Desert Seismic Low-frequency Noise Suppression Based On Random Field Prior

As the total demand for oil in China continues to increase,oil and gas exploration has become a hot issue in China's development process.Seismic exploration is one of the commonly used geophysical exploration methods,which is widely used in the intensive investigation stage of the later exploration period.This method is of great significance to probe the geological and tectonic characteristics of the hydrocarbon-bearing favorable areas.The Tarim Basin in northwestern China is one of the major areas for seismic exploration,and its desert areas are rich in oil and gas resources.However,due to the complex geographical environment of desert areas and the perennial wind and dry hot wind hazards,the records obtained from exploration often contain a large amount of low-frequency random noise.The noise has strong energy,low frequency,and serious spectrum aliasing with the effective signal,which greatly reduces the SNR of seismic exploration records.In order to obtain high-quality,reliable seismic information that meets the requirements of the "three highs",and provide intuitive and relevant geological basis for subsequent data interpretation,the study of highly feasible noise suppression methods is of great importance to the oil and gas exploration.At present,domestic and foreign experts have proposed many effective algorithms for the seismic data processing,such as median filter,wavelet denoising,multi-channel fitting algorithm and so on.Although these algorithms have achieved good results in practical,they still have certain limitations in some aspects.In recent years,machine learning has shown better performance and greater potential with the advent of the era of big data.The random field model is a widely used machine learning model.It focuses on establishing the probability distribution of unknown variables,uses undirected graphs to describe the correlation between variables,and reduces the learning task to the calculation task of inferring the probability distribution of variables.The model intuitively and efficiently defines the specific model for the specific task,and has the advantages of efficient inference and simple framework.Based on the excellent performance of this model,the paper applies it to the processing of seismic data.The denoising principle of the random field model is: first select the appropriate potential function to establish the prior distribution of the seismic signal,so as to capture the internal fine structure of the effective reflected wave;then select the appropriate estimation method to ensure the high-fidelity recovery of the effective signal based on the prior information.The denoising effect of the framework is mainly determined by the potential function and the parameters.Based on these two points,it is possible to extend the two algorithms used in this paper.On the one hand,for the potential function,the better the fit of the function to the probability distribution of seismic data,the higher the accuracy of the obtained prior information,and the better the denoising effect.However,this leads to the performance of model in practice being constrained by the potential function,that is,it presents the same function characteristics for signals with different characteristics.Therefore,the model has poor generalization ability.Based on this,the paper adopts the CSF method.This method uses RBFs to indirectly model the prior distribution,which gets rid of the limitation of the potential function and enhances the usability of the model in actual engineering.On the other hand,in order to design more suitable denoising parameters for seismic data,the sparsity measurement module is introduced to the GCRF-MMSE framework.Based on the difference in sparsity between the effective signal and desert low-frequency noise,this module uses the sparsity to determine the denoising parameters,which achieves the targeted removal of the desert low-frequency noise.In this paper,the above two algorithms are applied to the noise suppression of desert seismic exploration records.Both the synthetic and real experiments prove their effectiveness and reliability.By analyzing the denoising results in multiple aspects and angles,it can be concluded that compared with the comparison algorithms,the algorithms in this paper not only restore the energy continuity of the seismic events well,but also effectively suppress the desert low-frequency noise and surface waves.