Seismic Random Noise Attenuation Based on M-ResUNet

Improving the signal-to-noise ratio and suppressing random noise in seismic data is critical for high-precision processing. Although deep learning-based algorithms have gained popularity as denoising methods, they suffer from poor generalization ability, resulting in high training set construction cost and computation cost. To address this problem, we propose an unsupervised learning-based denoising method that includes an improved denoising strategy based on local similarity and replacement, a corresponding training method, and an improved network based on UNet. Our training method takes advantage of network convergence and allows direct training on the test region, effectively solving the problems associated with denoising methods using generalization ability while improving training performance. In addition, our network is specifically designed for the training method and incorporates various improvements that could further enhance the training effectiveness. Our method outperforms traditional denoising methods, as demonstrated by tests on synthetic and field data, with superior performance in random noise attenuation and reflection event reconstruction.

Many seismic denoising algorithms are easily trapped in the dilemma of signal leaking and noise remaining. But between the two situations, more complex and common situations happens that both signal leaking and noise remaining exist. We have presented a denoising framework for random seismic noise attenuation to mitigate the dilemma and achieve a better denoising performance. The framework consists of a denoising autoencoder (DAE) and data generator. The DAE attenuates random noise without ground truth and works with a vectoring patching technique to reduce time complexity. In the data generator, local correlation is first developed to nonlinear local correlation to detect and extract the signal leakage with noise components suppressed. Then, the extracted signal leakage is compensated back to the DAE output in a supersaturated way to generate a new record. After that, DAE is used again to suppress the remaining noise in the new record. The supersaturated compensation of the first signal leakage counteracts the signal leakage of the second DAE noise attenuation. Formally, the proposed framework can be implemented like a nonlinear geological inversion by thinking of the data generator as the forward operator. Four 2D/3D synthetic and field examples test the performance of the proposed framework. Experiments demonstrate that the proposed method is effective in random noise attenuation and signal leakage compensation, and thus significantly mitigates the mentioned dilemma.

The expected patch log likelihood (EPLL) is a patch-prior based denoising method which ensures denoising performance in ensemble approximation by local patch denoising. Basing on that, we propose a classification based EPLL method to attenuate seismic random noise, of which the level varies spatiotemporally. In EPLL method, the Gaussian mixture model (GMM) learned from samples is taken as a prior to statistically model patches, then the seismic signal is reconstructed by weighted averaging the noisy image and summation of denoised patches. Since the accuracy of the local statistic modeling and global signal reconstruction is related to the variance of the noise of the patches that have various noise variance in seismic image, we classify the patches into several groups in order to minimize the within-class variance. Therefore, appropriate regularization parameter and most likely prior are assigned to each patch according to the noise variance of the patch in EPLL. Experimental results of synthetic and field seismic data show that the classification based EPLL achieves a desired performance in seismic events preservation and nonstationary random noise attenuation.

Summary We have developed a novel random noise attenuation method for seismic data based on inversion in the frequency-space (f-x) domain. The method regards ‘clean’ seismic data in f-x domain as model parameters, and the noise attenuation is equivalent to inverting these parameters from the noisy data. The prediction error filter (PEF) in the f-x prediction method is used as lateral continuity constraints of the inversion. Since the PEF is calculated from the noisy data, it should be recalculated from the denoised seismic data. The process of calculating signal, then getting a new PEF should be iterated a few times. The choice of trade-off parameters also influences the performance of this method. The proposed method overcame the problem of the generation of spurious events and reduction of the signal amplitudes in the conventional f-x prediction method. Tests on synthetic and field data examples demonstrated that, compared with f-x domain prediction, the proposed method has a better performance in noise attenuation and signal-preserving.

The high-fidelity attenuation of random ground penetrating radar (GPR) noise is important for enhancing the signal-noise ratio (SNR). In this paper, a novel network structure for convolutional denoising autoencoders (CDAEs) was proposed to effectively resolve various problems in the noise attenuation process, including overfitting, the size of the local receptive field, and representational bottlenecks and vanishing gradients in deep learning; this approach also significantly improves the noise attenuation performance. We described the noise attenuation process of conventional CDAEs, and then presented the output feature map of each convolutional layer to analyze the role of convolutional layers and their impacts on GPR data. Furthermore, we focused on the problems of overfitting, the local receptive field size, and the occurrence of representational bottlenecks and vanishing gradients in deep learning. Subsequently, a network structure optimization strategy, including a dropout regularization layer, an atrous convolution layer, and a residual-connection structure, was proposed, namely convolutional denoising autoencoders with network structure optimization (CDAEsNSO), comprising an intermediate version, called atrous-dropout CDAEs (AD-CDAEs), and a final version, called residual-connection CDAEs (ResCDAEs), all of which effectively improve the performance of conventional CDAEs. Finally, CDAEsNSO was applied to attenuate noise for the H-beam model, tunnel lining model, and field pipeline data, confirming that the algorithm adapts well to both synthetic and field data. The experiments verified that CDAEsNSO not only effectively attenuates strong Gaussian noise, Gaussian spike impulse noise, and mixed noise, but it also causes less damage to the original waveform data and maintains high-fidelity information.

