A fluid game ocean ecological three-dimensional monitoring method and system based on multi-source remote sensing

Abstract

The present application relates to the technical field of ecological monitoring, in particular to a fluid game marine ecological stereoscopic monitoring method and system based on multi-source remote sensing. The method comprises multi-source remote sensing manifold alignment and cross-medium dynamic characteristic reconstruction based on obtained multi-source heterogeneous data, including multi-source satellite orbit period time series harmonic normalization processing and manifold space coordinate construction based on sea surface dynamics potential energy, remote sensing spectral semantic extraction and sea surface dynamics characteristic fusion, and vertical projection and three-dimensional field reconstruction based on SQG theory; physical constraint space-time hypergraph construction based on quasi-geostrophic flow theory based on reconstructed information, including dynamic neighborhood search based on Lagrange coherent structure and quasi-geostrophic flow line hyperedge correlation matrix construction; the present application breaks through the vertical detection blind area caused by the "skin effect" of satellite remote sensing, and significantly reduces the inversion error of deep sea ecological parameters.

Description

Technical Field

[0001] This invention relates to the field of ecological monitoring technology, and in particular to a three-dimensional monitoring method and system for marine ecology based on multi-source remote sensing and fluid game theory. Background Technology

[0002] Utilizing satellite remote sensing technology to acquire marine ecological parameters, such as chlorophyll concentration, primary productivity, and sea surface temperature, has become a core means of constructing a marine monitoring system. While advancements in Earth observation technology have led to the accumulation of massive amounts of multi-source remote sensing data, existing remote sensing retrieval and assessment technologies still have significant limitations in meeting the demands of three-dimensional monitoring of complex deep-sea ecosystems, making it difficult to satisfy the needs of precise marine governance.

[0003] However, existing marine remote sensing technologies fall short in the face of these complex scenarios, primarily due to technical bottlenecks such as vertical blind spots caused by the "skin effect" and difficulties in the physical fusion of multi-source data. On one hand, satellite remote sensing is essentially a two-dimensional observation of the sea surface. Optical sensors can only detect spectral information within a single attenuation depth of the sea surface, and microwave sensors can only sense the surface layer at the micrometer to centimeter level. In contrast, material transport and energy exchange in marine ecological processes often occur in a three-dimensional space from the sea surface to the deep sea. Existing vertical extrapolation techniques mostly employ empirical regression models based on statistics, lacking physical modeling of hydrodynamic stratification mechanisms. This results in the inability to accurately invert the maximum chlorophyll characteristics below the thermocline, leading to a significant "semantic gap" in the monitoring results across vertical profiles. On the other hand, there is a severe "spatiotemporal heterogeneity" and "source conflict" between multi-source remote sensing data and numerical models. Ocean color satellites, altimeter satellites, and sparse underwater in-situ observations differ greatly in temporal frequency and spatial resolution. Traditional data assimilation methods often only achieve shallow interpolation and stitching, failing to achieve deep coupling of heterogeneous data within a hydrodynamic framework. Furthermore, when remote sensing inversion results and numerical simulation models contradict each other in local sea areas, the existing assessment system lacks an intelligent game mechanism that can simulate expert cognition to dynamically determine the true value. As a result, the final monitoring products are often simple weighted averages, which are difficult to eliminate systematic errors and lack physical consistency and credibility of the evaluation results.

[0004] Based on this, this invention proposes a three-dimensional marine ecological monitoring method and system based on a large-scale fluid game model driven by multi-source remote sensing. This method innovatively introduces Quasi-Geotransformation (SQG) theory as the core physical constraint, and utilizes spatiotemporal hypergraph computation technology to construct a cross-media remote sensing dynamic manifold space, solving the problem of physical extrapolation from two-dimensional sea surface data to three-dimensional deep-sea space. Simultaneously, it constructs a multi-agent double-blind game mechanism based on a large model to simulate the adversarial and negotiation process between observational data and physical mechanisms, achieving a consensus on ecological parameters that conform to objective laws under the guidance of Nash equilibrium theory. This invention aims to achieve comprehensive, high-precision, and dynamic monitoring of the marine ecological environment through the deep integration of underlying physics and artificial intelligence technologies. Summary of the Invention

[0005] To address the issues of missing deep ecological information and vertical extrapolation distortion caused by the "skin effect" of remote sensing, the physical separation between multi-source heterogeneous remote sensing data and sparse in-situ data in spatiotemporal fusion, and the conflict between multi-source observation and numerical simulation models under complex sea conditions and the lack of an objective arbitration mechanism, this invention provides a fluid game-based three-dimensional marine ecological monitoring method and system.

[0006] In a first aspect, the present invention provides a three-dimensional monitoring method for marine ecology based on multi-source remote sensing and fluid game theory, RS-FluidGameNet, which adopts the following technical solution: A three-dimensional monitoring method for marine ecology based on multi-source remote sensing and fluid game theory includes: Acquire multi-source heterogeneous data; Based on the acquired multi-source heterogeneous data, multi-source remote sensing manifold alignment and cross-medium dynamic feature reconstruction are carried out, including time-series harmonic normalization processing of multi-source satellite orbital periods and manifold spatial coordinate construction based on sea surface dynamic potential energy, remote sensing spectral semantic extraction and sea surface dynamic feature fusion, and vertical projection and three-dimensional field reconstruction based on SQG theory. Based on reconstructed information, a physically constrained spatiotemporal hypergraph based on quasi-geotransfer theory is constructed, including dynamic neighborhood search based on Lagrange coherent structure and correlation matrix construction of quasi-geotransfer streamline hyperedges; hyperedge feature aggregation that satisfies potential vorticity conservation and cross-scale fluid information interaction and node feature backhaul. The consensus on the truth value of ecological parameters generated by the remote sensing-mechanism double-blind game theory is based on the construction results, including the construction of agent utility function based on hypergraph confidence and the generation of initial ecological parameter bids driven by utility; the calculation of game adversarial loss based on bid difference and the strategy iteration and bid update guided by loss gradient; Ecological evolution prediction based on game results using Lagrange coherent structures includes Lagrange particle scattering and initialization based on true value flow fields and particle position state updates under Runge-Kutta integrals; biochemical reaction source-sink state updates along the trajectory and Eulerian grid remapping based on full-state particles; Output the prediction results.

[0007] Secondly, a three-dimensional marine ecological monitoring system based on multi-source remote sensing and fluid game theory includes: The data acquisition module is configured to acquire multi-source heterogeneous data; The reconstruction module is configured to perform multi-source remote sensing manifold alignment and cross-medium dynamic feature reconstruction based on the acquired multi-source heterogeneous data, including time-series harmonic normalization processing of multi-source satellite orbital periods and manifold spatial coordinate construction based on sea surface dynamic potential energy, remote sensing spectral semantic extraction and sea surface dynamic feature fusion, and vertical projection and three-dimensional field reconstruction based on SQG theory. The hypergraph module is configured to construct a physically constrained spatiotemporal hypergraph based on quasi-geotransformation theory, based on reconstructed information. This includes dynamic neighborhood search based on Lagrange coherence structure and construction of the correlation matrix of quasi-geotransformation streamline hyperedges; aggregation of hyperedge features that satisfy potential vorticity conservation; and cross-scale fluid information interaction and node feature backhaul. The game theory module is configured to generate a consensus on the truth value of ecological parameters based on remote sensing-mechanism double-blind game theory, including the construction of agent utility function based on hypergraph confidence and utility-driven initial ecological parameter bid generation; game adversarial loss calculation based on bid difference and loss gradient-guided strategy iteration and bid update. The prediction module is configured to predict ecological evolution based on game results using a Lagrange coherent structure; including Lagrange particle scattering and initialization based on the true value flow field and particle position state update under Runge-Kutta integral; biochemical reaction source and sink state update along the trajectory and Eulerian grid remapping based on full-state particles; The output module is configured to output the prediction results.

