Transportability of a Convolutional Neural Network Seismic Denoising Model
Description:
Machine learning algorithms typically follow the paradigm of training a model on compiled datasets with characteristics of the type of data that the model is designed to then evaluate. This has often led to model performance problems when taking a previously trained model and attempting to utilize it to evaluate a new dataset which differs in characteristics when compared to the training dataset. This problem of model transportability is pronounced in the field of seismology, as we typically train our machine learning models on data specific to certain regions, networks, or time periods, which then experience significant performance drops when those models are utilized in other regions. Here, we assess the transportability of our previously constructed Multi-level Wavelet-based Convolutional Neural Network (MWCNN) seismic denoising model and investigate whether techniques such as transfer learning may provide solutions to said transportability problems. The MWCNN model was originally trained on earthquake and explosion data from the Utah seismic region and will now be used to analyze data from the Nevada seismic network. To assess the degree to which transportability is an issue for the MWCNN model we will first evaluate the denoising performance of the Utah-trained model using data associated with the 2020 Monte Cristo Range, Nevada, earthquake sequence. Then, we will compare the denoising capabilities of the Utah-trained MWCNN denoising model with that of a denoising model trained exclusively on Nevada earthquake and explosion data to measure whether a performance gap exists and what its magnitude is. Lastly, our goal is to assess whether transfer learning techniques may provide an avenue towards accounting for transportability issues by examining the denoising performance of a model with weights initially assigned using Utah-based data that is then further trained on data from the Nevada region.
Session: Advancements in Forensic Seismology and Explosion Monitoring [Poster Session]
Type: Poster
Date: 5/2/2024
Presentation Time: 08:00 AM (local time)
Presenting Author: Louis
Student Presenter: No
Invited Presentation:
Authors
Louis Quinones Presenting Author Corresponding Author laquino@sandia.gov Sandia National Laboratories |
Rigobert Tibi rtibi@sandia.gov Sandia National Laboratories |
|
|
|
|
|
|
|
Transportability of a Convolutional Neural Network Seismic Denoising Model
Category
Advancements in Forensic Seismology and Explosion Monitoring