Deep learning based sparse altimeter sea surface height short-term prediction method

Abstract

This invention relates to the fields of marine remote sensing data processing and artificial intelligence, and discloses a short-term sea surface height forecasting method based on deep learning using sparse altimeters. The method involves acquiring multi-source sub-satellite altimeter observation data, performing standardization processing, and preprocessing the daily average observation dataset with quality awareness to generate a quality-weighted observation tensor and a missing mask. The quality-weighted observation tensor and missing mask are then input into an end-to-end spatiotemporal neural network model, outputting a seamless sea surface height anomaly forecast field. Uncertainty estimation is performed using an ensemble sampling strategy and model dissimilarity indices to generate spatial pixel-level forecast confidence distribution data. Dynamic marine dynamic regions are divided based on multi-dimensional marine dynamic characteristics, and spatial filtering and bias correction are performed using corresponding post-processing parameter sets to infer the geostrophic current field. Evolutionary features are extracted to calculate a standardized anomaly index, generating a regional early warning signal. This invention skips the initial field interpolation and reconstruction stage, preventing the cumulative propagation of errors.

Description

Technical Field

[0001] This invention relates to the fields of marine remote sensing data processing and artificial intelligence technology, specifically to a method for short-term sea surface height forecasting based on sparse altimeter deep learning. Background Technology

[0002] Sea surface height is a key parameter characterizing ocean dynamic processes. Currently, multi-source nadir satellite altimeters are the main means of acquiring sea surface height observation data. However, due to the limited satellite orbital spacing, the along-orbit observation data acquired within a single day is spatially highly sparsely distributed. Existing short-term sea surface height forecasting methods typically employ a two-step processing flow: first, using methods such as objective analysis or optimal interpolation to interpolate and reconstruct a spatially continuous gridded initial field from the sparse observation data; then, inputting the initial field into numerical models or forecasting models to generate the future forecast field. However, traditional interpolation reconstruction methods usually rely on covariance matrix approximation, which can easily lead to a smooth transition of spatial features, making it difficult for the forecast results to accurately characterize the edge structure of small- and medium-scale ocean eddies. At the same time, the errors generated in the initial field interpolation reconstruction stage will accumulate and propagate to subsequent forecast stages, thereby reducing the absolute forecast accuracy of the final sea surface height forecast field.

[0003] Existing sea surface height forecasting systems mostly output single deterministic forecast results, lacking quantitative estimates of the uncertainty of forecast results. In actual marine environments, due to spatial differences in observation coverage density and marine dynamic system complexity, the forecasting difficulty and reliability of different spatial grid points vary significantly. Since existing methods cannot provide spatial pixel-level forecast confidence distribution data covering the target sea area, downstream users find it difficult to objectively assess the reliability of forecast results in different regions, thus limiting the data support role of forecast products in marine risk assessment and operational applications.

[0004] In terms of post-processing and application of forecast results, the physical characteristics of different ocean dynamic regions vary significantly. Traditional methods typically use globally unified post-processing parameters to perform error correction on the forecast field, without fully considering the spatial heterogeneity of the ocean dynamic environment. Using unified post-processing parameters makes it difficult to simultaneously achieve effective suppression of local high-frequency noise in the nearshore zone and elimination of systematic deviations in the high-energy zone of the open ocean. At the same time, existing forecast processing procedures often lack quantitative identification and hierarchical tracking mechanisms for anomalous evolution events that combine multi-dimensional ocean dynamic characteristics, resulting in insufficient environmental monitoring and dynamic warning capabilities of the forecast system under complex sea conditions. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a sparse altimeter-based short-term sea surface height forecasting method based on deep learning. This method solves the problems of error accumulation and propagation and smooth transition of spatial features caused by the reliance on initial field interpolation reconstruction in existing sea surface height forecasting methods. It also solves the problem of lacking pixel-level quantitative estimation of uncertainty in single deterministic forecasts, and the problem that using globally unified post-processing parameters is difficult to adapt to the spatial heterogeneity of the marine dynamic environment and the dynamic alarm of abnormal events.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] This invention provides a deep learning-based method for short-term sea surface height forecasting using sparse altimeters, comprising:

[0008] Acquire observation data from multi-source nadir altimeters and perform standardization processing to generate a daily average standardized sea surface height anomaly observation dataset covering the target sea area;

[0009] Quality-aware preprocessing is performed on the daily average standardized sea surface height anomaly observation dataset to identify and remove abnormal observation data. Quality scores are assigned to each spatial grid point based on the difference in observation quality, and a quality-weighted observation tensor and a missing mask indicating the coverage status of the observation location are generated.

[0010] The mass-weighted observation tensor and the missing mask are input into a pre-trained end-to-end spatiotemporal neural network model to output a seamless sea surface height anomaly prediction field.

[0011] Uncertainty estimation was performed on the seamless sea surface height anomaly forecast field using an integrated sampling strategy and model difference index. The corresponding forecast mean field and forecast variance field were calculated, and the forecast variance field was converted into spatial pixel-level forecast confidence distribution data.

[0012] Multidimensional ocean dynamics features are obtained, and dynamic ocean dynamic regions are divided based on these features. For different ocean dynamic regions, corresponding post-processing parameter sets are applied to perform spatial filtering and bias correction on the forecast mean field, generating a post-processed sea surface height anomaly field. The geostrophic velocity field is then inferred based on the post-processed sea surface height anomaly field.

[0013] The evolution characteristics of the post-processed sea surface height anomaly field relative to historical climatological data are extracted, the standardized anomaly index is calculated, and a regional early warning signal is generated based on the comparison results between the standardized anomaly index and the preset three-level early warning threshold.

[0014] The post-processed sea surface height anomaly field, geostrophic velocity field, forecast variance field, forecast confidence distribution data, and regional early warning signals are packaged into forecast product files for distribution.

[0015] In some embodiments, the quality-aware preprocessing of the daily average standardized sea surface height anomaly observation dataset, identifying and removing anomalous observation data, and generating a quality-weighted observation tensor and a missing mask includes: calculating the median and median absolute deviation of the time series corresponding to the spatial grid points; marking observation data with median absolute deviations exceeding a predetermined multiple as anomalous observation data and removing them; after removing anomalous observation data, constructing an observation quality weight matrix by combining signal-to-noise ratio, significant wave height, and backtracking distance deviation; and using the observation quality weight matrix to perform weighted processing on the daily average standardized sea surface height anomaly observation dataset to generate the quality-weighted observation tensor.

[0016] In some embodiments, the end-to-end spatiotemporal neural network model includes a four-dimensional variational data assimilation neural network branch and an encoder-decoder neural network branch. The step of inputting the quality-weighted observation tensor and the missing mask into the pre-trained end-to-end spatiotemporal neural network model to output a seamless sea surface height anomaly forecast field includes: acquiring historical quality-weighted sparse observation sequences and corresponding binary missing masks prior to the current forecast time; concatenating the historical quality-weighted sparse observation sequences and the binary missing masks along the feature channel dimension to generate an input tensor; and using the end-to-end spatiotemporal neural network model to receive the input tensor for forward inference and output a seamless sea surface height anomaly forecast field. During model training, the gradient norms of the state mean square error loss, spatial gradient mean square error loss, and dynamic regularization loss with respect to shared parameters are extracted. The moving average of the gradient norms is calculated, and the loss weight coefficients are adaptively updated based on the maximum value among the moving averages of the gradient norms.

[0017] In some embodiments, the process of performing uncertainty estimation on the seamless sea surface height anomaly forecast field using an integrated sampling strategy and model discrepancy index, calculating the corresponding forecast mean field and forecast variance field, and converting the forecast variance field into spatial pixel-level forecast confidence distribution data includes: calculating the spatial effective degrees of freedom using the eigenvalue spectrum of the historical sea surface height anomaly time series covariance matrix, determining the optimal random inactivation rate for each spatial region based on the effective degrees of freedom; performing multiple random sampling inferences according to the optimal random inactivation rate, combining the Monte Carlo random inactivation forecast variance with the inconsistency variance between the four-dimensional variational data assimilation neural network branch and the encoder-decoder neural network branch to generate the total forecast variance field, and using the sea surface height anomaly climatological variance as a reference benchmark to generate pixel-level forecast confidence distribution data through exponential decay mapping.

