Table of Contents
Seismic-NIST is a dataset of acoustic seismic waveforms and their underlying velocity profiles. The dataset is inspired by the work of Roeth and Tarantola 1994 where the authors tried to perform seismic inversion from raw acoustic waveforms at various levels of noise. Here we provide a reference dataset of such waveforms. The machine learning task to be solved is a regression problem of predicting synthetic p-Wave velocity profiles from given acoustic waveforms. The data can be generated completly from scratch using torch and libraries from the devito project.
The dataset is named after the outstanding deep-learning benchmark MNIST by Yann Le Cun, and is inspired by other projects such as FashionMNIST and KMNIST.
The dataset consists of 750 waveforms generated from 9-layer earth models of acoustic p-wave velocities.
The training set consists of 600 waveforms and the test sets consist of 150 waveforms.
There are three test sets - SNIST-0, SNIST-1 and SNIST-2.
The number corresponds to the level of noise added to the test set i.e. SNIST-0 has no noise added, SNIST-1 adds 1 sigma of noise, and SNIST-2 has 2 sigma of noise added. The noise is Gaussian uncorrelated noise.
Each waveform consists of 20 traces according to 20 offsets sampled at 8 ms time intervals. p-Wave velocities are capped at 4000 [m/s].
Here's what the waveform amplitudes and some of the velocity profiles (ground truth - black) look like.
The dataset was largely inspired by discussion on the software-underground slack channel and by Agile Geoscience's blog post on benchmark studies in the machine learning - geoscience domain.
While the realism and usefulness in terms of real seismic applications is limited, this benchmark may serve as a reference on what a realistic benchmark should include. Hence, this benchmark is very much a platform or sandbox as not (m)any reference benchmarks exist in the seismic deep-learning domain. It is up to the community to shape what we want out of such a reference benchmark and I hope to provide here a starting point for such a discussion.
If you would like to contribute or would like to raise an issue please do so and join the discussion on the slack-channel.
The data comes prepackaged as .npy files. Which you can either download manually or use the existing torch.dataset implementation found in utils/snist.
| File | Examples | Size | Download (NumPy format) |
|---|---|---|---|
| Training Amplitudes | 600 | 13 MB | train_amplitudes.npy |
| Training Velocities | 600 | 21 KB | train_velocities.npy |
| Testing Amplitudes 0 (no noise - SNIST-0) | 150 | 3 MB | test_amplitudes.npy |
| Testing Amplitudes 1 (1 sigma noise - SNIST-1) | 150 | 3 MB | test_amplitudes_noise_1.npy |
| Testing Amplitudes 2 (2 sigma noise - SNIST-2) | 150 | 3 MB | test_amplitudes_noise_2.npy |
| Testing Velocities | 150 | 5 KB | test_velocities.npy |
The following is an example on how to use the provided dataset in torch.
All the data will automatically be downloaded - in this case - from the directory and is ready for training.
You can try it out on 
from utils.snist import SNIST
snist_train = SNIST('./', train=True, download=True)
snist_0_test = SNIST('./', train=False, download=True, noise=0)
snist_1_test = SNIST('./', train=False, download=True, noise=1)
snist_2_test = SNIST('./', train=False, download=True, noise=2)A reference implementation is provided and here we collect the performance of methods that have been evaluated on the SeismicNIST dataset.
If you wish to contribute to this list please raise a pull-request and provide a link to a repository where your results can be reproduced.
| Model | SNIST-0 | SNIST-1 | SNIST-2 | Credit | Link |
|---|---|---|---|---|---|
| 1-Hidden Layer Benchmark | 242.42 [m\s] | 287.98 [m\s] | 428.59 [m\s] | @porestar | |
The data can be reproduced by running make build in the data_generation
directory.
This will run three scripts:
generate_velocities.py: creates the velocity models based on the paper by Roeth and Tarantolagenerate_amplitudes.sh: runs a docker container of devito and runs the forward model on the created velocitiesgenerate_noisy_test_set.py: creates the noisy SNIST versions SNIST-1 and SNIST-2
If you would like to contribute or would like to raise an issue please do so and join the discussion on the slack-channel.
MIT License
Copyright (c) 2019 Lukas Mosser
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.