Obtaining optimal velocity thresholds for deep-seated landslides using different fitting algorithms and receiver operating characteristic

Abstract

The Philippine Institute of Volcanology and Seismology - Project Dynaslope (PHIVOLCS - Dynaslope) developed an early warning system (EWS) for deep-seated landslide which uses threshold values as a basis in issuing alerts (i.e., A0, A2, and A3). Movements in the subsurface, which is measured through tilt sensors, is one of the parameters being monitored. An alert is raised every time the measurement exceeds velocity threshold. However, the sensor data exhibits frequent fluctuations and may cause false alerts. At present, simple moving average (SMA) is being used to attenuate noisy data. However, SMA exhibits delay at higher values which may also cause missed alerts. To prevent this issue, a suitable fitting algorithm should be implemented to reduce data noise and a site-specific optimal velocity threshold should be defined to capture impending alerts. This paper shows different techniques in fitting the sensor data and identifying the optimal velocity threshold (A2 alert) for the three sites: (1) Brgy. Sagasa, Dadong, Tarragona, Davao Oriental - dadtb, (2) Brgy. Magsaysay, Kibawe, Bukidnon - magta, and (3) Brgy. Tue, Tadian, Mt. Province - tuetb. Among the six algorithms tested, locally weighted scatterplot smoothing (LOWESS) found to be the most suitable algorithm due to its low root mean square error (RMSE) of 4.689 as compared to other algorithms. Using LOWESS and receiver operating characteristic (ROC) analysis, the optimal velocity thresholds for A2 alert are 0.73 cm/day, 0.50 cm/day, and 0.03 cm/day for magta, dadtb, and tuetb respectively. Also, LOWESS exhibits a favorable area under the curve (AUC) of 0.991. Thresholds will be further investigated to incorporate the response of the stakeholders at risk in order to maintain the integrity of the EWS.