Assessing Bikeability in Virginia: A Comparison of CHATGPT-4o and Traditional Models

Abstract

This study evaluates cycling friendliness by using Street View Images (SVIs) and deep learning models to assess street-level features associated with the cycling environment in Virginia. An image segmentation model, PixelLib, was used to extract the proportion of seven features, namely greenery, streetlights, roads, sidewalks, cars, sky, and buildings, from 1,727 Google Street View Images (SVIs). The estimated features were also used to calculate the scene complexity level (CL), reflecting the visual navigability of each location. Results revealed that many SVIs exhibit a moderate CL, indicating that the cycling environment offers clear visual conditions that may improve safety for cyclists. The CL, combined with three other factors: greenery, traffic volume (Average Annual Daily Traffic), and bike lane types, were used to develop the Bikeability Index (BI). The distribution of the BI varied across Virginia, with classes ranging from low to very high. Higher BI pointed to more bicycle-friendly environments, while lower BI indicated areas with conditions less supportive of cycling. The study also demonstrated the capability of a large language model (LLM), GPT-4o, to support transportation analysis through the automated classification of bike lane infrastructure. The model achieved a classification accuracy of 98.28% when used to categorize 294 SVIs into separated and shared bike lanes. In addition, GPT-4o was used to generate bikeabillity rating for ten SVIs based on a structured prompt. A correlation analysis revealed a weak but statistically significant negative relationship between the BI and the CDC's Social Vulnerability Index (SoVI), suggesting that socially vulnerable areas tend to have less bikeable environments.