A marine aquaculture risk prediction method based on marine multi-source observation data

Abstract

The present application provides a kind of marine aquaculture risk prediction method based on marine multi-source observation data, belong to marine aquaculture risk technical field, the present application is reconstructed three-dimensional marine environment field by sparse bayesian field reconstruction algorithm based on the constraint of ocean dynamics, then the risk benchmark distribution of long historical observation aquaculture area is migrated along the manifold coordinate to new aquaculture area by marine risk field manifold alignment migration algorithm, and the physical correction is carried out to bottom shape data and tidal harmonic constant, the risk index uncertainty interval is obtained by non-stationary generalized extreme value distribution combined with bayesian markov chain monte carlo sampling, drive physical constraint artificial intelligence to complete local high-resolution field reconstruction, according to the comparison result of risk index and early warning threshold, corresponding early warning level is triggered, finally form marine aquaculture insurance risk index product, solve the technical problem that new aquaculture area cannot establish reliable risk benchmark distribution due to lack of long-term historical observation data.

Description

Technical Field

[0001] This invention belongs to the field of marine aquaculture risk prediction technology, and specifically relates to a method for marine aquaculture risk prediction based on multi-source marine observation data. Background Technology

[0002] Risk prediction in marine aquaculture is a core technical aspect of insurance product design and claims trigger determination. Currently, for aquaculture areas with long-term historical observation data, extreme value statistics methods are typically used to model the risks of key environmental variables such as sea surface temperature and significant wave height. This is combined with numerical model outputs and observational data to construct a risk index sequence, and the warning threshold is determined using receiver operating characteristic (ROC) curve analysis. These methods have been widely applied in aquaculture areas with abundant historical data, providing a certain quantitative basis for claims verification.

[0003] However, the above methods rely on a sufficiently long historical observation sequence for the target area to support robust estimation of extreme value distribution parameters. For newly established aquaculture areas, the historical sequence is extremely short due to the recent deployment of observation stations. Extreme value statistical methods suffer from parameter estimation biases due to insufficient sample size, and numerical model outputs also exhibit systematic errors due to the lack of local observation correction. Consequently, the risk baseline distribution cannot be established through conventional methods.

[0004] In existing technologies, directly transferring risk parameters from adjacent historical regions to newly established aquaculture areas results in a mismatch between the transplanted risk benchmark distribution and the actual dynamic environment of the target area due to fundamental differences in the bottom structure, tidal dynamics, and local flow field characteristics of the two regions. This leads to significant deviations in the frequency and magnitude of extreme events in the risk index sequence, failing to meet the initial reliability requirements of insurance verification for risk assessment. In other words, existing technologies suffer from the technical problem of being unable to establish a reliable risk benchmark distribution for newly established aquaculture areas due to a lack of long-term historical observation data. Summary of the Invention

[0005] In view of this, the present invention provides a marine aquaculture risk prediction method based on multi-source marine observation data, which can solve the technical problem in the prior art that newly built aquaculture areas cannot establish a reliable risk benchmark distribution due to the lack of long-term historical observation data.

[0006] This invention is implemented as follows: This invention provides a method for predicting marine aquaculture risks based on multi-source marine observation data, comprising the following steps: Data on temperature, salinity, and depth profiles from marine stations, wave height and current velocity data from moored buoys, surface current field data from shore-based radar, and sea surface temperature and significant wave height data from satellite remote sensing were collected. After quality control, a sparse Bayesian field reconstruction algorithm based on ocean dynamics constraints was used to reconstruct the three-dimensional ocean environment field, and the posterior mean field and posterior variance field of the three-dimensional ocean environment field were output. The posterior mean field and posterior variance field of the three-dimensional marine environmental field are input into the marine risk field manifold alignment and transfer algorithm to perform cross-domain risk benchmark alignment between newly established aquaculture areas and long-historical observed aquaculture areas, and output the target domain risk benchmark distribution. Using the target domain risk baseline distribution as input, the extreme values ​​of sea surface temperature and significant wave height are statistically modeled using a non-stationary generalized extreme value distribution. The posterior distribution of the non-stationary generalized extreme value distribution parameters is obtained by combining Bayesian Markov chain Monte Carlo sampling, and the uncertainty interval of the risk index is output. Using temperature, salinity, current velocity, and significant wave height output from a regional ocean model as background field boundary conditions, a physically constrained neural network is driven to complete local high-resolution field reconstruction. The reconstruction results and the uncertainty interval of the risk index are input into the spatiotemporal prediction model of marine aquaculture risk, and the hourly risk index sequence and hourly risk index probability distribution of the aquaculture area are output. Calculate the dynamic resource scheduling function value, adjust the memory allocation ratio and batch size of the marine aquaculture risk spatiotemporal prediction model according to the interval to which the dynamic resource scheduling function value belongs, and trigger the corresponding risk warning level and execute the corresponding process based on the comparison results of the hourly risk index sequence with the yellow warning threshold, orange warning threshold and red warning threshold. The hourly risk index sequence, hourly risk index probability distribution, risk index uncertainty range, and risk warning level are output together to form a marine aquaculture insurance risk index product, which is used for insurance product design and compensation trigger determination.

[0007] Specifically, the quality control involves using three methods—climatological range test, internal consistency test, and spatiotemporal continuity test—to screen out outliers layer by layer. The removal threshold is determined by adding or subtracting three times the standard deviation from the mean of historical data for the same period.

[0008] The elimination threshold is specifically obtained by performing sliding window statistical analysis on historical observation data of no less than 20 years and combining it with manual verification and iterative correction. Each iteration is adjusted by step size until the sum of the false negative rate and the false positive rate is minimized.

[0009] Specifically, the sparse Bayesian field reconstruction algorithm based on ocean dynamics constraints describes the reconstruction of the three-dimensional ocean environment field as a regularized inverse problem. The horizontal direction adopts the sparse prior corresponding to the geostrophic current constraint, the vertical direction adopts the static equilibrium constraint, the expectation propagation algorithm is used to approximate the non-Gaussian posterior, and the sparsity is achieved by determining the prior through automatic correlation.

[0010] Specifically, the ocean risk field manifold alignment and migration algorithm uses diffusion mapping to extract the low-dimensional intrinsic structure of long-historical observed aquaculture area data and newly established aquaculture area data respectively, uses Procrustes analysis to rotate, scale and align the manifolds of the two domains to obtain the cross-domain coordinate transformation matrix, and then uses the bottom shape data and tidal harmonic constant to correct the transformation error.

[0011] Specifically, the diffusion mapping involves constructing a Gaussian kernel matrix between observed samples, normalizing the kernel matrix, calculating eigenvalue decomposition, and taking the eigenvectors corresponding to the top few eigenvalues ​​as low-dimensional embedding coordinates. The diffusion distance is defined by the weighted Euclidean distance.

[0012] Specifically, the Procrustean analysis involves solving for the rotation matrix and scaling factor that minimize the mean square error between two point sets through singular value decomposition. The rotation matrix and scaling factor together constitute the cross-domain coordinate transformation matrix. The alignment quality is determined by the Riemannian metric preservation evaluation, and iterative optimization is performed until the difference between the curvature tensors of the two manifolds is minimized.

[0013] The location and scale parameters of the non-stationary generalized extreme value distribution are expressed as linear functions of the covariates, with the Pacific Decadal Oscillation Index and the El Niño-Southern Oscillation Index as covariates. The Bayesian Markov chain Monte Carlo sampling uses a No-U-Turn sampler, and the convergence is determined by the Gehrman-Rubin diagnostic.

[0014] The loss function of the physical constraint neural network incorporates the residual terms of the continuity equation and the momentum equation as physical regularization constraints, thereby enabling the reconstruction of a local high-resolution field with a horizontal resolution of 50m to 200m.