[0018] In some embodiments, the process of acquiring multi-dimensional ocean dynamic features, dynamically dividing ocean dynamic regions based on these features, applying corresponding post-processing parameter sets to the predicted mean field for different divided ocean dynamic regions to perform spatial filtering and bias correction, generating a post-processed sea surface height anomaly field, and inferring the geostrophic velocity field based on the post-processed sea surface height anomaly field includes: constructing a multi-dimensional feature vector by combining sea surface height anomaly variability, distance from the shore, seabed topographic gradient, and seasonal variation coefficient of eddy kinetic energy to divide the target sea area into a nearshore zone, an open ocean high-kinetic-energy zone, and an open ocean low-kinetic-energy zone; extracting the spatial Gaussian filter kernel radius and linear bias correction regression coefficients corresponding to each region, and performing spatial filtering and bias correction on the predicted mean field; calculating the geostrophic equilibrium velocity using Coriolis parameters for non-equatorial regions, and introducing equatorial Rossby parameters to approximate the velocity for equatorial regions.

[0019] In some embodiments, the step of extracting the evolution characteristics of the post-processed sea surface height anomaly field relative to historical climatological data, calculating a standardized anomaly index, and generating a regional early warning signal based on the comparison result of the standardized anomaly index and a preset three-level early warning threshold includes: calculating the difference between the post-processed sea surface height anomaly field and the mean of climatological sea surface height anomalies and normalizing it using the climatological standard deviation to generate a standardized anomaly index; aggregating the partial derivatives of the standardized anomaly index in the spatial and temporal dimensions to generate a spatiotemporal gradient anomaly index; comparing the standardized anomaly index with a preset classification threshold to determine the early warning level; identifying rapidly evolving anomaly areas when the spatiotemporal gradient anomaly index exceeds the spatiotemporal gradient anomaly threshold; and generating a persistent anomaly early warning signal for areas that continuously reach a specified early warning level and a persistence time threshold.

[0020] This invention provides a short-term sea level forecasting method based on deep learning using sparse altimeters. It offers the following advantages:

[0021] 1. This invention establishes a direct mapping from sparse satellite altimeter observation data to a future seamless sea surface height anomaly prediction field by directly inputting the mass-weighted observation tensor and missing mask into an end-to-end spatiotemporal neural network model. This end-to-end prediction method skips the intermediate step of gridded interpolation reconstruction of the initial observation field in the traditional prediction process, avoids the smooth transition problem caused by covariance matrix approximation in traditional interpolation methods, and prevents the cumulative transmission of initial field reconstruction errors to the prediction stage. This improves the spatial correlation coefficient and absolute prediction accuracy of the prediction results in characterizing the edge structure of small- and medium-scale ocean eddies.

[0022] 2. This invention utilizes an integrated sampling strategy and model discrepancy indices to perform uncertainty estimation on sea surface height anomaly forecast results, and integrates the Monte Carlo random inactivation forecast variance with the inconsistency variance between multi-branch models to calculate the total forecast variance field. By performing an exponential decay mapping transformation between the total forecast variance field and historical climatological reference variance, spatial pixel-level forecast confidence distribution data covering the target sea area is generated, quantifying the reliability of forecast results at different spatial grid points, and providing objective data support for the operational application of forecast products and downstream marine risk assessment.

[0023] 3. This invention dynamically divides the target sea area into different ocean dynamic regions based on multi-dimensional ocean dynamic characteristics, and configures parameters to perform spatial filtering and bias correction according to the differences in physical characteristics of each region. This region-adaptive post-processing mechanism effectively suppresses local high-frequency noise in the nearshore zone and systematic biases in the high-energy zone of the open ocean. At the same time, by combining the calculation of standardized anomaly index and spatiotemporal gradient anomaly index, it realizes the evolution identification and graded early warning of ocean anomalies, and improves the forecasting system's environmental monitoring and dynamic warning capabilities under complex sea conditions. Attached Figure Description

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

[0025] Figure 2 This is a schematic diagram of the dual-branch neural network structure of the present invention;

[0026] Figure 3 This is a comparison chart of the nRMSE scores of the various methods of the present invention under different forecast lead times;

[0027] Figure 4 This is a comparison chart of the relative deviations of the RMSE from the DUACS benchmark for various methods with a forecast lead time of 0 days in three types of ocean dynamic regions according to the present invention;

[0028] Figure 5 This is an example diagram showing the output of the uncertainty quantification module of the present invention;

[0029] Figure 6 This is a comparison chart of the accuracy of geostrophic flow velocity forecasts for the various methods of the present invention under different forecast lead times. Detailed Implementation

[0030] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Please see Figure 1 This invention provides a short-term sea surface height forecasting method based on deep learning using sparse altimeters. This end-to-end deep learning-based method directly maps sparse satellite altimeter observation data to a complete future sea surface height anomaly forecast field, skipping the initial field interpolation and reconstruction stage and avoiding the problem of initial field error accumulation and propagation. The end-to-end deep learning-based short-term sea surface height forecasting method specifically includes the following steps:

[0032] Step S100: Collect observation data from multi-source nadir altimeters and perform standardization processing to generate a daily average standardized sea surface height anomaly observation dataset covering the target sea area.

[0033] Step S200: Perform quality-aware preprocessing on the daily average standardized sea surface height anomaly observation dataset, identify and remove abnormal observation data, assign quality scores to each spatial grid point according to the difference in observation quality, and generate a quality-weighted observation tensor and a missing mask indicating the coverage status of the observation location.

[0034] Step S300: Input the mass-weighted observation tensor and the missing mask into the pre-trained end-to-end spatiotemporal neural network model, and output a seamless sea surface height anomaly forecast field within the specified future forecast period.

[0035] Step S400: Uncertainty estimation is performed on the seamless sea surface height anomaly forecast field using an integrated sampling strategy and model difference index. The corresponding forecast mean field and forecast variance field are calculated, and the forecast variance field is converted into spatial pixel-level forecast confidence distribution data.

[0036] Step S500: Obtain multi-dimensional ocean dynamic characteristics of the target sea area, perform dynamic ocean dynamic region division based on multi-dimensional ocean dynamic characteristics, apply corresponding post-processing parameter groups to perform spatial filtering and bias correction on the forecast mean field for different ocean dynamic regions, generate post-processed sea surface height anomaly field, and infer geostrophic velocity field based on post-processed sea surface height anomaly field.

[0037] Step S600: Extract the evolution characteristics of the post-processed sea surface height anomaly field relative to historical climatological data, calculate the standardized anomaly index, detect marine anomaly events based on the comparison results between the standardized anomaly index and the preset three-level early warning threshold, and generate regional early warning signals.

[0038] Step S700: The post-processed sea surface height anomaly field, geostrophic velocity field, forecast variance field, forecast confidence distribution data, and regional early warning signals are packaged into a forecast product file of a predetermined format and distributed in real time.

[0039] Step S100 involves collecting observation data from multi-source nadir altimeters and standardizing it to generate a daily average standardized sea surface height anomaly observation dataset covering the target sea area. Step S100 establishes a unified spatiotemporal reference and a rule-based data structure to address the inconsistency of data characteristics across different satellite platforms. This is achieved through the following sub-steps:

[0040] Step S110: Acquire along-orbit sea surface height observation data from multiple nadir satellite altimeters. The specific selection of nadir altimeters includes one or more combinations of Jason-3, Sentinel-6 Michael Freilich, SARAL / AltiKa, HY-2B, and Cryosat-2. The along-orbit observation data output by each altimeter includes basic attributes such as longitude, latitude, observation time, and sea surface height measurement.

[0041] Step S120 involves performing spatiotemporal coordinate unification processing on the sea surface height observation data along the orbit of each altimeter. Due to differences in orbital parameters and reference ellipsoid design, different satellite platforms have deviations in their spatial reference. Through a coordinate system transformation algorithm, all observation data from each satellite platform are uniformly transformed to the WGS84 geodetic coordinate system to eliminate spatial reference errors during multi-source data fusion.

[0042] Step S130 involves performing a spatial resolution resampling operation on the sea surface height observation data along the orbit in a unified coordinate system. Since the observation points obtained from the satellite's orbital scanning are spatially irregularly distributed, this embodiment uses a bilinear interpolation algorithm as the resampling method to map the irregularly distributed observation points along the orbit onto a regular grid with a spatial resolution of 1 / 4° latitude and longitude. The spatial resolution resampling operation transforms the discretely distributed orbital data into a gridded data structure that facilitates subsequent matrix operations.

