First Arrival Auto Picking Based On Fully Convolutional Neural Network

At present,with the continuous deepening of seismic exploration and the constant improvement of acquisition technology,high-density single point acquisition is a particular direction of seismic data acquisition,and the amount of seismic data obtained by the unit of seismic exploration is also increasing.In seismic prospecting,geophone,which is arranged on the surface or underground,begins to receive the signal after the blasting point produces the seismic wave.We call the geophone's first useful seismic wave "first arrival".Picking up the first arrival is a necessary and essential work in seismic data processing.With the increasing complexity of the exploration terrain and the significant variation of the first arrival and the interference of all kinds of waves,the existing algorithm of first break picking is not working well.Many times by human-computer interaction,artificial auxiliary lines are added to pick up the first arrival.In the face of massive seismic data,the initial pick up is getting more and more heavy,which dramatically limits the speed of data processing.This paper uses deep learning technology widely to improve the stability and efficiency of the first automatic pickup and reduce the human resources to pick up the first arrival.After a brief analysis of the existing first arrival algorithm and the study of the convolutional network,a volume neural network is proposed to pick up the first arrival.We regard the first arrival pickup as the two classification problem,the first arrival is one kind,and the background is another class.With the successful application of full convolutional network in image semantic segmentation and edge detection,this paper build full convolutional network to pick up the first arrival after learning the existing network structure.It describes in detail the whole process of picking up the first arrival using fully convolutional neural network,including data acquisition,data processing,and data annotation.This paper tests three fully convolutional neural networks of different depths to get the optimal performance.Then compare the result with the commercial seismic data processing software TomoPlus to verify the efficiency and stability of using fully convolutional neural network.This paper applies the rapidly developing deep learning technology to the field of seismic exploration and proposes a new algorithm for picking up the first break using fully convolutional neural network.Through testing three networks of different depths,we choose the network with optimal performance to compare the result with TomoPlus.The result shows that it is efficient and stable to pick up of seismic data by fully convolutional neural network.