Integrating environmental conditions into machine learning models for predicting bridge maintenance deterioration

Abstract

Bridge management agencies face mounting pressure to maintain ageing infrastructure amidst intensifying climate extremes and limited budgets. This study develops and tests machine learning (ML) models integrating environmental data with traditional structural and inspection records to predict condition deterioration of Colorado's National Highway System bridges. An extensive database of 75,063 bridge-year observations and 97 features including deck, superstructure, and substructure condition ratings, traffic loads, freeze-thaw cycles, precipitation, temperature extremes, and humidity was compiled from the National Bridge Inventory (NBI) and PRISM climate archives (2014-2024). Decision Tree (DT), Random Forest (RF), and Gradient Boosting (GB) classifiers were trained using SMOTETomek resampling and optimized using cross-validated grid search. GB attained the most consistent top performance, achieving balanced accuracies between 0.972 and 0.978 and Matthews Correlation Coefficient (MCC) values around 0.97 across all bridge components. DTs achieved the single highest metric value for decks (balanced accuracy = 0.9875). The addition of climatic variables significantly improved performance over the baselines which has only structural variables in balanced accuracy and macro F1-scores ranging from 3.0 to 4.5 percentage points. This translated into a drop in the relative error rate by 6-15% for deterioration state forecasting. Feature importance analysis invariably revealed frequency of freeze-thaw cycles, annual rainfall, extreme temperatures, along with sufficiency rating, age, and traffic by trucks to be key predictors. A cost-benefit analysis by simulation indicated optimal allocation using these climate-aided models to generate up to 11.7% savings in life-cycle costs over a span of 30 years. The models resulting from this integration enable more accurate, risk-oriented inspection schedules and data-driven budgets while taking regional climatic stressors into account. This work proposes a structured yet flexible approach to improving climatic resilience in bridge maintenance planning