[0015] The marine aquaculture risk spatiotemporal prediction model uses marine observation stations and moored buoy nodes as a dynamic graph node set. The graph topology is reconstructed in real time from shore-based radar surface flow field data. The spatial layer uses a multi-head graph attention network, the temporal layer uses a temporal convolutional network, and a vine-style Copula joint distributed embedding layer is fused. The output layer uses a hybrid density head.

[0016] The multi-head graph attention network introduces a physically constrained attention mask to prohibit attention propagation against the ocean current direction. The physically constrained attention mask is determined by the cosine value threshold of the angle between the direction vector of the shore-based radar surface flow field data and the line connecting the node pairs.

[0017] Wherein, the dynamic resource scheduling function value Specifically, it is calculated by weighting the number of nodes in the current batch dynamic graph, the length of the input timing sequence, and the amount of available GPU memory, with the sum of each weighting coefficient being 1. Based on the interval to which the dynamic resource scheduling function value belongs, the GPU memory allocation ratio, the number of CUDA streams, and the training batch size are adjusted in segments.

[0018] Specifically, the risk warning level is triggered when the hourly risk index continuously exceeds the yellow warning threshold to reach the yellow duration threshold, triggers an orange warning when the duration of exceeding the orange warning threshold exceeds the orange duration threshold, and triggers a red warning when exceeding the red warning threshold.

[0019] The yellow, orange, and red warning thresholds are determined by analyzing historical aquaculture loss records and the subject operating characteristic curves of the hourly risk index for the corresponding time period, with the hourly risk index value corresponding to the maximum Youden index selected as the threshold for each level.

[0020] Specifically, the threshold for the Gehrman-Rubin diagnostic value is set to 1.1; the length of the Bayesian Markov chain Monte Carlo sampling chain is no less than 5000 steps, and the number of warm-up steps is no less than 1000 steps; the threshold for the cosine of the included angle is set to -0.2; the Copula joint distribution embedding layer encodes the multivariate correlation structure into a 32-dimensional risk covariance vector; the receptive field of the temporal convolutional network covers 72 time steps, and the inflation factor sequence is set to 1, 2, 4, and 8.

[0021] This invention uses a sparse Bayesian field reconstruction algorithm based on ocean dynamics constraints to reconstruct a three-dimensional ocean environment field. It then uses a marine risk field manifold alignment migration algorithm to migrate the risk benchmark distribution of aquaculture areas with long history observations along the manifold coordinates to newly established aquaculture areas. At the same time, it uses bottom shape data and tidal harmonic constants to physically correct the migration results. Finally, it drives the marine aquaculture risk spatiotemporal prediction model to output hourly risk index sequences and probability distributions.

[0022] Traditional methods of directly transplanting risk parameters neglect the fundamental differences in ocean dynamic geometry between two regions, leading to inconsistencies between the migrated distribution and the physical environment of the target region. This invention extracts the low-dimensional intrinsic structure of the data from both domains through diffusion mapping, and then uses Procrustes analysis to achieve manifold rotation and scaling alignment. This ensures that risk statistics are transmitted along physically meaningful manifold coordinates. Simultaneously, Riemannian metric preservation is introduced to iteratively optimize the alignment quality, guaranteeing geometric consistency during the migration process. Furthermore, the transformation error is corrected using bottom topography data from the newly established aquaculture area and tidal harmonic constants, ensuring that the migrated risk baseline distribution inherits both the extreme statistical characteristics of the long-historical region and retains the local dynamic characteristics of the target region.

[0023] In summary, this invention solves the technical problem mentioned in the background art that newly established aquaculture areas cannot establish a reliable risk benchmark distribution due to the lack of long-term historical observation data. Attached Figure Description

[0024] Figure 1 This is a flowchart of the method of the present invention.

[0025] Figure 2This is a schematic diagram of the low-dimensional embedding coordinate distribution of the two domains before and after manifold alignment migration.

[0026] Figure 3 A schematic diagram of the local high-resolution effective wave height field distribution reconstructed by a physically constrained neural network.

[0027] Figure 4 This is a schematic diagram showing the hourly risk index sequence and hourly risk index probability distribution during the typhoon's passage.

[0028] Figure 5 This diagram illustrates the correspondence between the triggering sequence of each warning level and the hourly risk index sequence.

[0029] Figure 6 A schematic diagram of the output structure of marine aquaculture insurance risk index products. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0031] like Figure 1 The diagram shown is a flowchart of a marine aquaculture risk prediction method based on multi-source marine observation data provided by the present invention. This method includes the following steps: S01. Collect ocean station temperature, salinity and depth profile data, moored buoy wave height and current velocity data, shore-based radar surface current field data, and satellite remote sensing sea surface temperature and significant wave height data. After quality control, use a sparse Bayesian field reconstruction algorithm based on ocean dynamics constraints to reconstruct the three-dimensional ocean environment field and output the posterior mean field and posterior variance field of the three-dimensional ocean environment field. S02. Input the posterior mean field and the posterior variance field of the three-dimensional marine environmental field into the marine risk field manifold alignment and transfer algorithm to perform cross-domain risk benchmark alignment between newly established aquaculture areas and long-historical observed aquaculture areas, and output the target domain risk benchmark distribution. S03. Using the target domain risk benchmark distribution as input, the extreme values ​​of sea surface temperature and significant wave height are statistically modeled using the non-stationary generalized extreme value distribution. The posterior distribution of the non-stationary generalized extreme value distribution parameters is obtained by combining Bayesian Markov chain Monte Carlo sampling, and the uncertainty interval of the risk index is output. S04. Using the temperature, salinity, current velocity, and significant wave height output by the regional ocean model as background field boundary conditions, a physical constraint neural network is driven to complete the local high-resolution field reconstruction with a horizontal resolution of 50m to 200m. The reconstruction results and the uncertainty interval of the risk index are input into the spatiotemporal prediction model of marine aquaculture risk, and the hourly risk index sequence and hourly risk index probability distribution of the aquaculture area are output. S05. Calculate the dynamic resource scheduling function value, adjust the memory allocation ratio and batch size of the marine aquaculture risk spatiotemporal prediction model according to the interval to which the dynamic resource scheduling function value belongs, and trigger the corresponding risk warning level and execute the corresponding process according to the comparison results of the hourly risk index sequence with the yellow warning threshold, orange warning threshold and red warning threshold. S06. The hourly risk index sequence, hourly risk index probability distribution, risk index uncertainty range, and risk warning level are output together to form a marine aquaculture insurance risk index product, which is used for insurance product design and compensation trigger determination.

[0032] The quality control steps employ three methods—climatological range testing, internal consistency testing, and spatiotemporal continuity testing—to progressively screen out outliers. The elimination threshold is determined by adding or subtracting three times the standard deviation from the mean of historical marine station temperature, salinity, and depth profile data, moored buoy wave height and current velocity data, shore-based radar surface current field data, and satellite remote sensing sea surface temperature and significant wave height data. The elimination threshold is obtained through sliding window statistical analysis of at least 20 years of historical observation data, combined with manual verification and iterative correction. Each iteration adjusts the threshold by 5% increments from the previous one until the sum of the false alarm rate and the missed alarm rate is minimized.

[0033] The sparse Bayesian field reconstruction algorithm based on ocean dynamics constraints describes the reconstruction of the three-dimensional ocean environment field as a regularized inverse problem. The prior distribution is taken from ocean dynamic constraints, with sparse priors corresponding to geostrophic constraints in the horizontal direction and hydrostatic constraints in the vertical direction. The boundary conditions are determined by coastline topographic data. The observation equations are the measurement models of the ocean station temperature, salinity, and depth profile data sensors, moored buoy sensors, shore-based radar sensors, and satellite remote sensing sensors, including instrument noise variance. The expectation propagation algorithm is used to approximate the non-Gaussian posterior. Sparsity is achieved by automatically determining the prior through correlation. The model automatically identifies areas with sufficient information and areas that require strong physical constraints to fill. The algorithm outputs the posterior mean field and the posterior variance field of the three-dimensional ocean environment field, and the posterior variance field of the three-dimensional ocean environment field simultaneously satisfies the mass conservation test. The sparse Bayesian field reconstruction algorithm based on ocean dynamics constraints fills the information gaps in sparse observation regions with ocean dynamic geometry, making the reconstructed field both statistically consistent and physically interpretable. The posterior variance field of the three-dimensional ocean environment field directly provides a quantitative basis for spatial uncertainty in insurance verification, avoiding the accumulation of systematic biases in traditional interpolation methods under sparse observation conditions, and improving the credibility of risk field reconstruction at the local scale of aquaculture areas.

