Multimodal Deep Learning Framework for Traffic Volume Estimation: A City-Scale Case Study in Edmonton, Alberta

Traffic volume is essential for transportation planning, safety analysis, and infrastructure investment. Achieving comprehensive spatial coverage across road networks faces budgetary and logistical constraints, necessitating reliable estimation for unmonitored segments. This study addresses this gap by developing a multimodal deep learning framework that fuses aerial imagery with parametric data to estimate annual average weekday traffic (AAWDT). Using 749 ground-truth locations spanning over 5,000 km of roadway, a citywide case-study is conducted in Edmonton, Alberta, representing one of the most spatially extensive and heterogeneous applications in literature to date. The proposed model is benchmarked against several established methodologies including multivariate linear regression, support vector regression, random forest, XGBoost, and neural networks. Results demonstrate that the framework reduces Mean Absolute percentage Error (MAPE) by 7%-85% relative to these models. Thus, the research yields a robust methodology that integrates multiple data sources to assist transportation agency decision-making by providing city-wide traffic volume estimation.