XDXintong Dong

Xintong Dong

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Xintong Dong has a total of 81 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • A transformer-based super-resolution method: A novel solution for improving the quality of stacked or migrated seismic data

    by Shiqi Dong, Xintong Dong, Kaiyuan Zheng, Ming Cheng, Tie Zhong, Hongzhou Wang

    Seismic data obtained after stacking or migration typically exhibit low dominant frequency, sparse spatial and temporal sampling, and may retain a low signal-to-noise ratio even after stacking. To enhance the resolution of seismic data, a deep-learning-based single-step super-resolution method was proposed that simultaneously performed denoising, interpolation, and frequency extrapolation. A seismic data super-resolution transformer (SDST) was designed to extract and fuse local and global features from low-resolution seismic data, with particular emphasis on the amplitude characteristics and spatial continuity of effective events (e.g., horizons, folds, and faults). Edge information from the input data was extracted using the Canny algorithm and was used as a mask to generate an edge map, which was then concatenated with the raw data to form a two-channel input, thereby improving amplitude preservation and noise suppression. The SDST backbone comprised residual groups containing Swin Transformer layers, which captured features within windows while reducing the computational cost of self-attention by restricting attention to finite-sized windows. Pixel shuffle layers were used to efficiently upsample the output feature maps from the SDST backbone, thereby enhancing the details of geologic structures. In addition, upsampling the input data through a skip connection helped preserve the output amplitudes, particularly for weak events. Three-dimensional synthetic seismic volumes with complex geologic structures were prepared, half of which contained a mixture of strong- and weak-amplitude events. Two-dimensional slices were then randomly selected from these volumes to construct training data sets that better approximate the characteristics of field data for supervised learning. Numerical tests on synthetic and field data from different exploration regions demonstrated the feasibility of the proposed method.

  • A Novel Training Strategy for the Interpolation of Consecutively Missing Case

    by Xinyang Wang, Jun Ma, Xintong Dong, Hongzhou Wang

    Self‐supervised learning (SSL) reduces the dependence on labelled training data by exploiting the structure of unlabelled data. However, since SSL struggles with capturing high‐level semantic information for complex tasks, it faces some difficulties in dealing with challenging tasks such as consecutively missing traces and weak signal recovery under strong noise. Although supervised learning can address these challenges, it demands high‐quality training sets, which are often unavailable. To achieve a good balance between the performance and reliance on the training set, we propose a pseudo‐self‐supervision training strategy‐based (PSL‐based) by introducing the idea of SSL into the classical supervised framework and using the three‐dimensional (3D) consecutive missing reconstruction to test its effectiveness. For incomplete seismic volumes, limited complete regions (sharing extremely similar properties as missing sections) can serve as some training data, yet they are typically too scarce to fulfil the requirements for effective training. Thus, we propose a 3D diffusion model for data augmentation and use the augmented data to construct paired training sets for network training. Since the proposed strategy still trains the network by learning the intrinsic features of the data themselves, we refer to it as the ‘PSL strategy’. Experimental results on both prestack and poststack field datasets demonstrate the effectiveness of the proposed PSL‐based training strategy compared with SSL‐based and supervised training strategies.

  • A globally oriented and lightweight seismic interpolation network: More global than convolutional neural networks and more efficient than transformers

