Seismic Data Processing Research Articles

Improving shallow and deep seismic-while-drilling with a downhole pilot in a desert environment

Processing seismic data from drillbit-generated vibrations requires a reliable source signature for correlation and deconvolution purposes. Recently, a land field trial has been conducted in a desert environment. A memory-based downhole vibration accelerometer has been used together with a more conventional top-drive sensor to continuously record the pilot signal from 590 to 8600 ft (180–2621 m). Past results indicate that seismic-while-drilling (SWD) data processed using the top-drive accelerometer exhibit good quality in the middle sections of the well but a reduced signal-to-noise ratio for shallow and deep sections. One of the main challenges in using the downhole pilot is a substantial drift of the downhole clock time. To resolve it, a novel automated time-alignment procedure using the GPS-synchronized signal of the top-drive sensor as a reference is applied. The downhole recording provides a source signature of better quality. In shallow sections of the well, it helps to overcome the intense surface-related vibrational noise, whereas, in deeper sections, it provides a cleaner extraction of weaker signals from the polycrystalline diamond compact bits. Processing with the downhole pilot results in better surface seismic data quality than with a conventional top-drive sensor. Therefore, enabling the use of the synchronized downhole pilot signal is of crucial importance for SWD applications. Modern cost-effective near-bit vibrational sensors widely used for different nonseismic applications could be an effective acquisition solution, as shown in this study.

Introduction to this special section: Seismic reservoir modeling

The seismic data set is a fundamental requirement for producing oil and gas fields. It provides understanding of the structure and stratigraphy of the reservoirs, and it is now routinely employed in reservoir modeling for advancing insights into how fields are being produced. Modern 3D seismic data were first developed in the 1960s, but it wasn't until the 1980s that seismic interpretation software enabled the building of gridded reservoir models from seismic interpretations. Reservoir modeling utilizing seismic interpretations drove insights into reservoir quality and performance, helping to understand the communication between reservoir units and wells, particularly in fields with many wells. But key challenges such as the cost of building or updating reservoir models and scale variance created barriers for early industry-wide adoption. Data integration required calibration to correct and account for the difference in measurement scales of seismic data and well data, as well as to create robust relationships between seismic properties and petrophysical properties in the model. Over time, technological advancements led to a reduction in the cost of reservoir modeling, while increased acquisition, processing, and utility of seismic data provided the means to drive innovation toward incorporating seismic. Today, 3D and 4D seismic data play pivotal roles in defining and updating reservoir models where hundreds to thousands of simulations can be realized in a reservoir model to explore history matching and model uncertainties.

Software for Interactive and Automated Seismic and Infrasonic Data Processing

Software for Interactive and Automated Seismic and Infrasonic Data Processing

Deep unfolding dictionary learning for seismic denoising

Seismic denoising is an essential step for seismic data processing. Conventionally, dictionary learning (DL) methods for seismic denoising always assume the representation coefficients to be sparse and the dictionary to be normalized or a tight frame. Current DL methods need to update the dictionary and the coefficients in an alternating iterative process. However, the dictionary obtained from the DL method often needs to be recalculated for different input data. Moreover, the performance of DL for seismic noise removal is related to the parameter selection and the prior constraints of dictionary and representation coefficients. Recently, deep learning demonstrates promising performance in data prediction and classification. Following the architecture of DL algorithms strictly, we have developed a novel and interpretable deep unfolding dictionary learning (DUDL) method for seismic denoising by unfolding the iterative algorithm of DL into a deep neural network (DNN). The proposed architecture of DUDL contains two main parts: the first is to update the dictionary and representation coefficients using least-squares inversion and the second is to apply a DNN to learn the prior representation of dictionary and representation coefficients, respectively. Numerical synthetic and field examples find the effectiveness of our method. More importantly, this method for seismic denoising obtains the dictionary for different seismic data adaptively and is suitable for seismic data with different noise levels.

