Lorenzo Nava , Antoinette Tordesillas, Guoqi Qian, Filippo Catani

This research presents the development of a Landslide Regime Shift Detector (LRSD) which integrates advanced prediction models to provide insights into regime shifts within substantial landslide bodies. Our primary focus lies in the identification of these shifts, emphasizing that LRSD does not merely detect accelerations but discerns exceptional accelerations, defined in our context as deviations significantly departing from historical relationships between predictions and actual accelerations. The study focuses on slow-moving landslides, employing a Vector Error-Correction Cointegration (VECC) model to predict displacement at several monitoring points, with rainfall as a crucial influencing factor. While the VECC model effectively predicts rainfall-triggered accelerations, certain situations exhibit unexpected acceleration magnitudes, resulting in extreme residuals. To establish the LRSD, 24-hour aggregated residuals between predicted and observed displacements are employed. Kernel density estimation (KDE) is applied to derive cumulative distribution function (CDF) values for the residuals, with a predetermined threshold of 0.999 CDF triggering alerts for each monitoring point independently. A last step, which includes the persistence in time of the threshold exceedance and the number of monitoring points that exceed the threshold at the same time is performed. This is to reduce the effect of noisy displacement time series and to encode the "group dynamics," crucial for identifying pre-failure indications. This approach offers several advantages, including the effective identification of critical time states, adaptability, and transferability. Furthermore, it introduces novel information into local landslide early warning systems (Lo-LEWS), which consists of strong changes in landslide trends and anomalous responses to external triggers, if any. This approach significantly enhances confidence in the resultant alert, particularly when integrated with conventional alert systems, thereby improving the reliability of Lo-LEWS.