Research On The Methods Of Random Noise Suppression In Seismic Exploration By Using Improved MIRNet Algorithm

In seismic acquisition,seismic waves are artificially excited in the ground and as they travel through the stratigraphic medium,they encounter different rock partitions with different medium properties,resulting in reflection and refraction phenomena.A geophone at the surface records the reflected seismic waves,and the recorded seismic signal is related to the characteristics of the source,the way the detection points are arranged,and the properties and configuration of the various subsurface media.Seismic signal processing is used to process and analyse the seismic signals received to infer the structure and distribution of subsurface rocks,providing information about the subsurface medium and reliable data for geological interpretation.However,in the actual seismic data acquired,there is often a lot of noise interference from the external environment.Therefore,the removal of background noise is a very important task in practical applications.Deep learning,as a data analysis method based on a large number of training samples,can better solve complex problems such as high dimensionality and redundancy that cannot be solved by traditional machine learning methods.In recent years,the most representative deep learning technique,Convolutional Neural Networks(CNN),has undergone rapid development,of which MIRNet is a convolutional neural network capable of retaining both high-resolution spatial detail and low-resolution contextual information.Compared to traditional convolutional neural networks,MIRNet can learn richer features and has achieved excellent results in areas such as image super-resolution and image restoration.However,as the depth of the network increases,simply increasing the depth of the network gradually leads to a loss of shallow information,resulting in the shallow information not being fully exploited.At the same time,most common convolutional neural networks,including MIRNet,train the network in a single time or frequency domain,ignoring the diversity of feature information extraction,resulting in a very limited number of effective features that can be extracted by the network,thus limiting the performance of the network.Based on this,this paper proposes to improve MIRNet and apply it to seismic random noise suppression.First,to address the inability of the deep MIRNet to fully exploit the shallow features extracted by the network,a dense connection approach based on an attention mechanism and a global feature fusion mechanism are introduced in the network,thus proposing a MIRNet with dense feature fusion(DFF-MIRNet),which can fully exploit both the extracted shallow and deep information and better achieve the suppression of random noise in seismic records.Second,although DFF-MIRNet fully exploits the shallow feature information and achieves a better denoising effect,DFF-MIRNet only performs time-domain feature extraction,ignoring the diversity of feature information.Therefore,the frequency domain branching and time-frequency domain information interaction mechanism are added to the DFF-MIRNet algorithm,and then the MIRNet with Time-Frequency Domain Dense Feature Fusion(TFDFF-MIRNet)is designed to enhance the diversity of feature extraction in the network by using a dense connection based on the attention mechanism to exchange information between time-frequency domain branches and extract feature information through both time-frequency domain branches.The interaction of information between the two branches strengthens the link between the time-frequency features,allowing the network to make full use of both extracted time-frequency features.To evaluate the effectiveness of the two improved MIRNet algorithms,the simulated seismic records and actual seismic records collected in the field will be processed and analysed by two trained network models,and the resulting data will be compared with traditional seismic denoising methods and the processing results of MIRNet.The experimental results show that these two improved MIRNet algorithms can achieve better denoising effects in seismic random noise suppression,effectively improving the signal-to-noise ratio of seismic data and providing a reliable theoretical basis for subsequent seismic data interpretation.