SEISMIC SIMULTANEOUS SOURCE SEPARATION VIA AN UNSUPERVISED DEEP LEARNING METHOD

Separation of blended seismic data acquired in simultaneous source acquisition is a key step in seismic data processing. In the context of sequential time-dithering firing, we construct the separation of the blended seismic data as an inverse problem, and present a novel unsupervised deep learning method. Neither the pre-training procedure nor the training dataset are required in our method, which is quite different from existing deep learning based deblending methods. In particular, Deep Image Prior (DIP) is introduced as the implicit regularization. The useful information of the recovering unblended seismic data can be iteratively captured by the generator network. Tests on synthetic and field data demonstrate that the recovery data obtained from our presented method has high separation accuracy.