A method and system for reconstructing a three-dimensional temperature field of the ocean

Abstract

This application discloses a method and system for reconstructing a three-dimensional ocean temperature field, relating to the field of data processing technology. The method includes: acquiring multi-source heterogeneous ocean observation data and a numerical model background field to generate an input feature set; extracting features from the input feature set using a dual-branch encoder; inputting the extracted features into a multi-head spatiotemporal channel attention fusion module for dynamic calibration and fusion to obtain deeply fused latent variables; inputting the deeply fused latent variables and the numerical model background field together into a decoder based on a conditional variational autoencoder to generate high-resolution three-dimensional temperature field grid data, and simultaneously outputting a three-dimensional uncertainty field. This application utilizes finite, sparse, multi-source, and heterogeneous ocean observation data, and through mathematical and physical methods, reconstructs a continuous, complete, and high-precision three-dimensional ocean temperature field over the entire study area, solving the problem that existing technologies are not accurate enough in generating ocean three-dimensional temperature fields.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and system for reconstructing a three-dimensional temperature field in the ocean. Background Technology

[0002] Ocean temperature field is a core physical field describing the distribution of ocean temperature in three-dimensional space (longitude, latitude, and depth) and time. It is crucial foundational data for numerous fields, including global ocean circulation, climate change research, fisheries resource management, underwater military operations, and early warning of marine disasters (such as typhoons and El Niño). Accurate, complete, and high-resolution ocean temperature fields are essential for understanding and predicting ocean system behavior. However, acquiring global, full-depth ocean temperature fields has always been a significant scientific challenge, primarily due to: 1. the sparsity and non-uniformity of observational data; 2. the multi-source heterogeneity of observational data; and 3. the complex multi-scale characteristics of ocean processes. Summary of the Invention

[0003] In view of this, embodiments of this application provide a method and system for reconstructing a three-dimensional ocean temperature field to generate a continuous, complete, and high-precision three-dimensional ocean temperature field.

[0004] One aspect of this application provides a method for reconstructing a three-dimensional ocean temperature field, the method comprising the following steps:

[0005] Acquire multi-source heterogeneous ocean observation data and numerical model background field, and perform collaborative analysis and standardization processing to generate an input feature set including a standardized physical field, an uncertainty feature field, and auxiliary features;

[0006] The input feature group is extracted using a dual-branch encoder; wherein, the first branch uses a VisionTransformer encoder to process the two-dimensional raster data of satellite remote sensing, and the second branch uses a multilayer perceptron and an aggregation layer to process the field profile observation point data, thereby obtaining the encoded feature tensor.

[0007] The two-dimensional raster data and the encoded feature tensor are input into the multi-head spatiotemporal channel attention fusion module for dynamic calibration and fusion to obtain the latent variables of deep fusion.

[0008] The deeply fused latent variables and the numerical model background field are input together into a decoder based on a conditional variational autoencoder to generate high-resolution three-dimensional temperature field grid data and simultaneously output a three-dimensional uncertainty field; wherein, the decoding process of the decoder introduces a differential equation arbitrator to perform physical compliance checks and corrections on the initial decoded samples.

[0009] In some embodiments, the step of acquiring multi-source heterogeneous ocean observation data and numerical model background field, and performing collaborative analysis and standardization to generate an input feature set including a standardized physical field, an uncertainty feature field, and auxiliary features includes the following steps:

[0010] Based on real-time data density and dynamic characteristics of satellite remote sensing inversion, a non-uniform three-dimensional spatiotemporal grid is dynamically generated.

[0011] A collaborative objective function is established, and a preliminary collaborative analysis field is obtained by solving based on the three-dimensional spatiotemporal grid. The residuals of each data source relative to the collaborative analysis field are calculated, and a three-dimensional dynamic confidence weight field of each data source is generated based on the residuals.

[0012] The synergistic analysis field and the numerical model background field are standardized to obtain the standardized physical field.

[0013] The three-dimensional dynamic credibility weight field is used as an uncertainty feature field;

[0014] The standardized physical field, the uncertain feature field, and the spatiotemporal coordinate features and data mask are combined to form the input feature set.

[0015] In some embodiments, the dual-branch encoder includes:

[0016] The field decoupling and feature distillation module is used to perform preliminary decoupling and spatial feature extraction on the input features through 1x1x1 convolution and depthwise separable convolution.

[0017] A cross-modal interactive spatiotemporal attention module is used to generate query vectors with the background field features of the numerical model, generate key vectors and value vectors with the collaborative analysis field and satellite observation features, and introduce a physical constraint mask based on the three-dimensional dynamic confidence weight field to calculate physically guided attention.

[0018] Vortex sensing relative position encoding is used to calculate dynamic distance based on sea surface height anomaly fields and is injected as a relative position encoding into the calculation of attention weights.

[0019] In some embodiments, the step of the multi-head spatiotemporal channel attention fusion module fusing the encoded feature tensor is a hierarchical decision fusion, specifically including:

[0020] The first layer, feature-level fusion, employs a gated cross-attention unit, where the gated vector is generated from the uncertain feature field, and adaptively mixes the physical prior features and observed features.

[0021] The second layer, decision-level fusion, simultaneously inputs the preliminary mixed features into the physical consistency verification channel and the uncertainty decision channel. The meta-arbitrator then selects either the data-driven mode or the physical arbitration mode based on the physical violation score output by the physical consistency verification channel, and outputs the final fusion result.

[0022] In some embodiments, the step of inputting the deeply fused latent variables and the numerical model background field together into a decoder based on a conditional variational autoencoder to generate high-resolution three-dimensional temperature field mesh data includes the following steps:

[0023] Using the deep fusion of latent variables and the numerical model background field as conditions, the probability distribution parameters of the initial latent vector are obtained through the decoder neural network mapping, and then sampled and decoded to generate the initial three-dimensional temperature field sample.

[0024] The physical residual of the initial three-dimensional temperature field sample relative to the predefined thermodynamic equation is calculated, and the correction amount of the initial potential vector is derived through mathematical operations. The physically feasible three-dimensional temperature field mesh data is generated using the corrected potential vector.

[0025] In some embodiments, generating high-resolution three-dimensional temperature field grid data and simultaneously outputting a three-dimensional uncertainty field includes the following steps:

[0026] The three-dimensional temperature field grid data and the three-dimensional uncertainty field are output as a structured decision support package, specifically including:

[0027] The first output level is the three-dimensional temperature field grid data, which serves as a high-confidence physical analysis field;

[0028] The second output level is the three-dimensional uncertainty field and the confidence field: by fusing the temperature variance calculated from multiple samplings, the dynamic confidence weight field, and the standardized value of the physical residual, a comprehensive confidence index field is generated; wherein, the physical residual is the residual of the initial three-dimensional temperature field sample relative to the predefined thermodynamic equation;

[0029] The third level of output is a feature extraction and anomaly diagnosis report: the physical analysis field is automatically analyzed, and a structured report is output that includes a list of mesoscale eddies, ocean front information, temperature anomaly warnings, and vertical structure diagnosis.

[0030] In some embodiments, the method further includes a step of training the decoder, specifically including:

[0031] An adversarial training strategy based on physical constraints is adopted, in which the residual calculated by the differential equation arbitrator is used as the training signal, and the training signal is backpropagated to the decoder neural network so that the initial output of the decoder conforms to physical laws.

[0032] Another aspect of this application embodiment provides a three-dimensional ocean temperature field reconstruction system, the system comprising:

[0033] The data acquisition unit is used to acquire multi-source heterogeneous ocean observation data and numerical model background fields, and to perform collaborative analysis and standardization processing to generate an input feature set including a standardized physical field, an uncertainty feature field, and auxiliary features.

