Developing a seismic texture analysis neural network for machine-aided seismic pattern recognition and classification

3-D seismic interpretation is essential for subsurface characterization and exploration. Many existing interpretation techniques have been developed for identifying certain seismic features (e.g. faults and salt domes) that are important for depositional facies and hydrocarbon system analysis, but at the expense of deemphasizing or removing other features. Interpreting all the important seismic patterns simultaneously is a long-time issue, although many authors have attempted with the aid of multiple seismic attributes and classification schemes. To resolve this problem, this study proposes developing a seismic texture network (StNet) for automated pattern recognition, classification, and interpretation from 3-D seismic data. The workflow begins with constructing a preliminary seismic texture data set (StData-12), which tentatively categorizes 12 commonly observed seismic patterns based on their signal texture of important geological features such as faults, salt bodies, gas chimneys, depositional facies and stratigraphic features. Then we build the StNet using the state-of-the-art architecture of fully convolutional neural networks and train it based on the constructed data set StData-12. For the seismic data sets we have tested, the StNet is proven to be capable of automatically recognizing and annotating the 12 defined seismic patterns in real time, which allows interpreters to quickly identify the important seismic features simultaneously from a seismic volume. Moreover, we demonstrate that the StNet can be utilized for deriving more task-oriented networks, such as a fault-detection neural network. It is concluded that the proposed StNet as an automated process for machine cognitive data analysis should have broad applications in geological and environmental sciences, mining and drilling engineering, and hydrocarbon exploration. In-depth seismic texture analysis and interdisciplinary collaboration are expected in the future for both enriching the StData-12 and correspondingly optimizing the StNet.