Methodology of Large-Scale Annual Average Daily Bicycle Traffic Estimation

Large-scale bicycle network design, safety, and crash estimates largely depend on network-wide bicycling volume estimation. Previous studies developed direct demand models on a smaller scale (e.g., city or county) where bicycle count and seasonality did not vary widely between count locations. The applicability of such smaller-scale methods is questionable for large-scale applications. This study aims to develop a technique that can be implemented in large-scale bicycle volume estimation. The study considered the State of California and used data from 187 unique permanent counters to develop three different types of expansion factors: hour of day for both weekdays and weekends, day of week, and month of year factors. The developed expansion factors were applied for imputing 2,589 (3,559 intersection legs) short-term site locations to estimate the annual average daily bicycle traffic (AADBT). Finally, the estimated AADBT from both permanent and short-term sites was used to develop Poisson and random forest (RF) regression model. The results indicate that the RF model performs [root mean square percentage error (RMSPE)=69] better than the Poisson regression model (RMSPE=110). The RF model RMSPE is 40 for high (>300) and 50 for mid-volume sites (101–300 AADBT); however, low-volume sites (≤100 AADBT) are overestimated with high RMSPE (129). The results also indicate that the crowdsourced (Strava) volume, bike and vehicle ownership status, intersection density, urban/rural status, location of bike parking, and counters within Yolo Country are crucial predictors of AADBT estimation. The developed methodology will provide planners and engineers guidelines for large-scale network-wide bicycle volume estimation.