Marine environmental information data assimilation method and system based on deep learning model

Abstract

The present invention discloses a method and system for assimilating marine environmental information data based on a deep learning model, and relates to the technical field of marine environmental information data processing. By acquiring the time series observation data of the marine environmental elements in the spatial distribution field in the observation area, and using the ensemble Kalman filter algorithm for assimilation, the iteratively updated data error is used as the input of the generator G, and the observed data is used as the real data to construct a generative adversarial network, the discriminator D and the generator G of the generative adversarial network are trained, and the trained generator G model is used to synthesize virtual observation data samples. By determining whether the assimilation performance evaluation index reaches the set expected value as an iterative determination condition to control the assimilation process, it can effectively solve the problem of realizing high-resolution prediction of the spatial distribution field of marine environmental elements in the observation area based on sparse observation data, thereby meeting the needs of high-precision dynamic system models for high-resolution marine environmental information.

Description

Technical Field

[0001] The present invention relates to the technical field of marine environment information data processing, and in particular to a method and system for assimilating marine environment information data based on a deep learning model. Background Art

[0002] The way to obtain marine environmental information is through direct observation by equipment. However, due to objective factors such as equipment and environment, the observation data often have errors or cannot be measured. Therefore, when using observation data, it is necessary to perform error analysis on the original data obtained to obtain the most credible observation data. Data assimilation is a bridge between system models and observation data. It is a sequential inference method for dynamic system models and static observation parameters. Under the existing ocean dynamic system framework, it combines the forecast field of the numerical model with high-precision observation information from different sources. Through optimization methods, the random errors of the numerical model itself are continuously reduced to obtain a more accurate or more complete analysis field than the original numerical model results and observation information, which better describes the true state of the ocean. It is widely used in the fields of ocean data reanalysis and numerical forecasting.

[0003] At present, the assimilation methods for marine environmental information mainly include variational method, ensemble Kalman filter method and ensemble optimal smoothing method. For the acquisition of observation data of marine environmental elements such as marine environmental temperature, salinity, and current velocity, it is usually adopted to set up monitoring points and arrange corresponding sensors for periodic collection. The acquired marine environmental observation data such as temperature, salinity, and current velocity are relatively sparse in time series spatial distribution and three-dimensional spatial distribution relative to the change trend of the marine environment. When the above assimilation methods are used based on sparse observation data, there will be a deviation between the analysis field of the dynamic system model and the actual ocean state. In order to solve the technical problem of how to achieve high-resolution prediction of the spatial distribution field of marine environmental elements such as marine environmental temperature, salinity, and current velocity in the observation area based on sparse observation data, so as to meet the high-precision dynamic system model's demand for high-resolution marine environmental information, we propose a marine environmental information data assimilation method and system based on a deep learning model. Summary of the invention

[0004] The main purpose of the present invention is to provide a method and system for assimilating marine environmental information data based on a deep learning model, which can effectively solve the problems in the background technology.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is:

[0006] The marine environment information data assimilation method based on the deep learning model includes the following steps:

[0007] S1: Obtaining time series observation data of any marine environmental element in the observation area in the spatial distribution field, wherein the marine environmental element includes temperature, salinity, and ocean current velocity;

[0008] S2: Classify the observed data into N sets according to the sampling time, and define is the state analysis value of the i-th set at time t, is the state prediction value of the i-th set at time t+1. The acquired time series observation data is expressed using the collective Kalman filter formula, and the state analysis value is iteratively updated. The collective Kalman filter formula expression is:

[0009]

[0010] Where M t,t+1 is the linear prediction factor from time t to time t+1; w i,t is the model error at time t, and w i,t ~N(0,Q t ), Q t is the covariance matrix of the set; is the predicted value of the state of the i-th set at time t; is the state analysis value of the i-th set at time t; α is the current state parameter; f is the predicted state parameter;

[0011] The state analysis value iterative update process expression is:

[0012]

[0013] In formula (1), is the state analysis value of the i-th set at time t+1; K t+1 is the gain matrix; is the observed data at time t+1; v i,t is the observation error at time t, and v i,t ~N(0,Q t );H t+1 is the observation factor at time t+1;

[0014] In formula (2), is the mean value of the state analysis value at time t+1; N is the total amount of data;

[0015] In formula (3), is the analysis field error variance matrix; H is the observation factor; R t is the measurement noise covariance matrix at time t;