[0034] The geostrophic flow constraint is the balance condition between pressure gradient force and Coriolis force; the static equilibrium constraint is the balance condition between vertical pressure gradient and buoyancy; the automatic correlation prior is a sparsified prior that assigns independent hyperparameters to each spatial basis function, and the hyperparameters are automatically determined by maximizing the marginal likelihood during the expectation propagation algorithm iteration process; the expectation propagation algorithm approximates the non-Gaussian posterior factor by factor into a Gaussian distribution family, iteratively updates the moment parameters of each factor, and obtains a global posterior approximation after convergence.

[0035] The marine risk field manifold alignment and migration algorithm treats the multivariate observation data of the long-historical aquaculture area and the multivariate observation data of the newly established aquaculture area as point sets on a high-dimensional manifold. First, it uses diffusion mapping to extract the low-dimensional intrinsic structure of the data from the long-historical aquaculture area and the newly established aquaculture area, and calculates the diffusion distance matrix. Then, it uses Procrustes analysis to rotate, scale, and align the manifolds of the two domains to obtain the cross-domain coordinate transformation matrix. The risk benchmark distribution of the long-historical aquaculture area is transferred to the newly established aquaculture area along the aligned manifold coordinates, while the bottom shape data and tidal harmonic constant of the newly established aquaculture area are used to correct the error after transformation. The alignment quality is evaluated by Riemannian metric preservation, and the cross-domain coordinate transformation matrix is ​​iteratively optimized until the difference in the curvature tensor of the two manifolds is minimized, and the target domain risk benchmark distribution is output. The proposed marine risk field manifold alignment and migration algorithm extrapolates statistical information from long-historical aquaculture areas along the ocean dynamic geometry in a physically meaningful manner. This allows the target domain risk benchmark distribution to inherit the extreme statistical characteristics of long-historical aquaculture areas, while preserving the physical corrections to the local bottom shape and tidal dynamics of newly established aquaculture areas. This avoids the dynamic inconsistencies caused by directly transplanting risk parameters and improves the initial reliability of risk assessment for newly established aquaculture areas.

[0036] The calculation steps of the diffusion mapping are as follows: construct a kernel matrix between observed samples, using a Gaussian kernel as the kernel function, and determining the bandwidth parameter by the square of the median distance between samples; after normalizing the kernel matrix, calculate its eigenvalue decomposition, and take the eigenvectors corresponding to the top few largest eigenvalues ​​as low-dimensional embedding coordinates; the diffusion distance is defined by the weighted Euclidean distance of each sample in the low-dimensional embedding coordinate system. The Procrustean analysis is as follows: given two sets of points of the same dimension, solve for the rotation matrix and scaling factor that minimize the mean square error of the two point sets through singular value decomposition, and the rotation matrix and scaling factor together constitute the cross-domain coordinate transformation matrix. The Riemannian metric preservation evaluation is as follows: calculate the geodesic distance ratio between the sampling point pairs on the manifold of the two domains before and after alignment, and determine alignment convergence when the variance of the geodesic distance ratio is lower than the convergence threshold; the convergence threshold is determined by repeated alignment experiments on no less than 30 sets of known paired aquaculture area historical data and statistically analyzing the inflection point of the geodesic distance ratio variance convergence curve.

[0037] The location and scale parameters of the non-stationary generalized extreme value distribution are expressed as linear functions of the Pacific Decadal Oscillation Index and the El Niño-Southern Oscillation Index as covariates, capturing the modulating effect of climate modes on the extreme value distribution. The Bayesian Markov chain Monte Carlo sampling uses a No-U-Turn sampler with a chain length of no less than 5000 steps and a preheating step count of no less than 1000 steps. Convergence is determined by the Gehrman-Rubin diagnostic measure. The threshold of the Gehrman-Rubin diagnostic measure is set to 1.1, which is a recognized statistical convergence standard. The posterior distribution of the output non-stationary generalized extreme value distribution parameters is used to calculate the uncertainty interval of the risk index.

[0038] The horizontal resolution of the regional ocean model is 5. ~25 The output temperature, salinity, flow velocity, and significant wave height are used as background field boundary conditions. Driven by the background field boundary conditions, the physical constraint neural network completes local high-resolution field reconstruction with a horizontal resolution of 50m to 200m. The loss function embeds the residual terms of the continuity equation and the residual terms of the momentum equation as physical regularization constraints.

[0039] The specific structure of the spatiotemporal prediction model for marine aquaculture risks is as follows: The model uses marine observation stations and moored buoy nodes as a dynamic graph node set. The graph topology is reconstructed in real time from the current shore-based radar surface flow field data. The edge weights are determined by the streamline integral connectivity and the horizontal vortex rotation angle. The graph topology is automatically reorganized when a typhoon passes. The spatial layer uses a multi-head graph attention network, introducing a physically constrained attention mask to prohibit attention propagation against the ocean current direction. The physically constrained attention mask is determined by the cosine threshold of the angle between the direction vector of the shore-based radar surface flow field data and the connection line between node pairs. The cosine threshold is set to -0.2. The temporal layer uses a temporal convolutional network with an expansion factor sequence of 1, 2, 4, and 8, covering 72 time steps. A Copula joint distributed embedding layer is fused to encode the multivariate correlation structure into a 32-dimensional risk covariance. Vectors are injected with dynamic graph node features; the graph topology is updated at each time step by the flow field prediction result output from the previous step, forming a dynamically self-consistent iterative cycle; the output layer adopts a hybrid density head, outputting a time-varying risk index probability distribution, and integrates extreme value regularization constraints, which use the negative log-likelihood of the generalized extreme value distribution as an additional loss term; the memory allocation strategy is as follows: the multi-head graph attention network module is allocated 40% to 55% of the total memory, the temporal convolutional network module is allocated 25% to 35%, the Copula joint distribution embedding layer is allocated 10% to 20%, and the hybrid density head is allocated 5% to 10%, and the above proportions are adjusted in real time by the dynamic resource scheduling function value; the CUDA stream allocation strategy is as follows: the graph topology reconstruction calculation and the forward propagation of the temporal convolutional network are allocated to different CUDA streams for parallel execution, and the number of CUDA streams is set to 2 to 4, the specific number of which is determined by the dynamic resource scheduling function value.

[0040] The steps for establishing the training dataset for the marine aquaculture risk spatiotemporal prediction model specifically include: collecting hourly historical observation sequences of at least 10 years of marine station temperature, salinity, and depth profile data, moored buoy wave height and current velocity data, shore-based radar surface current field data, and satellite remote sensing sea surface temperature and significant wave height data; constructing supervised sample pairs using historical aquaculture loss records as labels and corresponding hourly historical observation sequences as input; oversampling of historical periods of strong winds and waves, typhoon passage, and abnormal sea surface temperatures to ensure that the proportion of extreme event samples is not less than 20%, with the 20% lower limit determined by cross-validation experiments on the extreme value prediction bias of the marine aquaculture risk spatiotemporal prediction model under different oversampling ratios, with no less than 50 experimental groups; expanding the samples of newly established aquaculture areas using the marine risk field manifold alignment migration algorithm, mixing source domain migration samples and target domain measured samples at a 7:3 ratio; and splitting the data along the time axis, using the last two years' data as the test set, and the remainder as the training and validation sets.

