| Summary: | Today, urban computing has emerged as an interdisciplinary field connecting data science and urban planning, reflecting the growing integration of urban life with advanced computational methods. Urban computing has particularly benefited from deep learning owing to the spatiotemporal and multi-modal nature of data emerging from urban systems. Deep learning models have not only boosted predictive accuracies beyond traditional models, but also adeptly handled unstructured data. However, the application of deep learning in urban system analysis has a lot of challenges. Within the vast and complex scope of urban computing combined with deep learning, this dissertation zooms in on three emerging issues: uncertainty quantification, data fusion, and generative urban design while focusing on transportation systems and urban planning applications.
The first part of this dissertation proposes a framework of probabilistic graph neural networks (Prob-GNN) to quantify spatiotemporal uncertainty. This Prob-GNN framework is substantiated by deterministic and probabilistic assumptions and empirically applied to predict Chicago's transit and ridesharing demand. Prob-GNNs can accurately predict ridership uncertainty, even under significant domain shifts such as the COVID-19 pandemic. Among the family of Prob-GNNs, two-parameter distributions (e.g., heteroskedastic Gaussian) achieve the highest predictive performance, which is 20% higher in log-likelihood and 3-5 times lower in calibration errors compared to the one-parameter baseline distributions.
The second part addresses data fusion, in which a theoretical framework of deep hybrid models (DHM) was created to combine the numeric and imagery data for travel behavior analysis. DHM aims to enrich the family of hybrid demand models using deep architectures and enable researchers to conduct associative analysis for sociodemographics, travel decisions, and generated satellite imagery. Empirically, DHM is applied to analyze travel mode choice using the Chicago MyDailyTravel Survey as the numeric inputs and the satellite images as the imagery inputs. DHM can construct latent spaces that significantly outperform classical demand and deep learning models in predicting aggregate and disaggregate travel behavior. Such latent spaces can also be used to generate new satellite images that do not exist in reality and compute the corresponding travel behavior and economic information, such as substitution patterns and social welfare.
The last part develops a human-machine collaboration framework for generative urban design and then instantiates the framework with a model trained to generate satellite imagery from a land use text description and a constraint image depicting the unaltered major road networks and natural environments. The trained model can generate high-fidelity, realistic satellite images while retaining control over the land use patterns in generated images with natural language descriptions, producing alternate designs with the same inputs, respecting the built and natural environment, and learning and applying local contexts from different cities.
|
|---|