[0008] Thirdly, the present invention provides a computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the aforementioned method for three-dimensional monitoring of marine ecology based on multi-source remote sensing and fluid game theory.

[0009] Fourthly, the present invention provides a terminal device, including a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to provide a method for three-dimensional monitoring of marine ecology based on multi-source remote sensing and fluid game theory.

[0010] In summary, the present invention has the following beneficial technical effects: First, this invention overcomes the vertical detection blind spot caused by the "skin effect" in satellite remote sensing, significantly reducing the inversion error of deep-sea ecological parameters. Existing technologies, when using remote sensing data to extrapolate deep-sea structures, often rely on statistical regression of surface data, lacking physical modeling of the fluid vertical stratification mechanism, resulting in the inability to detect the maximum chlorophyll concentration (DCM) below the thermocline. This invention, through a "physically constrained spatiotemporal hypergraph construction module based on quasi-geotransfer theory," utilizes SQG theory to construct the geopotential vorticity correlation between the two-dimensional flow field at the sea surface and the three-dimensional structure of the deep sea, and introduces fluid dynamics conservation constraints. Experiments show that this design effectively solves the problem of physical mapping from the sea surface to the deep sea; even on the complex Deep-Blue-Dual-2025 stereo dataset, the root mean square error (RMSE) for chlorophyll concentration inversion at a depth of 75m underwater remains as low as 0.12 mg / m². 3 Compared to the traditional CNN-LSTM method (RMSE 0.68 mg / m), 3 This reduced the accuracy by approximately 82.3%, significantly improving the penetration accuracy and vertical structure reproduction of three-dimensional monitoring.

[0011] Secondly, this invention solves the problem of deep fusion of multi-source heterogeneous data within a fluid dynamics framework, significantly improving the integrity of the flow field structure. Given the significant differences in spatiotemporal resolution and physical properties between sea color satellites (optical), altimeters (radar), and in-situ buoys (point sources), existing technologies often employ simple spatiotemporal interpolation or pixel-level stitching, resulting in a dynamically fragmented flow field after fusion. This invention, through a "multi-source remote sensing manifold alignment and cross-medium dynamic feature reconstruction module," utilizes fluid potential energy manifold mapping and vortex adaptive position coding techniques to achieve deep alignment of cross-modal data in the "fluid manifold space." Experimental data show that this method achieves a structural similarity index (SSIM) of 0.92 when reconstructing mesoscale eddies and frontal structures, far exceeding the existing ST-GNN spatiotemporal map network model (0.76), greatly improving the self-consistency and logical coherence of the monitoring results in fluid dynamics.

[0012] Furthermore, this invention solves the challenge of truth value determination under multi-source conflict, achieving an extremely high objective consensus rate. This invention abandons the rigid processing method of simple weighted averaging in traditional data assimilation, utilizing a "remote sensing-mechanism double-blind game-based ecological parameter truth value consensus module" to introduce Nash equilibrium theory to simulate the dynamic negotiation process between the "observation agent" and the "mechanism agent." This mechanism can automatically perform dynamic arbitration based on the source confidence level under different sea conditions (e.g., reliance on the model in cloud-covered areas, reliance on remote sensing in active frontal areas), effectively filtering out unilateral extreme biases. Experimental verification shows that under complex sea conditions with frequent multi-source conflicts, the double-blind consensus rate of this method reaches 97.5%, while the OI (optimal interpolation) model without the game-theoretic mechanism can only reach about 60%-65%, demonstrating the superior ability of this invention to handle subjective rating conflicts and generate objective and fair evaluations.

[0013] Finally, this invention achieves a leap from static monitoring to Lagrange dynamic evolution, significantly extending the early warning period for ecological disasters. Based on the coupling law of material transport and biochemical reactions, this invention employs an "ecological evolution prediction module with a Lagrange coherent structure," utilizing particle tracking technology to accurately capture the transport trajectory and biochemical fluctuations of water masses. Combined with a dynamic early warning threshold mechanism, the system can issue timely warnings before red tides or hypoxia events occur. Experimental results show that the average early warning lead time for red tide outbreak trajectories using this method is as long as 7.5 days, nearly 3 days longer than the existing best Eulerian grid prediction model. This provides a valuable emergency response window for marine management departments, enabling process-oriented management and pre-emptive prevention of marine ecological monitoring. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of a three-dimensional monitoring method for marine ecology based on multi-source remote sensing and fluid game theory, according to Embodiment 1 of the present invention. Figure 2 This is a comparison chart of errors of different models in a three-dimensional monitoring task according to Embodiment 1 of the present invention; Figure 3 This is a comparison chart of the physical consistency residuals of each model in Embodiment 1 of the present invention.

[0015] Figure 4 This is a comparison chart of the consensus achievement rates of various game models in Embodiment 1 of the present invention.

[0016] Figure 5 This is a comparison chart of the disaster early warning lead time of various models in Embodiment 1 of the present invention.

[0017] Figure 6 This is the overall architecture diagram of the marine ecological three-dimensional monitoring system based on the fluid game model in Embodiment 1 of the present invention. Detailed Implementation

[0018] The present invention will be further described in detail below with reference to the accompanying drawings.

[0019] Example 1 Reference Figure 1 This embodiment of a three-dimensional monitoring method for marine ecology based on multi-source remote sensing and fluid game theory includes: This invention proposes "a method and system for three-dimensional monitoring of marine ecology based on a large-scale fluid game model driven by multi-source remote sensing" (RS-FluidGameNet). Its overall technical framework is designed according to a logical closed loop of "remote sensing manifold reconstruction - SQG physical topology - game truth value optimization - three-dimensional evolution prediction", and includes four core modules: Module 1: multi-source remote sensing manifold alignment and cross-medium dynamic feature reconstruction module; Module 2: physical constraint spatiotemporal hypergraph construction module based on quasi-geotransfer theory; Module 3: ecological parameter truth value consensus module based on remote sensing-mechanism double-blind game; Module 4: ecological evolution prediction module based on Lagrange coherence structure.

