Abstract

Machine learning, in particular deep learning, is rapidly gaining importance in data processing, including seismic processing and interpretation. It seems that convolutional neural networks (CNNs), which are derived from image processing, show the most promising results. However, different tasks require highly varying training setups and network architectures. We use convolutional encoder-decoder architectures, that is, autoencoders, to analyze and process seismic data at various steps in the processing chain, ranging from denoising and event separation to seismic interpretation. It turns out that autoencoders with similar architecture and training setups are able to deal with seismic data in classification as well as regression problems using both supervised and unsupervised schemes. Results show that they are able to generalize and can be applied to previously unseen seismic data, if trained properly. Nevertheless, they are limited, particularly by data shifts, such as covariate shifts  (Quionero-Candela et al., 2009), and Underspecification (D’Amour et al., 2020). Similar to other types of networks, they can only generalize to data with similar distributions. These claims are demonstrated by examples from seismic processing steps, e.g., denoising, first break picking and interpretation, as well as an approach for automatic in-field volcanic eruption monitoring.