[0041] The training steps for the spatiotemporal prediction model of marine aquaculture risks specifically include: the optimizer uses an adaptive moment estimation algorithm, and the initial learning rate is set to... ~ The weight decay coefficient is set to ~ The total loss function consists of a weighted sum of three terms: mixed density negative log-likelihood loss, extreme value regularization loss, and physical constraint residual loss. The weights of each term are determined by a hyperparameter grid search of the validation set loss. The training batch size is set to 32–128, determined by the available GPU memory and the dynamic resource scheduling function. After each training epoch, the continuous ranking probability score is evaluated on the validation set. Training is terminated when the continuous ranking probability score does not improve for 5 consecutive epochs. Cosine annealing learning rate scheduling is used, with a cycle length of 10–20 epochs.

[0042] The technical effects of the proposed marine aquaculture risk spatiotemporal prediction model are as follows: the dynamic graph topology enables the model to detect sudden changes in the surface flow field data of shore-based radar caused by typhoons; the physical constraint attention mask avoids the propagation of erroneous information in the reverse direction; the Copula joint distribution embedding layer captures the joint correlation structure of multivariate collaborative disasters; the parallel computing structure of the temporal convolutional network improves the training efficiency of long sequences; the hybrid density head provides an hourly risk index probability distribution rather than a single-point prediction; the extreme value regularization constraint ensures the physical rationality of the probability estimation of extreme events; and overall, the hourly risk index sequence has real-time performance, physical consistency, and statistical reliability of extreme events, providing a quantitative basis with confidence intervals for compensation trigger determination.

[0043] The cosine threshold of the included angle, -0.2, was determined through correlation analysis experiments on no fewer than 50 sets of historical flow field data and corresponding observation errors. Specifically, under different cosine thresholds, the increment of prediction error caused by attention propagation in the countercurrent direction was statistically analyzed, and the cosine value of the included angle corresponding to the inflection point where the increment of prediction error significantly increases was selected as the threshold. The dimension of the 32-dimensional risk covariance vector was determined by principal component analysis of multivariate observation data from different aquaculture areas. The minimum number of principal components when the cumulative variance contribution rate reached 95% was taken, and the mode was determined by taking the results of principal component analysis experiments on historical data from no fewer than 10 aquaculture areas.

[0044] Wherein, the dynamic resource scheduling function value The formula is expressed as follows: ;in This represents the number of nodes in the current batch of the dynamic graph. For reference node number, Input timing length, For reference timing length, This represents the current available video memory. This represents the total video memory of the device. , , The weighting coefficients and Take the average number of nodes in the dynamic graph of the training set. Take 72, , , Initial values ​​are all set to Through iterative adjustments using grid search experiments on memory overflow rate and computational utilization under no fewer than 100 different load scenarios; when At that time, the multi-head graph attention network module's memory allocation was set to 40%, the number of CUDA streams was set to 2, and the training batch size was set to 128; At that time, the multi-head graph attention network module's memory usage was set to 48%, the number of CUDA streams was set to 3, and the training batch size was set to 64; At that time, the memory usage ratio of the multi-head graph attention network module was set to 55%, the number of CUDA streams was set to 4, and the training batch size was set to 32; the thresholds of 0.4 and 0.7 were determined by piecewise fitting experiments on memory overflow probability curves under no less than 100 different computational loads.

[0045] The logic for switching risk warning levels is as follows: when the hourly risk index exceeds the yellow warning threshold for three consecutive time steps in the hourly risk index sequence, a yellow warning is triggered and the encrypted observation data collection process is initiated; when the hourly risk index exceeds the orange warning threshold and the duration exceeds six time steps, an orange warning is triggered and the insurance claim verification process is initiated; when the hourly risk index exceeds the red warning threshold, a red warning is triggered and a claim trigger report is automatically generated; the yellow, orange, and red warning thresholds are determined by analyzing the subject operating characteristic curves of historical aquaculture loss records and the hourly risk index for the corresponding period, and the hourly risk index value corresponding to the maximum Youden index is selected as the threshold for each level. The analysis is based on at least five years of historical paired data.

[0046] The Copula joint distribution embedding layer uses a vine-like Copula structure to decouple the multivariate marginal distributions. The marginal distribution parameters of each variable and the rank correlation coefficient matrix between variables are jointly encoded into a 32-dimensional risk covariance vector. This encoding is implemented by a fully connected layer, and the activation function is a linear rectified function. The continuous sorting probability score is an indicator that scores the integral error between the hourly risk index probability distribution and the hourly risk index sequence observations. The smaller the value, the higher the probability forecast quality, used to evaluate the distribution prediction accuracy of the mixed density head. The Youden index is the sum of sensitivity and specificity minus 1, a statistic for evaluating the comprehensive performance of the binary classification threshold. Maximizing the Youden index corresponds to the optimal classification threshold. The streamline integral connectivity is a scalar obtained by integrating the connectivity between two nodes along the streamlines of the coastal current field, reflecting the intensity of material transport between the two observation stations by the ocean current. The horizontal vorticity rotation angle is the arctangent of the ratio of horizontal velocity field divergence to curl, used to characterize the rotational characteristics of the local flow field and as a component of the dynamic graph edge weights.

[0047] Optionally, the present invention also provides a method for implementing a marine aquaculture risk prediction system based on multi-source marine observation data using a computer. The computer is equipped with a readable storage medium, which stores program instructions. When the program instructions are run on the computer, they can execute the above-described method.

[0048] The specific implementation of step S01 is as follows: Technicians first collect temperature, salinity, and depth profile data from oceanographic stations, wave height and current velocity data from moored buoys, surface current field data from shore-based radar, and sea surface temperature and significant wave height data from satellite remote sensing products. These four types of data are aligned with a unified timestamp before entering the quality control stage. Quality control sequentially employs three methods: climatological range verification, internal consistency verification, and spatiotemporal continuity verification. The elimination threshold is determined by adding or subtracting three times the standard deviation of the mean of at least 20 years of historical data for the same period. Iterative corrections are made through sliding window statistical analysis combined with manual verification, with each correction adjusting in 5% increments until the sum of the false negative and false positive rates is minimized. After quality control, multi-source observation data is input into a sparse Bayesian field reconstruction algorithm based on ocean dynamics constraints. This algorithm describes the reconstruction of the three-dimensional ocean environment field as a regularized inverse problem. The horizontal prior adopts a sparse prior corresponding to the geostrophic constraint, namely the balance condition between pressure gradient force and Coriolis force. The vertical prior adopts a static equilibrium constraint, namely the balance condition between vertical pressure gradient and buoyancy. The boundary conditions are determined by coastline topographic data. The observation equation includes instrument noise variance. Sparsity is achieved by automatically determining the prior through correlation and assigning independent hyperparameters to each spatial basis function. The hyperparameters are automatically determined by maximizing the marginal likelihood during the expectation propagation algorithm iteration. The expectation propagation algorithm approximates the non-Gaussian posterior factor by factor into a Gaussian distribution family, iteratively updating the moment parameters of each factor. After convergence, it outputs the posterior mean field and the posterior variance field of the three-dimensional ocean environment field. The latter also satisfies the mass conservation test, providing a quantitative basis for spatial uncertainty in subsequent steps.