[0034] The feature extraction unit is used to extract features from the input feature group using a dual-branch encoder; wherein, the first branch uses a Vision Transformer encoder to process two-dimensional raster data from satellite remote sensing, and the second branch uses a multilayer perceptron and an aggregation layer to process on-site profile observation point data to obtain the encoded feature tensor.

[0035] The feature fusion unit is used to input the two-dimensional raster data and the encoded feature tensor into the multi-head spatiotemporal channel attention fusion module for dynamic calibration and fusion to obtain the latent variables of deep fusion.

[0036] The feature decoding unit is used to input the deep fusion latent variables and the numerical model background field into the decoder based on the conditional variational autoencoder to generate high-resolution three-dimensional temperature field grid data and simultaneously output a three-dimensional uncertainty field; wherein, the decoding process of the decoder introduces a differential equation arbitrator to perform physical compliance checks and corrections on the initial decoded samples.

[0037] Another aspect of this application embodiment provides an electronic device, including a processor and a memory;

[0038] The memory is used to store programs;

[0039] The processor executes the program to implement any of the methods described above.

[0040] Another aspect of this application provides a computer-readable storage medium storing a program that is executed by a processor to implement the method described in any of the above embodiments.

[0041] This application includes at least the following beneficial effects:

[0042] This application can acquire multi-source heterogeneous ocean observation data and numerical model background fields, and perform collaborative analysis and standardization processing to generate an input feature set including a standardized physical field, an uncertainty feature field, and auxiliary features. A dual-branch encoder is used to extract features from the input feature set. The first branch uses a Vision Transformer encoder to process two-dimensional raster data from satellite remote sensing, while the second branch uses a multilayer perceptron and aggregation layer to process on-site profile observation point data, obtaining encoded feature tensors. The two-dimensional raster data and the encoded feature tensors are input into a multi-head spatiotemporal channel attention fusion module for dynamic calibration and fusion, resulting in deeply fused latent variables. The deeply fused latent variables and the numerical model background field are jointly input into a decoder based on a conditional variational autoencoder to generate high-resolution three-dimensional temperature field grid data, and simultaneously output a three-dimensional uncertainty field. The decoding process of the decoder incorporates a differential equation arbitrator to perform physical compliance checks and corrections on the initial decoded samples. This application utilizes limited, sparse, multi-source, and heterogeneous ocean observation data to reconstruct a continuous, complete, and high-precision three-dimensional ocean temperature field across the entire study area through mathematical and physical methods, thus solving the problem that existing technologies are not accurate enough in generating ocean three-dimensional temperature fields. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 A flowchart illustrating a method for reconstructing a three-dimensional ocean temperature field, provided in an embodiment of this application;

[0045] Figure 2 An example flowchart of a method for reconstructing a three-dimensional ocean temperature field provided in this application embodiment;

[0046] Figure 3 An example structural diagram of an encoder provided in an embodiment of this application;

[0047] Figure 4 An example structural diagram of a fusion device provided in an embodiment of this application;

[0048] Figure 5 A flowchart of a differential equation arbitrator provided in this application embodiment;

[0049] Figure 6 This is a structural block diagram of a marine three-dimensional temperature field reconstruction system provided in an embodiment of this application. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0051] Before providing a detailed description of the embodiments of this application, some related technologies involved in the embodiments of this application will be described first, as follows:

[0052] 1. Sparsity and Non-uniformity of Observational Data: Despite the increasing availability of global ocean observation methods, such as Argo buoys, CTD profiles, gliders, and satellite remote sensing, direct field observation points remain extremely sparse and non-uniform compared to the vast ocean volume. Argo buoys typically provide data at 100-kilometer grid intervals of 10-day intervals, and data is extremely scarce in certain areas (such as the deep sea and polar ice regions). While satellite remote sensing can provide large-scale, high spatiotemporal resolution sea surface temperature data, it cannot directly acquire temperature information from the subsurface or even the deep sea.

[0053] 2. Multi-source heterogeneity of observation data: Data acquired from different observation platforms (such as buoys, satellites, and shipborne instruments) vary significantly in terms of accuracy, resolution, spatiotemporal coverage, and error characteristics. How to effectively integrate these multi-source heterogeneous data is a core challenge.

[0054] 3. Complex multi-scale characteristics of ocean processes: The ocean temperature field is simultaneously influenced by large-scale circulation, mesoscale eddies (the "weather" system" in the ocean), and small-scale turbulent processes, exhibiting strong multi-scale characteristics. Accurate reconstruction requires the ability to simultaneously capture these physical processes at different scales, ranging from tens of kilometers to thousands of kilometers.

[0055] Therefore, the core scientific problem of ocean temperature field reconstruction boils down to: how to estimate the continuous, complete, and high-precision three-dimensional temperature field across the entire study area using limited, sparse, multi-source, and heterogeneous ocean observation data, through mathematical and physical methods. This is essentially an inverse problem of optimally estimating a complex physical field under incomplete observation conditions.

[0056] In the field of marine science, various technical solutions have been proposed to address the problem of temperature field reconstruction. The existing technical solutions most similar to this application can be broadly categorized into two types: methods based on statistical optimal interpolation and numerical model methods based on data assimilation. Optimal interpolation (OI) is one of the oldest and most widely used reconstruction techniques in oceanography. Its core idea is to interpolate sparse observation point information onto a regular grid in a statistically optimal manner based on the spatial correlation of observational data. Numerical model methods based on data assimilation are more advanced reconstruction methods. Their core is to combine observational data with numerical dynamics models, using the models to provide physical constraints so that the reconstructed field conforms to both observations and physical laws.

[0057] Disadvantages of existing technology:

[0058] The drawbacks of the optimal interpolation (OI) method are as follows:

[0059] 1. Reliance on Preset Covariance: The performance of OI is highly dependent on the correct setting of the background error covariance function (such as correlation scale and variance). These parameters are usually fixed and difficult to accurately represent the complex multi-scale structures that actually exist in the ocean and vary with time and space.

[0060] 2. Weak ability to handle nonlinearity: OI is a linear estimation method, and its ability to capture strongly nonlinear ocean processes (such as the dramatic changes of fronts and eddies) is limited.

[0061] 3. Insufficient dynamic constraints: OI is mainly based on statistical constraints and lacks explicit ocean dynamics and physics constraints, which may lead to physical inconsistencies in the reconstructed field (such as not satisfying conservation laws).

[0062] 4. High computational complexity: For large-scale data, the computational cost of inverting the covariance matrix is ​​very high.

[0063] The main drawbacks of numerical modeling methods based on data assimilation are as follows:

[0064] 1. Highly dependent on model accuracy: The accuracy of the reconstruction results is strongly correlated with the quality of the numerical model itself (parameterization scheme, resolution, forcing field accuracy, etc.). Deviations in the model itself will be directly introduced into the reconstructed field.

[0065] 2. Extremely high computational cost: Running high-resolution global or regional ocean models and coupling data assimilation requires the support of supercomputers, which is time-consuming and labor-intensive, and it is difficult to achieve rapid reconstruction in real time or near real time.

[0066] 3. Algorithm complexity: For example, EnKF (Ensemble Kalman Filter) requires maintaining and running dozens to hundreds of model replicas, which poses a huge challenge to algorithm implementation and computing resources.

[0067] 4. Observation operator error: During the assimilation process, it is necessary to map the model state variables (such as grid point temperature) to the observation space. The error of this "observation operator" will also affect the assimilation effect.

[0068] To address the above deficiencies, this application will utilize AI to improve ocean temperature field reconstruction, enhancing and improving it in three main aspects: "combining data-driven approaches with physical constraints," "quantifying uncertainty," and "improving computational efficiency."

