Normal Moveout Correction Research Articles

Velocity-independent normal moveout correction based on multiscale dynamic time warping

Normal moveout (NMO) correction is one of the common steps of seismic data processing. Conventional NMO correction is generally based on a velocity function, but velocity analysis is considered time consuming and labor intensive. Moreover, the field seismic data observed in complex areas can have a low signal-to-noise ratio, resulting in inaccurate picked velocity parameters and poor NMO correction. Conventional NMO correction has stretching issues, most significantly in shallow and large-offset areas. To address these problems, a new multiscale dynamic time-warping (MSD) algorithm for velocity-independent stretch-free NMO correction is investigated. This algorithm has higher accuracy and stronger noise resistance, as our comparisons with other applications of dynamic time warping show. In our algorithm, the original seismic data are decomposed into multiple morphological scales (using mathematical morphology theory), and alignment errors are calculated separately at each scale, followed by further adaptive weighted summation. Accumulation and backtracking operations based on alignment errors produce time-shift sequences for automatic flattening of the gathers. The effectiveness of the presented approach is tested on synthetic and field data. The results determine that compared with traditional velocity-based NMO correction methods, our method can accurately flatten seismic gathers without the NMO stretch and is robust to strong noise.

Unsupervised Ground Roll Attenuation via Implicit Neural Representations

Coherent noise attenuation presents a significant challenge compared to incoherent noise, particularly concerning ground roll in land seismic data. Traditional methods encounter limitations with the substantial overlap of ground roll and reflections, resulting in a compromise between signal distortion and residual noise. Recent deep learning strategies, while promising, predominantly rely on supervised learning, necessitating extensive paired training#xD;data, which is frequently scarce in real-world scenarios. Unsupervised approaches often struggle with convergence issues or excessive parameter tuning. We introduce an unsupervised deep learning framework for separating reflections from ground roll to address these hurdles. Leveraging the inherent low-frequency bias of implicit neural representations (INR), our method prioritizes extracting self-similarity features during training. For seismic data, the network learns self-similarity flattened events before those with deep dips and other incoherent noise. To enhance the self-similarity of reflections, we first apply normal moveout (NMO) correction to flatten reflections and then use the network to extract these flattened reflections from the NMO-corrected data. Moreover, to ensure convergence, we integrate a horizontal derivative regularization term into the loss function, penalizing horizontal variations. This regularization prevents the extraction of unflattened events through extended training, ensuring convergence with high fidelity of flattened reflections. It additionally streamlines parameter tuning, yielding stable outputs and obviating the need for early stopping. We validate our method using synthetic and real land data examples, comparing it with traditional f-k filters with real data and demonstrating its superiority in noise attenuation and signal preservation.

Nonstationary adaptive S-wave leakage suppression of ocean-bottom node data

Ocean-bottom nodes (OBNs) are widely used because of their wide azimuth, long-offset, and low-frequency advantages. However, in the vertical component of the OBN geophone, a significant amount of S-wave induced noise may be recorded. This significantly impacts the quality of the vertical component data and may affect the follow-up merging of dual-sensor data. Some commercially available methods apply normal-moveout correction, which requires velocity data. We avoid this with an adaptive matching method, which relies solely on seismic data. Therefore, we develop a novel adaptive subtraction method for S-wave leakage suppression using the horizontal component as the noise model. Our method effectively handles the nonstationarity of the input seismic data in all time and space directions, mitigating instability caused by manual selection of the smoothing radius. A method is introduced for estimating a global nonstationary smoothing radius using the noise model. Compared with commercial and stationary smoothing methods, our method can better suppress S-wave noise while balancing residual noise and signal leakage more effectively. Synthetic and field data examples demonstrate significant improvement with our method.

Nonstretching normal moveout correction via an extrapolated interferometry method

Nonstretching normal moveout correction via an extrapolated interferometry method

Optimizing a transformer-based network for a deep-learning seismic processing workflow

StorSeismic is a recently introduced model based on the transformer network to adapt to various seismic processing tasks through its pretraining and fine-tuning strategy. In the original implementation, StorSeismic uses a sinusoidal positional encoding (PE) and a conventional self-attention mechanism, borrowed from natural language processing applications. For seismic processing, they provide good results but also indicate limitations in efficiency and expressiveness. We develop modifications to these two key components, by using relative PE and low-rank attention matrices as replacements for the standard ones. Our changes are tested on processing tasks applied to realistic Marmousi and offshore field data as a sequential strategy, starting from denoising, direct-arrival removal, multiple attenuation, and finally root-mean-squared velocity ([Formula: see text]) prediction for normal moveout correction. We observe faster pretraining and competitive results on the fine-tuning tasks and, in addition, fewer parameters to train compared with the standard model.