[0020] The specific plan is as follows: I. Multi-source remote sensing manifold alignment and cross-medium dynamic feature reconstruction module This module, serving as the system's input, aims to construct a high-dimensional feature space that conforms to fluid dynamics constraints. Its core logic is as follows: First, it performs temporal and spatial dynamic alignment on multi-source heterogeneous data to generate a basic spatiotemporal vector; then, it uses a large model to extract semantic features from remote sensing spectra and combines these with the spatiotemporal vector to generate comprehensive sea surface features; finally, based on fluid vertical projection theory, it extrapolates the sea surface features to three-dimensional space, forming the input tensor required by the subsequent game theory module.

[0021] 1) Time-series harmonic normalization processing of multi-source satellite orbital periods To address the non-uniformity of observation time caused by the different revisit periods of various satellite payloads (such as MODIS and Sentinel-3), the system first calculates the temporal feature vector of the observed events. For any remote sensing pixel or in-situ observation point, its original sampling timestamp is denoted as... In order to generate time codes that reflect tidal and seasonal patterns. The calculation formula is defined as follows: , in, This indicates the starting reference time set for this monitoring task; This represents the total time span of the task, used to linearly normalize the absolute time to the [0,1] interval; It is used to capture short-period hydrodynamic changes, based on the main tidal cycle (such as the M2 tidal cycle, which is about 12.42 hours). It is a seasonal cycle (usually 365 days) used to capture long-term ecological evolution; This represents a vector concatenation operation; This is a learnable temporal feature projection matrix, used to map the concatenated temporal features to the hidden layer dimensions defined by the model. ; Output This is the standardized temporal feature vector for this observation point, which will serve as one of the basic inputs for semantic feature fusion in subsequent step 3). 2) Construction of manifold spatial coordinates based on sea surface dynamic potential energy To address the issue that traditional geographic coordinates (latitude and longitude) cannot reflect the "flow-line motion" characteristics of ocean fluids, this system utilizes sea surface height anomalies and sea surface temperature to construct dynamic manifold coordinates. For any observation point... Its original geographical location was Calculate its manifold position encoding vector The formula is as follows: , in, Normalized longitude and latitude; The anomaly value of sea surface height measured by the satellite altimeter corresponding to this point; The modulus of the sea surface temperature gradient is used to indicate the location of fronts; It is a stream function transformation operator used to transform a scalar field into streamline potential energy values, so that points on the same streamline have similar potential energy values; The relative vorticity, calculated based on geostrophic relationships, is used to characterize the rotational intensity of the fluid. This is a multilayer perceptron network used to nonlinearly map the aforementioned physical quantities into high-dimensional position codes; the output... This includes the fluid's geometric position and dynamic state, which will serve as another basic input for step 3).

[0022] 3) Fusion of remote sensing spectral semantic extraction and sea surface dynamic features Obtaining the time vector and space vectors This step then aims to extract ecological semantics from remote sensing spectral data and fuse spatiotemporal information to generate a comprehensive sea surface feature vector. The input data is a multi-band remote sensing reflectance vector. (If the data is in situ, pad with zeros). Generate comprehensive sea surface features using the large model encoder. The calculation formula is as follows: , in, These are the multi-band reflectance values ​​after atmospheric correction; Preset text prompt vectors for sensor types (such as "MODIS Ocean Color Mode") to guide large models to understand the data source; This indicates the addition of feature embeddings; This is the encoding layer of a large language model finely tuned with oceanographic knowledge, used to extract semantic information such as chlorophyll and suspended matter from the spectrum; The time vector output in step 1); The manifold position vector output in step 2); The spatiotemporal fusion weight matrix ensures that the spatiotemporal features and semantic features have the same dimension. This is a layer normalization operation. The output is... It is a two-dimensional sea surface feature vector that simultaneously contains "ecological semantics" and "spatiotemporal dynamic background".

[0023] 4) Vertical projection and 3D field reconstruction based on SQG theory To address the skin effect in remote sensing, this step utilizes the Quasi-Geotransfer (SQG) theory to transform the two-dimensional sea surface features generated in step 3) into... Perform physical deductions into deeper layers. Define the target depth layer as... ( ), calculate the inference features of this depth The formula is as follows: , in, This is the sea surface composite feature vector output in step 3); and These represent the two-dimensional Fourier transform and inverse Fourier transform, respectively, used to process convolutional relationships in the frequency domain; The buoyancy frequency represents the stability of vertical stratification of water. The horizontal wavenumber is obtained from the frequency domain analysis of the characteristic map; These are Coriolis parameters, which are latitude-dependent. The term is the vertical transfer function derived from SQG theory, which physically means that small-scale features of the sea surface decay rapidly with depth, while large-scale features decay slowly. This is a deep light attenuation correction factor used to simulate the natural decline in biomass below the euphotic layer. This step outputs values ​​for arbitrary depths. Deductive characteristics at the location .

[0024] 5) Tensor completion and initial field generation for multi-source heterogeneous data Following step 4), a quasi-three-dimensional field based on remote sensing inversion was obtained, but this field lacks correction from in-depth measured data. This step will deduce the characteristics... It is fused with sparse underwater in-situ observation data. The final module output tensor is constructed. The calculation formula is as follows: , in, The vertical inference feature field output in step 4); This is the actual value of underwater in-situ observation at that location (such as dissolved oxygen measured by the Argo buoy). It is a binary mask tensor, with a value of 1 at positions containing measured data and a value of 0 at positions without data; For numerical embedding layers, scalar observations are mapped to... Vectors of the same dimension; The Laplace smoothing coefficient; This is a second-order differential operator used to smooth the transition between measured and derived points, eliminating seams. The final output... It is a physically self-consistent initial three-dimensional ecological feature tensor that integrates remote sensing inference and in-situ measurement. It will serve as the direct input for the next module, "Physically Constrained Spatiotemporal Hypergraph Construction Module Based on Quasi-Geotransformation Theory".

[0025] II. Physically Constrained Spatiotemporal Hypergraph Construction Module Based on Quasi-Geomorphic Flow Theory This module immediately follows the initial three-dimensional ecological feature tensor output by module one. Although Module 1 solved the "skin effect" through SQG vertical projection, the derived 3D field is still composed of discrete grid points in the horizontal direction, lacking a holistic model of ocean fluid topology (such as mesoscale eddies, filaments, and fronts). This module utilizes hypergraph technology to connect all grid points belonging to the same hydrodynamic system (such as a cold vortex) into a "hyperedge," and forcibly introduces the conservation laws of hydrodynamics during information transmission to eliminate artifacts that violate physical laws.

[0026] 1) Dynamic neighborhood search based on Lagrange coherent structures To construct a hypergraph, it is first necessary to determine which nodes are dynamically "nearest neighbors." Traditional Euclidean distance cannot reflect the transport trajectory of fluid particles. This invention employs the Finite-Time Lyapunov Exponent (FTLE) to identify the Lagrangian Coherent Structure (LCS), using this as the neighborhood determination criterion. For the feature tensor... any node in Its dynamic neighborhood set The construction formula is as follows: , in, and Representing nodes respectively and nodes Three-dimensional coordinate vectors in manifold space; The Euclidean distance (L2 norm) between two vectors is used to measure their proximity in geometric space. The set basic spatial search radius; The logical AND operation means that neighboring nodes must simultaneously satisfy three conditions: distance, FTLE value, and flow direction. and They are nodes and The finite-time Lyapunov exponent value at that location; This represents the absolute value operator, used to calculate the absolute value difference between two FTLE values. The difference threshold for FTLE ridge determination requires that neighboring nodes must be located inside the same fluid structure and cannot cross vortex boundaries (i.e., cannot cross FTLE ridges). and Representing nodes respectively and nodes The velocity vector at the location (containing three components: u, v, and w); This represents the angle between the two velocity vectors; It is a cosine function; The flow direction consistency threshold (with a value close to 1) ensures that only particles with similar flow directions are connected.

