Restoring coastal images from high tide to low tide conditions is crucial for applications such as environmental monitoring, coastal planning, and disaster management. Image dewatering, the process of transforming high-tide images into their corresponding low-water-level states to reveal submerged objects, remains a challenging task. Unmanned aerial vehicle (UAV)-based image dewatering presents several unique difficulties. First, there is a scarcity of suitable pre-existing datasets for supervised learning and model training, necessitating extensive data collection to construct an appropriate dataset. Additionally, aligning and matching image pairs is complex. Due to the operational characteristics of UAVs, precisely aligning high-altitude images captured at the same geographic location is difficult, which negatively impacts model training and evaluation. To address these challenges, this study proposes Water2LandNet, an unsupervised UAV image dewatering model based on generative adversarial networks (GANs), which enables training with unpaired images. This approach effectively resolves the data-pairing problem present in our dewatering dataset, OUCD (OUC_UAV_DEWATER). Extensive experiments on the OUCD dataset show that the proposed model effectively removes water from UAV images, as evidenced by qualitative and quantitative evaluations. Furthermore, compared with traditional supervised methods, Water2LandNet demonstrates greater robustness to texture inconsistencies and illumination variations arising from dynamic marine environments. Experimental results confirm that the model can reconstruct visually realistic low-tide scenes even when trained on limited unpaired data, offering a practical solution for large-scale coastal mapping and post-disaster reconstruction.

Localization is a crucial technique for determining target locations in underwater acoustic internet of things (UAIoT) systems. However, the presence of stratified oceanic environments and anchor position uncertainties significantly undermine localization accuracy. To rigorously investigate their adverse effects on localization performance, this study presents a comprehensive theoretical analysis of hybrid RSS/TOA-based localization under stratified propagation and inaccurate anchors in UAIoT systems. First, a stratified acoustic propagation model is established, and the corresponding Cram

r-Rao lower bound (CRLB) is derived for RSS and TOA measurements. The analysis is then extended to a hybrid scenario incorporating anchor inaccuracy using the Banachiewicz-Schur theorem, resulting in a unified expression for the theoretical performance bound. Simulation results confirm the validity of the derived CRLB across diverse scenarios. Compared with using a single ranging modality, the derived lower-bound performance of the proposed hybrid RSS/TOA approach improves by at least 19.8% under accurate-anchor conditions and 24.1% under inaccurate-anchor conditions. The findings provide theoretical benchmarks and practical guidance for designing robust localization strategies in stratified oceanic environments.

Sea surface height (SSH) is a critical parameter for characterizing ocean dynamics and understanding mesoscale eddies, surface currents, and subsurface thermohaline structures. Using along-track data from Haiyang-2B (HY-2B) as a reference, this study evaluated three SSH products for the year 2022 globally and in specific regions: the Copernicus Marine and Environment Monitoring Service (CMEMS) delayed-time (DT) merged gridded product, the CMEMS near-real-time (NRT) merged gridded product, and the Global Ocean Reanalysis and Simulation System (GLORYS) reanalysis product. The CMEMS DT product outperformed the others, achieving the lowest global root mean square error (RMSE, 0.0267 m) and highest correlation, and demonstrated superior signal reconstruction, particularly in dynamically active regions such as the Kuroshio region. The NRT product offers reliable large-scale monitoring capability but shows higher errors in energetic regions, whereas GLORYS, despite its multivariate consistency, exhibits limited skill in resolving smaller-scale signals. For studies focusing on fine-scale and mesoscale features, the DT product is strongly recommended. For real-time large-scale applications, the NRT product is suitable with due regard to limitations in high-dynamic regions. Conversely, when employing GLORYS reanalysis in dynamically active areas, users should carefully evaluate its capacity to reproduce small- and mesoscale processes.

