Systems based on Time Delay Neural Networks (TDNNs) have achieved remarkable success in speaker verification due to their favorable balance between accuracy and computational cost. However, under short-utterance conditions, two limitations arise: limited information hinders the extraction of discriminative speaker features, and the inherent one-dimensional (1D) convolutions lack translation equivariance to frequency shifts, restricting robustness and feature utilization. To address these issues, we propose SRC-TDNN (Split Residual Collaborative TDNN), a lightweight architecture that integrates a two-dimensional (2D) convolutional component with a TDNN component. The 2D convolutional component employs parallel split residual convolutional (SplitResConv) blocks to capture multi-scale time–frequency features, followed by a Fourier Fusion Module (FFM) to enhance feature discriminability. By sharing weights along the frequency axis, this component achieves translation equivariance, alleviating the impact of frequency shifts in short utterances. The TDNN component further aggregates the fused features along the temporal axis while inheriting the translation equivariance from the 2D component. It also adopts parallel 1D SplitResConv blocks to capture both short-term phonetic details and long-term speaker characteristics across multiple temporal scales. A multi-branch architecture is employed throughout, enabling both network-level ( macro) and block-level (micro) multi-scale modeling to better exploit limited information. Evaluations on four test sets with varying utterance lengths demonstrate that the proposed method achieves the best overall performance, particularly under short-utterance conditions. Compared with ECAPA-TDNN, SRC-TDNN reduces the number of parameters by only 0.4 M while achieving at least a 30% relative improvement.

In recent years, Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) have become the mainstream segmentation methods for high-resolution remote sensing images (HRSIs). CNNs can quickly acquire the correlation between local neighboring pixels through convolutional operations, but it is difficult to establish global contextual relationships, resulting in limited segmentation accuracy. ViTs are able to establish reliable global semantic dependencies through the mechanism of self-attention, but the quadratic computational complexity of self-attention makes the ViTs present high accuracy but low efficiency. Therefore, in this letter, to balance the efficiency and accuracy of HRSIs segmentation, we combine the respective advantages of CNNs and ViTs to propose the FSegNet network. Specifically, we introduce FasterViT and utilize its efficient hierarchical attention to mitigate the surge in self-attention computation due to the high resolution of HRSIs. On this basis, we construct a lightweight decoder based on intensive computation, which achieves fast generation of segmentation results by reshaping and mapping multi-level features. Experiments on the ISPRS Potsdam and Vaihingen datasets show that the proposed FSegNet best balances performance and efficiency. The code is available at https://github.com/Rowan-L/FSegNet.

In geophysical exploration methods, the Semi-Airborne Transient Electromagnetic (SATEM) technique offers substantial exploration depth and the capability to acquire precise underground characteristics. However, the SATEM signal received by the induction coil has a low amplitude in the middle and late stages and is susceptible to various noise effects. Although a variety of conventional denoising methods are available today, their performance is often unsatisfactory, necessitating manual adjustments of the denoising outcomes. With the emergence of deep learning in various fields, several SATEM denoising neural network models have been proposed. However, these models are discriminative model and focus on modeling conditional probabilities, with insufficient generalization capabilities when faced with small data sets. On the other hand, the Variational Diffusion Model (VDM) is a generative model that captures the joint distribution and exhibits excellent generalization capabilities. Building upon this premise, we explored a VDM-based denoising approach for SATEM. Nevertheless, due to the uncontrollable nature of the results generated by VDM, it is not possible to use VDM directly for denoising SATEM signals. To address this issue, we introduce the VDM-based denoiser with constraints, which incorporates guiding conditions to constrain the model and generate desired denoising results, and proposed a new SATEM signal denoising network called TEMSGnet. To further enhance the practical applicability of our method in real-world scenarios, we incorporate a wavelet transform-based supervised fine-tuning strategy. Experimental results demonstrate that TEMSGnet achieves effective denoising performance on both synthetic and real datasets. Furthermore, through inversion, TEMSGnet can more accurately approximate the true subsurface characteristics, thereby effectively reflecting the denoised outcomes. The code is available at https://github.com/woldier/TEMSGnet.

