Self-similarity convolution neural network for seismic noise suppression in desert environment

Seismic signals are inevitably disturbed by random noise in the acquisition process, which greatly degrades seismic data. In order to improve the quality of seismic data, we propose a self-similarity convolutional neural network (SS-Net) for seismic data denoising by introducing the coherence of seismic events into convolutional neural network (CNN). The SS-Net consists of two modules, the directional matching module (DMM) and the denoising module. The DMM stacks similar seismic data blocks to generate three-dimensional (3D) groups by calculating the similarity between seismic data blocks with the same directional characteristics. For the 3D groups with redundant structural information, the following denoising module with the multi-channel convolution adaptively extracts and squeezes the structural feature characteristic of each 3D group, which enhances the characteristics of seismic signals and avoids confusion caused by local similarity of seismic signals and random noise. In addition, the skip connection is adopted by SS-Net to transport the sparse feature to the following denoising process, to reduce the loss of signal features extracted by multi-channel convolutional layers due to increased network depth. We validate the denoising performance of the SS-Net on the synthetic and field desert seismic data. The filtered results confirm that the SS-Net can suppress seismic random noise more thoroughly and recover the seismic events with complex morphology better than the competitive denoising methods.