We propose a new method to tackle the mapping challenge from time-series data to spatial image in the field of seismic exploration, i.e., reconstructing the velocity model directly from seismic data by deep neural networks (DNNs). The conventional way of addressing this ill-posed inversion problem is through iterative algorithms, which suffer from poor nonlinear mapping and strong nonuniqueness. Other attempts may either import human intervention errors or underuse seismic data. The challenge for DNNs mainly lies in the weak spatial correspondence, the uncertain reflection-reception relationship between seismic data and velocity model, as well as the time-varying property of seismic data. To tackle these challenges, we propose end-to-end seismic inversion networks (SeisInvNets) with novel components to make the best use of all seismic data. Specifically, we start with every seismic trace and enhance it with its neighborhood information, its observation setup, and the global context of its corresponding seismic profile. From the enhanced seismic traces, the spatially aligned feature maps can be learned and further concatenated to reconstruct a velocity model. In general, we let every seismic trace contribute to the reconstruction of the whole velocity model by finding spatial correspondence. The proposed SeisInvNet consistently produces improvements over the baselines and achieves promising performance on our synthesized and proposed SeisInv data set according to various evaluation metrics. The inversion results are more consistent with the target from the aspects of velocity values, subsurface structures, and geological interfaces. Moreover, the mechanism and the generalization of the proposed method are discussed and verified. Nevertheless, the generalization of deep-learning-based inversion methods on real data is still challenging and considering physics may be one potential solution.

Multichannel singular spectrum analysis (MSSA) is an effective algorithm for random noise attenuation in seismic data, which decomposes the vector space of the Hankel matrix of the noisy signal into a signal subspace and a noise subspace by truncated singular value decomposition (TSVD). However, this signal subspace actually still contains residual noise. We have derived a new formula of low-rank reduction, which is more powerful in distinguishing between signal and noise compared with the traditional TSVD. By introducing a damping factor into traditional MSSA to dampen the singular values, we have developed a new algorithm for random noise attenuation. We have named our modified MSSA as damped MSSA. The denoising performance is controlled by the damping factor, and our approach reverts to the traditional MSSA approach when the damping factor is sufficiently large. Application of the damped MSSA algorithm on synthetic and field seismic data demonstrates superior performance compared with the conventional MSSA algorithm.

The Cadzow rank-reduction method can be effectively utilized in simultaneously denoising and reconstructing 5D seismic data that depends on four spatial dimensions. The classic version of Cadzow rank-reduction method arranges the 4D spatial data into a level-four block Hankel/Toeplitz matrix and then applies truncated singular value decomposition (TSVD) for rank-reduction. When the observed data is extremely noisy, which is often the feature of real seismic data, traditional TSVD cannot be adequate for attenuating the noise and reconstructing the signals. The reconstructed data tends to contain a significant amount of residual noise using the traditional TSVD method, which can be explained by the fact that the reconstructed data space is a mixture of both signal subspace and noise subspace. In order to better decompose the block Hankel matrix into signal and noise components, we introduced a damping operator into the traditional TSVD formula, which we call the damped rank-reduction method. The damped rank-reduction method can obtain a perfect reconstruction performance even when the observed data has extremely low signal-to-noise ratio (SNR). The feasibility of the improved 5D seismic data reconstruction method was validated via both 5D synthetic and field data examples. We presented comprehensive analysis of the data examples and obtained valuable experience and guidelines in better utilizing the proposed method in practice. Since the proposed method is convenient to implement and can achieve immediate improvement, we suggest its wide application in the industry.

Attenuation of seismic random noise is considered an important processing step to enhance the signal-to-noise ratio of seismic data. A new approach is proposed to attenuate random noise based on a deep-denoising autoencoder (DDAE). In this approach, the time-series seismic data are used as an input for the DDAE. The DDAE encodes the input seismic data to multiple levels of abstraction, and then it decodes those levels to reconstruct the seismic signal without noise. The DDAE is pretrained in a supervised way using synthetic data; following this, the pretrained model is used to denoise the field data set in an unsupervised scheme using a new customized loss function. We have assessed the proposed algorithm based on four synthetic data sets and two field examples, and we compare the results with several benchmark algorithms, such as f-x deconvolution (f-x deconv) and the f-x singular spectrum analysis (f-x SSA). As a result, our algorithm succeeds in attenuating the random noise in an effective manner.

We used a novel iterative estimation scheme for separation of blended seismic data from simultaneous sources. The scheme is based on an augmented estimation problem that can be solved by iteratively constraining the deblended data using shaping regularization in the seislet domain. We formulated the forward modeling operator in the common-receiver domain, in which two sources were assumed to be blended using a random time-shift dithering approach. The nonlinear shaping-regularization framework offered some freedom in designing a shaping operator to constrain the model in an underdetermined inverse problem. We designed the backward operator and the shaping operator for the shaping-regularization framework. The backward operator can be optimally chosen as half of the identity operator in the two-source case, and the shaping operator can be chosen as coherency-promoting operator. The high performance deblending effect of the iterative framework was tested on three numerically blended synthetic data sets and one numerically blended field data set. Compared with alternative $f‐k$ domain thresholding and $f‐x$ predictive filtering, seislet-domain soft thresholding exhibits the most robust behavior.

We have introduced a novel time-frequency decomposition approach for analyzing seismic data. This method is inspired by the newly developed variational mode decomposition (VMD). The principle of VMD is to look for an ensemble of modes with their respective center frequencies, such that the modes collectively reproduce the input signal and each mode is smooth after demodulation into baseband. The advantage of VMD is that there is no residual noise in the modes and it can further decrease redundant modes compared with the complete ensemble empirical mode decomposition (CEEMD) and improved CEEMD (ICEEMD). Moreover, VMD is an adaptive signal decomposition technique, which can nonrecursively decompose a multicomponent signal into several quasi-orthogonal intrinsic mode functions. This new tool, in contrast to empirical mode decomposition (EMD) and its variations, such as EEMD, CEEMD, and ICEEMD, is based on a solid mathematical foundation and can obtain a time-frequency representation that is less sensitive to noise. Two tests on synthetic data showed the effectiveness of our VMD-based time-frequency analysis method. Application on field data showed the potential of the proposed approach in highlighting geologic characteristics and stratigraphic information effectively. All the performances of the VMD-based approach were compared with those from the CEEMD- and ICEEMD-based approaches.