Salt interpretation using seismic data is essential for structural interpretation and oil and gas exploration. Although deep learning has made great progress in automatic salt image segmentation, it is often difficult to obtain satisfactory results in complex situations. Thus, interactive segmentation with human intervention can effectively replace the fully automatic method. However, the current interactive segmentation cannot be directly applied to 3D seismic data and requires a lot of human interaction. Because it is difficult to collect 3D seismic data containing salt, we propose a workflow to simulate salt data and use a large amount of 3D synthetic salt data for training and testing. We use a 3D U-net model with skip connections to improve the accuracy and efficiency of salt interpretation. This model takes 3D seismic data volume with a specific size as an input and generates a salt probability volume of the same size as an output. To obtain more detailed salt results, we utilize a 3D graph-cut to ameliorate the results predicted by the 3D U-net model. The experimental results indicate that our method can achieve more efficient and accurate segmentation of 3D salt bodies than fully automatic methods.

Salt body extraction plays an important role in the analysis of salt structures and the exploration of oil and gas. Seismic attributes and edge detection algorithms, which require manual effort, are the conventional methods of extracting salt boundaries from seismic images. Convolutional neural networks (CNNs) have become the state-of-the-art automatic segmentation method for seismic interpretation. However, the fully automatic results of the extraction of salt boundaries may still need to be modified to become accurate and robust enough for practical production. We present a novel deep-learning-based interactive segmentation method for extracting salt boundaries. To incorporate the interaction points into our method, we transform positive and negative points into two Euclidean distance maps (EDMs), which are combined with seismic images to train our CNN model. The model is composed of a U-net and a pyramid pooling module, and it is trained on the TGS Salt Identification Challenge dataset. Then, we use a graph cut algorithm to refine the likelihood maps predicted by our CNN model and subsequently update the salt boundaries. Some field examples show that the proposed method outperforms fully automatic CNN methods with a higher matching degree of the ground truth.

Deep learning-based seismic horizon tracking methods have been extensively researched in the past few years. However, the predicted results of previous methods are currently unstable and require additional processing to obtain reasonable interpretations, limiting their applications to field seismic data. There are two potential reasons for this: (1) the training labels may not consider the uncertainty in horizon interpretation, and (2) the characteristics of the seismic data used for training may differ from those for prediction (domain shift). In addition, the previous deep learning methods are mostly based on point-by-point cost functions, which are not well-suitable for horizon tracking problem. To address these issues, we proposed a method that mimics the process of manual horizon interpretation to avoid the domain shift problem and regard horizon identification and tracking as a problem of conditional probability density estimation using deep learning. In the North Sea F3 seismic data experiments, our proposed method can predict horizons accurately and stably when we select only 1.6% of the seismic sections to create labels. Taking mean absolute error (MAE) and accuracy as evaluation indices, the MAE for single horizon prediction is as low as 2.0 ms, and the accuracy reaches 98.1%. A surprising finding revealed by the experiments is that the predicted results could provide a perspective of underground structures, such as faults and salt domes, demonstrating the feasibility of our method in interpreting complex field seismic data.

Fluvial facies exhibit higher porosity and permeability during sedimentation, making it crucial to study reservoir characteristics and assess the potential for oil and gas exploration. Traditional methods for identifying fluvial facies often rely on specific seismic attributes, which often require manual extraction of channel feature, especially when interpreting 3-D seismic data, which is inefficient. To improve the efficiency of channel interpretation in 3-D seismic data, we proposed a two-stage convolutional neural network to implement an interactive 3-D channel interpretation method. We generated 3-D seismic data with real channel structures and used their seismic attributes as inputs to the first stage network to automatically obtain initial and rough channel results. Then, based on this result, we added manual interaction to mark errors and combined the geodesic distance to transform the manual interaction information. Finally, we input the interaction information into the second-stage network to obtain high-quality identification results. Synthetic and field data examples demonstrated that this method retains good applicability even under complex geological conditions, providing valuable insights for the fields of oil and gas exploration and geological research.

In the exploration of oil and gas reservoirs, channels are important locations for storing oil and gas, and their distribution in the subsurface is usually heterogeneous. Therefore, interpreting channels from seismic data is very meaningful for oil and gas exploration. Currently, methods for extracting channels from seismic data mainly rely on seismic attributes, edge detection algorithms, and deep learning. However, these methods cannot fully and accurately delineate the boundary details and structural characteristics of channels when faced with poor- quality seismic data. To address this issue, we proposed an interactive interpretation method for 2D channels based on multi-attributes and a convolutional neural network (CNN) that could more accurately identify and segment channel bodies with fuzzy boundaries and poor continuity. First, we selected seed points from seismic data to indicate the presence of channels in the area. To highlight the channel structures and reduce the difficulties in identification, we used the geodesic distance map calculated from the seed points and two seismic attributes commonly used for channel identification as the inputs to the CNN model. Next, the probability map of the channels was output from the CNN model to obtain the preliminary results of the channel recognition. Finally, we judged whether additional seed points needed to be added according to the preliminary results, and we combined the conditional random field to fuse the geodesic distance map of the additional points with the probability map of the CNN model, ultimately obtaining accurate channel results. Compared to the automatic CNN method, this method extracted more complete channels and improved the continuity of the channel boundaries. In the case of complex seismic data, this method can effectively interpret channels and has important practical significance.