Detecting change of patterns in landslide displacements using machine learning, an example application.

Machine learning and signal processing can support the definition of landslide alert/alarm systems based on monitoring data. The possibility to rely on a straightforward and automatic procedure to identify hazardous situations could be very useful for risk management and decision makers. In this work, we propose a hierarchical clustering algorithm to identify changes of pattern in the displacements of monitored landslides. Our test site is a large, active Deep-seated Gravitational Slope Deformation (DGSD) in which secondary movements provide sediment for debris flows that threaten downstream settlements. An Automated Total Station (ATS) has been installed in 2012 to measure the three-dimensional displacements of several benchmarks distributed on the source area and to trigger alarms if superficial movements potentially leading to collapses are detected. Result show that the procedure allows to group benchmark with similar displacement patterns. The unsupervised definition of homogenous areas from a kinematics viewpoint supports an unbiased geomorphological characterization of the large landslide. Moreover, the method allows to trigger alert warnings if some monitored points change displacement pattern. The identification of possible hazardous situation is performed without imposing fixed and arbitrary thresholds and without calibration. The recognition of areas with new types of activity supports the definition of the sediment volumes available for transport for the next debris flow event, and assists the definition of reliable risk scenarios.