Strategies in picking training data for 3D convolutional neural networks in stratigraphic interpretation
Oddgeir Gramstad, Michael Nickel, Bartosz Goledowski, Marie Etchebes · 2020
In this work, we present a new machine learning workflow for stratigraphic interpretation examining different strategies of picking training data. Generating training data is a time consuming process and different aspects of the training data, such as the dimensionality, spatial extent, amount of data and distribution within the survey might affect the robustness and accuracy of the final surface prediction. We use the extracted training data to train a 3D convolutional neural network (3D CNN) which is demonstrated in two different examples. In the first one, an interpreter manually picks a set of seed points, automatically extracts training data in the vicinity of these points and predicts the specific surface. This result is then compared against the same surface extracted by an auto-tracker using the same manually picked input points used for the 3D CNN. The second example demonstrates how we can use well marker calibration points to automatically extract labeled training data. In both cases, the machine learning workflow shows improved accuracy and quality of the surface interpretation. The workflow also decreases the turnaround time of the interpretation process because it requires fewer user-defined input parameters. Presentation Date: Tuesday, October 13, 2020 Session Start Time: 9:20 AM Presentation Time: 10:35 AM Location: Poster Station 1 Presentation Type: Poster