In recent years, much attention has been paid to using deep learning techniques for land seismic first break (FB) picking. Among these, the deep learning-based 3-D FB picking method has shown stronger noise resistance than the 2-D method, as it simultaneously considers the correlation of FBs in both crossline and inline directions. In existing studies, whether using 2-D or 3-D methods, they are often regarded as a binary semantic segmentation problem of pre- and post-FB (PPFB). The FB time is determined by dividing the image with a binary classification mask. However, due to the instability of the decision threshold setting of the PPFB mask strategy, different thresholds can cause different FB picking results. Considering this, we propose a two-channel (2C) mask strategy that utilizes the feature interaction between a strip-like mask and a line-like mask to predict the confidence of FBs in seismic traces, thus realizing the FB picking with uniqueness. In addition, to enhance the global feature acquisition capability of the network, we construct a 3-D FB picking network USwinNet based on Swin Transformer, which establishes a larger receptive field through a multistage self-attention (SA) mechanism. Experimental results demonstrate that our method significantly improves the picking accuracy compared to existing methods. It also exhibits strong generalization capabilities, making it more suitable for practical engineering under complex geological conditions.

Seismic data often suffer from missing traces, and traditional reconstruction methods are cumbersome in parameterization and struggle to handle large-scale missing data. While deep learning has shown powerful reconstruction capabilities, convolutional neural networks’ (CNNs) point-to-point reconstruction may not fully cover the distribution of the entire dataset and may suffer performance degradation under complex missing patterns. In response to this challenge, we propose a novel diffusion model reconstruction framework tailored for 3-D seismic data. To facilitate 3-D seismic data reconstruction using diffusion models, we introduce conditional constraints into the diffusion model, constraining the generated data of the diffusion model based on the input data to be reconstructed. We introduce a 3-D neural network architecture into the diffusion model and refine the diffusion model’s generation process by incorporating existing parts of the data into the generation process, resulting in reconstructions with higher consistency. Through ablation studies determining optimal parameter values, although the sampling time is longer, our method exhibits superior reconstruction accuracy when applied to both field datasets and synthetic datasets, effectively addressing a wide range of complex missing patterns. Our implementation is available athttps://github.com/WAL-l/SeisFusion.

For a long time, the 2-D and 3-D Magnetotelluric (MT) forward modeling is mainly accomplished by computational methods. Traditional methods are time consuming due to the large amounts of discrete grids and slow solution of the matrix equation. Therefore, finding a fast forward modeling algorithm remains a major concern. In recent years, deep learning has provided new ways to accomplish this goal. Most existing deep learning-based MT forward modeling are performed on 2-D data, and there is a lack of research on the feasibility of 3-D problems. this paper constructs a large-scale 3-D MT dataset; employs a deep neural network suitable for 3-D MT data patterns, and improving the training efficiency through a transfer learning strategy for similar tasks, that can predict the apparent resistivity and phases in different polarization directions, and realizes fast and high-precision 3-D MT deep learning forward modeling. The experimental quantitative metrics show that the mean relative errors of apparent resistivity and phase are 0.6042% and 0.2423%, respectively, and the mean absolute errors are 1.6726 and 0.0994, respectively. When applying the method to geoelectric models that are more complex than the training set, accurate forward modeling results validate its generalization ability. The research may provide methodological and data support for larger-scale 3-D MT forward modeling in the future.

The application of neural networks to seismic first break (FB) picking research has been developed for many years. Numerous multitrace FB picking methods based on convolutional neural networks (CNNs) have been proposed. Among them, the pickup method of semantic segmentation networks based on fully convolutional networks (FCNs) is proven to have stronger noise immunity. However, when in the data with drastic variations in the FBs between local adjacent traces, because of the feature extraction method of FCNs for convergence information around the data, the network output feature edge tends to smooth, which leads to a flattening of the FBs of the multitrace pickup, which is not conducive to picking traces with drastic intertrace FB variations. Therefore, we use Transformer, a weight-transfer model that relies on the self-attention (SA) mechanism to calculate weights between input and output data, to extract FB features. The 2-D seismic data are flattened into a 1-D sequence along the time dimension of the trace and input to the network, and the spatial information of the FBs of the adjacent traces is considered without disrupting the time-series semantic information. The correlation between any element of the sequence and other elements is calculated to obtain the sequence feature weights. We use Swin Transformer as the backbone and combine the features of U-shaped networks to design an end-to-end FB picking network—STUNet. The results show that STUNet has higher FB picking accuracy than current FCNs and is more effective in local adjacent traces where the FB time variation is drastic.