Seismic spatially variant noise suppression method in the Tarim Basin based on FFDNet and transfer learning

Abstract

Due to a complex geologic structure and ultradeep reservoir location, noise distribution of prestack seismic data in the Tarim Basin is nonuniform. However, most of the current seismic random noise suppression methods lack the flexibility to deal with spatially variant random noise. To address this issue, we have developed an intelligent denoising method for seismic spatially variant random noise and applied it in the Tarim Basin. On the basis of denoising convolutional neural network (DnCNN), we add an extra channel to the input and introduce a tunable noise level map as input. The noise level map has the same dimensions as the input noisy seismic data, and each element in the noise level map corresponds to a denoising level. By adjusting the noise level map, a single model can handle noise with different levels as well as spatially variant noise. Due to the lack of labeled field data in the Tarim Basin, we introduce a transfer learning scheme that transfers features of effective signals learned from synthetic data to the denoiser for field data. The network learns the general and invariant features of an effective signal from a large number of easily obtained synthetic data and then learns the real effective signal characteristics from a small amount of approximately clean field data in the target area by fine-tuning. The processing results of synthetic and field data demonstrate that, compared with f- x deconvolution, dictionary learning, and DnCNN, our method exhibits high effectiveness in suppressing spatially variant random noise and preserves effective signals better.