[0016] In formula (4), is the mean of the state prediction values ​​of the i-th set at time t;

[0017] In formula (7), is the mean value of the state prediction at time t+1;

[0018] S3: Iteratively updated data error w i,t As the input of the generator G, the obtained observation data is used as the real data to construct a generative adversarial network, and the generator G is used to generate virtual observation data G(z). The discriminator D and the generator G of the generative adversarial network are trained by the observation data and the generated virtual observation data G(z), until the output result of the discriminator D is D(G(z))=1;

[0019] The training process of the generative adversarial network includes the following steps:

[0020] S31: The data error w is updated iteratively according to the input through the generative network of the generator G i,t Generate virtual observation data samples and input them into the discriminator D, and use the discriminant network of the discriminator D to make judgments;

[0021] S32: The discriminant network classifies the input virtual observation data samples and real data, and calculates the classification error;

[0022] S33: updating the parameters of the discriminant network according to the classification error to improve its classification accuracy;

[0023] S34: the generating network updates its own parameters according to the feedback information of the discriminating network to generate more realistic virtual observation data samples;

[0024] S35: Repeat the above steps S31 to S34 until a preset number of training rounds or convergence condition is reached;

[0025] The training process of the generator G includes the following steps:

[0026] S311: Obtain data samples of iteratively updated data errors;

[0027] S312: transforming the data sample of the data error;

[0028] S313: Determine the classification result of the discriminator D and classify the classification result as "true" or "false", wherein the output of the discriminator D is a binary variable D(x)=1 or D(x)=0; when the output D(x)=1, it indicates that the classification result of the discriminator D is true; when the output D(x)=0, it indicates that the classification result of the discriminator D is false;

[0029] S314: Calculate the loss of the discriminant network according to the classification result of the discriminator D;

[0030] S315: Back propagate through the discriminator D and the generator G to obtain the gradient of the discriminant network;

[0031] S316: Use the gradient of the discriminant network to change the generation network parameters of the generator G;

[0032] The training process of the discriminator D includes the following steps:

[0033] S321: Classify the input real data and the virtual observation data samples generated by the generator G;

[0034] S322: The loss function of the discriminant network of the discriminator D penalizes the mistakes it makes;

[0035] S323: Through back propagation, the discriminator D updates the weights of its discriminant network;

[0036] The discriminator D is used to discriminate the authenticity of the input data, that is, to distinguish the real data from the virtual observation data generated by the generator G; wherein, the input parameter of the discriminator D is set to be x, and the output parameter is D(x), wherein D(x) represents the probability that x is the real data, and if the output parameter D(x)=1, it represents that the input parameter x is 100% the real data, and if the output parameter D(x)=0, it represents that the input parameter x cannot be the real data;

[0037] S4: Use the trained generator G model to synthesize several virtual observation data samples, where the sample size of the virtual observation data and the sample size of the real data satisfy the following relationship:

[0038]

[0039] In the formula, Q i It is represented as the data sample size of the i-th set of observed data; Q v It is represented by the sample size of virtual observation data; Q max It is expressed as the maximum data sample size in the N set of observed data; ε is a constant coefficient not less than 1;

[0040] S5: Repeat steps S2-S4 until the assimilation performance evaluation index reaches the set expected value, where the assimilation performance evaluation index includes the root mean square error RMSE of the observed data and the ensemble mean RMSE of the predicted value of the observed data at time t. t , where the calculation formula for the root mean square error RMSE of the observed data is:

[0041]

[0042] RMSE of the ensemble mean of the predicted values ​​of the observed data t The calculation formula is:

[0043]

[0044] Where L is the length of assimilation time; X(t) is the variable function of the real state at time t; is the predicted value of the state of the i-th set at time t-1.

