Advances in Kriging-Based Modelling Approaches of Winter Weather Vehicular Collisions – A Region-Wide Geostatistical Investigation

  • Author / Creator
    Wong, Andy
  • The winter season is known for brisk cold weather and beautiful snowfalls, but it is also known for deteriorated driving conditions to where the risk of collisions becomes a major issue that plagues many municipalities around the world. As such, agencies are tasked with and strive to prioritize weather-related collision prone locations for an efficient mobilization of winter road maintenance services and to improve winter safety of motorists. Finding these locations is known as network screening and can be a challenge at times, requiring either a lot of data, or statistical background. Presented in this thesis is an alternative network screening method known as Regression Kriging (RK) that has recently gained some traction due to its wide applicability and excellent performance in modelling regionalized random variables. This thesis has three objectives and that is to (1) demonstrate the applicability and usability of RK models, (2) further enhance its predictive ability by substituting in network distances, and (3) characterize the underlying spatial structure of the winter collisions at various zonal scales to check the spatial continuity assumption, otherwise known as a second order stationarity. This thesis employs a case study within the state of Iowa where RK was utilized to model winter collisions using large-scale datasets of the entire state of Iowa that were collected over five (5) winter seasons from 2013 to 2018. The Winter Collision (WC) ratio was used as a surrogate measure for winter collisions as it represents a value used for relative comparison of collision sensitivity to winter conditions.
    The results from the case study resulted in some key findings. As an estimator, RK was shown to be a very effective providing predicted results that outperformed results from multiple linear regression and from ordinary kriging (OK), a precursor to RK. Five statistical measures were used to compare model performances with RK outperforming OK on all measures, though the predicted values were overestimated. In an exploratory analysis in an attempt to improve RK estimations, network distances were substituted into the kriging modelling process. Using the same five statistical measures, it was found that network distances provided marginal improvements to the predicted values, but the real improvement was in the level of uncertainty in those values. The model now underestimates the true value, but not to the extent that it had previously overestimated them, thus reducing the level of uncertainty of the values. As for the underlying spatial structure, it was found that the spatial variance that the model relies on was not continuous or stable, thus suggesting that models should be built on a zonal level to better capture unique regional spatial characteristics.
    The main contribution of this thesis can be broken down into three parts. For the first time in literature, RK has been shown to perform well as a network screening tool over a very large temporal and spatial scale. Secondly, network distances have now been shown to improve kriging results within network screening. And finally, it was determined that the spatial structure is highly sensitive to the area and its size. Applying RK over an extremely large spatial scale could possibly overlook regional factors that can affect the spatial structure as zonal analysis often gave different, often better, results.

  • Subjects / Keywords
  • Graduation date
    Fall 2021
  • Type of Item
  • Degree
    Master of Science
  • DOI
  • License
    This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.