Weakly Supervised Seismic Interpretation for Labeled Data Constrained Settings

Collaborators: Yazeed Alaudah, Motaz Alfarraj, Haibin Di, Shan Gao, Ghassan AlRegib

Goal/Motivation: Devise machine learning-based seismic interpretation frameworks based on weak labels

Challenges: Machine and deep learning has shown immense promise in automating interpretation of large seismic volumes. However, for progress on this front to be better realized, there needs to be an effort towards making present algorithms less data-hungry. In particular, present seismic interpretation techniques require several seismic sections to be labeled in a pixel-wise manner that defeats the purpose of using ML to avoid manual labeling in the first place.   

High Level Description of the Work: We propose a three-stage workflow to obtain dense, pixel level labels in seismic volumes from mere image-level descriptions of few representative structural patches. The first stage entails extraction of textural features in a similarity-based retrieval setup to collect a multitude of image patches representing respective exemplar modalities. Afterwards, we solve an orthogonal non-negative matrix factorization-based optimization problem to obtain per pixel confidence scores for the extracted image patches indicating their belongingness to various structure categories. Lastly, we train a convolutional segmentation model on the extracted patches and their labels to finally infer on the complete volume. 

Related Publications:

  1. Alaudah, Y. and G. AlRegib, "Weakly-Supervised Labeling of Seismic Volumes Using Reference Exemplars," in IEEE International Conference on Image Processing (ICIP), Phoenix, AZ, Feb. 2016. [PDF][Code]

  2. Y. Alaudah, H. Di, and G. AlRegib, "Weakly Supervised Seismic Structure Labeling via Orthogonal Non-Negative Matrix Factorization," in EAGE Annual Conference & Exhibition, Paris, France, Jun. 12-15 2017. [PDF][Code]

  3. Z. Long, Z. Wang, and G. AlRegib, "SeiSIM: Structural Similarity Evaluation for Seismic Data Retrieval," in IEEE International Conference on Communications, Signal Processing, and their Applications (ICCSPA), Sharjah, United Arab Emirates (UAE), Feb. 17-19 2015. [PDF][Code]