Studies Of 3D Seismic Horizon Tracking Method Based On Deep Learning

Through horizon tracking we can know the location of underground interface and the distribution of underground geological structure,which can help locate and explore oil and gas resources accurately.Horizon tracking is to determine the position of the interface between different strata.It is a basic work in seismic data interpretation.The accuracy of the tracking results will have an important impact on subsequent data interpretation.Traditional horizon tracking technology mainly depends on manual operations to interpret on two-dimensional seismic data.Interpreters track the seismic horizon according to the property of the horizon,which largely influenced by interpreters' experience and subjective judgment.The deep learning method is based on a large amount of data,and can extract the features of the data by constructing a proper network model,which can make data predictions.Based on the U-net network,this paper constructs a network model which is suitable for 3D seismic data by improving the network structure and loss function,and realizes the automatic tracking of 3D seismic horizons.Firstly,we test the two-dimensional automatic horizon tracking experiment based on the original U-net to verify that it is feasible to solve the horizon tracking problem via Unet,but the tracking results still is not so good.Then by adjusting the U-net network structure and improving the loss function,an improved U-net++ is proposed.Compared with the tracking result of U-net,the result of U-net++ has higher continuity.In order to make full use of the spatial morphological characteristics of the horizons in the threedimensional seismic data,based on the superior performance of the U-net++ in twodimensional tasks,this paper extends the U-net++ to the three-dimensional seismic data,and adjusts the network structure.Experiments prove that the 3D U-net++ network effectively improves the accuracy of the prediction results and also improves the noise resistance,and achieves better experimental results than the two-dimensional network,and better completes the automatic tracking task of the 3D seismic horizons.