Automated fault detection in the Arabian Basin

Abstract

In recent years, there has been a rapid development of the computer-aided interpretation of seismic data to reduce the otherwise intensive manual labor. A variety of seed detection algorithms for horizon and fault identification are integrated into popular seismic software packages. Recently, there has been an increasing focus on using neural networks for fully automatic fault detection without manually seeding each fault. These networks are usually trained with synthetic fault data sets. These data sets can be used across multiple seismic data sets; however, they are not as accurate as real seismic data, particularly in structurally complex regions associated with several generations of faults. The approach taken here is to combine the accuracy of manual fault identification in certain parts of the data set with a convolutional neural network that can then sweep through the entire data set to identify faults. We have implemented our method using 3D seismic data acquired from the Arabian Basin in Saudi Arabia covering an area of 1051 km2. The network is trained, validated, and tested with samples that included a seismic cube and fault images that are labeled manually corresponding to the seismic cube. The model successfully identifies faults with an accuracy of 96% and an error rate of 0.12 on the training data set. To achieve a robust model, we further enhance the prediction results using postprocessing by linking discontinued segments of the same fault line, thus reducing the number of detected faults. The postprocessing improves the prediction results from the test data set by 77.5%. In addition, we introduce an efficient framework to correlate the predictions and the ground truth by measuring their average distance value. Furthermore, tests using this approach also have been conducted on the F3 Netherlands survey with complex fault geometries and find promising results. As a result, fault detection and diagnosis are achieved efficiently with structures similar to the trained data set.