Synthetic Seismic Images Research Articles

Enhancing the seismic response of faults by using a deep learning‐based method

AbstractThe accuracy of fault interpretation is generally influenced by the quality of seismic images. Because of the blurring effect of the migration process, faults with small throws may not be clearly imaged in seismic images, which will impose limitations on the fault detection. To address this issue, we propose a deep learning‐based method to enhance faults in poststack seismic images. We generate abundant training samples by convolving the three‐dimensional point‐spread functions with the noisy reflectivity models. The corresponding labels are synthesized using the one‐dimensional seismic wavelet convolution method, simulating conditions with perfect illumination. To train the network for optimal performance, we investigate the impact of different loss functions. Ultimately, we employ a mixed loss function combining structural similarity index measure and gradient difference loss, since the gradient difference loss focuses more on geological edge information, and the structural similarity index measure possesses excellent image perceptual capability and optimization property. Results from one synthetic seismic image and three real seismic data demonstrate that our proposed method can effectively restore the sharpness of fault surfaces, particularly for faults with small displacements. Compared to the structural smoothing method, the network we trained achieves optimal fault enhancement. Furthermore, coherence‐based fault images indicate that seismic images enhanced using our method can improve the accuracy of fault interpretation and yield more continuous fault maps.

Iterative multitask learning and inference from seismic images

SUMMARY Seismic interpretation aims to extract quantitative and interpretable attributes from a seismic image produced using some migration method to inform characteristics of a subsurface reservoir or target of interest. Current paradigms for computing seismic attributes mostly rely on single-task algorithms. We develop an iterative, multitask machine learning method to learn and infer multiple attributes from a seismic image. This method is composed of two stages: a multitask inference stage and a multimodal, multitask refinement stage. The basic mechanism of this method is that we train a multitask inference neural network to estimate a set of attributes, including a relative geological time volume, a denoised higher-resolution seismic image and multiple fault attributes (including probability, dip and strike), from a low-resolution, noisy seismic image; then we input the inferred attributes to a multitask refinement NN to enhance the raw inference results iteratively. The two multitask neural networks are trained separately based on synthetic seismic images and associated labels generated by geological modelling. Applications of this multitask learning and inference method to synthetic and field seismic images show that our method can improve the structural consistency among output seismic attributes compared with single-task neural networks, leading to more reliable automatic interpretation and subsurface characterization.

Fault detection using a convolutional neural network trained with point-spread function-convolution-based samples

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.

D2UNet: Dual Decoder U-Net for Seismic Image Super-Resolution Reconstruction

Super-resolution reconstruction is an essential task of seismic inversion due to the low resolution and strong noise of field data. Popular deep networks derived from U-Net lack the ability to recover detailed edge features and weak signals. In this paper, we propose a dual decoder U-Net (D <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> UNet) to explore both detail and edge information of the data. The encoder inputs the low resolution image and the edge image obtained through the Canny algorithm. Edge image can provide rich shape and boundary information, which is helpful to generate more accurate and high-quality data. The dual decoder consists of a main decoder for high-resolution recovery and an edge decoder for edge contour detection. These two decoders interact with a texture warping module (TWM) with deformable convolution. TWM aims to distort realistic edge details to match the fidelity of low resolution inputs, especially the location of edges and weak signals. The loss function is a combination of <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">L</i> <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sub> loss and multi-scale structural similarity loss (MS-SSIM) to ensure perception quality. Results on synthetic and field seismic images show that D <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> UNet not only improves the resolution of noisy seismic images, but also maintains the image fidelity.

