An ocean temperature and salinity field reconstruction evaluation assimilation prediction method based on satellite remote sensing inversion

Abstract

This invention provides a method for reconstructing, evaluating, assimilating, and forecasting ocean temperature and salinity fields based on satellite remote sensing inversion, belonging to the field of marine observation technology. This invention employs a multi-source satellite data time-series flow fusion algorithm based on transmission optimality theory to time-align sea surface parameters at adjacent time points. Then, an adaptive dynamic weighted isodense surface optimal interpolation algorithm outputs a global three-dimensional temperature and salinity reconstructed field and point-by-point uncertainties. Subsequently, an artificial intelligence model for ocean temperature and salinity profile reconstruction, with a denoising diffusion probability model as its backbone, generates a three-dimensional temperature and salinity profile set and profile uncertainty distribution. The number of denoising steps is dynamically adjusted through a comprehensive uncertainty scoring function. Finally, the optimal interpolation assimilation system, using a masked profile set and the reconstructed field input set, completes the assimilation and drives an ocean numerical prediction model. This solves the problem of difficulty in effectively controlling three-dimensional temperature and salinity field inversion errors caused by time asynchrony in multi-source satellite observation data.

Description

Technical Field

[0001] This invention belongs to the field of marine observation technology, and specifically relates to a method for reconstructing, assessing, assimilating, and predicting marine temperature and salinity fields based on satellite remote sensing inversion. Background Technology

[0002] Three-dimensional reconstruction and numerical forecasting of ocean temperature and salinity fields are core tasks in physical oceanography. Current techniques typically employ optimal interpolation using isodense surfaces to fuse and invert sea surface height anomalies, sea surface temperature fields, and sea surface salinity fields from multi-source satellite remote sensing observations into a three-dimensional temperature and salinity profile. This profile is then used to drive ocean numerical forecasting models via an assimilation system to output forecast results. These methods have achieved certain application results in deep-water areas of the global ocean and have been developed into several operational forecasting systems.

[0003] However, due to differences in the transit times of different satellite sensors, sea surface parameter products from various sources are not synchronized in the time dimension. In rapidly evolving scenarios such as typhoon transits and strong mesoscale eddies, the time difference between products can lead to significant misalignment of the frontal position. This time inconsistency error is amplified step by step in subsequent inversion, ultimately significantly degrading the reconstruction accuracy of the three-dimensional temperature and salinity field. Most existing multi-source fusion schemes rely on numerical models to provide flow fields for flow correction, or simply use simple time linear interpolation, neither of which can eliminate the frontal misalignment caused by time difference from a physical mechanism without relying on models.

[0004] In existing technologies, the lack of a fully data-driven fusion method that can eliminate the time synchronization error of multi-source satellite products leads to a technical problem in the three-dimensional temperature and salinity field inversion: the time synchronization of multi-source satellite observation data makes it difficult to effectively control the inversion error. Summary of the Invention

[0005] In view of this, the present invention provides a method for reconstructing, evaluating, assimilating, and predicting ocean temperature and salinity fields based on satellite remote sensing inversion, which can solve the technical problem in the prior art where the error of three-dimensional temperature and salinity field inversion is difficult to control effectively due to the time asynchrony of multi-source satellite observation data.

[0006] This invention is implemented as follows: This invention provides a method for reconstructing, assessing, assimilating, and predicting ocean temperature and salinity fields based on satellite remote sensing inversion, comprising the following steps:

[0007] A time-series shift fusion algorithm based on transmission optimality theory is applied to the sea surface parameters of two adjacent time periods to output a time-aligned multi-source fused satellite observation field.

[0008] The multi-source fusion satellite observation field is input into the adaptive dynamic weighted isodense surface optimal interpolation algorithm. The effective isodense surface layer number of each grid point is counted. The observation error variance of the sea surface height anomaly field is dynamically adjusted according to the relationship between the effective isodense surface layer number and the preset layer threshold. The covariance matrix is ​​then inverted after applying adaptive regularization, and the global three-dimensional temperature and salinity reconstruction field and point-by-point uncertainty estimate are output.

[0009] Using sea surface parameters from multi-source fusion satellite observations as conditional vectors, the global 3D temperature and salinity reconstruction field is input into the ocean temperature and salinity profile reconstruction model, driving the denoising diffusion probability model for iterative sampling, and outputting a 3D temperature and salinity profile set and profile uncertainty distribution. Based on the profile uncertainty distribution, point-by-point uncertainty estimation, and local horizontal gradient, the comprehensive uncertainty score is calculated using a comprehensive uncertainty scoring function. Based on the interval to which the comprehensive uncertainty score belongs, the denoising step parameters of the ocean temperature and salinity profile reconstruction model are adjusted, and the ocean temperature and salinity profile reconstruction model is re-executed with the adjusted denoising step parameters, outputting the adjusted 3D temperature and salinity profile set and the adjusted profile uncertainty distribution.

[0010] Keeping the original latitude and longitude coordinates of the Argo buoy observation points unchanged, the global three-dimensional temperature and salinity reconstruction field is mapped to each Argo buoy observation point using an inverse interpolation strategy, and the bias and root mean square error are calculated. For interpolation points where the local horizontal gradient exceeds the gradient discrimination threshold, the interpolation is automatically switched to piecewise cubic Hermitian interpolation, and anomaly removal is performed on all interpolation points, outputting the root mean square error after anomaly removal.

[0011] The root mean square error after anomaly removal is used as the observation error parameter of the ensemble optimal interpolation assimilation system; a quality mask is generated and assimilation is blocked for regions where the root mean square error exceeds the assimilation error tolerance threshold; the ensemble of three-dimensional temperature and salinity profiles adjusted by the quality mask and the global three-dimensional temperature and salinity reconstruction field are input into the ensemble optimal interpolation assimilation system to complete the three-dimensional temperature and salinity field assimilation and output the forecast initial field.

[0012] The ocean numerical prediction model is driven by the predicted initial field, and the numerical prediction results of the temperature and salinity field are output.

[0013] The sea surface parameters include the sea surface height anomaly field, the sea surface temperature field, and the sea surface salinity field.

[0014] Specifically, the multi-source satellite data time-series transport fusion algorithm based on the transmission optimality theory uses the sea surface height anomaly fields of two adjacent time periods as the source and target distributions, respectively. It uses the Sinkhorn algorithm to solve the regularized optimal transport mapping to obtain the equivalent transport velocity field. The equivalent transport velocity field is then used to perform forward and backward transport corrections on the sea surface temperature field and sea surface salinity field. After interpolation to a unified reference time, the time residuals after transport are used as weights for weighted least squares fusion.

[0015] The regularization coefficients in the Sinkhorn algorithm are determined through iterative experiments on historical multi-source satellite datasets with the goal of minimizing frontal position deviations.

[0016] In the adaptive dynamic weighted isodense surface optimal interpolation algorithm, when the number of effective isodense surface layers is lower than the preset layer threshold, the observation error variance of the sea surface height anomaly field is dynamically amplified according to the inverse square law, and the fixed Gaussian kernel is replaced by the local Rossby deformation radius as the physical constraint of the anisotropic correlation length. The algorithm is iterated twice to converge the dynamic weight.

[0017] The regularization coefficient of the adaptive regularization automatically increases as the condition number of the covariance matrix exceeds the condition number discrimination threshold. The condition number discrimination threshold is determined experimentally by statistically analyzing the distribution of the ill-conditioned matrix condition number on a global grid and taking the high quantile.

[0018] The denoising backbone network of the ocean temperature and salinity profile reconstruction model adopts an improved U-Net structure. The encoder consists of downsampling blocks, each containing a residual convolutional layer and a channel attention module. A vertical physical prior layer is integrated into each vertical layer feature, injecting water depth, Brent-Visala frequency, and potential density as learnable depth position codes into the corresponding vertical layer features.

