuity over the Northeastern United States. The approach is evaluated against Aerosol Robotic Network (AERONET) observations, demonstrating a strong agreement (R = 0.91). During the June 2023 Canadian wildfire event, when over 70% of MAIAC AOD retrievals were unavailable due to cloud cover, the GAIN-based framework successfully reconstructed continuous AOD fields that preserved regional gradients and captured the spatiotemporal extent of dense smoke plumes. The results confirm that generative modeling can effectively bridge missing data gaps while maintaining physical consistency and interpretability in reconstructed aerosol distributions. 

The second component develops a spatiotemporal deep learning framework for high-resolution PM₂.₅ forecasting in California, employing a Spatio-Temporal Aware Transformer (STA-Transformer) with a generative decoding strategy to capture complex dependencies across spatial and temporal dimensions. Trained on multi-source data, including satellite AOD, meteorological, and geographical variables from 2021 to January 2025, the framework produces hourly surface PM₂.₅ at a 1 km × 1 km resolution. It achieves R² values of 0.92 and 0.87 for time-based and site-based evaluations, respectively, confirming strong temporal and spatial generalization. The model maintains high accuracy (R² > 0.90) at short lead times and sustains reliable performance (R² > 0.63) up to eight hours ahead, outperforming recurrent neural network baselines. The January 2025 Los Angeles wildfire case study highlights the framework’s capability to capture extreme pollution episodes, with robust retrievals across urban sites. These results demonstrate that transformer-based architectures can enhance fine-scale PM₂.₅ prediction by effectively modeling spatial heterogeneity and long-range dependencies. 

The third component addresses UQ in PM₂.₅ retrieval through probabilistic deep learning. Deep ensemble (DE) and Monte Carlo Dropout (MCD) methods are implemented to generate predictive distributions and evaluate confidence calibration across land-use types, air quality categories, and spatial domains. The probabilistic framework achieved strong coverage calibration (PICP ≈ 95%) in urban and grassland regions and revealed event-driven uncertainty patterns concentrated in wildfire-affected zones. While epistemic uncertainty was effectively captured, aleatoric uncertainty arising from measurement noise remains a focus for future improvement. By quantifying predictive confidence, this framework enhances the interpretability and trustworthiness of model outputs, supporting more informed decision-making in air quality management. 

This research demonstrates a unified and scalable framework that integrates generative AOD imputation, transformer-based PM₂.₅ forecasting, and probabilistic uncertainty quantification to improve the reliability of air quality prediction. The methods are validated through two high-pollution events—the June 2023 Canadian wildfire and the January 2025 Los Angeles wildfire—showing applicability across distinct atmospheric and geographic conditions. Beyond advancing methodological accuracy, the dissertation contributes to the foundation of an Air Quality Digital Twin (AQDT), capable of supporting real-time forecasting, uncertainty propagation analysis, and scenario-based evaluation of emission or climate policy impacts. Future research will extend these models to longer forecasting horizons, integrate aleatoric uncertainty and error propagation from auxiliary inputs, and incorporate out-of-distribution detection to ensure robust, adaptive, and transparent digital air quality monitoring systems.