Imaging of small-scale faults in seismic reflection data: Insights from seismic modelling of faults in outcrop

Faults with throws that fall below vertical seismic resolution are challenging to identify in reflection seismic datasets. Nevertheless, such small-scale faults may still affect the seismic images, and in this study, we build seismic models of outcrop analogues to investigate how. Using photogrammetry from faults affecting Oligocene to Miocene carbonate rocks in Malta, we build a series of geological models from which synthetic seismic images are produced. The resulting seismic images are analysed to elucidate the effects of varying geologic input, signal properties and introduction of noise, and compared to real seismic data from the SW Barents Sea, offshore Norway. Our results suggest that at signal peak frequencies of 30 Hz and higher, using the classic Ricker wavelet type and without introducing noise, graben forming faults with a combined displacement down to ∼5 m affect the seismic image by slight downwarping of reflections, whereas single faults with displacement down to ∼10 m show detectable non-discrete reflection offsets in form of a monoclinal geometry at signal peak frequencies at 60 Hz. Using an Ormsby wavelet, we get seismic images with a quality that lie in between that of the 30 Hz and 60 Hz Ricker, even though the peak frequency is lower. The identified structures can also be seen when noise is included, although the reflections are more irregular and harder to detect. This suggests that under relatively noise-free conditions in high-quality reflection seismic datasets, lower-throw faults (as low as 5 m in this study) that do not induce discrete reflection offsets in seismic images may still produce reflection distortions. Additionally, seismic modelling using the Ormsby wavelet, and its effect on the seismic image, is lacking in literature as of today. We suggest that the results and examples shown in this study may be used to geologically inform fault interpretations in real seismic datasets and may form an empirical basis for geologically concept-driven fault interpretation strategies.

Seismic volumetric dip estimation via a supervised deep learning model by integrating realistic synthetic data sets

Accurately estimating seismic volumetric dip is a crucial task for subsequent seismic processing and interpretation. The traditional window based dip estimation methods, such as gradient structure tensor (GST) and semblance based multiple window scanning strategy, usually struggle with complicated geological structures containing multiple reflections with different dips in the analysis window. Recently, deep learning has been utilized to estimate seismic volumetric dips. However, dip labels used for current convolutional neural network (CNN) based methods are usually created using traditional dip estimation methods, which are not the ground truth, or from specific geological structure models, which only cover a part of typical geological structures. We propose a supervised deep learning model to improve the accuracy of seismic dip estimation by integrating realistic synthetic data sets. Firstly, we propose a synthetic seismic data generation workflow for seismic volumetric dip estimation, which aims to simulate geological features from real seismic data. We then create numerous unique synthetic seismic images with realistic and diverse structural features by using the proposed workflow. To ensure that generated dip labels are exactly accurate and can be regarded as the ground truth, we create synthetic seismic images by deforming horizontal reflection images according to corresponding dip labels. We finally train an end-to-end supervised deep learning model, which focuses on the parallel processing and the extraction of various feature maps simultaneously, utilizing the created realistic synthetic data. The applications on synthetic and 3D real seismic data (Netherlands F3 block) effectually demonstrate the validity and effectiveness of our proposed model.

Exploring seismic detection and resolution thresholds of fault zones and gas seeps in the shallow subsurface using seismic modelling

Seismic resolution and illumination issues are sources of challenges in the detailed imaging and detection of subsurface fault architecture and fluid migration. Improved constraints on resolution can provide input into monitoring requirements and detectability of CO2 leakage, and fault-sealing properties during subsurface visualization of migration pathways. This study explores detection and resolution thresholds via synthetic seismic imaging of a detailed shallow normal fault model with realistic fault architecture including sub-seismic structures of damage zones. The base fault model is built from interpretations of high-resolution P-Cable seismic data and is further developed and conditioned by outcrop-based observations and empirical laws for fracture and deformation band distribution. Damage zones of sandstone layers host deformation bands contrary to shale layers with fractures, while mixed lithologies (shaley sandstone and sandy shale) are subjected to a combination of the two deformation mechanisms. We utilize a 2D point-spread function based convolution seismic modelling to produce the synthetic seismic images. Test scenarios include one baseline fault model without a damage zone (M1), and five more advanced/detailed fault models incorporating features known from outcrops (M2-M6; damage zones, an isolated fracture corridor, and gas seeps (CO2) along damage zones of faults and in the fracture corridor). Furthermore, sensitivity analyses on two selected models test the effect of changing the illumination angle and wavelet. The results show that: (1) faults with damage zones have larger disturbances in seismic signals than faults without damage zones, (2) stronger amplitudes are distinguished for models with CO2-filled fractures, (3) a fracture corridor (5 m at it widest) is clearly visible where it crosses horizons bounding horizontal layers, (4) sensitivity tests show good imaging for illumination ≥45°, which is the average dip of the main fault segment, and (5) learnings from fault modelling offer guidance for seismic monitoring.

