Sea-ground-air-space integrated ocean intelligent collaborative three-dimensional networking observation method

Abstract

The invention discloses a sea-ground-air-space integrated marine intelligent collaborative three-dimensional networking observation method, which belongs to the technical field of marine environment observation, is used for intelligent collaborative three-dimensional networking observation, and comprises the following steps of: generating an optimal navigation track covering a target area of dynamic characteristics by using a multi-target optimization algorithm; each observation platform executes a measurement task according to the received collaborative observation instruction, and each observation platform carries an embedded processor to analyze the collected temperature, salinity and flow field data characteristics in real time; and the data processing center completes data assimilation and ocean mode simulation by using supercomputing parallel calculation, and generates and outputs high-resolution three-dimensional temperature, salinity and flow field analysis results of dynamic characteristics. The coverage rate and the quality of observation information are obviously improved; the observation precision and efficiency are obviously improved, and optimal configuration and high-precision observation of resources are realized; a finer three-dimensional structure is captured with less navigation mileage, data redundancy and energy consumption are reduced, and observation efficiency is improved.

Description

Technical Field

[0001] This invention discloses an integrated marine intelligent collaborative three-dimensional network observation method that combines sea, land, air, and space, belonging to the field of marine environmental observation technology. Background Technology

[0002] Marine environmental dynamics are characterized by multi-scale, sudden, and highly nonlinear features. For example, mesoscale eddies carry the vast majority of ocean kinetic energy and have a significant impact on climate change and the marine ecological environment. Achieving high spatiotemporal resolution three-dimensional observation of key processes such as mesoscale eddies, internal ocean waves, and typhoons is an urgent need for marine scientific research and marine resource management. Currently, marine observation mainly relies on various individual or simple combinations of platforms, including sea-based observation platforms, shore-based radar, airborne unmanned aerial vehicles (UAVs), and space-based remote sensing satellites. Sea-based observation platforms use fixed platforms (such as underwater moorings, buoys, and seabed base stations deployed in the sea area), which can provide long-term continuous, full-depth in-situ data. However, their observation locations are fixed, their coverage is limited, and once deployed, they cannot be moved, resulting in insufficient responsiveness to dynamic processes. Mobile observation platforms (such as underwater gliders, autonomous underwater vehicles, and unmanned surface vessels) have flexible cruising capabilities and can conduct regional surveys, making them an effective means to compensate for the shortcomings of fixed platforms. Shore-based radar (such as high-frequency ground wave radar) can monitor sea surface current fields in real time over a range of tens to hundreds of kilometers near the coast, but its coverage is limited and cannot cover deep-sea areas. Airborne drones offer high-resolution sea surface observations and flexible deployment, but are limited by payload and endurance. Drones typically carry only lightweight sensors and have limited battery life, making sustained long-duration flights difficult. Space-based remote sensing satellites can perform large-scale sea surface scans, acquiring data such as sea surface height and temperature, but observations are limited to the sea surface and cannot obtain information about the ocean's interior; furthermore, they are affected by satellite orbital periods and weather conditions (such as cloud cover), resulting in blind spots and time delays.

[0003] Currently, technicians have recognized the importance of constructing an integrated air-space-ground-ocean observation network. However, existing network observation methods mostly remain at the level of simple function overlay or post-data fusion, lacking real-time and intelligent collaborative mechanisms. Typical problems include: delayed observation response—relying on post-event analysis to organize cruises, making it difficult to capture rapidly evolving processes in a timely manner; poor coordination between platforms—observation mission design mainly relies on human experience, lacking online scheduling and adaptive path replanning based on remote sensor information; and low utilization efficiency of observation resources—lacking intelligent mission planning, mobile platforms often use redundant or inefficient paths, resulting in insufficiently detailed three-dimensional structural characterization of key processes. Therefore, there is an urgent need for an intelligent, three-dimensional network observation method that can proactively sense sudden ocean processes and schedule multi-platform collaborative observations to meet the real-time and efficient observation needs of complex regional ocean dynamic processes. Summary of the Invention

[0004] The purpose of this invention is to provide an integrated marine intelligent collaborative three-dimensional network observation method that combines land, sea, air and space to solve the problem of poor collaborative effect in the combination of multiple marine observation platforms in the existing technology.

