We present a workflow to automate the channel interpretation in seismic images by using a 3D CNN (convolutional neural network). The main problem in applying the CNN to the channel interpretation is the inaccessibility of labeled training data. To settle this problem, we propose a workflow to simulate meandering channels and then integrate the simulated channel systems with folding structures to create realistic structure models. By randomly selecting the parameters in this workflow, we can generate numerous structure models with various geological features. Then we convert the structure models into comparable reflectivity models. We further convolve the reflectivity models with a frequency-varying wavelet to generate synthetic seismic images and the corresponding channel labels to train the CNN. During the training, we augment the training datasets to further enhance the diversity of data. Although trained on only synthetic seismic images, this CNN shows an outstanding performance on field seismic images. This suggests that synthetic seismic images created in the proposed workflow are realistic enough to enable the CNN to learn the channel characterization in field seismic images which are new to the CNN.
Presentation Date: Monday, October 12, 2020
Session Start Time: 1:50 PM
Presentation Time: 3:05 PM
Location: 351F
Presentation Type: Oral