Deep Learning for Simultaneous Seismic Image Super-Resolution and Denoising

Seismic interpretation is often limited by low resolution and strong noise data. To deal with this issue, we propose to leverage deep convolutional neural network (CNN) to achieve seismic image super-resolution and denoising simultaneously. To train the CNN, we simulate a lot of synthetic seismic images with different resolutions and noise levels to serve as training data sets. To improve the perception quality, we use a loss function that combines the <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\ell _{1}$ </tex-math></inline-formula> loss and multiscale structural similarity loss. Extensive experimental results on both synthetic and field seismic images demonstrate that the proposed workflow can significantly improve the perception of quality of original data. Compared to conventional methods, the network obtains better performance in enhancing detailed structural and stratigraphic features, such as thin layers and small-scale faults. From the seismic images super-sampled by our CNN method, a fault detection method can compute more accurate fault maps than from the original seismic images.

Recognition and characterization of small-scale sand injectites in seismic data: implications for reservoir development

Thin (&lt;20 m), sand injectites, too small to resolve in conventional seismic reflection data, can constitute a large part of the net-to-gross volume and affect fluid flow in the reservoir. However, they may also cause challenges for well placement and reservoir development because they are too small to be reliably imaged in seismic reflection data. It is therefore important to better understand how small-scale injectites influence seismic images, and how they may be recognized and characterized above reservoirs. The Grane Field (North Sea) hosts numerous small-scale sand injectites above the main reservoir unit, challengin well placement, volume estimates and seismic interpretation. Here, we investigate how such small-scale sand injectites influence seismic images and may be characterized by (1) using well, 3D seismic and outcrop data to investigate geometries of small-scale sand injectites (0–15 m thick), (2) performing seismic convolution modelling to investigate how these would be imaged in seismic data, and (3) comparing these synthetic seismic images with actual 3D seismic data from the well-investigated Grane Field. Our results show that despite injectites being below seismic resolution, small-scale sand injectites can be detected in seismic data. They are more likely to be detected if they have high thickness (&gt;5 m), steep dip (&gt;30°), and when densely spaced. Furthermore, as the fraction of sand injectites increases the top reservoir amplitude will decrease. Moreover, comparison of the synthetic seismic images with real seismic data from the Grane Field indicates that the low-amplitude anomalies and irregularities observed above the reservoir may be a result of the overlying sand injectites. Additionally, the comparison strongly suggests that the Grane Field hosts sand injectites that are thicker and located further away from the top reservoir than is indicated by well observations. These results may be used to improve well planning and develop reservoirs with overlying sand injectites. Supplementary material: A PDF file containing all the seismic modelling results allowing the reader to flip back and forth between the models is available at https://www.doi.org/10.6084/m9.figshare.14333102 and well logs from well 25/11-18T2 are available at https://factpages.npd.no/pbl/wellbore_documents/2358_25_1_18_COMPLETION_REPORT_AND_LOG.pdf

ChannelSeg3D: Channel simulation and deep learning for channel interpretation in 3D seismic images

