Comparison of Road Segment-Level Bicycle Volumes Estimation Models: Traditional Regression, Machine Learning, Large Language Models, and Novel Spatial Regression Approaches

Estimating road segment-level bicycle volumes is essential for exposure-based safety analysis and infrastructure planning for vulnerable road users. Existing approaches often require extensive data collection, including short-term counts expanded using continuous counts or direct-demand models based on roadway, socioeconomic, and land-use characteristics. Recent methods incorporating machine learning, deep learning, and crowdsourced data have shown promise, but are often limited by data bias and inconsistent availability across jurisdictions. This study evaluates several benchmark methods, in addition to newly proposed approaches based on spatial proximity effects, including linear interpolation, Bayesian spatial negative binomial, random forest, large language models, graph convolutional networks (GCNs), and spatial lag regression (SLR). A dataset consisting of 1,127 road segments with a full year of bicycle count data (2025), along with site characteristics from the cities of Waterloo and Kitchener, Canada, was used to compute seasonal average daily bicycle volumes. Model performance was assessed using root mean squared error and mean absolute error across varying levels of data availability. The results indicated that the proposed SLR model consistently outperformed all other models across all data availability levels, including the GCN model. These findings demonstrate that incorporating spatial proximity and roadway characteristics can improve estimation accuracy without requiring large training datasets or complex model structures, providing a practical and interpretable solution for bicycle volume estimation in data-constrained environments.