Learning a precipitation indicator from traffic speed variation patterns

It is common sense that traffic participants tend to drive slower under rain or snow conditions, which has been confirmed by many studies in the field of transportation research. When analyzing the relation between precipitation events and traffic speed observations, it was shown that by using extra weather information, road speed prediction models can be improved. Conversely, traffic speed variation patterns of multiple roads may also provide an indirect indication of weather conditions. In this paper, we attempt to learn such a model, which can detect the appearance of precipitation events, using only road speed observations, for the case of New York City. With a seasonal trend decomposition model Prophet, residuals between the observations and the model were used as features to represent the level of anomaly as compared to the normal traffic situation. Based on the timestamps of weather records on sunny days versus rainy or snowy days, features were extracted from traffic data and assigned to the corresponding labels. A binary classifier was then trained on six-month training data and achieved an accuracy of 91.74% when tested on the remaining two-month test data. We show that there is a significant correlation between the precipitation events and speed variation patterns of multiple roads, which can be used to train a binary indicator. This indicator can detect those precipitation events, which have a significant influence on the city traffic. The method has also a great potential to improve the emergency response of cities where massive real-time traffic speed observations are available.