: Noise attenuation performance was measured on the Joint Service Aircrew Mask (JSAM) Fixed Wing (FW) variant with the HGU-55/P, HGU-68/P, and HGU-84/P helmets at the Air Force Research Laboratory's (AFRL) Acoustics facilities at Wright-Patterson Air Force Base in March 2011. American National Standards Institute (ANSI) S12.6 and S12.42 methods were used to measure the passive attenuation of the systems. The objective of this study was to identify the noise attenuation performance and determine if the performance requirements were met as defined in the JSAM-FW Performance Specification. The addition of the JSAM-FW degraded the noise attenuation performance when compared to all the helmets alone: the objective (ANSI S12.42) passive insertion loss (across most frequencies) and the subjective noise attenuation (ANSI S12.6) performance (only low frequencies).

Time-frequency peak filtering (TFPF) is a new and effective tool for random-noise attenuation in the time-frequency domain. The conventional TFPF processes each channel of the seismic record independently with a fixed window length (WL). However, different frequency signals have different optimal WLs. Obviously, a fixed WL cannot effectively attenuate random noise for all frequency components at the same time. Depending on the geometry of the reflection, we can assume that the moveout of the reflected event is locally linear. With this in mind and taking the spatial correlation of the reflection events between adjacent channels in different layers into account, we propose a novel approach which is to do the TFPF along the local direction of the reflection event instead of along the channel. This method is called the local parallel radial-trace TFPF. It not only has the advantage of TFPF in denoising but also reduces the sensitivity of WL, making the filtering more flexible and effective. Both the synthetic model and seismic data have proved its better performance in noise attenuation and effective component preservation.

Self-adaptive denoising net: Self-supervised learning for seismic migration artifacts and random noise attenuation

The attenuation of seismic field noise using self-supervised deep learning has gained attention due to its label-free training process. However, common self-supervised methods are limited by the pixelwise independence assumption, which does not align with field seismic noise characteristics, and suffer from signal leakage due to receptive fields containing inherent blind spots or traces. In this paper, we propose a self-supervised random noise attenuation method based on the non-pixelwise independence assumption. By considering the spatial correlation map of field noise, we extend the blind spot to a generalized blind neighborhood, ensuring that the prediction pixel is not influenced by neighboring pixels with noise correlation greater than zero. The blind neighborhood size controls how much spatial correlation is disrupted, allowing our method to handle random noise with varying spatial correlation. Since larger blind neighborhoods may lead to signal loss, we introduce an automatic trade-off between noise correlation disruption and signal preservation during training. Experiments on real seismic noise attenuation (including random and tracewise coherent noise) demonstrate the superiority of our method in destroying the spatial coherence of noise and preventing useful signal leakage.

In recent years, deep learning dense matching techniques have developed rapidly. Among the various training methods such as supervised, unsupervised and semi-supervised, the matching accuracy of supervised training methods is still much higher than other training methods. However, this training method requires the use of manually labelled parallax maps as samples. The accuracy of the labelling will directly affect the matching accuracy of the trained network model. Therefore, it needs to be analyzed and studied. In this paper, the noise immunity of the deep learning supervised method is studied by simulating systematic error, random error and gross error, and the experimental results show that: (i) The deep learning method has anti-noise ability within a certain range and has a better anti-noise effect on random noise, but the matching accuracy decreases at an accelerated rate as the noise increases. (ii) The pre-training method by transfer learning can effectively improve the matching accuracy, increase the noise immunity, and make the original non-converging network converge.

