3D fault detection: Using human reasoning to improve performance of convolutional neural networks

Abstract

Depicting faults in seismic data is one of the key steps in seismic structure interpretation. However, the manual identification of faults is a time-consuming and tedious process. In conventional methods, seismic attributes associated with the reflection continuities or discontinuities of seismic data are extracted for fault detection. In recent years, convolutional neural networks (CNNs) have been introduced to solve geophysical problems. Herein, we have proposed powerful and efficient methods to enhance the performance of CNN-based fault detection. We first introduce human reasoning to improve the performance of a 3D neural network, which is trained with synthetic seismic data. We propose two methods that efficiently use human reasoning. Manually identified faults can be easily added to train an efficient spatial pyramid network (ESPNet) and contextual information can be obtained to improve the continuity of the faults. We propose an unsupervised fault registration framework to fuse the results from commonly used 3D CNNs and ESPNet, which simultaneously preserves high data accuracy and continuity. Furthermore, to use sparse manually identified faults to directly train a 3D CNN, we propose a weakly supervised learning method. Multiple examples find that faults can be more accurately predicted with the proposed methods than with previously established algorithms, including conventional seismic attributes and CNN-based methods.