Analysis of microseismicity in sea ice with deep learning and Bayesian inference: application to high-resolution thickness monitoring

Abstract

In the perspective of an upcoming seasonally ice-free Arctic, understanding the dynamics of sea ice in the changing climate is a major challenge in oceanography and climatology. In particular, the new generation of sea ice models will require fine parameterization of sea ice thickness and rheology. With the rapidly evolving state of sea ice, achieving better accuracy, as well as finer temporal and spatial resolutions of its thickness, will set new monitoring standards, with major scientific and geopolitical implications. Recent studies have shown the potential of passive seismology to monitor the thickness, density and elastic properties of sea ice with significantly reduced logistical constraints. For example, human intervention is no longer required, except to install and uninstall the geophones. Building on this approach, we introduce a methodology for estimating sea ice thickness with high spatial and temporal resolutions from the analysis of icequake waveforms. This methodology is based on a deep convolutional neural network for automatic clustering of the ambient seismicity recorded on sea ice, combined with a Bayesian inversion of the clustered waveforms. By applying this approach to seismic data recorded in March 2019 on fast ice in the Van Mijen Fjord (Svalbard), we observe the spatial clustering of icequake sources along the shoreline of the fjord. The ice thickness is shown to follow an increasing trend that is consistent with the evolution of temperatures during the 4 weeks of data recording. Comparing the energy of the icequakes with that of artificial seismic sources, we were able to derive a power law of icequake energy and to relate this energy to the size of the cracks that generate the icequakes.