Quantifying cities through user-generated geosocial content with GeoAI

Restricted (Penn State Only)
Author:
Liu, Xi
Graduate Program:
Geography
Degree:
Doctor of Philosophy
Document Type:
Dissertation
Date of Defense:
June 18, 2019
Committee Members:
  • Clio Maria Andris, Dissertation Advisor/Co-Advisor
  • Clio Maria Andris, Committee Chair/Co-Chair
  • Alan M Mac Eachren, Committee Member
  • Donna Peuquet, Committee Member
  • Bruce A Desmarais, Jr., Outside Member
Keywords:
  • GIScience
  • GeoAI
  • Cities
  • Computational Social Science
  • Machine Learning
  • Data Mining
Abstract:
Massive datasets describing human activities in cities, such as traveling for work, calling friends, and posting texts/images on social media apps, are being collected thanks to the development of information and communication technologies. In the last decade, the GIScience community has leveraged geospatial information in these "big data" to quantify cities based on how people interact with the built environment through their movements, connections, and perceptions. Recently, artificial intelligence (AI) has obtained impressive achievements with advanced deep learning techniques, revolutionizing research areas such as image classification, machine translation, and game playing. Integrating artificial intelligence with GIScience is now a research frontier in geography and has the potential to reveal knowledge of cities that has never been discovered before. In this dissertation, I propose three major tasks in geospatial artificial intelligence (GeoAI) research: efficiently adopting developed AI tools for geographical problem-solving, developing new GIScience models by incorporating deep neural networks, and aligning the usage of GeoAI with human values. Three studies were conducted using unconventional datasets and creative approaches to study cities, while addressing the challenges of these major research directions. The first study investigated the spatial patterns of residential ornamentation, both globally and locally, using millions of Airbnb images and transfer learning techniques. The second study proposed a representation learning-based model to measure colocation patterns of points of interest (POIs) integratively and quantified the regional variability of spatial colocation patterns across cities. The third study illustrated the correlation between the diverse human needs served in a neighborhood and its economic condition, which also addressed the potential biases in GeoAI research. The main contribution of this dissertation is twofold: first, it reveals new knowledge and patterns about cities with innovative GeoAI methods, which are insightful for improving our urban life; second, it provides showcases of how geographers can leverage domain knowledge to better formulate GeoAI research, tackling critical issues in the field.