[0027] 2) Construction of the correlation matrix of the hyperedge of the quasi-geotropic streamline After determining the dynamic neighborhood, the system constructs a "streamlined hyperedge". A hyperedge This represents a complete streamline or a closed vortex contour. Define the incidence matrix of the hypergraph. ( The total number of nodes. (total number of superedges), its elements The calculation formula is as follows: , in, Indicates the first The node belongs to the _th Membership weights of hyperedges; For the first Clustering centers of superedges (i.e., vortex core centers or mainstream axis centers). The potential function distance is defined based on the quasi-geotransfer (SQG) theory, specifically calculated at the node level. With the center The difference in potential vorticity (PV) between them; For the natural constant An exponential function with base 1 is used to convert distance into similarity probability; This indicates that for all nodes belonging to this superedge Summation is performed, and the sum is used as the normalized denominator; This is an indicator function; if the condition within the parentheses is true (i.e., the node...), it indicates that the function is valid. The spatial coordinates are located in the spatial region of the fluid structure. If the value is inside the matrix, it takes the value 1; otherwise, it takes the value 0. The probability membership of each remote sensing observation point to a specific water mass or eddy is described, forming the topological skeleton of the hypergraph.

[0028] 3) Hyperedge feature aggregation that satisfies potential vorticity conservation This is the core step in introducing physical constraints. When converging nodal features to a hyperedge to form a "global water mass feature," the conservation law of potential vorticity (PV) in geophysical fluid dynamics must be obeyed (under the adiabatic frictionless assumption). The calculation of the... Aggregation features of superedges Furthermore, physical residual correction is incorporated during the aggregation process, as shown in the following formula: , in, The first input for module one The feature vector of each node; This is the aggregated weight matrix, used to adjust the feature dimensions; This represents a weighted summation over all nodes; For non-linear activation functions (such as LeakyReLU); This is the physical constraint penalty coefficient, used to control the strength of the physical correction term; This indicates taking the absolute value, which is used here to calculate the magnitude of the matter derivative; The mass derivative operator in fluid mechanics (including local variation terms) Horizontal transport items ), used to calculate the rate of change of physical quantities as a function of fluid particle motion; For nodes Quasi-geotropic vorticity value; This represents the weighted average potential vorticity of the entire superedge (water mass); This formula corrects the direction vector for features that the network can learn. The physical meaning of this formula is: according to the conservation law, the PV mass derivative of the water mass should be 0; if the calculated result is not 0 (i.e., ...), then... If the error term is relatively large, it indicates a violation of physical laws, and the model will automatically subtract this error term.

[0029] 4) Cross-scale fluid information interaction and node feature feedback After hyperedge aggregation is complete, each hyperedge now carries macroscopic information about the entire vortex or front. This macroscopic information needs to be fed back to the microscopic grid nodes to utilize the large-scale flow field background to correct local ecological parameter estimates. The updated node features are then calculated. The formula is as follows: , in, The original features input to module one serve as the basis for residual connections; + indicates matrix addition; This is a random deactivation operation used to prevent overfitting; Indicates a node All superedges Perform summation; For elements of the correlation matrix; This is the linear projection matrix during feature backpropagation; The hyperedge features that conform to physical conservation are obtained from step 3); The scalar coefficients calculated for the attention mechanism represent the nodes. In the super-edge Importance weights are assigned (e.g., active nodes at the edge of a vortex may have higher weights). This step enhances the context of the micro-ecological inversion results with macro-fluid dynamics information.

[0030] 5) Reorganization and output of spatiotemporal hypergraph feature tensors After hypergraph convolution processing, each node incorporates information from its surrounding dynamic manifold. Finally, the system performs tensor reorganization on the updated node features to generate the final output tensor of this module. The flow field confidence plot is attached. The formula is as follows: , in, For the output layer, a multilayer perceptron network; The node features output in step 4); This is a layer normalization operation to ensure numerical stability; This indicates a splicing operation performed along the channel dimension of a tensor. The flow field confidence map tensor is calculated based on the physical residuals from step 3) (the smaller the residuals, the higher the confidence level). The final output is... It is a high-level feature tensor that not only contains ecological information retrieved from remote sensing, but also undergoes topological correction by fluid dynamics. It will serve as the trump card of the "observation agent" and participate in the game of the next module.

[0031] III. Ecological Parameter Truth Consensus Module Based on Remote Sensing-Mechanism Double-Blind Game Theory This module aims to resolve the source conflict problem arising from inconsistencies between remote sensing inversion and numerical simulation conclusions in ocean monitoring. The system constructs an observation agent and a mechanism agent, both utilizing the fluid hypergraph feature tensor output by module two. Based on the game theory framework, numerical consensus is reached through multiple rounds of data handshakes and strategy confrontations.

[0032] 1) Construction of agent utility function based on hypergraph confidence The first step in game theory is to establish the "bargaining leverage" (i.e., utility functions) of both parties. Observing the agents. The tensor output by module two must be parsed first. To determine one's confidence level in one's current position and convert this confidence level into a utility value. This serves as the basis for subsequent bidding weights. The calculation formula is as follows: in, Indexing spatial grid points; It is the flow field physical consistency confidence channel extracted directly from the output tensor of Module 2 (this value is determined by the physical residual of Module 2, and the smaller the residual, the higher the confidence). It is a sigmoid activation function used to normalize the confidence level to the (0,1) interval; It is the ecological feature vector channel output by module two; Used to transform feature vectors into probability distributions; Used to calculate the above The information entropy of the feature probability distribution output by the function; the smaller the entropy, the more significant the feature and the higher the utility. The output is a balanced weight between confidence level and feature significance. It is a scalar that quantifies the "confidence" of the observing agent in the current data. This value will be directly passed to step 2) to control the variance of the initial bid.

[0033] 2) Utility-driven initial ecological parameter bidding generation To obtain the utility value calculated in step 1). Then, the observation agent combines Generate the first round of ecological parameter estimates (i.e., the "initial bid"). The bid depends not only on features but also on the modulation of noise by the utility value: higher utility leads to a more certain bid; lower utility results in greater exploratory noise. Initial bid tensor. The calculation is as follows: , in, and These are the policy network weight matrix and bias vector of the observation agent, respectively, used to map the high-dimensional hypergraph features to specific physical quantities (such as chlorophyll concentration). This is the original output tensor of module two; This is a Gaussian noise generation term; Explore variance as a benchmark; The utility field calculated in step 1) appears as the denominator, meaning that when the confidence level given by module 2 is extremely high, the noise term approaches 0, and the agent will "firmly" give an exact value; otherwise, it will conduct a large-scale random exploration. This step generates... This will be used as a basic variable in step 3).