[0043] Step S140: Perform temporal resolution normalization on the observation data mapped to the regular grid, and aggregate to generate daily average values. For cases where multiple satellite observations of the same regular grid point exist within the same natural day in the target sea area, calculate the arithmetic mean of multiple sea surface height measurements of the same regular grid point within the same day, thereby normalizing the temporal resolution to the daily average dimension.

[0044] Step S150: Calculate sea surface height anomaly data and construct a standardized observation data tensor. Subtract the corresponding mean dynamic topographic reference value from the daily average sea surface height observation value at each grid point to calculate the sea surface height anomaly value. Combine the calculation results for all grid points to output a daily average standardized sea surface height anomaly observation dataset covering global ocean areas. In the formula, Represents the number of days in the time dimension. Indicates the number of grid points in the latitudinal direction. This indicates the number of grid points along the longitude direction. In the daily average standardized sea surface height anomaly observation dataset generated by the above processing, due to the limitations of satellite orbital spacing, there are grid areas that are not covered by observations. The data bits corresponding to the grid points without observation coverage are marked as missing values.

[0045] Step S200 involves performing quality-aware preprocessing on the daily average standardized sea surface height anomaly observation dataset, identifying and removing outlier observations, assigning quality scores to each grid point based on observation quality differences, generating a quality-weighted observation tensor, and a missing mask indicating the coverage status of the observation location. Step S200 is specifically implemented through the following sub-steps:

[0046] Step S210 involves outlier detection and removal on the daily average standardized sea surface height anomaly observation dataset. Outlier detection specifically employs the median absolute deviation method. For each spatial grid point, the median and median absolute deviation of the corresponding time series are calculated. Observational data deviating from the median by more than a preset multiple of the median absolute deviation are marked as outliers and removed.

[0047] The outlier removal formula is specifically expressed as follows:

[0048]

[0049] in, Indicates the status of the outlier; This indicates that the corresponding observation data is an outlier, and the corresponding observation data will be removed. Represents the first in a time series One observation value; Indicates the median of a time series; The absolute deviation of the median is defined as follows: ; This represents the outlier removal threshold. In this embodiment, The value of is 3.

[0050] Step S220: Calculate the observation quality score for the observation data after removing outliers. The observation quality score comprehensively considers three indicators: signal-to-noise ratio, significant wave height, and backtracking distance deviation.

[0051] The specific formula for calculating the quality score is as follows:

[0052]

[0053] in, This represents the quality score, and the range of values ​​for the quality score is... ; Indicates the signal-to-noise ratio; Indicates the significant wave height; Indicates the backtracking distance deviation; This represents the normalized upper bound of the signal-to-noise ratio; This represents the normalized upper bound of the significant wave height; This represents the normalized upper bound of the backtracking distance deviation; , , These are the corresponding weighting coefficients. In this embodiment, The value is 0.4. The value is 0.3. The value is 0.3.

[0054] Step S230: Generate an observation quality weight matrix based on the quality scores of each grid point. Establish an observation quality weight matrix with dimensions consistent with the daily average standardized sea surface height anomaly observation dataset. In the observation quality weight matrix, the weight value corresponding to the grid point with valid observation data is set as the calculated quality score. The weight value for grid points without observation coverage is set to 0.

[0055] Step S240: Generate a binary missing mask and construct a quality-weighted observation tensor. Establish a binary missing mask with dimensions consistent with the observed data. In the binary missing mask, the mask value for grid points with valid observation data is set to 1, and the mask value for grid points without observation coverage is set to 0. The daily average standardized sea surface height anomaly observation dataset is weighted using an observation quality weight matrix, and the output is a quality-weighted observation tensor. and missing mask .

[0056] The formula for weighted processing is:

[0057]

[0058] in, Represents the mass-weighted observation tensor; Represents the observation quality weight matrix; This represents the daily average standardized sea surface height anomaly observation dataset; This indicates the Hadamard element-wise multiplication operation.

[0059] Step S300 involves inputting the quality-weighted observation tensor and the missing mask into a pre-trained end-to-end spatiotemporal neural network model, outputting a seamless sea surface height anomaly prediction field within a specified future forecast lead time. The specific mapping dimensions between the model input and output are established through the following sub-steps:

[0060] Step S310: Construct the input tensor of the end-to-end spatiotemporal neural network model. Obtain the historical quality-weighted sparse observation sequence before the current forecast time and the corresponding missing mask as the network input. In specific implementation, extract past... The mass-weighted sparse observation sequence and the binary missing mask are concatenated along the feature channel dimension to generate the input tensor.

[0061] The specific mathematical expression of the input tensor is:

[0062]

[0063] The semicolon indicates a concatenation operation along the feature channel dimension; Indicates the current forecast time; This indicates the number of days in the time window of a historical observation sequence; Indicates from the first Heavenly Supreme The mass-weighted observation tensor sequence of the day; This represents a binary missing mask sequence corresponding to the time period; Indicates the number of grid points in the latitudinal direction; Indicates the number of grid points along the longitude direction; This indicates the total number of feature channels after splicing.

[0064] For operational configuration, the parameter range is set to 10 to 21 days, with a default value of 14 days. The time window size for historical observation sequences can be automatically adjusted according to the dynamic timescales of different sea areas.

[0065] Step S320: Establish the output prediction target of the end-to-end spatiotemporal neural network model. The output prediction target is set to the future after the current prediction time. The daily seamless sea surface height anomaly field.

[0066] The mapping relationship of end-to-end neural prediction is represented as follows:

[0067]

[0068] in, Indicated by network weight parameters Characterized end-to-end neural prediction operators; Indicates the forecast lead time in days; Indicates the number of times after the forecast time. Sea level height anomaly forecast field.

[0069] Forecast lead time days The value range is set to 5 to 10 days, with a default value of 7 days; forecast lead time days. It supports automatic adjustment based on sea area dynamics timescales. By configuring the input tensor and output forecast target, the end-to-end spatiotemporal neural network model has the ability to directly receive sparse observation data for forward inference, skipping the intermediate steps of interpolation and reconstruction.

[0070] In one specific embodiment of the present invention, the end-to-end spatiotemporal neural network model is implemented using a dual-branch neural network architecture. Please refer to [link to relevant documentation]. Figure 2 This dual-branch architecture includes an upper-level encoder-decoder neural network branch (UNet branch) and a lower-level four-dimensional variational data assimilation neural network branch (4DVarNet branch). By configuring a switch, the system supports parallel operation of the two branches and outputs a first deterministic forecast result and a second deterministic forecast result. Specifically, forward inference forecasting is performed through the following sub-steps:

[0071] Step S330: Configure the running branches of the end-to-end spatiotemporal neural network model. The end-to-end spatiotemporal neural network model includes a four-dimensional variational data assimilation neural network branch and an encoder-decoder neural network branch. The running states of the four-dimensional variational data assimilation neural network branch and the encoder-decoder neural network branch are switched and controlled by a configuration switch. During the inference phase, the configuration switch supports a single-branch independent running mode, as well as a dual-branch parallel running mode to generate complementary forecast modes.

[0072] Step S340: While the four-dimensional variational data assimilation neural network branch is running, iterative optimization prediction based on the variational data assimilation framework is executed. The four-dimensional variational data assimilation neural network branch embeds the variational data assimilation framework into the neural network structure. The four-dimensional variational data assimilation neural network branch contains three core components: a learnable variational cost function, a gradient descent solver, and a state update network. The parameter scale of the four-dimensional variational data assimilation neural network branch is within... to Magnitude.

[0073] Step S341: In each iteration of the solution of the four-dimensional variational data assimilation neural network branch, calculate the gradient of the current state estimate relative to the learnable variational cost function. Update the current state using the gradient descent step parameterized by the neural network. The four-dimensional variational data assimilation neural network branch is composed of learnable prior operators. With cost gradient update Alternatingly executing a preset number of steps achieves variational assimilation under sparse observations. After a set number of iterations, the final prediction result of the four-dimensional variational data assimilation neural network branch is output. The number of iterations is set to a range of 10 to 15.

[0074] Step S350: While the encoder-decoder neural network branch is running, perform spatial resolution prediction based on multi-scale feature extraction and recovery. The encoder-decoder neural network branch adopts a symmetrical encoder-decoder structure, with parameter scales ranging from... to Magnitude.