Near offset reconstruction for marine seismic data using a convolutional neural network

AbstractMarine seismic data is often missing near offset information due to separation between the source and receiver cables. To solve this problem, a convolutional neural network is trained on synthetic seismic data to reconstruct the near offset gap. The synthetic data is created using a two‐dimensional finite difference method within a heterogeneous velocity model. These synthetics are generated with a source‐over‐receiver acquisition geometry so that they contain complete near offset data. The convolutional neural network is then trained on input‐target synthetic pairs where the inputs are common midpoint gathers with the near offset section removed, and the targets are the same gathers with the near offset section retained. Following training, the robustness of the method is investigated with regards to common midpoint data sorting, normal moveout correction and changes in the velocity model. It is found that training on common midpoint‐sorted data results in 2.8 times lower error than training on shot gathers, that normal moveout correction of the training data makes no significant difference in error levels, and that the model can reconstruct realistic near offsets on synthetic data generated 10 km away within the heterogeneous velocity model. In field data testing, first a dataset with source‐over‐cable acquisition geometry from the Barents Sea is used to compare the reconstructed wavefields to ground truth values. Although the reconstructed amplitudes require minor scaling to match the true values, predictions on this dataset yield 2.5 times lower near offset reconstruction error compared to a simple Radon transform interpolation method. Furthermore, amplitude versus offset gradient and intercept sections from the Barents Sea dataset are estimated with half the error when including the convolutional neural network‐predicted near offset data, compared to only using the conventionally‐acquirable portion of the data (beyond 112.5 m of offset). In a secondary field data test, a conventional northern North Sea dataset is used to demonstrate how the method may be applied in practice. Here, the convolutional neural network generates more realistic predictions than the Radon method, and the gradient and intercept sections calculated using the convolutional neural network‐predicted traces have higher signal‐to‐noise ratios than the sections calculated using only the original data. The combination of high‐quality synthetic training data and interpolation in the common midpoint domain enables near offset reconstruction at significant depth (1 s of two‐way traveltime or more), which is demonstrated in both synthetic and field examples.

Oceanic crust—seismic structure, lithology and the cause of the 2A Event at borehole 504B

SUMMARY This study focuses on the 3-D velocity structure and thickness of ∼7-Myr-old oceanic crust surrounding borehole 504B, located ∼235 km from the intermediate-spreading Costa Rica Rift (Panama Basin). It investigates how well seismic structure determined by 3-D tomography compares with actual lithology and, consequently, what the origin and cause might be of an amplitude anomaly, the 2A Event, that is observed in multichannel seismic data. Our P-wave model shows an ∼0.3-km-thick sediment layer of velocity between ∼1.6 and 1.9 km s−1 (gradient 1.0 s−1), bound at its base by a velocity step to 4.8 km s−1 at the top of oceanic crustal Layer 2. Layer 2 itself is subdivided into two main units (2A and 2B) by a vertical velocity gradient change at 4.5 km depth, with a gradient of 1.7 s−1 above (4.8–5.8 km s−1) and 0.7 s−1 below (5.8–6.5 km s−1). The base of Layer 2, in turn, is defined by a change in gradient at 5.6 km depth. Below this, Layer 3 has a velocity range of 6.5–7.5 km s−1 and a gradient of ∼0.3 s−1. Corresponding S-wave igneous layer velocities and gradients are: Layer 2A, 2.4–3.1 km s−1 and 1.0 s−1; Layer 2B, 3.1–3.7 km s−1 and 0.5 s−1; Layer 3, 3.7–4.0 km s−1 and 0.1 s−1. The 3-D tomographic models, coupled with gravity modelling, indicate that the crust is ∼6 km thick throughout the region, with a generally flat-lying Moho. Although the P- and S-wave models are smooth, their velocities and gradients are remarkably consistent with the main lithological layering subdivisions logged within 504B. Thus, using the change in velocity gradient as a proxy, Layer 2 is interpreted as ∼1.8 km thick and Layer 3 as ∼3.8 km thick, with little vertical variation throughout the 3-D volume. However, the strike of lateral gradient variation is not Costa Rica Rift-parallel, but instead follows the orientation of the present-day adjacent Ecuador Rift, suggesting a reorientation of the Costa Rica Rift spreading ridge axis. Having determined its consistency with lithological ground-truth, the resulting P-wave model is used as the basis of finite difference calculation of wave propagation to find the origin of the 2A Event. Our modelling shows that no distinct interface, or transition, is required to generate this event. Instead, it is caused by averaging of heterogeneous physical properties by the seismic wave as it propagates through Layer 2 and is scattered. Thus, we conclude that the 2A Event originates and propagates exclusively in the lower part of Layer 2A, above the mean depth to the top of the dykes of Layer 2B. From our synthetic data we conclude that using the 2A Event on seismic reflection profiles as a proxy to determine the Layer 2A/2B boundary's depth will result in an overestimate of up to several hundred metres, the degree of which being dependent on the specific velocity chosen for normal moveout correction prior to stacking.

