Automatic channel detection using deep learning

Parihaka dataset

We apply the weights from training the synthetic data to a 135x201x301 subvolume of the Parihaka dataset in New Zealand (Figure 14). The relative coarse-grained channel deposits are at the base of the incisional channel systems, which is different from the Australian dataset where the coarse-grained channel deposits are vertically stacked. The dataset is in time with a sample rate of 4 ms. We also use nonstationary patching method (Claerbout, 2014) to divide the subvolume into small overlapping volumes of size 128x128x128 samples in order to eliminate edge artifacts. The trained neural network model successfully picks the channel bodies in the seismic volume with high probabilities (Figure 15a). The model uncertainty is calculated by using the variance of 100 samples from the posterior distribution of channel probability (Figure 16). Parihaka dataset is different from our synthetic training dataset so applying our trained model is hard to produce a clean probability volume. However, high channel probabilities follow the channel edges enhanced by plane wave destruction Sobel filter (Phillips and Fomel, 2017) (Figure 15b) with the addition of model uncertainty.

imageParinew3
Figure 14. Parihaka field dataset.

Parinew3,sobelnew3
Figure 15. (a) Channel probability in the Parihaka field dataset. (b) Channel boundaries enhancement in the Parihaka dataset by PWD Sobel filter.

Pariunnew3
Figure 16. Model uncertainty in the Parihaka field dataset.

Automatic channel detection using deep learning