Weakly Supervised Seismic Structure Labeling via Orthogonal Non-Negative Matrix Factorization
Y.K. Alaudah, H. Di and G. AlRegib
Event name: 79th EAGE Conference and Exhibition 2017
Session: Seismic Interpretation - Analytics and Machine Learning for Interpretation
Publication date: 12 June 2017
Info: Extended abstract, PDF ( 597.69Kb )
Price: € 20
With the growing demand of high-resolution subsurface characterization from 3D seismic surveying, the size of 3D seismic datasets has been dramatically increasing, and correspondingly, the process of interpreting a seismic dataset is becoming more time consuming and labor intensive. In addition, supervised machine learning has proved to be very successful for many applications in computational seismic interpretation. However obtaining training labels for large volumes of seismic data is a very demanding task. Furthermore, while the amount of data is continuously growing, the ability of human experts to label data remains limited. In this work, we propose a weakly-supervised framework for labeling seismic structures using Non-Negative Matrix Factorization (NMF) with additional sparsity and orthogonality constraints. We show that weakly-supervised learning requires a much smaller number of labels. Furthermore, we show that "rough" image-level labels of specific seismic structures can be mapped into finer more localized locations within the seismic volume. Results obtained by labeling fault regions and salt dome boundaries from the Netherlands F3 block prove to be very promising.