Predicting Culvert Deterioration Using Physical and Environmental Time-Independent Variables

Given a vast inventory of culverts with different physical and environmental characteristics, the task of condition assessment can become difficult or impossible. Using a database of approximately 8,000 culverts in South Carolina with varying sizes, types, and configurations, a model was produced to predict the condition of these culverts. This model combined only the physical culvert information given in the database with mapped environmental characteristics including historical temperature, precipitation, pH, and estimated runoff coefficient. The resulting model used the combination of inputs that produced the model with the best performance measures, primarily the area under a receiver operating characteristic curve. A separate model was created for each of the 6 culvert types commonly found in South Carolina and the 10 condition categories given to each culvert in the database: cracking, separation, corrosion, alignment, scour, sedimentation, vegetation, erosion, blockage, and piping. Both a logistic regression model and an artificial neural network model were used to combine the input variables into a prediction of the desired output variables. The models produced were shown to have a coefficient of determination of between 0.25 for poorly correlated models and 0.80 for better correlated models when comparing the predicted culvert score with the actual culvert score. With the ability to predict culvert health without inspection or any idea of the age of the culvert, the culverts in need of assessment can be prioritized without comprehensive inspections or more detailed information about the culverts.