[0069] Reference Figure 1 This application provides a method for reconstructing a three-dimensional ocean temperature field, specifically including the following steps S100~S130:

[0070] S100: Acquire multi-source heterogeneous ocean observation data and numerical model background field, and perform collaborative analysis and standardization processing to generate an input feature set including a standardized physical field, an uncertainty feature field, and auxiliary features;

[0071] S110: The input feature group is extracted using a dual-branch encoder; wherein, the first branch uses a Vision Transformer encoder to process the two-dimensional raster data of satellite remote sensing, and the second branch uses a multilayer perceptron and an aggregation layer to process the field profile observation point data, thereby obtaining the encoded feature tensor.

[0072] S120: Input the two-dimensional raster data and the encoded feature tensor into the multi-head spatiotemporal channel attention fusion module for dynamic calibration and fusion to obtain the latent variables of deep fusion;

[0073] S130: The latent variables of the deep fusion and the background field of the numerical model are input together into the decoder based on the conditional variational autoencoder to generate high-resolution three-dimensional temperature field grid data and output a three-dimensional uncertainty field simultaneously; wherein, the decoding process of the decoder introduces a differential equation arbitrator to perform physical compliance checks and corrections on the initial decoded samples.

[0074] Optionally, the step of acquiring multi-source heterogeneous ocean observation data and numerical model background field, and performing collaborative analysis and standardization to generate an input feature set including a standardized physical field, an uncertainty feature field, and auxiliary features includes the following steps:

[0075] Based on real-time data density and dynamic characteristics of satellite remote sensing inversion, a non-uniform three-dimensional spatiotemporal grid is dynamically generated.

[0076] A collaborative objective function is established, and a preliminary collaborative analysis field is obtained by solving based on the three-dimensional spatiotemporal grid. The residuals of each data source relative to the collaborative analysis field are calculated, and a three-dimensional dynamic confidence weight field of each data source is generated based on the residuals.

[0077] The synergistic analysis field and the numerical model background field are standardized to obtain the standardized physical field.

[0078] The three-dimensional dynamic credibility weight field is used as an uncertainty feature field;

[0079] The standardized physical field, the uncertain feature field, and the spatiotemporal coordinate features and data mask are combined to form the input feature set.

[0080] Optionally, the dual-branch encoder includes:

[0081] The field decoupling and feature distillation module is used to perform preliminary decoupling and spatial feature extraction on the input features through 1x1x1 convolution and depthwise separable convolution.

[0082] A cross-modal interactive spatiotemporal attention module is used to generate query vectors with the background field features of the numerical model, generate key vectors and value vectors with the collaborative analysis field and satellite observation features, and introduce a physical constraint mask based on the three-dimensional dynamic confidence weight field to calculate physically guided attention.

[0083] Vortex sensing relative position encoding is used to calculate dynamic distance based on sea surface height anomaly fields and is injected as a relative position encoding into the calculation of attention weights.

[0084] Optionally, the step of the multi-head spatiotemporal channel attention fusion module fusing the encoded feature tensor is a hierarchical decision fusion, specifically including:

[0085] The first layer, feature-level fusion, employs a gated cross-attention unit, where the gated vector is generated from the uncertain feature field, and adaptively mixes the physical prior features and observed features.

[0086] The second layer, decision-level fusion, simultaneously inputs the preliminary mixed features into the physical consistency verification channel and the uncertainty decision channel. The meta-arbitrator then selects either the data-driven mode or the physical arbitration mode based on the physical violation score output by the physical consistency verification channel, and outputs the final fusion result.

[0087] Optionally, the step of inputting the deeply fused latent variables and the numerical model background field into a decoder based on a conditional variational autoencoder to generate high-resolution three-dimensional temperature field grid data includes the following steps:

[0088] Using the deep fusion of latent variables and the numerical model background field as conditions, the probability distribution parameters of the initial latent vector are obtained through the decoder neural network mapping, and then sampled and decoded to generate the initial three-dimensional temperature field sample.

[0089] The physical residual of the initial three-dimensional temperature field sample relative to the predefined thermodynamic equation is calculated, and the correction amount of the initial potential vector is derived through mathematical operations. The physically feasible three-dimensional temperature field mesh data is generated using the corrected potential vector.

[0090] Optionally, generating high-resolution three-dimensional temperature field grid data and simultaneously outputting a three-dimensional uncertainty field includes the following steps:

[0091] The three-dimensional temperature field grid data and the three-dimensional uncertainty field are output as a structured decision support package, specifically including:

[0092] The first output level is the three-dimensional temperature field grid data, which serves as a high-confidence physical analysis field;

[0093] The second output level is the three-dimensional uncertainty field and the confidence field: by fusing the temperature variance calculated from multiple samplings, the dynamic confidence weight field, and the standardized value of the physical residual, a comprehensive confidence index field is generated; wherein, the physical residual is the residual of the initial three-dimensional temperature field sample relative to the predefined thermodynamic equation;

[0094] The third level of output is a feature extraction and anomaly diagnosis report: the physical analysis field is automatically analyzed, and a structured report is output that includes a list of mesoscale eddies, ocean front information, temperature anomaly warnings, and vertical structure diagnosis.

[0095] Optionally, the method further includes a step of training the decoder, specifically including:

[0096] An adversarial training strategy based on physical constraints is adopted, in which the residual calculated by the differential equation arbitrator is used as the training signal, and the training signal is backpropagated to the decoder neural network so that the initial output of the decoder conforms to physical laws.

[0097] The following section will provide a detailed introduction and explanation of the solutions in the embodiments of this application, using specific application examples.

[0098] This embodiment proposes a method and system for reconstructing a three-dimensional ocean temperature field. The method includes: acquiring multi-source heterogeneous ocean observation data; standardizing and quality-controlling the data; in the encoding stage, using a Vision Transformer-based encoder to extract global spatial features from satellite remote sensing sea surface information, while simultaneously processing profile point data using dedicated network branches; and dynamically calibrating and fusing multi-source features through a novel multi-head spatiotemporal channel attention fusion module. In the decoding stage, a decoder based on a conditional variational autoencoder is used to combine the fused latent variables with the background field provided by the three-dimensional numerical model to generate high-resolution, physically consistent three-dimensional temperature field grid data, and simultaneously outputting the uncertainty estimate of the prediction. This embodiment effectively overcomes the shortcomings of traditional methods in nonlinear processing, uncertainty quantification, and physical constraints, achieving high-precision and high-efficiency reconstruction of the three-dimensional ocean temperature field.

[0099] The core of this embodiment is a deep learning framework of encoder-fusionist-decoder, and its overall workflow is as follows: Figure 2 As shown, this embodiment is described in detail below:

[0100] 1. Multi-source data preprocessing and multi-source input.

[0101] The inputs in this embodiment include three types:

[0102] Satellite remote sensing data: Globally covered two-dimensional raster data such as sea surface temperature (SST) and sea surface height (SSH).

[0103] Field profile observation data: sparse and non-uniform temperature, salinity, and depth profile data from platforms such as Argo buoys and underwater gliders (Glider).

[0104] Background field of three-dimensional numerical model: Three-dimensional temperature field forecasts or reanalysis data provided by traditional ocean numerical models (such as HYCOM and ROMS) serve as physical priors.

[0105] All data were uniformly resampled to a standard spatiotemporal grid (e.g., 1 / 4° × 1 / 4°, 50 standard depth layers, daily scale) and subjected to quality control measures such as outlier detection and missing value imputation.

[0106] Existing preprocessing methods typically process data from different sources (satellite, Argo, model) independently (e.g., interpolating them separately to a unified grid) before fusing or assimilating them. This approach ignores the inherent physical correlations and error spatial structures between different data sources, leading to information loss and error propagation.

[0107] The core idea of ​​this embodiment is to abandon the traditional paradigm of "independent interpolation first, then fusion" and establish a new paradigm of "collaborative analysis first, then feature embedding". We elevate the preprocessing process from a passive data preparation stage to an active feature extraction and uncertainty quantification stage, providing subsequent AI models with standardized high-level features rich in physical information and error characteristics, rather than raw, coarse grid data.

