Sea surface height numerical simulation data correction method based on deep learning

Abstract

The invention provides a sea surface height numerical simulation data correction method based on deep learning, and relates to the technical field of data processing, and the method specifically comprises the steps: preparing data, and processing the data; constructing simulation satellite data based on the numerical simulation data; constructing a deep learning assimilation model; the simulation data, the background field data and the auxiliary physical quantity serve as input, numerical simulation data at a specified moment serve as a label, a deep learning assimilation model is trained, and a model capable of outputting a corrected high-resolution SSHA field is obtained; and putting a data set made of real high-resolution wide-swath satellite observation data, a low-resolution day-by-day satellite data product, numerical simulation data and auxiliary data into the trained model, and outputting a corrected sea surface height abnormal data product. According to the technical scheme, the problem that in the prior art, high-resolution satellite observation and low-resolution satellite products with good continuity cannot be fully fused is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data processing, in particular to a sea surface height numerical simulation data correction method based on deep learning. BACKGROUND

[0002] Sea surface height anomaly (SSHA) is an important physical quantity for describing the dynamic structure and thermal process of the ocean. The main ways to obtain SSHA include ocean numerical simulation and satellite remote sensing observation. High-resolution ocean models can provide continuous and spatially consistent SSHA simulation data, making up for the shortcomings of satellite observations, such as orbital limitations, data discontinuity, and limited spatial coverage. However, numerical models inevitably have error sources, are highly dependent on initial fields and boundary conditions, and have some simplifications and idealizations in the expression of physical processes. Therefore, systematic biases often occur in dynamic active regions such as boundary currents and tropical circulations, affecting the identification of key ocean phenomena such as mesoscale eddies and thermal content distribution. In addition, the sensitivity of numerical models to forcing terms such as wind stress and heat flux can further amplify errors, limiting their reliability in high-precision ocean prediction and intelligent business applications.

[0003] To reduce model errors, the common method is to introduce multi-source observation data into numerical models for error correction and state reconstruction. The current mainstream technology route is the data assimilation method based on physical mechanisms, such as three-dimensional variation (3D-Var), four-dimensional variation (4D-Var), and ensemble Kalman filter (EnKF). These methods incorporate observation data into the simulation system through the construction of error covariance structure between observation and model state, thereby improving the accuracy and reliability of model results. Although these methods have been successfully applied to atmospheric and oceanic prediction systems and deployed in some business systems, there are still significant limitations. First, physical assimilation algorithms require extremely high computing resources, usually relying on high-performance computing platforms, making it difficult to achieve rapid deployment and real-time operation. Second, these methods are highly dependent on the distribution of observation data, making it difficult to achieve stable assimilation in sparsely observed or discontinuous regions, especially in high-resolution ocean scenarios, where the assimilation effect will decrease significantly with spatial scale refinement. In addition, the system construction of traditional assimilation technology is complex, requiring fine tuning for different observation types, resolutions, and regional characteristics, resulting in poor generality and migration.

[0004] Although assimilation models based on deep learning have made good progress in the field of meteorology, they are still in the initial exploration stage in the field of ocean. There are the following challenges in the SSHA correction problem: firstly, the nonlinear, multiscale coupling of ocean dynamic processes leads to complex error propagation mechanism, and the model is difficult to fully capture; secondly, the limitations of satellite observations in temporal and spatial coverage and quality control restrict the construction of high-precision training samples; thirdly, most of the current assimilation models still rely on observations and historical simulation results for training, and lack of sufficient modeling of physical consistency and causal mechanism, which still has deficiencies in generalization ability, credibility and interpretability.

[0005] Therefore, there is a need for a deep learning-based sea surface height numerical simulation data correction method that fully integrates high-resolution satellite observations and low-resolution satellite products with good continuity to effectively correct high-resolution numerical simulation SSHA, thereby improving the reliability and applicability of ocean models in high-resolution and high-precision application scenarios. SUMMARY

[0006] The main purpose of the present application is to provide a deep learning-based sea surface height numerical simulation data correction method to solve the problem that the prior art cannot fully integrate high-resolution satellite observations and low-resolution satellite products with good continuity.

