An ai marine large model-based reanalysis data correction method and system
By using the reanalysis data correction method of AI ocean big data model, the high cost and physical field consistency problems of ocean reanalysis data are solved, and efficient and accurate data correction results are achieved.
Patent Information
- Application Number
- CN202610976443.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-02
- Publication Date
- 2026-08-25
AI Technical Summary
Existing ocean reanalysis data is computationally expensive and difficult to correct for nonlinear dynamic cumulative errors. Conventional deep learning predictions are prone to disrupting the consistency of the physical field, and global grid corrections are susceptible to land noise interference, resulting in insufficient data accuracy and consistency.
AI-powered ocean big data models are used for data reanalysis correction. Feature fusion and error correction are performed through resampling alignment, standardization, and deep learning correction models. The reliability of spatiotemporal features is quantified using ocean observation ground truth data, and error correction increments are generated and superimposed on the original data in the form of residuals.
It significantly improves the spatiotemporal accuracy and physical consistency of ocean reanalysis data, reduces computational costs and data dependence, and achieves efficient data correction.
Smart Images
Figure CN122634518A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of marine data processing and marine forecasting technology, specifically relating to a reanalysis data correction method and system based on an AI-based large marine model. Background Technology
[0002] Ocean reanalysis data is a global ocean state dataset that integrates multi-source observational data and numerical model results. While existing ocean reanalysis data can provide comprehensive ocean state information, it suffers from the following problems: First, it relies on traditional numerical model assimilation, resulting in high computational costs. Furthermore, due to its limited ability to fit complex spatiotemporal correlations, it struggles to capture nonlinear dynamic features such as mesoscale eddies in the ocean, failing to effectively correct for nonlinear dynamic cumulative errors and systemic biases. Second, in existing integrated reanalysis data correction schemes, using conventional deep learning models for end-to-end prediction can easily disrupt the flow field constraints of the original ocean physical field, leading to non-physical anomalous abrupt changes and a lack of physical consistency in the correction results. Third, due to the interweaving of global land and sea grids, existing correction methods fail to effectively isolate land topographic noise when processing global-scale grids, causing model attention to be scattered and making them highly susceptible to interference from invalid land data.
[0003] AI-powered ocean big data models have demonstrated superior performance in high-resolution ocean forecasting, efficiently capturing the spatiotemporal evolution of ocean variables and achieving computational efficiency more than a thousand times higher than traditional numerical models. However, current technologies do not fully utilize the forecasting capabilities of AI-powered ocean big data models for reanalysis data correction, and lack an integrated solution for efficiently fusing AI forecast results with reanalysis data. This makes it difficult to further improve the accuracy, physical consistency, and spatiotemporal consistency of the reanalysis data. Summary of the Invention
[0004] This invention addresses the problems in existing marine data correction methods, such as the high computational cost of traditional numerical models and the difficulty in correcting nonlinear dynamic cumulative errors, the tendency of conventional deep learning to directly predict and destroy the original physical field, resulting in a lack of physical consistency in the results, and the susceptibility of global grid correction to interference from invalid land data. It provides a reanalysis data correction method and system based on an AI-powered large-scale marine model, which solves the above problems.
[0005] The technical solution of the present invention is as follows:
[0006] This invention provides a reanalysis data correction method based on an AI-powered large-scale ocean model, comprising the following steps:
[0007] S1: Data Acquisition: Acquire AI ocean big model forecast data, ocean reanalysis data to be corrected, and ocean observation true data;
[0008] S2: Resampling Alignment: The ocean reanalysis data to be corrected is resampled to align it with the AI ocean big model prediction data in terms of spatiotemporal resolution and vertical depth stratification.
[0009] S3: Data preprocessing: The aligned AI ocean big model forecast data and the ocean reanalysis data to be corrected are standardized and subjected to land-sea masking to generate model forecast feature data and ocean reanalysis feature data to be corrected.
[0010] S4: Deep learning correction: Input the model-predicted feature data and the ocean reanalysis feature data to be corrected into a pre-trained deep learning correction model for processing, and output the corrected ocean reanalysis data;
[0011] The deep learning correction model includes a feature fusion layer, a feature extraction layer, an adaptive error correction layer, and an output layer connected in sequence.
