Seismic Data Through a ResNet18 Lens

Sergey Tsimfer27 May 2023

Problem

Gazprom Neft trains neural networks on seismic cubes, large 3D images of a field, to find horizons and faults. A model is trained on a few dozen cubes and then applied to new fields. Each cube was processed for years by different people on different equipment, so cubes look different and transfer to new data is uncertain. The company also stores each cube in about 10 systems under different names and slightly changed coordinate headers, so nobody knows which copies are the same cube.

Example from the article: a classifier trained to name the source cube of a small patch reaches about 98% accuracy on 10 cubes, so cubes are numerically distinguishable and their similarity can be measured.

Solution in brief

Train a small classifier (pruned ResNet18) to guess which dataset a patch comes from. If it succeeds, the datasets differ, and the distance between its embeddings measures how much. Confusion between two datasets exposes near-duplicates, and the same distance picks training cubes for a new field.

Questions this case answers
  • How to measure numerically how much two large sets of 3D images differ and which copies are near-duplicates
  • How to pick the training datasets closest to a new target dataset without labels
  • How to measure numerically how much two large datasets differ from each other
  • How to find duplicate or near-duplicate copies of large files when hashes fail because of metadata and filtering changes

Tools and methods

  • Pruned ResNet18 — patch classifier, about 4x reduced, trains in minutes
  • Embeddings and t-SNE — last-layer embeddings, one cluster per cube
  • Transformation ablation — a series of image distortions that shows what information still identifies a cube
Continue reading — it's free

Google or a one-time email code. No password.