De‐migration‐based supervised learning for interpolation and regularization of 3D offset classes

Regularization and interpolation of 3D offset classes prior to imaging are an important challenging step in the marine seismic data processing flow. Here we describe how perform this task using a deep neural network, explain overcome challenge creating suitable training set. The set is generated by de-migrating stacked pre-stack depth migration images. For each class volume, de-migrate migrated image into two configurations: (i) original survey configuration consisting recorded source/receiver positions (ii) ‘Ideal’ with constant azimuth for class. creates convolutional encoder–decoder model that will regularize interpolate data. trained on sliding windows cube map from (ii), i.e. irregular sparse sampling fully sampled regular cubes offset-based migration, such as Kirchhoff migration. Such algorithms rely sufficiently dense achieve constructive interference structures destructive suppress noise. We test new method one synthetic field example show it performs better than standard regularization/interpolation based anti-leakage Fourier transform, especially smallest classes. On data, also demonstrate preserves amplitude versus well method.