[0019] The ocean temperature and salinity profile reconstruction model introduces a cross-time step jump attention mechanism in the time step dimension of the denoising iteration, thereby increasing the denoising steps. With steps ( The hidden features (for the jump step) are interacted through a multi-head self-attention module. The jump step is determined through iterative experiments on the validation set with the goal of achieving the optimal combined score of convergence speed and front accuracy.

[0020] The comprehensive uncertainty scoring function is composed of a weighted sum of the variance of the temperature profile, the variance of the salinity profile, and the local horizontal gradient, with the sum of all weight coefficients being 1. It is determined experimentally on the validation set with the goal of minimizing the profile reconstruction error. When the comprehensive uncertainty score is greater than or equal to the high uncertainty threshold, the denoising step number parameter is increased to a high-magnification value of the preset denoising step number. When the comprehensive uncertainty score is less than the low uncertainty threshold, the denoising step number parameter is decreased to a low-magnification value of the preset denoising step number. In other cases, the denoising step number parameter remains unchanged.

[0021] In the inverse interpolation strategy, interpolation points where the local horizontal gradient exceeds the gradient discrimination threshold are automatically switched to piecewise cubic Hermitian interpolation. The gradient discrimination threshold is determined through iterative experiments in the global Argo buoy dataset with the goal of minimizing the overall interpolation error.

[0022] The anomaly removal process employs a double-precision accumulator to perform statistical calculations, and the physical threshold and deviation threshold are determined through iterative statistical analysis of the temperature and salinity deviation distribution of each standard vertical layer in the historical Argo buoy dataset.

[0023] The quality mask generates a binary labeled field based on whether the root mean square error after anomaly removal exceeds the assimilation error tolerance threshold. The assimilation error tolerance threshold is determined through iterative experiments in historical forecast experiments, with forecast skill scores as the target.

[0024] The training dataset for the ocean temperature and salinity profile reconstruction model includes global Argo buoy temperature and salinity profile data from 1990 to the present. After quality control, the data is interpolated to a unified standard vertical layer using the sea surface height anomaly field, sea surface temperature field, and sea surface salinity field at the corresponding time and location as conditional vectors. The dataset is then divided into training set, validation set, and test set.

[0025] The training of the ocean temperature and salinity profile reconstruction model uses the weighted sum of mean squared error loss and profile vertical gradient constraint loss as the total loss function, employs the AdamW optimizer, and uses a cosine decay of the initial learning rate after a warm-up strategy. Training stops when the validation set loss does not decrease for several consecutive cycles.

[0026] Wherein, the preset layer threshold is 5 layers; the regularization coefficient of the regularized optimal transport mapping is determined in the range of 0.01 to 1.0; the base ratio of the dynamic amplification coefficient is 16; the preset denoising step number parameter is 50, the high ratio value is 1.5 times the preset denoising step number, and the low ratio value is 0.7 times the preset denoising step number; the high uncertainty threshold corresponds to a ratio of comprehensive uncertainty score value to reference standard value greater than or equal to 0.8, and the low uncertainty threshold corresponds to a ratio of less than 0.5; the jump step size is determined in the range of 1 to 5.

[0027] This invention introduces a time-series transport fusion algorithm based on transport optimization theory for multi-source satellite data. Using the sea surface height anomaly fields of two adjacent time periods as the source and target distributions, the Sinkhorn algorithm is used to solve the regularized optimal transport mapping, extract the equivalent transport velocity field, and then use this to perform forward and backward transport corrections on the sea surface temperature and salinity fields. All products are then aligned to the reference time before fusion, thereby solving the technical problem that the inversion error of the three-dimensional temperature and salinity field is difficult to effectively control due to the time asynchrony of multi-source satellite observation data.

[0028] The flow correction of this invention is entirely driven by satellite data itself, without relying on any external numerical model flow fields. In rapidly evolving ocean scenarios, it can effectively eliminate frontal misalignment caused by time differences between sensor products, while preserving the original observation accuracy of each sensor. This provides high-quality, time-consistent input for subsequent isodense surface optimal interpolation and profile reconstruction models. Since the temporal consistency of the input field is guaranteed from the source, time asynchrony errors no longer propagate downstream in the 3D temperature and salinity inversion and assimilation forecasting link, effectively suppressing error accumulation throughout the forecasting link. In summary, this invention solves the technical problem mentioned in the background art where the 3D temperature and salinity field inversion error is difficult to control effectively due to time asynchrony in multi-source satellite observation data. Attached Figure Description

[0029] Figure 1 This is a flowchart of the method of the present invention.

[0030] Figure 2 This is a schematic diagram of the 72-hour temperature and salinity field numerical forecast results.

[0031] Figure 3 This is a spatial distribution diagram of the comprehensive uncertainty score. Detailed Implementation

[0032] 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.

[0033] like Figure 1 The diagram shows a flowchart of a method for reconstructing, assessing, assimilating, and predicting ocean temperature and salinity fields based on satellite remote sensing inversion, provided by this invention. This method includes the following steps:

[0034] S01. Perform a multi-source satellite data time-series flow fusion algorithm based on transmission optimality theory on the sea surface parameters of two adjacent time periods, and output a time-aligned multi-source fused satellite observation field;

[0035] S02. Input the multi-source fusion satellite observation field into the adaptive dynamic weighted isodense surface optimal interpolation algorithm, count the number of effective isodense surface layers at each grid point, dynamically adjust the observation error variance of the sea surface height anomaly field according to the relationship between the number of effective isodense surface layers and the preset layer threshold, and apply adaptive regularization to the covariance matrix and then invert it to output the global three-dimensional temperature and salinity reconstruction field and point-by-point uncertainty estimate.

[0036] S03. Using the sea surface parameters in the multi-source fusion satellite observation field as a condition vector, the global three-dimensional temperature and salinity reconstruction field is input into the ocean temperature and salinity profile reconstruction model to drive the denoising diffusion probability model to iteratively sample and output a three-dimensional temperature and salinity profile set and profile uncertainty distribution; based on the profile uncertainty distribution, the point-by-point uncertainty estimation, and the local horizontal gradient, a comprehensive uncertainty score is calculated using a comprehensive uncertainty scoring function; based on the interval to which the comprehensive uncertainty score belongs, the denoising step parameters of the ocean temperature and salinity profile reconstruction model are adjusted, and the ocean temperature and salinity profile reconstruction model is re-executed with the adjusted denoising step parameters to output an adjusted three-dimensional temperature and salinity profile set and an adjusted profile uncertainty distribution;

[0037] S04. Keeping the original latitude and longitude coordinates of the Argo buoy observation points unchanged, the global three-dimensional temperature and salinity reconstruction field is mapped to each of the Argo buoy observation points using an inverse interpolation strategy, and the deviation and root mean square error are calculated. For interpolation points where the local horizontal gradient exceeds the gradient discrimination threshold, the interpolation is automatically switched to piecewise cubic Hermite interpolation, and anomaly removal is performed on all interpolation points. The root mean square error after anomaly removal is output.

[0038] S05. The root mean square error after anomaly removal is used as the observation error parameter of the ensemble optimal interpolation assimilation system; a quality mask is generated and assimilation is blocked for regions where the root mean square error exceeds the assimilation error tolerance threshold; the adjusted three-dimensional temperature and salinity profile set through the quality mask and the global three-dimensional temperature and salinity reconstruction field are input into the ensemble optimal interpolation assimilation system to complete the three-dimensional temperature and salinity field assimilation and output the forecast initial field.

[0039] S06. Drive the ocean numerical prediction model with the predicted initial field and output the numerical prediction results of the temperature and salinity field.

