A method for evaluating deep-sea resource and environment carrying capacity based on spatiotemporal big data technology

Abstract

This invention provides a method for assessing the carrying capacity of deep-sea resources and environment based on spatiotemporal big data technology, belonging to the technical field of deep-sea resource and environmental carrying capacity assessment. This invention extracts multi-scale ocean dynamic mode data using a tidal harmonic-empirical mode hybrid decomposition algorithm, reconstructs a three-dimensional flow field using a domain decomposition parallel finite element method, applies non-stationary spatiotemporal Gaussian process regression for spatiotemporal interpolation, constructs a modular biogeochemical model to simulate nutrient cycling, inputs the processed data into an artificial intelligence-based deep-sea environmental carrying capacity assessment model, optimizes the model structure using a dynamic topology reconstruction sparse connectivity learning framework, and adjusts the pruning probability threshold using a topology adjustment function based on the comprehensive carrying capacity evaluation index. This solves the technical problem of achieving high-precision assessment and adaptive optimization of the model structure under multi-source heterogeneous ocean data conditions in deep-sea resource and environmental carrying capacity assessment models.

Description

Technical Field

[0001] This invention belongs to the field of deep-sea resource and environmental carrying capacity assessment technology, specifically, it relates to a method for assessing deep-sea resource and environmental carrying capacity based on spatiotemporal big data technology. Background Technology

[0002] Assessing the carrying capacity of deep-sea resources and the environment is a crucial foundation for marine resource development and ecological protection. Traditional assessment methods primarily rely on statistical models or empirical formulas, combining limited field observation data with expert scoring systems to determine carrying capacity. These methods are used in resource development planning in nearshore shallow water areas. However, traditional methods have significant drawbacks. Firstly, the complex and variable deep-sea environment, coupled with the vast differences in spatiotemporal resolution and physical quantity types between moored moorings and remote sensing satellite data, makes it difficult for traditional statistical models to effectively integrate this heterogeneous data, leading to biased assessment results. Secondly, existing assessment models employ fixed network topologies, failing to dynamically adjust model parameters and connectivity based on the environmental characteristics and data quality of different sea areas. Furthermore, their generalization ability is insufficient when training data is limited or unevenly distributed. In other words, existing technologies face the technical challenge of achieving high-precision assessment and adaptive optimization of model structure under conditions of multi-source heterogeneous marine data. Summary of the Invention

[0003] In view of this, the present invention provides a method for assessing the carrying capacity of deep-sea resources and environment based on spatiotemporal big data technology, which can solve the technical problem that existing deep-sea resource and environmental carrying capacity assessment models are difficult to achieve high-precision assessment and adaptive optimization of model structure under multi-source heterogeneous marine data conditions.

[0004] This invention is implemented as follows: It provides a method for assessing the carrying capacity of deep-sea resources and environment based on spatiotemporal big data technology. This includes deploying a multi-layered observation network in the target area of ​​the deep sea; collaboratively collecting temperature, salinity, current velocity, chlorophyll concentration, dissolved oxygen, and nutrient concentration data using moored moorings, autonomous underwater vehicles, and remote sensing satellites; simultaneously acquiring multibeam bathymetry data; generating a three-dimensional seabed topography model from the multibeam bathymetry data; reconstructing a three-dimensional density hierarchy structure from the temperature and salinity data; extracting multi-scale ocean dynamic modal data from the sea level height time series; solving the partial differential equations for reconstructing the three-dimensional flow field to output reconstructed three-dimensional flow field data; performing spatiotemporal interpolation on sparse observation data to output interpolated field data and uncertainty estimation data; and constructing a modular bio-earth system. The chemical model outputs biogeochemical process simulation data. Reconstructed 3D flow field data, 3D density stratification structure, 3D seabed topography model, multi-scale ocean dynamic modal data, interpolated field data, and biogeochemical process simulation data are input into the deep-sea environmental carrying capacity assessment model. The deep-sea environmental carrying capacity assessment model employs a dynamic topology reconstruction sparse connection learning framework to dynamically adjust the network connection structure based on the activation frequency and gradient magnitude of the connection edges, outputting a resource development suitability index, an ecological vulnerability index, and an environmental capacity threshold. A comprehensive carrying capacity evaluation index is calculated based on these indices. When the comprehensive carrying capacity evaluation index falls within different intervals, a topology adjustment function is used to adjust the pruning probability threshold in the dynamic topology reconstruction sparse connection learning framework.

[0005] The multi-level observation network consists of four observation layers in the vertical direction: surface observation layer, thermocline observation layer, middle observation layer and deep observation layer. In the horizontal direction, observation stations of different densities are set up according to the complexity of the terrain and dynamic characteristics to form a three-dimensional observation system.

[0006] The steps for generating a three-dimensional model of the seabed topography include applying a beamforming inverse problem-solving algorithm based on a physical acoustic model to the multibeam bathymetry data to correct the sound refraction error, using constrained triangulation to preserve the topographic structure features, introducing geological prior knowledge to construct a regularization term, and generating a three-dimensional model of the seabed topography through multi-scale surface fitting.

[0007] The step of reconstructing the three-dimensional density stratification structure involves applying an iterative algorithm based on the sound velocity profile to the temperature and salinity data to invert the seawater density stratification structure using the observed sound velocity profile. The density profile is then iteratively corrected to minimize the residual between the calculated sound velocity and the observed value, while physical constraints on the temperature-salinity relationship are applied.

[0008] The steps for extracting multi-scale ocean dynamic mode data specifically involve applying a tidal harmonic-empirical mode hybrid decomposition algorithm to the sea level height time series. First, the amplitude and phase parameters of the main tidal components are identified and extracted to reconstruct a deterministic tidal signal. The deterministic tidal signal is then subtracted from the original sea level time series to obtain a residual sequence. Finally, empirical mode decomposition is performed on the residual sequence to decompose it into a series of intrinsic mode functions.

[0009] Among them, the partial differential equations for solving the three-dimensional flow field reconstruction are solved by the domain decomposition parallel finite element method, which divides the computational domain into several subdomains and distributes them to the graphics processor cluster for parallel computation. Information exchange between subdomains is realized through alternating iteration, and the spatial resolution is improved in the region with large flow field gradient by combining adaptive mesh densification technology.

[0010] Among them, the domain decomposition parallel finite element method divides a large-scale computational domain into several overlapping or non-overlapping subdomains. Each subdomain is assigned to a computational node to solve the local problem independently. The subdomains exchange information through boundary conditions to achieve global coupling. Alternating iteration is used to make the solutions on the boundaries of adjacent subdomains gradually converge.

[0011] Among them, non-stationary spatiotemporal Gaussian process regression is used for spatiotemporal interpolation of sparse observation data. Local anisotropy is characterized by spatially variable coefficient covariance function. The dominant spatiotemporal modes are extracted by combining empirical orthogonal function decomposition, and the interpolated field data and uncertainty estimation data are output.

[0012] The modular biogeochemical model simulates nutrient cycling and primary production processes. It uses operator splitting technology to decouple physical transport from biochemical reactions, uses implicit numerical methods to solve rigid reaction equations, and calibrates biogeochemical model parameters by assimilating chlorophyll concentration data and dissolved oxygen data through ensemble filtering.

[0013] The input layer of the deep-sea environmental carrying capacity assessment model receives reconstructed three-dimensional flow field data, three-dimensional density hierarchical structure, three-dimensional seabed topography model, multi-scale marine dynamic modal data, interpolation field data, and biogeochemical process simulation data. The feature extraction layer consists of three parallel branches that process physical oceanographic data, topographic data, and biochemical data respectively. The outputs of the three parallel branches are weighted and fused in the fusion layer through an attention mechanism.

[0014] This invention employs a sparse connectivity learning framework based on spatiotemporal big data technology and dynamic topology reconstruction for assessing the carrying capacity of deep-sea resources and the environment. This addresses the technical challenge of achieving high-precision assessment and adaptive optimization of the model structure under conditions of multi-source heterogeneous marine data. The invention reconstructs the three-dimensional flow field using a region decomposition parallel finite element method, interpolates sparse observation data using non-stationary spatiotemporal Gaussian process regression, and simulates nutrient cycling using a modular biogeochemical model. It integrates physical oceanographic data, topographic data, and biochemical data into the deep-sea environmental carrying capacity assessment model, overcoming the difficulty in effectively integrating multi-source heterogeneous data. Furthermore, it introduces a sparse connectivity learning framework based on dynamic topology reconstruction, dynamically adjusting the network connection structure according to the activation frequency and gradient magnitude of the connection edges. Alternating pruning and growth optimizes the model topology, enabling the model to automatically discover key dependencies between input features and adapt to environmental differences in different sea areas, thus overcoming the insufficient generalization ability of fixed topology models. In summary, this invention solves the technical problem mentioned in the background art of achieving high-precision assessment and adaptive optimization of the model structure under conditions of multi-source heterogeneous marine data for deep-sea resource and environmental carrying capacity assessment models. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention.

[0016] Figure 2 This is a spatial distribution map of a three-dimensional model of the seabed topography.

[0017] Figure 3 This is a graph showing the instantaneous frequency distribution of the intrinsic mode functions.

[0018] Figure 4 This is a vertical cross-sectional velocity distribution diagram of the three-dimensional flow field.

[0019] Figure 5 This is a diagram illustrating the dynamic adjustment process of the pruning probability threshold.

[0020] Figure 6 This is a spatial distribution map of the comprehensive carrying capacity evaluation index. Detailed Implementation

[0021] 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.

[0022] like Figure 1 The diagram shown is a flowchart of a method for assessing the environmental carrying capacity of deep-sea resources based on spatiotemporal big data technology provided by this invention. This method includes the following steps:

[0023] S01. Deploy a multi-layered observation network in the deep-sea target area, and use moored buoys, autonomous underwater vehicles and remote sensing satellites to collect temperature data, salinity data, current velocity data, chlorophyll concentration data, dissolved oxygen data and nutrient concentration data, while acquiring multibeam bathymetry data.