[0049] The specific implementation of step S02 is as follows: Using the posterior mean and posterior variance fields of the three-dimensional marine environmental field output from step S01 as inputs, a marine risk field manifold alignment and transfer algorithm is executed. First, Gaussian kernel matrices are constructed for the multivariate observation data of both the long-historical aquaculture area and the newly established aquaculture area. The bandwidth parameter is determined by the square of the median distance between samples. After normalizing the kernel matrix, eigenvalue decomposition is performed, and the eigenvectors corresponding to the top few largest eigenvalues ​​are taken as low-dimensional embedding coordinates. The diffusion distance is defined by the weighted Euclidean distance of each sample in the low-dimensional embedding coordinate system. Subsequently, Procrustean analysis is used to rotate, scale, and align the manifolds of the two domains. The rotation matrix and scaling factor that minimize the mean square error of the two point sets are solved through singular value decomposition. These two factors together constitute the cross-domain coordinate transformation matrix. The risk benchmark distribution already calibrated in the long-historical aquaculture area is transferred to the newly established aquaculture area along the aligned manifold coordinates. Simultaneously, the transformation error is corrected using the bottom topography data and tidal harmonic constant of the newly established aquaculture area. Alignment quality is determined by the Riemannian metric preservation assessment, which calculates the geodesic distance ratio between the sampling point pairs of the two domain manifolds before and after alignment. Convergence is determined when the variance of the geodesic distance ratio is lower than the convergence threshold. The reference value of the convergence threshold is determined by the inflection point of the convergence curve of the variance of the geodesic distance ratio in repeated alignment experiments of no less than 30 sets of known paired aquaculture area historical data. Finally, the target domain risk benchmark distribution is output.

[0050] The specific implementation of step S03 is as follows: Using the target domain risk baseline distribution as input, non-stationary generalized extreme value distribution models are established for the extreme values ​​of sea surface temperature and significant wave height, respectively. Location and scale parameters are expressed as linear functions of the covariates, using the Pacific Decadal Oscillation Index and the El Niño-Southern Oscillation Index as covariates, to capture the modulating effect of climate modes on the extreme value distribution. Bayesian Markov chain Monte Carlo sampling uses a No-U-Turn sampler with a chain length of at least 5000 steps and a preheating step count of at least 1000 steps. Convergence is determined by the Gehrman-Rubin diagnostic metric, with a diagnostic threshold set to 1.1. After sampling convergence, the posterior distribution of the non-stationary generalized extreme value distribution parameters is obtained. The uncertainty interval of the risk index is obtained by integrating the posterior distribution, which is used to quantify the statistical reliability of the extreme event probability estimate.

[0051] The specific implementation of step S04 is as follows: Using temperature, salinity, current velocity, and significant wave height output from a regional ocean model (horizontal resolution 5km–25km) as background field boundary conditions, a physically constrained neural network is driven to reconstruct a local high-resolution field with a horizontal resolution of 50m–200m. The loss function incorporates residual terms from the continuity equation and momentum equation as physical regularization constraints to ensure the physical self-consistency of the reconstructed field. The reconstruction results and the uncertainty interval of the risk index output in step S03 are jointly input into the spatiotemporal prediction model for marine aquaculture risks. This model uses ocean observation stations and moored buoy nodes as a dynamic graph node set. The graph topology is reconstructed in real-time from the surface current field data from shore-based radar, and the edge weights are determined by the streamline integral connectivity and the horizontal vortex rotation angle. The spatial layer employs a multi-head graph attention network, introducing a physically constrained attention mask with a cosine threshold of -0.2 to prohibit attention propagation against the ocean current direction. The temporal layer uses a temporal convolutional network with an expansion factor sequence of 1, 2, 4, and 8, covering 72 time steps in the receptive field. The Copula joint distribution embedding layer encodes multivariate marginal distribution parameters and rank correlation coefficient matrices into a 32-dimensional risk covariance vector and injects it into the node features. The output layer employs a hybrid density head, incorporating extreme value regularization constraints, and outputs the hourly risk index sequence and the hourly risk index probability distribution.

[0052] The specific implementation of step S05 is: dynamic resource scheduling function value The result is calculated by weighted summation of three factors: the ratio of the number of nodes in the current batch dynamic graph to the number of reference nodes, the ratio of the input timing length to the reference timing length, and the memory consumption ratio. The reference timing length is set to 72, and the initial value of each weighting coefficient is set to 1 / 3. The results are iteratively adjusted through at least 100 sets of grid search experiments under various load scenarios. At that time, the multi-head graph attention network module's memory allocation was set to 40%, the number of CUDA streams was set to 2, and the training batch size was set to 128; At that time, the video memory usage was set to 48%, the number of CUDA streams was set to 3, and the batch size was set to 64; At that time, the VRAM allocation was set to 55%, the number of CUDA streams was set to 4, and the batch size was set to 32. The risk warning level was determined based on the hourly risk index sequence. A yellow warning was triggered if the risk index exceeded the yellow warning threshold for 3 consecutive time steps, an orange warning was triggered if the risk index exceeded the orange warning threshold for more than 6 time steps, and a red warning was triggered if the risk index exceeded the red warning threshold. Each threshold was determined by the maximum value of the Youden index in the analysis of the subject operating characteristic curve of at least 5 years of historical paired data.

[0053] The specific implementation of step S06 is as follows: the hourly risk index sequence, the hourly risk index probability distribution, the risk index uncertainty interval and the risk warning level are integrated and output to form a marine aquaculture insurance risk index product, which is used by the insurance product designer to determine the rate and as a quantitative basis for the determination of compensation trigger. The output content also includes the confidence interval to support the quantitative requirements of insurance verification for uncertainty.

[0054] It should be noted that the key technologies of this invention include: a sparse Bayesian field reconstruction algorithm based on ocean dynamics constraints fuses multi-source heterogeneous observation data into a physically consistent three-dimensional field and simultaneously outputs spatial uncertainty quantification information, overcoming the problem of systematic bias accumulation in sparse observation areas by traditional interpolation methods, thus providing a reliable input basis for subsequent statistical modeling; a marine risk field manifold alignment and transfer algorithm aligns the intrinsic structures of ocean dynamics in two domains in low-dimensional space through diffusion mapping and Procrustean analysis, enabling risk statistics to be transmitted along physically meaningful manifold coordinates, while compensating for local dynamic differences with bottom shape and tidal corrections, fundamentally solving the problem of insufficient data for establishing risk benchmarks in newly established aquaculture areas; and a marine aquaculture risk spatiotemporal prediction model, through the collaborative design of dynamic graph topology, physically constrained attention mask, vine-like Copula joint distribution embedding layer, and hybrid density head, realizes joint probabilistic modeling of multivariate collaborative disasters and real-time flow field perception, outputting an hourly risk index with confidence intervals. The above three key technologies form a complete data acquisition, knowledge transfer, and probability prediction chain. They are interdependent and indispensable, and together they ensure the physical consistency and statistical reliability of the entire process from observation data to insurance product output.

[0055] It should be noted that this invention also solves the following technical problem: In existing marine environmental monitoring, due to differences in sensor type, sampling frequency, and spatial coverage, direct fusion of multi-source heterogeneous observation data leads to the accumulation of systematic biases caused by information imbalance. Traditional interpolation methods lack physical constraints in sparse observation areas, resulting in insufficient reliability of the reconstructed field at local scales. This invention utilizes a sparse Bayesian field reconstruction algorithm based on ocean dynamics constraints. It embeds geostrophic and hydrostatic constraints into a sparse Bayesian framework, uses automatic correlation to determine prior information-rich areas and areas requiring strong physical constraints for filling. In sparse observation areas, ocean dynamic geometry is used to fill information gaps. The expectation propagation algorithm ensures the approximate accuracy of the non-Gaussian posterior, and the output three-dimensional marine environmental field posterior variance field satisfies the quality conservation test, providing quantifiable spatial uncertainty evidence for insurance verification. This solves the technical problem of systematic bias accumulation in sparse areas of the reconstructed field due to the lack of physical constraints during multi-source heterogeneous observation data fusion.