[0040] The sea surface parameters include the sea surface height anomaly field, the sea surface temperature field, and the sea surface salinity field.

[0041] The principle and implementation of the multi-source satellite data time-series transport fusion algorithm based on transport optimization theory are as follows: Using the sea surface height anomaly fields of two adjacent time periods as the source and target distributions respectively, the Sinkhorn algorithm is used to solve for the regularized optimal transport mapping, obtaining the equivalent transport velocity field. Forward and backward transport corrections are performed on the sea surface temperature and salinity fields using this equivalent transport velocity field, interpolating each product to a unified reference time. During fusion, the time residuals after transport are used as weights, and the weighted least squares method is used to fuse the products, outputting a time-aligned multi-source fused satellite observation field. This algorithm is entirely data-driven and does not rely on numerical model flow fields. In rapidly evolving scenarios such as typhoon passage, it effectively eliminates frontal misalignment caused by time differences in multi-source satellite products, preserves the observation accuracy of each sensor, and provides a high-quality, time-consistent input field for subsequent inversion, fundamentally reducing the transmission of time asynchrony errors to the three-dimensional temperature and salinity field.

[0042] The Sinkhorn algorithm is a numerical solution method for regularized optimal transmission. It applies an entropy regularization term to the transmission cost matrix between two discrete distributions, transforming the original linear programming problem into one that can be efficiently solved using alternating scaling iterations. Each iteration alternately normalizes rows and columns, and upon convergence, an approximately optimal transmission matrix is ​​obtained, thereby extracting the equivalent flow velocity field. The regularization coefficient is determined through iterative experiments on historical multi-source satellite datasets: using the frontal misalignment residual as the optimization objective, the regularization coefficient is searched at logarithmic intervals within the range of 0.01 to 1.0, and the coefficient that minimizes the frontal position deviation after weighted fusion is selected as the preset value.

[0043] The principle and implementation of the adaptive dynamic weighted isodense surface optimal interpolation algorithm are as follows. Based on the classic isodense surface optimal interpolation method, the effective isodense surface layers at the water depth of each grid point are first counted. The preset layer threshold is determined experimentally: grid points with 3, 5, 7, and 10 effective isodense surface layers are selected in typical sea areas around the world, and the root mean square error of the inversion results is examined one by one. Through multiple iterative experiments, the preset layer threshold is determined to be 5 layers. When the effective isodense surface layer number is less than 5 layers, the observation error variance of the sea surface height anomaly field is dynamically amplified according to the inverse square law. The benchmark magnification factor of 16 is determined through the following experiments: samples with 1 to 4 effective isodense surface layers are selected in the continental shelf area. Using Argo buoy observations as a reference, tests are conducted in steps of integer powers of 2 within the magnification range of 4 to 64. The magnification factor that minimizes the root mean square error is selected, and the median is obtained after multiple independent experiments, resulting in 16. Simultaneously, the local Rossby deformation radius is used as the physical constraint for the anisotropic correlation length, replacing the traditional fixed Gaussian kernel, and automatically shrinking the horizontal correlation scale of the background field error covariance. An adaptive regularization based on condition number estimation is applied before inverting the covariance matrix. The regularization coefficient automatically increases with the degree of condition number exceeding the limit. The condition number discrimination threshold is determined experimentally by statistically analyzing the condition number distribution of ill-conditioned matrices on a global grid and taking the 99th quantile. The entire process iterates twice to converge the dynamic weights, ultimately outputting the global 3D temperature and salinity reconstruction field and the pointwise uncertainty estimate. This algorithm solves the problem of numerical divergence caused by ill-conditioned covariance matrices in shallow sea areas, significantly improving the inversion stability of the continental shelf and nearshore regions while maintaining the accuracy of deep-sea inversion. Furthermore, the pointwise uncertainty estimate provides a physically meaningful error input for subsequent assimilation.

[0044] The effective number of isodense surface layers refers to the number of isodense surfaces at a given grid point whose density gradient within the water column depth range satisfies the stable stratification discrimination condition. It is used to measure the information constraint capability of the optimal isodense surface interpolation method at that point.

[0045] The local Rossby deformation radius refers to a characteristic horizontal scale determined by the stratification frequency and the Coriolis parameter in a local sea area, reflecting the typical spatial range of mesoscale motion. The calculation formula is as follows:

[0046] ;

[0047] in The local Rossby deformation radius ( ), Coriolis parameters ( ), For Brent-Vesala frequency ( ), For water depth ( ), , , Reference lengths ( Reference time () ), reference depth ( ), used to perform dimensionless processing on each item.

[0048] The specific structure of the ocean temperature and salinity profile reconstruction model is as follows: Using a denoising diffusion probability model framework as the backbone, the generation of the three-dimensional temperature and salinity profile is modeled as a conditional probability sampling process. The conditional vector is formed by concatenating the spatial features extracted from the sea surface height anomaly field, sea surface temperature field, and sea surface salinity field by independent two-dimensional convolutional encoders, and then injecting it into each vertical layer feature of the denoising backbone network. The denoising backbone network adopts an improved U-Net structure. The encoder consists of four downsampling blocks, each containing two layers of residual convolution and one layer of channel attention modules; the decoder is symmetrically set with four upsampling blocks, which are fused with the corresponding layers of the encoder through skip connections. A vertical physical prior layer is integrated into each vertical layer feature of the U-Net: water depth, Brent-Visala frequency, and potential density are encoded as learnable depth positions, linearly transformed, and added element-wise with the corresponding vertical layer features, enabling the network to perceive the local physical layering state at each depth layer. A cross-timestep skip attention mechanism is introduced in the time-step dimension of the denoising iteration: the denoising step... With steps ( To optimize the step size, hidden layer features (determined through iterative experiments within the range of 1-5 on the validation set, aiming for the optimal combined score of convergence speed and frontal accuracy) interact through a multi-head self-attention module. This allows different denoising steps to share high-frequency ocean structure information, accelerating convergence and improving the accuracy of mesoscale eddy fronts. The output layer consists of the mean, variance, mean, and variance of the temperature and salinity profiles across the entire vertical layer. The variances of the temperature and salinity profiles constitute the uncertainty distribution of the profiles, directly providing ensemble uncertainty for the dynamic assignment of the assimilation error matrix. The preset value for the denoising step number parameter is 50, determined through iterative experiments within the range of 20-200 on the historical Argo buoy profile dataset, aiming for the optimal combined score of sampling quality and computation time, with a step size of 10.

[0049] The steps for establishing the training dataset for the ocean temperature and salinity profile reconstruction model specifically include: collecting global Argo buoy temperature and salinity profile data from 1990 to the present, and retaining valid profiles after quality control; extracting sea surface height anomaly fields, sea surface temperature fields, and sea surface salinity fields from satellite products at the same time and location as the observation time and location corresponding to each valid profile as training samples for the condition vector; interpolating the valid profiles to a unified standard vertical layer; and dividing the training set, validation set, and test set in an 8:1:1 ratio.

[0050] The specific steps for training the ocean temperature and salinity profile reconstruction model include: using the weighted sum of the mean squared error loss and the profile vertical gradient constraint loss as the total loss function; the gradient constraint weights are determined experimentally on the validation set with the goal of minimizing the profile layering error; employing the AdamW optimizer; and using a warm-up strategy to adjust the initial learning rate from... linearly increasing to After cosine decay; training stops when the validation set loss does not decrease for 20 consecutive epochs; the weights with the lowest validation set loss are saved as the final model weights.

