Research And Application Of Self-similarity Convolution Neural Network For Desert Seismic Noise Suppression

Seismic exploration is one of the important means of oil,natural gas and other resources exploration.In the process of seismic exploration in desert area,the random noise is complicated because the vegetation cover area is small.The random noise has the characteristics of non-linear,non-Gaussian and non-stationary.At the same time,in some local areas,it shows similar properties to seismic signals,and complex random noise seriously interferes with the identification and accurate recovery of weak seismic signals.Therefore,utilizing the block structure similarity of seismic signals,by dividing seismic signals into blocks,utilizing the similarity and redundant information between signal blocks to enhance the correlation of their effective signals,laying the foundation for seismic signal denoising algorithms,contributing to high-quality seismic imaging and interpretation,and having significant significance for precise exploration of important resources such as oil and natural gas.The local similarity between random noise and seismic signals can cause significant interference when the convolutional neural network(CNN)based denoising method lacks the ability to extract features,leading to incomplete suppression of random noise and distortion of effective signal structures.To effectively suppress the desert random noise,a self-Similarity convolution neural network(SS-Net)is proposed by combining block matching algorithm and convolution neural network.SS-Net model is composed of the directional matching module(DMM)and the denoising module.Then,the data blocks in the three-dimensional(3D)group generated by the direction matching module are similar,containing the coherence of seismic events.With rich redundancy,this 3D group is used as the input of subsequent denoising modules,promoting the SS-Net model to improve feature extraction and denoising capabilities using the coherence of seismic events.Specifically,the denoising module utilizes multi-channel convolution to adaptively extract self-similarity features from 3D group,and aggregates redundant similar feature maps along channels to sparse similar feature maps.The enhanced features guide the subsequent denoising process to better recover the complex structure and detailed information of seismic events.In addition,the SS-Net model uses skip connection to introduce the sparse features extracted from the shallow layer into the subsequent network layer,avoiding the loss of effective structural features due to the deep network layer.This paper constructs a dataset suitable for suppressing desert random noise,and discusses the optimal selection of model parameters(channel number and block size)that have a significant impact on desert random noise suppression.The synthetic desert seismic data is used to verify the suppression ability of the SS-Net model on desert random noise,and the 3D groups and feature map extracted by the directional matching module and the denoising module are visualized respectively,further demonstrating that 3D groups with rich redundant generated the directional matching module can enhance the feature extraction ability of the SS-Net model.The effectiveness and practicability of the SS-Net model are verified by using two filed desert seismic data.The results show that SS-Net model can effectively suppress the desert random noise,and better recover the complex structure and details of seismic events.Compared with Curvelet,NLM,FK filter and Dn CNN module,the denoising effect is more significant.