Improving network level pedestrian activity prediction by accounting for spatial and longitudinal clustering in crowdsourced data

Abstract

Crowdsourced data are increasingly used to estimate active transportation volumes across large networks. Although bicycling data are widely available, comparable pedestrian data have been limited. Strava Metro recently began releasing link-level pedestrian activity data, but evaluations of these data remain scarce. This study is among the first to validate Strava’s pedestrian activity data against at-location counter measurements across Michigan and integrate them into a system-wide pedestrian volume model. We address a critical methodological gap where prevalent models ignore spatial and longitudinal clustering in crowdsourced datasets. By implementing a Mixed-Effects Gradient-Boosted Fixed Effects (MERF-XGB) predictive framework, we capture these hierarchical clustering structures to improve estimation accuracy. The tuned MERF-XGB model achieved high predictive performance, with a test RMSE of 100.32 and an R2 of 0.87. While predictive performance was comparable to a standard global model with cluster indicators, the mixed-effects formulation offers a more scalable representation of hierarchical clustering. SHAP analysis further reveals substantial spatial heterogeneity in feature contributions across different municipalities. Our findings introduce a validated pedestrian data source and a robust, generalizable predictive framework for network-level pedestrian volume estimation, which can guide active transportation planning and infrastructure investments.