[0056] Specifically, the principle of this invention is as follows: The fundamental reason why this invention can solve the above-mentioned technical problems lies in the fact that the core assumption of manifold alignment migration is consistent with the intrinsic structure of ocean dynamics. High-dimensional observational data of ocean environmental variables are not uniformly distributed in high-dimensional space, but are concentrated on low-dimensional manifolds determined by the constraints of dynamic equations. The data manifolds of different aquaculture areas have alignable geometric correspondences in their intrinsic dynamic structures. Diffusion mapping captures the global topological structure of the data through kernel matrix eigenvalue decomposition. The extracted low-dimensional coordinates reflect the intrinsic modes of ocean dynamics, rather than statistical noise. Therefore, when using these coordinates as a carrier to transmit risk statistics, the transmission path itself has physical interpretability. Procrustean analysis solves for optimal rotation and scaling through singular value decomposition, ensuring that the alignment error of the two manifolds in low-dimensional space is minimized, while Riemannian metric preservation assessment verifies the geometric fidelity of the alignment from the curvature tensor level. Bottom shape and tidal corrections further compensate for residual errors caused by local dynamic differences. The above mechanisms together ensure the dynamic consistency of the risk benchmark distribution during the migration process, which is precisely what traditional parameter translation methods lack. Therefore, the solution of this invention can logically solve the fundamental difficulty of establishing risk benchmarks in newly established aquaculture areas.

[0057] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0058] The specific implementation of step S01 is as follows: Data on temperature, salinity, and depth profiles from marine stations, wave height and current velocity data from moored buoys, surface current field data from shore-based radar, and sea surface temperature and significant wave height data from satellite remote sensing are collected. After quality control, these data are input into a sparse Bayesian field reconstruction algorithm based on ocean dynamics constraints. Quality control employs three methods: climatological range testing, internal consistency testing, and spatiotemporal continuity testing. The rejection threshold is determined by adding or subtracting three times the standard deviation from the mean of historical data from the same period. This threshold is obtained through sliding window statistical analysis of at least 20 years of historical observation data, combined with manual verification and iterative correction. Each iteration adjusts the threshold by a 5% step size until the sum of the false negative and false positive rates is minimized. The three-dimensional ocean environmental field reconstruction is expressed as a regularized inverse problem. Let the observation vector be... The state vector to be reconstructed is The observation equation is: ; In the formula, To synthesize the observation operator matrices of various sensor measurement models, This is the instrument noise vector. The instrument noise covariance matrix is ​​a diagonal matrix composed of the nominal noise variances of each sensor. The prior distribution is determined by ocean dynamic constraints, with geostrophic constraints applied in the horizontal direction, i.e., the pressure gradient force and the Coriolis force are balanced. ; In the formula, Coriolis parameters, in units of , This is the horizontal velocity vector, in units of... , This is a reference value for horizontal flow velocity, in units of... This is used to normalize both sides. For reference density, the unit is... , This represents the horizontal pressure gradient, in units of... This constraint is transformed into a precision matrix of sparse priors. A vertical static equilibrium constraint is applied, meaning the vertical pressure gradient is balanced with the buoyancy: ; In the formula, Pressure, unit: , Vertical coordinates, in units of , The density of seawater is expressed in units of 1000 kJ / m³. , This is a density reference value, in units of Used for two-sided normalization. This is the acceleration due to gravity, in units of 1. This constraint is transformed into a vertical accuracy matrix. Automatic correlation determines priors for each spatial basis function. ( Assign independent hyperparameters The prior accuracy matrix is hyperparameters The marginal likelihood is automatically determined by maximizing the marginal likelihood during the iterative process of the expectation propagation algorithm. The expectation propagation algorithm approximates the non-Gaussian posterior factorally as a Gaussian distribution family for each factor. Use Gaussian factor Approximate, where For the first The normalization constant of each factor, For the first The mean parameter of each approximation factor, For the first The variance parameter of each approximate factor is used to iteratively update the factors. and After convergence, the global posterior approximation is: The posterior mean and posterior variance are respectively: ; ; In the formula, Let be the prior mean vector. This is the output posterior mean field of the three-dimensional marine environmental field. This is the output three-dimensional marine environmental field posterior variance field, which simultaneously satisfies the mass conservation test.

[0059] The specific implementation of step S02 is as follows: the posterior mean field of the three-dimensional ocean environment field is... With posterior variance field Input the ocean risk field manifold alignment migration algorithm to analyze long-history observed aquaculture areas (source domain). ) and newly built aquaculture areas (target area) Cross-domain risk benchmark alignment is performed. First, the source domain sample sets are aligned. With the target domain sample set Construct a Gaussian kernel matrix, with the following elements: ; In the formula, and The first The and the first The feature vector of each sample point The Gaussian kernel bandwidth parameter is determined by the square of the median distance between samples. For the kernel matrix Line number Column elements. After normalizing the rows of the kernel matrix, calculate the eigenvalue decomposition and take the first few elements. The eigenvectors corresponding to the large eigenvalues As low-dimensional embedding coordinates, where For the sample size, To preserve the low-dimensional embedding dimension. The diffusion distance is defined as... ,in and The first The and the first The coordinate vector of each sample in the low-dimensional embedding coordinate system Let be the weighted Euclidean distance norm with corresponding feature values ​​as weights. The low-dimensional embedding coordinates of the source and target domains are denoted as . and By performing rotation, scaling, and alignment on the two-domain manifold using Procrustean analysis, the following results are obtained. Perform singular value decomposition: ; In the formula, and It is an orthogonal matrix. For diagonal singular value matrices, rotation matrices scaling factor ,in For matrix trace operations, For the Frobenius norm, the cross-domain coordinate transformation matrix The source domain's calibrated risk baseline distribution is transferred to the target domain along the aligned manifold coordinates, and the transformation error is corrected using the target domain's base shape data and tidal harmonic constant. Riemannian metric preservation is assessed by calculating the ratio of geodesic distances to manifold sampling points in the two domains before and after alignment. ,in and For the target domain manifold For sampling points, and These are the geodesic distances on the manifolds of the target and source domains, respectively, when the ratio variance Time-based alignment convergence determination, convergence threshold The inflection point of the variance convergence curve of the geodesic distance ratio is determined by iterative optimization in repeated alignment experiments using no fewer than 30 sets of known historical data from paired aquaculture areas. Until convergence, output the target domain risk baseline distribution.

[0060] The specific implementation of step S03 is as follows: using the target domain risk benchmark distribution as input, statistical modeling is performed on the extreme values ​​of sea surface temperature and significant wave height using a non-stationary generalized extreme value distribution, and location parameters are used. With scale parameters Pacific Decadal Oscillation Index With El Niño-Southern Oscillation Index As covariates, the linear expression is: ; ; In the formula, These are reference values ​​for the position parameters, used to normalize both sides of the position parameter equation. For the position parameter intercept term, , These are the regression coefficients of the covariates corresponding to the location parameters. This is a reference value for the scale parameter. The scale parameter is the logarithmic linear intercept term. , The scaling parameter corresponds to the logarithmic linear regression coefficients of the covariates. and These are the reference standard values ​​for the Pacific Decadal Oscillation Index and the El Niño-Southern Oscillation Index, respectively. For time index, shape parameter It is considered a time-independent constant. A no-back-up sampler is used for Bayesian Markov chain Monte Carlo sampling, with a chain length of at least 5000 steps and a warm-up step count of at least 1000 steps. Convergence is determined by the Gehrman-Rubin diagnostic measure. The decision is made with a threshold set to 1.1, and the posterior distribution of the output parameters is determined. Then, the uncertainty range of the risk index is calculated.

[0061] The specific implementation of step S04 is as follows: using the temperature, salinity, current velocity, and significant wave height output from a regional ocean model (horizontal resolution 5–25 km) as the background field boundary conditions. The physical constraint neural network is used to reconstruct local high-resolution fields with a horizontal resolution of 50m to 200m. The loss function is: ; In the formula, For the total loss, For data fitting loss, For the residual loss of the continuous equation, This represents the residual loss in the momentum equation. This serves as a reference value for loss, used to normalize various parameters. This is the normalized reference residual value for the continuity equation. The normalized reference residual value for the momentum equation is denoted as , both of which are dimensionless positive real numbers used to balance the relative weights of the loss terms. The reconstruction results and the uncertainty interval of the risk index are input into the spatiotemporal prediction model for marine aquaculture risk, outputting an hourly risk index sequence and probability distribution.

