Geostatistically based pavement performance prediction using universal kriging and artificial neural networks

Pavement performance is most often defined as the deterioration of a pavement condition rating over time. Thus, most empirically based pavement performance prediction models are primarily functions of time. Other parameters such as traffic volume or pavement type are used to group pavements into families. A family is a collection of pavements that are perceived to perform similarly. This method of determining pavement performance prediction models ignores the fact that a pavement is located at a non-random point in space.; The purpose of this study was to develop a pavement performance prediction method that accounts for both spatial and non-spatial variability. To develop such a model the definition of performance prediction was altered slightly. Instead of predicting the condition of a pavement at some point in time, the prediction of the entire pavement condition curve over time was used. Thus, the performance of a pavement was expressed as a single measure that can vary from both spatial and non-spatial factors.; By using this definition of performance, universal kriging can be implemented. Universal kriging is a type of geostatistics that predicts spatial phenomenon by decomposing the problem into three parts. First, pavement performance prediction models are developed to remove any identifiable trend from the data, or detrend the data. Second, the spatial autocorrelation of residual values are used to generate a residual response surface. Third, the predicted response from the trend is added to the predicted residual to determine the actual predicted value. Therefore, it was believed that universal kriging would enhance the predictability of pavement performance, because pavement performance has both spatial and non-spatial factors.; To determine if universal kriging could be implemented for predicting pavement performance, this study used data collected on Ohio's multi-lane divided highway system. To detrend the data, both multiple regression analysis and artificial neural networks were used. It was found that when using the statistical approach, accounting for spatial autocorrelation increased the abilities of regression to predict pavement performance. When using artificial neural networks there existed no spatial autocorrelation among the residuals. Also, artificial neural networks were more tolerant of noise in the data, but regression was found to be better at describing the trend. (Abstract shortened by UMI.)...