[0005] A method for integrated marine intelligent collaborative three-dimensional network observation encompassing sea, land, air, and space includes: S1. By comprehensively utilizing multi-source observation data, a GPU-based deep convolutional neural network model is deployed in the data processing center to process the raw data in real time, identify potential dynamic features, and extract the center position, radius, and intensity of the dynamic features. S2. Combine the real-time status information of each observation platform to generate a navigation trajectory covering the target area with dynamic characteristics, and issue collaborative observation commands to each observation platform according to the optimal navigation trajectory; S3. The data processing center uses supercomputing parallel computing to complete data assimilation based on a four-dimensional ensemble variational hybrid assimilation scheme and computational simulation based on the ROMS high-resolution ocean numerical model, generating and outputting high-resolution three-dimensional temperature, salinity and flow field analysis results of dynamic characteristics. S4. Analyze the improvement results of temperature, salinity and flow field data using information entropy, mean square error reduction or short-term forecast error reduction to obtain an effectiveness score. Input the effectiveness score and collaborative observation instructions into the multi-agent reinforcement learning model for training to optimize subsequent collaborative observation instructions.

[0006] The multi-source observation data includes observation data from buoys, underwater gliders, space-based remote sensing satellites, drones, and shore-based radar.

[0007] S2 includes generating a navigation trajectory covering the target region with dynamic characteristics using a multi-objective optimization algorithm based on a fusion framework of non-dominated sorting genetic algorithm III and potential field method, selecting the solution that satisfies the constraints and has the highest comprehensive score as the optimal navigation trajectory based on the objective function, and issuing collaborative observation commands to each observation platform according to the optimal navigation trajectory.

[0008] Each observation platform is equipped with an embedded processor to analyze the characteristics of the collected temperature, salinity, and flow field data in real time; When each observation platform executes measurement tasks according to the received collaborative observation instructions, if a sudden change in temperature-salinity gradient or flow velocity exceeding a preset threshold is detected during the measurement process, the observation platform will perform localized intensified observation.

[0009] The deep convolutional neural network model adopts a two-stage fusion architecture, including a physical constraint pre-screening layer and a multi-source feature fusion detection layer. The deep convolutional neural network model outputs the category confidence and bounding box coordinates of the dynamic features in real time, and uses the velocity gradient method, sea surface height gradient method, and eddy gradient method to determine the center position, radius, and intensity of the dynamic features.

[0010] The physical constraint pre-screening layer uses an improved Okubo–Weiss parameter method, combining current field data obtained from underwater moorings, shore-based radar, and buoys with sea surface height anomalies and sea surface temperature data obtained from space-based remote sensing satellites for preprocessing.

[0011] The multi-source feature fusion detection layer inputs dynamic features into a lightweight single-stage detection network based on the YOLO architecture. The backbone network of the lightweight single-stage detection network based on the YOLO architecture embeds a physical information attention module. The spatial branch inputs dynamic features to generate a weight map. The channel branch assigns feature channel weights according to the vortex, internal wave, and typhoon categories. The loss function incorporates parameter constraint terms.

[0012] The raw data includes sea surface height, three-dimensional ocean temperature, and current field data; The dynamic characteristics include mesoscale vortices, internal waves, and typhoons; The observation platform includes underwater moorings, buoys, underwater gliders, unmanned surface vessels, ground wave radar, unmanned aerial vehicles, and marine remote sensing satellites.

