Evaluation of data imputation techniques in pavement texture processing

The importance of roads in modern society is without a doubt incomparable and nowadays, federal, state and local highway agencies are increasing the demands on the performance and serviceability of transportation infrastructure. It is no longer sufficient to have a pavement with enough structural capacity to sustain the demands of traffic. There are also growing demands to increase the functional properties of the road that are highly correlated with texture, such as skid resistance, proper drainage and smoothness. To better assess, compare and improve the functional properties of roads, there has been an effort to standardize the measurement methods of texture at highway speeds, based on surface profiles. But even standardizing the measurement methods is not enough to ultimately improve road functional properties if the processing of these profile data changes depending on who the analysist is. Therefore, meticulous studies need to be performed to determine what are the best practices when processing pavement texture data. This thesis studied the process of data imputation to determine what is the best imputation method based on their accuracy and computation time. The case study explored ten popular imputation methods, explained how they work, tested each of by means of Monte Carlo (MC) simulation, and ranked their efficiency using the Analytical Hierarchical Process (AHP). A two-tailed hypothesis test was used to make the final decision and determine whether the gain in imputation accuracy (if any) was statistically significant compared to the same statistic computed with missing data. Data imputation for texture data processing was proven to significantly increase the accuracy of estimates of texture summary statistics when a good imputation method was implemented. This study found that linear interpolation imputation was the best imputation technique not only because of its robustness and efficiency but also because of its simplicity and ease of implementation. However, it was also proven that using poor imputation techniques such as spline interpolation for gaps of missing data that are greater than ten data points can potentially yield biased estimation of pavement texture statistics that are significantly worse than simply computing that statistics using the data with missing entries