[0034] 3) Calculation of Game-Against-Situation Losses Based on Bid Differences When the observing agent generates a bid At the same time, the mechanistic agent also generates bids. Next, the system needs to calculate the current "negotiation disagreement," i.e., the game loss function. This loss function... It directly depends on the bidding results from step 2) and measures the consistency between the two parties and the satisfaction of physical constraints. The calculation formula is as follows: , in, and These are the bid tensors generated in step 2) (or updated in subsequent rounds); This represents the squared L2 norm of the tensor, used to quantify the Euclidean distance (divergence magnitude) between the two. A spatially weighted mask, which is derived from the utility value in step 1). and Joint decisions (e.g., taking the maximum of the two) should be made to ensure that disagreements in high-reputation areas are amplified and given greater attention. For physical constraint coefficients; As a divergence operator, the second term mandates that the bid field of the observing agent must satisfy fluid continuity. The scalar output of this step... The "dissatisfaction" that accurately describes the current game state will be directly used for backpropagation in step 4).

[0035] 4) Loss gradient-guided policy iteration and bid update To eliminate the discrepancies calculated in step 3) The agent must adjust its policy. This step uses the loss value output from step 3) to update the policy parameters using gradients, thereby generating a better bid for the next round. The update logic is as follows: , in, For the previous round of bidding; For game learning rate; The partial derivative (gradient) of the loss function with respect to the bid tensor in step 3) indicates the direction of "how to modify the bid to reduce the divergence"; As a moderating factor based on the utility value in step 1), its physical meaning is extremely crucial: if the utility value at a certain position... If the confidence level is very high (i.e., high confidence in Module 2), the factor approaches 0, and the agent will refuse to modify its bid (stick to its own opinion); conversely, it will compromise significantly along the gradient direction. After rounds of iteration, the final convergent bid is obtained. .

[0036] 5) Truth-value consensus fusion based on game stability After After rounds of confrontation and compromise, step 4) outputs the final bids from both sides. This step utilizes the "path stability" principle in the game process to determine the final outcome. If an agent frequently and significantly modifies its bid during iteration in step 4), it indicates that its data is unreliable. The final fusion formula is as follows: , In this formula, and It is the final tensor after the iteration in step 4); This represents the variance tensor of the bidding sequence of the observing agent in all iteration rounds of step 4); This indicates the reciprocal operation. The smaller the variance, the larger the reciprocal, and the higher the weight. This formula automatically filters out the most "reliable" data sources through mathematical mechanisms, generating... It is the final product of this module and will be used directly as the initial field for the next module.

[0037] IV. Ecological Evolution Prediction Module of Lagrange Coherent Structures This module is based on the high-precision true field output from module three. This module uses the Lagrange perspective from fluid mechanics to predict the future evolution of marine ecosystems. The logical starting point for this module must be... This is because only fields that have undergone game-theoretic correction possess the physical stability required for long-term integration.

[0038] 1) Lagrange particle scattering and initialization based on true-value flow field To perform evolutionary prediction, we first need to consider the output of Module 3. The continuous fluid is discretized using virtual particles in the field. A set of particles is defined. For the first Each particle has an initial full-state vector. directly from sampling: , in, For the first At the initial moment, the particles The complete state vector encapsulates the particle's physical position and biochemical properties; These are the three-dimensional spatial coordinate components of the particle; The ecological concentration component carried by the particle (such as chlorophyll concentration, extracted from...) (biochemical pathways) For spatial trilinear interpolation functions, used to interpolate from discrete grid tensors Accurate extraction of random coordinates The attribute value at the location; These are the initial coordinate points generated within the monitoring area through random or uniform sampling. This step completes the process of extracting data from the Eulerian grid (…). ) to Lagrange particle objects ( The data format transformation prepares the data for tracking motion trajectories.

[0039] 2) Particle position and state update under Runge-Kutta integral Obtain the initial state Then, its position components need to be updated using fluid dynamics equations. Thus, the intermediate state is obtained. Due to the highly nonlinear nature of ocean current fields, a fourth-order Runge-Kutta (RK4) integral method is employed: , in, This represents the position component in the updated particle state vector; The position component from the previous moment; To predict the step size; These are the velocity slope vectors at four intermediate time points in the RK4 algorithm. Specifically, It is the flow rate at the current location. and It estimates the flow velocity at the predicted half-step position. This is an estimate of the flow velocity at the predicted full-step position. These velocity vectors all need to be accessed in real time. The flow field data is combined with the current particle position. The calculations show that this step only updates the spatial coordinates of the state vector, thus completing the simulation of the physical transport process.

[0040] 3) Update of biochemical reaction source and sink status along the trajectory After the particle's position changes, the concentration component in its state vector It will also change due to biochemical processes. A large-scale model is used to predict biochemical increments, and the state vector is updated a second time. , in, The updated particle ecological concentration component; This represents the initial concentration component; For the natural decay term, where This is the biological attenuation coefficient (such as phytoplankton mortality rate). It is an exponential function, and this term describes the physical dissipation of matter during transport. The ecological large-scale model prediction function explicitly calls the location components that were just updated in step 2) as input parameters. Using the environmental factors of the new location as context, the net primary productivity increment is output. At this point, the particle state vector... All components have been updated, incorporating both physical and biochemical evolution.

[0041] 4) Euler mesh remapping based on full-state particles Through the deductions in steps 2) and 3), the future time was obtained. The set of all states of particles To generate a visual monitoring graph, these state vectors need to be used to reconstruct the Eulerian mesh tensor. Kernel density estimation (KDE) is used: , in, For grid points The predicted value at that location; the formula molecule explicitly calls the concentration component of the particle state. As weight; This is a Gaussian kernel function used to calculate the contribution weights of particles to grid points; The bandwidth parameter of the kernel function determines the spatial range (smoothness) of the particle's influence. Represents the Euclidean distance (L2 norm), used to calculate grid points. With particle position Spatial distance between; denominator The normalization coefficient is the sum of the weights contributed by all particles to this grid point, ensuring the conservation of numerical density after resampling. This step transforms the discrete Lagrangian state object back into a continuous Eulerian field.

[0042] 5) Dynamic early warning threshold determination and monitoring report generation Finally, the system processes the data generated in step 4). Conduct disaster risk analysis. Calculate the risk for a specific monitoring area. Ecological anomaly index within : , in, To accumulate the ecological risk index; The designated monitoring sea area (integration domain); For logical indicator functions, when predicting concentration Greater than the historical warning threshold The value is 1 if it is true, and 0 otherwise. Indicates the extent to which the standard is exceeded; Let be a spatial integral infinitesimal element. If the calculated... If the preset alarm threshold is exceeded, the system will automatically generate a monitoring and early warning report that includes the location, range, and intensity.

[0043] Experimental verification To systematically verify the performance advantages of the method of this invention in general remote sensing scenarios and specific stereo monitoring scenarios, we constructed and used two datasets with different characteristics for comparative experiments: a publicly available benchmark remote sensing dataset and a self-built multi-source stereo heterogeneous dataset.

