Two methodologies to calibrate landslide runout models

Abstract

Extremely rapid, flow-like landslides pose a significant hazard worldwide; however, the analysis of the impact area and velocity of these flows is not routine. Semi-empirical numerical models are one tool that is available for performing this sort of analysis. These models are physically based; however, certain input parameters are determined through model calibration, using back-analysis of real landslide cases. Objective, repeatable calibration methods are needed for this approach to be useful for landslide runout prediction. The present analysis describes the application of optimization theory and Bayesian statistics to calibrate these types of models. Two complementary methods are presented. The first uses the Gauss-Marquardt-Levenberg optimization algorithm to efficiently determine a set of best-fit calibrated model parameters. The second uses a posterior analysis to quantify errors associated with parameter calibration, which can then be used for probabilistic forward analysis. Three case histories are presented to demonstrate how the new methods are able to rapidly calibrate a runout model, reduce subjectivity inherent in the calibration process, and provide information on parameter uncertainty.