Deep-learning-based methods have been successfully applied to seismic data random noise attenuation. Among them, the supervised deep-learning-based methods dominate the unsupervised ones. The supervised methods need accurate noise-free data as training labels. However, the field seismic data cannot meet this requirement. To circumvent it, some researchers utilized realistic-looking synthetic data or denoised results via conventional methods as labels. The former ones encounter the problem of weak generalization ability because it requires the same distribution of test and training data. The latter ones encounter the issue of insufficient denoising ability because its denoising ability is difficult to significantly exceed the conventional methods which were used to generate labels. To avoid preparing noise-free labels, we propose a novel deep learning framework for attenuating random noise of prestack seismic data in an unsupervised manner. The prestack seismic data, such as common-reflection-point (CRP) gathers and common-midpoint (CMP) gathers after normal moveout (NMO) correction, have high self-similarity. It is because their events are coherent in the time–space domain and approximately horizontal from shallow to deep layers. The generator convolutional neural network (GCN) first learns self-similar features before any learning. The useful signals are more self-similar than random noise, which is incoherent and randomly distributed. Therefore, the GCN extracts features of useful signals before random noise. We select the specific training iteration and adopt the early stopping strategy to suppress random noise. Both synthetic and field prestack seismic data examples demonstrate the validity of our methods.

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.

ABSTRACTTensor algebra provides a robust framework for multi‐dimensional seismic data processing. A low‐rank tensor can represent a noise‐free seismic data volume. Additive random noise will increase the rank of the tensor. Hence, tensor rank‐reduction techniques can be used to filter random noise. Our filtering method adopts the Candecomp/Parafac decomposition to approximates a N‐dimensional seismic data volume via the superposition of rank‐one tensors. Similar to the singular value decomposition for matrices, a low‐rank Candecomp/Parafac decomposition can capture the signal and exclude random noise in situations where a low‐rank tensor can represent the ideal noise‐free seismic volume. The alternating least squares method is adopted to compute the Candecomp/Parafac decomposition with a provided target rank. This method involves solving a series of highly over‐determined linear least‐squares subproblems. To improve the efficiency of the alternating least squares algorithm, we uniformly randomly sample equations of the linear least‐squares subproblems to reduce the size of the problem significantly. The computational overhead is further reduced by avoiding unfolding and folding large dense tensors. We investigate the applicability of the randomized Candecomp/Parafac decomposition for incoherent noise attenuation via experiments conducted on a synthetic dataset and field data seismic volumes. We also compare the proposed algorithm (randomized Candecomp/Parafac decomposition) against multi‐dimensional singular spectrum analysis and classical prediction filtering. We conclude the proposed approach can achieve slightly better denoising performance in terms of signal‐to‐noise ratio enhancement than traditional methods, but with a less computational cost.

Seismic data denoising is a critical component of seismic data processing, yet effectively removing erratic noise, characterized by its non-Gaussian distribution and high amplitude, remains a substantial challenge for conventional methods and deep-learning (DL) algorithms. Supervised learning frameworks typically outperform others, but they require pairs of noisy data sets alongside corresponding clean ground truth, which is impractical for real-world seismic data sets. In contrast, unsupervised learning (UL) methods, which do not rely on ground truth during training, often fall short in performance when compared with their supervised or traditional denoising counterparts. Moreover, current unsupervised DL methods fail to address the specific challenges posed by erratic seismic noise adequately. This paper introduces a novel zero-shot unsupervised DL framework designed specifically to mitigate random and erratic noise, with a particular emphasis on blended noise. Drawing inspiration from Noise2Noise (N2N) and data augmentation principles, we develop a robust self-supervised denoising network called robust Noiser2Noiser. Our approach eliminates the need for paired noisy and clean data sets as required by supervised methods or paired noisy data sets as in N2N. Instead, our framework relies solely on the original noisy seismic data set. Our methodology generates two independent recorrupted data sets from the original noisy data set, using one as the input and the other as the training target. Subsequently, we use a DL-based denoiser, denoising convolutional neural network, for training purposes. To address various types of random and erratic noise, the original noisy data set is recorrupted with the same noise type. Detailed explanations for generating training input and target data for blended data are provided. We apply our network to synthetic and real marine data examples, demonstrating significantly improved noise attenuation performance compared with traditional denoising methods and state-of-the-art UL Codes are available on https://github.com/Ji-seismic/N2N_deblending.

Seismic data usually contain various types of noise. Noise comes from signal collection environment, equipment (current noise), micro seismic, etc. Noise reduces the signal- to-noise ratio and resolution of seismic records, and affects the accuracy of seismic processing and stratigraphic interpretation. Improving the signal-to-noise ratio of seismic data is the primary task of seismic data processing. In practical research, many algorithms for suppressing random noise interference are proposed. In order to achieve better seismic data denoising technology, this paper introduces a new algorithm: improved shearlet transform algorithm. The basic principle of this algorithm is to perform shearlet transform on noisy seismic data, calculate the L2 norm of shearlet coefficients in different directions at each scale according to the distribution law of effective signals at different scales, and remove random noise according to the adaptive threshold in shearlet domain varying with scale and direction. Note: This paper was accepted into the Technical Program but was not presented at IMAGE 2022 in Houston, Texas.