[0062] The specific implementation of step S05 is: dynamic resource scheduling function value The formula is expressed as follows: ; In the formula, This represents the number of nodes in the current batch of the dynamic graph. The average number of nodes in the dynamic graph of the training set is used as the reference number. Input timing length, The experience value is 72. This represents the current available video memory. This represents the total video memory of the device. , , The weighting coefficients are and satisfy the following conditions: The initial values ​​were all 1 / 3, and were iteratively adjusted through grid search experiments on memory overflow rate and computational utilization under no less than 100 different load scenarios. At that time, the multi-head graph attention network module's memory allocation was set to 40%, the number of CUDA streams was set to 2, and the training batch size was set to 128; At that time, the video memory usage was set to 48%, the number of CUDA streams was set to 3, and the batch size was set to 64; At that time, the video memory ratio was set to 55%, the number of CUDA streams was set to 4, and the batch size was set to 32. Thresholds of 0.4 and 0.7 were determined through piecewise fitting experiments on video memory overflow probability curves under no fewer than 100 different computational loads. The risk warning level switching logic is as follows: exceeding the yellow warning threshold for 3 consecutive time steps triggers a yellow warning; exceeding the orange warning threshold for more than 6 consecutive time steps triggers an orange warning; exceeding the red warning threshold triggers a red warning and automatically generates a compensation trigger report. Each threshold was determined by analyzing historical aquaculture loss records and the subject operating characteristic curves of the hourly risk index for the corresponding time period, using the Youden index. The risk index value corresponding to the maximum point is used as the threshold for each level, where... For sensitivity, For specificity, the analysis is based on at least 5 years of historical paired data.

[0063] The specific implementation of step S06 is as follows: the hourly risk index sequence, the hourly risk index probability distribution, the risk index uncertainty range, and the risk warning level are output together to form a marine aquaculture insurance risk index product, which is used for insurance product design and compensation trigger determination.

[0064] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: To illustrate the technical effects of the invention, technicians set up a test environment and selected a near-shore aquaculture area as a newly established aquaculture area. This area only has about one year of local observation records. The adjacent aquaculture area with 23 years of continuous observation records is used as a long-history observation aquaculture area. The method of this invention is used to predict the full-process risk of the newly established aquaculture area.

[0065] In step S01, technicians collected data from four types of observation sources: eight temperature, salinity, and depth (TDM) profile observation stations were deployed at the marine station, three sets of moored buoys were used, the shore-based radar covered a surface current field with a radius of approximately 80 km, and satellite remote sensing data included sea surface temperature and significant wave height products with a spatial resolution of approximately 4 km. Quality control employed climatological range checks to remove outliers significantly exceeding the climatological mean plus or minus three standard deviations; internal consistency checks to remove observation points with contradictory TDM-SDM relationships within the same station; and spatiotemporal continuity checks to remove data where gradients between adjacent time points or adjacent stations exceeded physically reasonable limits. The iterative correction step size was set to 5%, and the sum of the false alarm rate and the missed alarm rate converged to a minimum. After quality control, the four types of data were fed into a sparse Bayesian field reconstruction algorithm, outputting a posterior mean field and a posterior variance field of the three-dimensional marine environmental field covering the target area. The posterior variance field of the three-dimensional marine environmental field passed the quality conservation check, proving that the reconstructed field was physically self-consistent.

[0066] In step S02, Gaussian kernel matrices are constructed for the multivariate observation data of both the newly established aquaculture area and the long-historical aquaculture area. The bandwidth parameter is the square of the median distance between samples in each region. After row normalization of the kernel matrix, eigenvalue decomposition is performed, and the first 8 eigenvectors are extracted as low-dimensional embedding coordinates. Procrustean analysis obtains the rotation matrix and scaling factor through singular value decomposition, forming a cross-domain coordinate transformation matrix. Riemannian metric preservation assessment calculates the ratio of geodesic distances between sampling points in the two domains before and after alignment. After 7 iterations, the variance of the geodesic distance ratio is lower than the convergence threshold, indicating alignment convergence. The transformation error is corrected using the bottom topography data of the newly established aquaculture area and 8 tidal harmonic constants, and the final output is the target domain risk baseline distribution, such as... Figure 2 The diagram shows the distribution of low-dimensional embedded coordinates of the two domains before and after manifold alignment and migration. It can be seen that the overlap of the coordinate point sets of the two domains in the low-dimensional space is significantly improved after alignment.

[0067] In step S03, using the target domain risk baseline distribution as input, non-stationary generalized extreme value distributions are established for the annual maximum sea surface temperature and the annual maximum significant wave height, respectively. The location and scale parameters are both covariated using the Pacific Decadal Oscillation Index and the El Niño-Southern Oscillation Index. The No-U-Turn sampler runs for 6000 steps with a 1500-step warm-up. The Gehrman-Rubin diagnostic values ​​for all four independent chains converge to below 1.1, indicating convergence. The relevant statistics for the posterior distribution of the parameters are shown in Table 1. Table 1. Posterior statistics of nonstationary generalized extreme value distribution parameters

[0068] As shown in Table 1, the posterior standard deviations of each parameter are relatively small compared to the mean, indicating that the target domain risk benchmark distribution has sufficient information after manifold alignment and transfer, supporting robust estimation of extreme value parameters, thus outputting a reliable risk index uncertainty interval.

[0069] In step S04, the horizontal resolution of the regional ocean model is set to 10 km, and the output temperature, salinity, current velocity, and significant wave height are used as background field boundary conditions. The physically constrained neural network uses the residual terms of the continuity equation and the momentum equation as physical regularization constraints to reconstruct a local high-resolution field with a horizontal resolution of 100 m, such as... Figure 3The diagram shows a local high-resolution significant wave height field distribution reconstructed by a physically constrained neural network. The reconstructed field exhibits a spatial gradient consistent with the bottom structure within the aquaculture area, demonstrating the effective regularization effect of physical constraints. The marine aquaculture risk spatiotemporal prediction model constructs a graph structure using 11 dynamic graph nodes (8 marine stations and 3 moored buoys). The graph topology is reconstructed in real-time from shore-based radar surface current field data. The threshold value of the cosine of the physical constraint attention mask angle is set to -0.2 to suppress the propagation of erroneous information in the countercurrent direction. The Fuji-style Copula joint distribution embedding layer encodes the marginal distribution parameters and rank correlation coefficient matrix of three variables—sea surface temperature, significant wave height, and current velocity—into a 32-dimensional risk covariance vector. The hybrid density head fuses extreme value regularization constraints, outputting an hourly risk index sequence, such as... Figure 4 The figure shows the hourly risk index sequence and the hourly risk index probability distribution during the passage of a typhoon. It can be seen that the risk index shows a significant peak before and after the typhoon makes landfall, and the probability distribution width expands as uncertainty increases, reflecting the model's dynamic response capability to extreme events.

[0070] In step S05, taking a continuous 72-hour period of strong winds and waves as an example, the number of nodes in the dynamic graph was 11, the input timing length was 72, and the available video memory was approximately 35% of the total video memory of the device. The dynamic resource scheduling function value was then calculated. The value is approximately 0.72, falling within the third interval. The multi-head graph attention network module's memory usage is automatically set to 55%, the number of CUDA streams is set to 4, and the training batch size is set to 32. The warning judgment results are shown in Table 2. Table 2 Risk Warning Level Trigger Records During Strong Wind and Wave Passage

[0071] As shown in Table 2, the risk warning level dynamically switches according to the wind and wave intensity. The orange warning is triggered after the duration exceeds 6 steps, which is consistent with the design logic. Figure 5 The diagram shows the correspondence between the triggering sequence of each warning level and the hourly risk index sequence.