[0024] S02. Apply a beamforming inverse problem solving algorithm based on a physical acoustic model to correct the sound ray refraction error for multibeam bathymetry data. Use constrained triangulation to preserve the topographic structure features. Introduce geological prior knowledge to construct a regularization term. Generate a three-dimensional model of the seabed topography through multi-scale surface fitting.

[0025] S03. Apply an iterative algorithm based on sound velocity profiles to the inverse problem of seawater stratification to reconstruct the three-dimensional density stratification structure. Apply a tidal harmonic-empirical mode hybrid decomposition algorithm to the sea level height time series to separate deterministic tidal signals from stochastic ocean dynamic processes and extract multi-scale ocean dynamic mode data.

[0026] S04. The partial differential equations for the reconstruction of the three-dimensional flow field are solved by the parallel finite element method of domain decomposition. The computational domain is divided into several subdomains and distributed to the graphics processor cluster for parallel computation. Information exchange between subdomains is realized through alternating iteration. Combined with adaptive mesh refinement technology, the spatial resolution is improved in the region with large flow field gradient, and the reconstructed three-dimensional flow field data is output.

[0027] S05. Apply non-stationary spatiotemporal Gaussian process regression to spatiotemporal interpolation of sparse observation data, characterize local anisotropy through spatial variable coefficient covariance function, extract dominant spatiotemporal modes by combining empirical orthogonal function decomposition, and output interpolated field data and uncertainty estimation data.

[0028] S06. Construct a modular biogeochemical model to simulate nutrient cycling and primary production processes. Use operator splitting technology to decouple physical transport and biochemical reactions. Use implicit numerical methods to solve rigid reaction equations. Use ensemble filtering to assimilate chlorophyll concentration data and dissolved oxygen data to calibrate biogeochemical model parameters and output biogeochemical process simulation data.

[0029] S07. Input the reconstructed three-dimensional flow field data, three-dimensional density layer structure, three-dimensional seabed topography model, multi-scale marine dynamic modal data, interpolation field data and biogeochemical process simulation data into the deep-sea environmental carrying capacity assessment model. The deep-sea environmental carrying capacity assessment model outputs a resource development suitability index, an ecological vulnerability index and an environmental capacity threshold.

[0030] S08. Calculate the comprehensive carrying capacity evaluation index based on the resource development suitability index, ecological vulnerability index, and environmental capacity threshold. When the comprehensive carrying capacity evaluation index is in different intervals, use the topology adjustment function to adjust the pruning probability threshold in the sparse connection learning framework of the dynamic topology reconstruction of the deep-sea environmental carrying capacity assessment model, optimize the recognition accuracy of the deep-sea environmental carrying capacity assessment model for different carrying capacity states, and output the final assessment report.

[0031] The multi-layered observation network refers to setting up four observation layers vertically: surface, thermocline, mesosphere, and deep sea. Horizontally, it involves setting up observation stations at varying densities based on topographic complexity and dynamic characteristics, forming a three-dimensional observation system. Moored buoys are buoy systems anchored to the seabed, carrying temperature, salinity, and depth sensors and acoustic Doppler current profilers to achieve long-term, continuous, fixed-point observations. Autonomous underwater vehicles (AUVs) are unmanned submersibles equipped with multi-parameter sensors that perform maneuvering observations along pre-set routes to acquire large-scale three-dimensional observation data. Multibeam bathymetry data refers to seabed topographic data obtained by simultaneously measuring water depth values ​​in different directions using a multibeam sonar system that emits multiple beams.

[0032] The beamforming inverse problem solving algorithm based on a physical acoustic model refers to an algorithm that inverts seabed topography based on echo signals received by multi-beam sonar. It establishes a physical model of sound wave propagation to describe the refraction of sound rays in seawater and the reflection process on the seabed. By solving the inverse problem, the depth and reflection characteristics of each point on the seabed are calculated from the observed echo delay and intensity. Constrained triangulation adds constraints during triangular mesh generation, requiring important feature edges such as topographic contours and fault lines to appear in the triangulation results, avoiding distortion of the topographic structure caused by traditional triangulation algorithms crossing important features. The regularization term is a penalty term added to the inverse problem solution to constrain the properties of the solution to satisfy prior information. Adjusting the regularization parameter balances the data fitting accuracy and the rationality of the solution. Multi-scale surface fitting refers to simultaneously using basis functions of different resolutions to fit surfaces. Coarse-scale basis functions capture the overall shape, while fine-scale basis functions characterize local details. The combination accurately represents complex surfaces with multi-scale characteristics. The three-dimensional seabed topography model refers to a digital model that describes the undulating shape of the seabed topography using a set of three-dimensional coordinate points and triangular mesh patches.

[0033] The principle of the iterative algorithm for the inverse problem of seawater stratification based on sound velocity profiles is to utilize the dependence of sound wave propagation speed in stratified seawater on temperature and salinity. By inverting the seawater density stratification structure through the observed sound velocity profile, an integral equation between sound velocity and density is established. Since the problem is ill-posed, regularization is required. An iterative method is used to gradually correct the density profile to minimize the residual between the calculated sound velocity and the observed value. At the same time, physical constraints on the temperature-salinity relationship are applied to avoid non-physical solutions. During the iteration process, ray tracing technology is used to simulate the sound ray path to calculate and predict the sound velocity. Empirical orthogonal functions are introduced as basis functions to restrict the dimension of the solution space. The density field is updated through gradient descent until convergence. The specific implementation of the seawater stratification inverse problem iterative algorithm based on sound velocity profile in the deep-sea environmental carrying capacity assessment model is as follows: the collected sound velocity observation data is used as input, the three-dimensional density field is initialized as the climatological average value, and the iterative loop first calculates the sound velocity distribution using the current three-dimensional density field through empirical formulas, uses the ray tracing algorithm to simulate the sound ray propagation path to obtain the predicted sound velocity profile, calculates the residual between the predicted sound velocity profile and the observed sound velocity profile, and calculates the gradient of the residual field. The gradient is projected onto the subspace spanned by the empirical orthogonal function, and constraints on the temperature-salinity map are applied to ensure that the updated temperature-salinity values ​​are within a reasonable range. The three-dimensional density field is updated according to the gradient direction, and the above process is repeated until the residual norm is less than the set threshold, and the converged three-dimensional density stratification structure is output. The proposed iterative algorithm for the inverse problem of seawater stratification based on sound velocity profiles obtains a refined three-dimensional density stratification structure through inversion, providing fundamental data for accurately calculating seawater stratification stability and vertical mixing intensity. The three-dimensional density stratification structure directly affects the vertical transport of nutrients and the growth environment of phytoplankton. The accurate three-dimensional density stratification structure enables modular biogeochemical models to more realistically simulate primary production processes, thereby improving the accuracy of assessment of the carrying capacity of deep-sea ecosystems. At the same time, the spatiotemporal variation characteristics of the three-dimensional density stratification structure reveal the modulating effect of internal marine dynamic processes on environmental carrying capacity.

[0034] The principle of the tidal harmonic-empirical mode hybrid decomposition algorithm is based on the characteristic that sea level time series contain both periodic tidal signals and non-periodic ocean dynamic processes. First, the amplitude and phase parameters of the main tidal components are identified and extracted to reconstruct a deterministic tidal signal. The deterministic tidal signal is subtracted from the original sea level time series to obtain a residual sequence. Empirical mode decomposition is performed on the residual sequence to adaptively decompose it into a series of intrinsic mode functions. Each intrinsic mode function represents an oscillation mode on a time scale. By calculating the instantaneous frequency and amplitude of each intrinsic mode function, different physical processes are identified. Based on the correlation between intrinsic mode functions, intrinsic mode functions with similar physical meanings are recombined, ultimately achieving the separation of complex sea level changes on multiple time scales. The specific implementation of the tidal harmonic-empirical mode hybrid decomposition algorithm in the deep-sea environmental carrying capacity assessment model involves inputting the sea level height observation time series into the algorithm, using harmonic analysis to fit the harmonic constants of major tidal constituents such as semi-diurnal and diurnal tides, calculating the theoretical tide level of each constituent during the observation period, and summing them to obtain the predicted tide level. The predicted tide level is then subtracted from the sea level height observation time series to obtain the de-tidal residual. Eclectic mode decomposition is applied to the de-tidal residual, adding white noise as an auxiliary sequence to suppress mode aliasing. Iterative sieving yields several intrinsic mode functions and a trend term. Hilbert transform is performed on each intrinsic mode function to obtain analytical signals and extract instantaneous frequencies. Based on the instantaneous frequency range, the intrinsic mode functions are classified into physical processes such as storm surge processes, mesoscale eddy processes, and seasonal variation processes. The identified physical processes are correlated with environmental factors to output multi-scale ocean dynamic mode data. The tidal harmonic-empirical mode hybrid decomposition algorithm decomposes sea-level changes into physical processes at different time scales, enabling the deep-sea environmental carrying capacity assessment model to evaluate the impact of phenomena such as tides, storm surges, and eddies on resource development activities. The accurate extraction of deterministic tidal signals provides a basis for selecting offshore operation time windows, while the identification of stochastic ocean processes helps assess the probability of extreme sea states. The multi-scale decomposition results reveal the contribution weights of different dynamic processes to environmental carrying capacity, providing scientific support for formulating differentiated resource development strategies and risk prevention and control measures.