[0007] To achieve the above-mentioned purpose, the present application provides a deep learning-based sea surface height numerical simulation data correction method, which specifically comprises the following steps: S1, preparing data includes: high-resolution wide swath data, low-resolution daily satellite data product, numerical simulation data and auxiliary data, and processing the data.

[0008] S2, constructing simulated satellite data based on numerical simulation data, including: high-resolution wide swath satellite simulation data and low-resolution daily satellite product simulation data.

[0009] S3, constructing a deep learning assimilation model, including an encoder, a processor and a decoder connected to each other.

[0010] S4, taking the simulation data, background field data and auxiliary physical quantities as input, and taking the numerical simulation data at a specified time as label, training the deep learning assimilation model to obtain a model capable of outputting corrected high-resolution SSHA field.

[0011] S5, putting the real high-resolution wide swath satellite observation data, low-resolution daily satellite data product, numerical simulation data and auxiliary data into the model trained in step S4, and outputting the corrected sea surface height anomaly data product.

[0012] Further, step S1 specifically comprises the following steps: S1.1, spatio-temporal matching of different source data, and unifying the spatial grid of all data to the high-resolution grid of numerical simulation data through spatial interpolation.

[0013] S1.2, outlier rejection and consistency processing of data.

[0014] Further, step S2 specifically comprises the following steps: S2.1, constructing high-resolution wide-swath satellite simulation data: selecting numerical simulation SSHA data with the same latitude and longitude range as the high-resolution wide-swath observation data as high-resolution wide-swath satellite simulation data, and simulating actual observation errors by adding noise.

[0015] S2.2, constructing low-resolution daily satellite product simulation data: interpolating numerical simulation SSHA data to the grid of low-resolution daily satellite data products, and then re-interpolating the interpolated data to the numerical simulation grid, introducing blur processing to approximate the low-resolution observation characteristics, and adding noise to simulate observation uncertainty.

[0016] Further, step S3 specifically comprises the following steps: S3.1, processing non-gridded observations and data missing through the SetConv layer of the encoder, as shown in formula (1): (1) ; wherein, represents the estimated value in the grid coordinate, represents the learnable kernel function, represents the target grid point, represents the spatial coordinate of the th observation data point, represents the observation value of the th observation data point.

[0017] Grid representation of multi-source observations is input into the visual Transformer module of the encoder for deep learning feature fusion: (2) ; wherein, represents the input original multi-source observation data set, represents the set convolution processing process, represents the feature extraction and fusion process of the encoder visual Transformer, represents the output feature of the visual Transformer module of the encoder.

[0018] S3.2, extracting features by using multiple visual Transformers cascaded in the processor, obtaining a predicted field by using formula (3), and capturing long-range space-time dependence: (3); wherein, denotes the global ocean state at time step , and is a specific output of the VIT network.

[0019] S3.3, reconstructing a high-resolution SSHA field by using upsampling in the decoder.

[0020] Further, in step S4, the background field data is numerical simulation SSHA data near a specified date; the auxiliary physical quantities include sea surface temperature, sea surface flow field and geopotential height, and the output label is a true numerical simulation SSHA field of the specified date.

[0021] The present application has the following beneficial effects: 1. The present application combines high-resolution wide-swath satellite data and low-resolution daily satellite data products organically, uses the spatial fine structure of the former and the time continuity of the latter to correct high-resolution numerical simulation SSHA data, and breaks through the limitation of traditional methods that cannot simultaneously consider spatial resolution and time continuity.

[0022] 2. The present application uses a deep learning assimilation model to correct numerical simulation results, avoids the dependence on error covariance matrix and large-scale iterative calculation in traditional physical assimilation methods, greatly improves the calculation efficiency, and can realize rapid correction and real-time application.

