The operation of underwater vehicles in deep waters is a very challenging task. The use of AUVs (Autonomous Underwater Vehicles) is the preferred option for underwater exploration activities. They can be autonomously navigated and controlled in real time underwater, which is only possible with precise spatio-temporal information. Navigation and positioning systems based on LBL (Long-Baseline) or USBL (Ultra-Short-Baseline) systems have their own characteristics, so the choice of system is based on the specific application scenario. However, comparative experiments on AUV navigation and positioning under both systems are rarely conducted, especially in the deep sea. This study describes navigation and positioning experiments on AUVs in deep-sea scenarios and compares the accuracy of the USBL and LBL/SINS (Strap-Down Inertial Navigation System)/DVL (Doppler Velocity Log) modes. In practice, the accuracy of the USBL positioning mode is higher when the AUV is within a 60° observation range below the ship; when the AUV is far away from the ship, the positioning accuracy decreases with increasing range and observation angle, i.e., the positioning error reaches 80 m at 4000 m depth. The navigational accuracy inside and outside the datum array is high when using the LBL/SINS/DVL mode; if the AUV is far from the datum array when climbing to the surface, the LBL cannot provide accurate position calibration while the DVL fails, resulting in large deviations in the SINS results. In summary, the use of multi-sensor combination navigation schemes is beneficial, and accurate position information acquisition should be based on the demand and cost, while other factors should also be comprehensively considered; this paper proposes the use of the LBL/SINS/DVL system scheme.

Detecting deep-sea megabenthic organisms is of foremost importance for seabed resource surveys, typical habitat protection, and biodiversity surveys. However, the complexity of the deep-sea environment, uneven illumination, and small biological targets that are easily obscured all increase target detection difficulty significantly. To address these, this paper proposes a deep-sea megabenthic detection algorithm, DS-YOLO, based on YOLOv5s. To improve the detection ability of the model for deep-sea megabenthic organisms, the space-to-depth module and the spatial pyramid pooling cross stage partial channel module are introduced in the Backbone layer to enlarge the receptive field and enhance the retention of small-scale features. Then, the space-to-depth and normalization-based attention modules and the Add and Concat functions of the bidirectional feature pyramid network are introduced in the Neck layer to increase the multiscale fusion ability of the model and highlight the insignificant features. Finally, the two branches of the decoupling header output the category and location of the target, which causes the model to utilize the feature information to the maximum extent. Experiments showed that DS-YOLO improved mAP0.5 from 89.6% to 92.4% and mAP0.5:0.95 from 65.7% to 72.3% compared to the original YOLOv5s on the homemade dataset and outperformed other algorithms in the YOLO series. DS-YOLO reaches 84.7 FPS for deployment on mobile platforms. In addition, the combined DS-YOLO and DeepSORT algorithm can be used to calculate the abundance and community structure of deep-sea megabenthos. The model outperforms general target detection models for deep-sea megabenthos detection and is suitable for use in complex deep-sea environments.

The western Pacific seamount area is abundant in both biological and mineral resources, making it a crucial location for international investigation of regional seabed resources. An essential stage in comprehending and advancing seamounts is gaining knowledge about the distribution characteristics and laws governing the seabed substrate. Deep-sea geological sampling is challenging because of the intricate nature of the deep-sea environment, resulting in increased difficulty in identifying and evaluating substrates. This study addresses the aforementioned issues by utilizing in-situ video footage obtained from the “Jiaolong” manned deep submersible and shipborne deep-water multibeam data. This data is used as a foundation for constructing a Western Pacific seamount areas substrate classification point set. Additionally, the paper introduces the mRMR-XGBoost substrate classification model. Substrate categorization in deep sea and mountainous regions has been successfully accomplished, yielding a classification accuracy of 92.5%. The classification experiments and box sampling results demonstrate that the mRMR-XGBoost substrate classification model proposed in this paper can efficiently use acoustic and optical data to accurately divide the substrate types in seamount areas, with better classification accuracy, when compared with commonly used machine learning models. It has a significant application value and the best classification effect on the two types of substrates: nodules and gravel substrates.

Ensuring the safe landing of deep-sea mining vehicles (DSMVs) on soft seabed sediments is critical for the stability and operational reliability of subsea mineral extraction. However, deep-sea sediments, particularly in polymetallic nodule regions, are characterized by low shear strength, high compressibility, and rate-dependent behavior, posing significant challenges for full-scale experimental investigation and predictive modeling. To address these limitations, this study develops a high-fidelity finite element simulation framework based on the Coupled Eulerian–Lagrangian (CEL) method to model the landing and penetration process of full-scale DSMVs under various geotechnical conditions. To overcome the high computational cost of FEM simulations, a data-driven surrogate model using the random forest algorithm is constructed to predict the normalized penetration depth based on key soil and operational parameters. The proposed hybrid FEM–ML approach enables efficient multiparameter analysis and provides actionable insights into the complex soil–structure interactions involved in DSMV landings. This methodology offers a practical foundation for engineering design, safety assessment, and descent planning in deep-sea mining operations.

In this paper, a method for extracting the on-the-fly (OTF) GNSS tide level was designed by combining variational modal decomposition (VMD) and a wavelet thresholding (WT) method to improve the extraction accuracy of the OTF GNSS tide level. First, the energy difference ratio method was used to determine the number of layers for the VMD. Subsequently, the VMD performed a second decomposition of the IMF1 obtained from the first VMD to achieve an efficient separation of signal and noise. The normalized cross-correlation coefficient (NCC) was applied to determine the number of layers for the WT method. Finally, experimental results showed that the VMD–WT method outperformed the other seven filtering methods in three metrics: maximum error, the root-mean-square error (RMSE), and error distribution. Therefore, the VMD–WT method was able to extract extremely accurate on-the-fly GNSS tide level and additionally obtain more accurate bathymetry data after tidal correction of the bathymetry data.