A reliability evaluation of four landslide failure forecasting methods in real-time monitoring applications

Abstract

Early warning systems (EWSs) for landslides are becoming a pivotal tool to safeguard assets and stakeholders. With this mission, an EWS should be capable of reliably forecasting the failure time when the ground accelerates. There are analytical methods developed to this end that use time-series kinematics: inverse velocity (INV), minimum inverse velocity (MINV), slope gradient (SLO), and velocity over acceleration (VOA). Although an abundant number of studies applied these methods, they have been majorly examined in a back-analysis context where all the measurements are incorporated into the forecasting process. A successful operation of EWSs in raising meaningful alarms calls for an examination in which the forecasting method is evaluated synchronously. This study evaluates the ability of the four mentioned methods to provide reliable forecasts in real time using a comprehensive database including 75 historical failures. For the first time, the methods are evaluated using a quantitative metric called reliability fitness index (RFI) that measures the portion of forecasts meeting an accuracy threshold. For accuracy thresholds of 50, 75 and 90%, INV showed the highest RFI values of 16, 7, and 4% followed by SLO values of 12, 5, and 2%, respectively. Opposing reliability values for SLO and INV suggest EWSs should take advantage of hybrid models that consider both methods.