Reflection seismic imaging across a greenstone belt, Abitibi (Ontario), Canada

AbstractEstablished 2D seismic data processing methods such as Kirchhoff pre‐stack time migration and Kirchhoff pre‐stack depth migration function relatively well for regular acquisition geometries and well‐constrained velocity models. Recently developed focusing pre‐stack depth migration methods have the potential to enhance image quality in the case of sparse and non‐regular source–receiver distribution. We have tested the performance of the coherency migration method as one of these focusing migration approaches in comparison to standard dip‐moveout and Kirchhoff pre‐stack time migration techniques by applying them to the Swayze East seismic profile acquired in the Abitibi greenstone belt of Canada. This seismic profile represents a crooked‐line survey that intersects several metal‐bearing deformation zones, providing good target geometries to examine various pre‐stack migration methods. Analysis of the seismic data indicates reflectivity associated with shallowly dipping reflections appears relatively well preserved over the entire 0–6 km offset range for the frequency range between 20 to 90 Hz. Although most reflections are visible already in either the dip‐moveout or the Kirchhoff pre‐stack time migration results, the coherency migration method delivers the most improved image showing all reflective structures inferred for this area. The comparisons suggest the coherency migration method can be considered as superior in terms of resulting seismic image quality compared with conventional approaches for this type of crooked‐line seismic survey in such a complex geological setting.

Algorithms and Measuring Complex for Classification of Seismic Signal Sources, Determination of Distance and Azimuth to the Point of Excitation of Surface Waves

Machine learning and digital signal processing methods are used in various industries, including in the analysis and classification of seismic signals from surface sources. The developed wave type analysis algorithm makes it possible to automatically identify and, accordingly, separate incoming seismic waves based on their characteristics. To distinguish the types of waves, a seismic measuring complex is used that determines the characteristics of the boundary waves of surface sources using special molecular electronic sensors of angular and linear oscillations. The results of the algorithm for processing data obtained by the method of seismic observations using spectral analysis based on the Morlet wavelet are presented. The paper also describes an algorithm for classifying signal sources, determining the distance and azimuth to the point of excitation of surface waves, considers the use of statistical characteristics and MFCC (Mel-frequency cepstral coefficients) parameters, as well as their joint application. At the same time, the following were used as statistical characteristics of the signal: variance, kurtosis coefficient, entropy and average value, and gradient boosting was chosen as a machine learning method; a machine learning method based on gradient boosting using statistical and MFCC parameters was used as a method for determining the distance to the signal source. The training was conducted on test data based on the selected special parameters of signals from sources of seismic excitation of surface waves. From a practical point of view, new methods of seismic observations and analysis of boundary waves make it possible to solve the problem of ensuring a dense arrangement of sensors in hard-to-reach places, eliminate the lack of knowledge in algorithms for processing data from seismic sensors of angular movements, classify and systematize sources, improve prediction accuracy, implement algorithms for locating and tracking sources. The aim of the work was to create algorithms for processing seismic data for classifying signal sources, determining the distance and azimuth to the point of excitation of surface waves.

A ray perturbation method for the acoustic attenuating transversely isotropic eikonal equation

Modeling complex-valued traveltimes is helpful for developing attenuation-associated techniques for seismic data processing. Transverse isotropy can explain the directional variation of velocity and attenuation anisotropy of long-wavelength seismic waves in many sedimentary rocks. The acoustic attenuating transversely isotropic eikonal equation can be used to accurately calculate P-wave complex-valued traveltimes under such a geologic condition. However, no ray-tracing system for this eikonal equation could be found in the literature until now. We have developed a ray perturbation method to solve this eikonal equation. Unlike all existing ray perturbation methods, our newly proposed method does not perturb the exact ray-tracing system but splits the acoustic attenuating transversely isotropic eikonal equation into a nonlinear partial differential equation (PDE) and a first-order PDE. These two PDEs can be solved by the method of characteristics. This gives rise to two sets of ray-tracing equations for the real and imaginary parts of the complex-valued traveltimes, respectively. Both sets of ray-tracing equations share the same raypath, which allows us to merge them into a complete ray-tracing system for complex-valued traveltimes. Numerical examples are used to demonstrate the high accuracy of the newly proposed ray-tracing system, analyze the complex-valued traveltimes of diving P waves, and compare the modeled complex-valued traveltimes with those extracted from constant-[Formula: see text] viscoelastic waveforms.

Machine Learning in Earthquake Seismology

Machine learning (ML) is a collection of methods used to develop understanding and predictive capability by learning relationships embedded in data. ML methods are becoming the dominant approaches for many tasks in seismology. ML and data mining techniques can significantly improve our capability for seismic data processing. In this review we provide a comprehensive overview of ML applications in earthquake seismology, discuss progress and challenges, and offer suggestions for future work. ▪ Conceptual, algorithmic, and computational advances have enabled rapid progress in the development of machine learning approaches to earthquake seismology. ▪ The impact of that progress is most clearly evident in earthquake monitoring and is leading to a new generation of much more comprehensive earthquake catalogs. ▪ Application of unsupervised approaches for exploratory analysis of these high-dimensional catalogs may reveal new understanding of seismicity. ▪ Machine learning methods are proving to be effective across a broad range of other seismological tasks, but systematic benchmarking through open source frameworks and benchmark data sets are important to ensure continuing progress.

