Research On High Resolution Seismic Data Based On Deep Learning

Improving the resolution of seismic data is one of the important links in the seismic data processing process,which plays a vital role in the subsequent structural interpretation.The traditional methods to improve seismic resolution are often based on certain assumptions,which lead to the difficulty in obtaining ideal processing results under different environmental conditions,and the complexity of the operation and the processing time are too long.In the process of field seismic data acquisition,due to the influence of the earth environment,geological structure,acquisition instruments and other factors,the collected seismic data often has problems of missing channels,bad channels,data missing,and a lot of environmental noise will be added in the process of acquisition and processing,which has an impact on improving the resolution of seismic data.Deep learning is a very popular research direction in the field of artificial intelligence,with the characteristics of high generalization and operability.Convolution neural network,as a branch of deep learning,is particularly prominent in the field of image processing.It completely relies on data-driven,does not need to make any conditional assumptions,and only needs to provide enough input data and label data for the neural network to learn,so that it can learn the relationship between data independently,establish a complex nonlinear mapping relationship between data,and finally predict the ideal data we want.In this thesis,the convolution neural network is applied to improve the resolution of seismic data.First,many factors affecting the resolution of seismic data are briefly analyzed,and the ideal seismic wavelet commonly used in seismic exploration is given;Then the basic structure and super parameters of the convolution neural network are introduced,and some common parameter adjustment references are given;Then it introduces the principle and process of improving the resolution of seismic data based on deep learning,analyzes in detail the matters needing attention in making the training set and the training parameters of the network model,gives the performance comparison of the commonly used convolutional neural network based on the UNet framework,and tests the high-resolution processing effect on different types of geological models;Finally,through the actual data test,the convolutional neural network based on RAUNet can achieve good results in high-resolution processing.