[0051] The technical effects of the ocean temperature, salinity, and thermal profile reconstruction model are as follows: The denoising diffusion probability model gradually recovers the probability distribution of the profile from Gaussian noise through multi-step iterative denoising. Compared to direct regression methods, its output naturally includes uncertainty information, enabling the assimilation system to obtain physically meaningful estimates of observation errors and avoiding the subjectivity of manually specifying error parameters. The vertical physical prior layer directly encodes physical constraints such as Brent-Vissara frequency and potential density into the network features, ensuring that the model remains subject to physical constraints even in data-sparse, deep layers, suppressing non-physical interpretations. The cross-time-step attention mechanism preserves high-frequency ocean structures such as fronts through feature sharing between different denoising steps, improving the ability to reconstruct mesoscale features.

[0052] The design of the comprehensive uncertainty scoring function is as follows: Comprehensive uncertainty score value. The calculation formula is expressed as follows: ;in The variance of the temperature profile ( ), The variance of the salinity profile ( ), For local horizontal gradient ( ), , , , To correspond to the standard reference value, , , These are weighting coefficients, and the sum of the three is 1. They are determined experimentally on the validation set with the objective of minimizing the profile reconstruction error. When When the noise reduction step count parameter is increased to 1.5 times the preset value; when When, keep the denoising step count parameter unchanged; when When this happens, the number of denoising steps is reduced to 0.7 times the preset value.

[0053] The reverse interpolation strategy refers to keeping the original latitude and longitude coordinates of the Argo buoy observation points unchanged, interpolating the global three-dimensional temperature and salinity reconstruction field to the locations of each Argo buoy observation point, rather than averaging the Argo buoy observation data onto the grid, thereby eliminating the smoothing error and positioning deviation caused by gridding, and preserving the small- and medium-scale information of Argo buoy observations.

[0054] Piecewise cubic Hermite interpolation is an interpolation method that uses cubic polynomial interpolation between adjacent data nodes while constraining the first derivative at each node to keep the interpolation result monotonic. It suppresses spurious extrema generated by bilinear interpolation in strong gradient regions.

[0055] The gradient discrimination threshold refers to the threshold used to determine whether the local horizontal gradient exceeds the standard for a strong frontal gradient region. It is determined through the following experiment: Samples with known frontal locations are selected from the global Argo buoy dataset. The optimization objective is the cross-comparison of bilinear interpolation residuals and piecewise cubic Hermite interpolation residuals, within a temperature gradient of 0.05–0.5. Salinity gradient 0.01–0.1 Within the range, iterative experiments are conducted with uniform step sizes, and the gradient value that minimizes the overall interpolation error is taken as the gradient discrimination threshold.

[0056] The outlier removal refers to the process of marking and excluding interpolation points whose deviations exceed the physical or deviation thresholds. A double-precision accumulator is used to perform statistical calculations to prevent result distortion due to precision overflow. The physical and deviation thresholds are determined through hierarchical statistical analysis of historical Argo buoy datasets: the mean and standard deviation of temperature and salinity deviation distributions for each standard vertical layer are calculated, and the initial threshold is set at the mean plus or minus three times the standard deviation. After multiple iterations to remove outliers, the thresholds are recalculated until convergence, yielding the physical and deviation thresholds for each layer.

[0057] The ensemble optimal interpolation assimilation system refers to a numerical assimilation system that estimates the background error covariance matrix using ensemble statistical methods and assimilates the observation information into the background field using the optimal interpolation formula. Its observation error parameter is dynamically assigned by the root mean square error after anomaly removal.

[0058] The quality mask refers to a binary label field generated based on the evaluation results of S04 for regions where the root mean square error after anomaly removal exceeds the assimilation error tolerance threshold. Regions with a value of 0 do not participate in assimilation. The assimilation error tolerance threshold is determined through iterative experiments in historical forecast experiments, with forecast skill scores as the target and the threshold being gradually tightened.

[0059] The denoising diffusion probability model refers to a probabilistic sampling process that starts with Gaussian noise and gradually removes noise to restore the data distribution by defining a forward-noising Markov chain and a reverse denoising neural network. The training objective is to minimize the difference between the predicted noise and the actual noise in the reverse denoising step.

[0060] The Brunt-Väisälä frequency is a physical quantity that characterizes the frequency of vertical buoyancy oscillations in the ocean, reflecting the vertical stability of the water column. The larger the value, the more stable the stratification. It is calculated from the vertical gradient of the potential density, gravitational acceleration, and reference density.

[0061] The potential density refers to the density of a seawater sample after it has been adiabatically compressed or expanded to a reference pressure surface. It is used to eliminate the influence of pressure on density and to reflect the true stratification state of seawater.

[0062] Optionally, the present invention also provides a computer-based method for forming an ocean temperature and salinity field reconstruction assessment and assimilation forecasting system based on satellite remote sensing inversion. The computer is equipped with a readable storage medium, which stores program instructions. When the program instructions are run on the computer, they can execute the above-described method.

[0063] The specific implementation of step S01 is as follows: First, the sea surface height anomaly field, sea surface temperature field, and sea surface salinity field for two adjacent time periods are read from the satellite product database. Using the sea surface height anomaly field from the previous time period as the source probability distribution and the sea surface height anomaly field from the next time period as the target probability distribution, a cost matrix is ​​constructed between the two distributions. The elements of the cost matrix are the squared geographical distances between corresponding grid points. An entropy regularization term is applied to the cost matrix, transforming the optimal transport problem into a form that can be solved iteratively using the Sinkhorn algorithm. The rows and columns of the cost matrix are alternately normalized until the transport matrix converges. The regularization coefficient is determined by searching logarithmically in the range of 0.01 to 1.0 on historical multi-source satellite datasets with the objective of minimizing frontal position deviation. An equivalent current velocity field is extracted from the converged optimal transport matrix. This velocity field is used to perform forward and backward current corrections on the sea surface temperature field and sea surface salinity field, respectively, and each product is interpolated to a unified reference time. Finally, using the time residuals of each product after migration as weights, the multi-source products are fused using the weighted least squares method to output a time-aligned multi-source fused satellite observation field. The purpose of this step is to eliminate frontal misalignment caused by different transit times of the multi-source satellite products, providing a time-consistent input field for subsequent inversion.

[0064] The specific implementation of step S02 is as follows: Using the sea surface height anomaly field, sea surface temperature field, and sea surface salinity field from the multi-source fusion satellite observation field as input, and based on the classical isodense surface optimal interpolation method, the number of isodense surfaces satisfying the stable stratification discrimination condition within the water column depth range is counted at each grid point; this is the effective isodense surface layer number. When the effective isodense surface layer number is less than a preset layer threshold of 5 layers, the observation error variance of the sea surface height anomaly field is dynamically amplified according to the inverse square law, with a base multiplier of 16, thereby suppressing matrix ill-conditioning caused by insufficient information in shallow sea areas. The local Rossby deformation radius is used as the physical constraint for the anisotropic correlation length. The calculation formula involves the vertical integral of the Brent-Vesala frequency and the Coriolis parameter, used to replace the traditional fixed Gaussian kernel, allowing the correlation length to adaptively change with the local stratification state. Before inverting the background error covariance matrix, adaptive regularization is applied based on condition number estimation. The regularization coefficient automatically increases as the condition number exceeds the discrimination threshold, with the condition number discrimination threshold being the 99th quantile of the global grid condition number distribution. The above dynamic weight adjustment process is iterated twice to converge the weights, and finally outputs the global three-dimensional temperature and salinity reconstruction field and point-by-point uncertainty estimate.