[0035] The domain decomposition parallel finite element method divides a large-scale computational domain into several overlapping or non-overlapping subdomains. Each subdomain is assigned to a computational node to independently solve a local problem. Subdomains exchange information through boundary conditions to achieve global coupling. An iterative method is used to gradually make the solutions on the boundaries of adjacent subdomains converge, ultimately obtaining an approximate solution for the entire computational domain. Alternating iteration refers to updating the solutions of each subdomain in turn during each iteration. When solving the current subdomain, the boundary values ​​of the previous iteration of the adjacent subdomains are used as boundary conditions. Through multiple iterations, the inconsistencies between subdomains are reduced. Adaptive mesh refinement technology dynamically adjusts the mesh density based on the gradient or error estimation index of the solution. In regions where the solution changes drastically, the mesh is refined to improve local resolution, while in regions where the solution is smooth, a coarse mesh is used to save computational resources. Reconstructing three-dimensional flow field data refers to reconstructing the complete three-dimensional fluid velocity field distribution data from sparse observation points using numerical calculation methods.

[0036] Non-stationary spatiotemporal Gaussian process regression is a probabilistic interpolation method. It assumes that the spatiotemporal field to be interpolated follows a Gaussian process prior distribution. The parameters of the covariance function of the Gaussian process prior distribution vary with spatial location to characterize the non-stationary nature of the spatiotemporal field. The prior distribution is updated using observed data to obtain the posterior distribution. The mean of the posterior distribution is used as the interpolation estimate, and the variance of the posterior distribution provides a measure of uncertainty. A spatially variable coefficient covariance function refers to a covariance function where parameters such as length scale and variance are not constants but functions of spatial coordinates. This allows for different characteristics of spatial correlation at different locations and is used to describe anisotropy and non-stationarity. Empirical orthogonal function decomposition decomposes the spatiotemporal field into the sum of products of time coefficients and spatial modes. Spatial modes are the eigenvectors of the data covariance matrix, and time coefficients are the projections of the data onto the corresponding eigenvectors. The first few principal modes contain the main variation information of the spatiotemporal field. Interpolated field data refers to the spatial distribution data of physical quantities estimated at unobserved locations using interpolation algorithms. Uncertainty estimation data refers to statistical data that quantifies the degree to which the interpolation result may deviate from the true value. Sparse observation data refers to observation data such as temperature, salinity, flow velocity, chlorophyll concentration, dissolved oxygen, and nutrient concentration, where the sampling points are few and unevenly distributed in both space and time.

[0037] The modular biogeochemical model refers to decomposing marine biogeochemical processes into multiple functional modules. Each module describes a single process such as primary production, organic matter degradation, nutrient cycling, and microbial metabolism. The complete model is formed through the coupling of input-output relationships between modules. Operator splitting technology decomposes partial differential equations containing multiple physical processes into several sub-equations containing only a single process. Each sub-equation is solved sequentially at each time step, and the output of each sub-equation serves as the initial condition for the next sub-equation. This method decouples the coupled problem into several relatively simple problems that are handled separately. Implicit numerical methods use unknown variables from the current time step to express the time derivative during time discretization. This requires solving a system of nonlinear algebraic equations to obtain the solution for the next time step, offering better numerical stability compared to explicit methods. Ensemble filtering is a data assimilation method that estimates the probability distribution of the modular biogeochemical model's state by running multiple instances of the modular biogeochemical model to form an ensemble. When new observational data arrives, the state values ​​of the ensemble members are updated using Bayes' theorem, allowing the modular biogeochemical model's predictions to gradually approximate the observed values. Biogeochemical model parameters refer to the numerical parameters describing biochemical reaction rates, half-saturation constants, growth rates, etc., in a modular biogeochemical model. Biogeochemical process simulation data refer to the spatiotemporal distributions of nutrient concentrations, primary productivity rates, and organic matter fluxes output by the modular biogeochemical model.

[0038] The specific structure of the deep-sea environmental carrying capacity assessment model is as follows: The input layer receives reconstructed three-dimensional flow field data, three-dimensional density hierarchical structure, three-dimensional seabed topography model, multi-scale marine dynamic modal data, interpolation field data, and biogeochemical process simulation data. After normalization, the data is sent to the feature extraction layer. The feature extraction layer consists of three parallel branches that process physical oceanographic data, topographic data, and biochemical data respectively. Each parallel branch contains multiple convolutional layers and pooling layers to extract multi-scale spatial features. The outputs of the three parallel branches are weighted and fused in the fusion layer through an attention mechanism. The fused feature vector is input into a fully connected layer for nonlinear transformation. The fully connected layer uses a sparse connection learning framework with dynamic topology reconstruction to optimize the connection structure. The output layer contains three output nodes that correspond to the resource development suitability index, ecological vulnerability index, and environmental capacity threshold, respectively. The activation value of each output node is normalized and mapped to a standard interval. The steps for establishing the training dataset for the deep-sea environmental carrying capacity assessment model specifically include: extracting time series of environmental parameters for deep-sea areas where resource development activities have been carried out from historical observation databases; labeling the carrying capacity level of each deep-sea area according to actual development and ecological monitoring results; using the environmental parameter time series as input features and the carrying capacity level as supervision labels; generating training samples by sliding window segmentation of the environmental parameter time series; generating synthetic data by numerical model simulation for deep-sea areas lacking measured labels; assigning labels to the synthetic data through expert knowledge rules; and dividing the measured samples and synthetic samples into training sets, validation sets, and test sets according to proportions. The specific steps for training the deep-sea environmental carrying capacity assessment model include: initializing network weights to small random values; setting hyperparameters such as batch size and learning rate; randomly sampling a batch of samples from the training set and inputting them into the deep-sea environmental carrying capacity assessment model for forward propagation to calculate the predicted output; calculating the mean square error between the predicted output and the supervision label as the loss function; calculating the gradient of the loss function with respect to the network weights of each layer using the backpropagation algorithm; updating the network weights based on the gradient using an adaptive moment estimation optimizer; applying a sparse connection learning framework with dynamic topology reconstruction to adjust the connection structure during the network weight update process; evaluating the performance of the deep-sea environmental carrying capacity assessment model on the validation set after each training cycle; adjusting the learning rate based on the validation error and determining whether to stop training early; and evaluating the generalization ability of the final deep-sea environmental carrying capacity assessment model on the test set after training is completed.

[0039] The principle of the sparse connection learning framework for dynamic topology reconstruction is that, during the training of the neural network, the connection topology is not fixed. Instead, connections are dynamically added or deleted based on the activation statistics of neurons. Specifically, the importance of each connection is evaluated by monitoring the activation frequency and gradient magnitude of each connection in multiple training batches. The mutual information criterion in information theory is used to quantify the contribution of connections to the output of the neural network. Connections with a contribution value below a threshold are pruned with a certain probability to remove redundant connections. At the same time, in order to prevent the decline of the neural network's expressive power during the sparsification process, a grouping regularization constraint is introduced to keep the weights of connections within the same group correlated. A small number of new connections are randomly initialized in the pruned neural network to explore possible effective paths. By alternating between pruning and growth, the topology of the neural network is gradually optimized, so that the neural network can achieve near-dense network performance while maintaining a small number of parameters. The specific implementation of the sparse connection learning framework for dynamic topology reconstruction in the deep-sea environmental carrying capacity assessment model is as follows: During the training of the fully connected layer, a topology reconstruction operation is performed every fixed number of training steps. First, the variance of the activation values ​​of the corresponding neurons in the most recent batches of each connection edge weight is calculated. The product of the connection edge weight and the output error gradient is calculated as the importance score. After sorting all the connection edges according to their importance scores, a certain proportion of the connection edges with the lowest importance scores are marked as candidate pruning objects. The Bernoulli distribution is used to randomly decide whether to remove the candidate pruning objects to maintain the randomness of pruning. A group lasso regularization term is applied to the remaining connection edges to encourage the joint activation or inhibition of connection edges in the same group. After pruning, a small number of pairs are randomly selected from the unconnected neuron pairs to establish new connection edges and assign small random weights. The current topology structure is recorded and conventional gradient descent training continues. The sparse connection learning framework based on dynamic topology reconstruction enables the deep-sea environmental carrying capacity assessment model to automatically discover key dependencies between input features. By pruning redundant connections, the number of parameters in the deep-sea environmental carrying capacity assessment model is significantly reduced, thereby reducing computational resource consumption and the risk of overfitting. The sparse neural network structure improves the interpretability of the deep-sea environmental carrying capacity assessment model because the retained connections reflect the core influence paths of environmental elements on carrying capacity. Dynamically adjusting the topology structure allows the deep-sea environmental carrying capacity assessment model to adapt to the differences in environmental characteristics of different sea areas. With limited training data, structural optimization enhances generalization ability. At the same time, grouping regularization constraints ensure that the neural network still maintains sufficient expressive power to capture nonlinear relationships after sparsification. Ultimately, a comprehensive optimization of the accuracy, efficiency, and interpretability of the deep-sea environmental carrying capacity assessment model is achieved.

[0040] The comprehensive carrying capacity evaluation index is obtained by weighting and comprehensively calculating the resource development suitability index, ecological vulnerability index and environmental capacity threshold. The weight coefficients are set according to different assessment objectives and policy orientations. The comprehensive carrying capacity evaluation index reflects the comprehensive carrying capacity of the deep-sea area to human activities.

[0041] The topology adjustment function is used to dynamically adjust the pruning probability threshold of the sparse connection learning framework of dynamic topology reconstruction in the deep-sea environmental carrying capacity assessment model according to the numerical range of the comprehensive carrying capacity evaluation index. The inputs include the mean of the comprehensive carrying capacity evaluation index for the current batch, the standard deviation of the comprehensive carrying capacity evaluation index, and the current sparsity of the model. The output is the adjusted pruning probability threshold. The calculation steps of the topology adjustment function are as follows: first, normalize the comprehensive carrying capacity evaluation index by dividing the mean by the full-scale value of the evaluation index to obtain the normalized evaluation index; second, normalize the standard deviation by dividing the standard deviation by the full-scale value of the evaluation index to obtain the normalized standard deviation; third, normalize the model's current sparsity by the target sparsity to obtain the normalized sparsity; fourth, calculate the adjustment factor, which is equal to the sum of the normalized evaluation index and the normalized standard deviation minus the normalized sparsity; fifth, multiply the adjustment factor by the baseline pruning probability to obtain the pruning probability adjustment amount; and sixth, add the baseline pruning probability to the pruning probability adjustment amount to obtain the adjusted pruning probability threshold. When the adjustment factor is in the range [0, 0.3), a conservative pruning strategy is used to reduce the weight coefficient of the pruning probability adjustment to 0.5 times. When the adjustment factor is in the range [0.3, 0.7), a standard pruning strategy is used to keep the weight coefficient of the pruning probability adjustment at 1 times. When the adjustment factor is in the range [0.7, 1.0], an aggressive pruning strategy is used to increase the weight coefficient of the pruning probability adjustment to 1.5 times. When the adjustment factor is greater than 1.0, a saturated pruning strategy is used to limit the pruning probability threshold to the maximum allowable value. When the adjustment factor is less than 0, a suppressive pruning strategy is used to set the pruning probability threshold to the minimum allowable value.