[0012] In processing, the deep learning correction model concatenates and fuses the model's predicted feature data with the ocean reanalysis feature data to be corrected through the feature fusion layer to generate a fused feature tensor; the feature extraction layer extracts the spatiotemporal correlation information of the fused feature tensor to generate spatiotemporal features; the adaptive error correction layer quantifies the prediction reliability of the spatiotemporal features based on a set time window and the ocean observation ground truth data, and adaptively generates time attention weights to perform weighted fusion and mapping of the spatiotemporal features to generate error correction increments; finally, the output layer upsamples the error correction increments and superimposes them on the ocean reanalysis data to be corrected in the form of residuals to output the corrected ocean reanalysis data.
[0013] Preferably, the standardization and land-sea masking processing of the aligned AI ocean big model forecast data and the ocean reanalysis data to be corrected, as described in S3, to generate model forecast feature data and ocean reanalysis feature data to be corrected, specifically includes:
[0014] The aligned AI ocean big model forecast data and the ocean reanalysis data to be corrected were processed using the Z-score normalization method to eliminate the dimensional differences of different ocean elements.
[0015] A preset land-sea mask is applied to the standardized data, retaining only the valid data of the ocean grid points and removing the invalid values of the land grid points, thereby generating the model prediction feature data and the ocean reanalysis feature data to be corrected.
[0016] Preferably, the step S4, which involves splicing and fusing the model-predicted feature data with the ocean reanalysis feature data to be corrected through the feature fusion layer to generate a fused feature tensor, specifically includes:
[0017] The model-predicted feature data and the ocean reanalysis feature data to be corrected are concatenated along the channel dimension.
[0018] Global layer normalization is applied to the spliced data to generate the fused feature tensor.
[0019] Preferably, the step S4, which involves extracting the spatiotemporal correlation information of the fused feature tensor through the feature extraction layer to generate spatiotemporal features, specifically includes:
[0020] The feature extraction layer includes three alternating convolutional blocks and two max pooling layers;
[0021] The fused feature tensor is sequentially processed by convolution through the convolution block and downsampled through the max pooling layer to gradually extract multi-scale spatial correlation information and reduce the dimension of the feature tensor, and output the spatiotemporal features.
[0022] Preferably, the quantification of the spatiotemporal characteristics based on the set time window and the true ocean observation data in step S4 specifically includes:
[0023] The spatiotemporal features are divided into multiple consecutive time windows according to a set time step to obtain the temporal features corresponding to each time window.
[0024] For each time window, the spatial mean absolute error between the temporal characteristics of that time window and the true ocean observation data is calculated. The forecast reliability of each time window is quantified by the spatial mean absolute error. The smaller the spatial mean absolute error, the higher the forecast reliability.
[0025] Preferably, the adaptive generation of temporal attention weights, and the weighted fusion and mapping of the spatiotemporal features to generate error correction increments, specifically include:
[0026] The spatial mean absolute error of each time window is normalized using the Softmax function to generate the temporal attention weights in the channel dimension.
[0027] Based on the time attention weights, the temporal features of each time window are summed using channel weights to obtain fused temporal features;
[0028] The fused temporal features are sequentially input into two fully connected layers, and nonlinear mapping is performed using the Sigmoid activation function to generate the error correction increment whose value is constrained to the [0,1] interval.
[0029] Preferably, the step S4, which involves upsampling the error correction increment through the output layer and then superimposing it as a residual onto the ocean reanalysis data to be corrected, to output the corrected ocean reanalysis data, specifically includes:
[0030] The error correction increment is upsampled by a deconvolution layer to restore its spatial resolution to be consistent with the ocean reanalysis data to be corrected, thus obtaining the upsampled error increment.
[0031] The upsampling error increment is added to the ocean reanalysis data to be corrected in the form of residuals to obtain preliminary correction results;
[0032] The preliminary correction results are fine-tuned using a 1×1 convolutional layer to output the corrected ocean reanalysis data, ensuring the physical consistency of the corrected data.
[0033] Preferably, the training process of the pre-trained deep learning correction model described in S4 includes:
[0034] The deep learning correction model is trained end-to-end using dimensional weighted Charbonnier loss as the training loss function.
[0035] The formula for calculating the training loss function is as follows:
[0036]
[0037] in, The weighting factor at latitude Φi Y represents the corrected ocean reanalysis data output by the deep learning correction model, C represents the number of ocean element channels, H and W represent the number of spatial grid points of the data, and ϵ represents the minimum constant that ensures the differentiability of the function.
[0038] Preferably, the training process further includes:
[0039] The AdamW optimizer is used to update the model parameters;
[0040] A cosine annealing learning rate scheduling strategy is adopted to control the learning rate decay during the training process, so as to balance the convergence speed in the early stage of model training with the stability in the later stage.
