| In seismic exploration,horizon tracing is the basic work of analyzing and processing seismic data,and accurate horizon information can provide better groundwork for data processing and analysis of seismic exploration.Seismic facies analysis technology is used to interpret the geological stratigraphic structure in seismic data,predict the storage of oil and gas,and provide reference for processing seismic data.With the continuous development of data acquisition methods and technologies in recent years,the information contained in seismic data has increased.Traditional methods of horizon tracing and seismic facies identification are laborious,inefficient,and poor in accuracy,and the interpretation process often relies on the subjective empirical knowledge of interpreters.Deep learning is an efficient data-based analysis method that can accurately extract the correlation information and effective features about horizon and seismic facies in seismic data.In this paper,we study the neural network model of deep learning to analyze seismic data and investigate the mapping relationship with seismic stratigraphic levels and seismic facies.To address the problems of high workload,low efficiency and poor accuracy in existing methods,tasks based on computer vision in deep learning are studied to achieve seismic stratum tracking and seismic facies identification.The research is as follows:(1)Seismic facies identification model based on improved residual networkSeismic facies identification is a fundamental task in seismic data interpretation,and using deep learning-based seismic facies identification can greatly improve the efficiency of seismic data interpretation.The large-scale features of existing semantic segmentation models are usually obtained from deep low-resolution feature mapping.It not only sacrifices the spatial resolution and ignores the detail information,but also the simple up-sampling from low resolution to high resolution loses accuracy,resulting in the model’s poor portrayal of the edge fineness between seismic facies categories.To address these problems,a seismic facies recognition model(Res Net-RRM)based on improved edge accuracy is proposed.It adds the RRM(Residual Refinement Module,RRM)module at the end of the improved Res Net network to emphasize feature reconstruction and refinement to capture global information and obtain good edge portrayal.Applying the method to the F3 work area data,the improved semantic segmentation model of the residual network has a higher accuracy and better edge portrayal results than the UNet-based semantic segmentation model in the prediction of the main seismic facies profile.(2)Res NeXt-based seismic horizon tracking modelTo address the problem of insufficient seismic horizon data,we use a 224×224 window to take small Patch images of the main seismic line and contact line profiles respectively to expand the marker data.To address the problem that the number of Res Net parameters in the encoder part of the semantic segmentation model is too high,which leads to a slow training process and increases the convergence speed of the model,we propose the layer tracking model(Att-Res Ne Xt),which uses the Res Ne Xt50 classification network in the encoder part to improve the feature extraction ability with the same number of parameters,simplifies Deep Lab V3+,introduces the channel Space(Convolutional Block Attention Module,CBAM)lightweight attention mechanism to improve model performance.The layer tracking results in F3 work area data show that Att-Res Ne Xt has better tracking results than UNet,PSPNet,Deep Lab V3 and other models,and also has some improvement in accuracy. |
