Machine Learning for Spatio-Temporal Urban Traffic Estimation

Traffic volume, the number of road users passing a street segment, is a key input for data-driven, climate-oriented mobility policy planning. However, most cities rely on only a limited number of traffic sensors, resulting in sparse and fragmented observations that fail to capture citywide traffic patterns across transport modes. Recent advances in machine learning (ML) offer new opportunities to address this limitation by combining available data sources and leveraging spatial and temporal dependencies to estimate traffic volumes at unmeasured locations. This cumulative dissertation examines complementary model-centric and data-centric ML strategies for citywide traffic volume interpolation across three empirical studies. Chapter 1 investigates which supplementary datasets, in combination with various ML models, improve bicycle volume interpolation in Berlin. By combining stationary sensor data with various datasets, such as infrastructure, motorized traffic, and crowdsourced trajectories, the study finds the latter to be especially informative. XGBoost performs best among the compared ML models, and simulations show that ten days of short-term counts at target locations can nearly halve prediction errors. Chapter 2 introduces GNNUI, a spatio-temporal Graph Neural Network for urban traffic volume interpolation, explicitly designed to address key urban challenges such as heterogeneous street networks, sparse and zero-inflated observations, and complex spatial dependencies. The model is evaluated on two publicly released datasets collected as part of this research, covering citywide Strava cycling volumes in Berlin and taxi volumes in Manhattan. Across both case studies, GNNUI consistently outperforms existing interpolation methods, with particularly strong gains under conditions of extreme sensor scarcity. Chapter 3 analyses how sensors should be placed to collect meaningful data for interpolation. Using the same Berlin and Manhattan datasets as in Chapter 2, this work evaluates commonly used sensor placement strategies. Broad spatial coverage and active learning achieve the lowest errors. The chapter also shows that short-term rotating sensors, spread across many locations and weekdays, can approach the performance of permanent sensors. Together, the three studies provide an integrated perspective on urban traffic volume interpolation: they identify valuable supplementary data, develop models suited to urban complexity, and guide sensor deployment to enable accurate, citywide estimation.