Learning-based geospatial data analysis for urban traffic and transportation improvement

In this modern era, geospatial data have experienced massive growth. Research in geospatial data encompass all applications where spatial coordinates of entities are recorded and analyzed, but they are mostly focused on urban applications. Two topics in general, urban traffic and transportation improvement both receive substantial attention. Both urban traffic and transportation improvement consist of many tasks. One of these tasks is similar trajectory search, a fundamental geospatial task. Similar trajectory search cannot stand by itself and there are different tasks specific to urban traffic and public transport that add to the completeness of the research. For urban traffic, one of the most important task is traffic prediction, which seeks to predict future traffic based on historical traffic data. For urban transport, public bus routing and scheduling both play an important part; and is usually combined under the umbrella term ``bus network optimization". The increasing volume, variety, and accessibility of traffic and transportation-related datasets, coupled with the increasing prominence of deep learning methods that can learn complex patterns from large input datasets, have enabled breakthroughs in these three tasks. The insights gained from this research has benefitted society as a whole. Despite these advances, there are existing research gaps, whether it is from the perspective of introducing new and novel problems, improving the efficiency and effectiveness of current solutions to existing problems, or to include the commonly overlooked social equity factors. In this thesis, our mission is to identify and fill the research gaps of these three tasks to advance the research fields of urban traffic and transportation improvement as a whole. In the case of similar trajectory search, we identified that current research only explores spatial-only trajectory search, overlooking the important temporal factors. The addition of temporal information not only increases accuracy, but also provides an extra dimension for downstream tasks, allowing for more complex and useful search queries. Thus, we explore the topic of spatio-temporal similar trajectory search. We utilize a learning-based approach that not only provides better accuracy compared to existing approaches, but also enables faster querying time and robustness against common dataset errors. In the traffic prediction task, we found that current research only focus on short-term traffic prediction that is limited to few-hours prediction at most. Thus, we address the research gap of long-term traffic prediction, extending the prediction horizon from few-hours to one-day ahead prediction. This longer horizon provides crucial information for the public and government alike, allowing them to take more proactive actions to combat issues such as traffic congestion. We introduce a state-of-the-art deep learning model and show that our method is more effective than the existing literature on this task. Finally, in the bus network optimization task, we tackle an unexplored task of equitable bus network optimization. This is an important task, as inequities in bus network optimization have adverse impact on economically, socially or geographically disadvantaged population. We formulate the first-ever notion of bus network equity, and propose a new, comprehensive bus network efficiency metric that takes into account four important route network quality factors: route directness, service coverage, network efficiency and number of transfers required. We then explore the intricacies and difficulties of the bus network optimization problem through experiments on our novel optimization algorithms based on deep reinforcement learning.