[0075] Step S351 involves extracting multi-scale spatiotemporal features from the multi-source input data using the encoder through successive convolutional layers and downsampling operations. The encoder comprises four downsampling modules, each containing two 3×3 convolutional layers and one 2×2 max-pooling layer. The downsampling modules include downsampling layers E1 to E3 and a bottleneck layer (Bot). The number of channels in each layer is configured as 64, 128, 256, and 512, respectively.

[0076] Step S352: The spatial resolution of the feature map is restored using the decoder through upsampling operations and skip connections. The decoder is symmetrically configured with four upsampling modules. The upsampling modules include upsampling layers D1 to D3, with the number of channels in each layer configured as 256, 128, and 64 respectively. A channel attention module is added at the skip connection between the encoder and decoder. The channel attention module is used to suppress noise propagation in the missing measurement regions.

[0077] Step S360: Obtain the forecast results output by the neural network branch in the running state, use the forecast results as a seamless sea surface height anomaly forecast field within the specified future forecast period, and send it to the subsequent uncertainty quantification module.

[0078] In one specific embodiment of the present invention, the end-to-end spatiotemporal neural network model needs to undergo supervised learning training before deployment and inference. The training phase of the end-to-end spatiotemporal neural network model is configured, specifically through the following sub-steps to implement an adaptive model training mechanism:

[0079] Step S370: Construct a training sample set and perform sparse observation data augmentation. The daily average global sea level height anomaly field from historical ocean reanalysis products is used as the ground truth label data. By simulating the satellite's along-orbit sampling geometry, sampling masks are randomly extracted from a historical real orbit database to generate synthetic sparse observation data as input samples for the end-to-end spatiotemporal neural network model.

[0080] Step S371: During the training iteration, data augmentation is performed on the input synthetic sparse observation data. The retention rate of the sampling mask is randomly adjusted to the range of 0.02 to 0.10 to simulate different observation densities. Zero-mean heavy-tailed noise is superimposed on the observations to improve the noise robustness of the end-to-end spatiotemporal neural network model. When dealing with extreme anomalous batches of data where the input is completely empty, the system reverts to the prediction results of the previous time step and records alarm information.

[0081] Step S380: Configure a combination of multiple loss functions for supervised learning. Depending on the selected network architecture, the loss function terms include the state mean squared error loss, the spatial gradient mean squared error loss, and the dynamic regularization loss. When running the four-dimensional variational data assimilation neural network branch, the total training loss function integrates these three losses. The specific mathematical expression of the total training loss function is:

[0082]

[0083] in, Indicates the total training loss; This represents the weighting coefficient corresponding to the state mean square error loss; This represents the weighting coefficient corresponding to the spatial gradient mean square error loss; This represents the weighting coefficient corresponding to the dynamic regularization loss; This represents the state mean square error loss; This represents the spatial gradient mean square error loss; This represents the dynamic regularization loss. When running the encoder-decoder neural network branch, the total training loss function only includes the state mean squared error loss and the spatial gradient mean squared error loss.

[0084] Step S381: Calculate the state mean square error loss. The specific mathematical expression for the state mean square error loss is:

[0085]

[0086] in, This represents the state mean square error loss; Indicates the total number of days in advance for the forecast; Indicates the forecast time The next The true sea level height anomaly field of the sky; This represents the sea surface height anomaly field predicted by the corresponding end-to-end spatiotemporal neural network model; Indicates the first The time weight of each forecast lead time.

[0087] Step S382: Calculate the time weights for different forecast lead times. Time weights A lesson-based, progressive strategy is employed for dynamic value assignment. In the initial training phase, only dynamic values ​​are assigned to... The first day's forecast is assigned a non-zero weight. Every set number of iterations... Increase the validity period of the valid assignment by 1 day until all values ​​are covered. Forecast lead time; set iteration rounds The value is taken over two training epochs. After entering the stable training period, the time weight is calculated according to the exponential decay formula: The attenuation coefficient The value is 0.1.

[0088] Step S383: Calculate the spatial gradient mean square error loss. Replace the actual sea surface height anomaly field and the predicted sea surface height anomaly field with the corresponding spatial gradient values. Using the same formula structure and curriculum-based asymptotic strategy as the state mean square error loss, calculate the spatial gradient mean square error loss. .

[0089] Step S390: Perform runtime adaptive loss weight balancing update based on gradient norm. Instead of using fixed empirical weight coefficients, after each backpropagation calculation, extract the gradient norms of the state mean squared error loss, spatial gradient mean squared error loss, and dynamic regularization loss for the shared parameters of the end-to-end spatiotemporal neural network model, and calculate the moving average of the gradient norms.

[0090] Update the weight coefficients of each loss term according to the following rules:

[0091]

[0092] in, Indicates the first During the nth iteration The weighting coefficients corresponding to the item loss; Indicates the first During the nth iteration The weighting coefficients corresponding to the item loss; This represents the momentum coefficient, with a value of 0.05. Indicates the first The moving mean of the gradient norm of the term loss; This represents the maximum value among the moving averages of the three loss gradient norms; This is a numerically stable term, taking the value of Initial weighting coefficients , , All values ​​are set to 1, and the update results of the weight coefficients are pruned and restricted within the logarithmic domain. Within the range.

[0093] Step S391: Configure the optimizer to perform iterative updates of the end-to-end spatiotemporal neural network model parameters. The optimizer uses an adaptive moment estimation optimizer, and the base learning rate employs a warm-up-cosine annealing strategy. The learning rate is controlled to linearly increase to the peak learning rate within the first 5000 steps. It then gradually decreases according to the cosine curve to Peak learning rate The value is automatically determined using the learning rate probing method and its range is defined within [a certain range]. to between.

[0094] Step S400 involves performing uncertainty estimation on the seamless sea surface height anomaly forecast field using an ensemble sampling strategy and a model dissimilarity index. While the end-to-end spatiotemporal neural network model outputs deterministic forecast results, the forecast variance is output using an adaptive Monte Carlo Dropout (MCDropout) method based on spatiotemporal covariance analysis. The specific calculation of the forecast variance is implemented through the following sub-steps:

[0095] Step S410: Extract the sea surface height anomaly time series for each spatial region in the training set, and calculate the covariance matrix of the sea surface height anomaly time series for each spatial region in the training set. Perform eigenvalue decomposition on the covariance matrix of each spatial region to obtain the eigenvalue spectrum of the covariance matrix. Calculate the effective degrees of freedom corresponding to each spatial region based on the eigenvalue spectrum.

[0096] The formula for calculating the effective degrees of freedom is:

[0097]

[0098] in, This represents the effective degrees of freedom of the calculated spatial region; The first element representing the spatial region covariance matrix is... Each feature value.

[0099] Step S420: Based on the calculated effective degrees of freedom of each spatial region, determine the optimal random inactivation rate of the end-to-end spatiotemporal neural network model in the corresponding spatial region.

[0100] The formula for calculating the optimal random inactivation rate is:

[0101]

[0102] in, This represents the optimal random inactivation rate for the corresponding spatial region; This represents the maximum effective degrees of freedom across the entire target sea area. By constraining the upper bound to 0.5 and the lower bound to 0.05, the configuration rule ensures that regions with high signal complexity are matched with a lower random inactivation rate to preserve model parameter capacity, while regions with simple signals are matched with a higher random inactivation rate to enhance the network's regularization effect.

[0103] Step S430: During the forward inference phase of the end-to-end spatiotemporal neural network model, the randomly deactivated layer within the network model remains active. For the same received input tensor, multiple random sampling inferences are performed according to the optimal random deactivated rate set in step S420. Each inference produces different sea surface height anomaly prediction results due to the randomness of the random deactivated layer. The number of random sampling inferences for the same input tensor is set to... Next, the output is... A set of forecast results In this embodiment, the number of random sampling inferences... The value is set to 50.

[0104] Step S440, gather The set of sea surface height anomaly forecast results output by random sampling inference is used to calculate the mean and variance of the MCDropout ensemble forecast for each grid point.

[0105] The formula for calculating the mean of the MCDropout ensemble forecast is:

[0106]

[0107] in, This represents the mean of the MCDropout ensemble forecasts; Indicates the number of random sampling inferences; Indicates the first The sea surface height anomaly prediction results output by random sampling inference.

[0108] The formula for calculating the variance of MCDropout predictions is:

[0109]

[0110] in, This represents the MCDropout forecast variance calculated based on Monte Carlo random sampling. This forms the basis for the calculation of the total forecast variance.