[0042] The resource development suitability index reflects the favorable conditions for resource development activities in deep-sea areas. It comprehensively considers factors such as water depth, current velocity, seabed topography slope, and resource abundance; a higher value indicates more favorable development conditions. The ecological vulnerability index characterizes the sensitivity and resilience of marine ecosystems to external disturbances. It is calculated based on indicators such as biodiversity, key species abundance, and food web structure stability; a higher value indicates a more vulnerable ecosystem. The environmental carrying capacity threshold refers to the maximum amount of pollutant emissions or resource development intensity that the environment can withstand while maintaining the health and function of the ecosystem. Exceeding the environmental carrying capacity threshold will lead to significant deterioration of environmental quality or loss of ecological function.

[0043] The full-scale value of the evaluation index refers to the theoretically achievable maximum and minimum value of the comprehensive carrying capacity evaluation index. The normalized evaluation index is a dimensionless value mapping the mean of the comprehensive carrying capacity evaluation index to a range of 0 to 1. The normalized standard deviation is a dimensionless value mapping the standard deviation of the comprehensive carrying capacity evaluation index to a range of 0 to 1. The current sparsity of the model refers to the proportion of currently pruned edges in the fully connected layer of the deep-sea environmental carrying capacity assessment model to the total number of edges. The target sparsity refers to the expected sparsity value during the optimization process of the deep-sea environmental carrying capacity assessment model. The normalized sparsity is a dimensionless value mapping the current model sparsity to a range of 0 to 1. The adjustment factor is a dimensionless parameter used to adjust the pruning strategy, calculated by comprehensively considering the normalized evaluation index, normalized standard deviation, and normalized sparsity. The baseline pruning probability is the default pruning probability value set under the standard pruning strategy. The pruning probability adjustment amount is the increment used to correct the baseline pruning probability based on the adjustment factor. The adjusted pruning probability threshold refers to the probability threshold used to guide edge pruning operations after calculation by the topology adjustment function. A conservative pruning strategy reduces the pruning probability to retain more edges. A standard pruning strategy uses a baseline pruning probability for regular pruning. An aggressive pruning strategy increases the pruning probability to remove more edges. A saturated pruning strategy limits the pruning probability threshold to its maximum allowable value to prevent over-pruning. A suppressive pruning strategy sets the pruning probability threshold to its minimum value to pause pruning operations. The maximum allowable value is the upper limit achievable by the adjusted pruning probability threshold. The minimum allowable value is the lower limit achievable by the adjusted pruning probability threshold.

[0044] Optionally, the present invention also provides a computer-based method for forming a deep-sea resource and environmental carrying capacity assessment system, wherein the computer is provided with a readable storage medium, the readable storage medium storing program instructions, and the program instructions being used to execute the above-described method when running in the computer.

[0045] The specific implementation methods of the above steps are described in detail below.

[0046] The specific implementation of step S01 involves first deploying observation equipment in the deep-sea target area according to a layered deployment principle. Vertically, the water body is divided into four layers from the sea surface to the seabed: a surface observation layer, a thermocline observation layer, a mesosphere observation layer, and a deep observation layer. The surface observation layer is located at a depth of 0 to 50 meters below the sea surface; the thermocline observation layer is located at a depth of 50 to 200 meters; the mesosphere observation layer is located at a depth of 200 to 1000 meters; and the deep observation layer is located from 1000 meters to the seabed. Moored buoys are deployed at each observation layer. These buoys are fixed to the seabed by an anchoring system and suspended at different depths by temperature, salinity, and depth sensors and acoustic Doppler current profilers. The sensor sampling frequency is set to once per hour. Simultaneously, an autonomous underwater vehicle (AUV) is deployed to conduct maneuvering observations along a preset grid route. The grid spacing is determined based on the terrain complexity. The autonomous underwater vehicle (AUV) is positioned at depths ranging from 5 to 20 km. Equipped with temperature sensors, conductivity sensors, chlorophyll fluorescence sensors, dissolved oxygen sensors, and a nutrient analyzer, it collects real-time data on temperature, salinity, current velocity, chlorophyll concentration, dissolved oxygen, and nutrient concentration at various depths during navigation. Time-series data on sea surface temperature and height are acquired using remote sensing satellites. A multibeam echo sounder system emits fan-shaped acoustic beams along the survey line to cover the seabed area, receiving reflected echo signals and recording the round-trip time and intensity of each beam. The round-trip time is converted to depth values ​​through sound velocity correction. All observation data is transmitted to a data center for storage and preprocessing via satellite or acoustic communication. The purpose of these steps is to obtain multi-element, multi-level, three-dimensional observation data of the deep-sea target area to provide a data foundation for subsequent analysis.

[0047] The specific implementation of step S02 involves first performing quality control on the received multibeam echo sounding data to remove outliers; using sound velocity profile data to correct the curvature of the sound wave propagation path in different water layers; establishing a physical model of sound ray propagation to describe the refraction law of sound waves in stratified seawater; calculating the refraction angle of sound waves at the interface of water layers with different densities based on Snell's law; simulating the complete path of sound waves from sonar emission to seabed reflection and back to the receiver using ray tracing technology; recalculating the true position coordinates of each measuring point on the seabed based on the actual propagation path; and applying a constrained triangulation algorithm to construct a triangular mesh on the corrected depth data point set, forcibly preserving the edges of seabed canyons and seamount ridges during the triangulation process. Topographic feature lines are used as constraint edges. A regularization constraint term is constructed using geological prior knowledge to penalize unreasonable topographic gradient changes. The regularization parameter is set to 0.01 to 0.1 according to the data noise level. Multi-scale basis functions are used to fit the surface. The wavelength of the coarse-scale basis function is set to 10 to 50 km to express regional topographic undulations, and the wavelength of the fine-scale basis function is set to 100 m to 1 km to characterize local topographic details. The basis function coefficients are solved by least squares to minimize the sum of squared residuals between the fitted surface and the observed data. Finally, a three-dimensional model of the seabed topography containing a three-dimensional coordinate point set and triangular mesh patches is generated. The purpose of these steps is to eliminate acoustic refraction errors and retain topographic structural features to generate a high-precision seabed topography model.

[0048] The specific implementation of step S03 is as follows: First, the seawater sound speed profile is calculated using temperature and salinity data through empirical formulas. The three-dimensional density field is initialized as the monthly average density distribution of historical climatic states. An iterative loop is then initiated. In each iteration, the corresponding sound speed distribution is calculated based on the current three-dimensional density field. A ray tracing algorithm is used to simulate the propagation path of sound rays in the three-dimensional density layer structure. The sound speed integral along the sound ray path is calculated to obtain the predicted sound speed profile. The difference between the predicted and observed sound speed profiles is used to obtain the residual field. The gradient of the residual field is calculated and projected onto an empirical orthogonal function basis to reduce the dimension of the solution space. The first 20 modes of the empirical orthogonal function typically contain more than 95% of the variance contribution. Temperature-salinity constraints are applied to ensure that the updates of temperature and salinity satisfy the temperature-salinity map characteristics of the region. The three-dimensional density field is updated along the negative gradient direction using the conjugate gradient method. The convergence criterion is set as a relative change in the residual norm of less than 0.001. After the iteration terminates, the converged three-dimensional density layer structure is output. The sea level height time series is then analyzed. First, harmonic analysis is applied to identify the main tidal components. The amplitudes and phase angles of the semi-diurnal main lunar equinox, semi-diurnal main solar equinox, and diurnal main lunar equinox are fitted using the least squares method. The theoretical tide level time series is calculated by superimposing the components. The theoretical tide level is subtracted from the observed sea level height time series to obtain the de-tidal residual series. Eclectic mode decomposition is applied to the de-tidal residual series, and Gaussian white noise with an amplitude of 10% of the residual standard deviation is added to suppress mode aliasing. The residual series is decomposed into several intrinsic mode functions through a sieving process. Hilbert transform is performed on each intrinsic mode function to obtain analytical signals. Instantaneous frequencies and amplitudes are extracted from the analytical signals. Based on the statistical characteristics of the instantaneous frequencies, the intrinsic mode functions are classified into physical processes such as storm surge processes, mesoscale eddies, and seasonal variations. The classified multi-scale ocean dynamic mode data is output. The purpose of these steps is to reconstruct a fine three-dimensional density stratification structure and separate ocean dynamic processes at different time scales.