[0108] Detailed implementation steps:

[0109] Step 1: Create a dynamic adaptive spatiotemporal grid.

[0110] Unlike existing fixed global grids, this embodiment dynamically creates a processing grid based on real-time data density and ocean dynamic characteristics. Data density awareness: Calculates the spatial distribution density of observational data (Argo, Glider) within each time window (e.g., 24 hours). In data-sparse areas (e.g., the open ocean), a coarser base grid is used; in data-dense areas (e.g., shipping routes, Argo-dense areas), the grid is automatically densified to ensure no high-resolution information is lost. Dynamic characteristic guidance: Integrates geostrophic flow fields or sea surface height anomalies (SSHA) products from real-time satellite remote sensing to identify the boundaries of dynamic characteristics such as mesoscale eddies and fronts. Near these characteristic boundaries, the grid is adaptively densified to ensure that key dynamic processes are characterized at high resolution, while a coarser grid is maintained in the calm ocean interior to improve efficiency. Output: A non-uniform, dynamically changing three-dimensional spatiotemporal grid system, which itself contains prior knowledge of the ocean state.

[0111] Step 2: Multi-source data collaborative analysis and error field generation.

[0112] The method proposed in this embodiment does not generate "analysis values" for each data source, but instead generates "collaborative analysis values" and their "credibility weight fields".

[0113] ① Establish a collaborative objective function:

[0114] Simultaneously considering all data sources (satellite SST, satellite SSH, Argo temperature profile, and model background field Xb). The goal is not simply to interpolate Argo points onto the grid, but to find an optimal 3D temperature field analysis increment δX that minimizes the temperature difference between this field and all Argo / Glider observations, minimizes the sea surface temperature difference with satellite SST, minimizes the difference between its sea surface dynamic height (calculated from the temperature field through thermal relations) and satellite SSH observations, and minimizes the deviation from the model background field Xb (as a constraint). This is a lightweight, local 3D variational (3D-Var) co-analysis process, but its purpose is not to produce a final product, but rather to produce intermediate features.

[0115] ② Generate the synergistic analysis field and error field:

[0116] Solving the objective function yields a preliminary, high-quality collaborative analysis field. The key innovation is the simultaneous computation of the residuals (innovation vectors) of this collaborative analysis field relative to each data source. These residual fields are calculated and stored, clearly revealing the spatial systematic biases and inconsistencies among the different data sources.

[0117] ③ Generate Dynamic Confidence Weight Fields:

[0118] Based on the residuals from step 2, a spatialized confidence weight field is generated for each data source. The calculation principle is that for a given data source, a small residual in a certain region indicates high consensus and reliability with other sources, resulting in a large weight W; a large residual suggests potential errors or the source's inability to reflect specific phenomena, leading to a smaller weight W. For example, in cloud-covered areas where satellite SSTs typically fail, the residuals may be incalculable, and the weight W_sat is automatically set to 0; in dense Argo buoy areas, W_argo is close to 1; and in regions where model bias persists (such as the western boundary current region), W_model will be lower. The output not only includes a collaborative analysis field X_analysis, but more importantly, a set of three-dimensional dynamic weight fields: W_sat(x,y,z), W_argo(x,y,z), and W_model(x,y,z).

[0119] Step 3: Feature engineering and embedding for AI models.

[0120] The above output is transformed into a feature form suitable for input to a deep learning model.

[0121] ① Input feature group one: Standardized physical field.

[0122] The co-analysis field X_analysis is normalized to obtain a three-dimensional "normalized temperature anomaly field". The model background field Xb is normalized and used as the input for the "physical prior field". Two-dimensional fields such as satellite SST and SSH are normalized and then input.

[0123] ② Input feature group two: Uncertainty feature field.

[0124] The credibility weight fields W_sat, W_argo, and W_model generated in step two are directly fed as additional input channels to the subsequent encoder. This explicitly and quantitatively informs the AI ​​model of the uncertainties and data quality information in the preprocessing process. The model no longer blindly learns and merges data, but clearly knows "at which location and depth, which data source is more reliable," thereby greatly enhancing the model's learning efficiency and the physical rationality of the final reconstruction.

[0125] ③ Input feature group three: auxiliary features.

[0126] Provide spatiotemporal coordinate features, such as longitude, latitude, depth, annual cycle, and daily cycle, encoded as feature maps to help the model capture periodicity and geographical effects. Provide data masks: clearly label which grid points have actual observations and which have missing values, avoiding the model's blind guessing of missing values.

[0127] 2. Feature extraction and encoding.

[0128] This step uses a dual-branch encoder structure to process different types of data separately:

[0129] Satellite Data Branch (ViT Branch): Treats 2D field data such as SST and SSH as multi-channel "images." It employs the Vision Transformer (ViT) as the core encoder. ViT segments the image into patches and efficiently captures long-range global dependencies of features such as sea surface temperature and mesoscale eddies through a self-attention mechanism, overcoming the limitation of the limited receptive field of CNNs. The output is a feature vector representing the global state of the sea surface. Profile Point Data Branch: Processes sparse profile point data. Each point contains information such as longitude, latitude, depth, and temperature. First, a multilayer perceptron (MLP) with shared weights is used to independently encode each observation point, mapping it to a high-dimensional feature space. Then, interpolation / pooling layers are used to aggregate these discrete point features into a feature tensor consistent with the scale of the target 3D grid, capturing the vertical structure information directly revealed by the observation points.

[0130] Existing methods typically stack different data sources into a single multi-channel input or process them in separate branches before stitching them together. This approach fails to explicitly model the complex, nonlinear spatiotemporal physical relationships between different modalities, such as the dynamic correlation between sea level height (SSH) anomalies and lower-level temperature field anomalies, or the relationship between sea surface temperature (SST) fronts and vertical temperature structures.

[0131] The core idea of ​​this embodiment is to abandon the simple paradigm of "channel stacking" or "independent processing" and establish a new interactive coding paradigm "centered on the physical field and using attention as a bridge". We no longer treat data as independent images or point sets, but as an interconnected physical signal system. Through a series of innovative attention mechanisms, the model can autonomously discover and utilize these physical rules for efficient coding.

[0132] Detailed structure and steps:

[0133] The encoder receives the multi-channel three-dimensional feature tensor output from the preprocessing stage, and its structure is as follows: Figure 3As shown, it includes the following core innovative modules:

[0134] Module 1: Field Decoupling and Characteristic Distillation Module.

[0135] The goal is to replace the initial part of a standard CNN encoder by performing preliminary decoupling and purification of different channels of the input tensor (representing data from different physical sources) before deep feature extraction. The innovative structure consists of: a 1x1x1 convolutional layer operating on the channel dimension; its function is not fusion, but linear projection and decoupling. It recombines features from different sources such as the co-analysis field, weight field, and background field in a high-dimensional space, initially separating feature groups representing different physical meanings (such as "average state features," "uncertainty features," and "dynamic anomaly features"). Depthwise separable convolutions are then performed on the decoupled feature maps along the spatial dimensions (longitude, latitude, and depth). This efficiently extracts local spatial features (such as gradients and edges) while significantly reducing the number of parameters and avoiding early information confusion that might occur with channel convolutions. The GELU activation function provides a non-linear transformation.

[0136] This module implements "decoupling first, then extraction," avoiding the problem of traditional CNNs mixing all data sources together for convolution at the beginning, which leads to blurred physical information and provides a purer and more physically meaningful feature base for subsequent attention mechanisms.

[0137] Module 2: Cross-modal interactive spatiotemporal attention module.