[0111] The end-to-end spatiotemporal neural network model is configured for dual-branch parallel operation during the inference phase. After obtaining the MCDropout prediction variance calculated based on Monte Carlo random sampling, the following sub-steps are used to perform fusion uncertainty calculation and pixel-level confidence transformation:

[0112] Step S450: Calculate the forecast inconsistency variance introduced by different neural network structures. Extract the deterministic forecast results from the output of the four-dimensional variational data assimilation neural network branch, and the deterministic forecast results from the output of the encoder-decoder neural network branch. Calculate the squared difference between the two sets of deterministic forecast results at spatial grid points to obtain the inter-model inconsistency variance.

[0113] The formula for calculating the variance of inconsistency between models is:

[0114]

[0115] in, Indicates the variance of inconsistency between models; This represents the deterministic prediction result output by the branch of the four-dimensional variational data assimilation neural network independently performing forward inference; This represents the deterministic prediction result output by the forward inference branch of the encoder-decoder neural network, which executes independently.

[0116] Step S460: Integrate multiple sources of uncertainty to calculate the total forecast variance. The MCDropout forecast variance and the inter-model inconsistency variance are weighted and fused to output the total forecast variance field covering the target sea area.

[0117] The formula for calculating the total forecast variance is:

[0118]

[0119] in, This represents the total forecast variance in the calculated output; This represents the MCDropout forecast variance calculated based on Monte Carlo random sampling; This represents the calculated variance of inconsistency between models; This represents the fusion weighting coefficients configured for the MCDropout forecast variance. This represents the fusion weight coefficients configured to account for inconsistencies in variance between models. In this embodiment, The value is 0.6. The value is 0.4.

[0120] Step S470: Obtain the climatological variance of sea surface height anomalies for the corresponding spatial grid points, and calculate the pixel-level reference variance. To avoid numerical anomalies with zero denominators in the confidence transformation calculation, a piecewise function is used to configure the reference variance:

[0121]

[0122] in, This represents the calculated reference variance; This represents the climatological variance of sea surface height anomalies at the corresponding spatial grid points; This represents the lower bound of the minimum reference variance, and its value is [value missing]. .

[0123] Step S480: Based on the calculated total forecast variance and reference variance, the pixel-level confidence level corresponding to each spatial grid point is calculated using an exponential decay function.

[0124]

[0125] in, This represents the pixel-level confidence score of the calculated output, with a value range of [value range missing]. interval; This represents the total forecast variance. Confidence level. A value close to 1 indicates high forecast reliability, while a value close to 0 indicates low forecast reliability.

[0126] Step S490: Set the confidence threshold In this embodiment, the value is taken as 0.3. The pixel-level confidence level is compared. With threshold Confidence level Below the threshold The spatial region is identified as a low-confidence region.

[0127] Step S491, when determining the climatological variance of the spatial grid points This indicates a lack of effective historical benchmarks, and the corresponding pixel-level confidence level will be adjusted accordingly. Assign a value of 0 and record the exception flag.

[0128] Step S492: Output the results data from the uncertainty quantification stage. The output dataset includes the forecast mean field. Total forecast variance field and spatial confidence distribution map .

[0129] Step S500 involves performing regional adaptive post-processing on the forecast results. Specifically, this is achieved through the following sub-steps: multi-dimensional dynamic region segmentation, spatial filtering, bias correction, and flow velocity estimation.

[0130] Step S510: Obtain multi-dimensional ocean dynamic features of the target sea area and construct a feature vector. These multi-dimensional ocean dynamic features include the historical time variability standard deviation of sea surface height anomalies, distance from the shore, seafloor topographic gradient, and seasonal variation coefficient of eddy kinetic energy. Extract the multi-dimensional ocean dynamic features to form the corresponding four-dimensional feature vector.

[0131] The specific mathematical expression of a four-dimensional eigenvector is:

[0132]

[0133] in, This represents the constructed four-dimensional feature vector; The standard deviation of the historical time variability of sea level anomalies; Indicates the distance from the shore; Indicates the gradient of seabed topography; The coefficient of variation of eddy kinetic energy represents the seasonal variation of eddy kinetic energy.

[0134] Step S520: Based on the four-dimensional feature vector, a fuzzy C-means clustering algorithm is used to perform soft region division of the target sea area. The fuzzy C-means clustering algorithm is used to calculate the membership weight of each spatial grid point to each preset ocean dynamic region category. Membership weights are applied to avoid abrupt changes in post-processing parameters at hard boundaries.

[0135] Step S530 involves dynamically classifying the target sea area based on the historical time variability standard deviation of sea surface height anomalies and the distance from the shore. Specific categories for this classification include the nearshore zone, the open ocean high-energy zone, and the open ocean low-energy zone.

[0136] Step S531, the judgment condition for region classification is:

[0137]

[0138] in, This indicates the classification results of the delineated marine dynamic regions; This represents the threshold distance from the shore to the nearshore zone, which is set to 200 km in this embodiment. The standard deviation threshold for sea surface height anomalies, representing the boundary between high and low variability, is set to 8 cm in this embodiment. The regional classification results for the target sea area are set to be updated monthly.

[0139] Step S540: After completing the dynamic region classification, configure corresponding post-processing parameter sets for each of the different ocean dynamic regions obtained. The post-processing parameter sets specifically include the spatial Gaussian filter kernel radius, bias-corrected regression coefficients, and confidence threshold.

[0140] Step S550: Extract the spatial Gaussian filter kernel radius corresponding to each ocean dynamic region and perform spatial filtering operation on the predicted mean field.

[0141] The specific calculation formula for spatial filtering is as follows:

[0142]

[0143] in, This represents the spatially filtered sea surface height anomaly field. The standard deviation is expressed as A two-dimensional Gaussian filter kernel; Indicates the radius of the spatial Gaussian filter kernel; Represents a two-dimensional convolution operation; This represents the input forecast mean field.

[0144] Step S551: Adaptively configure the spatial Gaussian filter kernel radius based on the classification results of the ocean dynamics region. The numerical value. In the nearshore zone region, the spatial Gaussian filter kernel radius. The value is configured as 25km; in the open ocean high-energy zone region, the spatial Gaussian filter kernel radius is... The value is set to 50km; in the open ocean low-energy zone region, the radius of the spatial Gaussian filter kernel is... The value is configured as 75km.

[0145] Step S560: Based on the predetermined deviation correction regression coefficients, perform linear deviation correction on the spatially filtered sea surface height anomaly field to generate a post-processed sea surface height anomaly field.

[0146] The specific calculation formula for deviation correction is as follows:

[0147]

[0148] in, This indicates the post-processed sea surface height anomaly field; Represents the regression coefficient; Represents the regression coefficient; This represents the spatially filtered sea surface height anomaly field.

[0149] Step S561: Configure the mechanism for obtaining and updating the deviation-corrected regression coefficients. Regression coefficients and regression coefficients The regression coefficients were determined based on cross-validation calculations using historical forecast data and independent observation data. and regression coefficients The value is set to update monthly.

[0150] Step S570: After generating the post-processed sea surface height anomaly field, the geostrophic velocity field is calculated based on the post-processed sea surface height anomaly field. According to the latitude of the spatial grid points, the target sea area is divided into equatorial and non-equatorial regions, and different dynamic calculation logics are used for the equatorial and non-equatorial regions respectively.

[0151] Step S580: In non-equatorial regions, calculate the sea surface geostrophic current velocity based on the geostrophic balance relationship. Calculate the corresponding zonal and meridional current velocities using the post-processed sea surface height anomaly field.

[0152] The specific formula for calculating the zonal velocity is:

[0153]

[0154] The specific formula for calculating the meridional velocity is:

[0155]

[0156] in, Indicates zonal flow velocity; Indicates meridional velocity; Represents gravitational acceleration, with a value of ; The Coriolis parameter is defined as follows: ; This represents the Earth's angular velocity of rotation; Indicates latitude; This represents the post-processed sea surface height anomaly field after bias correction; This represents the climatological component of the zonal geostrophic velocity corresponding to the average dynamic topography. This represents the climatological component of the meridional geostrophic velocity corresponding to the average dynamic topography.

[0157] Step S581: Calculate the climatological component of the geostrophic flow velocity and perform spatial differentiation operation. and The computational logic is defined as follows: and ,in This represents the average dynamic terrain. In the extrapolation process, the spatial differential uses a central difference scheme.

