A spatiotemporal deep learning approach for citywide short-term crash risk prediction with multi-source data

The primary objective of this study is to investigate how the deep learning approach contributes to citywide short-term crash risk prediction by leveraging multi-source datasets. This study uses data collected from Manhattan in New York City to illustrate the procedure. The following multiple datasets are collected: crash data, large-scale taxi GPS data, road network attributes, land use features, population data and weather data. A spatiotemporal convolutional long short-term memory network (STCL-Net) is proposed for predicting the city-wide short-term crash risk. A total of nine prediction tasks are conducted and compared, including weekly, daily and hourly models with 8 x 3, 15 x 5 and 30 x 10 grids, respectively. The results suggest that the prediction performance of the proposed model decreases as the spatiotemporal resolution of prediction task increases. Moreover, four commonly-used econometric models, and four state-of-the-art machine-learning models are selected as benchmark methods to compare with the proposed STCL-Net for all the crash risk prediction tasks. The comparative analyses suggest that in general the proposed STCL-Net outperforms the benchmark methods for different crash risk prediction tasks in terms of higher prediction accuracy rate and lower false alarm rate. The results verify that the proposed spatiotemporal deep learning approach performs better at capturing the spatiotemporal characteristics for the citywide short-term crash risk prediction. In addition, the comparative analyses also reveal that econometric models perform better than machine-learning models in weekly crash risk prediction tasks, while they exhibit worse results than machine-learning models in daily crash risk prediction tasks. The results can potentially guide transportation safety engineers to select appropriate methods for different crash risk prediction tasks.