Seismic channel interpretation involves detecting channel structures, which often appear as meandering shapes in 3D seismic images. Many conventional methods are proposed for delineating channel structures using different seismic attributes. However, these methods are often sensitive to seismic discontinuities (e.g., noise and faults) that are not related to channels. We have adopted a convolutional neural network (CNN) method to improve automatic channel interpretation. The key problem in applying the CNN method into channel interpretation is the absence of labeled field seismic images for training the CNNs. To solve this problem, we adopt a workflow to automatically generate numerous synthetic training data sets with realistic channel structures. In this workflow, we first randomly simulate various meandering channel models based on geologic numerical simulation. We further simulate structural deformation in the form of stratigraphic folding referred to as “folding structures” and combine them with the previously generated channel models to create reflectivity models and the corresponding channel labels. Convolved with a wavelet, the reflectivity models can be transformed into learnable synthetic seismic volumes. By training the designed CNN with synthetic seismic data, we obtain a CNN that learns the characterization of channel structures. Although trained on only synthetic seismic volumes, this CNN shows outstanding performance on field seismic volumes. This indicates that the synthetic seismic images created in this workflow are realistic enough to train the CNN for channel interpretation in field seismic images.

Fault Detection on Seismic Structural Images Using a Nested Residual U-Net

Automatic identification of faults on seismic structural images is a challenging yet crucial task in quantitative seismic interpretation. Human picking or attribute-based fault detection methods may misidentify faults on noisy, complex seismic images. We develop a new automatic fault detection method using a nested residual U-shaped convolutional neural network. Each of the encoders and decoders in this neural network is a residual U-Net, leading to a nested architecture. The final fault map results from the fusion of three fault maps with low, medium, and high fault resolutions. We demonstrate the excellent fault-detection capability of our nested neural network using a series of synthetic and field seismic images. We find that our approach produces clearer and more interpretable fault maps than the current state-of-the-art U-Net fault detection method, particularly on noisy seismic images. Our new automatic fault detection method can facilitate reliable quantitative seismic interpretation on field seismic images.

Time-lapse image registration by high-resolution time-shift scan

Seismic image registration is crucial for the joint interpretation of multivintage seismic images in time-lapse reservoir monitoring. Time-shift analysis is a commonly used method to estimate the warping function by creating a time-shift map, where the energy of each time-shift point in the 3D map indicates the probability of a correct registration. We have adopted a new method to obtain a high-resolution time-shift analysis spectrum, which can help with manual and automatic picking. The time-shift scan map is obtained by trying different local shifts and calculating the local similarity attributes between the shifted and reference images. We adopt a high-resolution calculation of the time-shift scan map by applying the nonstationary model constraint in solving the local similarity attributes. The nonstationary model constraint ensures the time-shift scan map to be smooth in all physical dimensions, for example, time, local shift, and space. In addition, it permits variable smoothing strength across the whole volume, which enables the high resolution of the calculated time-shift scan map. We use an automatic-picking algorithm to demonstrate the accuracy of the high-resolution time-shift scan map and its positive influence on the time-lapse image registration. Synthetic (2D) and real (3D) time-lapse seismic images are used for demonstrating the better registration performance of the proposed method.

Syn-rift carbonate platforms in space and time: testing and refining conceptual models using stratigraphic and seismic numerical forward modelling

Abstract Understanding and predicting architecture and facies distribution of syn-rift carbonates is challenging owing to complex control by climatic, tectonic, biological and sedimentological factors. CarboCAT is a three-dimensional stratigraphic forward model of carbonate and mixed carbonate–siliciclastic systems that has recently been developed to include processes controlling carbonate platform development in extensional settings. CarboCAT has been used here to perform numerical experiment investigations of the various processes and factors hypothesized to control syn-rift carbonates sedimentation. Models representing three tectonic scenarios have been calculated and investigated, to characterize facies distribution and architecture of carbonate platforms developed on half-grabens, horsts and transfer zones. For each forward stratigraphic model, forward seismic models have also been calculated, so that modelled stratal geometries presented as synthetic seismic images can be directly compared with seismic images of subsurface carbonate strata. The CarboCAT models and synthetic seismic images corroborate many elements of the existing syn-rift and early-post-rift conceptual model, but also expand these models by describing how platform architecture and spatial facies distributions vary along-strike between hanging-wall, footwall and transfer zone settings. Synthetic seismic images show how platform margins may appear in seismic data, showing significant differences in overall seismic character between prograding and backstepping stacking patterns.

