Towards understanding common features between natural and seismic images

Abstract

In this paper, we propose an unsupervised learning framework that aims at evaluating the applicability of the broad domain knowledge from natural images and videos in assisting seismic interpretation, such as seismic attributes, structural automation, and seismic image processing. Specifically, we propose a novel approach based on a data-driven sparse autoencoder architecture that can automatically recognize and extract salient geologic features from unlabeled 3D seismic volumes. It is superior in learning sparse features from natural images, which is not limited by the lack of labeled seismic images. By developing models based on prevalent features in both domains, we can not only automate the process of seismic interpretation but also develop new seismic attributes that highlight areas of interest in seismic sections and convey the most useful information in a compact manner. We show that the proposed approach can effectively detect salient areas within real and synthetic seismic datasets. The experimental results demonstrate the potential of the proposed method in highlighting important geological structures such as horizons, faults, salt domes, and seismic reflections at different orientations and can be effectively used for computer-aided extraction of other geologic features as well.