Spatiotemporal analysis of urban traffic crash risk using a bagging-optimized dynamic mode decomposition framework.

Abstract

Objective: This study aims to analyze the spatiotemporal evolution of traffic crash risk in Manhattan and to improve one- to seven-day-ahead crash prediction through a Bagging Optimized Dynamic Mode Decomposition (BOPDMD) framework. By examining two full years of daily crash data (2019-2020), the study further investigates how latent crash patterns differ between a typical pre-pandemic year and a disruption-dominated pandemic year.

Methods: Daily crash counts from 69 Manhattan neighborhoods were aggregated into zone-day matrices for 2019 and 2020, as the daily scale provides a practical compromise between short-term responsiveness and data stability for neighborhood-level crash modeling. A unified preprocessing pipeline was applied, including square-root variance stabilization and zone-wise standardization. For prediction experiments, a fixed 212-day window from April 1 to October 30 was used in each year to ensure identical sample length. Within this window, the first 160 days were used for training and the remaining 52 days were used for testing. For interpretation, BOPDMD was applied to the complete 2019 and 2020 matrices, with 365 and 366 days respectively, to extract spatial modes, temporal coefficients, and modal frequencies that characterize underlying crash dynamics. Its forecasting performance was compared with standard Dynamic Mode Decomposition (DMD) and several representative baseline models under one- to seven-day prediction horizons.

Results: BOPDMD achieved the lowest or near-lowest Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) across most prediction horizons in both years and exhibited the slowest error accumulation in multi-step forecasting. The spatiotemporal mode analysis revealed clear cross-year differences. In 2019, dominant modes captured stable high-risk corridors and regular weekly and seasonal oscillations, indicating a mobility-driven and quasi-periodic crash regime. In contrast, the 2020 modes exhibited abrupt spatial reconfiguration, rapid temporal decay, and weakened periodic structure, reflecting pandemic-induced disruptions in travel demand and risk allocation. Eigenvalue patterns confirmed that 2020 dynamics were more transient and less cyclic than those of 2019.

Conclusions: The findings demonstrate that BOPDMD provides both accurate one- to seven-day-ahead crash forecasts and interpretable representations of underlying risk dynamics. The revealed modal structures highlight how urban crash risk shifts between stable mobility patterns and externally driven disruptions. These insights can support proactive safety management by enabling zone level risk monitoring, prioritization of high risk areas, and the design of targeted interventions for both persistent hotspots and disruption induced risk shifts in complex urban environments.