Improving fault surface construction with inversion-based methods

Constructing fault surfaces is a key step for seismic structural interpretation and building structural models. We automatically construct fault surfaces from oriented fault samples scanned from a 3D seismic image. The main challenges of the fault surface construction include the following: Some fault samples are locally missing, the positions and orientations of the fault samples may be noisy, and surfaces may form complicated intersections with each other. We adopt the Poisson equation surface method (PESM) and the point-set surface method (PSSM) to automatically construct complete fault surfaces from fault samples and their corresponding orientations. Our methods can robustly fit the noisy fault samples and reasonably fill holes or missing samples, thus improving fault surface construction. By formulating fault surface construction as an inverse problem, we estimate a scalar function to approximate the fault samples in the least-squares sense. In PESM, we estimate the scalar function by solving a weighted Poisson equation. In PSSM, the scalar function is derived by fitting local algebraic spheres based on moving least-squares approximations. Then, the fault surfaces can be approximated by zero isosurfaces of the resulting scalar function. To handle complicated cases of crossing faults, we first classify the fault samples according to their orientations, and we take each class of samples as input of our inversion-based approaches to independently construct the crossing faults. We determine the ability of our methods in robustly building the complete fault surface using synthetic and real seismic images complicated by noise and complexly intersecting faults.

Distilling Knowledge From an Ensemble of Convolutional Neural Networks for Seismic Fault Detection

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.

Karhunen Loeve Transform with adaptive dictionary learning for coherent and random noise attenuation in seismic data

Seismic reflection data are used by geologists to identify the best site for oil and gas explorations. The raw seismic image cannot be directly used for explorations because the noises may hinder the primary reflections and result in misinterpretations. In the proposed research work, the denoising method with adaptive dictionary learning is attempted to attenuate the coherent and random noises which affect the seismic reflections. In this method, Karhunen Loeve Transform (KLT) is combined with K-means Singular Value Decomposition (KSVD) to improve the Signal to Noise Ratio (SNR) of the seismic data. The combination of fixed transform using KLT and learning-based dictionary using KSVD remove the redundant data while retaining the necessary data for further interpretations. KLT is applied on the whole seismic image to decorrelate the coefficients and retain the primary reflections. The horizontal events in the seismic image are preserved as they represent the eigen images with large energy. The KSVD is applied to the KLT resultant data, which denoises the patch of data while simultaneously updating the dictionary. The noise is eliminated using local sparsity of the image where mutually overlapping small image blocks are learned to yield the self-adaptive redundant dictionary. It is then used to obtain the sparse representation of the image blocks by eliminating noise. The algorithm is tested for both synthetic and field seismic images, and the results indicate a better reduction of random and coherent noise with the acceptable execution time compared to other denoising algorithms used. The KLT combined with KSVD method decorrelate the seismic data from noise by preserving the primary reflections and discarding the redundant noise.

Deep Learning for Characterizing Paleokarst Collapse Features in 3‐D Seismic Images

AbstractPaleokarst systems are found extensively in carbonate‐prone basins worldwide. They can form large reservoirs and provide efficient pathways for hydrocarbon migration, but they can also create serious engineering geohazards. The full delineation of potentially buried paleokarst systems plays an important role for reservoir characterization, oil and gas production, and other engineering tasks. We propose a supervised convolutional neural network (CNN) to automatically and accurately characterize paleokarst and associated collapse features from 3‐D seismic images. To avoid time‐consuming manual labeling for training the CNN, we propose an efficient workflow to automatically generate numerous 3‐D training image pairs including synthetic seismic images and the corresponding label images of the collapsed paleokarst features simulated in the seismic images. With this workflow, we are able to simulate realistic and diverse geologic structures and collapsed paleokarst features in the training images from which the CNN can effectively learn to recognize the collapsed paleokarst features in real field seismic images. Two field examples from the Fort Worth Basin demonstrate that our CNN‐based method is superior to conventional automatic methods in delineating paleokarst collapse features from seismic images. From the CNN‐based paleokarst characterization, we can further automatically extract 3‐D collapsed paleokarst systems and quantitatively measure their geometric parameters. Our CNN‐based method is highly efficient and takes only seconds to classify collapsed paleokarst features a 3‐D seismic image with 320 × 1, 024 × 1, 024 samples (approximately 268 km2) by using one graphics processing unit.