[0044] Dataset Descriptions: Global-Sat-2024 (Public Benchmark Dataset): Collected from the internationally available GlobalOceanColor satellite remote sensing product library. This dataset primarily contains daily products of sea surface chlorophyll concentration (Chl-a) and sea surface temperature (SST) in a single modality. It features wide coverage and regular time series, including 50,000 remote sensing image sequences of key sea areas and corresponding field buoy calibration points. It is mainly used to verify the basic performance and generalization ability of the model when handling standard two-dimensional sea surface monitoring tasks. Since this dataset only contains sea surface data and does not involve deep-sea three-dimensional structures or complex numerical model games, some indicators (such as Depth-RMSE and Consensus-Rate) are not applicable to this dataset (N / A). Deep-Blue-Dual-2025 (Self-built 3D Dataset): This is a self-built, high-difficulty dataset used in this study. It simulates and measures the entire "sky-sea-diving" process data of a typical sea area (including the continental shelf and deep-sea areas) over a complete year. This dataset contains typical dual-stream heterogeneous data. Stream 1 consists of high-frequency, wide-coverage multi-source satellite remote sensing data (MODIS / Sentinel-3 / Jason-3), while Stream 2 consists of low-frequency, high-precision underwater Argo buoy profiles and moored array data. Simulated fields from the concurrent ROMS numerical model are also included as a game opponent. The dataset contains 100,000 remote sensing images and 20,000 underwater profile data, with fine-grained physical process annotations by oceanographic experts. The aim is to verify the core advantages of this invention in handling "skin effect" removal, deep-sea inversion, and multi-party game theory.

[0045] Experimental setup: To comprehensively evaluate the performance, we selected four representative methods for comparison with the method of this invention.

[0046] The CNN-LSTM method: Remote sensing images are input into a convolutional neural network in chronological order to extract features, and then LSTM is used for temporal prediction. This represents a pure data-driven deep learning benchmark that ignores fluid physics laws. The OI-Merge (Optimal Interpolation) method: A basic multi-source fusion method, it uses the Gauss-Markov theorem to perform statistical weighted interpolation on multi-source data, representing the traditional static assimilation approach. The ST-GNN (Spatiotemporal Graph Neural Network) method: Uses a standard spatiotemporal graph convolutional network to process gridded data, representing a strong AI competitor in this field, but it does not incorporate fluid dynamics constraints. The HAN (Heterogeneous Graph Attention Network): Capable of handling different types of nodes, but lacks in physical conservation. Finally, the RS-FluidGameNet method proposed in this invention.

[0047] Evaluation Metrics: The experiment used five core metrics for evaluation. Surface-RMSE: Measures the root mean square error of sea surface parameter predictions; lower values ​​are better. Depth-RMSE: Specifically measures the accuracy of ecological parameter inversion at 75m underwater (DCM layer); lower values ​​are better. Phys-Consistency: Measures whether the prediction results violate the law of conservation of mass (e.g., sourceless creation and destruction); it uses normalized physical residuals; lower values ​​are better. Consensus-Rate: Measures the degree of agreement between remote sensing inversion and numerical simulation after game theory; higher values ​​are better. LeadTime: Measures the average number of days in advance the model predicts the trajectory of disasters such as red tides and issues warnings; higher values ​​are better. Time: Records the processing time for a single area sample.

[0048] Table 1. Performance comparison of different methods on the Global-Sat-2024 and Deep-Blue-Dual-2025 datasets. The experimental results are shown in Table 1. Figure 2 , Figure 3 , Figure 4 and Figure 5 As shown, through in-depth comparative analysis of the experimental data of each model in Table 1 on the dual datasets, the technical advantages of this invention in handling complex three-dimensional marine ecological monitoring tasks can be clearly revealed. The specific analysis is as follows: First, compared with benchmark models such as CNN-LSTM and OI-Merge, this invention demonstrates an overwhelming advantage in handling deep-sea three-dimensional inversion tasks. Experimental data shows that on the complex Deep-Blue-Dual-2025 dataset, the depth-RMSE of these two benchmark models is as high as 0.68 and 0.72, which is basically unacceptable for actual deep-sea monitoring needs. This is because CNN-LSTM and OI-Merge only use simple statistical mapping or linear interpolation, ignoring the complex nonlinear hydrodynamic relationship between the sea surface and the deep sea, resulting in severe distortion when extrapolating the thermocline structure. In contrast, this invention introduces a "physically constrained spatiotemporal hypergraph construction module based on quasi-geotransfer theory," and utilizes SQG theory and large model instruction encoding technology to achieve deep physical mapping of cross-medium data in the underlying feature space, thereby reducing the depth-RMSE to 0.12 and fundamentally solving the vertical blind zone problem caused by the "skin effect."

[0049] Secondly, compared with the spatiotemporal graph models ST-GNN and HAN, this invention significantly surpasses them in terms of physical consistency and early warning timeliness. Although ST-GNN and HAN perform reasonably well on the single-modal Global-Sat-2024 dataset (Surface-RMSE 0.22 and 0.25, respectively), their errors are relatively high (0.32-0.38) when facing the "physical consistency" metric, which requires strict adherence to conservation laws. This indicates that the generated flow field contains a large number of non-physical artifacts (such as material being generated out of thin air). This invention not only offers superior prediction accuracy but, more importantly, explicitly embeds fluid continuity constraints and material derivative penalties into the "Lagrange coherent structure ecological evolution prediction module." This mechanism forces the model to follow the Navier-Stokes equations, reducing the Phys-Consistency error to 0.05. Thanks to this physical stability, this invention can pinpoint the transport trajectory of ecological anomalies earlier, extending the LeadTime to 7.5 days, nearly 3 days longer than ST-GNN.

[0050] Furthermore, compared to its strong competitor HAN (Heterogeneous Graph Attention Network), this invention achieves a breakthrough improvement in consensus rate in multi-source game theory. While HAN can handle heterogeneous nodes, its consensus rate is only 78.9% when dealing with conflicts between remote sensing and model data. Its main drawback is the lack of a mechanism that can simulate the dialectical analysis of human experts. This invention, through the "remote sensing-mechanism double-blind game theory ecological parameter truth consensus module," introduces Nash equilibrium theory to construct a double-blind game theory between observation and mechanism, effectively eliminating the systematic bias of a single source and increasing the Consensus-Rate to 97.5%.

[0051] In summary, the method of this invention not only demonstrates robustness in general sea surface monitoring tasks, but also comprehensively surpasses existing mainstream technical solutions in complex scenarios such as multi-source three-dimensional fusion, deep-sea physical inversion, multi-party game collaboration, and dynamic evolution prediction, proving the effectiveness of the "multi-source remote sensing + fluid game model" technical approach.

[0052] Example 2 This embodiment provides a three-dimensional monitoring system for marine ecology based on multi-source remote sensing and fluid game theory.

[0053] A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device, the aforementioned method for three-dimensional monitoring of marine ecology based on multi-source remote sensing and fluid game theory.

[0054] A terminal device includes a processor and a computer-readable storage medium, the processor being used to implement various instructions; the computer-readable storage medium being used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor as described in the multi-source remote sensing-based fluid game-theoretic three-dimensional monitoring method for marine ecology.