An Interactive System Based on the IASP91 Earth Model for Earthquake Data Processing

System software for interactive human–computer data processing based on the IASP91 Earth model was designed. An interactive data processing system for visualizing earthquake data was designed and implemented via the Intel Fortran platform. The system reads and processes broadband seismic data acquired by field stations, mainly including the reading and import of raw data, pre-processing, identification of seismic phases and inter-correlation traveltimes picking. In the data processing step, shortcomings have been improved and functions have been gradually refined and enhanced, making it easier and faster to process data. It has already processed more than 1000 large seismic events received by the station from 2013 to 2018. The practical application shows that the human–computer interaction system is easy to operate, accurate, fast and flexible, and is an effective tool for processing seismic data.

3D MODEL DESİGNED FOR FOOT OF HORİZON I OF MAYKOP İN NAFTALAN FİELD APPLYİNG SEİSMİC ATTRİBUTE ANALYSİS

The paper is devoted to design of 3D model of the study area by use of attribute analysis of 3D seismic data applied to study the geology of Naftalan oil field in more detail. The field is located in Naftalan-North Naftalan area of Ganja Oil and Gas Province, Yevlakh-Aghjabadi depression. The paper expounds the data on Naftalan field considered an ancient brachianticline type of field in Azerbaijan. The history of study of the field using geological and geophysical techniques has been briefly given. Geological evaluation of the cube derived by processing of 3D seismic data acquired in Naftalan area in 2012 made it possible to outline a confidently traced interval reflecting the foot of Horizon I of Maykop. Through selection of areas featured by variation of wavefield characteristics, we have calculated several attributes in this interval and analyzed the acquired results gaining more accurate seismic data and confident tracing of seismic horizons. To avoid repetition, some poorly informative attributes or attributes providing similar results have not been applied in further studies. Our study aims to define attributes (RMS amplitude (Root mean square), Variance/Edge method, Relative Acoustic Impedance) that are more effective for outlining disjunctive dislocations of various amplitude and can be applied for 3D model design of the geology of study area. Processing results of all 3D data have been given in form of cubes of seismic attributes. Further analysis of these attribute cubes enables us to study the geological setting of the foot of Horizon I of Maykop and select the most effective attributes. As a result, the 3D model of the study area for seismic horizon reflecting the foot of Horizon I of Maykop has been designed. Keywords: attribute analysis, 3D seismic survey, Naftalan, seismic horizon, Maykop deposits, dislocations.

Seismic data IO and sorting optimization in HPC through ANNs prediction based auto-tuning for ExSeisDat

ExSeisDat is designed using standard message passing interface (MPI) library for seismic data processing on high-performance super-computing clusters. These clusters are generally designed for efficient execution of complex tasks including large size IO. The IO performance degradation issues arise when multiple processes try accessing data from parallel networked storage. These complications are caused by restrictive protocols running by a parallel file system (PFS) controlling the disks and due to less advancement in storage hardware itself as well. This requires and leads to the tuning of specific configuration parameters to optimize the IO performance, commonly not considered by users focused on writing parallel application. Despite its consideration, the changes in configuration parameters are required from case to case. It adds up to further degradation in IO performance for a large SEG-Y format seismic data file scaling to petabytes. The SEG-Y IO and file sorting operations are the two of the main features of ExSeisDat. This research paper proposes technique to optimize these SEG-Y operations based on artificial neural networks (ANNs). The optimization involves auto-tuning of the related configuration parameters, using IO bandwidth prediction by the trained ANN models through machine learning (ML) process. Furthermore, we discuss the impact on prediction accuracy and statistical analysis of auto-tuning bandwidth results, by the variation in hidden layers nodes configuration of the ANNs. The results have shown the overall improvement in bandwidth performance up to 108.8% and 237.4% in the combined SEG-Y IO and file sorting operations test cases, respectively. Therefore, this paper has demonstrated the significant gain in SEG-Y seismic data bandwidth performance by auto-tuning the parameters settings on runtime by using an ML approach.