[0049] The specific implementation of step S04 is as follows: First, based on the three-dimensional model of the seabed topography and the distribution of observation stations, the computational domain is divided into several subdomains. An overlapping area of ​​2 to 5 grid points is set between the boundaries of each subdomain to facilitate information exchange. Each subdomain is assigned to a computing node in the graphics processing unit cluster. Within each subdomain, three-dimensional flow field control equations, including continuity and momentum equations, are established. The finite element method is used to spatially discretize the control equations, and tetrahedral elements are selected as the basic grid elements. In areas with small flow field gradients, the grid size is set to 5 to 10 km. In areas with large flow field gradients, such as near steep seabed slopes and frontal regions, adaptive mesh refinement technology is applied. The mesh density is dynamically adjusted based on the local estimates of the velocity gradient and vorticity. When the gradient exceeds a threshold of 0... When the flow rate is 0.01 m / s, the mesh is refined to one-quarter of its original size. An implicit time-stepping scheme is used to discretize the time derivative, and the time step is set to 300 to 600 s to ensure numerical stability. In each time step, an alternating iterative method is used to solve the subdomain coupling problem. First, each subdomain solves its local problem independently, using the boundary values ​​of the previous iteration of the adjacent subdomain as boundary conditions. After the solution is completed, the flow velocity and pressure values ​​on the boundary of the subdomain are transferred to the adjacent subdomain. The iteration process is repeated until the difference between the boundary values ​​of the adjacent subdomains is less than the set tolerance of 0.001 m / s. After the iteration converges, the solutions of each subdomain are combined to form the reconstructed three-dimensional flow field data of the entire computational domain. The purpose of the above steps is to efficiently solve the large-scale three-dimensional flow field reconstruction problem through parallel computing and adaptive mesh technology.

[0050] The specific implementation of step S05 involves inputting temperature, salinity, flow velocity, chlorophyll concentration, dissolved oxygen, and nutrient concentration data as sparse observation data into a non-stationary spatiotemporal Gaussian process regression model. It is assumed that the interpolation field follows a Gaussian process prior distribution with a zero mean function. A spatially variable coefficient covariance function is constructed, where the length scale parameter and variance parameter are expressed as functions of spatial coordinates. A gradient-enhanced decision tree is used to learn the nonlinear mapping relationship between the length scale parameter and environmental variables, including water depth, seabed slope, and distance from the shore. Historical observation data is then processed. Empirical orthogonal function decomposition extracts the top 30 dominant spatial modes. These dominant spatial modes serve as the basis for the covariance function to reduce computational dimensionality. The posterior distribution of the Gaussian process is updated using Bayesian inference based on observed data. The mean function of the posterior distribution is used as the interpolation estimate for unobserved locations. The covariance function of the posterior distribution provides a quantitative assessment of interpolation uncertainty. For large-scale datasets, variational inference is used to approximate the posterior distribution to avoid directly solving for the high-dimensional covariance matrix. The interpolated field data and uncertainty estimation data are output. The purpose of these steps is to reconstruct the complete spatiotemporal field and quantify interpolation uncertainty under sparse observation conditions.

[0051] The specific implementation of step S06 involves constructing a modular biogeochemical model comprising phytoplankton, zooplankton, nutrient, and organic matter modules. These modules are coupled through mass flux calculations. Operator splitting techniques are used to decompose each time step into transport and reaction steps. In the transport step, convection-diffusion equations are solved to calculate the spatial transport of nutrients and phytoplankton, and reconstructed three-dimensional flow field data is used to provide the velocity field. In the reaction step, a set of biochemical reaction equations is solved to describe processes such as photosynthesis, respiration, feeding, and degradation. This set of biochemical reaction equations is a rigid set of ordinary differential equations, solved using the Rosenbrock implicit method, with adaptive adjustment of the time step. The range is 60 to 3600 s. Ensemble filtering is used to assimilate chlorophyll concentration data and dissolved oxygen data to calibrate biogeochemical model parameters. Ensemble filtering estimates the probability distribution of state variables and parameters by running 50 to 100 model instances. When chlorophyll concentration data from satellite remote sensing arrives, the information between the ensemble forecast and the observation is calculated. The state and parameters of the ensemble members are updated according to the Kalman gain matrix. Simulation data of biogeochemical processes such as spatiotemporal distribution of nutrient concentration, spatiotemporal distribution of primary production rate, and spatiotemporal distribution of organic matter flux are output. The purpose of these steps is to simulate marine biogeochemical processes and improve simulation accuracy through data assimilation.

[0052] The specific implementation of step S07 involves inputting reconstructed 3D flow field data, 3D density hierarchical structure, 3D seafloor topography model, multi-scale ocean dynamic modal data, interpolated field data, and biogeochemical process simulation data into the deep-sea environmental carrying capacity assessment model. First, the input layer normalizes each data source to eliminate dimensional differences. Normalization uses a maximum-minimum standardization method to map the data to the 0-1 range. The normalized data is then fed into the feature extraction layer. The physical oceanographic data branch of the feature extraction layer performs convolution operations on the reconstructed 3D flow field data and 3D density hierarchical structure to extract spatial pattern features of the flow field. The topographic data branch performs convolution operations on the 3D seafloor topography model to extract topographic relief features. The biochemical data branch performs convolution operations on the biogeochemical process simulation data to extract biological distribution pattern features. Each branch contains 3 to 5 convolutional layers and pooling layers. The convolutional kernel size is set to 3×3×3, the pooling window size is set to 2×2×2, and the activation function is a modified linear unit function. The output feature maps of the three branches are then processed through an attention mechanism in the fusion layer. Weighted fusion is performed, with attention weights adaptively determined by learning the contribution of each feature map to the final output. The fused feature vector is input into a fully connected layer, which applies a sparse connection learning framework with dynamic topology reconstruction. During training, topology reconstruction is performed every 100 training steps. The variance of the activation values ​​of the corresponding neurons in the most recent 50 batches of connection edge weights is calculated, and the product of the connection edge weight and the output error gradient is used as the importance score. The 20% of connections with the lowest importance scores are marked as candidate pruning objects, and candidate connections are randomly pruned with a 50% probability. A grouped lasso regularization term is applied to the remaining connections, with the regularization coefficient set to 0.001 to 0.01. After pruning, 5% of the unconnected neuron pairs are randomly selected to establish new connections. The output layer contains three output nodes that calculate the resource development suitability index, ecological vulnerability index, and environmental capacity threshold, respectively. The activation values ​​are mapped to the range of 0 to 1 using the sigmoid function. The purpose of these steps is to integrate multi-source data and intelligently assess the carrying capacity of the deep-sea environment through a deep learning model.

[0053] The specific implementation of step S08 is as follows: First, calculate the comprehensive carrying capacity evaluation index based on the resource development suitability index, ecological vulnerability index, and environmental capacity threshold. The calculation process involves multiplying the resource development suitability index by a weighting coefficient of 0.4, multiplying the reciprocal of the ecological vulnerability index by a weighting coefficient of 0.3, and normalizing the environmental capacity threshold by a weighting coefficient of 0.3. The three factors are then weighted and summed to obtain the comprehensive carrying capacity evaluation index. The comprehensive carrying capacity evaluation index ranges from 0 to 1, with a larger value indicating stronger carrying capacity. The mean comprehensive carrying capacity evaluation index of all samples in the current batch is then calculated. The sparsity of the deep-sea environmental carrying capacity assessment model is obtained by summing the standard deviation and calculating the sparsity as the number of pruned edges divided by the total number of edges. The mean of the comprehensive carrying capacity evaluation index, the standard deviation of the comprehensive carrying capacity evaluation index, and the current sparsity of the model are input into the topology adjustment function. The topology adjustment function first divides the mean of the comprehensive carrying capacity evaluation index by the full-scale value of the evaluation index (1) to obtain the normalized evaluation index, then divides the standard deviation of the comprehensive carrying capacity evaluation index by the full-scale value of the evaluation index (1) to obtain the normalized standard deviation, and finally divides the current sparsity of the model by the target sparsity of 0.7 to obtain the normalized sparsity. The adjustment factor equals the normalized evaluation index plus the normalized standard deviation minus the normalized sparsity. A pruning strategy is selected based on the adjustment factor's range: a conservative pruning strategy with a weighting coefficient of 0.5 is used when the adjustment factor is between 0 and 0.3; a standard pruning strategy with a weighting coefficient of 1 is used when the adjustment factor is between 0.3 and 0.7; an aggressive pruning strategy with a weighting coefficient of 1.5 is used when the adjustment factor is between 0.7 and 1.0; a pruning probability threshold is set to 0.8 when the adjustment factor is greater than 1.0; and a pruning probability threshold is set to 0.1 when the adjustment factor is less than 0. The adjustment factor is multiplied by the baseline pruning probability of 0.2 and then by the weighting coefficient to obtain the pruning probability adjustment amount. The baseline pruning probability of 0.2 plus the pruning probability adjustment amount gives the adjusted pruning probability threshold. The adjusted pruning probability threshold is applied to the next round of training of the deep-sea environmental carrying capacity assessment model to optimize the model's recognition accuracy for different carrying capacity states. Finally, an assessment report is generated that includes a resource development suitability index, an ecological vulnerability index, an environmental capacity threshold, and a comprehensive carrying capacity evaluation index. The purpose of these steps is to dynamically adjust the model parameters based on the assessment results and output the final assessment report.

[0054] Specifically, the principle of this invention is: This invention can solve the technical problem that the deep-sea resource and environmental carrying capacity assessment model is difficult to achieve high-precision assessment and adaptive optimization of model structure under the condition of multi-source heterogeneous marine data. Its principle is to achieve deep fusion of multi-source heterogeneous data and adaptive adjustment of model structure by constructing a complete technical chain from data acquisition to model optimization. First, multi-level observation data were collected collaboratively using moored underwater buoys, autonomous underwater vehicles, and remote sensing satellites. Depth measurement errors were corrected using a beamforming inverse problem algorithm based on a physical acoustic model. Density stratification structure was reconstructed using an iterative algorithm based on sound velocity profiles for seawater stratification inverse problem. A tidal harmonic-empirical mode hybrid decomposition algorithm was applied to separate ocean dynamic processes at different time scales, ensuring the accuracy and consistency of heterogeneous data from a physical mechanism perspective. Second, a three-dimensional flow field was reconstructed through parallel computing. Non-stationary spatiotemporal Gaussian process regression interpolation was used for sparse data, and a modular biogeochemical model was used to simulate ecological processes. The processed physical, topographic, and biochemical data were input into a deep-sea environmental carrying capacity assessment model with three parallel branches. Weighted fusion using an attention mechanism effectively integrated the heterogeneous data. Finally, a dynamic topology reconstruction sparse connection learning framework dynamically pruned and grew connections based on the importance scores of the connection edges, enabling the model's topology to adaptively adjust with data features, thereby improving generalization performance while maintaining the model's expressive power.