[0013] When the number of observation platforms in the collaborative observation command is ≤10, the multi-agent reinforcement learning model adopts an efficiency-enhanced deep Q-network; When the number of observation platforms in the collaborative observation command is greater than 10, the multi-agent reinforcement learning model adopts a multi-agent PPO-performance feedback architecture.

[0014] The real-time status information of each observation platform includes location, motion attitude, battery level, and sensor status.

[0015] Compared to existing technologies, this invention offers the following advantages: Proactive sensing enhances observation timeliness, automatically detecting and locking onto rapidly evolving ocean processes, achieving second-level monitoring of target events and optimal resource allocation, significantly improving observation timeliness. Platform collaboration efficiency is multiplied, significantly improving the coverage and quality of observational information; observation accuracy and efficiency are significantly improved, achieving optimal resource allocation and high-precision observation; capturing more detailed three-dimensional structures with less navigation distance reduces data redundancy and energy consumption, improving observation efficiency. A complete closed-loop and self-evolving capability continuously optimizes collaborative strategies, maintaining the long-term vitality and adaptability of the technical solution. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0017] A method for integrated marine intelligent collaborative three-dimensional network observation encompassing sea, land, air, and space includes: S1. By comprehensively utilizing multi-source observation data, a GPU-based deep convolutional neural network model is deployed in the data processing center to process the raw data in real time, identify potential dynamic features, and extract the center position, radius, and intensity of the dynamic features. S2. Combine the real-time status information of each observation platform to generate a navigation trajectory covering the target area with dynamic characteristics, and issue collaborative observation commands to each observation platform according to the optimal navigation trajectory; S3. The data processing center uses supercomputing parallel computing to complete data assimilation based on a four-dimensional ensemble variational hybrid assimilation scheme and computational simulation based on the ROMS high-resolution ocean numerical model, generating and outputting high-resolution three-dimensional temperature, salinity and flow field analysis results of dynamic characteristics. S4. Analyze the improvement results of temperature, salinity and flow field data using information entropy, mean square error reduction or short-term forecast error reduction to obtain an effectiveness score. Input the effectiveness score and collaborative observation instructions into the multi-agent reinforcement learning model for training to optimize subsequent collaborative observation instructions.

[0018] The multi-source observation data includes observation data from buoys, underwater gliders, space-based remote sensing satellites, drones, and shore-based radar.

[0019] S2 includes generating a navigation trajectory covering the target region with dynamic characteristics using a multi-objective optimization algorithm based on a fusion framework of non-dominated sorting genetic algorithm III and potential field method, selecting the solution that satisfies the constraints and has the highest comprehensive score as the optimal navigation trajectory based on the objective function, and issuing collaborative observation commands to each observation platform according to the optimal navigation trajectory.

[0020] Each observation platform is equipped with an embedded processor to analyze the characteristics of the collected temperature, salinity, and flow field data in real time; When each observation platform executes measurement tasks according to the received collaborative observation instructions, if a sudden change in temperature-salinity gradient or flow velocity exceeding a preset threshold is detected during the measurement process, the observation platform will perform localized intensified observation.

[0021] The deep convolutional neural network model adopts a two-stage fusion architecture, including a physical constraint pre-screening layer and a multi-source feature fusion detection layer. The deep convolutional neural network model outputs the category confidence and bounding box coordinates of the dynamic features in real time, and uses the velocity gradient method, sea surface height gradient method, and eddy gradient method to determine the center position, radius, and intensity of the dynamic features.

[0022] The physical constraint pre-screening layer uses an improved Okubo–Weiss parameter method, combining current field data obtained from underwater moorings, shore-based radar, and buoys with sea surface height anomalies and sea surface temperature data obtained from space-based remote sensing satellites for preprocessing.

[0023] The multi-source feature fusion detection layer inputs dynamic features into a lightweight single-stage detection network based on the YOLO architecture. The backbone network of the lightweight single-stage detection network based on the YOLO architecture embeds a physical information attention module. The spatial branch inputs dynamic features to generate a weight map. The channel branch assigns feature channel weights according to the vortex, internal wave, and typhoon categories. The loss function incorporates parameter constraint terms.

