In the prestack depth migration, obtaining a high-quality seismic image requires an accurate common image gather (CIG) where the diffractions are focused, the reflection events are flat and the depth of the events are correctly migrated. In practice, such an accurate CIG, however, is hard to obtain due to the errors of acquisition, processing, and the migration velocity model. We propose a kernel prediction neural network (KPN) to flatten the events, correct the depth of the improperly migrated events, and remove the noise and unfocused artifacts in CIGs, and further yield an optimally stacked image. In particular, KPN produces spatially varying kernels that can align and denoise the CIGs. The application of the KPN includes two steps: training and prediction. In the training process, we design a basic loss function to fit the stacked image of the predicted CIGs and the target image, and we also incorporate an annealed loss to match the image slice at each offset with the target image which helps correct CIGs. The training samples are the CIGs migrated with inaccurate velocity models. The correspondingly targeted images are obtained by the convolutions between the Ricker wavelet and synthetic reflectivity models. The trained network is applied to stack new input CIGs to obtain a stacked image, which is automated and not subject to human bias. We demonstrate the effectiveness of the new method using the synthetic Marmousi dataset.
Presentation Date: Monday, October 12, 2020
Session Start Time: 1:50 PM
Presentation Time: 3:05 PM
Location: Poster Station 7
Presentation Type: Poster
