Study of Small Dataset of Images for Seismic Fault Imaging with a Supervised Convolutional Neural Network

Recognizing faults in seismic images is crucial for structural modeling, prospect delineation, reservoir characterization, and well placement. Basically, faults have the appearance of lateral reflection discontinuities in seismic images and are interpreted using seismic attributes that measure those discontinuities such as coherence, and curvature. However, methods based on seismic attributes are often more challenging, time-consuming, and may suffer from noises and sensitivity of stratigraphic features, which also tie in reflection discontinuities. Therefore, we propose a solution for delineating faults from 3D seismic images using a supervised fully convolutional neural network (CNN). This approach uses a pixel-by-pixel prediction in 3D seismic images to classify whether a given pixel is a fault or a non-fault. The trained model learned to bank on rich and proper features that are important for the recognition of faults and achieved 97% of accuracy. To test the effectiveness of our model, we used new 3D seismic images, and the results displayed clean and accurate recognition of faults within only milliseconds, saving time and optimizing the accuracy. In this paper, we showed that by using only a few 3D seismic images from a given seismic volume to train the model, not only do we handle one of the difficulties encountered by researchers to obtain a sufficient amount of data needed to train common CNN models but also, interpreters can successfully predict faults in any other seismic image from the same volume