Better Routing in Developing Regions: Weather and Satellite-Informed Road Speed Prediction

Inaccurate digital road networks significantly complicate the use of analytics in developing, data scarce, environments. For routing purposes, the most important characteristic of a digital road network is the information about travel times/speeds of roads. In developing regions, these are often unknown, and heavily dependent on the weather (e.g., rainfall). This may, for instance, cause vehicles to experience longer travel times than expected. Current methods to predict the travel speeds are designed for the short upcoming period (minutes or hours). They make use of data about the position of the vehicle, the average speed on a given road (section), or patterns of trafic flow in certain periods, which are typically not available in more developing regions. This paper presents a novel deep learning method that predicts the travel speeds for all roads in a data scarce environment using GPS trajectory data and open-source satellite imagery. The method is capable of predicting speeds for previously unobserved roads and incorporates specific circumstances, which are characterized by the time of the day and the rainfall during the last hour. In collaboration with the organization PemPem, we perform a case study in which we show that our proposed procedure predicts the average travel speed of roads in the area (that may not exist in the GPS trajectory data) with an average RMSE of 8.5 km/h.