Data-driven seismic inversion has achieved certain advancements. However, these methods often require a large number of expensive well logs, limiting their application only to mature or synthetic data. This article presents ContrasInver, a method that achieves seismic inversion using as few as two or three well logs, significantly reducing the current requirements. In ContrasInver, two key innovations are proposed to address the challenges of applying semi-supervised learning to regression tasks with ultra-sparse labels: 1) the region-growing training (RGT) strategy leverages the inherent continuity of seismic data, effectively propagating accuracy from closer to more distant regions based on the proximity of well logs. To realize this concept, a multidimensional sample generation (MSG) method is also proposed that produces a large number of diverse samples from a single well, while establishing lateral continuity within the seismic data; 2) the impedance vectorization projection (IVP) vectorizes impedance values and performs semi-supervised learning in a compressed space. The Jacobian matrix derived from this space can filter out some outlier components in pseudo-label vectors, thereby solving the value confusion issue in semi-supervised regression learning. In the experiments, ContrasInver achieved state-of-the-art performance on the synthetic SEAM I data. In the field data with two or three well logs, only the methods based on the components proposed in this article were able to achieve reasonable results. It is the first data-driven approach yielding reliable results on the Netherlands F3 and Delft, using only three and two well logs, respectively.

Interpolation and reconstruction of seismic data are critical steps in geophysical exploration, with results largely dependent on the performance of the interpolation techniques and the available feature information in the data. The task becomes particularly challenging when faced with complex data loss scenarios, such as high proportions of random discrete missing data and large amounts of random continuous missing data. To address this challenge, we propose a new method: generative nonlocal a priori augmented Wasserstein generative adversarial network (GNP-WGAN). The method uses a non local prior extraction (NLE) module improved by an edge detection algorithm to capture the structural information of seismic data, and a generative multidimensional attention restorer (GMAR) designed based on causal and axial attention to generate a smooth and accurate generative nonlocal prior (GNP). The Wasserstein GAN with gradient penalty is then augmented with GNP for finer and more accurate seismic data reconstruction. Finally, the MS-SSIM-$L_{1}$loss function is introduced to improve the quality of the generator reconstruction. Experiments on synthetic and field seismic datasets demonstrate the superior performance of GNP-WGAN in reconstructing seismic data with complex missing cases. In addition, subsequent experiments show that our GNP can be easily integrated as a plug-in into most of the currently popular reconstruction models to improve the accuracy and structural integrity of the reconstruction results and also exhibits enhanced robustness.

The interpolation and reconstruction of missing traces are crucial steps in seismic data processing; moreover, it is also a highly ill-posed problem, especially for complex cases such as high-ratio random discrete missing, continuous missing, and missing in fault-rich or salt body surveys. These complex cases are rarely mentioned in current works. To cope with complex missing cases, we propose multidimensional adversarial generative adversarial network (MDA GAN), a novel 3-D GAN framework. It keeps the anisotropy and spatial continuity of the data after 3-D complex missing reconstruction using three discriminators. The feature splicing module is designed and embedded in the generator to retain more information of the input data. The tanh cross entropy (TCE) loss is derived, which provides the generator with the optimal reconstruction gradient to make the generated data smoother and continuous. We experimentally verified the effectiveness of the individual components of the study and then tested the method on multiple publicly available data. The method achieves reasonable reconstructions for up to 95% of random discrete missing and 100 traces of continuous missing. In fault and salt body enriched surveys, MDA GAN still yields promising results for complex cases. Experimentally, it has been demonstrated that our method achieves better performance than other methods in both simple and complex cases. Moreover, our network does not require training weights for each survey, the same weights it uses are applied to multiple surveys, significantly reducing time and computational costs, and we make the model publicly available onhttps://github.com/douyimin/MDA_GAN.

Salt bodies are significant reservoir structures, and there are still difficulties in interpreting them end-to-end from 3-D seismic data. Conventional semi-supervised learning struggles with obtaining high-quality pseudo labels early on, affecting subsequent model performance. Moreover, complex background noise hinders the accuracy of salt body predictions, while a strategy of gradually feeding training blocks leads to fragmented and confusing results. To address these challenges and restore realistic subsurface salt profiles, we have proposed an innovative, fully automated, and refined 3-D salt interpretation method called 3D Salt-HSM. In this method, we have designed a hybrid semi-supervised training paradigm based on stable pseudo labels and multilevel consistency constraints. This approach allows us to obtain high-quality pseudo labels for salt bodies and fully explore their features in unlabeled segmented blocks. We have also introduced a multitask learning strategy for fine interpretation of salt bodies, ranging from image level to pixel level. This strategy helps alleviate the adverse impact of interfering textures on salt body prediction. In addition, we have incorporated a contextual feature fusion module (CFFM) based on the multiscale context of salt bodies. This module enables the network to capture the global information of seismic images and achieve fine-grained salt body interpretation. In our experiments on the SEAM and F3 seismic datasets, we utilized only 3% of the labels for supervised learning, while the remaining data were used for unsupervised learning and validation. The experimental results demonstrate that 3D Salt-HSM outperforms previous state-of-the-art (SOTA) methods in terms of salt body segmentation performance, producing highly satisfactory results.

Data-driven fault detection has been regarded as a 3D image segmentation task. The models trained from synthetic data are difficult to generalize in some surveys. Recently, training 3D fault segmentation using sparse manual 2D slices is thought to yield promising results, but manual labeling has many false negative labels (abnormal annotations), which is detrimental to training and consequently to detection performance. Motivated to train 3D fault segmentation networks under sparse 2D labels while suppressing false negative labels, we analyze the training process gradient and propose the Mask Dice (MD) loss. Moreover, the fault is an edge feature, and current encoder-decoder architectures widely used for fault detection (e.g., U-shape network) are not conducive to edge representation. Consequently, Fault-Net is proposed, which is designed for the characteristics of faults, employs high-resolution propagation features, and embeds Multi-Scale Compression Fusion block to fuse multi-scale information, which allows the edge information to be fully preserved during propagation and fusion, thus enabling advanced performance via few computational resources. Experimental demonstrates that MD loss supports the inclusion of human experience in training and suppresses false negative labels therein, enabling baseline models to improve performance and generalize to more surveys. Fault-Net is capable to provide a more stable and reliable interpretation of faults, it uses extremely low computational resources and inference is significantly faster than other models. Our method indicates optimal performance in comparison with several mainstream methods.

Detection faults in seismic data are a crucial step for seismic structural interpretation, reservoir characterization, and well placement. Some recent works regard it as an image segmentation task. The task of image segmentation requires huge labels, especially 3-D seismic data, which has a complex structure and lots of noise. Therefore, its annotation requires expert experience and a huge workload. In this study, we presented$\lambda $-binary cross-entropy (BCE) and$\lambda $-smooth$L_{1}$loss to effectively train 3D-CNN by some slices from 3-D seismic volume label, so that the model can learn the segmentation of 3-D seismic data from a few 2-D slices. In order to fully extract information from limited data and suppress seismic noise, we proposed an attention module that can be used for active supervision training and embedded in the network. The attention map label is generated by the original label and letting it supervise the attention module using the$\lambda $-smooth$L_{1}$loss. The experimental results demonstrate that the proposed loss function can extract 3-D seismic features from a few 2-D slice labels. And it also shows the advanced performance of the attention module, which can significantly suppress the noise in the seismic data while increasing the sensitivity of the model to the foreground. Finally, on the public test set, the proposed method achieved similar performance to using 3-D volume labels by using only 3.3% of the slices.