[0041] This invention also provides a reanalysis data correction system based on an AI-powered large ocean model, comprising:
[0042] The data acquisition module is used to acquire AI ocean big model forecast data, ocean reanalysis data to be corrected, and ocean observation true data;
[0043] The resampling alignment module is used to resample the ocean reanalysis data to be corrected, so that it is aligned with the AI ocean big model prediction data in terms of spatiotemporal resolution and vertical depth stratification.
[0044] The data preprocessing module is used to perform standardization and land-sea masking on the aligned AI ocean big model forecast data and the ocean reanalysis data to be corrected, and to generate model forecast feature data and ocean reanalysis feature data to be corrected.
[0045] The deep learning correction module is used to input the model prediction feature data and the ocean reanalysis feature data to be corrected into a pre-trained deep learning correction model for processing, and output the corrected ocean reanalysis data.
[0046] The deep learning correction model includes a feature fusion submodule, a feature extraction submodule, an adaptive error correction submodule, and an output submodule connected in sequence.
[0047] The feature fusion submodule is used to splice and fuse the model-predicted feature data with the ocean reanalysis feature data to be corrected to generate a fused feature tensor.
[0048] The feature extraction submodule is used to extract the spatiotemporal correlation information of the fused feature tensor and generate spatiotemporal features;
[0049] The adaptive error correction submodule is used to quantify the forecast reliability of the spatiotemporal features based on a set time window and the true ocean observation data, adaptively generate time attention weights, and perform weighted fusion and mapping on the spatiotemporal features to generate error correction increments.
[0050] The output submodule is used to upsample the error correction increment and then superimpose it onto the ocean reanalysis data to be corrected in the form of residuals, and output the corrected ocean reanalysis data.
[0051] The beneficial effects of this invention are as follows:
[0052] This invention provides a reanalysis data correction method and system based on AI ocean big model, which breaks through the limitations of traditional statistical correction and numerical model correction, and overcomes the inherent defects in the field of traditional ocean data correction.
[0053] First, this invention uses ground truth ocean observation data (S1) as a highly reliable supervisory benchmark. Then, in the deep learning correction forward inference (S4), the adaptive error correction layer of the deep learning correction model efficiently quantifies the forecast reliability of spatiotemporal features based on a set time window and the ground truth ocean observation data. This adaptively generates time attention weights to weightedly fuse and map the spatiotemporal features, generating error correction increments. This mechanism completely solves the problems of high computational costs in traditional numerical models and their limited ability to fit complex spatiotemporal correlations, making it difficult to capture nonlinear dynamic features such as ocean mesoscale eddies and western boundary currents, and unable to effectively correct nonlinear dynamic cumulative errors and system biases. The intelligent feature extraction of this invention enables the model to accurately capture complex spatiotemporal evolution patterns, significantly improving the spatiotemporal accuracy of ocean reanalysis data at sub-daily and eddy-resolved scales, with an average latitudinal weighted RMSE reduction of over 30% for core ocean elements.
[0054] Secondly, after generating the error correction increment in the aforementioned steps, in S4, the upsampled error correction increment is superimposed as a residual onto the original background field described in S1, i.e., the ocean reanalysis data to be corrected, through the output layer of the deep learning correction model, and the final corrected ocean reanalysis data is output. Preferably, in some embodiments of the present invention, a 1×1 convolutional layer in the output layer is used to fine-tune the preliminary correction results across channels. This mechanism perfectly solves the problem that conventional deep learning models, when directly performing end-to-end predictions, are prone to disrupting the flow field constraints of the original ocean physical field and generating non-physical anomalous abrupt changes. Using residual correction as a bottom-line constraint not only maintains the large-scale structural balance of the original ocean physical field but also ensures the stability and authenticity of the corrected data, meeting the stringent requirements of physical boundary constraints in actual marine meteorological forecasting.
[0055] Furthermore, this invention relies solely on core ocean element data as input in step S1, eliminating the need for additional complex atmospheric forcing data and significantly reducing the system's data dependency. In step S2, spatiotemporal resampling and alignment are performed, and in step S3, rigorous standardization and land-sea masking are introduced to accurately eliminate land topographic noise interference, enabling the model to fully focus on the features of pure seawater grid points. In step S4, secondary correction development is directly based on high-precision forecast data provided by the already trained AI ocean model, without needing to retrain the AI ocean model itself with billions of parameters. This lightweight, integrated mechanism fully leverages the advantages of data-driven intelligent solutions while completely shielding coastal topographic noise interference. After model training, rapid batch correction of ocean reanalysis data can be achieved, with a single correction inference taking only a few seconds, far exceeding the correction efficiency of traditional numerical models, significantly reducing the computational and time costs of model development.