Ultrahigh-resolution 9C seismic survey in a landslide prone area in southwest of Sweden

SUMMARY We studied the benefits of a nine-component (9C) seismic survey over a landslide-prone area in southwest of Sweden to retrieve ultrahigh-resolution shear wave reflection images of the subsurface as well as crucial information on physical properties of the sediments. A complete, 1 m shot and receiver spacing, multicomponent 2-D seismic profile was acquired using three-component microelectromechanical-system-based landstreamer receivers, and a 5-kg sledgehammer strike in vertical and horizontal orientations as three-component seismic source. Given the rich number of shear wave reflections observed on all the 9C data, the processing work focused on their retrievals. It revealed three distinct reflections, two of which associated with coarse-grained materials and one with an extremely undulating bedrock surface. Given the extremely slow shear wave velocities on the order of 60–100 m s−1, we obtained ultrahigh-resolution shear wave sections avoiding temporal and spatial aliasing. Imaging results suggest vertical-source and horizontal-radial receiver (V–HR), and horizontal-transverse source–receiver orientations (HT–HT) provided the most optimum images of the subsurface. A non-hyperbolic algorithm was applied to the normal-moveout corrections justified by the traveltime differences of the bedrock reflection in different shear wave sections. The improved images by incorporating the anisotropy term suggest that the data set reveals moderate shear wave anisotropy along some portions of the profile. The Vp/Vs ratio obtained by using bedrock reflection in P- and S-wave sections suggests values ranging 10–16, which implies high water content. Areas with lower Vp/Vs coincides with greater anisotropic parameters and this can indicate disturbed clays or presence of sensitive clays.

Two-dimensional complex wavelet transform for linear noise attenuation and image decomposition

Abstract For developing a high-fidelity, high-resolution seismic denoising method, we use the two-dimensional complex wavelet transform (2D CWT) to analyze noise and signals. By investigating a surface wave's features and evaluating factors affecting the fidelity of the method, the best practice for the wavelet transform-based denoising has been established. First, static and normal moveout correction are applied on shot gathers. Then, 2D CWT is used to attenuate linear noises. The results demonstrate that the proposed method and practice significantly attenuate noises and preserve the signal's amplitudes and frequency band. In addition to denoising, we also apply the 2D CWT to decompose a seismic image into multiscale images with different resolutions. Multiscale decomposed images derive more detailed information for subsurface structures and fault networks. The decomposed images depict sharper structures and reveal detailed features of faults more significantly than the original images.