To address the long-standing professional knowledge bottlenecks in scientific marine research and aquaculture, this paper proposes a marine reasoning large language model construction framework based on structured reasoning chain-of-thought (SRCoT) fine-tuning and a knowledge graph (KG). To implement the framework, an indent-driven article heuristic search method is first adopted to construct a marine-domain-specific dataset, followed by the development of a sliding window and weight-matrix-based strategy for dataset deduplication. Subsequently, a marine-domain KG is constructed, and an entity entailment method based on pointwise mutual information vectors is designed. Finally, a model post-training approach integrating SRCoT and three-stage direct preference optimization (DPO) is proposed. The base model is fine-tuned on the marine-domain SRCoT dataset and post-trained using the three-stage DPO strategy. During deployment, the custom-built marine-domain KG is used as an external reference to enhance the model responses. The experimental results demonstrate that the model trained with the proposed framework achieves performance improvements in marine-domain complex reasoning tasks and is effective in mitigating over-reasoning and refining model responses.

Underwater images play an increasingly important role in scientific research and industrial fields such as marine military, marine environmental protection, and marine engineering. However, owing to nonuniform lighting conditions, the quality of underwater imaging is often degraded by remarkable color distortion and detail loss. Although existing traditional underwater image enhancement methods have advanced, they are still limited by scarce and low-quality samples, making it difficult to achieve satisfactory results. In this study, a novel network based on the structure-guided former is proposed to effectively address the challenges of color correction and illumination enhancement in underwater images. The proposed cross-axial compression transformer block preserves the powerful global modeling capacity of transformers while significantly enhancing local feature extraction. In addition, the introduction of structural prior information not only guides the feature reconstruction process in the decoder but also effectively corrects color casts and enhances high-frequency details. Comparative experiments and ablation experiments on publicly available datasets have validated the effectiveness of the proposed method.

As air pollution increases and the energy crisis intensifies, marine renewable energy technologies have rapidly become crucial. Compared with traditional ships, the integrated onboard energy microgrid system enables pollution-free, renewable, and efficient energy utilization. However, the integration of electricity, hydrogen, and heat within an integrated-energy shipborne microgrid system presents challenges to existing optimization methods. Therefore, given that traditional ship energy models struggle to effectively handle the uncertainty of photovoltaic output, this paper proposes a novel two-level optimization framework based on the multi-objective artificial hummingbird algorithm to achieve multi-energy collaborative scheduling. This model integrates diesel generators, hydrogen fuel cells, photovoltaic systems, energy storage systems (ESSs), and heat storage devices, and it responds to spatiotemporal fluctuations in the marine environment through an electric-hydrogen-thermal coupling mechanism. The upper-level model achieves the optimal scheduling of power generation equipment and loads, and the lower-level optimization model is established to reduce the lifetime loss of an ESS. An improved multi-objective artificial hummingbird algorithm is introduced to obtain the optimal scheduling solution of the two-level optimization scheduling model. Simulation results demonstrate that the proposed optimization method not only reduces greenhouse gas emissions by 44.6% but also increases the cycle life of the ESS by 8.06%.

The Arctic Ocean ecosystem is undergoing dramatic changes as the sea ice cover retreats, underscoring the need for technologies that can monitor biological responses in the upper ocean, particularly beneath sea ice, where traditional ship-based investigations are severely constrained. Fish such as polar cod, which depend strongly on under-ice habitats, are widely regarded as key indicator species of Arctic warming, making their observation crucial for understanding ecosystem changes. In this study, we developed and deployed an integrated, fully automated observation system for long-term fish monitoring under the Arctic sea ice. The system combines a custom-designed underwater multi-focal automatic camera system (UMACS), a robust computing platform, and a satellite communication module to realize an autonomous ‘detection-to-transmission’ workflow under extreme polar conditions. A deep learning-based detection model trained on a multi-source dataset was implemented to address challenges such as low illumination, turbidity, and blurred backgrounds. A three-month continuous deployment in the Central Arctic Ocean demonstrated the robust engineering performance of the system under realistic field conditions. Although no fish were unambiguously confirmed, highlighting the intrinsic difficulty of discriminating small biological targets against a pure water background in this region, the system successfully achieved persistent, unattended, under-ice visual observation with data return. Therefore, this study provides a practical and transferable engineering framework for the scalable, technology-driven ecological monitoring of one of the planet’s most remote and fragile marine environments.