[0055] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for three-dimensional monitoring of marine ecology based on fluid game by multi-source remote sensing, characterized in that, include: Acquire multi-source heterogeneous data; Based on the acquired multi-source heterogeneous data, multi-source remote sensing manifold alignment and cross-medium dynamic feature reconstruction are carried out, including time-series harmonic normalization processing of multi-source satellite orbital periods and manifold spatial coordinate construction based on sea surface dynamic potential energy, remote sensing spectral semantic extraction and sea surface dynamic feature fusion, and vertical projection and three-dimensional field reconstruction based on SQG theory. Based on reconstructed information, a physically constrained spatiotemporal hypergraph based on quasi-geotransfer theory is constructed, including dynamic neighborhood search based on Lagrange coherent structure and correlation matrix construction of quasi-geotransfer streamline hyperedges; hyperedge feature aggregation that satisfies potential vorticity conservation and cross-scale fluid information interaction and node feature backhaul. The consensus on the truth value of ecological parameters generated by the remote sensing-mechanism double-blind game theory is based on the construction results, including the construction of agent utility function based on hypergraph confidence and the generation of initial ecological parameter bids driven by utility; the calculation of game adversarial loss based on bid difference and the strategy iteration and bid update guided by loss gradient; Based on the game results, ecological evolution prediction is performed using the Lagrange coherence structure, including Lagrange particle scattering and initialization based on the true value flow field and particle position state update under Runge-Kutta integral; biochemical reaction source and sink state update along the trajectory and Eulerian grid remapping based on full-state particles; Output the prediction results.

2. The method for three-dimensional monitoring of marine ecology based on multi-source remote sensing and fluid game theory as described in claim 1, characterized in that, The time-series harmonic normalization processing of the multi-source satellite orbital periods and the construction of manifold spatial coordinates based on sea surface dynamic potential energy include, to address the non-uniformity of observation time caused by different revisit periods of different satellite payloads, firstly calculating the time feature vector of the observed event. For any remote sensing observation pixel or in-situ observation point, its original sampling timestamp is denoted as... To generate time codes that reflect tidal and seasonal patterns ,definition: in, This indicates the starting reference time set for this monitoring task; The total time span of the task; The main tidal cycle; It is a seasonal cycle; This represents a vector concatenation operation; This is a learnable temporal feature projection matrix, used to map the concatenated temporal features to the hidden layer dimensions defined by the model. ; Output This involves creating a standardized time feature vector for the observation point. Then, addressing the issue that traditional geographic coordinates cannot reflect the characteristics of ocean fluid movement along streamlines, a dynamic manifold coordinate system is constructed using sea surface height anomalies and sea surface temperature. For any observation point... Its original geographical location was Calculate its manifold position encoding vector ,formula: in, Normalized longitude and latitude; The anomaly value of sea surface height measured by the satellite altimeter corresponding to this point; The modulus of the sea surface temperature gradient; For stream function transformation operators; The relative vorticity is calculated based on the geostrophic relationship; It is a multilayer perceptron network; the output is It includes the fluid's geometric position and dynamic state.

3. The method for three-dimensional monitoring of marine ecology based on multi-source remote sensing and fluid game theory as described in claim 2, characterized in that, The remote sensing spectral semantic extraction and sea surface dynamics feature fusion, along with vertical projection and three-dimensional field reconstruction based on SQG theory, include obtaining the time vector. and space vectors Subsequently, ecological semantics were extracted from remote sensing spectral data, and spatiotemporal information was fused to generate a comprehensive sea surface feature vector. The input data was a multi-band remote sensing reflectance vector. Large model encoder is used to generate comprehensive sea surface features. Calculation formula: in, These are the multi-band reflectance values ​​after atmospheric correction; For sensor type; This indicates the addition of feature embeddings; The encoding layer of a large language model, finely tuned with oceanographic knowledge; It is a time vector; Let the position vector of the manifold be denoted by . This is the spatiotemporal fusion weight matrix; For layer normalization operation, It is a two-dimensional sea surface feature vector; then, to solve the skin effect in remote sensing, the two-dimensional sea surface features generated by quasi-geothermal currents are used. To perform a deeper physical deduction, the target depth layer is defined as... Calculate the inference features of this depth. ,formula: in, This represents the comprehensive feature vector of the sea surface. and These represent the two-dimensional Fourier transform and inverse Fourier transform, respectively; The frequency of buoyancy; The horizontal wavenumber; For Coriolis parameters; The term is the vertical transfer function derived from SQG theory; This is a correction factor for deep illumination attenuation, outputting at any depth. Deductive characteristics at the location Finally, the characteristics will be deduced. By fusing with sparse underwater in-situ observation data, the final module output tensor is constructed. Calculation formula: in, For vertical derivation of characteristic fields; These are the actual values ​​from in-situ underwater observations. It is a binary mask tensor; For numerical embedding layers, scalar observations are mapped to... Vectors of the same dimension; The Laplace smoothing coefficient; For a second-order differential operator, the final output is It is the initial three-dimensional ecological feature tensor.

4. The method for three-dimensional monitoring of marine ecology based on multi-source remote sensing and fluid game theory as described in claim 3, characterized in that, The dynamic neighborhood search and correlation matrix construction based on the Lagrange coherent structure and the quasi-geotransfer streamline hyperedge include using the finite-time Lyapunov exponent FTLE to identify the Lagrange coherent structure (LCS) for the feature tensor. any node in Its dynamic neighborhood set The construction formula: in, and Representing nodes respectively and nodes Three-dimensional coordinate vectors in manifold space; This represents the Euclidean distance between two vectors. The set basic spatial search radius; and They are nodes and The finite-time Lyapunov exponent value at that location; The absolute value operator is represented by the symbol. The difference threshold for FTLE ridge determination; and Representing nodes respectively and nodes The velocity vector at that location; This represents the angle between the two velocity vectors; Define the consistency threshold for the flow direction; then define the association matrix of the hypergraph. , The total number of nodes. The total number of superedges, elements The calculation formula is as follows: in, Indicates the first The node belongs to the _th Membership weights of hyperedges; For the first Cluster centers of superedges; Let be the potential function distance defined based on the quasi-geotransfer SQG theory, and be the node. With the center The difference in potential vorticity (PV) between them; This indicates that for all nodes belonging to this superedge Perform summation; For indicator functions, spatial regions .

