Seismic Data Interpolation Via Deep Convolutional Networks

In the exploration,many factors can lead to the lack of seismic data,which affecting the subsequent interpretation of the data.Therefore,we need to reconstruct the missing seismic data,that is seismic data interpolation.In this thesis,the method of deep convolutional network is used for interpolation.The reconstruction effect of deep convolutional network is better.Its error is smaller than other interpolation methods.The convolutional network has little pre-processing and post-processing.It is more convenient to use.The real problem is often so complex that it cannot be solved by a linear model,or using the linear model maybe lead to very large error.Thus,the result will be not good.In order to solve this problem,the activation function is introduced.But the common activation function has some disadvantages.If we use the sigmoid activation function will have the problems of the gradient disappearance and zero center.Using Hyperbolic Tangent activation function can solve the zero center problem.But it cannot solve the problem of the gradient disappearance.In this thesis,ReLU activation function is selected to solve the problem of limitation of linear model.It can solve the problems of gradient disappearance and zero center.The applied scope of ReLU is wider.The traditional method based on sparse coding is to deal with each layer separately.On contrast,the method of deep convolutional network is to optimize all layers together.The deep convolutional network in this thesis is a lightweight structure,which is end-to-end mapping.It ensures good results and reduces redundant information without tedious operation.