Combining classical processing with Deep Learning
Daniel O. Trad
Machine learning (ML) is a powerful tool that has become very useful in many areas of science and seismic applications are not an exception. For example, the uses of ML in interpretation have proven to be very helpful. Also, in seismic processing, we see a significant effort for using ML techniques to help or replace traditional processing. Processing applications, however, present a more difficult challenge. In this report, we will discuss what some of those challenges are, and also we will discuss an environment to research this type of technique, where ML does not replace but rather cooperates with traditional signal processing and inversion. We will present this methodology with several examples: migration, multiple attenuation with Radon transform, near-surface noise (ground roll) suppression and interpolation. For the purpose of illustrating the flexibility of machine learning and the importance of data preprocessing, we will address all these problems by using the same deep learning network but changing only inputs and outputs. The goal of the report will be to emphasize the importance of combining conventional and neural network techniques.