[0158] Step S590: Perform an approximate calculation of the geostrophic velocity in the equatorial region. (The last part, "when latitude...", appears to be incomplete and requires further context.) At that time, Coriolis parameters Approaching zero. When the Coriolis parameter The absolute value is less than the threshold When this occurs, the corresponding geostrophic flow velocity is marked as an invalid value, where the threshold value is... .

[0159] Step S591: Apply equatorial force to spatial grid points marked as invalid values. A planar approximation method is used for substitution calculations. Near the equator, the Coriolis parameters are... Linearization processing .

[0160] in, The Rossby parameter for the equator is defined as follows: ; Indicates the Earth's radius; Indicates the reference latitude; This represents the north-south distance from the equator. By substituting the linearized Coriolis parameters into the calculation formulas for zonal and meridional current velocities, the approximate geostrophic current velocity in the equatorial region is obtained.

[0161] Step S592: Integrate the calculation results from non-equatorial and equatorial regions, and output the post-processed sea surface height anomaly field. and geostrophic flow velocity field .

[0162] Step S600 involves extracting the evolution characteristics of the post-processed sea surface height anomaly field relative to historical climatological data, calculating the standardized anomaly index, detecting marine anomalies, and generating regional early warning signals. Specifically, this is achieved through the following sub-steps: spatiotemporal anomaly event detection and early warning generation.

[0163] Step S610: Extract historical climatological statistics corresponding to the post-processed sea surface height anomaly field. These historical climatological statistics are calculated based on historical observation data and specifically include the mean and standard deviation of the climatological sea surface height anomaly for the target spatial grid points on the corresponding dates. In this embodiment, the mean and standard deviation of the climatological sea surface height anomaly are calculated based on historical data from 2010 to 2019.

[0164] Step S620: Combine the post-processed sea surface height anomaly field with historical climatological statistics to calculate the standardized anomaly index of the forecast results relative to historical climatological data.

[0165] The specific formula for calculating the standardized anomaly index is as follows:

[0166]

[0167] in, Represents spatial grid points At any moment Standardized anomaly index; Represents spatial grid points At any moment Post-processing of sea surface height anomaly field numerical values; Represents spatial grid points The first of the year The average climatological sea level height anomaly of the day; Represents spatial grid points The first of the year The standard deviation of the climate state of a day.

[0168] Step S621: Configure a processing mechanism for a climatological standard deviation of zero. When the climatological standard deviation of the corresponding spatial grid point is zero... When the value is 0, the standardized anomaly index of the corresponding spatial grid point is... The value is assigned to 0, and the corresponding anomaly identifier is recorded in the forecast system.

[0169] Step S630: Based on the calculated standardized anomaly index, perform a classification judgment of marine anomalies. Set three-level early warning thresholds based on the standardized anomaly index and generate corresponding early warning levels.

[0170] The criteria for determining the warning level are as follows:

[0171]

[0172] in, Indicates the warning level as determined; This represents the standardized anomaly index of the corresponding spatial grid point.

[0173] Step S640: Simultaneously, based on the calculated standardized anomaly index, the spatiotemporal gradient anomaly index of the target sea area is calculated. The partial derivatives of the standardized anomaly index in the spatial and temporal dimensions are calculated using the central difference scheme, and then aggregated to generate the spatiotemporal gradient anomaly index.

[0174] The specific formula for calculating the spatiotemporal gradient anomaly index is as follows:

[0175]

[0176] in, Indicates a spatiotemporal gradient anomaly index; This represents the spatial differential of the standardized anomaly index in the meridional direction; This represents the spatial differential of the standardized anomaly index in the latitudinal direction. This represents the time derivative of the standardized anomaly index in the time dimension. During the calculation, both the spatial and temporal derivatives are discretized using the central difference scheme.

[0177] Step S650: Identify rapidly evolving anomalous regions based on spatiotemporal gradient anomaly indicators. Configure spatiotemporal gradient anomaly thresholds. Spatiotemporal gradient anomaly indicators for each spatial grid point. With spatiotemporal gradient anomaly threshold Compare them. When the spatiotemporal gradient anomaly index... Exceeding the spatiotemporal gradient anomaly threshold At that time, the corresponding spatial grid points are marked as rapidly evolving anomalous regions.

[0178] Step S651: Configure the spatiotemporal gradient anomaly threshold The rules for determining the value of the spatiotemporal gradient anomaly threshold. The value is the 95th percentile of the historical spatiotemporal gradient anomaly index of the corresponding spatial grid point.

[0179] Step S660: Continuously track spatial regions that continuously exceed the warning threshold. When a spatial region continuously... When the system remains at the warning or alarm level, a persistent abnormality warning is generated. The value is 3 days.

[0180] Step S670: Output the complete results data for marine anomaly event detection. The output dataset includes anomaly index spatial distribution map. A tiered early warning indicator map and a continuous abnormal warning signal.

[0181] Step S700 involves the operationalization of forecast product generation and deployment after anomaly detection is completed. Specifically, forecast product generation and deployment are performed through the following sub-steps:

[0182] Step S710: Collect the output data from the sea surface height anomaly forecasting stage, the uncertainty quantification stage, and the anomaly event detection stage to construct a multidimensional ocean forecasting dataset.

[0183] The tensor structure representation of the multidimensional ocean forecast dataset is as follows:

[0184]

[0185] in, Indicates at time Spatial grid points The forecast data set; This indicates the post-processed sea surface height anomaly field; Indicates zonal flow velocity; Indicates meridional velocity; Indicates the total forecast variance; Indicates pixel-level confidence level; Indicates the warning level.

[0186] Step S720: Perform format conversion and standardization encapsulation on the multidimensional ocean forecast dataset. Encode the multidimensional ocean forecast dataset according to climate and forecast metadata conventions to generate a standardized network-wide data format file.

[0187] Step S721: Write geospatial coordinate variables, time variables, and corresponding attribute metadata into the network general data format file.

[0188] Step S730: Deploy standardized network-wide data format files to the operational forecasting system server. Configure the application programming interface and Web map service to publish the post-processed sea surface height anomaly field, geostrophic flow velocity field, and graded early warning labels as spatial visualization layers.

[0189] Step S731: Configure a timed triggering task to control the end-to-end spatiotemporal neural network model to receive the latest sparse altimeter observation data daily. By cyclically executing forward inference forecasting, regional adaptive post-processing, and spatiotemporal anomaly detection, an updated multidimensional ocean forecast dataset is generated and automatically pushed to the operational forecast system server.

[0190] After deploying the end-to-end spatiotemporal neural network model, multiple sets of comparative experiments were configured to evaluate forecast performance, regional adaptability, accuracy of geostrophic dynamics, and uncertainty quantification effects. The experimental setup and comparison data for quantitative evaluation are as follows:

[0191] Training data were generated from the historical daily average global sea level height anomaly field of the Global Ocean Reanalysis Simulation System (GLORYS12), with a spatial resolution of 1 / 4°. Synthetic sparse observation data was generated as network input by simulating the along-track sampling patterns of multiple nadir altimeters. The independent observation data used for evaluation included time-delayed along-track sea level height anomalies and the trajectory velocity of a drifting buoy at a depth of 15m.

[0192] Quantitative evaluation indicators include: normalized root mean square error score, accuracy of flow direction, and accuracy of flow magnitude.

[0193] The specific formula for calculating the normalized root mean square error score is as follows:

[0194]

[0195] in, This represents the forecast value. Represents the observed value. This represents the sample size. The normalized root mean square error score ranges from [value missing]. The closer the value is to 1, the higher the forecast accuracy.

[0196] The accuracy rate of flow velocity direction is defined as the percentage of samples in which the angle between the predicted flow velocity direction and the observed flow velocity direction is less than 30° out of the total number of samples; the accuracy rate of flow velocity magnitude is defined as the percentage of samples in which the relative error between the predicted flow velocity magnitude and the observed flow velocity magnitude is less than 30% out of the total number of samples.

[0197] The comparison methods include: the GLO12 operational assimilation prediction system, the GloNet neural ocean simulator, and the XiHe neural ocean simulator. The optimal interpolation product with DUACS delay time is used as a reference benchmark. The method of this invention tests the neural network branch of four-dimensional variational data assimilation and the encoder-decoder neural network branch.

[0198] Normalized root mean square error (RMSE) scores were extracted for forecast leads of 1 to 10 days to perform baseline model comparison experiments. Table 1 shows the RMSE scores of each method for forecast leads of 0 to 7 days. The neural network branch and encoder-decoder neural network branch of the four-dimensional variational data assimilation outperformed the GLO12 assimilation forecast system and the two comparative neural ocean simulators across all forecast leads.