Nonstationary seismic reflectivity inversion based on prior-engaged semisupervised deep learning method

Reflectivity inversion methods based on a stationary convolution model are essential for seismic data processing. They compress the seismic wavelet, and by broadening the bandwidth of seismic data, they assist the interpretation of seismic sections. Unfortunately, they do not apply to realistic nonstationary deconvolution cases in which the seismic wavelet varies as it propagates in the subsurface. Deep learning techniques have been proposed to solve inverse problems in which networks can behave as the regularizer of the inverse problem. Our goal is to adopt a semisupervised deep learning approach to invert reflectivity when the propagating wavelet is considered unknown and time-variant. To this end, we have designed a prior-engaged neural network by unrolling an alternating iterative optimization algorithm, in which convolutional neural networks are used to solve two subproblems. One is to invert the reflectivity, and the other is to estimate the time-varying wavelets. In general, it is well known that, when working with geophysical inverse problems such as ours, one has limited access to labeled data for training the network. We circumvent the problem by training the network via a data-consistency cost function in which seismic traces are honored. Reflectivity estimates also are honored at spatial coordinates in which true reflectivity series derived from borehole data are available. The cost function also penalizes time-varying wavelets from varying abruptly along the spatial direction. Experiments are conducted to find the effectiveness of our method. We also compare our approach to a nonstationary blind deconvolution algorithm based on regularized inversion. Our findings reveal that the developed method improves the vertical resolution of seismic sections with noticeable correlation coefficient improvements over the nonstationary blind deconvolution. In addition, our method is less sensitive to initial estimates of nonstationary wavelets. Moreover, it needs less human intervention when setting parameters than regularized inversion.

Fleet's Geode: A Breakthrough Sensor for Real-Time Ambient Seismic Noise Tomography over DtS-IoT.

As most of the outcropping and shallow mineral deposits have been found, new technology is imperative to finding the hidden critical mineral deposits required to transition to renewable energy. One such new technique, called ambient seismic noise tomography, has shown promise in recent years as a low-cost, low environmental impact method that can image under cover and at depth. Wireless and compact nodal seismic technology has been instrumental to enable industry applications of ambient noise tomography, but these devices are designed for the active seismic reflection method and do not have the required sensitivity at low frequencies for ambient noise tomography, and real-time data transmission in remote locations requires significant infrastructure to be installed. In this paper, we show the development and testing of the Geode-a real-time seismic node purpose-built by Fleet Space Technologies for ambient seismic noise tomography on exploration scales. We discuss the key differences between current nodal technology and the Geode and show results of a field trial where the performance of the Geode is compared with a commercially popular nodal geophone. The use of a 2 Hz high sensitivity geophone and low noise digitiser results in an instrument noise floor that is more than 30 dB lower below 5 Hz than nodes that are commonly used in the industry. The increased sensitivity results in signal-to-noise ratios in the cross-correlation functions in the field trial that are more than double that of commercially available nodal geophone at low frequencies. When considering the full bandwidth of retrievable correlations in our study, using the Geode would reduce the required recording time from 75 h to 32 h to achieve an average signal-to-noise ratio in the cross-correlation functions of 10. We also discuss the integration of a real-time direct-to-satellite Internet of Things (DtS-IoT) modem in the Geode, which, together with edge processing of seismic data directly on the Geode, enables us to image the subsurface in real-time. During the field trial, the Geodes successfully transmitted more than 90% of the available preprocessed data packets. The Geode is compact enough so that several devices can be carried and installed by one field technician, whilst the array of stations do not require a base station to transmit data to the cloud for further processing. We believe this is the future of passive seismic surveys and will result in faster and more dynamic seismic imaging capabilities analogous to the medical imaging community, increasing the pace at which new mineral deposits are discovered.

