Deep Embedded Clustering of a Cryo-Data-Cube

Cryoseismological records consist of numerous signals generated by various sources within or surrounding glacial ice, including icequakes, water flow, avalanches, rockfalls, wind, or precipitation. This results in a notably high noise level within the data, posing a significant challenge in detecting and distinguishing individual seismic events and sources.

Our research employs Deep Embedded Clustering (DEC) to address this challenge, focusing on the analysis of a Distributed Acoustic Sensing (DAS) dataset acquired on Rhonegletscher (Switzerland) in 2020.

To visualize and efficiently streamline the DEC processing of this substantial volume of data, we reorganize the numerous continuous DAS channels as a 3D data cube featuring the three dimensions: time, space, and frequency. The DEC approach involves first transforming high-dimensional seismic data into a more manageable lower-dimensional latent space using an autoencoder. This transformation is vital in emphasizing the essential characteristics of the data, thereby enabling more effective clustering. Subsequently, the DEC algorithm autonomously categorizes these seismic signals into distinct clusters based on their unique spatio-temporal characteristics, without the prerequisite of manual annotation.

The primary aim of this approach is to utilize DEC for the effective mapping of clearly defined spatio-temporal clusters within cryoseismological records. This approach is geared towards achieving a more nuanced understanding of the various sources contributing to these records and their complex dynamics. By successfully segregating these clusters, the aim is to reveal new insights into the complex processes and interactions in glacial environments.

Both the DAS data and the clustering results can be explored interactively using the data cube viewer Lexcube. Come find us at the poster stand!