Improving Spatiotemporal Transferability of Real-Time Crash Likelihood Prediction Models Using Transfer-Learning Approaches

Abstract

A real-time crash likelihood prediction model is an important component of the proactive traffic safety management system. Over the past decades, numerous models have been proposed for predicting real-time crash likelihood and achieved promising results. However, most studies have ignored the model transferability, especially for deep-learning models. The transferability of a model could be defined as applying the model to new data which obtain from different locations or periods. The purpose of this study is to improve the spatiotemporal transferability of a deep-learning crash likelihood prediction model by using transfer-learning approaches. Trajectory and crash data from five arterials in Florida were used in this study. A two-layer long short-term memory (LSTM) model was developed to predict real-time crash likelihood. Two scenarios were created for investigating spatial and temporal transferability of the developed model. Experimental results suggested that the model could be accurately transferred to new data by using the fine-tuning transfer-learning approach. In the framework of this research, the transferred models achieved higher predictive accuracy than models which were directly developed on the new data. Moreover, the spatial transfer learning outperformed the temporal transfer learning in predictive accuracy. For example, when the percentage of test data was 70%, the spatial transferred model had an average area under the curve (AUC) of 0.74. while the temporal transferred model had an average AUC of 0.65. Results from this study could be applied to transfer pretrained crash likelihood prediction models to new locations when few crashes are available.