Attenuation Of Coherent Noise Based On Convolutional Neural Network

Seismic technology has always been the most important means to explore the geological structure.However,during seismic data acquisition,geophones often receive abundant noise due to poor source excitation,receiving conditions,surface complex geological conditions,and the heterogeneity of subsurface lithology.Noise always damages the continuity of reflection events and interferes with the processing and interpretation of seismic data.According to the law of different noise propagation,they can be divided into two categories:coherent noise and random noise.Compared with random noise,coherent noise has more serious interference to signals because of its strong energy.Therefore,the removal of this noise represents an essential step in the processing of seismic data.Traditional coherent noise suppression methods are mostly based on separating the signals from noise in the frequency-wavenumber domain or other transform domains.Although most of these methods can effectively suppress coherent noise,they are mostly based on some assumptions,such as coherent noise is approximately straight.When the hypotheses are false,satisfactory results can not be obtained.Moreover,when coherent noise overlaps with signals in transform domains,it is easy to damage effective signals while removing noise.The denoised methods based on convolutional neural network algorithms are not subject to certain conditions,and they do not require prior knowledge of geologic information or the propagation laws of seismic waves.The convolutional neural network is a relatively recent development in the field of seismic data processing including noise attenuation.However,the research on coherent noise attenuation based on convolutional neural networks is less and thus worthy to be launched.The problem of coherent noise attenuation based on convolution neural networks is studied in this paper.Here,the network structure is designed and modified according to the characteristic of coherent noise.Firstly,we add the asymmetric convolutional kernels and form blocks to replace the regular square kernels for extracting features of coherent noise.Then,synthetic,numerical simulated,and real land seismic data are used to make training datasets.Through the adjust of hyper parameters,the influence of different learning rates and convolutional kernels numbers is demonstrated in the paper,and the value of hyper parameters is determined.Finally,the attenuation performance is compared with the denoised results off-k filtering,median filtering,and the denoising convolutional neural network(DnCNN).Moreover,the denoised result of data from another geological area without retraining demonstrates that the proposed network has generalization ability.