Deep learning seismic waveforms in sea ice to understand its dynamics
5 to 6 months, starting between January and March 2025, followed by a PhD (ANR funding)
Understanding the dynamics of sea ice in the changing climate is crucial and represents a major challenge in the perspective of upcoming seasonally ice-free Arctic. In particular, accurate monitoring of the sea ice thickness and its rheology is crucial for understanding the fragmentation of the ice cover by waves, in view of forecasting the future state of polar to mid-latitude regions for both climate and short-term timescales.
Recently, it has been shown that essential sea ice parameters can be extracted from the seismic noise recorded with autonomous geophones on sea ice (Moreau, Boué et al., 2020 ; Moreau, Weiss et al., 2020, Serripierri et al., 2022), thus significantly reducing the logistics associated with seismic acquisitions in this hostile environment. However, the processing of thousands of icequakes (Figure 1), requires to perform waveform inversions that are computationally heavy. Using thousands of waveform inversions previously performed on seismic data acquired on sea ice, the goal of the internship is to train a deep convolutional neural network that will replace the inversion of ice properties.
As a first step, we will train a two-dimensional CNN that uses waveforms recorded at several stations as input for a regression prediction of the ice thicknesses, as well as source coordinates (three parameters in total). As the training data, we will start by using the catalog of waveforms containing more than 3000 icequakes and recorded on 250 stations on fast ice (representing more than 750000 waveforms to train from) in Svalbard during the Icewaveguide project. Each of these waveforms is associated with an ice thickness and a source-receiver distance (figure 1). The training dataset will also be completed with synthetic waveforms that can be convolved with different types of seismic noise recorded on fast ice, following the idea by Zhu et al. (2019). Various types of seismic noise (anthropogenic, transients, swell etc.) can be extracted and clustered following Seydoux et al. (2020). This was for example applied to the data recorded in Svalbard in Moreau et al. (2023). More training data will also be available from new field experiments. Once the CNN is operational, we will train another two-dimensional CNN that makes use of the same datasets, but this time based on the waveforms of the longitudinal and shear waves (on the horizontal channels), which are mainly sensitive to Young’s modulus. If both these CNN prove successful, we will finally train a CNN to generalize all six parameters at once, because using simultaneously the waveforms of the three types of waves (flexural, longitudinal and shear) better constrains each parameter individually. This is due to the fact that there is always at least some redundancy of information between each type of wave.
There is ANR funding to follow up the internship with a doctorate.
References
Moreau, L., Boué, et al. : Sea ice thickness and elastic properties from the analysis of multimodal guided wave propagation measured with a passive seismic array, JGR : Oceans, 125, e2019JC015 709, 2020.
Moreau, L., Weiss, J., et al.. : Accurate estimations of sea-ice thickness and elastic properties from seismic noise recorded with a minimal number of geophones : from thin landfast ice to thick pack ice, JGR : Oceans, 125, e2020JC016 492, 2020.
Moreau, L. et al.. : Analysis of micro-seismicity in sea ice with deep learning and Bayesian inference : application to high-resolution thickness monitoring, The Cryosphere, https://doi.org/10.5194/tc-2022-212, 2023.
Serripierri, A., et al. : Recovering and monitoring the thickness, density, and elastic properties of sea ice from seismic noise recorded in Svalbard, The Cryosphere, 16, 2527–2543, https://doi.org/10.5194/tc-16-2527-2022, 2022.
Seydoux, L. et al. : Clustering earth- quake signals and background noises in continuous seismic data with unsupervised deep learning, Nature Communications, 11, 3972, https://doi.org/10.1038/s41467-020-17841-x, 2020.
Zhu, W. et al. : Seismic Signal Denoising and Decomposition Using Deep Neural Networks, IEEE Transactions on Geoscience and Remote Sensing, 57, 9476–9488, https://doi.org/10.1109/TGRS.2019.2926772, 2019.
Mis à jour le 6 novembre 2024