Integrating coastal segmentation on predictive modeling for data-driven bridge asset management

The number of deteriorating bridges is increasing, particularly in coastal areas that are frequently exposed to strong winds and severe storms, presents critical challenge for infrastructure asset management. This study introduces a machine learning-based framework for automating the prediction of bridge conditions, focusing on three key components: substructure, deck, and superstructure. To account for environmental variability, the dataset was segmented into coastal and inland subsets. The Boruta algorithm was applied to identify significant factors. Eight machine learning models, K-nearest neighbors, Artificial Neural Networks, Support Vector Machine, Random Forest, Decision Tree, Catboost, XG-Boost, and Gradient Boosting Machines, were subsequently developed for bridge condition prediction. Among the eight machine learning models, the KNN model achieved the highest prediction accuracies of 92.38% for substructure, 91.48% for deck, and 90.45% for superstructure. Optimal segmentation distances were identified as 12 km for the substructure, 16 km for the deck condition prediction model, and 6 km for the superstructure condition prediction model. The t-test results confirmed statistically significant differences between coastal and inland models, emphasizing the varying impacts of environmental factors on bridge conditions. This study provides transportation agencies a reliable, data-driven decision support tool for maintenance planning, contributing to more efficient and targeted bridge asset management.