Ground roll interpolation and attenuation

Ji Li, Daniel O. Trad

Ground-roll attenuation is an essential step in land seismic data processing. Simple methods like bandpass filter or f-k filter can remove the ground roll based on the difference in the frequency domain. However, the performance is often limited due to the spatial aliasing of the ground roll, which causes the overlap of the ground roll with reflections in the frequency domain. An interpolation step of the seismic data can improve the final ground roll attenuation performance. We adopt a convolutional neural network-based framework named Residual dense networks (RDN) to interpolate the seismic data with a strong ground roll and weak reflections. The purpose is to interpolate the strong ground roll and weak reflections simultaneously. We first create a training dataset via the finite different method to train the model. Then, we test seismic data interpolation on another dataset, which is not apart of the training dataset. We compare the interpolated shot gather with and without the ground roll. The interpolated results prove that the proposed approach can interpolate the strong ground roll and keep the weak reflections simultaneously. After the interpolation,a simple f-k filter is applied to both the synthetic and real data examples to attenuate the ground roll. The result is also compared with the traditional F-X interpolation approach.