Precise dynamic positioning of autonomous underwater vehicles (AUVs) faces significant challenges because of actuator nonlinearities such as input saturation and backlash, which degrade system performance and stability in subsea operations. This work proposes a disturbance observer-based event-triggered prescribed-time dynamic positioning (DOB-ETPTDP) control framework to address these issues. The scheme reformulates the AUV kinematics as a dynamic position error system, modeling actuator nonlinearities as a bounded lumped disturbance. A key innovation is the prescribed-time disturbance observer that ensures rapid and accurate disturbance estimation, enhancing robustness against unknown actuator dynamics. A prescribed-time backstepping-based control law uses these estimates, to guarantee the asymptotic convergence of the dynamic position error to zero, independent of initial conditions. An adaptive prescribed-time event-triggered mechanism further optimizes control efficiency by reducing update rates and preventing Zeno behavior. Numerical simulations verified the effectiveness of the proposed DOB-ETPTDP scheme. In terms of convergence speed, the proposed method achieved improvements of approximately 55% and 67% compared with the sliding mode and backstepping approaches, respectively. Regarding computational complexity, the proposed method reduced the average computational load by about 20%. Moreover, the average data transmission ratio was significantly reduced, conserving more than 65% of the communication resources relative to the backstepping strategy. Rigorous stability analysis validated the theoretical guarantees, and extensive simulations confirmed that the proposed DOB-ETPTDP approach ensures high-accuracy dynamic positioning with enhanced robustness under complex actuator constraints.

Deep learning models that combine convolutional neural networks (CNNs) and long short-term memory (LSTM) networks have demonstrated strong capabilities in spatiotemporal feature extraction, proving effective for applications such as ocean environment monitoring and forecasting. Specialized artificial intelligence (AI) processors are often required for marine equipment with constrained computational resources and energy budgets to handle AI workloads. However, the distinct computational and memory access patterns of CNNs and LSTMs present significant challenges for designing efficient edge AI processors; existing hardware accelerators often struggle to efficiently support the heterogeneous computational patterns, irregular dataflow, and dynamic precision requirements of such hybrid models. To address these challenges, this paper proposes a dynamically reconfigurable field-programmable gate array (FPGA)-based accelerator tailored for parallel CNN-LSTM computation. The proposed architecture integrates a mixed-precision computation array, multilevel reconfigurable processing elements, and a triple-mode dataflow controller supporting weight-stationary/output-stationary/row-stationary dataflow, thereby enabling adaptive resource allocation and enhanced data reuse under diverse computation patterns. The accelerator is designed to efficiently execute both individual and hybrid CNN-LSTM workloads. Experimental evaluation on a representative ConvLSTM-based sea surface temperature prediction task demonstrates that the proposed design achieves high throughput and energy efficiency in both convolutional and recurrent computation phases.

The rapid obsolescence of electronic products highlights the critical need for efficient resource recovery through disassembly to support sustainable development. However, conventional disassembly methods often fail to handle complex disassembly sequences and dynamic operational constraints. To address the challenge of disassembly line balancing under worker fatigue, this study proposes an attention-based double deep Transformer Q-network (DDTQN) for minimizing disassembly time. It develops a disassembly time optimization model that incorporates fatigue-induced efficiency decay to enable the simulation of realistic operational conditions. By integrating an attention mechanism into the DDQ framework, the proposed approach enhances the capacity of the model to capture intricate task dependencies, thereby improving state representation, exploration efficiency, and long-term decision-making. Experimental results across three disassembly cases indicate that DDTQN reduces the average disassembly time by 19.37% compared with benchmark algorithms—including DDQN, deep Q-network (DQN), and advantage actor-critic. The successful application of DDTQN to marine equipment disassembly demonstrates its broad applicability and effectiveness, offering a robust solution for both general disassembly lines and specialized contexts such as ship recycling.