[0024] The raw data includes sea surface height, three-dimensional ocean temperature, and current field data; The dynamic characteristics include mesoscale vortices, internal waves, and typhoons; The observation platform includes underwater moorings, buoys, underwater gliders, unmanned surface vessels, ground wave radar, unmanned aerial vehicles, and marine remote sensing satellites.

[0025] When the number of observation platforms in the collaborative observation command is ≤10, the multi-agent reinforcement learning model adopts an efficiency-enhanced deep Q-network; When the number of observation platforms in the collaborative observation command is greater than 10, the multi-agent reinforcement learning model adopts a multi-agent PPO-performance feedback architecture.

[0026] The real-time status information of each observation platform includes location, motion attitude, battery level, and sensor status.

[0027] The specific implementation process of the embodiments of the present invention is described below.

[0028] T1. By comprehensively utilizing three types of observation data—space-based remote sensing satellites, sea and air drones, and shore-based radar—a GPU-based deep convolutional neural network model is deployed in the data processing center to process the raw data in real time, identify potential dynamic features, and extract the center position, radius, and intensity of the dynamic features. T2. Combining the real-time status information of each observation platform, use a multi-objective optimization algorithm to generate the optimal navigation trajectory covering the target area for dynamic characteristics, and issue collaborative observation commands to each observation platform according to the optimal navigation trajectory; T3. Each observation platform executes measurement tasks according to the received collaborative observation instructions. Each observation platform is equipped with an embedded processor to analyze the characteristics of the collected temperature, salinity and flow field data in real time. If a sudden change in temperature-salinity gradient or flow velocity exceeding the preset threshold is detected during the measurement process, the observation platform performs local densified observation. T4. The data processing center uses supercomputing parallel computing to complete data assimilation and ocean model simulation, generating and outputting high-resolution three-dimensional temperature, salinity and flow field analysis results of dynamic characteristics; T5. Analyze the improvement results of temperature, salinity and flow field data using information entropy, mean square error reduction, or short-term forecast error reduction to obtain an effectiveness score. Input the effectiveness score and collaborative observation instructions into a multi-agent reinforcement learning model for training to optimize subsequent collaborative observation instructions.

[0029] T2 includes scheme generation, system querying of platform status databases, and generation of optimal navigation trajectories covering the target area using customized path planning algorithms. For example, circular or radial densified path planning algorithms are applied to underwater glider swarms to efficiently describe the three-dimensional structure of vortices; vertical profile route strategies are used for internal wave forecasting. These algorithms ensure observation coverage while minimizing travel time and achieving optimal resource allocation. Coordinated command is then issued to each observation platform according to the established scheme. For underwater glider swarms, commands are issued to navigate along predetermined routes such as "radial ray densification" or "concentric stacked rings" and conduct densified profile observations; for fixed platforms such as moorings and buoys, commands are issued to switch to event-driven high-frequency sampling modes (e.g., adjusting the CTD sampling interval from the usual 1 time / hour to 1 time / minute) in response to the arrival of the target process. All commands are transmitted in real time via satellite links or sea-land wireless networks to ensure coordinated execution by all platforms.

[0030] T3 comprises observation platforms that automatically execute measurement tasks according to received instructions. The platforms possess edge computing capabilities (such as embedded processors for real-time analysis of the characteristics of acquired temperature, salinity, and flow data). During observation, if a temperature-salinity gradient exceeding a preset threshold or a sudden change in flow velocity is detected (e.g., a temperature gradient greater than 0.5 degrees Celsius per 100 meters), the platform can autonomously adjust its flight path or perform locally intensified observations to finely characterize the key process structures. All acquired data is transmitted to the data center in real time via low-Earth orbit satellites, with a communication link latency of less than 50ms, achieving data feedback at the minute or even second level.

