Leveraging the spatial-temporal resolution of crowdsourced cycling data to improve the estimation of hourly bicycle volume

Abstract

Traditional methods for measuring bicycle volume have proven to be challenging and costly, especially when planning for non-motorized facilities at network level. One potential enhancement to traditional bicycle volume measurement methods is to use crowdsourced cycling data in conjunction with other data. This study explored the potential of incorporating crowdsourced data in bicycle volume estimation methods to improve the spatial-temporal coverage and resolution. To account for the potential biases associated with crowdsourced data, the study used additional predictors to enhance estimations. Different probabilistic and Machine Learning models were tested, including the Negative Binomial (NB), K-Nearest Neighbors (KNN), Random Forest (RF), Support Vector Machines (SVM) and Artificial Neural Networks (ANN). Both RF and ANN model were found to have a better prediction capability compared to other models. For the RF model, the addition of crowdsourced bicycle counts from Strava, which had an average observed penetration rate of 7 percent across study sites, improved the model significantly by increasing its ability to explain variations in hourly bicycle volume from 65 percent (R2 = 0.65) to 71 percent (R2 = 0.71). A simulation study to assess the change in model performance based on different simulated penetration rates found that a penetration rate of about 40 percent would improve prediction capability to 96 percent. To demonstrate a practical way for integrating crowdsourced data in practice, a web-based tool that can be used to estimate bicycle counts was developed. The results of this study can assist planners and engineers to make informed decisions by providing them with a quick method for estimating bicycle volume over a network when they have counts from crowdsourced data sources.