Data-driven edge detectors for seismic data interpretation
We present a novel method to assist in seismic interpretation. The algorithm learns data-driven edge-detectors for structure enhancement when applied to time slices of 3D poststack seismic data. We obtain the operators by distilling the local and structural information retrieved from patches taken randomly from the input time slices. The filters conform to an orthogonal family that behaves as structure-aware Sobel-like edge detectors, and the user can set their size and number. The results from marine Canada and New Zealand 3D seismic data demonstrate that the proposed algorithm allows the semblance attribute to improve the delineation of the subsurface channels. This fact is further supported by testing the method with realistic synthetic 2D and 3D data sets containing channeling and meandering systems. We contrast the results with standard plain Sobel filtering, multidirectional Sobel filters of variable size, and the dip-oriented plane-wave destruction Sobel attribute. The proposed method gives results that are comparable or superior to those of Sobel-based approaches. In addition, the obtained filters can adapt to the geological structures present in each time slice, which reduces the number of unwanted artifacts in the final product.