Geomorphology and deep learning: Homogenization of a database of strike-slip fault markers for the evaluation of neural networks.

Contexte et objectifs du stage

Deep learning has great potential for automatically solving numerous problems in geosciences. This is particularly the case in morphotectonics, which aims to study the impact of earthquakes on the landscape. However, the main challenge to these applications lies in the difficulty of obtaining annotated data in sufficient quantity and quality to train and evaluate deep neural networks.

This internship is part of the "StrikeLearn" project, which aims to develop a neural network for the task of characterizing strike-slip faults from topographic data in the form of digital elevation models (DEM). This task aims to estimate the offsets along stream channels, which show evidence of past seismic events (Reitman et al., 2019) (see Figure 1). In this context, the annotation of these offsets by experts is difficult and costly. Given the rarity of real data, we have trained a neural network on synthetic data, made of synthetic DEMs and seismic offsets simulated with the Landlab software (Hobley et al., 2017). The model was evaluated on real data located along the San Andreas fault, USA (Visage et al., 2024). These preliminary results encourage us to now evaluate the model’s ability to generalize to different faults in different geographical regions, with varied climates and reliefs.
In this perspective, we have created a database compiling strike-slip fault offsets published in the literature, and for which topographic data are publicly available. However, these datasets derive from variable methodologies and are provided in heterogeneous formats.
In order to be able to exploit all these data, we therefore need to process and homogenize them, so that they are all grouped under a common format and quality level that makes them ready to be used for the evaluation of the neural network

The objectives of the internship are:

1. Analyze the data from previously-collected heterogeneous datasets.
2. Establish a protocol for filtering, cleaning, and homogenizing these data, determining a common data format and a conversion procedure for each dataset.
3. Develop the code responding to the problem and produce the homogenized data.
4. Evaluate the previously trained neural network on the produced data, analyze the performances, and compare them on the different faults.

References

Hobley, D. E. J., Adams, J. M., Nudurupati, S. S., Hutton, E. W. H., Gasparini, N. M., Istanbulluoglu, E., & Tucker, G. E. (2017). Creative computing with Landlab : An open-source toolkit for building, coupling, and exploring two-dimensional numerical models of Earth-surface dynamics. Earth Surface Dynamics, 5(1), 21‑46. https://doi.org/10.5194/esurf-5-21-2017

Reitman, N. G., Mueller, K. J., Tucker, G. E., Gold, R. D., Briggs, R. W., & Barnhart, K. R. (2019). Offset channels may not accurately record strike‐slip fault displacement: Evidence from landscape evolution models. Journal of Geophysical Research: Solid Earth, 124(12), 13427-13451. https://doi.org/10.1029/2019JB018596

Visage, S., Pousse, L., Giffard-Roisin, S., Mouchené, M., Audin, L., & Perrinel, S. (2024). Characterisation of strike-slip fault offsets using convolutional neural networks (EGU24-16516). EGU24. Copernicus Meetings. https://doi.org/10.5194/egusphere-egu24-16516