[0199] Please see Figure 3 This paper presents the evolution trend of the normalized root mean square error score for each method within a forecast lead time of 0 to 7 days. The corresponding quantitative values ​​are recorded in Table 1.

[0200] Table 1 Comparison of Normalized Root Mean Square Error Scores for Different Methods under Different Forecast Lead Times

[0201]

[0202] Combination Figure 3 As shown in the data analysis in Table 1, the four-dimensional variational branch provided by this invention remains optimal under all forecast lead times. When the forecast lead time is 7 days, the accuracy improvement of this invention relative to the GLO12 system (0.056) is greater than that when the forecast lead time is 0 days (0.044). Therefore, it can be concluded that the longer the forecast lead time, the more significant the relative advantage of the end-to-end mapping mechanism of this invention in suppressing error accumulation.

[0203] Please see Figure 4 This table shows the comparison of the root mean square error deviations of each method relative to the DUACS benchmark in three types of ocean dynamic regions when the forecast lead time is 0 days. The corresponding detailed percentage data are shown in Table 2.

[0204] Table 2. Root mean square error deviation (%) of each method relative to DUACS in three types of ocean dynamic regions with a forecast lead time of 0 days.

[0205]

[0206] according to Figure 4 As shown in Table 2, the four-dimensional variational branch of this invention achieved a negative bias (-7.3%) in the high-energy zone of the open ocean, indicating that its prediction accuracy surpasses that of the optimal interpolation product. This result demonstrates the suppressive effect of dynamic regional parameter configuration on local high-frequency noise and systematic bias, enabling this invention to exhibit robust adaptability in complex ocean dynamic environments.

[0207] Please see Figure 6 The small figures (a) and (b) show the decay curves of the accuracy of the flow direction and the accuracy of the flow magnitude, respectively. The corresponding data can be found in Tables 3 and 4.

[0208] Table 3. Accuracy of flow direction (%) of each method under different forecast lead times

[0209]

[0210] Table 4. Accuracy (%) of flow velocity magnitude for each method under different forecast lead times

[0211]

[0212] Experimental results show that when the forecast lead time is 0 days, the accuracy of the branch velocity direction of the neural network for four-dimensional variational data assimilation reaches 45.8%, which is 7.6 percentage points higher than that of GLO12; the accuracy of the velocity magnitude reaches 39.2%, which is 6.7 percentage points higher than that of GLO12.

[0213] Furthermore, the output effect of the integrated uncertainty quantization module is verified. Please refer to [link / reference]. Figure 5It showcases an example of the output from the uncertainty quantification module, including the SLA forecast field, forecast variance field, and confidence level distribution map. By combining the obtained forecast variance and confidence level of the forecast grid points, we can observe... Figure 5 From (b) and (c), we can see that the variance of the forecast variance field increases to [value missing] in the high-variance region of the western boundary flow system. In the above, the variance remains stable in low-variance regions such as the inner tropical ocean. The magnitude of the predicted variance is approximately 8 times that of the high-variance region and the low-variance region. The confidence level distribution is strongly negatively correlated with the spatial model of predicted variance, and regions with low confidence levels are spatially highly consistent with regions with high predicted variance. Furthermore, the rank correlation coefficient between predicted variance and absolute prediction error reaches 0.82, confirming that the uncertainty estimate provided by this invention is highly consistent with the actual error distribution.

[0214] The overall normalized root mean square error score improved by 0.044 for a forecast lead time of 0 days and by 0.056 for a forecast lead time of 7 days. In terms of the accuracy of geostrophic flow velocity direction and magnitude, it has shown a continuous percentage improvement compared to the existing operational system. The uncertainty quantification index can effectively characterize the forecast difficulty and achieve robust short-term forecasts for complex sea surface height changes.

[0215] Based on the same inventive concept, this invention provides a short-term sea surface height forecasting system based on end-to-end deep learning using sparse altimeters, corresponding to the aforementioned method. Since the principle underlying the problem solved by the forecasting system is similar to that of the aforementioned method, its specific implementation is the same as described above, and repeated details will not be elaborated further. The forecasting system specifically includes the following functional modules:

[0216] The observation data acquisition module is configured to collect observation data from multi-source sub-satellite altimeters and perform standardization processing to generate a daily average standardized sea surface height anomaly observation dataset covering the target sea area.

[0217] The quality-aware preprocessing module is configured to perform quality-aware preprocessing on the daily average standardized sea surface height anomaly observation dataset, identify and remove abnormal observation data, and generate a quality-weighted observation tensor and a missing mask indicating the coverage status of the observation location.

[0218] The end-to-end network forecasting module is configured to input the quality-weighted observation tensor and the missing mask into a pre-trained end-to-end spatiotemporal neural network model and output a seamless sea surface height anomaly forecast field.

[0219] The forecast uncertainty quantification module is configured to perform uncertainty estimation on the seamless sea surface height anomaly forecast field using an integrated sampling strategy and model difference index, calculate the forecast mean field and forecast variance field, and convert them into spatial pixel-level forecast confidence distribution data.

[0220] The regional adaptive post-processing module is configured to perform dynamic ocean dynamics region division based on multi-dimensional ocean dynamics characteristics, apply the corresponding post-processing parameter set to perform spatial filtering and bias correction on the forecast mean field, generate a post-processed sea surface height anomaly field, and calculate the geostrophic velocity field.

[0221] The marine anomaly detection module is configured to extract the evolution characteristics of the post-processed sea surface height anomaly field relative to historical climatological data, calculate the standardized anomaly index and perform threshold comparison, and generate regional early warning signals.

[0222] The forecast product generation and distribution module is configured to encapsulate post-processed sea surface height anomaly field, geostrophic velocity field, forecast variance field, forecast confidence distribution data and regional early warning signals into forecast product files for distribution.

[0223] Based on the same inventive concept, this invention provides an electronic device, specifically including a processor, a memory, a communication interface, and a communication bus. The processor, memory, and communication interface communicate with each other via the communication bus. The memory internally stores a computer program executable by the processor, and the memory specifically includes high-speed random access memory and non-volatile memory. When the processor executes the computer program, it implements the sparse altimeter sea surface height short-term forecasting method based on end-to-end deep learning, as described in any of the foregoing embodiments.

[0224] Based on the same inventive concept, this invention provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, it implements the sparse altimeter sea surface height short-term forecasting method based on end-to-end deep learning of any of the foregoing embodiments. The computer-readable storage medium specifically includes hardware media configured to store computer program code, such as a universal serial bus flash drive, portable hard drive, read-only memory, random access memory, magnetic disk, or optical disk.

[0225] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A short-term sea surface height forecasting method based on sparse altimeter using deep learning, characterized in that, include: Collect observation data from multi-source nadir altimeters and perform standardization processing to generate a daily average standardized sea surface height anomaly observation dataset covering the target sea area; The daily average standardized sea surface height anomaly observation dataset is subjected to quality-aware preprocessing to identify and remove abnormal observations. Quality scores are assigned to each spatial grid point based on the difference in observation quality, and a quality-weighted observation tensor and a missing mask indicating the coverage status of the observation location are generated. The mass-weighted observation tensor and the missing mask are input into a pre-trained end-to-end spatiotemporal neural network model. The parameterized time stepping operator embedded in the end-to-end spatiotemporal neural network model is used to perform a mapping operation and output a seamless sea surface height anomaly forecast field within a specified future forecast period. Uncertainty estimation is performed on the seamless sea surface height anomaly forecast field using an integrated sampling strategy and model difference index. The forecast mean field and forecast variance field corresponding to each grid point are calculated and converted into spatial pixel-level forecast confidence distribution data. The target sea area is obtained in multiple dimensions of ocean dynamics characteristics. Based on these multiple dimensions of ocean dynamics characteristics, the target sea area is dynamically divided into different ocean dynamic regions. For each classified region, the corresponding post-processing parameter group is applied to perform spatial filtering and bias correction on the predicted mean field to generate a post-processed sea surface height anomaly field. The corresponding geostrophic velocity field is calculated based on the geostrophic balance relationship. The evolution characteristics of the post-processed sea surface height anomaly field relative to historical climatological data are extracted, a standardized anomaly index is calculated, and marine anomaly events are detected based on the comparison results between the standardized anomaly index and the preset three-level early warning threshold, generating regional early warning signals of different levels. The post-processed sea surface height anomaly field, the geostrophic velocity field, the forecast variance field, the forecast confidence distribution data, and the regional early warning signal are encapsulated into a forecast product file of a predetermined format, and the forecast product is distributed in real time through a network interface.