Low-rank-based residual statics estimation and correction

Surface consistency forms the basis for short-wavelength statics estimation. When raypaths in the near surface diverge from a normal incidence or when the normal moveout (NMO) velocity is inaccurate, surface-consistent methods may fail to estimate accurate statics. Existing nonsurface-consistent techniques can be prone to errors due to the need to construct pilot traces or pick horizons while imposing additional computational costs. To overcome these limitations and correct for the surface- and nonsurface-consistent statics, we have developed a low-rank-based residual statics (LR-ReS) estimation and correction framework. The method makes use of the redundant nature of seismic data by using its low-rank structure in the midpoint-offset-frequency domain. Due to the near-surface effect, the low-rank structure is destroyed. Therefore, we estimate the statics by means of low-rank approximation and crosscorrelation. To alleviate the need for accurate rank selection for low-rank approximation and improved statics estimation, we implement the method in an iterative and multiscale fashion. Because the low-rank approximation deteriorates at high frequencies, we use its better performance at low frequencies and exploit the common statics among the different frequency bands. The LR-ReS estimation and correction can be applied to data without an NMO correction, which makes statics estimation independent of the NMO velocity errors. Consequently, it can reduce the multiple iterations of the NMO velocity estimation and short-wavelength statics correction commonly needed for conventional methods to improve their performance. Moreover, the LR-ReS estimation does not require windowing of a noise-free area containing aligned primaries or mute to avoid the NMO stretch effect, which enables statics correction of the wavefield of all offsets. To evaluate the performance of our method, we apply it to simulated data and a challenging field data set affected by complex weathering layers and noise, which indicate a substantial improvement compared with conventional short-wavelength statics correction.

Estimating S-wave statics from a PS-receiver-stack image

Estimating converted-wave (S-wave) statics is a difficult challenge in reality, but it is crucial for the final success of PS-wave processing and imaging. A novel and practical method is introduced to directly estimate the S-wave statics from a PS-receiver-stack image. Before a PS data set is stacked to a PS-receiver-stack image, available source-side P-wave statics and common-conversion-point normal-moveout (NMO) correction are applied, so we only need to estimate an S-wave static value for each trace in this PS-receiver-stack image. For the corresponding PP data set, the PP-receiver stack is generated by stacking the NMO-corrected gathers with source- and receiver-side P-wave statics applied. We select a target event on the PP-receiver-stack image and convert it to a PS-guidance structure via available P-to-S velocity conversions. Because the target PS event on our PS-receiver-stack image only has source-side P-wave statics applied, the time differences between this PS-guidance structure and the target PS event on the PS-receiver-stack image are actually the outstanding receiver-side S-wave statics. By aligning this PS target event to its corresponding PS-guidance structure, the S-wave statics for this PS data set are retrieved. A challenging field data set is used as an example to demonstrate the robustness of our new S-wave statics estimation method.

Seismic profile denoising based on common-reflection-point gathers using convolution neural networks

Abstract With the development of seismic surveys and the decline of shallow petroleum resources, high resolution and high signal-to-noise ratio have become more important in seismic processing. To improve the quality of seismic data, stationary-phase migration based on dip-angle gathers can be used to separate the reflected waves and noise. However, this method is very computationally intensive and heavily dependent on expert experience. Neural networks currently have powerful adaptive capabilities and great potential to replace artificial processing. Certain applications of convolution neural networks (CNNs) on stack profiles lead to a loss of amplitude information. Therefore, we have developed CNNs for noise reduction based on common-reflection-point (CRP) gathers. We used CRP gathers of stationary-phase migration as labels and CRP gathers of conventional prestack time migration as inputs. In addition, we analyzed the seismic amplitude properties and demonstrated the neural network optimization process and results. The results showed that our methods can achieve fast and reliable denoising and produce high-quality stack profiles that contain true amplitude information. Furthermore, the predicted high-quality CRP gathers can be used for further processing steps, such as normal moveout correction and amplitude variation with offset.

Reduction of normal-moveout stretch using nonstationary scaling transformation in time-frequency domain

The normal-moveout (NMO) stretch causes decrease in the dominant frequency of seismic wavelet after conventional NMO correction and severely damages the quality of the stacked data for shallower reflectors at far offsets. Muting, which is commonly used to handle this problem, reduces seismic fold and negatively affects results of the amplitude-variation-with-offset analysis within the stretched area. We found a novel approach to reduce the stretching phenomenon through compensating the lost frequencies by increasing the dominant frequency of the seismic wavelet before applying the NMO correction. The added so-called compensated frequencies are defined according to the difference between the dominant frequency of the original seismic wavelet and the assumed stretched wavelet after the NMO correction. The corresponding procedure considers frequency content of each time sample along each trace in the time-frequency domain using the Gabor transform. As such, the dominant frequency of the seismic wavelets is increased in a nonstationary manner. Performance of our method is evaluated by applying it on the synthetic and field data examples. The obtained results suggest that this approach provides common-midpoint (CMP) gathers with reduced stretching effect, with the potential to be considered as another alternative for nonstretch NMO correction. However, it should be noted that the presented method resolves neither the problem of intersecting events nor the multiples in the CMP gather. This method also cannot handle highly contaminated noise data and does not contribute in removing multiples during the frequency compensation process.

