Cluster Analysis of Slope Hazard Seismic Recordings Based Upon Unsupervised Deep Embedded Clustering

Abstract

Slope disasters, such as landslides and rockfalls, can cause significant safety issues for land and road use in many countries. Therefore, it is important to have effective monitoring and early warning systems in place. Although using images or closed-circuit televisions have some limitations, using seismic data for monitoring can provide real-time information, are useful at night, and are less affected by weather conditions. However, a lack of seismic recordings for different types of slope disasters can limit the use of the seismic method. To collect more seismic recordings, we deployed seismometers in the Luhu mountain area of Miaoli, Taiwan, to record rockfalls, earthquakes, and other natural and human-induced events. To cluster the different types of seismic waveforms present in these recordings, we used deep embedded clustering (DEC), an unsupervised clustering algorithm, to group the recordings based on their features. We selected waveforms with significantly larger amplitudes than background noise and converted them to spectrograms using the short-time Fourier transform as input for DEC. About 45,000 seismic recordings were selected, of which 1951 were manually labeled source classes as external information for finding an optimal number of clusters and evaluating clustering results. The DEC clustering results showed that most seismic recordings of different source classes had distinct features and could be grouped into different clusters. However, different classes of seismic recordings may be assigned to the same cluster if they have similar features on spectrograms but can still be distinguished by other conditions. By utilizing unsupervised clustering to categorize a large number of seismic recordings based on their features, the time required for manual classification can be reduced. The identified seismic recordings of rockfalls will be useful for monitoring and analyzing rockfall hazards using seismic data.

RockNet: Rockfall and Earthquake Detection and Association via Multitask Learning and Transfer Learning

“…Unsupervised clustering methods have recently shown promise in grouping seismic recordings according to their waveform features without relying on manual labeling. This approach could be useful in providing a more objective and efficient identification of recordings of rockfalls, reducing the need for manual identifications and potentially improving the quality of datasets [43]. Additionally, advanced machine-learning techniques can be adopted to further improve the accuracy of the model.…”