Seismic inversion is a process to obtain the spatial structure and physical properties of underground rock formations by using surface acquired seismic data, constrained by known geological laws, drilling and logging data. The principle of seismic inversion based on deep learning is to learn the mapping between seismic data and rock properties by training a neural network using logging data as labels. However, due to high cost, the number of logging curves are often limited, leading to a trained model with poor generalization. Multi-task learning (MTL) provides an effective way to mitigate this problem. Learning multiple related tasks at the same time can improve the generalization ability of the model, thereby improving the performance of the main task on the same amount of labeled data. However, the performance of multi-task learning is highly dependent on the relative weights for the loss of each task, and manual tuning of the weights is often time-consuming and laborious. In this paper, a method based on homoscedastic uncertainty of the Bayesian model is used to balance the weights of the loss function for multiple tasks, and a Fully convolutional residual network (FCRN) is used to achieve seismic impedance inversion and seismic data reconstruction simultaneously. The test results on the synthetic dataset of Marmousi2 model show that the proposed method can automatically determine the approximate optimal weight of the two tasks, and predicts impedance with higher accuracy than single-task FCRN model.

Due to the constraints of natural environments, acquired prestack seismic data is usually not complete, which seriously affects subsequent seismic data processing. With the progress of deep learning, many neural networks with different structures have been applied to missing seismic data interpolation. Among them, U-net can efficiently interpolate the regularly and irregularly missing seismic traces with small gap. While for consecutively missing seismic traces with big gap, the interpolation results for low amplitude missing components need to be further improved. In this letter, we analyze the variation of interpolation results for consecutively missing seismic traces during the traditional U-net training process, and find that U-net tends to only interpolate the high amplitude missing components. Meanwhile, due to the distribution difference between low and high amplitude seismic data, one U-net model is insufficient to interpolate both high and low amplitude missing components with a wide amplitude range. To improve the interpolation results of single U-net, we propose a multistage training process to train multiple U-net models. Each U-net model focuses on interpolating different missing components with a small amplitude range. In this way, more accurate interpolation results for low amplitude missing components can be obtained. Comparison experiments conducted on synthetic and field seismic data show that, under the same number of training epochs, the proposed training process can produce more accurate interpolation results comparing with traditional single U-net.

Missing traces interpolation is a basic step in the seismic data processing workflow. Recently, many seismic data interpolation methods based on different neural networks have been proposed. The existing research shows that when the seismic data are consecutively missing, the larger gap for missing traces, the more difficult task of interpolation, due to convolution operation in the neural network can only capture local relations. In this letter, we incorporate the coordinate attention block to the Unet for 2-D successive missing traces interpolation. The hybrid loss function combined with structural similarity (SSIM) and$\text {L}_{ {1}}$norm is used as the loss function to further improve the interpolation performance of the designed network. Comparison experiments on 2-D synthetic and field seismic data show that the interpolation results obtained by the proposed method are more accurate and reasonable compared with Unet and Unets equipped state-of-the-art similar modules.

Denoising is an indispensable step in seismic data processing. Deep-learning-based seismic data denoising has been recently attracting attentions due to its outstanding performance. In this letter, we investigate the architecture of deep Convolutional Networks (ConvNets) for seismic data denoising. The untrained ConvNets are served as a generative network to a single seismic data profile with Gaussian noise. Starting with random initialized parameters, the generative networks with various handcrafted architectures have different ability to map the seismic data at iterations and can separate the Gaussian noise as residuals. For the purpose of exploring the ability of Gaussian noise separation, the depth, width, and skip connection as the main components of generative network are assembled as various architectures to fit Gaussian noise, clean, and noisy seismic data, respectively. Then, the favorable network architecture with high and low impedance (an ability to hinder data reconstruction) to noise and seismic data is adopted as prior model to seismic data denoising task. Furthermore, a stopping criterion is designed for the data fitting process to obtain the latent clean seismic data automatically. The proposed method does not need data sets for training and it makes use of network architecture as prior. Extensive experiments both on synthetic and field data demonstrate the effectiveness of the selected ConvNet and the advantages are evaluated by comparing the denoising results with f-x multi-channel singular spectrum analysis (MSSA) and state-of-the-art unsupervised neural network (NN)-based method.

Fault detection is a crucial task in seismic structure interpretation. Convolutional neural network (CNN)-based methods, in general, require large amount of labeled data for network training. One way to build the labeled data is to create synthetic seismic images with corresponding fault labels. However, it is hard to ensure that the synthetic data have the same fault feature distributions as the field data, which may lead to inaccurate and unreliable prediction results. Another way is to manually label the faults, which is time-consuming and subjective. In this letter, we propose that using knowledge distillation (KD) to improve the performance of fault detection by integrating the features from large number of synthetic samples and a small number of field samples. We distill knowledge from an ensemble of two teacher CNNs to train a student CNN (applied to final target) for seismic fault detection. In our work, one segmentation teacher CNN is trained on synthetic samples with known ground truth fault labels and another classification teacher CNN is trained on field samples with manually picked labels. Then, a classification student network is trained on samples generated by voting the results from two teacher models. The student CNN learns not only the general fault characteristics in the synthetic data but also the specific fault features of the target field data. Test on the field data shows that the student CNN highlights seismic fault more accurately with higher resolution than the teacher CNNs.

Semblance velocity analysis plays an indispensable role in seismic data processing. In order to avoid the huge time-cost when performed manually, some deep learning methods are proposed for automatic velocity picking from semblance. However, the application of existing deep learning methods is still restricted by the shortage of labels in practice. To solve this problem, we take semblance velocity analysis as a point-to-point regression problem at each time sample. A time window on semblance which can extract the block corresponding to a time-velocity (t-v) pair and the reference velocity curve (RVC) which can transform semblance randomly are employed together to augment the labeled data. We divide the development of data augmentation strategy into three progressive modes. The datasets from three modes are prepared for training designed attention neural network. The field experiments show that the attention neural network can produce reasonable results and the data augmentation strategy can effectively improve the velocity picking accuracy.