[0138] The goal is to establish global dependencies between features from different modalities, locations, and times (if the input is a sequence), but these dependencies are guided by physical knowledge. An innovative mechanism is designed—physics-guided Query-Key-Value generation: In traditional Transformers, Q, K, and V come from the same input. This embodiment differentiates the generation of Q (query vector, representing "what is wanted" or "what is the current focus"), K (key vector, representing the label or identifier of the information source "what it is"), and V (value vector, representing the actual content contained in the information source "). The query (Q) mainly originates from the background field features of the numerical model. Why does the model background field contain prior state information constrained by physical equations? Using it as the "questioner" (Q) means allowing the physical prior to actively ask: "What happened where observations and my predictions are inconsistent?". Key (K) & Value (V): mainly originate from the co-analysis field and satellite observation features. They act as the "questioned parties" (K, V), providing observational evidence to answer the physical prior's question. Attention Calculation: Attention(Q, K, V) = Softmax( (QK^T) / √d_k + M ) · V. Mask Matrix M (Innovation): A physically constrained mask is introduced here. For example, based on the confidence weight field W_argo generated in the preprocessing stage, we can construct a mask that significantly reduces the attention weight of surface satellite information on the deep-seated influence below the depth not observed by the Argo buoy, as this does not conform to the fundamental physical constraints of ocean vertical remote sensing. This is equivalent to adding physical rules to the attention calculation. Multi-head Mechanism: Multi-head attention is adopted, allowing different "heads" to focus on learning different kinds of physical relationships (e.g., one head focuses on the thermal effects of eddies, and another head focuses on the dynamic adjustments of fronts).

[0139] Module 3: Vortex Sensing Relative Position Encoding.

[0140] The aim is to replace standard absolute position encoding and inject more ocean dynamics-compliant positional information into the attention mechanism. The relative positional offsets (Δx, Δy, Δz) between any two points (i, j, k) and (i', j', k') in the grid are calculated. In addition to using Euclidean distance, dynamic distance is introduced: utilizing the preprocessed sea surface height anomaly (SSHA) field, the geostrophic velocity shear or vorticity difference between the two points is calculated as an additional relative positional encoding. These relative physical quantities are then encoded using a small MLP and incorporated into the calculation of attention weights. The innovation lies in enabling the model to understand that "two points within the same vortex, even if far apart in absolute distance, have a stronger physical correlation than two points that are close in absolute distance but located inside or outside the vortex." This is a crucial step in deeply embedding core concepts of ocean dynamics into the model structure.

[0141] Module 4: Uncertainty-aware feedforward network.

[0142] The aim is to incorporate uncertainty information into feature transformation. A standard feedforward network (FFN) consists of two linear layers and an activation function. This embodiment introduces a gating mechanism and modulation into the FFN: the confidence weight field (such as W_argo) in the preprocessed input is passed through a lightweight network to generate a set of modulation coefficients Λ. The calculation of FFN becomes: FFN(X) = (Linear2( GELU( Linear1(X) ) )) ⊙ Λ + b, where ⊙ represents element-wise multiplication. The innovation lies in modulating the strength of the feature transformation with uncertainty information. In high confidence regions (large Λ values), the features are fully transformed; in low confidence or data-deficient regions (small Λ values), the feature transformation is suppressed, relying more on information passed from the high confidence region. This greatly enhances the rationality and robustness of the model output.

[0143] 3. Multi-source feature fusion (multi-head spatiotemporal channel attention mechanism).

[0144] The feature tensors generated from the two branches are concatenated and fed into a multi-head spatiotemporal channel attention fusion module. Channel attention: Learns the relative importance of different data sources (such as SST, SSH, and point observations), automatically assigning greater weight to data sources with higher credibility. Spatial attention: Focuses on regions with greater information content (such as vortex edges and frontal regions) or regions lacking observations that require focused inference in the horizontal dimension. Temporal attention (if the input is a time series): Focuses on historical moments most relevant to the prediction time. The "multi-head" mechanism allows the model to simultaneously focus on information from different representation subspaces, thereby capturing the complex relationships between features more richly and comprehensively. The output of this module is a deeply fused latent variable containing multi-source information.

[0145] Existing fusion methods are essentially "static" or "passive." Concatenation / summation assumes all features contribute equally, failing to distinguish source reliability. Simple attention, while calculating weights, relies solely on data similarity, lacking physical rules and easily leading to physically inconsistent fusion results. The core idea of ​​this embodiment is to abandon the passive paradigm of "one-time weighted fusion" and establish a new proactive fusion paradigm of "hierarchical decision-making, active verification, and physical arbitration." We construct the fusion process as a data-driven decision-making system supervised by physical rules. This system determines, hierarchically and step-by-step, when, where, and how to trust which data source, ultimately ensuring the output is physically self-consistent.

[0146] The detailed structure and steps are as follows:

[0147] This fusion unit receives multimodal features from the encoder. Its core is a hierarchical decision-making process, the structure of which is as follows: Figure 4 As shown, it specifically includes:

[0148] First layer: Feature-level fusion (adaptive fusion based on gated cross-attention).

[0149] The goal is to perform preliminary and refined feature interaction and compensation, replacing simple concatenation operations. The innovative structure is the Gated Cross-Attention Unit (GCAU). Multiple GCAUs are used, with core units (such as those between physical prior features and observed features) performing cross-attention calculations. The query (Q) comes from the physical prior features. The key (K) and value (V) come from the observed features. The cross-attention output is calculated as: Cross_Attn = Attention(Q, K, V). The innovation lies in the gating mechanism: a learnable gating vector is introduced, generated from uncertain features through a sigmoid activation function, with a value range of [0,1]. The final output is: GCAU_Output = Gate ⊙ Cross_Attn + (1 - Gate) ⊙ Q.

[0150] Second layer: Decision-level fusion (final decision under physical rule arbitration).

[0151] It inputs the initial mixed features into a decision network, which contains two parallel channels and an arbitrator:

[0152] ①Physical consistency verification channel.

[0153] The goal is to proactively assess whether the initial fusion results violate known physical rules. This channel includes a lightweight physics operator layer (e.g., calculating the horizontal gradient of the temperature field, vertical stratification parameters, or calculating the residuals of thermodynamic equations by simplifying equations). A Physics Violation Score is generated to quantify the degree of physical inconsistency in the current fusion result.

[0154] ② Uncertainty Decision Pathway.

[0155] The goal is to make decisions based on data credibility. This channel receives initial mixed features and raw uncertainty features, and through a small neural network, generates a data-driven decision feature.

[0156] ③ Meta-Arbiter.

[0157] The goal is to receive the outputs from the two channels mentioned above and make a final decision. The decision logic is as follows: if the physical violation score is below a threshold (τ), it indicates that the fusion result is physically reasonable, and the meta-arbitrator primarily adopts the output of the uncertainty decision channel (data-driven mode). If the physical violation score is above the threshold (τ), it indicates that a physically impossible fusion result has occurred (such as the sudden disappearance of the thermocline), and the meta-arbitrator will ignore the data-driven decision and initiate a physical correction procedure. This procedure generates a correction vector based on the physical violation score, forcibly adjusting the initial mixed features to bring them back to a physically reasonable range (physical arbitration mode).

[0158] 4. Decoding and physical constraints.

[0159] The decoder employs a Conditional Variational Autoencoder (C-VAE) architecture, its role being to decode the fused latent variables into the final 3D temperature field grid. Probabilistic deep learning: The latent variables in a C-VAE are not fixed values, but rather a probability distribution (typically Gaussian). This means that for the same set of inputs, the decoder can sample and generate slightly different but plausible outputs, thus naturally providing an estimate of the uncertainty in the prediction. We can calculate the mean and variance of the temperature at each grid point through multiple samplings; a larger variance indicates higher uncertainty.

[0160] Incorporating physical constraints:

[0161] Numerical model as prior: The 3D background field provided by the numerical model is input as a condition into the C-VAE decoder. This is equivalent to providing a strong physical prior for the data-driven model, guiding the reconstruction results to approach a physically reasonable state. Physical loss function: In the loss function during the training phase, in addition to the regular data reconstruction loss (such as mean squared error), an additional physical constraint loss is added. For example, the gradient of the reconstructed temperature field is calculated to ensure that it satisfies the basic constraints of the thermodynamic equations, and its residual is minimized. This ensures that the generated results basically conform to physical laws even in regions with scarce data.

