Deep learning for seismic data: a practical guide for model development and evaluations

What is it about?

Deep learning allows us to train predictive models that can solve complex tasks, such as detecting specific patterns in noisy signals. These models however do not always work well when applied to new data that is different from the data originally used to train them, which is a common case in practice. We propose a dataset and a protocol to train and evaluate predictive models that allows us to observe how well they would perform in such scenarios. The dataset we use is composed of multiple hardrock seismic surveys. Our protocol suggests that we train models on some surveys and then evaluate on others. Following this protocol will help future researchers build better models that are more likely to work in real-world applications.

Featured Image

Photo by Clint Adair on Unsplash

Photo by Clint Adair on Unsplash

Why is it important?

We propose a machine learning model training and evaluation methodology for first break picking based on a new multi-survey dataset. In contrast with prior works, our methodology can be used to assess how well predictive models would work in realistic settings where first breaks must be provided for surveys without annotations. We benchmark and discuss the performance of a well-known deep neural network architecture (U-Net) following this methodology, and provide ideas for future model improvements.

Perspectives