Urban visual AI for travel behavior prediction: A comparative analysis of satellite and street view imagery

The increasing availability of satellite and street-view images offers new opportunities for visual AI in commuting mode share analysis. However, few studies have systematically examined how different visual representations of the built environment encode census tract-level commuting mode shares. This study evaluates the predictive and interpretative performance of satellite and street-view imagery for modeling census tract-level commuting mode shares in Florida, USA. We apply semantic segmentation to extract land-use and streetscape features from both imagery types and integrate them with socio-demographic data in multiple machine learning models. We further apply SHAP (SHapley Additive exPlanations) to interpret how visual features contribute to model predictions across commuting modes. Results show that satellite imagery generally outperforms street-view imagery in terms of out-of-sample R2, although the magnitude of improvement varies across models, travel modes, and satellite zoom levels. For the tree-based models, the relative improvement of satellite imagery over street-view imagery typically ranges from approximately 2% to 8%. Lower zoom levels with a broader spatial context better predict driving and transit use, whereas higher zoom levels with finer details improve walking and biking prediction. Despite the performance differences, SHAP analysis reveals largely consistent predictive associations: lower-density and more vegetated environments are associated with higher car dependence, while denser and pedestrian-oriented environments are associated with higher transit and active mode shares. This study demonstrates the value of interpretable GeoAI in linking visual urban form to commuting mode shares. The results highlight the importance for planners of using both satellite and street-view imagery, as the two data sources offer complementary capacities for interpreting neighborhood-level commuting mode-share patterns. The proposed approach can help practitioners compare neighborhoods and flag areas where commuting patterns or visible built-environment conditions warrant closer transportation or land-use analysis.