Unsupervised seismic data deblending based on the convolutional autoencoder regularization

Simultaneous source technology can provide high-quality seismic data with lower acquisition costs. However, a deblending algorithm is needed to suppress the blending noise. The supervised deep learning methods are effective, but are usually limited by the lack of labels. To solve the problem, we propose an unsupervised deep learning method based on acquisition system. A convolutional autoencoder (CAE) network is employed to predict the deblending results of the input pseudo-deblended data. And then, the deblending results will be re-blended using the given blending operator. The parameters of CAE will be optimized by the difference between re-blended data and input data, which is defined as 'blending loss.' The blending problem is ill-posed but the CAE can be regarded as an implicit regularization term which constrains the solving process to obtain the desire solution. A numerical test on synthetic data demonstrates that the proposed method can converge to correct results and two field data experiments verify the flexibility and effectiveness of our model. The transfer training method is also used to improve model performance.