[0031] T4 comprises a data center that receives heterogeneous observational data from multiple platforms across sea, shore, air, and space, and performs quality control and fusion on data such as temperature, salinity, current velocity, and sea surface height. The system utilizes high-performance computing clusters (e.g., equipped with multi-core CPUs and GPUs) to run high-resolution regional oceanographic numerical models and employs data assimilation techniques (such as ensemble Kalman filtering or variational assimilation) to integrate observational data into model analysis. These calculations are performed on a parallel computing platform, enabling the rapid output of high-resolution three-dimensional environmental field products (e.g., temperature, salinity, and current fields with 1km resolution) for the target process area within 4 to 6 hours, with significantly higher accuracy than products from a single data source.

[0032] T5 involves a comprehensive evaluation of the improvements in observation results and numerical model predictions after each collaborative observation task. Specifically, it defines observation performance metrics such as information entropy, reduction in mean squared error, or decrease in short-term forecast error, and calculates performance scores based on the forecast improvements before and after data assimilation. The observation strategy (path scheme, platform configuration, etc.) used in the current task and its corresponding performance score are input into a multi-agent reinforcement learning model for training. A performance feedback closed-loop training architecture is employed: when the number of observation platforms in the collaborative task is ≤10, an performance-enhanced deep Q-network (EE-DQN) is used; when the number of observation platforms in the collaborative task is >10, a multi-agent PPO-performance feedback architecture (MAPPO-EF) is used. Through this associative learning mechanism, the selection of collaborative schemes for future tasks is optimized, enabling adaptive evolution and continuous improvement of the observation system.

[0033] The example illustrates intelligent collaborative observation of a mesoscale cold vortex. The system involved in this example includes: 1) Space-based observation subsystem, including passing ocean remote sensing satellites, provides data such as sea surface height anomaly (SSHA) and sea surface temperature (SST); 2) Sea-based observation subsystem, including underwater mooring systems deployed in the sea area and a cluster of underwater gliders on standby (e.g., 3 gliders equipped with temperature and salinity probes, inertial navigation systems and satellite communication modules). 3) Data processing and control center, equipped with high-performance computing servers (such as multi-core CPUs and GPU cards), and deployed software modules such as vortex detection, task planning, and data assimilation; 4) Communication network, low-Earth orbit satellite communication links, used to transmit commands and data between the data center and various observation platforms.

[0034] Examples of parameters in this embodiment include: glider group speed of approximately 0.4 m / s and endurance of approximately 30 days; satellite communication latency of <50 ms; data assimilation adopts a four-dimensional ensemble variational hybrid assimilation scheme, with a horizontal grid of approximately 1 / 10° (approximately 10 km) and a vertical resolution of approximately 50 m in the regional model.

[0035] During T1 execution, the data center receives and processes real-time SSH and SST data provided by the space-based subsystem, and runs a pre-trained deep learning vortex detection model and an improved Okubo-Weiss algorithm on the GPU cluster.

[0036] A modified Okubo-Weiss (OW) isoparametric method is used for preprocessing, the improvement being the introduction of an adaptive threshold mechanism: ; In the formula, α and β are adaptive coefficients for sea area characteristics. The standard deviation of the absolute value of the OW parameter is used to mark the region that satisfies |OW|>λ as a candidate region for dynamic features, and the ratio γ of the target sea area vorticity ζ to strain rate s is calculated as the physical feature vector. The multi-source feature fusion detection layer inputs the physical feature data of the target sea area into a lightweight YOLOv5s network, which has been improved as follows: the backbone network embeds a Physical Information Attention (PIAM) module; the spatial branch inputs γ to generate a weight map; and the channel branch assigns feature channel weights according to the vortex / internal wave / typhoon category. An OW consistency constraint term is added to the loss function. Through GPU parallel acceleration, the class confidence and bounding box coordinates of the dynamic features are output in real time.