    by Changxin Wei, Ming Cheng, Xintong Dong, Shaoping Lu, Zhenbo Guo, Zhiqi Wei

    Due to certain environmental limitations and economic considerations, seismic interpolation is occasionally used to compensate for sparse and incomplete geometries. In recent years, many deep learning (DL)-based methods have shown great promise in seismic data interpolation. Most DL-based interpolation methods adopt architectures that rely on convolutional neural networks, which are inherently limited in capturing global features, owing to their local receptive fields. Transformer-based architectures have been recently proposed in several attempts to capture such global features and thus improve the interpolation performance. However, the core operation, multihead self-attention (MSA), imposes a substantial computational burden. The purpose of this study was to explore a lightweight and globally-oriented DL architecture for seismic data interpolation that achieves performance comparable to, or even surpassing, transformers while substantially reducing the heavy computational cost. To this end, the Mamba module was, for the first time, adopted in the context of seismic data interpolation. Different from transformers, the Mamba uses the selective state space model (SSM) to capture the global features instead of the MSA. The selection mechanism and hardware-aware algorithm in SSM could significantly alleviate the heavy computational burden by compressing features and reducing memory swaps. Therefore, several Mamba modules were incorporated into a feed-forward neural network, with dense connections introduced to enhance the feature interactions among different layers. In synthetic and field examples, the proposed dense Mamba-based interpolation network achieved a superior trade-off between the computational cost and interpolation performance, compared with two existing architectures: U-Net and an encoder-stacked transformer.

  • 3-D Seismic Random Noise Attenuation via Self-Supervised Conditional Diffusion Model

    by Xinyang Wang, Qianyu Ge, Xintong Dong

    Effective noise suppression and data quality enhancement are critical in seismic data processing. Over the years, many traditional and deep learning methods have been proposed and achieved some success. However, traditional methods often struggle to achieve satisfactory denoising precision and require significant manual effort for parameter optimization, making them less efficient for large-scale applications. Deep learning methods commonly require a huge number of training data and lack of some generalization abilities in different applications. Additionally, conventional diffusion models are typically designed under the assumption of Gaussian noise, which might not hold in field seismic data. To address these limitations, we introduce a Self-Supervised Conditional Diffusion model (SSCDiff) for 3-D seismic random noise attenuation. In SSCDiff, we introduce the conditional diffusion model to solve the problem of random noise attenuation rather than Gaussian noise only. Moreover, SSCDiff incorporates a self-supervised reverse process that allows the network to leverage seismic data priors and optimize using a loss function without pre-training. Clean results are obtained through iterative variational distribution sampling. Furthermore, we integrate fast fourier convolution (FFC) modules into the network, enabling the effective utilization of frequency domain information to enhance the denoising ability. Comprehensive synthetic and field seismic experiments prove the superior performance of SSCDiff in handling seismic data with across different datasets and noise levels.

  • Open Access

    Light-weighted foundation model for seismic data processing based on representative and non-redundant pre-training dataset

    by Xintong Dong, Wenshuo Yu, Jun Lin, Zhenbo Guo, Hongzhou Wang, Jianhao Yang

    In the fields of computer vision (CV) and remote sensing (RS), foundational models typically follow the "big data + large model parameters" paradigm. However, the application of this strategy in seismic data processing faces several challenges: seismic data is difficult to obtain and the scarcity of publicly available datasets make it difficult to construct large-scale datasets. Additionally, the high computational cost associated with a large number of model parameters restricts widespread research in this domain. Therefore, we propose a lightweight seismic processing foundational model paradigm (SPFM), which aims to overcome the limitations of traditional methods by data engineering and network architecture innovation. Specifically, we propose an innovative dataset construction strategy that generates more seismic data by data augmentation techniques, including collecting publicly available field data and using generative diffusion models (GDM) for data enhancement. Furthermore, we optimize the data distribution by employing dimensionality reduction, cluster analysis, and stratified sampling methods, reducing redundant information while preserving important seismic features, thus constructing a comprehensive dataset. In terms of network architecture design, we introduce the selective structured state-space model (Mamba) structure, which effectively captures global features of seismic data and alleviates the quadratic growth of computational complexity inherent in Transformer-based models, thereby improving computational efficiency. This model, pre-trained with only four A800 GPUs, outperforms traditional methods across multiple tasks, including denoising, interpolation, frequency-band extrapolation, and resolution enhancement. The lightweight paradigm provides an solution for seismic data processing, advancing the generalization and accessibility of seismic data processing.

