Residual Moveout Correction with Deep Learning Models

Residual MoveOuts (RMO) relates to the misaligned primary reflectors after NMO-correction or prestack migration. Post migration alignment can improve subsequent Amplitude Versus Offset (AVO) analysis and inversion. In this work, we proposed a supervised deep learning methodology to train Convolutional Neural Networks (CNNs) to predict time-shifts (TS) for residual moveout correction. We create synthetic training data to address this processing task. We suggest particular adjustments to train CNNs to map misaligned primary reflectors to the corresponding time-shifts needed for alignment. First, we scale the time-shift outputs by a factor of 100 to reduce the disparity between the absolute values of inputs and outputs. Second, we evaluate various CNN architectures to identify those that enhance time-shift prediction accuracy. Our test on synthetic data reports that an encoder-decoder architecture (UNet) without skip connections is effective in generating residual time-shifts for correcting misaligned primaries. We further validate the generalization of our proposed network on a field dataset. The proposed method works in a parameter-free manner during inference time, relieving the user from any manual task or optimal parameter search. As a result, it can provide a new powerful tool for residual moveout correction in existing workflows.