Presentation Information

[AI-10]Graph Convolutional Networks for Network-Level Forecasting of Pavement Surface Condition Indicators

*Artem Yushchenko1, Jan-Hinrik Borchers1, Léo Francoso Dal Piccol Sotto1, Bernardo Mota Lontra2, Gina Torres-Alves2 (1. Federal Highway and Transport Research Institute (BASt), Bergisch Gladbach (Germany), 2. Netherlands Organisation for Applied Scientific Research (TNO), Delft (Netherlands))

Keywords:

pavement management,pavement surface condition forecasting,machine learning,graph learning,Graph Convolutional Networks

Forecasting of pavement surface condition indicators at the network level supports the decision making of highway asset managers in terms of maintenance strategy and cost efficiency. Data-driven methods based on machine- and deep learning have been shown to improve forecasting results compared to traditional approaches due to their ability to discover complex patterns from historical data. The inherent network structure of interconnected highways additionally allows spatial patterns to be considered when modeled accordingly. This study expresses the highway network as a graph with individual highway segments representing graph nodes and examines the effectiveness of Graph Convolutional Networks (GCN) on the forecasting of pavement surface condition indicators. The model is evaluated on the task of forecasting rutting and unevenness in asphalt pavements with a four-year horizon based on the German highway network 100m segment dataset. In addition to historical surface condition measurements, input data includes estimated traffic volume, pavement structure, as well as meteorological data. Current GCN optimization trends are utilized to achieve optimal performance. Stratified 5-fold cross validation is performed to ensure the robustness of the experiment results. The GCN model exhibits comparable forecasting performance relative to contemporary approaches that forgo modeling the spatial structure of a highway network, hinting at the potential relevance of discovering the data's inherent spatial patterns. The effectiveness of the GCN model is additionally validated via ablation studies and the investigation of graph connectivity’s influence on model performance. This study’s results demonstrate that incorporating the spatial structure of highway networks through GCNs can provide more accurate forecasts of pavement surface condition indicators. These findings highlight the potential of graph-based learning to enhance data-driven pavement management systems and support informed, cost-effective maintenance strategies at the network level.