Assessing regression methods to estimate network-wide bicycle traffic volumes based on crowdsourced GPS and permanent counter data

GPS data offer an up-to-date, available, and easily processable database for bicycle traffic planning. Unlike permanent counters, they generally represent wide parts of the bicycle network. However, GPS data is derivable only from a subset of the cycling population and thus provides a limited overview of existing bicycle traffic volumes in a city at best. For planning or dimensioning of cycling infrastructure the data is only partially sufficient. Values such as the (annual) average daily number of bicycles (ADB/AADB) are more suitable. Using regression methods, GPS data in combination with (permanent) counter data can be utilized to model network-wide ADB. So far however, related studies mostly deal with only few counters in individual cities or metropolitan regions. Due to different modelling approaches and input variables, the results are neither comparable nor transferable. Therefore, no conclusion as to which models are most suited can be drawn. This study investigates the extrapolation of GPS data from a nationwide data set in Germany. First, six different types of regression models are trained based on the data set. Second, the trained models are utilized for network-wide AADB estimation in six municipalities. Thereby, this study provides a framework for comparable error metrics and investigates the suitability of the tested models for (1) estimation at permanent counters and (2) network-wide estimation. The models are divided into three classes: linear, tree-based and neural network models. We used 452 data points from permanent counters across Germany for model training. After assessing the model performances at the counters, they are applied to municipality-wide network sections. Comparing the overall performance, Support Vector Regression currently proves to be the most promising for extrapolating traffic volumes from GPS data to network-wide AADB.