[0037] The system automatically identified a developing mesoscale cold vortex with a diameter of about 150 km near 19°N and 118°E, and extracted the center location, radius and intensity information of the vortex to generate a "vortex densification observation" mission instruction.

[0038] During T2 execution, the mission planning system, once triggered, performs the following operations: The system queries the platform status database, confirming that the three gliders (Glider_01, Glider_02, and Glider_03) are located approximately 80 km southeast of the vortex and have sufficient power. It also identifies a mooring (Mooring_A) on the predicted path of the vortex. In the scheme generation phase, the system evaluates various observation schemes based on the vortex characteristics, ultimately selecting the "radial ray densification" scheme. An optimization algorithm is used to generate three radial trajectories traversing the vortex center. During coordinated scheduling, the system sends instructions to the underwater glider group: Glider_01, Glider_02, and Glider_03 are instructed to depart from the east, south, and west directions respectively, traversing the vortex center along the aforementioned radial paths, and in the core area, to densify the profile measurement interval from the original 10 km to 5 km. The system also sends a command to the mooring (Mooring_A) to adjust the sampling interval of its CTD probe from the usual 1 time / hour to 1 time / 10 minutes in order to capture the drastic fluctuations in temperature and salinity as the vortex passes by.

[0039] During T3, three underwater gliders received instructions and automatically navigated to the target area, commencing intensive profile observations. During the observation process, as Glider_02 crossed the eddy boundary, its edge computing module detected in real-time that the vertical temperature gradient exceeded a set threshold (e.g., 0.5 degrees Celsius / 100 meters), indicating it was approaching a water temperature front. It then autonomously decided to perform an additional "zigzag" small-range scan to perform detailed measurements of the frontal structure. All temperature, salinity, and depth data collected by the underwater gliders and moorings were transmitted back to the data center in real-time via a low-Earth orbit satellite link, with communication latency typically below 100ms, achieving near real-time data feedback.

[0040] During T4 execution, the data center received on-site profile data transmitted from underwater gliders and moorings, and input it along with raw satellite remote sensing data into the data assimilation system. Based on ensemble Kalman filtering, the system assimilated the data from each platform into a high-resolution regional ocean model. Utilizing supercomputing parallel computing, data fusion and model simulation were completed in approximately 6 hours, generating and outputting high-resolution three-dimensional temperature, salinity, and flow field analysis products for this mesoscale cold eddy. This invention clearly reveals the thermal structure and circulation characteristics of the eddy from the sea surface to a depth of 1000m, with an accuracy far exceeding that of conventional model outputs based on single-platform data.

[0041] During T5 execution, the system automatically initiates a performance evaluation: calculating the information entropy of the data collected in this observation and the degree to which the model reduces the error in predicting the eddy position for the next 24 hours after assimilating this data, which is used as the "performance score" for this collaborative observation. This score, along with the "radial ray densification" strategy adopted in this study, is input into a deep reinforcement learning model (such as dual-depth DQN). Through continuous iterative training, the model learns to select the optimal observation strategy under similar sea states and eddy characteristics. After multiple mission iterations, the observation system can automatically optimize the collaborative scheme, achieving self-evolution and continuous improvement of the observation strategy.

[0042] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for integrated marine intelligent collaborative three-dimensional network observation combining sea, land, air, and space, characterized in that: include: S1. By comprehensively utilizing multi-source observation data, a GPU-based deep convolutional neural network model is deployed in the data processing center to process the raw data in real time, identify potential dynamic features, and extract the center position, radius, and intensity of the dynamic features. S2. Combine the real-time status information of each observation platform to generate a navigation trajectory covering the target area with dynamic characteristics, and issue collaborative observation commands to each observation platform according to the optimal navigation trajectory; S3. The data processing center uses supercomputing parallel computing to complete data assimilation based on a four-dimensional ensemble variational hybrid assimilation scheme and computational simulation based on the ROMS high-resolution ocean numerical model, generating and outputting high-resolution three-dimensional temperature, salinity and flow field analysis results of dynamic characteristics. S4. Analyze the improvement results of temperature, salinity and flow field data using information entropy, mean square error reduction or short-term forecast error reduction to obtain an effectiveness score. Input the effectiveness score and collaborative observation instructions into the multi-agent reinforcement learning model for training to optimize subsequent collaborative observation instructions.

