Location-Aware Machine Learning to Forecast Urban Foot Traffic

Smart cities increasingly rely on intelligent tools to manage urban mobility, reduce congestion, and support evidence-based public governance. This study examines how machine learning (ML) models can support urban mobility management by predicting footfall and multimodal traffic flows. Five ML models were trained and evaluated on a dataset of 261,539 hourly observations collected from a 16-sensor monitoring network across four urban locations in Turku, Finland. The modelling framework integrates temporal, meteorological, socio-economic, and spatial covariates to capture location-specific mobility flows. To prevent temporal leakage, model evaluation used a strict chronological train–test split, supported by five-fold time-series cross-validation; all metrics are reported with empirical standard deviations (±SD). XGBoost recorded the strongest predictive performance (R² = 0.720 ±0.071 in the City Centre, 0.839 ±0.099 in the University corridor, and 0.570 ±0.105 in the cultural district). Location 4, a low-volume remote sensor, yielded R² = 0.179 ±0.231, reflecting the inherent predictability limits of a near-zero-volume sensor (avg. 2.4 pedestrians/h). The findings guide sensor deployment planning, real-time mobility management, and data infrastructure design in smart city contexts.