[0023] 3. The numerical simulation data method of sea surface height proposed in the present application introduces background fields and auxiliary physical quantities in the model training process, so that the model can learn the pattern deviation law while enhancing the representation ability of complex dynamic processes, effectively reduce the systematic deviation of the model, and improve the accuracy, stability and scalability of the prediction results.

[0024] In summary, the present application proposes a numerical simulation data correction method of sea surface height based on deep learning, which combines high-resolution wide-swath satellite observations and low-resolution satellite products with good time continuity, and considers spatial resolution and time continuity while ensuring data accuracy. This method effectively reduces the systematic deviation of the model in dynamic active areas, significantly improves the reliability of SSHA data, and can generate a high-resolution data set covering a long time sequence. Compared with traditional assimilation methods, the present application has higher calculation efficiency, simpler structure, good robustness and scalability, and provides reliable support for ocean prediction, climate change research and intelligent ocean system. BRIEF DESCRIPTION OF DRAWINGS

[0025] In order to more clearly illustrate the technical solutions in the specific embodiments or prior art of the present application, the drawings required to be used in the description of the specific embodiments or prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor based on these drawings. In the drawings: Figure 1 A flowchart of a deep learning-based sea surface height numerical simulation data correction method of the present application is shown.

[0026] Figure 2 A sea surface height anomaly map of the numerical simulation data before correction is shown.

[0027] Figure 3 A real sea surface height anomaly map is shown.

[0028] Figure 4 A sea surface height anomaly map of the numerical simulation data after correction is shown.

[0029] Figure 5 A sea surface height anomaly difference distribution map before and after correction is shown.

[0030] Figure 6 A sea surface height anomaly difference distribution map after correction and real sea surface height anomaly data is shown. DETAILED DESCRIPTION

[0031] The technical solutions of the present application will be described clearly and completely below in combination with the drawings. Obviously, the described embodiments are some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor belong to the scope of protection of the present application.

[0032] As shown in a deep learning-based sea surface height numerical simulation data correction method, the method specifically comprises the following steps: Figure 1 S1, preparing data, including high-resolution wide swath data, low-resolution daily satellite data products, numerical simulation data and auxiliary data, and processing the data.

[0033] S2, constructing simulated satellite data based on numerical simulation data, including high-resolution wide swath satellite simulation data and low-resolution daily satellite product simulation data.

[0034] S3, constructing a deep learning assimilation model, including an encoder, a processor and a decoder connected to each other.

[0035] ​S4, training the deep learning assimilation model with the simulation data, background field data and auxiliary physical quantity as input and the numerical simulation data at the specified time as label to obtain a model capable of outputting the corrected high-resolution SSHA field; the auxiliary physical quantity is ERA5 auxiliary data.

[0036] S5, putting the dataset of real high-resolution wide-swath satellite observation data, low-resolution daily satellite data product, numerical simulation data and auxiliary data into the model trained in step S4 to output the corrected sea surface height anomaly data product.

[0037] The present application proposes a numerical simulation data correction method based on deep learning to solve the problems of difficult to balance spatial resolution and time continuity, systematic deviation of numerical simulation results and complex calculation of traditional assimilation methods in sea surface height anomaly (SSHA) data. The method provided by the present application fuses high spatial resolution but time sparse wide-swath satellite observation and time continuous but low spatial resolution satellite product, significantly improves the spatial resolution and time continuity of SSHA data under the premise of ensuring data accuracy, effectively corrects the model data, and can generate high-resolution SSHA data set covering long time sequence through data reconstruction, providing more reliable and complete data support for marine dynamic process analysis and business forecast.

[0038] Specifically, step S1 specifically includes the following steps: S1.1, time and space matching of different source data, and unification of spatial grid of all data to high-resolution grid of numerical simulation data through spatial interpolation to ensure data comparability and fusibility.