Deep learning for relative geologic time and seismic horizons

Constructing a relative geologic time (RGT) image from a seismic image is crucial for seismic structural and stratigraphic interpretation. In conventional methods, automatic RGT estimation from a seismic image is typically based on only local image features, which makes it challenging to cope with discontinuous structures (e.g., faults and unconformities). We have considered the estimation of 2D RGT images as a regression problem, where we design a deep convolutional neural network (CNN) to directly and automatically compute an RGT image from a 2D seismic image. This CNN consists of three parts: an encoder, a decoder, and a refinement module. We train this CNN by using 2080 pairs of synthetic input seismic images and target RGT images, and then we test it on 960 testing seismic images. Although trained with only synthetic images, the network can generate accurate results on real seismic images. Multiple field examples show that our CNN-based method is significantly superior to conventional methods, especially in dealing with complex structures such as crossing faults and complicatedly folded horizons, without the need of any manual picking.

Building realistic structure models to train convolutional neural networks for seismic structural interpretation

Seismic structural interpretation involves highlighting and extracting faults and horizons that are apparent as geometric features in a seismic image. Although seismic image processing methods have been proposed to automate fault and horizon interpretation, each of which today still requires significant human effort. We improve automatic structural interpretation in seismic images by using convolutional neural networks (CNNs) that recently have shown excellent performances in detecting and extracting useful image features and objects. The main limitation of applying CNNs in seismic interpretation is the preparation of many training data sets and especially the corresponding geologic labels. Manually labeling geologic features in a seismic image is highly time-consuming and subjective, which often results in incompletely or inaccurately labeled training images. To solve this problem, we have developed a workflow to automatically build diverse structure models with realistic folding and faulting features. In this workflow, with some assumptions about typical folding and faulting patterns, we simulate structural features in a 3D model by using a set of parameters. By randomly choosing the parameters from some predefined ranges, we are able to automatically generate numerous structure models with realistic and diverse structural features. Based on these structure models with known structural information, we further automatically create numerous synthetic seismic images and the corresponding ground truth of structural labels to train CNNs for structural interpretation in field seismic images. Accurate results of structural interpretation in multiple field seismic images indicate that our workflow simulates realistic and generalized structure models from which the CNNs effectively learn to recognize real structures in field images.

Generating Sketch-Based Synthetic Seismic Images With Generative Adversarial Networks

The characterization of the subsurface is paramount for the exploration and production life cycle and, more specifically, for the identification of potential hydrocarbon accumulations. In recent years, the oil and gas industry has increased its interest in applying machine learning to accelerate the seismic interpretation process, which is regarded as a time-consuming and human-centered task. Although machine learning has been successfully used in many applications ranging from stratigraphic segmentation to salt dome detection, a usual bottleneck is the need for a large amount of high-quality annotated data. To overcome this, data augmentation approaches are commonly used, and one of the most powerful is synthetic data generation. In addition, sketch-based synthetic seismic images can be used to support image retrieval applications to help oil companies leverage petabytes of seismic data sets. In this context, this letter investigates the generation of synthetic seismic images based on sketches using generative adversarial networks (GANs). To the best of our knowledge, this is the first work to propose such an approach. Experiments with five different sketch types in a public seismic data set indicate that realistic seismic images can be synthesized using rather simple sketches.