[0055] 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.

[0056] The specific implementation of step S01 is the same as described above, and will not be repeated in detail here.

[0057] The specific implementation of step S02 is as follows: Apply a beamforming inverse problem solving algorithm based on a physical acoustic model to the multibeam bathymetry data to correct the acoustic refraction error. The formula for correcting the acoustic refraction error is expressed as follows:

[0058] ;

[0059] In the formula, This is a dimensionless normalized depth correction value. The measured depth is in meters. This is the corrected depth value obtained through ray tracing calculations, in meters; This represents the maximum depth of the observed area, in meters. This represents the minimum depth of the observation area, in meters. Constrained triangulation is used to preserve topographic structure features, and prior geological knowledge is introduced to construct a regularization term. The formula for the regularization term is as follows:

[0060] ;

[0061] In the formula, This is a normalization regularization term, dimensionless; This is the smoothness weighting coefficient, with an empirical value of 0.6, and it is dimensionless. This is the geological prior weighting coefficient, with an empirical value of 0.4, and is dimensionless. The total number of grid nodes, dimensionless; The number of nodes with prior geological information, dimensionless; For the first The depth value of each grid node, in meters; For the first A Laplace operator with a depth of n nodes, in negative first power of meters; For the first The depth value of a node with prior information, in meters; For the first The geological prior depth value of each node, in meters; The reference depth value, in meters, is typically taken as the average depth of the observation area. A 3D model of the seabed topography is generated through multi-scale surface fitting. The formula for multi-scale surface fitting is as follows:

[0062] ;

[0063] In the formula, coordinates The depth of the seabed at that location, in meters; The scale layer number is typically 3 to 5 and is dimensionless. For the first The number of basis functions for each scale layer is dimensionless. For the first The first scale layer The coefficients of the basis functions, in meters; For the first The first scale layer One basis function, dimensionless; The coordinates are eastward, and the unit is meters. The coordinates are north-facing, and the unit is meters.

[0064] The specific implementation of step S03 is as follows: A three-dimensional density stratification structure is reconstructed by applying an iterative algorithm based on the sound velocity profile-driven inverse seawater stratification problem to the temperature and salinity data. The formula for updating the density field is expressed as follows:

[0065] ;

[0066] In the formula, For the first Coordinates after the second iteration The density value at that location, expressed in kilograms per cubic meter; For the first The density value of the next iteration, in kilograms per cubic meter; The learning rate is empirically estimated to be between 0.01 and 0.05, expressed in kilograms per cubic meter. The modal number of the empirical orthogonal function is typically 10 to 20 and is dimensionless. For the first The projection coefficients of an empirical orthogonal function, dimensionless; For the first A spatial mode of an empirical orthogonal function, dimensionless; The coordinates are eastward, and the unit is meters. The coordinates are north-facing, and the unit is meters. This is a vertical coordinate system, in meters, with upward being positive. Projection coefficient. The calculation formula is expressed as follows:

[0067] ;

[0068] In the formula, For the first The projection coefficients of an empirical orthogonal function, dimensionless; This is the sound velocity residual field, measured in meters per second. This represents the maximum value of the sound velocity residual field, expressed in meters per second. For the first An empirical orthogonal function space mode, dimensionless; Let the volume be a infinitesimal element. Empirical orthogonal functions. The covariance matrix is ​​obtained by performing eigenvalue decomposition on the covariance matrix of historical density field data. The calculation formula is expressed as follows:

[0069] ;

[0070] In the formula, The first covariance matrix is ​​the first... Line number Column elements, dimensionless; The number of time steps for the historical sample is dimensionless. For the first At the time step Density values ​​at each spatial location, in kilograms per cubic meter; For the first The time-average density at a spatial location, in kilograms per cubic meter; For the first The standard deviation of the density values ​​at each spatial location, in kilograms per cubic meter; For the first The standard deviation of the spatial location density values ​​is given in kilograms per cubic meter. The tidal harmonic empirical mode hybrid decomposition algorithm is applied to the sea level height time series to separate the deterministic tidal signal from the stochastic ocean dynamic process. The formula for reconstructing the tidal signal is as follows:

[0071] ;

[0072] In the formula, for The tide level at any given time, in meters; The number of tidal constituents is typically taken as 8 to 15, and is dimensionless. For the first The amplitude of each tidal component, in meters; For the first The angular frequency of each tidal constituent, measured in radians per second; Time, in seconds; For the first The phase of each tidal constituent, in radians. The formula for the tidal residual is as follows:

[0073] ;

[0074] In the formula, To normalize and remove residual moisture, it is dimensionless; for The constantly observed sea level height, in meters; for The tide level at any given time, in meters; This represents the maximum sea level height during the observation period, in meters. This represents the minimum sea level height during the observation period, expressed in meters.

[0075] The specific implementation of step S04 is as follows: The partial differential equations for reconstructing the three-dimensional flow field are solved using the domain decomposition parallel finite element method. The computational domain is divided into several subdomains and distributed to a graphics processor cluster for parallel computation. Information exchange between subdomains is achieved through alternating iterations. Adaptive mesh refinement technology is used to improve spatial resolution in regions with large flow field gradients, and the reconstructed three-dimensional flow field data is output. The formula for velocity coordination at the subdomain interfaces is as follows:

[0076] ;

[0077] In the formula, For the first The flow velocity on the subdomain interface after the next iteration is expressed in meters per second. For subdomain In the The interface flow velocity calculated in each iteration is expressed in meters per second. For subdomain In the The interface flow rate for each iteration is measured in meters per second. The relaxation factor has an empirical value of 0.5 to 0.7 and is dimensionless. The iteration number is dimensionless.

[0078] The specific implementation of step S05 is as follows: Spatiotemporal interpolation is performed on the sparse observation data using a non-stationary spatiotemporal Gaussian process regression. The formula for the spatially variable coefficient covariance function is as follows:

[0079] ;

[0080] In the formula, For position and The covariance between them is dimensionless; For position The variance parameter at point is dimensionless. For the first The coordinate vectors of a spatial location, in meters; For the first The coordinate vectors of a spatial location, in meters; For position and The Euclidean distance between them, in meters; For position The spatially relevant length scale at a given location is expressed in meters. The formula for the interpolation estimate is as follows:

[0081] ;

[0082] In the formula, For position The interpolated estimates at the location are in the same units as the observations; Here are the coordinate vectors of the position to be interpolated, in meters; For the observed location and the predicted location The covariance vector between them is dimensionless; Let be the covariance matrix between observation locations, which is dimensionless; To observe the noise variance, dimensionless; It is an identity matrix, dimensionless; This is the observation vector, with the same units as the observed values. The observation vector... for 3D column vector, The total number of observation points is dimensionless; the covariance matrix is... for 3D matrix; covariance vector for A column vector of dimensions. The formula for estimating interpolation uncertainty is as follows:

[0083] ;

[0084] In the formula, For position The interpolation variance at the point is dimensionless. For position The prior variance at the location is dimensionless. For the observed location and the predicted location The covariance vector between them is dimensionless; Let be the covariance matrix between observation locations, which is dimensionless; To observe the noise variance, dimensionless; It is an identity matrix, dimensionless.

[0085] The specific implementation of step S06 is as follows: A modular biogeochemical model is constructed to simulate nutrient cycling and primary production processes. Operator splitting technology is used to decouple physical transport and biochemical reactions. An implicit numerical method is used to solve the rigid reaction equations. The biogeochemical model parameters are calibrated by assimilating chlorophyll concentration and dissolved oxygen data through ensemble filtering, and the simulated biogeochemical process data is output. The formula for updating the concentration field after operator splitting is expressed as follows:

[0086] ;

[0087] In the formula, This refers to the concentration of nutrients or other biochemical components, expressed in millimoles per cubic meter. Time, in seconds; For physical input operations, the unit is millimoles per cubic meter per second; This is a biochemical reaction operator, with units of millimoles per cubic meter per second. Within the time step... First, the physical transport process is solved to obtain the intermediate state. :

[0088] ;

[0089] Then, the state at the next time step is obtained by solving the biochemical reaction process. :

[0090] ;

[0091] In the formula, For the first The concentration at each time step, in millimoles per cubic meter; The concentration is at an intermediate state, expressed in millimoles per cubic meter. For the first The concentration at each time step, in millimoles per cubic meter; The time step is in seconds. Here, represents the time step number, which is dimensionless. The formula for the primary production rate is as follows:

[0092] ;

[0093] In the formula, The normalized primary production rate is dimensionless. The normalized maximum primary production rate is dimensionless and defaults to 1. Light intensity, expressed in watts per square meter; For reference light intensity, the unit is watts per square meter, usually taken as 100; Nutrient concentration, expressed in millimoles per cubic meter; For reference nutrient concentration, the unit is millimoles per cubic meter, and the value is usually taken as 10; is the light half-saturation constant, measured in watts per square meter, with an empirical value of 50; The nutrient half-saturation constant is expressed in millimoles per cubic meter, with an empirical value of 1. The formula for parameter updates in ensemble filtering assimilation is as follows:

[0094] ;

[0095] In the formula, This is the assimilated model parameter vector; This is the predicted parameter vector before assimilation; The Kalman gain matrix is ​​dimensionless. For observation vectors; The observation operator matrix is ​​dimensionless. The Kalman gain matrix is ​​also present. The calculation formula is expressed as follows:

[0096] ;

[0097] In the formula, The Kalman gain matrix is ​​dimensionless. The forecast error covariance matrix; The transpose of the observation operator matrix is ​​dimensionless; The observation operator matrix is ​​dimensionless; Let be the observation error covariance matrix.