[0162] The core idea is to abandon the passive paradigm of "post-generation correction" or "soft constraints" and establish a new paradigm of active physical constraints that is "embedded in the process, sample arbitration, and hard guarantees." We transform physical laws from an external "suggestion" to an internal "arbitration rule," directly enforcing physical constraints during the decoding and sampling generation process. This ensures that every generated sample is physically feasible and ultimately provides a probabilistic output.

[0163] Detailed structure and steps:

[0164] This decoder receives fused feature vectors from the fusion generator. Its core innovation lies in introducing a differential equation arbitrator (DEA) to guide the sampling process. Its workflow is as follows: Figure 5 As shown, it specifically includes:

[0165] Module 1: Probabilistic Framework for Conditional Variational Autoencoders (CVAEs)

[0166] The aim is to provide probabilistic generation capabilities and output uncertainty estimates. In this embodiment, the CVAE's condition includes not only the fused feature vector but also explicitly the numerical pattern background field. This means the generation process is strongly guided towards a physically plausible initial region. Initial sampling: The decoder neural network maps the fused features to parameters (μ, σ), defining a Gaussian distribution. An initial latent vector Z is obtained from this distribution, and an initial three-dimensional temperature field sample T is generated through the decoder network.

[0167] Module 2: Differential Equation Arbitrator

[0168] This is the core innovation of the entire decoding process; it is a differentiable mathematical operation module that is independent of the main decoder network.

[0169] The purpose is to perform physical compliance checks and hard corrections on the initial sample T*.

[0170] Workflow:

[0171] ① Physical field construction: Receive the initial sample T* output by the decoder.

[0172] ② Checking the physical laws:

[0173] DEA internally predefines core physical laws that must be strictly followed, such as the simplified thermodynamic equation: ℵ(T) =∂T / ∂t + u·∇T - ∇·(k∇T), where ℵ is the differential operator. DEA calculates the residual R of this equation: R = |ℵ(T*) - S|, where S is the source term (such as solar radiation heating), which can be ignored on short timescales or is used as a known input.

[0174] ③ Residual analysis and correction:

[0175] The core innovation of this module lies in the fact that it does not directly use R as the loss function, but rather as a mathematical constraint.

[0176] By using automatic differentiation (Autograd) or solving the adjoint equation, DEA calculates the correction δZ to the initial latent vector Z*, so that the temperature field T_final corresponding to the corrected sample Z_final = Z* + δZ can significantly reduce or even eliminate the physical residual R.

[0177] That is, the purpose of DEA is to find a δZ such that ℵ( Decoder(Z* + δZ) ) ≈ S.

[0178] The output is a physically feasible final sample T_final = Decoder(Z_final). Since the correction δZ is usually very small, T_final preserves the features learned from the data to the greatest extent possible while satisfying physical laws.

[0179] Module 3: Adversarial Training Strategies for Physical Constraints

[0180] The goal is to train the decoder neural network so that its initial output T* is as close as possible to physical reality, thereby reducing the correction magnitude of the DEA and improving efficiency and stability. The residual R calculated by the DEA is used as a training signal and backpropagated to the decoder neural network. From a game theory perspective, the decoder attempts to generate a sample that both conforms to the data characteristics and is unassailable by the DEA (i.e., a small residual R). The DEA, acting as a fixed "physical referee," rigorously examines each physical law and provides gradient information on how to correct it. This training method allows the decoder to gradually learn the internalized representation of physical laws, ultimately making its initial sampling highly physically reliable.

[0181] 5. System output.

[0182] The system ultimately outputs two products: ① High-resolution 3D temperature field analysis field: optimally estimated 3D grid data. ② 3D uncertainty field: estimated variance of temperature values ​​at each grid point, clearly indicating the spatial distribution of prediction confidence (e.g., low uncertainty in shipping lanes, high uncertainty in unobserved areas in the open ocean).

[0183] The core idea of ​​this embodiment is to abandon the paradigm of "passively providing data" and establish a new intelligent output paradigm of "proactively providing insights, quantifying risks, and supporting decision-making." The system outputs not just data, but a "decision support package" containing multi-dimensional, actionable, and interpretable information, enabling even non-expert users to quickly understand the current ocean state and make reliable decisions. The system's output is a structured data packet (such as NetCDF / HDF5 format or a JSON object via API interface), containing the following four inseparable core layers:

[0184] Output Level 1: High-Confidence Physical Analysis Field.

[0185] The physically consistent 3D temperature field, generated and arbitrated by the PCSD-DEA decoder, represents the optimal estimate of the temperature state. This field strictly satisfies the physical conservation laws (such as thermodynamic equations) integrated into the DEA, eliminating non-physical oscillations or structures. This is a fundamental difference from the output of any traditional data assimilation or pure AI method. It represents the state that best matches the multi-source observation data among all physically possible scenarios.

[0186] Output Level 2: Three-dimensional uncertainty and credibility field.

[0187] A three-dimensional field is completely isomorphic to the physical analysis field, but the value of each grid point is no longer temperature, but a composite confidence index (CCI) fused from three pieces of information, with a value range of [0,1]. The temperature variance calculated through multiple samplings represents the cognitive uncertainty of the model itself. The dynamic confidence weight field generated in the preprocessing stage represents the reliability of the data source. The magnitude of the physical residuals calculated by the differential equation arbitrator, after standardization, represents the degree of agreement between the state of the region and the physical laws. The smaller the residuals, the higher the confidence. CCI is the first to quantify and integrate the three major sources of uncertainty—data, model, and physics—into an intuitive indicator. Users can directly view the CCI field, just like viewing the "probability of precipitation" in a weather forecast, and clearly identify high-risk, low-confidence areas (such as unobserved open ocean areas or dynamically complex frontal zones) and high-risk, high-confidence areas (such as confirmed warm vortices).

[0188] Output Level 3: Feature Extraction and Anomaly Diagnosis Report.

[0189] The system automatically analyzes the output physical analysis field in real time, generating a structured diagnostic report. Innovative outputs include: ① Automatic identification and inventory of mesoscale eddies: including eddy center location, radius, intensity (rotation speed, thermonuclear temperature anomaly), type (cyclone / anticyclone), and vertical structure. ② Quantification of ocean front location and intensity: automatically identifies and marks the location of major temperature fronts and calculates their intensity gradients. ③ Temperature anomaly early warning: compares with climatological averages to identify significant positive / negative temperature anomaly areas (such as oceanic heat waves) and calculates their area, volume, intensity, and confidence level. ④ Vertical structure diagnosis: automatically outputs key vertical parameter fields such as mixing layer depth and isotherm depth.

[0190] The system represents a leap from "data provision" to "information extraction," directly outputting quantitative summaries of marine phenomena of interest to users, significantly reducing post-processing workload. The report, presented as structured data (such as JSON), can be automatically accessed and processed by downstream decision-making systems, early warning systems, or robotic processes, enabling fully automated operations.

[0191] Output level four: Dynamic adaptive product generation interface.

[0192] A built-in, lightweight, interactive API. User-defined threshold queries: Users can send a request via the API without downloading the entire data package, for example: "Return the volume of water bodies with a temperature above 25°C in all regions with CCI > 0.9" or "Return the route area with frontal intensity greater than 0.15°C / km". On-demand slicing and aggregation: Users can request the average or integral values ​​of any vertical profile, horizontal layer, or specific spatial region; the system calculates and returns the results instantly. "What-if" scenario simulation: Users can input a hypothetical observation (e.g., "If an observation point with temperature T is added at coordinates (X,Y)"), and the system uses a built-in encoder to quickly infer the impact of this hypothetical observation on the analysis field and uncertainty field, thus supporting adaptive observation planning.