Seismic velocity field estimation using sonic well logs decomposed by Singular Spectrum Analysis (SSA) and Exponentially Weighted Moving Average (EWMA) - A case of study in Recôncavo Basin, Brazil

The determination of an accurate seismic wavefield velocity propagating in the subsurface is a problem faced currently by geoscientists. However, velocity analysis on land seismic data is often compromised due to lack of a good signal-to-noise ratio, absence of guide reflectors, and low fold, which significantly affects the quality of the final velocity model and, consequently, the seismic image. This paper introduces an improvement in the root mean square (RMS) velocity model obtained from the Polo-Miranga seismic cube, located in the Recôncavo sedimentary basin in northeast of Brazil, using slowness well log data. The available vagarosity logs were smoothed using the Exponentially Weighted Moving Average (EWMA) and Singular Spectrum Analysis (SSA) to determine a new measurement that reflects the low frequency of the p-wave seismic velocity to be employed in the velocity analysis. Estimating the best parameter of the EWMA algorithm required some tests to smooth the slowness. At the same time, the first component of the sonic decomposition into three other curves was a reasonable way to estimate the velocity. Those techniques showed robust results and enhanced some reflectors in the 3D stack. Furthermore, the proposed methodology’s performance is validated through the normal moveout correction (NMO) simulation and on the time-slice image.

Active and Passive-Source Underground Seismic Data Acquisition

As part of the European research project Seismic Imaging Techniques for Mineral Exploration (SIT4ME), in-mine seismic active-source and continuous noise measurements were performed within an underground mine gallery of a former radioactive waste repository – the Asse II salt mine (Lower Saxony, Germany) to investigate its geological conditions. Inspired by recent underground active-seismic surveys in the Cote Blanche salt mine and former In-seam seismic surveys in the German hard-coal district of the Ruhr area, we apply conventional exploration and processing methods to an image of the subsurface. Among others, these include data sorting, bandpass filtering, normal moveout correction, static correction and depth (distance) conversion. To process the passive seismic data, we perform an illumination diagnosis for the retrieval of body-wave arrivals and apply passive-source seismic interferometry by cross-correlation (PSICC) on noise data dominated by S-waves. We show that active-source seismic measurements can be used from underground mine galleries for the identification of geological structures. Furthermore, we demonstrate that PSICC can be used to produce virtual-source underground seismic surveys resembling an active-source seismic survey.

NMO stretch in vertical transverse isotropic media presented by nonhyperbolic approximations

Normal-moveout (NMO) correction is an inevitable part of routine seismic data processing and imaging. NMO correction suffers from an undesirable phenomenon called NMO stretching, which is more pronounced in larger emerging angles. Frequency distortions caused by NMO stretching for isotropic media have fully been studied in the literature. We express the distortions for the simplest case of anisotropy, which is vertical transverse isotropic media. This is done through different nonhyperbolic approximations, including shifted-hyperbola approximation, rational approximation, and generalized moveout approximation (GMA). To evaluate the predicted distortions, we define a trace-muting threshold and use it in the trace-muting procedure for synthetic data, as well as a field recorded example. Comparison between the results proves that the GMA has the highest accuracy among the other presented approximations. The proposed approach has an immediate potential to be used in industry and also to help to develop innovative nonstretch NMO correction workflows.

Reflection residual statics estimated using a high-resolution neural network

Performing reflection statics corrections for improving seismic imaging is an important step in land and shallow marine data processing. The reflection residual statics solution is commonly derived from stack power maximization with hyperbolic assumptions that are valid for normal moveout (NMO) correction and stacking. We have developed a deep learning method for deriving surface-consistent reflection residual statics directly from common-shot, common-receiver, or common-midpoint gathers without velocity analysis and NMO correction. To mitigate overfitting and improve generalization ability, we augment training data by using only a few gathers but massive synthetic statics values. Because the magnitude of residual statics is often small, we apply a high-resolution neural network as a backbone to learn detailed features. We also design the head of the neural network to allow multiscale training and multiscale testing, which enables different acquisition geometries for prediction. Training with labeled samples from synthetic or real data has been tested with new real data as prediction inputs. In both cases, residual statics help improve stacked sections. If the NMO assumption is valid, our method produces results comparable to stack power maximization with much less computation time. If the NMO assumption is invalid, our method produces better results than stack power maximization.