2. The integrated marine intelligent collaborative three-dimensional network observation method according to claim 1, characterized in that, The multi-source observation data includes observation data from buoys, underwater gliders, space-based remote sensing satellites, drones, and shore-based radar.

3. The integrated marine intelligent collaborative three-dimensional network observation method according to claim 1, characterized in that, S2 includes generating a navigation trajectory covering the target region with dynamic characteristics using a multi-objective optimization algorithm based on a fusion framework of non-dominated sorting genetic algorithm III and potential field method, selecting the solution that satisfies the constraints and has the highest comprehensive score as the optimal navigation trajectory based on the objective function, and issuing collaborative observation commands to each observation platform according to the optimal navigation trajectory.

4. The integrated marine intelligent collaborative three-dimensional network observation method according to claim 1, characterized in that, Each observation platform is equipped with an embedded processor to analyze the characteristics of the collected temperature, salinity, and flow field data in real time; When each observation platform executes measurement tasks according to the received collaborative observation instructions, if a sudden change in temperature-salinity gradient or flow velocity exceeding a preset threshold is detected during the measurement process, the observation platform will perform localized intensified observation.

5. The integrated marine intelligent collaborative three-dimensional network observation method according to claim 1, characterized in that, The deep convolutional neural network model adopts a two-stage fusion architecture, including a physical constraint pre-screening layer and a multi-source feature fusion detection layer. The deep convolutional neural network model outputs the category confidence and bounding box coordinates of the dynamic features in real time, and uses the velocity gradient method, sea surface height gradient method, and eddy gradient method to determine the center position, radius, and intensity of the dynamic features.

6. The integrated marine intelligent collaborative three-dimensional network observation method according to claim 5, characterized in that, The physical constraint pre-screening layer uses an improved Okubo–Weiss parameter method, combining current field data obtained from underwater moorings, shore-based radar, and buoys with sea surface height anomalies and sea surface temperature data obtained from space-based remote sensing satellites for preprocessing.

7. The integrated marine intelligent collaborative three-dimensional network observation method according to claim 5, characterized in that, The multi-source feature fusion detection layer inputs dynamic features into a lightweight single-stage detection network based on the YOLO architecture. The backbone network of the lightweight single-stage detection network based on the YOLO architecture embeds a physical information attention module. The spatial branch inputs dynamic features to generate a weight map. The channel branch assigns feature channel weights according to vortex, internal wave, and typhoon categories. The loss function incorporates parameter constraint terms.

8. The integrated marine intelligent collaborative three-dimensional network observation method according to claim 1, characterized in that, The raw data includes sea surface height, three-dimensional ocean temperature, and current field data; The dynamic characteristics include mesoscale vortices, internal waves, and typhoons; The observation platform includes underwater moorings, buoys, underwater gliders, unmanned surface vessels, ground wave radar, unmanned aerial vehicles, and marine remote sensing satellites.

9. The integrated marine intelligent collaborative three-dimensional network observation method according to claim 1, characterized in that, When the number of observation platforms in the collaborative observation command is ≤10, the multi-agent reinforcement learning model adopts an efficiency-enhanced deep Q-network; When the number of observation platforms in the collaborative observation command is greater than 10, the multi-agent reinforcement learning model adopts a multi-agent PPO-performance feedback architecture.

10. The integrated marine intelligent collaborative three-dimensional network observation method according to claim 1, characterized in that, The real-time status information of each observation platform includes location, motion attitude, battery level, and sensor status.