[0193] The system is transformed from a "data publisher" to a "problem-solving service," providing users with unprecedented flexibility. The fourth-level output is directly linked to the design of the observation system and action decisions, forming a complete closed loop from perception to decision-making.

[0194] In summary, this embodiment includes the following key technical solutions:

[0195] (1) Breaking the traditional fixed global grid paradigm, a method for dynamically generating non-uniform grids based on real-time data density and ocean dynamic activity is proposed, achieving a high degree of matching between computing resources and information density.

[0196] (2) A lightweight variational assimilation framework was innovatively adopted as a preprocessing step. Its purpose is not to produce a final product, but to simultaneously analyze the inconsistencies of all data sources and generate a residual field that reveals the error space structure, thus realizing collaborative analysis and pre-quantification of uncertainty.

[0197] (3) The dynamic credibility weight field generated by preprocessing is provided as an explicit input feature to the core AI model. This enables the AI ​​model to change from "blind learning" to "informed learning," knowing where to trust which data source, greatly improving the physical rationality and efficiency of the fusion. It realizes explicit embedding of uncertainty information, which is the fundamental difference from all existing preprocessing methods.

[0198] (4) The output of the entire preprocessing process is not an isolated, interpolated data field, but a composite feature tensor tailored for the AI ​​model, containing a standardized physical field, an uncertain weight field and auxiliary information. It achieves seamless and efficient connection with the deep learning model and is an end-to-end feature output.

[0199] (5) Heterogeneous QKV generation mechanism: The core innovation of differentiated generation of Q, K, and V is clarified. That is, Q originates from physical priors (mode field), while K and V originate from observational evidence (cooperative field, satellite field). This breaks the tradition of homogeneity of Q, K, and V in self-attention and establishes a new paradigm of "physical questioning-observational response", which is the fundamental reason why the method has physical guidance.

[0200] (6) Embedded physical constraints: The physical constraint mask matrix (M) and the vortex sensing relative position encoding deeply embed the prior knowledge of the ocean field (such as remote sensing limitations and vortex dynamics) into the network structure in a computable form, realizing the organic combination of data-driven and physical-driven approaches, and ensuring the physical rationality of the output results.

[0201] (7) Uncertainty Flow Modulation: The uncertainty-aware feedforward network clarifies how uncertainty information (credibility weight field) participates in and regulates the deep feature extraction process throughout (through modulation coefficient Λ). This enables the model's behavior to be synchronized with data quality in real time, making the output results more reliable. This is a deep realization of the idea of ​​uncertainty quantification.

[0202] (8) Specialized structural design for marine fields: The field decoupling and feature distillation module (1x1x1 convolution + depth separable convolution) is a feature extraction starting step tailored to the multi-source and heterogeneous characteristics of marine data. It is more efficient than simply stacking channels and then directly using standard convolution or Transformer, and can retain clearer physical information.

[0203] (9) Paradigm shift from “weighting” to “decision-making”: The HD-PIFN proposed in this invention is no longer a simple feature weighted summation function, but an intelligent decision-making system with hierarchical judgment logic. It simulates the thought process of domain experts when dealing with contradictory data: first try to merge, then check the rationality, and finally make arbitration when rules and evidence conflict.

[0204] (9) Gated Cross Attention (GCAU): By introducing cross attention driven by uncertainty, a delicate, adaptive, and conditional feature interaction between physical priors and observational evidence is achieved, which is fundamentally superior to simple dot product attention or splicing.

[0205] (10) Explicit physical rule arbitration mechanism: The system innovatively sets up a physical consistency verification channel and a meta-arbitrator, transforming physical rules from an implicit constraint (such as through a loss function) into an explicit, proactive, and interventionist decision factor. This is a key breakthrough in ensuring the physical credibility of the fusion results and eliminating the possibility of AI models generating "physical illusions".

[0206] (11) Uncertainty-driven global decision-making: Uncertain information is no longer just the weight of a certain attention point, but runs through the entire process of gating (first layer) and decision-making (second layer), becoming the core variable controlling the fusion behavior, making the system behavior highly robust and self-aware.

[0207] (12) A paradigm leap from soft constraints to hard arbitration: This invention introduces a differential equation arbitrator (DEA), which is not a suggested loss term, but a mandatory, differentiable mathematical arbitration unit. It intervenes and corrects directly on the sampling path, theoretically guaranteeing the physical consistency of the final output sample, which is the fundamental difference from soft constraint methods such as PINNs.

[0208] (13) Physical guarantees embedded in the process: Physical constraints are embedded in the generation process, rather than being processed afterward. This “generation-check-correction” closed loop ensures that the generation steps from the first to the last are under the supervision of physical laws, achieving a true “physically guaranteed output”.

[0209] Efficient gradient correction: DEA utilizes automatic differentiation techniques to directly calculate how to correct the latent variable Z to satisfy the physical equations, which is a gradient-based, highly efficient optimization method. It is more accurate and efficient than post-processing corrections in high-dimensional physical spaces.

[0210] (14) Adversarial internalization learning: Through an adversarial training strategy with physical constraints, the decoder network is forced to learn physical laws during training, eventually becoming a generator with "extensive physical knowledge". This greatly reduces the reliance on DEA correction during the inference stage and improves efficiency.

[0211] (15) Unity of probabilistic output and physical guarantee: This invention successfully unifies two seemingly contradictory goals: probabilistic deep learning (providing uncertainty estimation) and hard physical constraints (ensuring sample rationality). The multiple samples obtained by the user are both possible data realities and rigorous physical constraints, making the uncertainty estimation results more credible and useful.

[0212] (16) A fundamental shift in output paradigm: from outputting “data” to outputting “insights + services”. Providing complete solutions that integrate data products, uncertainty quantification, phenomenon diagnosis and interactive services, rather than a single document.

[0213] (17) Quantification of uncertainty through multi-source fusion: The proposed Comprehensive Credibility Index (CCI) unifies the three major sources of uncertainty—data, model, and physics—for the first time, generating the world's first truly meaningful "ocean credibility map" that can be directly used for risk decision-making.

[0214] (18) Embedded intelligent diagnostic function: The system has built-in automated marine phenomenon recognition and feature extraction algorithms, and outputs them directly in the form of structured reports, which solidifies domain knowledge in the system, greatly reducing the threshold for use and professional dependence.

[0215] (19) Interactive and customizable output interface: The provided dynamic adaptive product generation interface enables the system to have "question and answer" and "scenario simulation" capabilities, supporting true interactive scientific analysis and adaptive observation planning, which is impossible for static data products.

[0216] Reference Figure 6 This application provides a three-dimensional ocean temperature field reconstruction system, comprising:

[0217] The data acquisition unit is used to acquire multi-source heterogeneous ocean observation data and numerical model background fields, and to perform collaborative analysis and standardization processing to generate an input feature set including a standardized physical field, an uncertainty feature field, and auxiliary features.

[0218] The feature extraction unit is used to extract features from the input feature group using a dual-branch encoder; wherein, the first branch uses a Vision Transformer encoder to process two-dimensional raster data from satellite remote sensing, and the second branch uses a multilayer perceptron and an aggregation layer to process on-site profile observation point data to obtain the encoded feature tensor.

[0219] The feature fusion unit is used to input the two-dimensional raster data and the encoded feature tensor into the multi-head spatiotemporal channel attention fusion module for dynamic calibration and fusion to obtain the latent variables of deep fusion.

[0220] The feature decoding unit is used to input the deep fusion latent variables and the numerical model background field into the decoder based on the conditional variational autoencoder to generate high-resolution three-dimensional temperature field grid data and simultaneously output a three-dimensional uncertainty field; wherein, the decoding process of the decoder introduces a differential equation arbitrator to perform physical compliance checks and corrections on the initial decoded samples.

[0221] It is understood that the content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0222] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this application are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.

