next up previous [pdf]

Next: Acknowledgments Up: Pham et al.: DL Previous: Parihaka dataset

Conclusions

We propose a method for automatic detection of channel bodies in seismic images using an encoder-decoder convolutional neural network. The network is trained on synthetic training data and is then applied to field data.

We test the model on field datasets from offshore Australia and New Zealand. With only training on the synthetic dataset, the model succesfully identifies the channel bodies in the field datasets. The prediction uncertainty is computed simultaneously and can help an interpreter judge and enhance the channel detection results.

We believe the proposed method has a high potential in the future for automatic interpretation and quantitative analysis. Neural network models are trained with synthetic datasets created by the knowledge of experts from geologists, geophysicists, and petroleum engineers, and then the trained models are applied to field datasets to perform interpretation tasks such as faults, salt, and channel geobodies detection. Our results can be improved using more diverse labeled training datasets. Future research will also combine object detection and semantic segmentation to clearly image individual channels.


next up previous [pdf]

Next: Acknowledgments Up: Pham et al.: DL Previous: Parihaka dataset

2022-04-29