Deblending using convolutional neural networks

Zhan Niu, Daniel O. Trad

Machine learning has been a booming subject in computer science and its applications have been made in various subjects including geophysics. Convolutional Neural Networks (CNNs) have great potential for solving image processing problems like denoising and interpolation. Deblending, considered as an underdetermined denoising problem, falls into this category. In this report, we use CNN to replace the deblending operator and its performance is analyzed. We use a 4-layer U-Net to perform deblending on synthetically blended shots from a wedge velocity model with point scatterers. We test out different hyper-parameters and the trained model could successfully remove the noise and preserve diffractions from the scatterers with some tolerance. The generality of the model is evaluated by testing the model on an easier 2-layer velocity model. The model can successfully identify and recover most part of the primaries but fails to deal with some interferences and leaves them muted.