Dissertation Defense - Xiqi Fei - Department of Geography and Geoinformation Science

PhD Candidate: Xiqi Fei
PhD of Science, Earth Systems and Geoinformation Sciences
Department of Geography and Geoinformation Science 

Dissertation Title: Machine Learning for Spatio-temporal Analysis of Urban Mobility Data and Satellite Imagery

Dissertation Co-Chairs: Dr. Andreas Züfle (Emory, Computer Science) 
                             Dr. Ruixin Yang (GMU, Geography and Geoinformation Science) 

Committee Members: 
Dr. Dieter Pfoser (GMU, Geography and Geoinformation Science) 
Dr. Arie Croitoru (GMU, Computational and Data Sciences) 

Abstract:
As the volume of spatial-temporal data collected grows at an unprecedented rate, driven by the rapid development and improvement in technology and the increasing availability of data sources, the challenge of efficiently managing, processing, and extracting meaningful insights from this vast volume of information becomes more critical. Our research is crucial in this regard, as it not only identifies and confronts the inherent complexities of handling spatial-temporal data but also provides innovative solutions. 

This dissertation demonstrates spatial-temporal analysis for urban application studies with two types of data: structured vector data and raster data. The research presented in this dissertation advances the domains of urban transportation analysis and remote sensing machine learning modeling, which improves the effectiveness and application of data-driven solutions. Specifically, the first part of the study analyzes the bus tracking data from the Washington Metropolitan Area Transit Authority to identify speed patterns and traffic congestion, employing Dynamic Time Warping for clustering stops based on their temporal speed profiles. The result can be used to generate a spatiotemporal route profile that can provide actionable intelligence for route planning purposes. For raster data, convolutional neural networks (CNN) for model transfer across different locations within the Chesapeake Bay Catchment are employed, demonstrating the potential of machine learning models to generalize effectively across distances up to 600 km with minimal loss in accuracy and confirming the influence of geographic distance on model performance. Furthermore, a novel semi-supervised learning approach incorporating spectral, spatial, and temporal information is developed for satellite image time series segmentation, outperforming traditional methods, especially at low label rates.  

This dissertation highlights the complexity of spatial and temporal dynamics in data analysis, emphasizing the need for holistic views to fully understand and leverage geospatial concepts in model development. The main contributions include using urban transportation sensor data for insights into mobility patterns and decision support, validating the generalization ability of convolutional neural networks in land cover classification with a focus on the impact of geographic distance, and developing a graph-based semi-supervised learning method that incorporates spatial and temporal factors for improved accuracy in time series image segmentation. Together, this dissertation demonstrates the vast potentials of machine learning in the spatio-temporal analysis of mobility data and remote sensing data. Machine learning techniques, through their ability to handle large volumes of data and uncover hidden patterns, are revolutionizing how we understand and optimize mobility in urban environments.