Training data versus deep learning architectures in the seismic fault attribute computation

Abstract

Seismic coherence is one of the seismic attributes that can be used to aid in seismic fault and stratigraphic interpretation. Traditionally, seismic coherence is computed by measuring the similarity between waveforms of 2D seismic section or 3D seismic volumes. More recently, the computation of a seismic attribute that may be significant in seismic fault interpretation or a seismic fault attribute is treated as an image segmentation problem and using different deep learning (DL) architectures. For doing this, the researchers mainly concentrate on applying cutting-edge DL architectures in computing seismic fault attribute. The available literature on the topic discusses exploring factors, which can be classified into data and method categories, that affect the performance of seismic image segmentation by using DL methods. To explore such factors in computing a seismic fault attribute, we compare the computed fault probability using DL architectures under different scenarios. The designed scenarios aim to highlight the leading factor that may affect the accuracy and resolution of seismic image segmentation. The proposed comparisons are applied to one marine seismic survey from New Zealand and one land seismic survey from China. The results demonstrate that properly preparing training data is far more important than choosing a cutting-edge DL architecture in computing seismic fault attribute. We also propose a practical workflow that can include real seismic data and corresponding labels in training data for a specific seismic survey.