[0056] In summary, this invention achieves refined and intelligent correction of ocean reanalysis data. The corrected, high-quality ocean reanalysis data can be directly used as the initial field for high-precision ocean forecasting, exhibiting extremely high engineering practicality, scalability, and broad commercial application prospects. Attached Figure Description
[0057] Figure 1 The flowchart illustrates a reanalysis data correction method based on an AI-powered large ocean model, as provided in this invention.
[0058] Figure 2 This is a schematic diagram of the deep learning correction model structure in this invention. Detailed Implementation
[0059] To make the objectives, technical solutions, and advantages of this invention clearer, the following detailed description of a reanalysis data correction method and system based on an AI-powered large-scale ocean model, in conjunction with the accompanying drawings and specific embodiments, will further illustrate the present invention. The advantages and features of this invention will become clearer from the following description. It should be noted that the accompanying drawings are in a very simplified form and use non-precise proportions, used only to facilitate and clearly illustrate the purpose of the embodiments of this invention. Please refer to the accompanying drawings for a clearer understanding of the objectives, features, and advantages of this invention. It should be understood that the structures, proportions, sizes, etc., depicted in the accompanying drawings are only for illustrative purposes and to enable those skilled in the art to understand and read the content disclosed in the specification. They are not intended to limit the implementation conditions of this invention and therefore have no substantial technical significance. Any modifications to the structure, changes in proportions, or adjustments to the size, without affecting the effects and objectives achieved by this invention, should still fall within the scope of the technical content disclosed in this invention.
[0060] The flowchart of a reanalysis data correction method based on an AI-powered large ocean model provided by this invention is as follows: Figure 1 As shown, the specific steps include:
[0061] S1: Data Acquisition: Acquire AI ocean big model forecast data, ocean reanalysis data to be corrected, and ocean observation true data.
[0062] In some embodiments of the present invention, the forecast data has a spatial resolution of 1 / 12° and a temporal resolution of 6 hours, covering a vertical depth of 0-1500 meters and divided into 20 layers (0m, 2m, 4m, 6m, 10m, 20m, 30m, 40m, 50m, 60m, 70m, 80m, 100m, 125m, 150m, 200m, 300m, 500m, 1000m, 1500m). The core ocean elements include temperature (T), salinity (S), zonal current (U), meridional current (V), and sea surface height (SSH). HYCOM ocean reanalysis data is used as the data to be corrected, and ground truth ocean observation data fused from Argo buoys and satellite altimeters is used as the supervision label, with the element types of the three being consistent. A summary of the AI ocean large model forecast data variables is shown in Table 1.
[0063] Table 1
[0064]
[0065] S2: Resampling Alignment: The ocean reanalysis data to be corrected is resampled to align it with the AI ocean big model forecast data in terms of spatiotemporal resolution and vertical depth stratification.
[0066] In some embodiments of the present invention, the alignment includes: fully aligning its spatial resolution, temporal resolution, and vertical depth layering with the AI ocean big model forecast data.
[0067] The nearest neighbor interpolation method was used to resample the HYCOM ocean reanalysis data to make its spatial grid number, time step, and vertical depth stratification completely match the AI ocean big model forecast data, thus achieving full spatiotemporal alignment of the two types of data.
[0068] S3: Data Preprocessing: The aligned AI ocean big model forecast data and the ocean reanalysis data to be corrected are standardized and subjected to land-sea masking to generate model forecast feature data and ocean reanalysis feature data to be corrected.
[0069] In some embodiments of the present invention, the aligned AI ocean large model forecast data and the ocean reanalysis data to be corrected are respectively subjected to Z-score standardization, with the following formula: Where X represents the original data, μ represents the data mean, σ represents the data standard deviation, and X′ represents the standardized data to eliminate dimensional differences among different ocean elements. Standardized model forecast data and ocean reanalysis data to be corrected are obtained, with ocean observation ground truth data serving as supervisory labels for model training and validation.
[0070] A land-sea mask is applied to the standardized model forecast data and the ocean reanalysis data to be corrected, retaining only the valid data of the ocean grid points and removing the invalid values of the land grid points.
