Classifying Landslide Seismic Signals With Unsupervised Machine Learning From Multiple Locations
Description:
Creating a reliable landslide early warning system is one of the main goals of landslide scientists. Seismometers are used to monitor landslides since they have high temporal resolution and an exceptional ability to record subtle vibrations from internal cracks caused by landslide movements. Because landslides are complex, it should be monitored at multiple locations. When this data is available, it is often impossible to understand every signal recorded. Recent advances on unsupervised machine learning mainly classify the time periods of continuous seismic signals of a single station but when multiple stations have different classifications at the same time, it can be challenging to interpret the overall behavior of an area since it is preferable to be consistent between stations. For slow-moving large-scale landslides, this is particularly true because different blocks of the landslide could behave distinctly. For this reason, we combine signals from seismometers of an array to build a spatial feature for classification, as a network response rather than a single-station response. We used an unsupervised machine learning method called K-means to do the classification and cross-validate our results. We used collocated GPS stations to study classes of seismic signals related to landslide movement and investigated the cause of movement from external forces. Finally, we compared our results to the common single-station classification approach. Our results reveal classes of seismic signals related to landslide movement and rainfall, where we can identify first-order spatiotemporal characteristics. We believe our approach has the potential to issue landslide alerts based on multiple stations in a landslide region.
Session: Detecting, Locating, Characterizing and Monitoring Non-earthquake Seismoacoustic Sources
Type: Oral
Date: 4/19/2023
Presentation Time: 08:45 AM (local time)
Presenting Author: Kyle Smith
Student Presenter: No
Invited Presentation:
Authors
Kyle Smith Presenting Author Corresponding Author kyle36smih85@gmail.com Institute of Earth Sciences, Academia Sinica |
Hsin-Hua Huang hhhuang@earth.sinica.edu.tw Institute of Earth Sciences, Academia Sinica |
|
|
|
|
|
|
|
Classifying Landslide Seismic Signals With Unsupervised Machine Learning From Multiple Locations
Category
Detecting, Locating, Characterizing and Monitoring Non-earthquake Seismoacoustic Sources