Deep Learning-Based Spatiotemporal Modeling of Bicycle Crash Risk Using Mobile Sensing and Urban Built Environment Data

Abstract

Bicycle transportation is increasingly recognized as a sustainable urban mobility mode, yet pervasive safety concerns inhibit realization of its full societal potential. Conventional crash analysis methods rest on aggregate police-reported datasets that suffer from underreporting, spatiotemporal sparseness, and limited characterization of near-miss events. This paper presents a system-level investigation of deep learning architectures designed for spatiotemporal bicycle crash risk modeling by integrating heterogeneous streams of mobile sensing data and detailed urban built environment information. The discussion is oriented around structural trade-offs inherent in sensor fusion, computational graph construction, attention-driven memory mechanisms, and adversarial robustness. Particular emphasis is placed on the governance of large-scale mobility data pipelines, including privacy preservation, algorithmic fairness, infrastructure interoperability, and long-term deployment sustainability. Drawing from cross-domain comparisons with motor vehicle crash prediction and urban flow forecasting, we analyze the interplay between model complexity, interpretability, and real-world policy actionability. We further examine how varying spatial resolutions, temporal update frequencies, and built environment feature encodings influence distributive equity in risk assessments across sociodemographic groups. The paper concludes with an exploration of anticipatory governance frameworks that reconcile the promise of proactive infrastructure safety interventions with the public accountability demands of municipal planning systems, offering a forward-looking perspective on deep learning as an infrastructural component for sustainable urban safety systems.