[0071] S4: Deep learning correction: Input the model-predicted feature data and the ocean reanalysis feature data to be corrected into a pre-trained deep learning correction model for processing, and output the corrected ocean reanalysis data;
[0072] The deep learning correction model includes a feature fusion layer, a feature extraction layer, an adaptive error correction layer, and an output layer connected in sequence.
[0073] In processing, the deep learning correction model concatenates and fuses the model's predicted feature data with the ocean reanalysis feature data to be corrected through the feature fusion layer to generate a fused feature tensor; the feature extraction layer extracts the spatiotemporal correlation information of the fused feature tensor to generate spatiotemporal features; the adaptive error correction layer quantifies the prediction reliability of the spatiotemporal features based on a set time window and the ocean observation ground truth data, and adaptively generates time attention weights to perform weighted fusion and mapping of the spatiotemporal features to generate error correction increments; finally, the output layer upsamples the error correction increments and superimposes them on the ocean reanalysis data to be corrected in the form of residuals to output the corrected ocean reanalysis data.
[0074] In some embodiments of the present invention, the specific structure of the deep learning correction model based on hybrid temporal attention is as follows: Figure 2 As shown, the details are as follows:
[0075] The model is an end-to-end deep learning network. The input is a fused feature tensor, and the output is the corrected ocean reanalysis data. The network has a four-level structure: feature fusion layer, feature extraction layer, adaptive error correction layer, and output layer.
[0076] Feature Fusion Layer: The core task of this layer is to integrate the feature information of AI ocean big model forecast data and ocean reanalysis data to be corrected, so as to provide complete input for subsequent feature extraction and error correction.
[0077] Let the AI ocean big model forecast data be... The ocean reanalysis data to be corrected is Where C = 81 (corresponding to 4 vertical variables × 20 depth layers + 1 sea surface height variable), H = 2160, and W = 4320 are the global grid dimensions corresponding to 1 / 12° spatial resolution.
[0078] The feature fusion layer concatenates Xmodel and Xreana along the channel dimension to form a fused feature tensor. The splicing formula is:
[0079]
[0080] Where Concat(·) represents the concatenation operation along the channel dimensions. After concatenation, global layer normalization is applied to Xfusion. The normalization formula is:
[0081]
[0082] In the formula, μ is the mean of Xfusion in the channel dimension, and σ 2 The variance is ϵ=10−6, which is a very small constant to prevent the denominator from being zero. Normalization improves the stability and convergence speed of model training.
[0083] Feature Extraction Layer: This layer consists of alternating convolutional blocks and max-pooling layers, used to progressively extract multi-scale spatial correlation information from fused features, while reducing feature dimensionality to improve computational efficiency. The parameters and processing flow of each component are as follows:
[0084] The first convolutional block contains one 3×3 convolutional layer (64 kernels, stride 1, padding 1) and one ReLU activation function. Perform shallow feature extraction and output feature tensor The convolution formula is:
[0085]
[0086] Where W1 and b1 are the kernel weights and bias terms, respectively, and Conv3d(⋅) represents the three-dimensional convolution operation.
[0087] First max-pooling layer: Uses a 2×2 pooling kernel (stride 2, no fill), downsamples F1, and outputs... Pooling preserves key features and reduces computational cost.
[0088] The second convolutional block has the same structure as the first convolutional block, but the number of kernels is increased to 128. It performs deep feature extraction on the F1pool and outputs... .
[0089] Second max-pooling layer: uses a 2×2 pooling kernel (step size 2, no fill), output .
[0090] The third convolutional block maintains a kernel count of 128 to further enhance feature extraction capabilities, outputting the final extracted feature tensor, i.e., the spatiotemporal features. , as input to the adaptive error correction layer;
[0091] Adaptive Error Correction Layer: This is the core layer of the model. A Mixed-Time (MoT) attention module is introduced, which adaptively generates error correction increments by learning the feature reliability weights for different time windows. The specific process and parameters are as follows:
[0092] Time window partitioning: The historical time series data corresponding to the input feature Fext is divided into 4 consecutive time windows (each window corresponds to 6 hours, totaling 24 hours), resulting in a time series feature set. Where t is the current time, t −i (i=0,1,2,3) represents historical moments.
[0093] Reliability Quantification: For the characteristics of each time window, calculate its spatial average MAE with the true ocean observation Y to quantify the forecast reliability of that window. The formula is:
[0094]
[0095] In the formula, MAEi is the spatial average MAE value of the i-th time window. The smaller the value, the higher the reliability of the window features.