[0098] The specific implementation method of step S07 is the same as described above, and will not be repeated in detail here.

[0099] The specific implementation method of step S08 is as follows: Calculate the comprehensive carrying capacity evaluation index based on the resource development suitability index, ecological vulnerability index, and environmental capacity threshold. The formula for the comprehensive carrying capacity evaluation index is as follows:

[0100] ;

[0101] In the formula, This is a dimensionless index for comprehensive carrying capacity evaluation. For resource development suitability index; This represents the maximum value of the resource development suitability index. As an ecological vulnerability index; This represents the maximum value of the ecological vulnerability index. This is the environmental capacity threshold; This represents the maximum value of the environmental capacity threshold. The resource development suitability weighting coefficient has an empirical value of 0.4 and is dimensionless. This is the ecological vulnerability weighting coefficient, with an empirical value of 0.3, and is dimensionless. The environmental capacity weighting coefficient has an empirical value of 0.3 and is dimensionless. First, calculate the adjustment factor. The formula is expressed as follows:

[0102] ;

[0103] In the formula, The regulating factor is dimensionless. This is a normalized evaluation index, dimensionless. The normalized standard deviation is dimensionless. Normalized sparsity, dimensionless. Normalized evaluation index. The calculation method is as follows:

[0104] ;

[0105] In the formula, It is a normalized evaluation index, dimensionless; This is a dimensionless index for comprehensive carrying capacity evaluation. To evaluate the full-scale value of the exponent, which is dimensionless. Normalized standard deviation. The calculation method is as follows:

[0106] ;

[0107] In the formula, The normalized standard deviation is dimensionless. The standard deviation of the comprehensive bearing capacity evaluation index for the current batch is dimensionless. The exponent is dimensionless and its full-scale value is used to evaluate the normalized sparsity. The calculation method is as follows:

[0108] ;

[0109] In the formula, The normalized sparsity is dimensionless. The sparsity of the model is dimensionless. The target sparsity is dimensionless. Then, based on the adjustment factor... The adjusted pruning probability threshold is calculated using the following formula:

[0110] ;

[0111] In the formula, The adjusted pruning probability threshold is dimensionless. The baseline pruning probability has an empirical value of 0.1 to 0.3, and is dimensionless. These are weighting coefficients, dimensionless. The regulating factor is dimensionless. When When they belong to different intervals, the weighting coefficient The value can be: when hour, ;when hour, ;when hour, ;when hour, ,in This is the maximum permissible value, typically taken as 0.8, and is dimensionless; when hour, ,in This is the minimum allowable value, usually taken as 0.01, and is dimensionless.

[0112] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: To verify the effectiveness of the invention, technicians built a test environment and conducted a resource and environmental carrying capacity assessment in a deep-sea target area. The water depth of this area ranged from 2800 to 4200 meters, and the area was approximately 15,000 square meters. Historically, preliminary resource exploration activities had been conducted, but systematic carrying capacity assessment data was lacking. Following the method of this invention, technicians deployed a multi-layered observation network in the target area. Vertically, observation levels were set at 5m above the surface, 150m at the thermocline, 800m at the middle layer, and 3000m at the depth. Horizontally, 42 observation stations were set up according to the seabed topographic features, including 12 moored underwater buoys. An autonomous underwater vehicle (AUV) conducted 8 patrol missions, and remote sensing satellites acquired continuous sea surface data for 90 days. During the observation period, 126,580 sets of temperature data, 118,340 sets of salinity data, 93,270 sets of current velocity data, 78,450 sets of chlorophyll concentration data, 82,190 sets of dissolved oxygen data, and 71,680 sets of nutrient concentration data were collected. Multibeam bathymetry data covered 87% of the total area.

[0113] Technicians processed multibeam bathymetry data using a beamforming inverse problem algorithm based on a physical acoustic model. The sound velocity profile showed a surface sound velocity of 1532 m / s, decreasing to 1485 m / s at depth. The maximum sound ray refraction angle reached 18 degrees. After 15 iterations, the sound ray refraction error decreased from an initial 23 m to 1.8 m. Constrained triangulation was used to divide the seabed topography into 328,560 triangular mesh units, preserving 1247 important topographic features such as seamounts and canyons. Introducing geological prior knowledge to construct a regularization term improved the depth estimation accuracy at abrupt topographic changes by 42%. Multi-scale surface fitting employed three scale levels: 580 coarse-scale basis functions, 2340 meso-scale basis functions, and 9860 fine-scale basis functions. The resulting 3D seabed topography model achieved a spatial resolution of 50 m. Figure 2 As shown, the three-dimensional model of the seabed clearly displays the topographic undulations of the target area, with a maximum elevation difference of 1380m.

[0114] Technicians reconstructed the three-dimensional density stratification structure using an iterative algorithm based on the inverse problem of seawater stratification, employing temperature and salinity data. The three-dimensional density field was initialized with historical climatological averages, with a surface density of 1024.8. The density at depth is 1027.6. During the iterative process, the ray tracing algorithm simulated 18,650 sound ray paths. After 23 iterations, the root mean square error between the predicted and observed sound velocity profiles decreased from the initial 3.7 m / s to 0.4 m / s. Empirical orthogonal function decomposition extracted eight principal modes, with the first three principal modes contributing 78% of the cumulative variance. Temperature-salinity constraints ensured that temperature-salinity values ​​at all locations were within the historical observation range. The reconstructed three-dimensional density stratification structure revealed a distinct thermocline structure in the target region, with a thickness of approximately 180 m. The stratification stability parameter reached its maximum value at the center of the thermocline. This provides the basic data for subsequent calculations of vertical mixing intensity.

[0115] Technicians processed the sea level height time series using a tidal harmonic-empirical mode hybrid decomposition algorithm, and identified the semi-diurnal tides through harmonic analysis. The tidal amplitude is 0.68m, and the diurnal tide is... The amplitude of the first tidal constituent is 0.32 m, while the amplitudes of other major tidal constituents range from 0.08 to 0.24 m, as shown in Table 1. The correlation coefficient between the predicted and observed tide levels is 0.94. Subtracting the predicted tide level from the sea level observation time series yields the de-tidal residual series, with a standard deviation of 0.15 m. Ensemble empirical mode decomposition is applied to the de-tidal residuals, and a white noise auxiliary sequence with an amplitude of 0.02 m is added. Iterative sieving yields seven intrinsic mode functions and one trend term, as shown in Table 1. Figure 3 As shown, the instantaneous frequency distribution characteristics of each intrinsic mode function are significantly different. Based on the instantaneous frequency range, the intrinsic mode functions are classified into two storm surge processes, three mesoscale eddy processes, and two seasonal variation processes. The multi-scale ocean dynamic mode data provides dynamic background information for subsequent environmental carrying capacity assessment.

[0116] Table 1 Main Tidal Parameters

[0117]

[0118] Technicians employed a domain decomposition parallel finite element method to solve the partial differential equations for reconstructing the three-dimensional flow field. The computational domain was divided into 16 subdomains, distributed across a graphics processing unit (GPU) cluster for parallel computation. Each subdomain contained approximately 45,000 mesh nodes. Information was exchanged between subdomains through boundary conditions. Using an alternating iterative method, after 34 iterations, the velocity difference at the boundaries of adjacent subdomains decreased from an initial 0.27 m / s to 0.008 m / s. Adaptive mesh refinement technology was applied when the flow field gradient exceeded a threshold of 0.015. The automatic mesh refinement in the region increased the mesh density by 3 times. The refined regions are mainly distributed around the seamount and at the canyon entrance, accounting for 18% of the total computational domain area. The reconstructed 3D flow field data shows that the target region is mainly controlled by westward flow, with surface velocity of 0.35–0.52 m / s, middle layer velocity decreasing to 0.08–0.15 m / s, and deep layer velocity only 0.02–0.05 m / s. Figure 4 As shown, the vertical cross-section of the three-dimensional flow field clearly demonstrates the attenuation law of the flow velocity with depth.

[0119] Technicians applied non-stationary spatiotemporal Gaussian process regression to perform spatiotemporal interpolation on sparse observation data. The length scale parameter of the spatially variable coefficient covariance function varied from 8 to 35 km at different locations, reflecting the anisotropic characteristics of spatial correlation in the target area. Empirical orthogonal function decomposition extracted 12 dominant spatiotemporal modes. The variance contribution rates of the first five modes were 32%, 18%, 14%, 11%, and 8%, respectively, with a cumulative variance contribution rate of 83%. The interpolated temperature field achieved a spatial resolution of 5 km and a temporal resolution of 1 day. The interpolated field data filled in the gaps in the original observations, increasing the coverage from 43% to 100%. Uncertainty estimation data showed that the standard deviation of the interpolation error was 0.12℃ in areas with dense observation stations, increasing to 0.48℃ in areas with sparse observation stations, providing a quantitative basis for the credibility analysis of the evaluation results.

[0120] Technicians constructed a modular biogeochemical model to simulate nutrient cycling and primary production processes. The model includes six state variables: phytoplankton, zooplankton, detritus, nitrates, phosphates, and dissolved oxygen. Operator splitting techniques were used to decouple physical transport from biochemical reactions, with a physical transport step size of 1 hour and a biochemical reaction step size of 0.1 hours. Implicit numerical methods were used to solve the rigid reaction equations, achieving a 98% satisfaction rate of the numerical stability criterion. Chlorophyll concentration and dissolved oxygen data were assimilated using ensemble filtering, with the ensemble size set to 50 members. After assimilation, the maximum phytoplankton growth rate increased from the initially estimated value of 0.8. Adjusted to 0.65 The half-saturation constant starts from an initial value of 1.2. Adjusted to 0.87 The root mean square error between the model's predicted chlorophyll concentration and the observed values ​​decreased from 0.34 before assimilation. Reduced to 0.09 Biogeochemical process simulation data show that the primary productivity rate of the surface layer in the target area is 180–320. The vertical transport flux of nutrients is 0.8–1.5. This provides a biochemical basis for assessing ecological carrying capacity.

