This paper explores the application of Deep Reinforcement Learning (DRL) to optimize adaptive observation strategies for multi-AUV systems in complex marine environments. Traditional algorithms struggle with the strong coupling between environmental information and observation modeling, making it challenging to derive optimal strategies. To address this, we designed a DRL framework based on the Dueling Double Deep Q-Network (D3QN), enabling AUVs to interact directly with the environment for more efficient 3D dynamic ocean observation. However, traditional D3QN faces slow convergence and weak action–decision correlation in partially observable, dynamic marine settings. To overcome these challenges, we integrate a Gated Recurrent Unit (GRU) into the D3QN, improving state-space prediction and accelerating reward convergence. This enhancement allows AUVs to optimize observations, leverage ocean currents, and navigate obstacles while minimizing energy consumption. Experimental results demonstrate that the proposed approach excels in safety, energy efficiency, and observation effectiveness. Additionally, experiments with three, five, and seven AUVs reveal that while increasing platform numbers enhances predictive accuracy, the benefits diminish with additional units.

Unmanned surface vehicles (USVs) rely on multi-sensor perception, such as radar, LiDAR, GPS, and vision, to ensure safe and efficient navigation in complex maritime environments. Traditional ant colony optimization (ACO) for path planning, however, suffers from premature convergence, slow adaptation, and poor smoothness in cluttered waters, while the dynamic window approach (DWA) without global guidance can become trapped in local obstacle configurations. This paper presents a sensor-oriented hybrid method that couples an improved ACO for global route planning with an enhanced DWA for local, real-time obstacle avoidance. In the global stage, the ACO state–transition rule integrates path length, obstacle clearance, and trajectory smoothness heuristics, while a cosine-annealed schedule adaptively balances exploration and exploitation. Pheromone updating combines local and global mechanisms under bounded limits, with a stagnation detector to restore diversity. In the local stage, the DWA cost function is redesigned under USV kinematics to integrate velocity adaptability, trajectory smoothness, and goal-deviation, using obstacle data that would typically originate from onboard sensors. Simulation studies, where obstacle maps emulate sensor-detected environments, show that the proposed method achieves shorter paths, faster convergence, smoother trajectories, larger safety margins, and higher success rates against dynamic obstacles compared with standalone ACO or DWA. These results demonstrate the method’s potential for sensor-based, real-time USV navigation and collision avoidance in complex maritime scenarios.

Marine physical data are currently derived mainly from satellite remote sensing, sparsely distributed buoys, and mobile observation platforms. The buoys and mobile observation platforms are more suitable for acquiring deep ocean data. Traditional methods for dealing with such observations rely on linear and Gaussian assumptions in the assimilation of nonlinear marine characteristics, which may introduce bias in analysis and forecasting. The particle-filter assimilation method is of increasing interest because it has advantages in dealing with nonlinear assimilation problems, although it has limitations with sparsely distributed observations. To address this problem, we introduced inverse distance-weighted interpolation into the localization scheme of the localized equivalent-weights particle filter with time-distributed statistical observations method based on data from sparsely distributed buoys and mobile platforms. This improved the utilization rate of observations and increased forecast accuracy. Theoretical experiments were undertaken to highlight the characteristics of the improved method, using reanalysis data from the European Centre for Medium-Range Weather Forecasts as evaluation criteria in comparing the method with the localized weighted ensemble Kalman filter method—a hybrid particle-filter method. Results indicate that the method effectively improves assimilation and the accuracy of state estimation and forecasts of ocean temperature and salinity. Significance Statement: This paper addresses the problem of assimilating sparsely distributed observations in processing ocean elements. We improved the localization scheme to deal with the sparse-observation problem based on the localized equivalent-weights particle filter with time distribution statistical observation (LEWPF–T–Sobs). This method can effectively handle the assimilation problem with sparse observations while retaining the advantages of traditional particle filtering assimilation, thus improving the accuracy of estimates of physical ocean data.

Data assimilation has been widely applied in atmospheric and oceanic forecasting systems and particle filters (PFs) have unique advantages in dealing with nonlinear data assimilation. They have been applied to many scientific fields, but their application in geoscientific systems is limited because of their inefficiency in standard settings systems. To address these issues, this paper further refines the statistical observation and localization scheme which used in the classic localized equivalent-weights particle filter with statistical observation (LEWPF-Sobs). The improved method retains the advantages of equivalent-weights particle filter (EWPF) and the localized particle filter (LPF), while further refinements incorporate the effect of time series on the reanalyzed data into the statistical observation calculations, in addition to incorporating the statistical observation proposal density into the localization scheme to further improve the assimilation accuracy under sparse observation conditions. In order to better simulate the geoscientific system, we choose an intermediate atmosphere-ocean-land coupled model (COAL-IC) as the experimental model and divide the experiment into two parts: standard observation and sparse observation, which are analyzed by the spatial distribution results and root mean square error (RMSE) histogram. In order to better analyze the characteristics of the improved method, this method was chosen to be analyzed in comparison with the localized weighted ensemble Kalman filter (LWEnKF), the LPF and classical LEWPF-Sobs. From the experimental results, it can be seen that the improved method is better than the LWEnKF and LPF methods for various observation conditions. The improved method reduces the RMSE by about 7% under standard observation conditions compared to the traditional method, while the advantage of the improved method is even more obvious under sparse observation conditions, where the RMSE is reduced by about 85% compared to the traditional method. In particular, this improved filter not only combine the advantage of the two algorithms, but also overcome the computing resources.

Fine-scale detailed estimation of sea ice concentration (SIC) is pivotal for maritime safety, scientific exploration, and environmental surveillance. However, current datasets frequently present challenges due to their limited resolution, thereby hindering fine-scale analysis of sea ice conditions. This paper introduces a novel Deep Self-Attention Residual U-Net (Deep SARU-Net) architecture to address the limitations inherent in existing super-resolution estimation techniques. By harnessing distinctive multi-stage self-attention mechanisms, orthogonal rectangular convolutional kernels, and residual modules, this architecture significantly augments both the precision and generalizability of SIC super-resolution estimation tasks. Experimental results demonstrate that in the vicinity of the Chukchi Sea, the Deep SARU-Net method exhibits superior performance in terms of both RMSE and SSIM values compared to other models, showcasing its efficacy. Furthermore, generalization analyses across diverse sea regions confirm the model’s universality.