2. The method for short-term sea surface height forecasting based on sparse altimeter according to claim 1, characterized in that, The process of performing quality-aware preprocessing on the daily average standardized sea surface height anomaly observation dataset, identifying and removing anomalous observations, assigning quality scores to each spatial grid point based on observation quality differences, and generating a quality-weighted observation tensor and a missing mask indicating the coverage status of the observation location includes: Calculate the median and absolute deviation of the median for the time series corresponding to the spatial grid points. Mark the observation data that deviates from the median by more than a preset multiple of the absolute deviation of the median as abnormal observations and remove them. The observation quality score is calculated based on three indicators: signal-to-noise ratio, significant wave height, and backtracking distance deviation, after removing outlier observations. An observation quality weight matrix is ​​generated based on the observation quality scores of each grid point, and a binary missing mask consistent with the dimension of the observation data is established. The daily average standardized sea surface height anomaly observation dataset is weighted using the observation quality weight matrix, and the quality-weighted observation tensor and missing mask are output.

3. The method for short-term sea surface height forecasting based on sparse altimeter according to claim 1, characterized in that, The process of inputting the mass-weighted observation tensor and the missing mask into a pre-trained end-to-end spatiotemporal neural network model to output a seamless sea surface height anomaly prediction field within a specified future forecast period includes: Obtain the historical quality-weighted sparse observation sequence before the current forecast time and the corresponding binary missing mask, and concatenate the historical quality-weighted sparse observation sequence and the binary missing mask along the feature channel dimension to generate the input tensor of the end-to-end spatiotemporal neural network model; The end-to-end spatiotemporal neural network model receives the input tensor and performs forward inference to output a daily seamless sea surface height anomaly forecast field for the future forecast lead time days after the current forecast time.

4. The method for short-term sea surface height forecasting based on sparse altimeter according to claim 3, characterized in that, The end-to-end spatiotemporal neural network model includes a four-dimensional variational data assimilation neural network branch and an encoder-decoder neural network branch; The four-dimensional variational data assimilation neural network branch includes a learnable variational cost function, a gradient descent solver, and a state update network. By alternately executing the learnable prior operator and the cost gradient update, it outputs the first deterministic prediction result. The encoder-decoder neural network branch adopts a symmetrical encoder-decoder structure, and a channel attention module is provided at the jump connection between the encoder and the decoder. The second deterministic prediction result is output through multi-scale feature extraction and recovery.

5. The method for short-term sea surface height forecasting based on sparse altimeter according to claim 4, characterized in that, The loss function for pre-training the end-to-end spatiotemporal neural network model includes the state mean square error loss, the spatial gradient mean square error loss, and the dynamic regularization loss; pre-training the end-to-end spatiotemporal neural network model includes: After each backpropagation calculation of the end-to-end spatiotemporal neural network model, the gradient norms of the state mean square error loss, spatial gradient mean square error loss, and dynamic regularization loss with respect to the shared parameters are extracted respectively, and the moving average of the gradient norms is calculated. The weight coefficients corresponding to each loss are adaptively updated based on the maximum value among the gradient norm moving averages of the state mean square error loss, the spatial gradient mean square error loss, and the dynamic regularization loss.

6. The short-term sea surface height forecasting method based on deep learning using sparse altimeters according to claim 4, characterized in that, The uncertainty estimation of the seamless sea surface height anomaly forecast field is performed using an integrated sampling strategy and model difference index, and the forecast mean field and forecast variance field corresponding to each grid point are calculated, including: Based on the eigenvalue spectrum of the covariance matrix of the time series of sea surface height anomalies in each spatial region of the training set, the effective degrees of freedom corresponding to each spatial region are calculated. Based on the effective degrees of freedom corresponding to each spatial region and the maximum effective degrees of freedom in the global scope of the target sea area, the optimal random inactivation rate of the end-to-end spatiotemporal neural network model in the corresponding spatial region is determined; During the forward inference phase of the end-to-end spatiotemporal neural network model, multiple random sampling inferences are performed according to the optimal random inactivation rate. The sea surface height anomaly forecast results from multiple random sampling inferences are aggregated, and the mean and variance of the Monte Carlo random inactivation ensemble forecasts are calculated.

7. The method for short-term sea surface height forecasting based on sparse altimeter according to claim 6, characterized in that, The method of performing uncertainty estimation on the seamless sea surface height anomaly forecast field using an integrated sampling strategy and model difference index, calculating the forecast mean field and forecast variance field corresponding to each grid point, and converting it into spatial pixel-level forecast confidence distribution data, also includes: The squared difference between the first deterministic forecast result and the second deterministic forecast result is calculated to obtain the variance of inconsistency between models; The Monte Carlo random inactivation forecast variance and the inter-model inconsistency variance are weighted and fused to generate the total forecast variance field. Pixel-level reference variance is determined based on the climatological variance of sea surface height anomalies at the corresponding spatial grid points; The ratio of the total forecast variance field to the pixel-level reference variance is mapped using an exponential decay function to generate spatial pixel-level forecast confidence distribution data.

8. The method for short-term sea surface height forecasting based on sparse altimeter according to claim 1, characterized in that, The multidimensional ocean dynamics features include the historical time variability standard deviation of sea surface height anomalies, distance from the shore, seabed topographic gradient, and seasonal variation coefficient of eddy kinetic energy. The process involves acquiring multi-dimensional ocean dynamics characteristics of the target sea area, dynamically dividing the target sea area into different ocean dynamic regions based on these characteristics, and applying corresponding post-processing parameter sets to perform spatial filtering and bias correction on the predicted mean field for each classified region to generate a post-processed sea surface height anomaly field, including: Based on the historical time variability standard deviation of sea surface height anomalies and the distance from the shore, the target sea area is divided into the nearshore zone, the open ocean high-energy zone, and the open ocean low-energy zone. Extract the spatial Gaussian filter kernel radius corresponding to each ocean dynamic region, perform a two-dimensional convolution operation on the predicted mean field, and obtain the spatially filtered sea surface height anomaly field; Linear deviation correction is performed on the spatially filtered sea surface height anomaly field using predetermined deviation correction regression coefficients to generate a post-processed sea surface height anomaly field.

9. The method for short-term sea surface height forecasting based on sparse altimeter according to claim 8, characterized in that, The geostrophic velocity field calculated based on the geostrophic balance relationship includes: Based on the latitude of the spatial grid points, the target sea area is divided into equatorial and non-equatorial regions; In non-equatorial regions, the zonal and meridional velocities of the geostrophic flow are calculated using the post-processed sea surface height anomaly field, mean dynamic topography, gravitational acceleration, and Coriolis parameters. In the equatorial region, when the absolute value of the Coriolis parameter is less than the set Coriolis parameter threshold, the approximate geostrophic velocity in the equatorial region is calculated by using the equatorial Rossby parameter and the north-south distance from the equator. Integrate the calculation results from non-equatorial and equatorial regions to output the geostrophic velocity field.

10. The method for short-term sea surface height forecasting based on sparse altimeter according to claim 1, characterized in that, The process involves extracting and processing the evolution characteristics of the sea surface height anomaly field relative to historical climatological data, calculating a standardized anomaly index, and detecting marine anomaly events based on the comparison results between the standardized anomaly index and preset three-level warning thresholds. This generates regional warning signals of different levels, including: By combining the post-processed sea surface height anomaly field with the mean and standard deviation of the climatological sea surface height anomaly of the target spatial grid points on the corresponding dates, the standardized anomaly index of the forecast results relative to historical climatological data is calculated. The standardized anomaly index is compared with preset attention level thresholds, warning level thresholds, and alarm level thresholds to determine the corresponding early warning level. The partial derivatives of the standardized anomaly index in the spatial and temporal dimensions are calculated using the central difference scheme, and the spatiotemporal gradient anomaly index is aggregated to generate a spatiotemporal gradient anomaly index. When the spatiotemporal gradient anomaly index exceeds the set spatiotemporal gradient anomaly threshold, the corresponding spatial grid point is marked as a rapidly evolving anomaly region. A continuous abnormality early warning signal is generated for spatial areas that continuously reach the warning or alarm level and reach the duration threshold.

Images