Divergence of covariance matrices as a discontinuity attribute for 2D seismic fault detection

Seismic fault interpretation is a crucial step which requires considerable manual intervention by experts. In this work, we aim to develop an efficient fault detection framework. We adopt a multi-stage approach for the task. Initially, we pre-process the seismic data using an iterative anisotropic diffusion filter. Further, patch-based covariance descriptors are computed. A novel discontinuity metric is proposed using exponential-Euclidean Frobenius distance between two adjacent covariance descriptors. The proposed discontinuity metric significantly improves the fault detection performance. Next, Hough transform is applied to extract the possible fault lines, followed by false fault line removal. Finally, optimized fault is detected using support vector regression. The overall efficiency of the fault interpretation framework on both synthetic and real field seismic datasets is discussed in this article.