Automatic fault detection in seismic images using deep learning-based methods has attracted great interest in recent years. Detecting faults by deep learning methods is generally considered a supervised learning task, which requires numerous diverse training samples and corresponding accurate labels. Because field data with faults labeled by experienced interpreters are difficult to acquire, training samples are usually generated synthetically, in which the 1D seismic wavelets are convolved with reflectivity models. However, the 1D-convolution-based synthetic seismic images still differ from the real seismic images. Therefore, the network trained by such samples may fail to detect some faults in the field data. Full-wavefield approaches are the optimal seismic modeling methods, but they cannot be widely used at present due to the high cost. In this paper, we present an efficient approach to generate realistic synthetic seismic images by using a point-spread function (PSF)-based convolution. Because the size of a seismic sample used for deep learning training is a small range relative to the field seismic data, assuming the background velocity is close to homogeneous in such a small range, we can efficiently construct the PSF using the analytical Green’s function. With an Intel Core i9-10900K CPU, the PSF approach takes approximately 15 min to generate a seismic image of size 128 × 128 × 128, whereas it takes approximately 20 h to generate a seismic image of the same size using reverse time migration. The examples of one synthetic image and two field seismic images demonstrate that the network trained with the PSF convolution samples can predict more accurate and continuous faults than that trained with the 1D convolution samples.

Deep-learning-based seismic data interpretation has received extensive attention and focus in recent years. Research has shown that training data play a key role in the process of intelligent seismic interpretation. At present, the main methods used to obtain training data are synthesizing seismic data and manually labeling the real data. However, synthetic data have certain feature differences from real data, and the manual labeling of data is time-consuming and subjective. These factors limit the application of deep learning algorithms in seismic data interpretation. To obtain realistic seismic training data, we propose label-to-data networks based on cycle-consistent adversarial networks in this work. These networks take random labels and unlabeled real seismic data as input and generate synthetic seismic data that match the random labels and have similar features to the real seismic data. Quantitative analysis of the generated data demonstrate the effectiveness of the proposed methods. Meanwhile, test results on different data indicate that the generated data are reliable and can be applied for seismic fault detection.

In the marine seismic acquisition, hydrophones are distributed along streamers that are towed under the sea surface, recording both the up-going primary reflections and the down-going free-surface reflections. The interferences between them result in notches in the frequency-domain amplitude spectra, causing missing frequencies and destroying the broadband signature. These free-surface reflections, referred to as ghost waves, are usually treated as noises and are supposed to be suppressed or removed before depth migration imaging, namely the deghosting processing. If the ghost waves are not effectively removed, they will be projected to the subsurface image domain, further causing missing wavenumbers and influencing resolutions. The missing information can hardly be fully compensated through traditional deconvolution-type filters in the data or image domain. Recently, deep-learning-based methods have been introduced to data-domain deghosting. However, generating a sufficiently large number of training samples with and without ghost waves is quite costly, and it is difficult to guarantee the diversity, leading to the limited generalization capability of neural networks trained with these data sets. Therefore, we are motivated to use the point spread function (PSF) to project the near-surface ghost-related parameters to the subsurface local image domain. In this way, large numbers of depth images with and without ghost waves can be constructed efficiently through convolutions between PSFs and local model perturbations. The diversity of training samples can be easily satisfied. Using these paired samples, we train U-nets for deghosting processing in the image domain. Various examples validate the new deep-learning-based method.