[0121] Technicians input reconstructed 3D flow field data, 3D density hierarchical structure, 3D seabed topography model, multi-scale ocean dynamic modal data, interpolated field data, and biogeochemical process simulation data into the deep-sea environmental carrying capacity assessment model. The feature vector received by the model input layer has a dimension of 1568, which is then normalized and sent to the feature extraction layer. The feature extraction layer has three parallel branches, each containing four convolutional layers with kernel sizes of 7×7, 5×5, 3×3, and 3×3, respectively. The pooling layer uses max pooling with a pooling window size of 2×2. The fusion layer uses an attention mechanism to weight and fuse the outputs of the three parallel branches: the physical oceanographic data branch has a weight of 0.42, the topographic data branch has a weight of 0.31, and the biochemical data branch has a weight of 0.27. The fully connected layer adopts a sparse connection learning framework with dynamic topology reconstruction. The initial number of connection edges is 286,400. During training, a topology reconstruction operation is performed every 500 steps. The variance of the connection edge weights is calculated as the importance score. The 20% of connection edges with the lowest importance scores are selected as candidate pruning objects. A Bernoulli distribution is used to randomly prune with a probability of 0.15. After pruning, 3% of new connection edges are randomly initialized. After 15,000 training steps, the sparsity of the model stabilizes at 0.68, and the number of remaining connection edges is 91,648.

[0122] Technicians calculated the comprehensive carrying capacity evaluation index based on the resource development suitability index, ecological vulnerability index, and environmental capacity threshold output by the deep-sea environmental carrying capacity assessment model. The resource development suitability index ranged from 0.52 to 0.78, the ecological vulnerability index ranged from 0.31 to 0.64, and the environmental capacity threshold ranged from 45 to 120 t / d. The comprehensive carrying capacity evaluation index was obtained through weighted comprehensive calculation, with weight coefficients set to 0.4, 0.35, and 0.25, respectively. The mean of the calculated comprehensive carrying capacity evaluation index was 0.61, and the standard deviation was 0.13. Based on the distribution characteristics of the comprehensive carrying capacity evaluation index, technicians dynamically adjusted the pruning probability threshold using a topological adjustment function. The full-scale value of the evaluation index was set to 1.0, the normalized evaluation index to 0.61, the normalized standard deviation to 0.13, the current sparsity of the model to 0.68, the target sparsity to 0.70, and the normalized sparsity to 0.97. The calculated adjustment factor was 0.61 + 0.13 - 0.97 = -0.23. Since the adjustment factor was less than 0, a suppression pruning strategy was adopted, setting the pruning probability threshold to the minimum allowable value of 0.05, pausing further pruning operations to maintain the model's expressive power. Figure 5 As shown, the dynamic adjustment process of the pruning probability threshold reflects the adaptive optimization mechanism of the model for different bearing capacity states.

[0123] Technicians conducted spatial analysis of the assessment results for the target area and found that the comprehensive carrying capacity evaluation index of the western sea area was relatively high, with an average value of 0.73. This area has moderate water depth, stable current velocity, and sufficient nutrient supply, making it suitable for resource development activities. The comprehensive carrying capacity evaluation index of the eastern sea area was relatively low, with an average value of 0.48. This area has complex topography, strong mesoscale eddy activity, and a relatively fragile ecosystem, requiring careful planning of development activities. The comprehensive carrying capacity evaluation index of the central sea area showed a gradient change, gradually increasing from north to south, reflecting the impact of spatial heterogeneity of environmental factors on carrying capacity. Figure 6 As shown, the spatial distribution of the comprehensive carrying capacity evaluation index provides a scientific basis for the spatial layout of resource development.

[0124] It should be noted that the detailed explanations of the variables involved in this invention are shown in Tables 2, 3, and 4.

[0125] Table 2. Variable Explanation Table (Part 1)

[0126]

[0127] Table 3. Variable Explanation Table (Part Two)

[0128]

[0129] Table 4. Variable Explanation Table (Part 3)

[0130]

[0131] 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 assessing the carrying capacity of deep-sea resources and environment based on spatiotemporal big data technology, characterized in that, This involves deploying a multi-layered observation network in the deep-sea target area, using moored buoys, autonomous underwater vehicles, and remote sensing satellites to collaboratively collect data on temperature, salinity, current velocity, chlorophyll concentration, dissolved oxygen, and nutrient concentration. Simultaneously, multibeam bathymetry data is acquired. A three-dimensional seabed topography model is generated from the multibeam bathymetry data. An iterative algorithm based on the inverse problem of seawater stratification, using sound velocity profiles, is applied to the temperature and salinity data. The seawater density stratification structure is inverted through the observed sound velocity profiles, and physical constraints on the temperature-salinity relationship are applied to reconstruct the three-dimensional density stratification structure. The time-series data of sea level height are also analyzed. The sequence employs a tidal harmonic-empirical mode hybrid decomposition algorithm. First, it identifies and extracts the amplitude and phase parameters of the main tidal components to reconstruct a deterministic tidal signal. The deterministic tidal signal is then subtracted from the original sea-level time series to obtain a residual sequence. Empirical mode decomposition is then applied to the residual sequence to extract intrinsic mode functions. Based on the instantaneous frequency range of each intrinsic mode function, they are categorized into physical processes such as storm surge, mesoscale eddy processes, and seasonal variations. Multi-scale ocean dynamic mode data is output, and the partial differential equations for reconstructing the three-dimensional flow field are solved to output reconstructed three-dimensional flow field data. This paper describes the continuity and momentum equations of fluid motion based on reconstructed 3D flow field data. It performs spatiotemporal interpolation on sparse observation data to output interpolated field data and uncertainty estimation data. A modular biogeochemical model is constructed to output biogeochemical process simulation data. The reconstructed 3D flow field data, 3D density hierarchical structure, 3D seabed topography model, multi-scale ocean dynamic modal data, interpolated field data, and biogeochemical process simulation data are input into a deep-sea environmental carrying capacity assessment model. The fully connected layer of the deep-sea environmental carrying capacity assessment model adopts a sparse connection learning framework based on dynamic topology reconstruction. The importance of each connection edge is assessed by monitoring its activation frequency and gradient magnitude in multiple training batches. Connection edges with importance below a threshold are pruned with a certain probability to remove redundant edges, dynamically adjusting the network connection structure. The model outputs a resource development suitability index, an ecological vulnerability index, and an environmental capacity threshold. A comprehensive carrying capacity evaluation index is calculated based on these indices. When the comprehensive carrying capacity evaluation index falls within different intervals, a topology adjustment function is used to adjust the pruning probability threshold in the dynamic topology reconstruction sparse connection learning framework. The main tidal components include semi-diurnal tides and diurnal tides; The topology adjustment function takes the mean of the comprehensive carrying capacity evaluation index, the standard deviation of the comprehensive carrying capacity evaluation index, and the current sparsity of the model as inputs, and outputs the pruning probability threshold in the sparse connection learning framework. The modular biogeochemical model breaks down marine biogeochemical processes into phytoplankton, zooplankton, nutrient, and organic matter modules, which are then coupled together through material flux to form a complete model.

2. The method according to claim 1, characterized in that, The multi-level observation network consists of four vertical observation layers: surface observation layer, thermocline observation layer, middle observation layer, and deep observation layer. Horizontally, it sets up observation stations of different densities according to the complexity of the terrain and dynamic characteristics to form a three-dimensional observation system.

3. The method according to claim 2, characterized in that, The steps for generating a three-dimensional model of the seabed topography are as follows: applying a beamforming inverse problem-solving algorithm based on a physical acoustic model to correct the sound refraction error from the multibeam bathymetry data; using constrained triangulation to preserve the topographic structure features; introducing geological prior knowledge to construct a regularization term; and generating a three-dimensional model of the seabed topography through multi-scale surface fitting.

4. The method according to claim 3, characterized in that, The steps to reconstruct the three-dimensional density stratification structure are as follows: Specifically, an iterative algorithm based on the sound velocity profile is applied to the temperature and salinity data to solve the inverse problem of seawater stratification. The seawater density stratification structure is inverted through the observed sound velocity profile. An iterative method is used to gradually correct the density profile so that the residual between the calculated sound velocity and the observed value is minimized. At the same time, physical constraints on the temperature-salinity relationship are applied.

5. The method according to claim 4, characterized in that, The partial differential equations for reconstructing the three-dimensional flow field are solved using the domain decomposition parallel finite element method. The computational domain is divided into several subdomains and distributed to a graphics processor cluster for parallel computation. Information exchange between subdomains is achieved through alternating iterations. Combined with adaptive mesh refinement technology, the spatial resolution is improved in regions with large flow field gradients.

6. The method according to claim 5, characterized in that, The domain decomposition parallel finite element method divides a large-scale computational domain into several overlapping or non-overlapping subdomains. Each subdomain is assigned to a computational node to solve the local problem independently. Subdomains exchange information through boundary conditions to achieve global coupling. Alternating iterations are used to make the solutions on the boundaries of adjacent subdomains gradually converge.

7. The method according to claim 6, characterized in that, Spatiotemporal interpolation of sparse observation data is performed using non-stationary spatiotemporal Gaussian process regression. Local anisotropy is characterized by spatially variable coefficient covariance function. The dominant spatiotemporal modes are extracted by combining empirical orthogonal function decomposition, and the interpolated field data and uncertainty estimation data are output.

8. The method according to claim 7, characterized in that, The input layer of the deep-sea environmental carrying capacity assessment model receives reconstructed three-dimensional flow field data, three-dimensional density hierarchical structure, three-dimensional seabed topography model, multi-scale marine dynamic modal data, interpolation field data, and biogeochemical process simulation data. The feature extraction layer consists of three parallel branches that process physical oceanographic data, topographic data, and biochemical data respectively. The outputs of the three parallel branches are weighted and fused in the fusion layer through an attention mechanism.

Images