Skip to content

Using convolutional autoencoders to remove random noise from seismic data.

Notifications You must be signed in to change notification settings

Greveley/AutoEncoderImageDenoise

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

81 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

This repo contains code that uses a deep convolutional autoencoder to automatically detect and remove random noise from seismic images.

Methodology

The methodology consists of 2 stages:

Stage 1: Synthetic Data Training

Generate Data:

Synthetic seismograms of a random number of seismic events are generated using a ricker wavelet. Random characteristics of these seismic events include:

  • Frequency (between 10 and 70Hz)
  • Angle (beween +/- 50 degrees)
  • Amplitude (between +/- 2) These seismograms are the target output of the network. A random level of Gaussian noise is then added to the clean seismic data; this dataset is the input to the network.

Model generation:

A number of models have been (and still are being) tested to give optimum results. These models include:

  • Convolutional autoencoders applied over windows of the input images. *** results shown below from this model type ***
  • Unet architecture model.
  • 1D Artificial neural network.
  • 1D Convolutional neural network.

Fitting models:

These models are then fit to the synthetic data to effectively remove the random noise. Optimizer = SGD, loss function = MSE.

Stage 2: Real Data Training

The models trained on the synthetic data are then used as a starting point to train on real data. However, the ground truth clean seismic image is not know for the real seismic data. Therefore a psuedo-unsupervised training scheme is implemented by editing the loss function to minimise the correllation coefficient between the input image and the output image.

Above shows two zoomed sections of the input seismic data, output of the denoising process and the removed noise.

Known Issues

There is noise still left within the data that I think could be removed by refining this method. In particular, some more hyperparameter tuning of batchsize, window size and possibly the optimizer.

Having observed results in the other papers I'm sure that the ones I'm getting can be improved. But this was initially meant as a PoC so improvements can be made!

Real seismic data displayed here can be found here: https://walrus.wr.usgs.gov/namss/

The methodology brings together those found in two papers:

  • Saad, O. and Chen, Y., 2020. Deep denoising autoencoder for seismic random noise attenuation. GEOPHYSICS, 85(4), pp.V367-V376.
  • Zhang, M., Liu, Y. and Chen, Y., 2019. Unsupervised Seismic Random Noise Attenuation Based on Deep Convolutional Neural Network. IEEE Access, 7, pp.179810-179822

About

Using convolutional autoencoders to remove random noise from seismic data.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published