Deep learning strategy for salt model building

Velocity models are crucial intermediate products generated in seismic data processing, and the model’s accuracy is essential for constructing quality seismic images. Conventional approaches to velocity-model building (VMB) use a family of inversion methods, among which are ray-based tomography and full-waveform inversion. These methods have been highly optimized throughout the years but are still heavily dependent on continuous human curation of the results, which leads to an overall high time cost, especially in areas with high structural complexity, such as those containing salt tectonics. We have investigated a deep learning (DL) approach that accurately defines salt geometries for VMB. We train our convolutional neural network on synthetic shot gather data, explore a manner of leveraging information through summation of shot data, and determine the influence that the choice of loss function has on the quality and aspect of predicted velocity models. Our residual U-Net model trained on data containing only randomly shaped salt bodies can estimate geologically complex salt geometries such as those in 2D SEG/EAGE salt model slices. Our results find that deeper encoder-decoder models with shortcut connections resolve velocity model structures better than shallower models. Moreover, network models trained with a composite loss function — combining mean absolute error and the multiscale structural similarity index — better delineate the contours of areas with high-velocity contrast and better recover regions with a uniform velocity trend than network models trained with conventional loss functions like the mean squared error. The residual U-Net and loss functions that we use are not task-specific and can be extended to other DL approaches to VMB.

Perturbation-based moveout approximation in the elastic transversely isotropic media with a vertical symmetry axis

Explicit approximations in traveltime are usually used in seismic data processing such as simulation through ray tracing, velocity analysis for model building, and imaging from time migration. Accurate approximation is key to achieving high-resolution imaging of the structure. We have proposed a perturbation-based approximation for P- and S-waves traveltime in a transversely isotropic medium with a vertical symmetry axis. The perturbation coefficients in this approximation are derived directly from the implicit parametric offset-traveltime equation. The comparison with the generalized moveout approximation (GMA) for P and S waves is applied in numerical examples. We find that our proposed approximation indicates higher accuracy than the GMA counterpart for P and S waves. The error sensitivity of P and S waves in the depth, vertical velocity ratio, and the anellipticity parameter also is analyzed to examine the influence of these factors. In addition, we investigate the anisotropy estimation through two approximations for P and S waves and find that our proposed perturbation method results in high accuracy, which can significantly improve the data processing quality. For the S wave, our approximation also is more accurate than the GMA counterpart and could be implemented for joint tomographic inversions.

3-D quantitative seismic imaging of tectonically-influenced Middle-Eocene carbonates stratigraphic traps of SE-Asian basins using spectral decomposition: Implications for hydrocarbon exploration

Carbonates form energetic stratigraphic petroleum systems. The tectonic uplifting and subsidence significantly affect the overall reservoir configurations. During these stratigraphic scenarios, the prediction of accurate thickness for subsidence and uplifting zones becomes challenging. The band-limited seismic data interpretation tools fail to detect the zones of thinning and thickening of sediments along with uplifting and subsidence. The band-limited seismic does not have the tuning frequency, that can be used to predict the accurate vertical thickness of the thin-bedded carbonates. The constant wavelet-based transformation tool (WT) and inverted reservoir simulations have robust implications for the quantification of any depositional system. Especially, when the reservoir is shallow marine, the application of WT has resolved a variety of stratigraphic traps in terms of their seismic and sedimentological aspects. The stratigraphic traps are comprised of thin beds, which are not resolvable on conventional seismic. These traps require a specific tuning frequency, which is the key ingredient for the prediction of oil and gas inside them and can only be obtained by the implementation of the WT tool. Therefore, the WT, static stratigraphic thickening wedge models (SSTWM), and WT–based instantaneous spectral thickness modeling (ISTM) tools are applied on a gas field in Indus Basin, Pakistan. The 43-Hz CWT images the densely fractured and faulted network within the shoal reservoir limestone lenticular lens (SRLL) with an aerial extent of 208 km2 in southwest-northeast orientations. The band-limited attributes revealed poor imaging of sub-surface carbonate systems. The 43-Hz wavelet-based SSTWM predicts a 15-m-thick coarse-grained SRLL along subsidence and a 12-m-thick fine-grained SRLL along uplifting zones. Post-stack processing at 43-Hz CWT tuning frequency predicts the 18 and 10 m thick sediments for SRLL along the subsidence and uplifting zone. The ISTM predicts a thickness of 31 m during subsidence within the SRLL and 14 m for uplifting zones. The 43-Hz tuning frequency-based processing at ISTM predicts enhanced thicknesses of 41 and 18 m for subsiding and uplifting sediments that are trapped inside the shoal limestone petroleum systems. The quantitative visualization of AI with the predicted thicknesses at ISTM has provided robust clues for imaging the internal stratigraphic architecture of the SRLL. The linear regression analysis remained a very successful tool for assessing the quantitative aspects of SRLL with a strong R2 > 0.95, which has implicated the real-time imagining of sea-level fall and rise. Only post-stack seismic data processing at 43-Hz found a very effective exploration tactic in predicting the enhanced thicknesses, and lateral continuity of the reservoir, and improved the S/N by removing the tuning effect of ambiguous lithology/fluids, which have accurately predicted the exact location along the subsidence zones. This workflow can act as an analogue stratigraphic exploration of remaining gas fields within Asian Basins and similar geological settings.