[0223] Furthermore, although this application is described in the context of functional modules, it should be understood that, unless otherwise stated to the contrary, one or more of the described functions and / or features may be integrated into a single physical system and / or software module, or one or more functions and / or features may be implemented in a separate physical system or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding this application. Rather, given the properties, functions, and internal relationships of the various functional modules in the system disclosed herein, the actual implementation of the module will be understood within the scope of conventional technology for an engineer. Therefore, those skilled in the art can implement the application set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and not intended to limit the scope of this application, which is determined by the full scope of the appended claims and their equivalents.

[0224] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0225] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, system, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, system, or device). For the purposes of this specification, "computer-readable medium" can mean any system that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, system, or device.

[0226] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic systems) with one or more wires, portable computer disk drives (magnetic systems), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic systems, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0227] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0228] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0229] Although embodiments of this application have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the claims and their equivalents.

[0230] The above is a detailed description of the preferred embodiments of this application, but this application is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A method for reconstructing a three-dimensional temperature field in the ocean, characterized in that, The method includes the following steps: Acquire multi-source heterogeneous ocean observation data and numerical model background field, and perform collaborative analysis and standardization processing to generate an input feature set including a standardized physical field, an uncertainty feature field, and auxiliary features; The input feature group is extracted using a dual-branch encoder; wherein, the first branch uses a VisionTransformer encoder to process the two-dimensional raster data of satellite remote sensing, and the second branch uses a multilayer perceptron and an aggregation layer to process the field profile observation point data, thereby obtaining the encoded feature tensor. The two-dimensional raster data and the encoded feature tensor are input into the multi-head spatiotemporal channel attention fusion module for dynamic calibration and fusion to obtain the latent variables of deep fusion. The deeply fused latent variables and the numerical model background field are input together into a decoder based on a conditional variational autoencoder to generate high-resolution three-dimensional temperature field grid data and simultaneously output a three-dimensional uncertainty field; wherein, the decoding process of the decoder introduces a differential equation arbitrator to perform physical compliance checks and corrections on the initial decoded samples.

2. The method for reconstructing a three-dimensional ocean temperature field according to claim 1, characterized in that, The process of acquiring multi-source heterogeneous ocean observation data and numerical model background fields, and performing collaborative analysis and standardization to generate an input feature set including a standardized physical field, an uncertainty feature field, and auxiliary features includes the following steps: Based on real-time data density and dynamic characteristics of satellite remote sensing inversion, a non-uniform three-dimensional spatiotemporal grid is dynamically generated. A collaborative objective function is established, and a preliminary collaborative analysis field is obtained by solving based on the three-dimensional spatiotemporal grid. The residuals of each data source relative to the collaborative analysis field are calculated, and a three-dimensional dynamic confidence weight field of each data source is generated based on the residuals. The synergistic analysis field and the numerical model background field are standardized to obtain the standardized physical field. The three-dimensional dynamic credibility weight field is used as an uncertainty feature field; The standardized physical field, the uncertain feature field, and the spatiotemporal coordinate features and data mask are combined to form the input feature set.

3. The method for reconstructing a three-dimensional ocean temperature field according to claim 2, characterized in that, The dual-branch encoder includes: The field decoupling and feature distillation module is used to perform preliminary decoupling and spatial feature extraction on the input features through 1x1x1 convolution and depthwise separable convolution. A cross-modal interactive spatiotemporal attention module is used to generate query vectors with the background field features of the numerical model, generate key vectors and value vectors with the collaborative analysis field and satellite observation features, and introduce a physical constraint mask based on the three-dimensional dynamic confidence weight field to calculate physically guided attention. Vortex sensing relative position encoding is used to calculate dynamic distance based on sea surface height anomaly fields and is injected as a relative position encoding into the calculation of attention weights.

4. The method for reconstructing a three-dimensional ocean temperature field according to claim 2, characterized in that, The step of fusing the encoded feature tensor in the multi-head spatiotemporal channel attention fusion module is a hierarchical decision fusion, specifically including: The first layer, feature-level fusion, employs a gated cross-attention unit, where the gated vector is generated from the uncertain feature field, and adaptively mixes the physical prior features and observed features. The second layer, decision-level fusion, simultaneously inputs the preliminary mixed features into the physical consistency verification channel and the uncertainty decision channel. The meta-arbitrator then selects either the data-driven mode or the physical arbitration mode based on the physical violation score output by the physical consistency verification channel, and outputs the final fusion result.

5. The method for reconstructing a three-dimensional ocean temperature field according to claim 1, characterized in that, The process of inputting the deeply fused latent variables and the numerical model background field into a decoder based on a conditional variational autoencoder to generate high-resolution three-dimensional temperature field mesh data includes the following steps: Using the deep fusion of latent variables and the numerical model background field as conditions, the probability distribution parameters of the initial latent vector are obtained through the decoder neural network mapping, and then sampled and decoded to generate the initial three-dimensional temperature field sample. The physical residual of the initial three-dimensional temperature field sample relative to the predefined thermodynamic equation is calculated, and the correction amount of the initial potential vector is derived through mathematical operations. The physically feasible three-dimensional temperature field mesh data is generated using the corrected potential vector.

6. The method for reconstructing a three-dimensional ocean temperature field according to claim 1, characterized in that, The process of generating high-resolution three-dimensional temperature field grid data and simultaneously outputting a three-dimensional uncertainty field includes the following steps: The three-dimensional temperature field grid data and the three-dimensional uncertainty field are output as a structured decision support package, specifically including: The first output level is the three-dimensional temperature field grid data, which serves as a high-confidence physical analysis field; The second output level is the three-dimensional uncertainty field and the confidence field: by fusing the temperature variance calculated from multiple samplings, the dynamic confidence weight field, and the standardized value of the physical residual, a comprehensive confidence index field is generated; wherein, the physical residual is the residual of the initial three-dimensional temperature field sample relative to the predefined thermodynamic equation; The third level of output is a feature extraction and anomaly diagnosis report: the physical analysis field is automatically analyzed, and a structured report is output that includes a list of mesoscale eddies, ocean front information, temperature anomaly warnings, and vertical structure diagnosis.

7. The method for reconstructing a three-dimensional ocean temperature field according to claim 1, characterized in that, The method further includes a step of training the decoder, specifically including: An adversarial training strategy based on physical constraints is adopted, in which the residual calculated by the differential equation arbitrator is used as the training signal, and the training signal is backpropagated to the decoder neural network so that the initial output of the decoder conforms to physical laws.

8. A three-dimensional ocean temperature field reconstruction system, characterized in that, The system includes: The data acquisition unit is used to acquire multi-source heterogeneous ocean observation data and numerical model background fields, and to perform collaborative analysis and standardization processing to generate an input feature set including a standardized physical field, an uncertainty feature field, and auxiliary features. The feature extraction unit is used to extract features from the input feature group using a dual-branch encoder; wherein, the first branch uses a Vision Transformer encoder to process two-dimensional raster data from satellite remote sensing, and the second branch uses a multilayer perceptron and an aggregation layer to process on-site profile observation point data to obtain the encoded feature tensor. The feature fusion unit is used to input the two-dimensional raster data and the encoded feature tensor into the multi-head spatiotemporal channel attention fusion module for dynamic calibration and fusion to obtain the latent variables of deep fusion. The feature decoding unit is used to input the deep fusion latent variables and the numerical model background field into the decoder based on the conditional variational autoencoder to generate high-resolution three-dimensional temperature field grid data and simultaneously output a three-dimensional uncertainty field; wherein, the decoding process of the decoder introduces a differential equation arbitrator to perform physical compliance checks and corrections on the initial decoded samples.

9. An electronic device, characterized in that, The electronic device includes a processor and a memory; The memory is used to store programs; The processor executes the program to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a program that is executed by a processor to implement the method as described in any one of claims 1 to 7.

Images