Optimizing Visual and Automated Pavement Condition Surveys Using Analysis of Transverse Variability and Artificial Neural Networks (05-1678)
Alan Reggin, EBA Engineering Consultants, Ltd., Canada
Ahmed Shalaby, University of Manitoba, Canada

The paper deals with two approaches to optimizing pavement condition surveys for the urban pavement network of the City of Winnipeg, Manitoba, Canada. First, a non-parametric statistical test is made on a randomly selected set of visual survey data in order to assess the transverse variability of the data. The test compares the ratings for one lane with the rating of all lanes of each segment. While the medians are considered the same with 92% confidence, there were some observed biases in the data, which can be eliminated if the surveyed lane is selected randomly. The second approach is to use artificial neural networks to predict visual survey scores from automated (laser-based) measurement of rut depth and international roughness index (IRI). The training data consisted of 80% of parallel visual and automated surveys conducted over the entire network while the remaining 20% of the data is used for testing. Two neural network architectures are examined and the number of neurons in the hidden layer is optimized to minimize the root mean square of errors. A resilient back propagation algorithm is selected and the Kappa coefficient is used to examine the strength of agreement. The results showed that only fair to moderate agreement was achieved and that additional data elements are required to improve the predictive power of the neural network models.