A Machine Learning Approach to Identify Landslides With Seismic Waves Using Support Vector Machine Method

Date: 4/24/2019

Time: 06:00 PM

Room: Grand Ballroom

Landslides are a common natural geologic hazard that can cause significant damage to human life and properties. The typical ways to study landslides and evaluate their susceptibilities are through in-situ measurement and remote sensing techniques (i.e., satellite imagery, GPS, InSAR). These measurements, in general, provide good constraints on bulk properties of landslides, such as the area, overall shapes, and mass volumes, but not as much on their dynamic processes or the occurrence times, due to the lack of temporal resolution. On the other hand, landquake, a term that refers to landslide signals being recorded by seismometers, can provide information on the evolution time history as well as the dynamic processes, and even source mechanisms. Identification of landquakes on seismograms is challenging, as they have emerging onsets and complicated long-duration wave trains with relatively low SNR. One can potentially identify them in a very low-frequency band (e.g., 0.02~0.05 Hz) where the phase is cleaner, or in a higher frequency band (e.g., 1~3 Hz) when the events contain more high-frequency energy. But it’s easy to misclassify them with teleseismic earthquakes, or regional earthquakes/tectonic tremors without manually analyzing the origin of the source carefully. Taking the advantage of recent developments in machine learning techniques, we built an automatic detector based on a support vector machine (SVM) to detect and classify various types of events (e.g., local earthquake, teleseismic, tectonic tremor, and landquakes) using the Broadband Array in Taiwan for Seismology (BATS) network. The model can help with identifying more landslides and separating landquakes from background noise and other natural or anthropogenic sources. Our next step is to apply the SVM-based detector to continuous seismic data recorded in landslide-prone regions to build a more complete landslide catalog. We hope that complete landslide catalogs can contribute to better understanding of landslide dynamics and source mechanisms and eventually help in mitigating landslide hazards.

Presenting Author: Lindsay Y. Chuang

Authors