Automatic channel detection using deep learning

Training Data

For the training data, we choose a 3D convolutional synthetic depth model created by James Jennings at Bureau of Economic Geology, Austin, Texas, in collaboration with Chevron (Figure 5a). The data simulates a complex deep-water stacked channel system in Africa with correlated noise in porosity. On top of the channel is an overburden with stochastically generated velocity fluctuations and correlated noise in porosity (Fomel et al., 2007). The dominant frequency of the seismic wavelet is 40 Hz. The data is created by three pieces of information: a 3D shallow high-resolution seismic data is used together with an analytical curve to simulate the shape of channels, a group of geologists study the channel properties distribution at an analog outcrop in California, and the background information is created by geostatistics.

We eliminate the noise in the channel bodies and subtract the result from original data to obtain the location of channels. We create the labels by simply masking the channel location with 1 and everywhere else with 0 (Figure 6). We modify different channel properties, such as amalgamated sand cross-section shape parameter, porosity, dominant frequency, and channels thickness to create a diverse training dataset (Figure 5b). Because of limited computational resources, a training batch has 10 seismic volumes with a size of 128x128x128 samples (Figure 7). Examples in the training data overlap with one another, but it is a way of augmenting the data. We generate a total of 1140 training examples with 300 examples for validating the network.

mt3d-40,mt3d-40-2
Figure 5. (a) Synthetic training data. (b) An example of synthetic training data with thin channels.

diff-mask2
Figure 6. Training label.

prdw9
Figure 7. Training cuboid.

Automatic channel detection using deep learning