A Potential Solution to Insufficient Target-Domain Noise Data: Transfer Learning and Noise Modeling
Xintong Dong, Ming Cheng, Hongzhou Wang, Guanghui Li, Jun Lin, Shaoping Lu
Abstract
Recently, a number of deep learning (DL) methods are developed to attenuate the noise in seismic data. Most of them show good performance under a common precondition: the training and testing data are drawn from the same distribution. However, it is challenging to acquire sufficient noise training data whose distribution is same as the noise presented in the testing seismic data; we call such noise data with same distribution as target-domain noise. To address this issue, we propose a promising DL paradigm for seismic data denoising based on transfer learning and seismic noise modeling. We firstly utilize a Green-function-based modeling method for seismic noise to generate a massive amount of synthetic noise which is similar to real seismic noise. Secondly, the high-authenticity synthetic noise is used as the pre-training data in source domain. Finally, we utilize limited real target-domain noise data to fine-tune the partial trainable parameters of pre-trained model and thus transferring it into target domain. This proposed DL paradigm gets rid of the need for enough target-domain noise data, so as to extend the application scope of DL-based method in seismic data denoising. Moreover, we design a novel network architecture based on multi-cascade structure and attention mechanism. This DL paradigm shows extremely similar denoising performance to that of using a large amount of target-domain noise data in both synthetic and real examples, demonstrating its potential in mitigating the dependence of DL-based seismic denoising methods on target-domain noise data.