[0065] The specific implementation of step S03 is as follows: The sea surface height anomaly field, sea surface temperature field, and sea surface salinity field are extracted for spatial features by independent two-dimensional convolutional encoders and then concatenated to form a conditional vector, which is then injected into each vertical layer feature of the denoising backbone network. The denoising backbone network adopts an improved U-Net structure. The encoder contains four downsampling blocks, each containing two layers of residual convolution and one channel attention module. The decoder is symmetrically set and fused with the corresponding layer of the encoder through skip connections. A vertical physical prior layer is incorporated into each vertical layer feature, encoding water depth, Brent-Visala frequency, and potential density as learnable depth position codes, which are then linearly transformed and added element-wise to the corresponding vertical layer feature. A cross-time step skip attention mechanism is introduced in the denoising iteration to... With steps The hidden features interact through a multi-head self-attention module, with a jump step size. The range of 1–5 was determined by validation set experiments. The model outputs the mean temperature profile, variance of the temperature profile, and mean and variance of the salinity profile across the entire vertical layer. The variances constitute the profile uncertainty distribution. A comprehensive uncertainty score is then calculated. From temperature profile variance Salinity profile variance With local horizontal gradient The weighted sum calculation, weight coefficients , , The sum is 1, as determined by the validation set experiments. According to... Compared with reference standard value The ratio range is determined by the following conditions: when the ratio is greater than or equal to 0.8, the number of denoising steps is increased to 1.5 times the preset value of 50; when the ratio is less than 0.5, the number of denoising steps is reduced to 0.7 times the preset value; otherwise, the ratio remains unchanged. The model sampling is re-driven with the adjusted number of denoising steps, and the adjusted three-dimensional temperature-salinity profile set and profile uncertainty distribution are output.

[0066] The specific implementation of step S04 is as follows: Keeping the original latitude and longitude coordinates of all Argo buoy observation points unchanged, the global three-dimensional temperature-salinity reconstruction field is mapped to the locations of each observation point through inverse interpolation. That is, interpolation is performed on the reconstruction field grid using the observation point coordinates, rather than averaging the observation data across the grid. For interpolation points where the local horizontal gradient exceeds the gradient discrimination threshold, piecewise cubic Hermitian interpolation is automatically switched to suppress spurious extrema generated by bilinear interpolation in strong gradient regions. The gradient discrimination threshold is determined through iterative experiments in the global Argo buoy dataset with the goal of minimizing the overall interpolation error. The reference range for the temperature gradient is 0.05–0.5℃ / km, and the reference range for the salinity gradient is 0.01–0.1℃ / km. Anomaly removal is performed on all interpolation points. The physical threshold and bias threshold are determined iteratively by using the mean plus or minus three times the standard deviation as initial values ​​for the standard vertical layer temperature-salinity deviation distribution of the historical Argo buoy dataset. A double-precision accumulator is used for statistical calculations to prevent precision overflow. The final output is the root mean square error after anomaly removal, which is used for subsequent assimilation error assignment.

[0067] The specific implementation of step S05 is as follows: The root mean square error (RMSE) after anomaly removal is directly and dynamically assigned to the observation error parameters of the ensemble optimal interpolation assimilation system, replacing the traditional fixed error parameters. For regions where the RMSE exceeds the assimilation error tolerance threshold, a mass mask with a value of 0 is generated to block the assimilation process in that region. The assimilation error tolerance threshold is determined through iterative experiments in historical forecast experiments, gradually tightening the threshold with forecast skill scores as the target. For regions that pass through the mass mask, the adjusted 3D temperature and salinity profile set and the global 3D temperature and salinity reconstruction field are jointly input into the ensemble optimal interpolation assimilation system. The background error covariance matrix is ​​estimated using ensemble statistical methods, and the optimal interpolation formula assimilates the observation information into the background field, completing the 3D temperature and salinity field assimilation and outputting the initial forecast field.

[0068] The specific implementation of step S06 is as follows: using the forecast initial field output in S05 as the driving field, inputting it into the ocean numerical prediction model, and outputting the numerical prediction results of the temperature and salinity field after model integration. The input of this step is the assimilated forecast initial field, and the output is the global or regional temperature and salinity field forecast values ​​for future time periods, providing quantitative results for marine environmental forecasting.

[0069] It should be noted that the key technologies of this invention include: a time-series transport fusion technology based on transport optimality theory solves the optimal transport mapping between two sea surface height anomaly fields using the Sinkhorn algorithm. The extracted equivalent transport velocity field most faithfully describes the temporal evolution of the sea surface height anomaly field in terms of probability distribution. After correcting the sea surface temperature and salinity fields in this way, the time difference misalignment of the front position is eliminated at the physical mechanism level, rather than relying on the approximate substitution of the model flow field. Therefore, the correction accuracy is not affected by model errors. The adaptive dynamic weighted isodense surface optimal interpolation technology dynamically adjusts the observation error by sensing the number of effective isodense surface layers. The differential weighting and the constraint of the correlation length by the Rossby deformation radius effectively control the condition number of the covariance matrix in information-poor regions such as shallow seas, avoiding numerical divergence. The profile reconstruction technology based on the denoising diffusion probability model denoises the natural output probability distribution through multi-step iteration, combined with the physical constraint of the vertical physical prior layer on the sparse region of deep data, so that the profile uncertainty estimation has physical meaning. The three technologies work together to form a closed loop in the same link, with time-consistent high-quality input, stable 3D reconstruction, reliable uncertainty quantification and dynamic error assignment. The overall forecast accuracy is better than the sum of the effects of individual improvements to each module.

[0070] It should be noted that this invention also solves the following technical problem: In shallow sea and nearshore areas, due to limited water depth and insufficient number of isodense surfaces, traditional optimal interpolation methods are prone to generating ill-conditioned matrices when constructing the covariance matrix, leading to numerical divergence during the inversion process. This results in non-physical oscillations or divergence in the 3D temperature and salinity inversion results in shallow sea areas, making it impossible to provide reliable input for the assimilation system. This invention introduces a dynamic error variance amplification mechanism based on the number of effective isodense surface layers into the adaptive dynamic weighted isodense surface optimal interpolation algorithm. When the number of effective isodense surface layers is lower than a preset layer threshold, the observation error variance is amplified according to the inverse square law, reducing the ill-conditioned nature of the covariance matrix. At the same time, a local Rossby deformation radius is used instead of the fixed Gaussian kernel constraint correlation length, and adaptive regularization based on condition number estimation is applied before matrix inversion. The regularization strength increases adaptively with the degree of condition number exceeding the limit, thereby ensuring the numerical stability of matrix inversion in shallow sea areas. This prevents non-physical oscillations in the temperature and salinity inversion results in nearshore and shelf areas, significantly improving the inversion stability in shallow sea areas while maintaining the accuracy of deep-sea inversion.

[0071] Specifically, the principle of this invention is as follows: The reason this invention can solve the aforementioned technical problems is that the optimal transport theory provides a mathematical framework for measuring the minimum cost transformation between two distributions in the sense of probability distribution. When the sea surface height anomaly fields of adjacent time periods are taken as the source and target distributions respectively, the optimal transport mapping obtained is physically equivalent to the equivalent flow field driving the sea surface height anomaly field from one time period to the next. The Sinkhorn algorithm, by applying entropy regularization to the cost matrix, transforms the original linear programming into a form that can be solved by alternating iterative normalization, making the optimal transport calculation in large-scale ocean scenarios engineering feasible. After performing transport correction on the sea surface temperature and salinity fields using this equivalent transport velocity field, each sensor product is mapped to a unified reference time. The time difference misalignment of the front position is eliminated at the physical mechanism level, rather than relying on the approximate substitution of the model flow field. Therefore, the correction accuracy is not affected by model errors, ensuring the temporal consistency of the multi-source fusion satellite observation field from the source, providing reliable input for subsequent inversion and assimilation.

[0072] 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.

[0073] The specific implementation method of step S01 is as follows.