[0039] S1.2, outlier rejection and consistency processing of data to reduce the influence of observation error and noise on model training.

[0040] Specifically, since the real high-resolution wide-swath satellite observation data and low-resolution daily satellite product have deficiencies in time and space coverage and quality control, it is difficult to be directly used as large-scale training sample, the present application first constructs two types of simulation satellite data based on numerical simulation data to approximate the real observation characteristics and simulate the error distribution, so as to ensure that the difference relationship between observation and model can be fully learned in the training process. This simulation strategy can not only alleviate the problem of sparse and discontinuous observation data, but also provide systematic high-quality training samples for deep learning model to ensure that the model can effectively capture the regularity of model bias. The SSHA data of numerical simulation is processed to construct two types of satellite simulation data. Step S2 specifically includes the following steps: S2.1, Constructing high-resolution wide-swath satellite simulation data: Selecting the numerical simulation SSHA data with the same latitude and longitude range as the high-resolution wide-swath observation data as the high-resolution wide-swath satellite simulation data, and simulating the actual observation error by adding noise.

[0041] S2.2, Constructing low-resolution daily satellite product simulation data: Interpolating the numerical simulation SSHA data to the grid of the low-resolution daily satellite data product, and then re-interpolating the interpolated data to the numerical simulation grid, introducing blur processing to approximate the low-resolution observation characteristics, and adding noise to simulate observation uncertainty. The blur processing is as follows: interpolating the 2-kilometer resolution model data to the 0.125-degree grid, and then re-interpolating the data interpolated to the 0.125-degree grid back to the model data grid, which is low in resolution, the same as the resolution of the low-resolution daily satellite product, and has been blurred.

[0042] Specifically, step S3 specifically includes the following steps: S3.1, processing non-gridded observations and data missing through the SetConv layer of the encoder, as shown in formula (1): learning a learnable kernel function to aggregate sparse observations into a regular grid, enhancing the robustness and generalization ability of the model; (1) ; wherein, represents the estimated value in the grid coordinates, represents the learnable kernel function, represents the target grid point, represents the spatial coordinates of the th observation data point, represents the observation value of the th observation data point.

[0043] The gridded representation of the multi-source observation obtained by processing through formula (1) is input into the visual Transformer module of the encoder for deep learning feature fusion; the module captures long-range spatiotemporal dependencies through its self-attention mechanism and outputs a high-quality gridded initial state, i.e., the output feature of the visual Transformer module of the encoder , the process of which is shown in formula (2): (2) ; wherein, represents the input set of original multi-source observation data, represents the set convolution processing process, represents the feature extraction and fusion process of the encoder visual Transformer.

[0044] The encoder consists of interconnected SetConv layers and a visual Transformer module.

[0045] S3.2, extract features using multiple cascaded visual Transformers (i.e., VIT networks) in the processor, i.e., use formula (3) to obtain the prediction field and capture long-distance spatiotemporal dependencies: the cascaded visual Transformers form an autoregressive prediction chain, and the prediction field is obtained using formula (3), which improves the accuracy and reliability of the model.

[0046] (3); in, It is a high-dimensional tensor, representing the time step. The overall state of the ocean It is the specific output of the VIT network, representing the predicted state change.

[0047] S3.3 utilizes upsampling in the decoder to reconstruct a high-resolution SSHA field.