[0096] Attention weight generation: The MAEi values of the four windows are normalized using the softmax function to generate an attention weight matrix along the channel dimension. The formula is:
[0097]
[0098] In the formula, c is the feature channel index, exp (−MAEi(c)) The MAE value is mapped to a weight through a negative exponential transformation (the higher the reliability, the greater the weight).
[0099] Weighted feature fusion: Based on the weight matrix V, the features of the four time windows are summed by channel weights to obtain the fused time series features. The formula is
[0100]
[0101] Error increment generation: Ftime is mapped to error correction increments using two fully connected layers (256 hidden layers and 128 output layers) and a sigmoid activation function. The formula is:
[0102]
[0103] FC1 and FC2 are the first and second fully connected layers, respectively. The Sigmoid function constrains the increment value to the range [0,1] to ensure the stability of the correction.
[0104] Output layer: This layer is responsible for upsampling the error correction increment to the original resolution and fusing it with the ocean reanalysis data to be corrected, outputting the final correction result. The specific process is as follows:
[0105] First deconvolutional layer: Uses a 3×3 deconvolution kernel (64 kernels, stride 2, padding 1), upsamples ΔF, and outputs... The formula for deconvolution is:
[0106]
[0107] Where Wt1 and bt1 are the deconvolution kernel weights and bias terms.
[0108] The second deconvolutional layer uses a 3×3 deconvolution kernel (number C, stride 2, padding 1) to further upsample to the original spatial resolution, outputting the error increment. .
[0109] Residual Fusion and Output: ΔFfinal is superimposed as residuals onto the ocean reanalysis data Xreana to be corrected, yielding the preliminary correction result Xtemp. Then, a 1×1 convolutional layer (number of kernels C, stride 1) is used for feature fine-tuning, outputting the final corrected ocean reanalysis data. The formula is:
[0110]
[0111]
[0112] Wfinal and bfinal are the weights and biases of the 1×1 convolutional layer. The residual connections ensure the physical consistency of the corrected data and avoid non-physical outliers.
[0113] In some embodiments of the present invention, the training method of the model is as follows:
[0114] AI ocean big data model forecasts, HYCOM reanalysis data, and corresponding ocean observation ground truth data from 2006 to 2014 were selected as the training set, and data from 2015 were selected as the validation set.
[0115] The AdamW optimizer was used for end-to-end training of the model. The optimization objective was to minimize the training loss function. The optimizer parameters were set as follows: β1 = 0.9 (first-order momentum decay coefficient), β2 = 0.95 (second-order momentum decay coefficient), and weight decay coefficient λ = 10. −5(To prevent model overfitting).
[0116] The learning rate uses a cosine annealing scheduling strategy, and the formula for how the learning rate changes with the number of iterations is:
[0117]
[0118] In the formula, lrmax = 2.5 × 10 −4 (Initial maximum learning rate), lrmin=10 −8 (Minimum learning rate), iter is the current iteration number, itermax=60000 (total iterations). Cosine annealing is used to achieve a smooth decay of the learning rate, balancing the convergence speed in the early stage of model training with the stability in the later stage.
[0119] Regarding the loss function, the latitude-weighted Charbonnier loss was chosen as the training loss function. This function is robust to outliers in ocean data and adapts to the characteristics of the global ocean grid through the latitude weighting factor. The training set data was input into the correction model for multiple rounds of iterative training. After each round of training, the accuracy of the model was verified using the validation set. When the latitude-weighted RMSE of the validation set converged to a stable value, training was stopped, and the completed correction model was obtained.
[0120] In some embodiments of the present invention, the method for evaluating the effectiveness of the reanalysis data correction method and system based on AI ocean big model is as follows:
[0121] Latitude-weighted RMSE is used as the core evaluation indicator, while mean absolute error (MAE) and correlation coefficient (CC) are calculated as auxiliary indicators.
[0122]
[0123] In the formula, t0 is the initial time of assessment, τ is the forecast time step, D is the assessment dataset, c is the ocean element channel (temperature, salinity, zonal current, meridional current, sea surface height), and αi is the latitude weighting factor. This index comprehensively reflects the overall error level of the data.
[0124]
[0125] This indicator is not sensitive to outliers and provides a supplementary reflection of the average error level of the data.
[0126]
[0127] In the formula, Cov(·) is the covariance and Var(·) is the variance. This index reflects the degree of linear correlation between the corrected data and the true value. The closer the value is to 1, the better the consistency.