[0074] A multi-source satellite data time-series flow fusion algorithm based on transmission optimality theory is applied to sea surface parameters at two adjacent time points. The sea surface height anomalies at two adjacent time points are used as the source distributions. With target distribution The equivalent flow velocity field is solved using entropy-regularized optimal transport. The objective function for regularized optimal transport is expressed as follows:

[0075] ;

[0076] In the formula, To regularize the optimal transmission cost ( ), Reference value for transmission cost ( ), For grid points To grid point Transmission cost ( ), For elements of the transfer matrix (dimensionless). This is the reference value for the transmission matrix (dimensionless). is the entropy regularization coefficient (dimensionless). A set of transfer matrices that satisfy the boundary conditions. The frontal position deviation was minimized by searching logarithmically at intervals between 0.01 and 1.0 on historical multi-source satellite datasets. The Sinkhorn algorithm was used to solve the problem through alternating iterative row and column normalization. The row normalization and column normalization of the next iteration are described as follows:

[0077] ;

[0078] ;

[0079] In the formula, For the first The row scaling vector of the nth iteration 1 element (dimensionless) For the first The column scaling vector of the nth iteration 1 element (dimensionless) For the source distribution The mass of each grid point (dimensionless). For the target distribution The mass of each grid point (dimensionless). The elements of the Gibbs kernel matrix (dimensionless) are defined as follows: After convergence, the elements of the transfer matrix are: Based on this, the equivalent flow velocity field is extracted. , The The components are obtained by weighting and removing bits from the transmission matrix using the time difference. sea ​​surface temperature field With sea surface salinity field Perform forward and backward current shift corrections separately, and interpolate each product to a unified reference time. During fusion, the time residual after the current transfer is used. Using weights as values, the products are fused using weighted least squares method. The fused sea surface temperature field is described as follows:

[0080] ;

[0081] In the formula, To integrate the sea surface temperature field ( ), Sea surface temperature reference value ( ), For the first The time residual weight (dimensionless) of each product is defined as follows: , For the first Time residuals after flow correction for each product ( ), For the first Sea surface temperature field after product migration correction ( Sea surface salinity field By merging in the same way, Replace with , To integrate the sea surface salinity field ( ), For the first Sea surface salinity field after product migration correction ( ), outputting a time-aligned multi-source fused satellite observation field.

[0082] The specific implementation method of step S02 is as follows.

[0083] The multi-source fusion satellite observation field is input into the adaptive dynamic weighted isodense surface optimal interpolation algorithm. The number of effective isodense surface layers at each grid point is counted. Preset layer threshold Experiments confirmed that it has 5 layers. At that time, the variance of the sea surface height anomaly field observation error is dynamically amplified according to the inverse square law, and the dynamic amplification factor is... The statement is as follows:

[0084] ;

[0085] In the formula, The dynamic amplification factor (dimensionless) represents the variance of the observation error. This is a reference value for the magnification factor (dimensionless). The base multiplier (dimensionless) was determined to be 16 based on the median obtained from experiments. For the preset layer threshold, This represents the number of effective isodense surface layers at the current grid point. The value is the local Rossby deformation radius. As a physical constraint on anisotropic correlation length, The calculation formula is expressed as follows:

[0086] ;

[0087] In the formula, The radius of local Rossby deformation ( ), Reference length ( ), Coriolis parameters ( ), For reference time ( ), For depth Brent-Vesara frequency ( ), For water depth ( ), For reference depth ( ), Vertical depth coordinates ( Before inverting the covariance matrix, adaptive regularization is applied. The regularized background error covariance matrix is ​​expressed as follows:

[0088] ;

[0089] In the formula, The background error covariance matrix after regularization ( ), The background error covariance reference value ( ), The original background error covariance matrix ( ), For adaptive regularization coefficients (dimensionless). It is an identity matrix (dimensionless). Depending on the condition number Exceeding the discrimination threshold The degree automatically increases, as described below:

[0090] ;

[0091] In the formula, The regularization baseline coefficient (dimensionless) has an empirical value of [value missing]. , is the condition number (dimensionless) of the covariance matrix. The condition number threshold (dimensionless) is set at the 99th percentile of the global grid condition number distribution. The entire process iterates twice to converge the dynamic weights, outputting the global 3D temperature-salinity reconstruction field and point-by-point uncertainty estimates. , The local uncertainty of temperature-salinity inversion at each grid point ( or ).

[0092] The specific implementation method of step S03 is as follows.

[0093] Sea surface parameters are used as conditional vectors and input into the ocean temperature, salinity, and thermal profile reconstruction model to drive iterative sampling of the denoising diffusion probability model. The training objective of the denoising diffusion probability model is stated as follows:

[0094] ;

[0095] In the formula, The training loss (dimensionless) is used for the denoising diffusion probability model. This is a reference value for loss (dimensionless). The real Gaussian noise (dimensionless) introduced into the forward noise addition process. The noise (dimensionless) predicted by the denoising neural network. For the first The latent variables (dimensionless) of the denoising step. This is the denoising step number (dimensionless). It is a conditional vector (dimensionless). For expectation operator, This represents the set of learnable parameters (dimensionless) for the denoising neural network. The overall uncertainty score is... The calculation formula is expressed as follows:

[0096] ;

[0097] In the formula, This is the overall uncertainty score (dimensionless). This is the reference standard value for the overall uncertainty (dimensionless). The variance of the temperature profile ( ), Reference value for the variance of the temperature profile ( ), The variance of the salinity profile ( ), Reference value for the variance of the salinity profile ( ), For local horizontal gradient ( ), For local horizontal gradient reference values ​​( ), , , The weighting coefficients are denoted by , and the sum of the three is 1. These are determined experimentally on the validation set with the objective of minimizing the profile reconstruction error. When... When the noise reduction step count parameter is increased to 1.5 times the preset value; when When, keep the number of denoising steps constant; when At this time, the denoising step count parameter is reduced to 0.7 times the preset value, which is 50 by default. The model is re-executed with the adjusted denoising step count parameter, and the adjusted three-dimensional temperature-salinity profile set and profile uncertainty distribution are output.

[0098] The specific implementation method of step S04 is as follows.

[0099] While maintaining the original latitude and longitude coordinates of the Argo buoy observation points, an inverse interpolation strategy is employed to map the global 3D temperature and salinity reconstruction field to each Argo buoy observation point. For interpolation points where the local horizontal gradient exceeds the gradient discrimination threshold, piecewise cubic Hermitian interpolation is automatically switched. The gradient discrimination threshold is determined by minimizing the overall interpolation error within the global Argo buoy dataset, within a temperature gradient range of 0.05 to 0.5. Salinity gradient 0.01 to 0.1 The results were determined through iterative experiments with uniform step sizes within the specified range. Anomaly removal was performed on all interpolation points. The physical threshold and deviation threshold were initially set based on the mean plus or minus three standard deviations of the deviation distributions of each standard vertical layer, and were determined after iterative convergence. The root mean square error after anomaly removal was then calculated. The statement is as follows:

[0100] ;

[0101] In the formula, The root mean square error after outlier removal (RMSE) or ), The root mean square error reference value ( or ), This represents the number of interpolation points remaining after anomaly removal (dimensionless). For the first The deviation of each interpolation point ( or ), defined as the difference between the interpolated value and the Argo buoy observation.

[0102] The specific implementation method of step S05 is as follows.

[0103] by As the observation error parameter of the ensemble optimal interpolation assimilation system, Regions exceeding the assimilation error tolerance threshold are generated using a quality mask, while regions with a value of 0 are excluded from assimilation. The assimilation error tolerance threshold is determined iteratively by gradually tightening it through historical forecast experiments. The adjusted 3D temperature-salinity profile set, processed through the quality mask, is then combined with the optimal interpolation assimilation system of the global 3D temperature-salinity reconstruction field input set to complete the 3D temperature-salinity field assimilation, outputting the initial forecast field.

