Identification of Icequakes and Earthquakes on the Ross Ice Shelf, Antarctica using Unsupervised Deep Embedded Clustering

Abstract:

Rapid advances in machine learning methods and computational capacity, coupled with the prevalence of increasingly large datasets, have yielded new techniques with which to process, sort, and analyze seismic observations. We present an implementation of a recently developed technique called deep embedded clustering (DEC), which enables separation of different types of icequakes and earthquakes through unsupervised clustering of seismic signals. The DEC algorithm employs spectrograms of seismic time series as input, encodes them to an N-dimensional representation of their latent features using a convolutional neural network, then seeks to separate signal types using their latent space representations. To initialize the encoder in the DEC method, a convolutional autoencoder composed of encoding and decoding layers is constructed. The autoencoder is trained until the latent features of the input spectrograms are learned and the spectrograms can be sufficiently reconstructed, preserving amplitude and time-frequency structure. Clusters are initialized using the K-means algorithm. Once initialized, the DEC algorithm updates clusters by minimizing the Kullback-Leibler divergence with a pre-selected distribution, while simultaneously fine-tuning the encoder parameters. We apply the DEC method to data recorded by a 34-station broadband seismic array deployed on the Ross Ice Shelf (RIS) from 2014 to 2016, containing a broad variety of seismic signals. Using DEC, we demonstrate the ability to group different types of RIS seismic signals, icequakes, and earthquakes, without the need for manual labeling or prior knowledge of signals.