[0128] The corrected ocean reanalysis data from 2016 to 2020 were compared with the true observation data fused from Argo buoys and satellite altimeters. The latitude-weighted RMSE of temperature, salinity, zonal current, meridional current, and sea surface height were calculated at 6-hour, 24-hour, 120-hour, and 240-hour time scales, as well as at vertical depths of 0m, 100m, 500m, and 1500m. At the same time, the latitude-weighted RMSE of the original HYCOM reanalysis data at the same spatiotemporal scales was calculated and compared.
[0129] In some embodiments of the present invention, the reanalysis data correction method and system based on AI ocean big model described in the present invention can reduce the average latitude-weighted RMSE of the core element of ocean reanalysis data; the physical consistency of the corrected reanalysis data is significantly better than that of the original reanalysis data, and it can effectively capture the spatiotemporal evolution characteristics of the ocean at the sub-daily scale.
[0130] It will be understood by those skilled in the art that the above descriptions are merely preferred examples of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the invention should be included within the scope of protection of the invention. All technical features in this embodiment can be freely combined according to actual needs.
Claims
1. A reanalysis data correction method based on an AI-powered large-scale ocean model, characterized in that, Includes the following steps: S1: Data Acquisition: Acquire AI ocean big model forecast data, ocean reanalysis data to be corrected, and ocean observation true data; S2: Resampling Alignment: The ocean reanalysis data to be corrected is resampled to align it with the AI ocean big model prediction data in terms of spatiotemporal resolution and vertical depth stratification. S3: Data preprocessing: The aligned AI ocean big model forecast data and the ocean reanalysis data to be corrected are standardized and subjected to land-sea masking to generate model forecast feature data and ocean reanalysis feature data to be corrected. S4: Deep learning correction: Input the model-predicted feature data and the ocean reanalysis feature data to be corrected into a pre-trained deep learning correction model for processing, and output the corrected ocean reanalysis data; The deep learning correction model includes a feature fusion layer, a feature extraction layer, an adaptive error correction layer, and an output layer connected in sequence. In processing, the deep learning correction model concatenates and fuses the model's predicted feature data with the ocean reanalysis feature data to be corrected through the feature fusion layer to generate a fused feature tensor; the feature extraction layer extracts the spatiotemporal correlation information of the fused feature tensor to generate spatiotemporal features; the adaptive error correction layer quantifies the prediction reliability of the spatiotemporal features based on a set time window and the ocean observation ground truth data, and adaptively generates time attention weights to perform weighted fusion and mapping of the spatiotemporal features to generate error correction increments; finally, the output layer upsamples the error correction increments and superimposes them on the ocean reanalysis data to be corrected in the form of residuals to output the corrected ocean reanalysis data.
2. The reanalysis data correction method based on an AI-powered large-scale ocean model according to claim 1, characterized in that, S3 describes the standardization and land-sea masking of the aligned AI ocean big model forecast data and the ocean reanalysis data to be corrected, generating model forecast feature data and ocean reanalysis feature data to be corrected. Specifically, this includes: The aligned AI ocean big model forecast data and the ocean reanalysis data to be corrected were processed using the Z-score normalization method to eliminate the dimensional differences of different ocean elements. A preset land-sea mask is applied to the standardized data, retaining only the valid data of the ocean grid points and removing the invalid values of the land grid points, thereby generating the model prediction feature data and the ocean reanalysis feature data to be corrected.
3. The reanalysis data correction method based on an AI-powered large-scale ocean model according to claim 1, characterized in that, The process described in S4, which involves concatenating and fusing the model-predicted feature data with the ocean reanalysis feature data to be corrected through the feature fusion layer to generate a fused feature tensor, specifically includes: The model-predicted feature data and the ocean reanalysis feature data to be corrected are concatenated along the channel dimension. Global layer normalization is applied to the spliced data to generate the fused feature tensor.
4. The reanalysis data correction method based on an AI-powered large-scale ocean model according to claim 1, characterized in that, The step S4, which describes extracting the spatiotemporal correlation information of the fused feature tensor through the feature extraction layer to generate spatiotemporal features, specifically includes: The feature extraction layer includes three alternating convolutional blocks and two max pooling layers; The fused feature tensor is sequentially processed by convolution through the convolution block and downsampled through the max pooling layer to gradually extract multi-scale spatial correlation information and reduce the dimension of the feature tensor, and output the spatiotemporal features.