[0104] The specific implementation method of step S06 is as follows.

[0105] The ocean numerical prediction model is driven by the initial forecast field, and the numerical prediction results of temperature and salinity field are output, thus completing the entire ocean temperature and salinity field reconstruction, assessment, assimilation, and prediction process based on satellite remote sensing inversion.

[0106] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: To illustrate the effect of the invention, the technicians set up a test environment, using a strong mesoscale eddy active sea area in the Northwest Pacific as the test area, and used the method of the invention to reconstruct, evaluate, assimilate and predict a mesoscale eddy event in three dimensions, verifying the technical effect of the whole process.

[0107] During the selected time period in the test area, the frontal gradient near the vortex center was significant, and the transit times of multiple satellites differed by approximately 6 to 9 hours within 24 hours, representing a typical scenario where time asynchrony errors were prominent. In step S01, technicians read the sea surface height anomaly field, sea surface temperature field, and sea surface salinity field at two adjacent time intervals, respectively. A cost matrix was constructed using the sea surface height anomaly fields at the two time intervals, with a regularization coefficient of 0.08. After iterative convergence using the Sinkhorn algorithm, the equivalent current velocity field was extracted. The equivalent current velocity vector near the vortex center highly matched the actual geostrophic flow direction. After correcting the current flow in the sea surface temperature and salinity fields, the time difference misalignment of the frontal position was eliminated, and the continuity of the fused observation field near the front was significantly improved.

[0108] In step S02, adaptive dynamic weighted isodense surface optimal interpolation is performed on the fused observation field. The effective isodense surface layer number in the shelf edge region surrounding the vortex center is 3 layers, lower than the preset threshold of 5 layers. The observation error variance is amplified according to the inverse square law, with a base factor of 16. The Rossby deformation radius is approximately 80 km in the vortex core region and approximately 25 km in the outer shelf region, with the relevant length shrinking accordingly. This causes the condition number of the shelf region covariance matrix to decrease from the value before regularization. The magnitude dropped to Below the order of magnitude, the matrix is ​​stable upon inversion. Table 1 shows the point-by-point uncertainty estimates for each layer of the output global 3D temperature-salinity reconstruction field, illustrating the average uncertainty level at typical depths.

[0109] Table 1. Mean Uncertainties of Vertical Temperature-Salt Reconstruction for Each Standard

[0110]

[0111] In step S03, sea surface parameters from the multi-source fusion satellite observation field are used as condition vectors, and the global three-dimensional temperature and salinity reconstruction field is input into the ocean temperature and salinity profile reconstruction model. The preset denoising step number is 50, and the jump step size is... Take 3. The model outputs the mean and variance of the temperature-salinity profiles across the entire vertical layer, and the comprehensive uncertainty score. The distribution in the vortex core region and surrounding area is as follows: Figure 3 As shown, the vortex core region The value was 0.91, exceeding the high uncertainty threshold of 0.8. The denoising step count was adjusted to 75 steps. After resampling, the vertical gradient of the profile set near the thermocline was clearer, and the frontal structure was better restored; the peripheral low-dynamic region... The uncertainty value was 0.38, which is lower than the low uncertainty threshold of 0.5. The number of denoising steps was reduced to 35, and the output quality did not decrease significantly while improving computational efficiency. The changes in the profile uncertainty distribution before and after adjustment are shown in Table 2.

[0112] Table 2 Comparison of profile uncertainty before and after denoising step count adjustment in typical areas

[0113]

[0114] In step S04, keeping the original latitude and longitude of all Argo buoy observation points unchanged, the global three-dimensional temperature and salinity reconstruction field is inversely interpolated to each observation point. The local horizontal temperature gradient at the edge of the vortex core region reaches 0.31℃ / km, exceeding the gradient discrimination threshold of 0.15℃ / km. In this region, the interpolation method automatically switches to piecewise cubic Hermitian interpolation, effectively suppressing spurious extrema near the thermocline. After anomaly removal, the root mean square error of temperature for all Argo buoy observation points within the test area is 0.41℃, and the root mean square error of salinity is 0.062 psu.

[0115] In step S05, the observation error parameters of the optimal interpolation assimilation system based on the aforementioned root mean square error dynamic assignment set are used to replace the traditional fixed error parameters. The root mean square errors near the vortex core region do not exceed the assimilation error tolerance threshold, the mass mask passes through completely, and the adjusted three-dimensional temperature and salinity profile set and the global three-dimensional temperature and salinity reconstruction field both participate in assimilation, outputting the initial forecast field. In step S06, the initial forecast field drives the ocean numerical prediction model, and the 72-hour temperature and salinity field numerical prediction results are output after model integration, such as... Figure 2 As shown, the depth of the thermocline and the intensity of the vortex center are reasonably characterized in the forecast results.

[0116] This invention brings the following technological advancements compared to traditional methods: Traditional methods only use simple time linear interpolation in the multi-source fusion stage, which cannot eliminate frontal time lag misalignment from a physical mechanism perspective. In contrast, this invention extracts the equivalent flow velocity field through transmission optimal theory and drives flow correction with optimal mapping in the sense of probability distribution, ensuring the temporal consistency of frontal positions from the source. Traditional isodense surface optimal interpolation exhibits numerical divergence in shallow sea areas due to matrix ill-conditioning. This invention suppresses ill-conditioning by jointly suppressing it through dynamic error variance amplification and adaptive regularization, significantly improving the inversion stability in shallow sea areas. Traditional profile reconstruction methods cannot automatically output uncertainty. This invention utilizes the natural probability sampling mechanism of the denoising diffusion probability model to output the profile uncertainty estimate and reconstruction result simultaneously, providing a physically clear basis for the error assignment of the assimilation system and avoiding subjective errors caused by manually specifying error parameters.

[0117] It should be noted that the variables involved in this invention are explained in detail in Tables 3 and 4.

[0118] Table 3. Variable Explanation Table (Part 1)

[0119]

[0120] Table 4. Variable Explanation Table (Part Two)

[0121]

[0122] 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 reconstructing, assessing, assimilating, and predicting ocean temperature and salinity fields based on satellite remote sensing inversion, characterized in that, Includes the following steps: A time-series shift fusion algorithm based on transmission optimality theory is applied to the sea surface parameters of two adjacent time periods to output a time-aligned multi-source fused satellite observation field. The multi-source fusion satellite observation field is input into the adaptive dynamic weighted isodense surface optimal interpolation algorithm. The effective isodense surface layer number of each grid point is counted. The observation error variance of the sea surface height anomaly field is dynamically adjusted according to the relationship between the effective isodense surface layer number and the preset layer threshold. The covariance matrix is ​​then inverted after applying adaptive regularization, and the global three-dimensional temperature and salinity reconstruction field and point-by-point uncertainty estimate are output. Using sea surface parameters from multi-source fusion satellite observations as conditional vectors, the global 3D temperature and salinity reconstruction field is input into the ocean temperature and salinity profile reconstruction model, driving the denoising diffusion probability model for iterative sampling, and outputting a 3D temperature and salinity profile set and profile uncertainty distribution. Based on the profile uncertainty distribution, point-by-point uncertainty estimation, and local horizontal gradient, the comprehensive uncertainty score is calculated using a comprehensive uncertainty scoring function. Based on the interval to which the comprehensive uncertainty score belongs, the denoising step parameters of the ocean temperature and salinity profile reconstruction model are adjusted, and the ocean temperature and salinity profile reconstruction model is re-executed with the adjusted denoising step parameters, outputting the adjusted 3D temperature and salinity profile set and the adjusted profile uncertainty distribution. Keeping the original latitude and longitude coordinates of the Argo buoy observation points unchanged, the global three-dimensional temperature and salinity reconstruction field is mapped to each Argo buoy observation point using an inverse interpolation strategy, and the bias and root mean square error are calculated. For interpolation points where the local horizontal gradient exceeds the gradient discrimination threshold, the interpolation is automatically switched to piecewise cubic Hermitian interpolation, and anomaly removal is performed on all interpolation points, outputting the root mean square error after anomaly removal. The root mean square error after anomaly removal is used as the observation error parameter of the ensemble optimal interpolation assimilation system; a quality mask is generated and assimilation is blocked for regions where the root mean square error exceeds the assimilation error tolerance threshold; the ensemble of three-dimensional temperature and salinity profiles adjusted by the quality mask and the global three-dimensional temperature and salinity reconstruction field are input into the ensemble optimal interpolation assimilation system to complete the three-dimensional temperature and salinity field assimilation and output the forecast initial field. The ocean numerical prediction model is driven by the predicted initial field, and the numerical prediction results of the temperature and salinity field are output.