5. A three-dimensional monitoring method for marine ecology based on multi-source remote sensing and fluid game theory, as described in claim 4, is characterized in that... The aggregation of hyperedge features that satisfy the conservation of potential vorticity and the cross-scale fluid information interaction and node feature backhaul include calculating the first... Aggregation features of superedges And physical residual correction is added during the aggregation process, formula: ,in, For the first The feature vector of each node; This is the aggregate weight matrix; This represents a weighted summation over all nodes; It is a non-linear activation function; This is the physical constraint penalty coefficient; This is the mass derivative operator in fluid mechanics. For nodes Quasi-geotropic vorticity value; This represents the weighted average potential vorticity of the entire hyperedge; The direction vectors are corrected for the network's learnable features. Then, the macroscopic information after hyperedge aggregation is fed back to the microscopic grid nodes to calculate the updated node features. ,formula: ,in, Original features; This is a random deactivation operation; Indicates a node All superedges Perform summation; For elements of the correlation matrix; This is the linear projection matrix during feature backpropagation; To conform to the hyperedge characteristics of physical conservation; The scalar coefficients calculated for the attention mechanism represent the nodes. In the super-edge The importance weights are assigned to each node; finally, after hypergraph convolution, each node incorporates information from its surrounding dynamic manifold, and the updated node features are reconstructed into tensors to generate the final output tensor. Attached is a flow field confidence plot: , in, For the output layer, a multilayer perceptron network; For node features; For layer normalization operation; This indicates a splicing operation performed along the channel dimension of a tensor. For the flow field confidence map tensor.

6. The method for three-dimensional monitoring of marine ecology based on multi-source remote sensing and fluid game theory as described in claim 5, characterized in that, The construction of the agent utility function based on hypergraph confidence and the generation of utility-driven initial ecological parameter bids include, to address the source conflict problem between the results of ocean monitoring remote sensing inversion and numerical simulation, the construction of observation agents and mechanism agents. First, analyze the tensor. To determine one's confidence level in one's current position and convert this confidence level into a utility value. As the basis for subsequent bid weighting, the calculation formula is as follows: in, Indexing spatial grid points; It is a confidence channel for the physical consistency of the flow field; It is a sigmoid activation function; It is an ecological feature vector channel; Used for calculation Information entropy of the feature probability distribution output by the function; The output is a balanced weight between confidence level and feature significance. It is a scalar; in obtaining utility value Then, the observation agent combines Generate the first round of ecological parameter estimates and the initial bid tensor. The calculation is as follows: in, and These are the policy network weight matrix and bias vector of the observing agent, respectively; For the original output tensor; This is a Gaussian noise generation term; Explore variance as a benchmark; For utility field.

7. A three-dimensional monitoring method for marine ecology based on multi-source remote sensing and fluid game theory, as described in claim 6, is characterized in that... The game adversarial loss calculation based on bid difference and the strategy iteration and bid update guided by loss gradient, including when the observing agent generates a bid... At the same time, the mechanistic agent also generates bids. Then, calculate the current game loss function. Based on the bidding results and measuring the consistency between the two parties and the satisfaction of physical constraints, the calculation formula is as follows: ,in, and These are the bid tensors for both sides; This represents the squared L2 norm of the tensor; Use a space-weighted mask; For physical constraint coefficients; For divergence operators; to eliminate computational divergences The agent uses the output loss value to update the policy parameters using gradients, thereby generating a better bid for the next round. Update logic: in, For the previous round of bidding; For game learning rate; This is the partial derivative of the loss function with respect to the bid tensor; As a moderating factor for utility value, through After rounds of iteration, the final convergent bid is obtained. After After rounds of confrontation and compromise, the final offers from both sides are determined by the final fusion formula: , in, and It is the final tensor after the iteration is completed; The variance tensor of the bid sequence of the observing agent across all iteration rounds; the generated It is the final product.

8. A three-dimensional monitoring method for marine ecology based on multi-source remote sensing and fluid game theory, as described in claim 7, is characterized in that... The Lagrange particle scattering and initialization based on the true-value flow field and the particle position and state update under Runge-Kutta integral include, firstly, for evolution prediction, in the output... The continuous fluid is discretized using virtual particles in the field, and a particle set is defined. For the first Each particle has an initial full-state vector. directly from sampling: in, For the first At the initial moment, the particles The complete state vector; These are the three-dimensional spatial coordinate components of the particle; The ecological concentration component carried by the particle; It is a spatial trilinear interpolation function; These are the initial coordinate points generated through random or uniform sampling within the monitoring area; the initial state is obtained. Then, update the location components. Obtaining the intermediate state The fourth-order Runge-Kutta RK4 integration method is used: in, This represents the position component in the updated particle state vector; The position component from the previous moment; To predict the step size; These are the velocity slope vectors at four intermediate time points in the RK4 algorithm. Specifically... It is the flow rate at the current location. and It is the flow velocity estimate at the predicted half-step position.

9. A three-dimensional monitoring method for marine ecology based on multi-source remote sensing and fluid game theory, as described in claim 8, is characterized in that... The state update of biochemical reaction source and sink terms along the trajectory and the Eulerian grid remapping based on full-state particles include the concentration component in the state vector after the particle position changes. Changes will occur due to biochemical processes. A large model is used to predict biochemical increments, and the state vector is updated a second time. in, The updated particle ecological concentration component; This represents the initial concentration component; For the natural decay term, where The biological attenuation coefficient; For the ecological large-scale model prediction function, location component Future moments were obtained through deduction. The set of all states of particles To generate a visual monitoring map, the Eulerian mesh tensor is reconstructed using state vectors. Kernel density estimation of KDE is used: in, For grid points Predicted values ​​at the location; concentration components As weight; The Gaussian kernel function; This is the bandwidth parameter of the kernel function; Represents the Euclidean distance; denominator part These are the normalization coefficients; finally, for the generated... Conduct disaster risk analysis and calculate specific monitoring areas Ecological anomaly index within : in, To accumulate the ecological risk index; The designated monitoring sea area; For logical indicator functions; Indicates the extent to which the standard is exceeded; It is a spatial integral infinitesimal element.

10. A three-dimensional marine ecological monitoring system based on multi-source remote sensing and fluid game theory, characterized in that, include: The data acquisition module is configured to acquire multi-source heterogeneous data; The reconstruction module is configured to perform multi-source remote sensing manifold alignment and cross-medium dynamic feature reconstruction based on the acquired multi-source heterogeneous data, including time-series harmonic normalization processing of multi-source satellite orbital periods and manifold spatial coordinate construction based on sea surface dynamic potential energy, remote sensing spectral semantic extraction and sea surface dynamic feature fusion, and vertical projection and three-dimensional field reconstruction based on SQG theory. The hypergraph module is configured to construct a physically constrained spatiotemporal hypergraph based on quasi-geotransformation theory, based on reconstructed information. This includes dynamic neighborhood search based on Lagrange coherence structure and construction of the correlation matrix of quasi-geotransformation streamline hyperedges; aggregation of hyperedge features that satisfy potential vorticity conservation; and cross-scale fluid information interaction and node feature backhaul. The game theory module is configured to generate a consensus on the truth value of ecological parameters based on remote sensing-mechanism double-blind game theory, including the construction of agent utility function based on hypergraph confidence and utility-driven initial ecological parameter bid generation; game adversarial loss calculation based on bid difference and loss gradient-guided strategy iteration and bid update. The prediction module is configured to predict ecological evolution based on the game results using the Lagrange coherence structure. This includes Lagrange particle scattering and initialization based on true-value flow fields and particle position and state updates under Runge-Kutta integrals; biochemical reaction source and sink state updates along the trajectory; and Eulerian grid remapping based on full-state particles. The output module is configured to output the prediction results.

Images