Predicting traffic volumes on road infrastructures in the context of multi-risk assessment frameworks

In multi-risk assessment frameworks involving road infrastructures, measures of exposure to natural hazards include traffic volumes. However, traffic counts are usually collected through traffic counter/radar stations which only cover a small part of the road network. In this study, country-wide Annual Average Daily Traffic (AADT) prediction models based on Italian data were developed to provide direct risk exposure measures both in terms of traffic volumes (continuous variable) and traffic volume discrete classes, using province-/municipality-related geographic, socio-economic and road-related variables as predictors. To ease transferability and applicability of the models, only publicly available predictors were selected. Traditional statistical techniques (generalized linear models for predicting traffic values and ordered logistic models for traffic classes) and Machine Learning (ML) approaches (XGBoost for both regression and classification problems) were used. Both the direct estimation of traffic volumes and the classification into traffic ranges provided satisfactory results in terms of goodness-of-fit and predictive accuracy metrics. Results show that population, occupation, tourism, density, number of lanes, urban environment, complex intersections and ring roads were generally related to an increase in traffic volumes. Distance from large cities and accessibility metrics are inversely related to traffic instead. The application of the XGBoost ML approach proved to be more accurate than traditional approaches only for heavy vehicles. It was discussed how the obtained models can be used as input modules for overall multi-risk assessment frameworks involving road infrastructures.