Uncovering Stick-Slip Events: Denoising Cryoseismological Distributed Acoustic Sensing Data with an Autoencoder

One major challenge in cryoseismology is that signals of interest are often buried within the high noise level emitted by a multitude of environmental processes. Specifically, basal sources such as stick-slip events often stay unnoticed due to long travel paths to surface sensors and accompanied wave attenuation. Yet, stick-slip events play a crucial role in understanding glacier sliding and therefore, it is of great interest to investigate their spatio-temporal evolution, across the entire glacier from its ablation to its accumulation zone.
Distributed Acoustic Sensing (DAS) is a technology for measuring strain rate by using common fiber-optic cables in combination with an interrogation unit. This technology enables us to acquire seismic data over an entire glacier with great spatial and temporal resolution. To unmask stick-slip events, new techniques are required that effectively and efficiently denoise large cryoseismological DAS data sets. 
Here, we propose an autoencoder, a type of deep neural network, which is able to separate the incoherent environmental noise from the temporally and spatially coherent signals of interest (e.g., stick-slip events or crevasse formations). We trained the autoencoder in order to denoise a DAS data set acquired on Rhonegletscher, Switzerland, in July 2020. Due to the highly active and dynamic cryospheric environment as well as non-ideal cable-ground coupling the collected DAS data are characterized by a low signal to noise ratio compared to classical point sensors.
Several models were trained on a variety of data subsets, differing in recording positions (ablation or accumulation zone), event types (stick-slip event or surface event) and the quantity of training events. We compare and discuss the denoising capabilities of these models with several metrics, such as inter-channel coherence, similarity between seismometer and DAS recordings, and visual assessment. This evaluation is conducted while considering different data types in a qualitative and quantitative manner. All models show an increase in inter-channel coherence of the seismic records after denoising. Further, all models uncover previously undetected stick-slip events, whereby models trained on manually picked training data perform better than models trained on randomly picked training data. We believe that the application of our models can improve the understanding of basal stick-slip information in cryoseismological DAS datasets, potentially uncovering previously hidden information.