[0048] Specifically, in step S4, during the training of the deep learning assimilation model, it is necessary to ensure that the deep learning assimilation model can learn the error relationship between numerical simulation and observation, and also identify the bias characteristics of the model itself in the time evolution. To this end, this invention uses numerical simulation SSHA data from a specified date as a label to provide a clear training objective, i.e., the idealized real field at that moment; simultaneously, it selects numerical simulation results near that date as background field input, ensuring that the deep learning assimilation model can utilize the model error information contained in the temporally adjacent data, thereby achieving correction more effectively. Step S4 includes the following steps:

[0049] S4.1, Creating the training dataset The numerical simulation SSHA data of a specified date is taken as the label data, i.e., as the real sea surface height field. The observation input is constructed by step S2, i.e., two types of simulated satellite data are first generated based on the numerical simulation data: one type is high-resolution wide-swath simulated observation obtained by intercepting the same latitude and longitude region data as the wide-swath satellite coverage range and adding noise; the other type is low-resolution daily satellite product simulated observation obtained by first interpolating the numerical simulation data to the grid of the low-resolution daily satellite product, then interpolating back to the original model grid and performing blur processing and noise adding operation. At the same time, in order to provide the bias information of the model itself, the numerical simulation SSHA data near the specified date is also selected as the background field input; and further, auxiliary physical quantities of the same date as the background field are introduced, including sea surface temperature (SST), sea surface flow field (U, V) and geopotential, as additional input features to enhance the constraint on the model error characteristics. After time matching, spatial alignment and data cleaning processing, the finally formed input data is composed of two types of simulated satellite observation, background field data and auxiliary physical quantity, and the output label is the real numerical simulation SSHA field of the specified date. Based on this process, the complete training set, validation set and test set are constructed.

[0050] S4.2, the deep learning assimilation model constructed by step S3 dynamically fuses low-resolution daily satellite data products, high-resolution wide-swath data and numerical simulation data, efficiently extracts deep semantic features and accurately restores spatial details.

[0051] Specifically, in step S4, the background field data is the numerical simulation SSHA data near the specified date; the auxiliary physical quantities include sea surface temperature, sea surface flow field and geopotential, and the output label is the real numerical simulation SSHA field of the specified date. After the training of the deep learning assimilation model is completed, the data set composed of real high-resolution wide-swath satellite observation data, low-resolution daily satellite data product, numerical simulation data and auxiliary data is put into the model trained by step S4. The model corrects and reconstructs the input satellite observation data in spatial and temporal dimensions according to the multi-source data features and error correction rules learned in the training stage, and outputs the corrected high-resolution sea surface height anomaly data product with strong temporal continuity and significantly reduced error.

[0052] The application designs a dynamic standardization preprocessing procedure. Mean and standard deviation of background field, observation field and true value field are independently calculated based on a training set, and fixed standardization parameters are generated and synchronized to a validation set and a test set to eliminate distribution deviation of multi-source data; satellite simulation data and numerical simulation data are used to train a model, sea surface height field of the numerical simulation data is obtained through the satellite simulation data, and after the model is trained, real satellite observation data high-resolution wide swath data and low-resolution daily satellite data products can be used to obtain real sea surface height field. The encoder-processor-decoder cascade architecture is used to realize the whole process prediction, the efficiency is significantly improved compared with the traditional scheme, the end-to-end joint fine-tuning is supported, and the error is reduced. The training loss function only calculates the error of the effective observation area, and the dynamic learning rate scheduling (ReduceLROnPlateau) is combined, and finally the high spatial and temporal resolution numerical simulation data sea surface height anomaly product is output.

[0053] In order to verify the method provided by the application, taking the data of a sea area on January 12, 2004 as an example, the abscissa represents longitude, and the ordinate represents latitude. Figure 2 is a sea surface height anomaly map before the numerical simulation data is corrected; Figure 3 is a real sea surface height anomaly map; Figure 4 is a sea surface height anomaly map after the numerical simulation data is corrected; by comparing the three maps, it can be found that the area near the land has been significantly corrected.

[0054] As shown in Figure 5 , the sea surface height anomaly difference map of the numerical simulation data before and after correction, the deeper the color, the greater the difference, the root mean square difference is 0.0933, indicating that the sea surface height is effectively corrected by deep learning; as shown in Figure 6 , the difference distribution map of the corrected numerical simulation data and the real sea surface height anomaly data, the deeper the color, the greater the difference between the two, although there is a difference, but the difference is small, the root mean square difference is 0.0017, indicating that the numerical simulation data corrected by the deep learning method provided by the application has a small difference with the real sea surface height field.

