| Seismic data interpretation is an important part of seismic exploration,and seismic horizon tracing is one of the most basic and critical techniques in seismic data interpretation.The accuracy of its tracing results will have an important impact on the accuracy of subsequent tectonic interpretation and reservoir prediction.Manual tracking of seismic horizon is time-consuming and labor-intensive,and it takes up most of the workload in the seismic data interpretation and analysis stage.Intelligent seismic horizon tracking has become a frontier hotspot and development trend in the industry.In order to improve the tracking efficiency and accuracy,this paper carries out a study on automatic seismic horizon tracking based on deep learning.The main research contents are as follows:First,this paper conducts experiments on automatic horizon tracking on U-Net,an image semantic segmentation network,and verifies the feasibility of using U-Net to realize the automatic seismic horizon tracking problem,but the accuracy of the tracking results still needs to be improved.Then,we propose a horizon tracking method based on U~2-Net,a two-layer nested U structure,in which the ordinary convolutional blocks of U-Net are replaced by Re Sidual U(RSU)blocks to fuse the features of different scales of perceptual fields,to address the characteristics of layers in seismic data and the shortcomings of the original U-Net feature learning.Experimental results show that the U~2-Net-based horizon tracking method effectively improves the accuracy of horizon tracking,but its network parameters increase more than the original U-Net.Finally,in order to obtain higher horizon tracking accuracy using lightweight convolutional neural networks,this paper proposes a knowledge distillation-based horizon tracking method with the knowledge distillation method as the core.The core idea of knowledge distillation is to use the complex network as the teacher network and the simple network as the student network,and use the teacher network to guide the training of the student network so that the lightweight student network can get the approximate performance of the teacher network.In this paper,we use U~2-Net as the teacher network and U-Net as the student network,and the final experimental results show that the horizon tracking method based on knowledge distillation ensures the accuracy of horizon tracking and the model complexity is lower than that of U~2-Net. |
