Search

Detecting subsurface salt structures from seismic images is important for seismic structural analysis and subsurface modeling. Recently, deep learning has been successfully applied in solving salt segmentation problems. However, most of the studies focus on supervised salt segmentation and require numerous accurately labeled data, which are usually laborious and time-consuming to collect, especially for the geophysics community. We have adopted a semisupervised framework for salt segmentation, which requires only a small amount of labeled data. In our method, adopting the mean teacher method, we train two models sharing the same network architecture. The student model is optimized using a combination of supervised loss and unsupervised consistency loss, whereas the teacher model is the exponential moving average of the student model. We introduce the unsupervised consistency loss to better extract information from unlabeled data by constraining the network to give consistent predictions for the input data and its perturbed version. We train and test our novel semisupervised method on synthetic and real data sets. Results demonstrate that our semisupervised salt segmentation method outperforms the supervised baseline when there is a lack of labeled training data.

Salt boundary interpretation is important for the understanding of salt tectonics and velocity model building for seismic migration. Conventional methods consist of computing salt attributes and extracting salt boundaries. We have formulated the problem as 3D image segmentation and evaluated an efficient approach based on deep convolutional neural networks (CNNs) with an encoder-decoder architecture. To train the model, we design a data generator that extracts randomly positioned subvolumes from large-scale 3D training data set followed by data augmentation, then feed a large number of subvolumes into the network while using salt/nonsalt binary labels generated by thresholding the velocity model as ground truth labels. We test the model on validation data sets and compare the blind test predictions with the ground truth. Our results indicate that our method is capable of automatically capturing subtle salt features from the 3D seismic image with less or no need for manual input. We further test the model on a field example to indicate the generalization of this deep CNN method across different data sets.

Relating well-log data, measured in depth, to seismic data, measured in time, typically requires estimating well-log impedance and a time-to-depth relationship using available sonic and density logs. When sonic and density logs are not available, it is challenging to incorporate wells into integrated reservoir studies because the wells cannot be tied to seismic. We have developed a workflow to estimate missing well-log information, automatically tie wells to seismic data, and generate a global well-log property volume using data matching techniques. We first used the local similarity scan to align all logs to constant geologic time and interpolate missing well-log information. Local similarity is then used to tie available wells with seismic data. Finally, log data from each well are interpolated along local seismic structures to generate global log property volumes. We use blind well tests to verify the accuracy of well-log interpolation and seismic well ties. Applying our workflow to a 3D seismic data set with 26 wells achieves consistent and verifiably accurate results.

A 3D seismic image contains structural and stratigraphic features such as reflections, faults, and channels. When smoothing such an image, we want to enhance all of these features so that they are easier to interpret. Most smoothing methods aim to enhance reflections but may blur faults and channels in the image. A few methods smooth seismic reflections while preserving faults and channel boundaries. However, it has not well-discussed to smooth simultaneously along the seismic reflections and channels, which are linear features apparent within dipping reflections. In addition, to interpret faults and channels, extra steps are required to compute attributes or mappings of faults and channels from a seismic image. Such fault and channel attributes are often sensitive to noise because they are typically computed as discontinuities of seismic reflections. In this paper, we have developed methods to simultaneously enhance seismic reflections, faults, and channels while obtaining mappings of the faults and channels. In these methods, we first estimate the orientations of the reflections, faults, and channels directly in a seismic image. We then use the estimated orientations to control the smoothing directions in an efficient iterative diffusion scheme to smooth a seismic image along the reflections and channels. In this iterative scheme, we also efficiently compute mappings of faults and channels, which are used to control smoothing extents in the diffusion to stop smoothing across them. This diffusion scheme iteratively smooths a seismic image along reflections and channels while updating the mappings of faults and channels. By doing this, we will finally obtain an enhanced seismic image (with enhanced reflections and channels and sharpened faults) and cleaned mappings of faults and channels (discontinuities related to noise are cleaned up). We have examined the methods using 2D and 3D real seismic images.

Predictive painting can effectively spread information in 3D volumes following the local structures (dips) of seismic events. However, it has trouble spreading information across faults with significant displacement. To address this problem, we incorporate fault-slip information into predictive painting to correctly spread information across faults. The fault slip is obtained using a local similarity scan to measure local shifts of the different sides of a fault. We have developed three methods to use the fault-slip information: (1) the area partition method, which uses the fault slip to correct the painting result after predictive painting in each divided area; (2) the fault-zone replacement method, which replaces fault zones with smooth transitions calculated with the fault slip information to avoid sharp jumps; and (3) the unfaulting method, in which we use the fault slip information to unfault the volume, perform predictive painting in the unfaulted domain, and then map the painting result back to the original space. Our methods are tested in application of predictive painting to horizon picking. Numerical examples demonstrate that predictive painting after incorporating fault slip information can correctly spread information across faults, which makes the proposed three approaches of using fault-slip information effective and applicable.

Extracting fault, unconformity, and horizon surfaces from a seismic image is useful for interpretation of geologic structures and stratigraphic features. Although others automate the extraction of each type of these surfaces to some extent, it is difficult to automatically interpret a seismic image with all three types of surfaces because they could intersect with each other. For example, horizons can be especially difficult to extract from a seismic image complicated by faults and unconformities because a horizon surface can be dislocated at faults and terminated at unconformities. We have proposed a processing procedure to automatically extract all the faults, unconformities, and horizon surfaces from a 3D seismic image. In our processing, we first extracted fault surfaces, estimated fault slips, and undid the faulting in the seismic image. Then, we extracted unconformities from the unfaulted image with continuous reflectors across faults. Finally, we used the unconformities as constraints for image flattening and horizon extraction. Most of the processing was image processing or array processing and was achieved by efficiently solving partial differential equations. We used a 3D real example with faults and unconformities to demonstrate the entire image processing.