Wavelet estimation and nonstretching NMO correction

Normal-moveout (NMO) correction is an important step applied to common-midpoint (CMP) gathers for subsequent stacking. However, conventional NMO correction methods often suffer from the problem of NMO stretching, which nonlinearly increases with offsets and decreases with zero-offset traveltime. The NMO stretching can be quantified by frequency distortion, so stretching is confined mainly to large offsets and shallow times. To solve this problem, we have proposed a wavelet-based method with the following four steps. First, we estimate a wavelet from the CMP gather by using the NMO stretching appearing in the conventional NMO correction. Second, we deconvolve the original CMP gather based on the estimated wavelet. This step of removing the wavelet from the CMP gather is helpful for the next steps of NMO velocity scan and NMO correction. Third, we apply an improved NMO correction to the deconvolved CMP gather and obtain flattened reflectivities. Finally, we convolve the flattened and deconvolved gather with the estimated wavelet back to obtain an NMO-corrected gather without stretching artifacts. In our method, by using a deconvolved CMP gather, we are able to calculate a high-resolution semblance velocity spectrum that benefits from the NMO velocity picking. In addition, applying the NMO correction to the deconvolved CMP gather, instead of the original gather, is helpful in reducing the NMO stretching related to the wavelet distortion. Tests on synthetic and field data find that our new NMO correction method can estimate an accurate seismic wavelet and obtain an NMO-corrected gather without NMO stretching. Reducing the NMO stretching can significantly improve the resolution of shallow layers at far offsets and preserve the spectral bandwidth.

The Inverse Fresnel Beam XSP-CDP Stack Imaging in Crosswell Seismic

In order to overcome the shortcomings of serious arc drawing and low computational efficiency in the crosswell seismic migration method and the problems of the inaccurate velocity model and sparse distribution of reflection points in the traditional stack imaging method, the article proposes an inverse Fresnel beam XSP-CDP stack imaging method based on first-arrival wave velocity tomography combined with the characteristics of crosswell seismic wave field. Firstly, an accurate crosswell velocity model is established by the first-arrival wave tomography inversion method based on the characteristics of high energy and easy pick-up of the first-arrival wave in crosswell seismic. Secondly, the velocity model is optimized, and the energy contribution weights of effective rays to the receiver point are calculated through the crosswell seismic Fresnel beam wave field forward numerical simulation method. Then, the reflected wave field is dynamically migrated to the reflection points within the first Fresnel zone according to the weight function, and the intensive common reflection point (CRP) gather after normal moveout (NMO) correction is generated. Finally, an appropriate bin is selected for stacking. In this article, the inverse Fresnel beam method is used to decompose the single-channel seismic wave field into the effective reflection points in the Fresnel zone, which makes the fold of the reflection point more uniform and improves the imaging accuracy. The model test and actual data processing results proved the validity and robustness of this method.

Nonhyperbolic normal moveout stretch correction with deep learning automation

Normal-moveout (NMO) correction is a fundamental step in seismic data processing. It consists of mapping seismic data from recorded traveltimes to corresponding zero-offset times. This process produces wavelet stretching as an undesired by-product. We have addressed the NMO stretching problem with two methods: (1) an exact stretch-free NMO correction that prevents the stretching of primary reflections and (2) an approximate post-NMO stretch correction. Our stretch-free NMO produces parallel moveout trajectories for primary reflections. Our post-NMO stretch correction calculates the moveout of stretched wavelets as a function of offset. Both methods are based on the generalized moveout approximation and are suitable for application in complex anisotropic or heterogeneous environments. We use new moveout equations, modify the original parameter functions to be a constant over the primary reflections, and then interpolate the seismogram amplitudes at the calculated traveltimes. For fast and automatic modification of the parameter functions, we use deep learning. We design a deep neural network (DNN) using convolutional layers and residual blocks. To train the DNN, we generate a set of 40,000 synthetic NMO-corrected common-midpoint gathers and the corresponding desired outputs of the DNN. The data set is generated using different velocity profiles, wavelets, and offset vectors, and it includes multiples, ground roll, and band-limited random noise. The simplicity of the DNN task — a 1D identification of primary reflections — improves the generalization in practice. We use the trained DNN and find successful applications of our stretch-correction method on synthetic and different real data sets.