Sonar is a valuable tool for ocean exploration since it can obtain a wealth of data. With the development of intelligent technology, deep learning has brought new vitality to underwater sonar image classification. However, due to the difficulty and high cost of acquiring underwater sonar images, we have to consider the extreme case when there are no available sonar data of a specific category, and how to improve the prediction ability of intelligent classification models for unseen sonar data. In this work, we design an underwater sonar image classification method based on Image Disentanglement Reconstruction and Zero-Shot Learning (IDR-ZSL). Initially, an image disentanglement reconstruction (IDR) network is proposed for generating pseudo-sonar samples. The IDR consists of two encoders, a decoder, and three discriminators. The first encoder is responsible for extracting the structure vectors of the optical images and the texture vectors of the sonar images; the decoder is in charge of combining the above vectors to generate the pseudo-sonar images; and the second encoder is in charge of disentangling the pseudo-sonar images. Furthermore, three discriminators are incorporated to determine the realness and texture quality of the reconstructed image and feedback to the decoder. Subsequently, the underwater sonar image classification model performs zero-shot learning based on the generated pseudo-sonar images. Experimental results show that IDR-ZSL can generate high-quality pseudo-sonar images, and improve the prediction accuracy of the zero-shot classifier on unseen classes of sonar images.

Underwater target detection is pivotal for marine exploration, yet it faces significant challenges because of the inherent complex underwater environment. Sonar images are generally degraded by noise, exhibit low resolution, and lack prominent target features, making the extraction of useful feature information from blurred and complex backgrounds particularly challenging. These limitations hinder highly accurate autonomous target detection in sonar imagery. To address these issues, this paper proposes the ABFP-YOLO model, which was designed to enhance the accuracy of underwater target detection. Specifically, the bi-directional feature pyramid network (BiFPN) structure is integrated into the model to efficiently fuse the features of different scales, significantly improving the capability of the network to recognize targets of varying scales, especially small targets in complex scenarios. Additionally, an attention module is incorporated to enhance feature extraction from blurred images, thereby boosting the detection accuracy of the model. To validate the proposed model’s effectiveness, extensive comparative and ablation experiments were conducted on two datasets. The experimental results demonstrate that the ABFP-YOLO model achieves mean average precision (mAP0.5) scores of 0.988 and 0.866, indicating its superior performance in target detection tasks within complex underwater environments.

Side-scan sonar is essential to underwater target detection, yet its effectiveness is hindered by scarce annotated data and complex acoustic artifacts. This study systematically evaluates four YOLO variants, YOLOv8n, YOLOv10n, YOLOv11n, and the newly released YOLOv13n, on two public side-scan sonar datasets with limited samples and severe class imbalance. We assess detection accuracy, computational efficiency, inference speed, and transfer learning using COCO pre-trained weights, as well as the impact of optimizer choice between SGD and AdamW. The results reveal distinct strengths: YOLOv8n achieves the fastest inference at 60.98 FPS, with a competitive mAP50 of 0.906, ideal for real-time applications. YOLOv11n offers the best accuracy–efficiency balance, attaining the highest recall of 0.859 and mAP50 of 0.917. YOLOv13n demonstrates exceptional precision of 0.993 and high-IoU localization, with an mAP75 of 0.760. Transfer learning consistently boosts performance, with average mAP50:95 gains exceeding 54% on the more challenging dataset, highlighting its critical role in overcoming data scarcity. SGD generally outperforms AdamW, confirming its suitability as the default optimizer. These findings provide practical guidelines: YOLOv8 for real-time needs, YOLOv11 for balanced performance, and YOLOv13 for precision-critical tasks with ample resources. This work also establishes a benchmark for future underwater autonomous system research.

This study presents an innovative solution method for ultra-fine group slowing-down equations tailored to stochastic media with double heterogeneity (DH), focusing on advanced nuclear fuels such as fully ceramic microencapsulated (FCM) fuel and Mixed Oxide (MOX) fuel. Addressing the limitations of conventional resonance calculation methods in handling DH effects, the proposed UFGSP method (the ultra-fine group slowing-down method with the Sanchez–Pomraning method) integrates the Sanchez–Pomraning technique with the ultra-fine group transport theory to resolve spatially dependent resonance cross-sections in both matrix and particle phases. The method employs high-fidelity geometric modeling, iterative cross-section homogenization, and flux reconstruction to capture neutron self-shielding effects in stochastically distributed media. Validation across seven FCM fuel cases, four poison particle configurations (BISO/QUADRISO, Bi/Tri-structural Isotropic), and four plutonium spot problems demonstrated exceptional accuracy, with maximum deviations in effective multiplication factor k_eff and resonance cross-sections remaining within ±138 pcm and ±2.4%, respectively. Key innovations include the ability to resolve radial flux distributions within TRISO particles and address resonance interference in MOX fuel matrices. The results confirm that the UFGSP method significantly enhances computational precision for DH problems, offering a robust tool for next-generation reactor design and safety analysis.

Traditional computational methods for pressurized water reactors are unable to handle dispersed fuel particles as the double heterogeneity and the direct volumetric homogenization can result in significant errors. In contrast, reactivity-equivalent physical transformation techniques offer high precision for addressing the double heterogeneity introduced by dispersed fuel particles. This approach converts the double heterogeneity problem into a single heterogeneity problem, which is then subsequently investigated by using the conventional pressurized water reactor computational procedure. However, it is currently empirical and takes a lot of time to obtain the right k_∞. In this paper, we train the RPT model by using the existing dataset of plate-dispersed fuel and rod-dispersed fuel by a machine learning method based on a linear regression model, and we then use the new data to make predictions and derive the corresponding similarity ratios. The burnup verification, density verification, fission rate verification, and neutron energy spectrum analysis are calculated through the OpenMC program. For plate-type fuel elements, the method maintains an accuracy within 200 pcm during depletion, with deviations in the ²³⁵U density and ²³⁵U fission rate within 0.1% and neutron energy spectrum errors within 6%. For rod-type fuel elements, the method maintains an accuracy within 100 pcm during depletion, with deviations in ²³⁵U and ²³⁹Pu density within 1.5% and neutron energy spectrum errors within 1%. The numerical validation indicates that the reactivity-equivalent physical transformation method based on the linear regression model not only greatly improves the computational efficiency, but also ensures a very high accuracy to deal with double heterogeneity in nuclear reactors.

Sub-bottom profile data have the potential to characterize sediment properties but are seldom used for offshore site investigations because of uncertainties in rock-physics models. Deep-learning techniques appear to be poised to play very important roles in our processing flows for the interpretation of geophysical data. In this paper, a novel deep learning-based method for this task is proposed in which a nonlinear mapping between the observed data and sediment types is learned using a multi-attribute temporal convolution network (MATCN). Firstly, empirical mode decomposition (EMD) is employed for the original data, and intrinsic mode functions (IMFs) with multiple time scales are generated. Based on different IMFs, instantaneous frequency (IF) data under different IMFs can be obtained, while instantaneous phase (IP) and instantaneous amplitude (IA) data are obtained based on the original data. IF, IA and IP data are called attribute data, and are highly related to the attenuation, reflection, and interior structure of the sediment. Thus, IA, IF, and IP are used as the inputs, and a 1D convolutional neural network (CNN) and a time convolution network (TCN) are used to extract sequential features. Different feature representations are then fused. Combining cross-entropy loss function and class-edge loss function, the network is encouraged to produce classified results with more continuous sediment distributions compared with the traditional loss function. The real-data experiments demonstrate that the proposed MATCN has achieved good performance with an F measure greater than 70% in all cases, and greater than 80% in most cases.