[0055] Of course, the above description is not a limitation of the application, and the application is not limited to the above examples. Changes, modifications, additions or replacements made by those skilled in the art within the essential scope of the application should also be within the protection scope of the application.

Claims

1. A method for correcting sea surface height numerical simulation data based on deep learning, characterized in that, Specifically, the steps include the following: S1. Prepare data including: high-resolution wide-swath data, low-resolution daily satellite data products, numerical simulation data and auxiliary data, and process the data; S2, based on numerical simulation data, constructs simulated satellite data, including: high-resolution wide-swath satellite simulation data and low-resolution daily satellite product simulation data; S3 builds deep learning assimilation models, including interconnected encoders, processors, and decoders; S4 takes simulation data, background field data, and auxiliary physical quantities as inputs, and numerical simulation data at a specified time as labels to train a deep learning assimilation model, resulting in a model that can output a corrected high-resolution SSHA field. S5 inputs the dataset created from real high-resolution wide-swath satellite observation data, low-resolution daily satellite data products, numerical simulation data, and auxiliary data into the model trained in step S4, and outputs the corrected sea surface height anomaly data product.

2. The method for correcting sea surface height numerical simulation data based on deep learning according to claim 1, characterized in that, Step S1 specifically includes the following steps: S1.1 performs spatiotemporal matching on data from different sources and unifies the spatial grid of all data to the high-resolution grid of the numerical simulation data through spatial interpolation. S1.2, perform outlier removal and data consistency processing.

3. The method for correcting sea surface height numerical simulation data based on deep learning according to claim 1, characterized in that, Step S2 specifically includes the following steps: S2.1, Constructing high-resolution wide-swath satellite simulation data: Select numerical simulation SSHA data with the same latitude and longitude range as the high-resolution wide-swath observation data as the high-resolution wide-swath satellite simulation data, and simulate actual observation errors by adding noise; S2.2, Constructing low-resolution daily satellite product simulation data: Interpolate the numerical simulation SSHA data onto the grid of the low-resolution daily satellite data product, and then re-interpolate the interpolated data onto the numerical simulation grid. Introduce fuzzing to approximate the low-resolution observation characteristics, and add noise to simulate observation uncertainties.

4. The method for correcting sea surface height numerical simulation data based on deep learning according to claim 1, characterized in that, Step S3 specifically includes the following steps: S3.1, the problem of non-meshable observations and missing data is handled through the SetConv layer of the encoder, as shown in formula (1): （1）； in, This represents the estimated value in grid coordinates. This represents a learnable kernel function. Indicates the target grid point. Indicates the first Spatial coordinates of each observation data point Indicates the first The observed values ​​of each observation data point; The multi-source observations are represented in a gridded format and then input into the visual Transformer module of the encoder for deep learning feature fusion. （2）； in, This represents the original set of multi-source observation data input. This represents the set convolution process. The feature extraction and fusion process of the encoder's visual Transformer. This represents the output characteristics of the encoder's visual Transformer module; S3.2, features are extracted using multiple cascaded visual Transformers in the processor, and the prediction field is obtained using formula (3) to capture long-distance spatiotemporal dependencies: （3）； in, Indicates time step The overall state of the ocean It is the specific output of the VIT network; S3.3 utilizes upsampling in the decoder to reconstruct a high-resolution SSHA field.

5. The method for correcting sea surface height numerical simulation data based on deep learning according to claim 1, characterized in that, In step S4, the background field data is numerical simulation SSHA data near a specified date; the auxiliary physical quantities include: sea surface temperature, sea surface current field and geopotential height, and the output label is the real numerical simulation SSHA field of the specified date.

Images