Automatic channel detection using deep learning

Browse Basin dataset

We apply the weights from training the synthetic data to a 312x312x100 subvolume of a field dataset from offshore Australia (Figure 11). The dataset is a 3D marine seismic survey located in 2500 m water depth with a sample rate of 2 ms and a dominant frequency of 120 Hz. The dataset is in depth with a sampling interval of 2 m. The seismic data hosts numerous stacked deep-water channel-levee complexes. We divide the subvolume into 16 small overlapping volumes of size 128x128x128 samples using nonstationary patching method (Claerbout, 2014) and padding along the depth dimension, to eliminate the edge artifacts and test each volume independently. The testing output volumes are stitched together using the inverse of non-stationary patching method with weighted boundaries. Figure 12a shows that channel bodies are clearly picked in the seismic volume. We analyze the prediction uncertainty by using the variance of 100 samples from the posterior distribution of channel probability (Figure 13).

When there are multiple channels in the dataset (around crossline 4000 and inline 2780 in Figure 11), the trained model cannot distinguish individual channels very well and the prediction uncertainty is high. The trained model can detect thin channels in the dataset with not too high probabilities, but the uncertainty map displays high values in these regions. Therefore, the prediction uncertainty has useful information for the channels detection task and interpreters can repick the regions with high uncertainty to enhance the detection result from neural network. Our result follows the channel edges enhanced by plane wave destruction Sobel filter (Phillips and Fomel, 2017) (Figure 12b), with the addition of model uncertainty.

imageausnew3
Figure 11. Australia field dataset.

ausnew3,aussobel3
Figure 12. (a) Channel probability in the Australia field dataset. (b) Channel boundaries enhancement in the Australia dataset by PWD Sobel filter.

ausunnew3
Figure 13. Model uncertainty in the Australia field dataset.

Automatic channel detection using deep learning