  • A multitask deep-learning model for the denoising, interpolation, and wavefield separation of distributed acoustic sensing-vertical seismic profiling data

    by Ming Cheng, Jun Lin, Xintong Dong, Tie Zhong

    Distributed acoustic sensing (DAS) is an innovative data acquisition method and is usually combined with vertical seismic profiling (VSP) techniques for downhole seismic exploration, offering some advantages such as low cost, high resistance to temperature and pressure, and a high spatial sampling rate. A preprocessing workflow can significantly improve the quality of prestack seismic data for subsequent inversion and high-resolution imaging. Denoising, interpolation, and wavefield separation are three necessary preprocessing steps for prestack DAS-VSP records. Traditionally, these three steps are carried out independently, resulting in certain shortcomings, such as the amplitude attenuation of recovered signals, higher computational costs, and extensive parameter adjustments. In this work, we develop a multitask preprocessing model (MTPM) to combine the three steps. It can simultaneously perform denoising, interpolation, and wavefield separation operations on DAS-VSP data using only a single well-trained model. The network architecture of MTPM is a combination of two classical frameworks: a convolutional neural network (CNN) and a transformer. Specifically, this developed MTPM comprises a front-Net and a post-Net. The front-Net is a transformer-based encoder-decoder structure enhanced by a U-Net++ block. This hybrid CNN-transformer structure can simultaneously extract local and global features using convolution and multihead self-attention operations, which are important for denoising and interpolation tasks. The post-Net is a pure transformer that focuses only on global features and thus enhances the performance of wavefield separation. Furthermore, we design a two-stage training strategy for MTPM to combine the three tasks. In our experiments, we use synthetic and field DAS-VSP records to test the effectiveness of MTPM, which demonstrates better denoising, interpolation, and wavefield separation performance compared with other methods.


  • Deep-learning-based global feature capture for seismic data reconstruction

    by Xintong Dong, Changxin Wei, Jun Lin, Shaoping Lu, Ming Cheng, Rongzhe Zhang

    Compared with regularly and irregularly missing traces, consecutively missing traces (also called large gaps) create a more challenging task in seismic data interpolation. Recently, deep-learning-based methods have proven to be an effective tool for seismic data interpolation. However, widely adopted convolutional neural network (CNN) architectures, which are constrained by their limited receptive fields, emphasize local features while overlooking global features. This inherent characteristic limits their ability to reconstruct the more challenging large gaps. Compared with CNN-based architectures, transformers can capture more global features using a multihead self-attention mechanism. Therefore, we develop a novel transformer-based architecture, the dense double-branch attention transformer (D2AT), to interpolate large gaps in seismic data when the model is properly trained. The backbone of D2AT is a double-branch structure based on the Swin transformer, which enables it to learn low- and high-resolution features. Moreover, in the high-resolution upper branch, we use dense connections to facilitate feature interactions at different stages, and a spatial attention block is used to refine the extracted features. Experimental results on synthetic and field seismic records demonstrate that D2AT is a potential solution for reconstructing the consecutively missing traces in 2D seismic data.


  • Self-supervised diffusion model for 3D seismic data reconstruction

    by Xinyang Wang, Qianyu Ge, Xintong Dong, Shiqi Dong, Tie Zhong

    Seismic data reconstruction is an effective tool for compensating for nonuniform and incomplete seismic data. Compared with the methods for 2D seismic data, 3D reconstruction methods can consider more spatial correlation in seismic data. However, most existing 3D reconstruction methods are mainly theory driven and have some limitations due to their prior assumptions about seismic data. To overcome these limitations, deep-learning-based reconstruction methods emerged and have shown some potential in practical applications. However, there are mainly two shortcomings of the existing deep-learning methods. On the one hand, most of the existing deep-learning-based methods adopt convolutional neural networks, which have some difficulties in dealing with data that have complex or time-varying distributions. Recently, the diffusion model has been reported to have the capability to handle data with complex distributions. On the other hand, existing methods need sufficient paired data to train the network, which are difficult to obtain, especially in three dimensions. Deep prior-based unsupervised and sampling-based self-supervised networks offer a solution to this problem. Motivated by these, we develop a self-supervised diffusion model for 3D seismic data reconstruction (3D-SSDM). Our model mainly contains a diffusion restoration model and a variational time-spatial module. Synthetic and field data experiments demonstrate the superiority of our 3D-SSDM algorithm.