In this study, we discuss the uncertainty involved in machine learning-based seismic image segmentation. Using a salt body detection example, we demonstrate that while often referred to as probability, the output from a single machine learning prediction run provides neither an adequate estimate of the true probability nor uncertainty. We use Monte Carlo dropout as the method to estimate epistemic uncertainty, while also providing a more appropriate estimate of the prediction probability. These results help us better understand the machine learning output.

Presentation Date: Monday, October 12, 2020

Session Start Time: 1:50 PM

Presentation Time: 2:40 PM

Location: 351F

Presentation Type: Oral

Keywords:
neural network, dropout, artificial intelligence, posterior distribution, machine learning, deep learning, epistemic uncertainty, prediction, classification result, cnn model
Subjects:
Reservoir Characterization, Information Management and Systems, Seismic processing and interpretation, Artificial intelligence, Neural networks
This content is only available via PDF.
2020. Society of Exploration Geophysicists
You can access this article if you purchase or spend a download.