The Development of a Site Specific Early Warning System for Rainfall Induced Landslides

Abstract

This thesis describes the development of a site specific early warning system for rainfall induced landslides. In New Zealand alone, rainfall induced landslides cause millions of dollars’ worth of damage annually. Worldwide the toll is much greater, with thousands of fatalities and billions of dollars’ worth of damage annually. The early warning system developed in this research is a means to mitigate the risk of such landslides. This thesis describes the underlying theory, the methodology used and results of the development of this early warning system. The early warning system is site specific, and based on the assumed failure mechanism of a slope subject to rainfall events of variable magnitude and intensity including prolonged events. The selected site, which the prototype of this early warning system was developed for, is located in Silverdale, New Zealand. A slope was formed at the site by a previous road cut operation to form State Highway One, which lies at the toe of the slope. A landslide occurred at the site in the winter of 2008 following a period of prolonged rainfall. The soil at the site consists of residual soil weathered from the Northland Allochthon formation. Previous research and experience within this soil group suggests it is particularly susceptible to rainfall induced landslides. A variety of laboratory tests were undertaken in this research to better understand the shear strength characteristics of the soil. The results obtained from consolidated drained and constant shear stress drained triaxial tests indicate that this soil may exhibit different shear strength parameters depending on the stress path associated with failure. Many site specific early warning systems have been developed in the literature, however they are usually based on explicitly stating a level which a measured parameter (pore-pressure for instance) must reach before a warning is given. The aim of this research was to develop an early warning system which alerts the user (a) to decrease of the factor of safety to a level defining overall failure or landslide occurrence and (b) the time-frame in which this failure may happen. For the purpose of this research, a factor of safety of one is defined as overall failure. This early warning system utilises field monitoring to determine the factor of safety of the slope against slope failure. Two sites are referred to throughout this thesis. One is the landsite site; the location of the 2008 landslide event. The other is the monitored site. Volumetric water content sensors were installed at various depths and locations along the same cross section of the site, approximately 40m away from the 2008 landslide site. A tipping bucket rain-gauge was used to monitor rainfall events. The rainfall record captured at the site was input into a finite element model (SEEP/W) to replicate the fluctuating water content observed at the monitored site. Once a good agreement was obtained, the matric suction/pore-water pressure profile was coupled with a limit equilibrium analyses (SLOPE/W). Thus, the factor of safety at each time step of the finite element model was obtained. An artificial neural network was trained to predict this factor of safety, using the corresponding readings of the volumetric sensors at the site as inputs. Next, the rate of change of this factor of safety (change of factor of safety with respect to time during a rainfall event) was used to estimate the time until failure. Finally, another artificial neural network was trained to predict the factor of safety at the site in the future, using rainfall forecasts for the site. The user of the early warning system can then use these predictions of the time until failure as a basis for taking any necessary action. For the given monitored site, it is recommended that if failure is predicted to occur within 5 hours, then the warning should consist of lowering speed limits. If failure is predicted to occur within 1 hour, than the warning should consist of diverting traffic to avoid the landslide site. The finite element model was reasonably successful at replicating the observed water content fluctuations in the field. A factor of safety of one was obtained for the 2008 landslide site, using the rainfall record leading up to the landslide failure (using the stratigraphy at the location of the 2008 landslide), verifying the modelling process. The artificial neural network could predict the factor of safety of the monitored site to a reasonable level, using the field monitoring data and rainfall forecasts as inputs, which can form the basis of an early warning system as a means to mitigate the risk of rainfall induced landslides.