2. The method for ocean temperature and salinity field reconstruction assessment assimilation prediction based on satellite remote sensing inversion according to claim 1, characterized in that, The sea surface parameters include the sea surface height anomaly field, the sea surface temperature field, and the sea surface salinity field.

3. The method for ocean temperature and salinity field reconstruction assessment assimilation prediction based on satellite remote sensing inversion according to claim 2, characterized in that, The multi-source satellite data time-series transport fusion algorithm based on transport optimization theory specifically uses the sea surface height anomaly fields of two adjacent time periods as the source and target distributions, respectively. It uses the Sinkhorn algorithm to solve the regularized optimal transport mapping to obtain the equivalent transport velocity field. The equivalent transport velocity field is then used to perform forward and backward transport corrections on the sea surface temperature field and sea surface salinity field. After interpolation to a unified reference time, the time residuals after transport are used as weights for weighted least squares fusion.

4. The method for ocean temperature and salinity field reconstruction assessment assimilation prediction based on satellite remote sensing inversion according to claim 3, characterized in that, The regularization coefficients in the Sinkhorn algorithm were determined through iterative experiments on historical multi-source satellite datasets with the goal of minimizing frontal position deviations.

5. The method for ocean temperature and salinity field reconstruction assessment assimilation prediction based on satellite remote sensing inversion according to claim 4, characterized in that, In the adaptive dynamic weighted isodense surface optimal interpolation algorithm, when the number of effective isodense surface layers is lower than the preset layer threshold, the observation error variance of the sea surface height anomaly field is dynamically amplified according to the inverse square law, and the fixed Gaussian kernel is replaced by the local Rossby deformation radius as the physical constraint of the anisotropic correlation length. The algorithm is iterated twice to converge the dynamic weight.

6. The method for ocean temperature and salinity field reconstruction assessment assimilation prediction based on satellite remote sensing inversion according to claim 5, characterized in that, The regularization coefficient of the adaptive regularization automatically increases as the condition number of the covariance matrix exceeds the condition number discrimination threshold. The condition number discrimination threshold is determined experimentally by statistically analyzing the distribution of the ill-conditioned matrix condition number on a global grid and taking the high quantile.

7. The method for ocean temperature and salinity field reconstruction assessment assimilation prediction based on satellite remote sensing inversion according to claim 6, characterized in that, The denoising backbone network of the ocean temperature and salinity profile reconstruction model adopts an improved U-Net structure. The encoder consists of downsampling blocks, each containing a residual convolutional layer and a channel attention module. A vertical physical prior layer is incorporated into each vertical layer feature, and water depth, Brent-Vesala frequency, and potential density are injected as learnable depth location codes into the corresponding vertical layer features.

8. The method for ocean temperature and salinity field reconstruction assessment assimilation prediction based on satellite remote sensing inversion according to claim 7, characterized in that, The ocean temperature and salinity profile reconstruction model introduces a cross-time-step jump attention mechanism in the time-step dimension of the denoising iteration, thus enabling the denoising steps to... With steps ( The hidden features (for the jump step) are interacted through a multi-head self-attention module. The jump step is determined through iterative experiments on the validation set with the goal of achieving the optimal combined score of convergence speed and front accuracy.

9. The method for ocean temperature and salinity field reconstruction assessment assimilation prediction based on satellite remote sensing inversion according to claim 8, characterized in that, The comprehensive uncertainty scoring function is composed of a weighted sum of the variance of the temperature profile, the variance of the salinity profile, and the local horizontal gradient, with the sum of all weight coefficients being 1. It is determined through experiments on the validation set with the goal of minimizing the profile reconstruction error. When the comprehensive uncertainty score is greater than or equal to the high uncertainty threshold, the denoising step number parameter is increased to a high-magnification value of the preset denoising step number. When the comprehensive uncertainty score is less than the low uncertainty threshold, the denoising step number parameter is decreased to a low-magnification value of the preset denoising step number. In other cases, the denoising step number parameter remains unchanged.

10. The method for ocean temperature and salinity field reconstruction assessment assimilation prediction based on satellite remote sensing inversion according to claim 9, characterized in that, In the inverse interpolation strategy, interpolation points where the local horizontal gradient exceeds the gradient discrimination threshold are automatically switched to piecewise cubic Hermitian interpolation. The gradient discrimination threshold is determined through iterative experiments in the global Argo buoy dataset with the goal of minimizing the overall interpolation error.

11. The method for ocean temperature and salinity field reconstruction assessment assimilation prediction based on satellite remote sensing inversion according to claim 10, characterized in that, The anomaly removal process employs a double-precision accumulator to perform statistical calculations. The physical threshold and deviation threshold are determined through iterative statistical analysis of the temperature and salinity deviation distribution of each standard vertical layer in the historical Argo buoy dataset.

12. The method for ocean temperature and salinity field reconstruction assessment assimilation prediction based on satellite remote sensing inversion according to claim 11, characterized in that, The quality mask generates a binary labeled field based on whether the root mean square error after anomaly removal exceeds the assimilation error tolerance threshold. The assimilation error tolerance threshold is determined through iterative experiments in historical forecast experiments, with forecast skill scores as the target.

13. The method for ocean temperature and salinity field reconstruction assessment assimilation prediction based on satellite remote sensing inversion according to claim 12, characterized in that, The training dataset for the ocean temperature and salinity profile reconstruction model includes global Argo buoy temperature and salinity profile data from 1990 to the present. After quality control, the data is interpolated to a unified standard vertical layer using the sea surface height anomaly field, sea surface temperature field, and sea surface salinity field at the corresponding time and location as conditional vectors. The dataset is then divided into training set, validation set, and test set.

14. The method for ocean temperature and salinity field reconstruction assessment assimilation prediction based on satellite remote sensing inversion according to claim 13, characterized in that, The training of the ocean temperature and salinity profile reconstruction model uses the weighted sum of mean squared error loss and profile vertical gradient constraint loss as the total loss function, employs the AdamW optimizer, and uses a cosine decay of the initial learning rate after a warm-up strategy. Training stops when the validation set loss does not decrease for several consecutive cycles.

15. The method for ocean temperature and salinity field reconstruction assessment assimilation prediction based on satellite remote sensing inversion according to claim 14, characterized in that, The preset layer threshold is 5 layers; the regularization coefficient of the regularized optimal transport mapping is determined in the range of 0.01 to 1.0; the base ratio of the dynamic amplification coefficient is 16; the preset denoising step number parameter is 50, the high ratio value is 1.5 times the preset denoising step number, and the low ratio value is 0.7 times the preset denoising step number; the high uncertainty threshold corresponds to a ratio of comprehensive uncertainty score to reference standard value greater than or equal to 0.8, and the low uncertainty threshold corresponds to a ratio of less than 0.5; the jump step size is determined in the range of 1 to 5.

Images