Post-processing of migrated common image gathers can be a time-consuming task, as it requires careful parameterization of post-processing workflows. Additionally, this step must be taken before residual moveout (RMO) picking or velocity-based pore pressure analysis. We train a deep Residual network architecture with an adaptive stochastic gradient descent optimizer using back-propagation to attenuate noise. To achieve compelling performance, we incorporate bottleneck residual blocks, batch normalization, and ReLU pre-activations into the network architecture. We use field seismic from multiple Gulf of Mexico datasets to train and test the 2D neural network. Further, to preserve the timing of events, RMO curvatures, frequencies, and amplitudes, we propose a composite loss function based on the short-time Fourier transform. Experimental results show that the designed loss helps to mitigate the introduction of any edge artifacts, low or high-frequency artifacts, amplitudes shifts, and spurious kinks in the data.

This content is only available via PDF.
You can access this article if you purchase or spend a download.