Seismic Coherent Noise Removal of Source Array in the NSST Domain

The technique of the source array based on the vibroseis can provide the strong energy of a seismic wave field, which better meets the need for seismic exploration. The seismic coherent noise reduces the signal-to-noise ratio (SNR) of the source array seismic data and affects the seismic data processing. The traditional coherent noise removal methods often cause some damage to the effective signal while suppressing coherent noise or cannot suppress the interference wave effectively at all. Based on the multi-scale and multi-direction properties of the non-subsampled Shearlet transform (NSST) and its simple mathematical structure, the seismic coherent noise removal method of source array in NSST domain is proposed. The method is applied to both the synthetic seismic data and the filed seismic data. After processing with this method, the coherent noise of the seismic data is greatly removed and the effective signal information is greatly protected. The analysis of the results demonstrates the effectiveness and practicability of the proposed method on coherent noise attenuation.

Cross-Well Reflection Imaging at the Verkhnekamskoe Potash Deposit

Abstract —We propose the technique of extended processing of cross-well seismic data for reflection imaging. Based on the analysis of the wave field and synthetic modeling, a graph for processing cross-well data is developed to separate the reflections in the presence of a boundary of a sharp change in the velocity of elastic waves. The migration of the reflected waves from the time scale to the depth is performed by solving a forward problem for each source–receiver pair. As a result, the position of the reflection points is calculated, after which all the traces are stacked based on the binning grid. The input information for performing migration and solving a forward problem is the velocity characteristic of the massif, which is calculated using traveltime tomography on direct body waves obtained from the same data set of cross-well survey and vertical seismic profiling. The resulting depth seismic section has a much higher resolution than that of vertical seismic profiling and ground-based shallow seismic studies. This opens up the opportunity of identifying different small natural or technogenic objects in the interwell space.

Structure-Preserving Random Noise Attenuation Method for Seismic Data Based on a Flexible Attention CNN

The noise attenuation of seismic data is an indispensable part of seismic data processing, directly impacting the following inversion and imaging. This paper focuses on two bottlenecks in the AI-based denoising method of seismic data: the destruction of structural information of seismic data and the inferior generalizability. We propose a flexible attention-CNN (FACNN) and realized the denoising work of seismic data. This paper’s main work and advantages were concentrated on the following three aspects: (i) We propose attention gates (AGs), which progressively suppressed features in irrelevant background parts and improved the denoising performance. (ii) We added a noise level map M as an additional channel, making a single CNN model expected to inherit the flexibility of handling noise models with different parameters, even spatially variant noises. (iii) We propose a mixed loss function based on MS_SSIM to improve the performance of FACNN further. Adding the noise level map can improve the network’s generalization ability, and adding the attention structure with the mixed loss function can better protect the structural information of the seismic data. The numerical tests showed that our method has better generalization and can better protect the details of seismic events.

Ground roll intelligent suppression based on spatial domain synchrosqueezing wavelet transform convolutional neural network

AbstractGround roll is usually considered as a common linear noise in land seismic data. The existence of the ground roll often masks the effective reflection information of underground media, resulting in the deterioration of seismic data quality. Therefore, ground roll suppression is one of the main tasks in seismic data processing. A large number of previous studies have proved that the time‐frequency signal processing method based on mathematical transformation has shown excellent performance in ground roll attenuation and still has development potential. Meanwhile, a convolutional neural network, as one of the popular deep learning technologies, has also been widely used in the field of seismic signal processing. In this paper, we combine the convolutional neural network with the time‐frequency signal processing method based on mathematical transformation, that is, spatial domain synchrosqueezing wavelet transform, and propose a complete ground roll suppression workflow of shot gathers in spatial wavenumber domain, realizing high‐precision and automatic ground roll removal. Field data examples show that compared with bandpass filtering, FK filtering, time domain synchrosqueezing wavelet transform, spatial domain synchrosqueezing wavelet transform and the convolutional neural network, the spatial domain synchrosqueezing wavelet transform convolutional neural network has achieved satisfactory results in effectively attenuating ground roll and retaining valid information.