Previously, in our series on OpendTect Machine Learning workflows, we showed an unsupervised workflow for seismic facies analysis. That workflow clustered seismic waveforms to generate a segmentation volume consisting of 50 different segments enabling detailed interpretation of seismic facies.
Today, we show one of several possible workflows in OpendTect using supervised learning. This workflow uses the Thalweg tracker for labeling target positions. In total 8 different label sets were created representing positive and negative amplitude classes of meandering channels, unconfined channels, splays and floodplains.
Next, we trained a LeNet CNN-type of Machine Learning model to classify the 3D seismic response surrounding each labeled position into these 8 classes. This is an image-to-point classification approach, which takes considerable time in the application phase. To speed up the process, we only generated a small output volume, which we subsequently used as target for an image-to-image classification run. The model used in this phase is a 3D U-Net.
Training this model is time-consuming but we make up for the lost time in the application phase, which is blazingly fast. In the video we show the classification volume in the Wheeler (HorizonCube-flattened) domain.
For details about this workflow, please see our publication in First Break: de Groot, P., Pelissier, M., and van Hout, M., . Seismic classification: A Thalweg tracking/machine learning approach. First Break, Vol. 39, pg. 59-64, March 2021.
The workflow shown in the video is recorded in version 7, which we are now rigorously testing for release. We will be giving live demonstrations of the upcoming release at the EAGE in Vienna booth 4314 . To replicate this workflow in version 6, please be aware of user interface differences.
Last but not least, we will be organizing a free webinar on data conditioning today at 4 pm CET.
We hope to see you online this afternoon!