Deep Learning Method for Latent Space Analysis

Abstract

Seismic attributes are a well-established method for highlighting subtle features buried in seismic data in order to improve interpretability and suitability for quantitative analysis. Seismic attributes are a critical enabling technology in such areas thin bed analysis, 3D geobody extraction, and seismic geomorphology. When it comes to seismic attributes, we often suffer from an "abundance of riches" as the high dimensionality of seismic attributes may cause great difficulty in accomplishing even simple tasks. Spectral decomposition, for instance, typically produces 10's and sometimes 100's of attributes. However, when it comes to visualization, for instance, we are limited to visualizing three or at most four attributes simultaneously. My co-authors and I first proposed the use of latent space analysis to reduce the dimensionality of seismic attributes in 2009. At the time, we focused upon the use of non-linear methods such as self-organizing maps (SOM) and generative topological maps (GTM). Since then, many other researchers have significantly expanded the list of unsupervised methods as well as supervised learning. Additionally, latent space methods have been adopted in a number of commercial interpretation and visualization software packages. In this paper, we introduce a novel deep learning-based approach to latent space analysis. This method is superior in that it is able to remove redundant information and focus upon capturing essential information rather than just focusing upon probability density functions or clusters in a high dimensional space. Furthermore, our method provides a quantitative way to assess the fit of the latent space to the original data. We apply our method to a seismic data set from the Canterbury Basin, New Zealand. We examine the goodness of fit of our model by comparing the input data to what can be reproduced from the reduced dimensional data. We provide an interpretation based upon our method.