[0072] In step S06, the final output marine aquaculture insurance risk index product includes four types of information: hourly risk index sequence, hourly risk index probability distribution, risk index uncertainty interval, and risk warning level. Figure 6 The diagram shows the output structure of the marine aquaculture insurance risk index product. The product output also marks the confidence interval for each time step, providing a quantitative reference with statistical basis for determining the payout trigger.

[0073] Compared to traditional methods, this invention achieves several technological advancements. Traditional methods, when directly transplanting risk parameters from neighboring regions, neglect the fundamental differences in the ocean dynamic geometry between the two regions, leading to a mismatch between the risk baseline distribution and the physical environment of the target area. This invention, however, uses manifold alignment migration to transfer risk statistics along physically meaningful manifold coordinates. The geometric consistency of the migration process is guaranteed by Riemannian metric preservation assessment, thus enabling the risk baseline distribution of newly established aquaculture areas to possess statistical reliability comparable to that of long-historical regions, without waiting for years of local observation data accumulation. Furthermore, traditional extreme value statistical methods typically employ the stationarity assumption, neglecting the modulating effect of climate modes such as the Pacific Decadal Oscillation Index and the El Niño-Southern Oscillation Index on the extreme value distribution. This results in systematic biases in extreme value probability estimation during climate transition periods. In contrast, the non-stationary generalized extreme value distribution of this invention uses climate indices as covariates, directly capturing this modulating relationship and improving the physical rationality of extreme event probability estimation.

[0074] It should be noted that the variables involved in this invention are explained in detail in Tables 3 and 4.

[0075] Table 3. Variable Explanation Table (Part 1)

[0076] Table 4. Variable Explanation Table (Part Two)

[0077] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting marine aquaculture risks based on multi-source marine observation data, characterized in that, Includes the following steps: Data on temperature, salinity, and depth profiles from marine stations, wave height and current velocity data from moored buoys, surface current field data from shore-based radar, and sea surface temperature and significant wave height data from satellite remote sensing were collected. After quality control, a sparse Bayesian field reconstruction algorithm based on ocean dynamics constraints was used to reconstruct the three-dimensional ocean environment field, and the posterior mean field and posterior variance field of the three-dimensional ocean environment field were output. The posterior mean field and posterior variance field of the three-dimensional marine environmental field are input into the marine risk field manifold alignment and transfer algorithm to perform cross-domain risk benchmark alignment between newly established aquaculture areas and long-historical observed aquaculture areas, and output the target domain risk benchmark distribution. Using the target domain risk baseline distribution as input, the extreme values ​​of sea surface temperature and significant wave height are statistically modeled using a non-stationary generalized extreme value distribution. The posterior distribution of the non-stationary generalized extreme value distribution parameters is obtained by combining Bayesian Markov chain Monte Carlo sampling, and the uncertainty interval of the risk index is output. Using temperature, salinity, current velocity, and significant wave height output from a regional ocean model as background field boundary conditions, a physically constrained neural network is driven to complete local high-resolution field reconstruction. The reconstruction results and the uncertainty interval of the risk index are input into the spatiotemporal prediction model of marine aquaculture risk, and the hourly risk index sequence and hourly risk index probability distribution of the aquaculture area are output. Calculate the dynamic resource scheduling function value, adjust the memory allocation ratio and batch size of the marine aquaculture risk spatiotemporal prediction model according to the interval to which the dynamic resource scheduling function value belongs, and trigger the corresponding risk warning level and execute the corresponding process based on the comparison results of the hourly risk index sequence with the yellow warning threshold, orange warning threshold and red warning threshold. The hourly risk index sequence, hourly risk index probability distribution, risk index uncertainty range, and risk warning level are output together to form a marine aquaculture insurance risk index product, which is used for insurance product design and compensation trigger determination.

2. The marine aquaculture risk prediction method based on multi-source marine observation data according to claim 1, characterized in that, The quality control specifically employs three methods—climatological range testing, internal consistency testing, and spatiotemporal continuity testing—to screen out outliers layer by layer. The removal threshold is determined by adding or subtracting three times the standard deviation from the mean of historical data for the same period.

3. The marine aquaculture risk prediction method based on multi-source marine observation data according to claim 2, characterized in that, The elimination threshold is specifically obtained by performing sliding window statistical analysis on historical observation data of no less than 20 years and combining it with manual verification and iterative correction. Each iteration is adjusted by step size until the sum of the false negative rate and the false positive rate is minimized.

4. The marine aquaculture risk prediction method based on multi-source marine observation data according to claim 3, characterized in that, The sparse Bayesian field reconstruction algorithm based on ocean dynamics constraints specifically describes the reconstruction of the three-dimensional ocean environment field as a regularized inverse problem. The horizontal direction adopts the sparse prior corresponding to the geostrophic current constraint, the vertical direction adopts the static equilibrium constraint, the expectation propagation algorithm is used to approximate the non-Gaussian posterior, and the sparsity is achieved by automatically determining the prior through correlation.

5. The marine aquaculture risk prediction method based on multi-source marine observation data according to claim 4, characterized in that, The ocean risk field manifold alignment and migration algorithm specifically uses diffusion mapping to extract the low-dimensional intrinsic structure of long-historical observed aquaculture area data and newly established aquaculture area data, respectively. It then uses Procrustes analysis to rotate, scale, and align the manifolds of the two domains to obtain a cross-domain coordinate transformation matrix. Finally, it uses the base shape data and tidal harmonic constant to correct the transformation error.

6. The marine aquaculture risk prediction method based on multi-source marine observation data according to claim 5, characterized in that, The diffusion mapping specifically involves constructing a Gaussian kernel matrix among the observed samples, normalizing the kernel matrix, calculating the eigenvalue decomposition, and taking the eigenvectors corresponding to the top few eigenvalues ​​as low-dimensional embedding coordinates. The diffusion distance is defined by the weighted Euclidean distance.

7. The marine aquaculture risk prediction method based on multi-source marine observation data according to claim 6, characterized in that, The Prokrustes analysis specifically involves solving for the rotation matrix and scaling factor that minimize the mean square error between two point sets through singular value decomposition. The rotation matrix and scaling factor together constitute the cross-domain coordinate transformation matrix. The alignment quality is determined by the Riemannian metric preservation evaluation, and iterative optimization is performed until the difference between the curvature tensors of the two manifolds is minimized.

8. The marine aquaculture risk prediction method based on multi-source marine observation data according to claim 7, characterized in that, The location and scale parameters of the non-stationary generalized extreme value distribution are expressed as linear functions of the covariates, with the Pacific Decadal Oscillation Index and the El Niño-Southern Oscillation Index as covariates. The Bayesian Markov chain Monte Carlo sampling uses a No-U-Turn sampler, and the convergence is determined by the Gehrman-Rubin diagnostic.

9. The marine aquaculture risk prediction method based on multi-source marine observation data according to claim 8, characterized in that, The loss function of the physical constraint neural network embeds the residual terms of the continuity equation and the residual terms of the momentum equation as physical regularization constraints, thereby completing the reconstruction of local high-resolution fields with a horizontal resolution of 50m to 200m.

10. The marine aquaculture risk prediction method based on multi-source marine observation data according to claim 9, characterized in that, The proposed marine aquaculture risk spatiotemporal prediction model uses marine observation stations and moored buoy nodes as a dynamic graph node set. The graph topology is reconstructed in real time from shore-based radar surface flow field data. The spatial layer uses a multi-head graph attention network, the temporal layer uses a temporal convolutional network, and a vine-style Copula joint distributed embedding layer is fused. The output layer uses a hybrid density head.

Images