Research On The Fault Detection Method Of Seismic Image Based On Semantic Segmentation

Seismic image fault detection is of great significance in hydrogeology,solid mineral resources investigation,petroleum geology investigation,seismology-related research,and earthquake prediction.The traditional fault detection method has a manual marking,but this method is time-consuming and prone to errors.The method of seismic attribute detection is based on the discontinuity of seismic attribute profile caused by fault,such as coherence,coherence,variance,and curvature.In addition,some researchers propose to use automatic tracking and interpretation methods,such as ant algorithm for fault detection.However,these methods have some disadvantages,such as unclear fault detection and inaccurate location.With the development of artificial intelligence in computer vision,the deep learning method is widely used in various fields.The semantic segmentation method can classify the image at the pixel level,separate the segmentation target from the background,and accurately locate and classify the target.Seismic image fault detection is of great significance in the fields of hydrogeology,solid mineral resources investigation,petroleum geology investigation,seismology related research,and earthquake predictionIn this paper,Channel-UNet semantic segmentation network is proposed based on SENET and UNet,and the fault detection of the seismic image is carried out.The network uses the U-shaped basic framework of UNet network,including downsampling and upsampling.Downsampling is used for feature extraction,and upsampling is used for feature fusion.U-shaped network structure can get deep information and shallow information.Deep information can reflect the characteristics of the relationship between the segmentation target and its environment and can be used for classification.The shallow information can provide more detailed features for the segmentation task.Channel-UNet network also integrates the channel attention module of SENET.SENET focuses on the relationship between channels and proposes a feature re-calibration method.The importance of each feature channel is obtained through autonomous learning,and then the valuable features are emphasized,and the unimportant features are suppressed according to the importance.It is difficult to obtain enough seismic images containing apparent faults.It is a time-consuming task to train a neural network with real seismic data,which involves the confidentiality of land and resources.The interpretation of seismic fault depends on the interpreter's experience to a great extent,which is prone to errors.Therefore,this paper uses an amplitude-based seismic image generation platform to automatically generate training data sets,customize fault parameters and generate corresponding fault labels.Then the synthetic seismic dataset was used to train multiple semantic segmentation networks,including UNet,Resnet34?UNet,R2U-Net,Attention UNet,and the Channel-UNet proposed in this paper with Channel Attention mechanism.Then the trained network was used to verify the synthetic dataset.The experimental results show that the fault detection effect of the Channel-UNET network is better than other semantic segmentation networks,and the IOU and dice values calculated by comparing label data are higher than other semantic segmentation networks.At the same time,Channel-UNet network is also used for fault detection of real seismic data.In this paper,the seismic data of the F3 block in the Holland Beihai area are obtained and visualized using Opendt,a seismic interpretation software.Then the Channel-UNet network is used to detect the fault of the F3 real seismic image.In addition,to make the fault easier to observe and connect the fault lines that should belong to the same fault,the Hough transform is used to post-process the detection results of the neural network.In this paper,the traditional geophysical interpretation method is used to mark the original data of artificial fault.The fault predicted by the network is compared with the artificially marked fault.The fault detection accuracy is high by using FAD,which indicates that Channel-UNet network can effectively detect the real seismic image.