5. The reanalysis data correction method based on an AI-powered large-scale ocean model according to claim 1, characterized in that, The forecast reliability of the spatiotemporal characteristics quantified based on the set time window and the true ocean observation data as described in S4 specifically includes: The spatiotemporal features are divided into multiple consecutive time windows according to a set time step to obtain the temporal features corresponding to each time window. For each time window, the spatial mean absolute error between the temporal characteristics of that time window and the true ocean observation data is calculated. The forecast reliability of each time window is quantified by the spatial mean absolute error. The smaller the spatial mean absolute error, the higher the forecast reliability.
6. The reanalysis data correction method based on an AI-powered large-scale ocean model according to claim 5, characterized in that, The adaptive generation of temporal attention weights, and the weighted fusion and mapping of the spatiotemporal features to generate error correction increments, specifically include: The spatial mean absolute error of each time window is normalized using the Softmax function to generate the temporal attention weights in the channel dimension. Based on the time attention weights, the temporal features of each time window are summed using channel weights to obtain fused temporal features; The fused temporal features are sequentially input into two fully connected layers, and nonlinear mapping is performed using the Sigmoid activation function to generate the error correction increment whose value is constrained to the [0,1] interval.
7. The reanalysis data correction method based on an AI-powered large ocean model according to claim 1, characterized in that, S4 describes upsampling the error correction increment through the output layer, then superimposing it as a residual onto the ocean reanalysis data to be corrected, and outputting the corrected ocean reanalysis data. Specifically, this includes: The error correction increment is upsampled by a deconvolution layer to restore its spatial resolution to be consistent with the ocean reanalysis data to be corrected, thus obtaining the upsampled error increment. The upsampling error increment is added to the ocean reanalysis data to be corrected in the form of residuals to obtain preliminary correction results; The preliminary correction results are fine-tuned using a 1×1 convolutional layer to output the corrected ocean reanalysis data, ensuring the physical consistency of the corrected data.
8. The reanalysis data correction method based on an AI-powered large ocean model according to claim 1, characterized in that, The training process of the pre-trained deep learning correction model described in S4 includes: The deep learning correction model is trained end-to-end using dimensional weighted Charbonnier loss as the training loss function. The formula for calculating the training loss function is as follows: in, The weighting factor at latitude Φi Y represents the corrected ocean reanalysis data output by the deep learning correction model, C represents the number of ocean element channels, H and W represent the number of spatial grid points of the data, and ϵ represents the minimum constant that ensures the differentiability of the function.
9. The reanalysis data correction method based on an AI-powered large-scale ocean model according to claim 8, characterized in that, The training process also includes: The AdamW optimizer is used to update the model parameters; A cosine annealing learning rate scheduling strategy is adopted to control the learning rate decay during the training process, so as to balance the convergence speed in the early stage of model training with the stability in the later stage.
10. A reanalysis data correction system based on an AI-powered large-scale ocean model, characterized in that, include: The data acquisition module is used to acquire AI ocean big model forecast data, ocean reanalysis data to be corrected, and ocean observation true data; The resampling alignment module is used to resample the ocean reanalysis data to be corrected, so that it is aligned with the AI ocean big model prediction data in terms of spatiotemporal resolution and vertical depth stratification. The data preprocessing module is used to perform standardization and land-sea masking on the aligned AI ocean big model forecast data and the ocean reanalysis data to be corrected, and to generate model forecast feature data and ocean reanalysis feature data to be corrected. The deep learning correction module is used to input the model prediction feature data and the ocean reanalysis feature data to be corrected into a pre-trained deep learning correction model for processing, and output the corrected ocean reanalysis data. The deep learning correction model includes a feature fusion submodule, a feature extraction submodule, an adaptive error correction submodule, and an output submodule connected in sequence. The feature fusion submodule is used to splice and fuse the model-predicted feature data with the ocean reanalysis feature data to be corrected to generate a fused feature tensor. The feature extraction submodule is used to extract the spatiotemporal correlation information of the fused feature tensor and generate spatiotemporal features; The adaptive error correction submodule is used to quantify the forecast reliability of the spatiotemporal features based on a set time window and the true ocean observation data, adaptively generate time attention weights, and perform weighted fusion and mapping on the spatiotemporal features to generate error correction increments. The output submodule is used to upsample the error correction increment and then superimpose it onto the ocean reanalysis data to be corrected in the form of residuals, and output the corrected ocean reanalysis data.