/CNN-for-ASI

Tutorial: Convolutional Neural Networks for Automated Seismic Interpretation

Primary LanguagePythonOtherNOASSERTION

CNN for ASI

Code for the paper:
Convolutional Neural Networks for Automated Seismic Interpretation,
A. U. Waldeland, A. C. Jensen, L. Gelius and A. H. S. Solberg
The Leading Edge, July 2018

EAGE E-lecture: Seismic interpretation with deep learning (YouTube)

This repository contains python/pytorch code for applying Convolutional Neural Networks (CNN) on seismic data. The input is a segy-file containing post-stack seismic amplitude data. The training labels are given as images of the training slices, colored to indicate the classes.

Setup to get started

CNNs requires time-demanding computations, consider using a computer with a fast NVIDIA GPU with at least 6GB ram.

  • Clone this repository
  • Install the required python packages by using the command pip install requirments.txt If you have Anaconda, do the following:
  • Create a conda virtual enviroment by using the following command, replacing with a name for your virtual enviroment : conda create --name --file spec-file.txt
  • Download the demo data set here. This is the F3 Netherlands dataset - originally made available by OpenDTect - and available via the MalenoV Project.
  • Locate the '.segy'-file, rename it to 'data.segy' and put it in the 'F3'-folder.

How to use tensorboard (for visualization)

  • Open a terminal
  • cd to the code-folder
  • run: tensorboard --logdir='log'
  • Open a web-browser and go to localhost:6006
    More information can be found here.

Usage

  • train.py - train the CNN
  • test.py - Example of how the trained CNN can be applied to predict salt in a slice or the full cube. In addition it shows how learned attributes can be extracted.

Files

In addition, it may be useful to have a look on these files

  • texture_net.py - this is where the network is defined
  • batch.py - provide functionality to generate training batches with random augmentation
  • data.py - load/save data sets with segy-format and labeled slices as images
  • tb_logger.py - connects to the tensorboard functionality
  • utils.py - some help functions
  • test_parallel.py - An implemenation of test.py supporting multi-gpu prediction (thanks to Max Kaznady).

Using a different data set and custom training labels

If you want to use a different data set, do the following:

  • Make a new folder where you place the segy-file
  • Make a folder for the training labels
  • Save images of the slices you want to train on as 'SLICETYPE_SLICENO.png' (or jpg), where SLICETYPE is either 'inline', 'crossline', or 'timeslice' and SLICENO is the slice number.
  • Draw the classes on top of the seismic data, using a simple image editing program with the class colors. Currently up to six classes are supported, indicated by the colors: red, blue, green, cyan, magenta and yellow.

Contact

Email: anders.u.waldeland@gmail.com