Abstract:

Reverse-time migration (RTM) is widely employed for its ability to produce high-resolution imaging results. Nevertheless, wavefield extrapolation based on the two-way wave equation often leads to pronounced low-frequency noise under cross-correlation imaging condition. This issue significantly impacts the quality of the resulting images. In this study, we introduce a convolution-deconvolution neural network (DeCNN) built upon the U-Net architecture to mitigate the presence of low-frequency noise in RTM. We utilize source-normalized imaging results with  low-frequency noise as training data， and the Laplace filtering results   as labels to abtain  the neural network model, grounded in a data-driven approach. The comparison results of   model trials and transfer learning demonstrations highlight the superior denoising proficiency of  DeCNN. It outperforms  U-Net, effectively suppressing noise in scenarios such as the transfer  SEG/EAGE standard salt dome model and  Marmousi model. In contrast to conventional techniques aimed at mitigating low-frequency noise through up and down traveling wave decomposition, DeCNN  can input the source-normalized imaging results into the network after training, and the network can output high resolution imaging results in seconds. The efficiency is much higher than that of conventional low-frequency noise suppression methods using wavefield decomposition, and it has certain advantages in noise suppression effect.

Key words: neural networks, convolution-deconvolution, reverse-time migration, low-frequency noise, filtering, wavefield decomposition