[0045] The marine environment information data assimilation system based on the deep learning model includes data acquisition module, data classification module, data processing module, generative adversarial network construction module, data iteration module, and performance determination module;

[0046] The data acquisition module is used to obtain time series observation data of any marine environmental element in the observation area in the spatial distribution field, wherein the marine environmental element includes temperature, salinity, and current velocity;

[0047] The data classification module is used to classify the observed data into N sets according to the sampling time, defining is the state analysis value of the i-th set at time t, is the predicted value of the state of the i-th set at time t+1;

[0048] The data processing module is used to express the acquired time series observation data using the ensemble Kalman filter formula and iteratively update the state analysis value; the generative adversarial network construction module is used to iteratively update the data error w i,t As the input of the generator G, the obtained observation data is used as the real data to build a generative adversarial network, and the generator G is used to generate virtual observation data G(z). The discriminator D and the generator G of the generative adversarial network are trained by the observation data and the generated virtual observation data G(z). When the output result of the discriminator D is D(G(z))=1, a number of virtual observation data samples are synthesized using the trained generator G model;

[0049] The data iteration module is used to express the acquired virtual observation data samples using a collective Kalman filter formula and iteratively update the state analysis value;

[0050] The performance judgment module is used to evaluate the data assimilation process and determine whether the data assimilation process continues to be iteratively updated according to the set expected value of the assimilation performance evaluation index, wherein the assimilation performance evaluation index includes the root mean square error RMSE of the observed data and the ensemble mean RMSE of the predicted value of the observed data at time t. t ;

[0051] The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor.

[0052] The present invention has the following beneficial effects:

[0053] Compared with the existing technology, by obtaining the time series observation data of any marine environmental element in the observation area in the spatial distribution field, the observation data are classified into N sets according to the sampling time, and the definition is the state analysis value of the i-th set at time t, is the state prediction value of the i-th set at time t+1. The acquired time series observation data is expressed using the collective Kalman filter formula, and the state analysis value is iteratively updated to obtain the iteratively updated data error w i,t As the input of the generator G, the acquired observation data is used as the real data to construct a generative adversarial network, and the generator G is used to generate virtual observation data G(z). The discriminator D and the generator G of the generative adversarial network are trained through the observation data and the generated virtual observation data G(z), and the trained generator G model is used to synthesize a number of virtual observation data samples. The iterative process of the assimilation process is controlled by determining whether the assimilation performance evaluation index reaches the set expected value, which effectively solves the problem of realizing high-resolution prediction of the spatial distribution field of marine environmental elements in the observation area based on sparse observation data, thereby meeting the demand of high-precision dynamic system models for high-resolution marine environmental information. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a flow chart of the marine environment information data assimilation method based on the deep learning model of the present invention;

[0055] Figure 2 It is a structural block diagram of the marine environment information data assimilation system based on the deep learning model of the present invention;

[0056] Figure 3 A block diagram of the structure of the generative adversarial network constructed for the present invention. DETAILED DESCRIPTION

[0057] The present invention will be further described below in conjunction with specific implementation methods, wherein the accompanying drawings are only used for exemplary descriptions and represent only schematic diagrams rather than actual drawings, and should not be understood as limiting the present invention. In order to better illustrate the specific implementation methods of the present invention, some parts of the accompanying drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product.

[0058] The specific implementation process of the technical solution of the present invention includes the following steps:

[0059] Step 1: Obtain the time series observation data of any marine environmental element in the observation area in the spatial distribution field, where the marine environmental elements include temperature, salinity, and current velocity.

[0060] Step 2: Classify the observed data into N sets according to the sampling time and define is the state analysis value of the i-th set at time t, is the predicted value of the state of the i-th set at time t+1.

[0061] Step 3: The acquired time series observation data is expressed using the ensemble Kalman filter formula, and the state analysis value is iteratively updated. The ensemble Kalman filter formula expression is:

[0062]

[0063] Where M t,t+1 is the linear prediction factor from time t to time t+1; w i,t is the model error at time t, and w i,t ~N(0,Q t ), Q t is the covariance matrix of the set;

[0064] The state analysis value iteration update process expression is:

[0065]

[0066] In formula (1), is the state analysis value of the i-th set at time t+1; K t+1 is the gain matrix; is the observed data at time t+1; v i,t is the observation error at time t, and v i,t ~N(0,Q t );

[0067] In formula (2), is the mean value of the state analysis value at time t+1;

[0068] In formula (3), is the analysis field error variance matrix; H is the observation factor; R t is the measurement noise covariance matrix at time t;

[0069] In formula (4), is the mean of the state prediction values ​​of the i-th set at time t.

[0070] Step 4: Iteratively update the data error w i,t As the input of the generator G, the observed data obtained is used as the real data to construct a generative adversarial network, where the structure of the generative adversarial network is as follows Figure 3As shown, the generator G is used to generate virtual observation data G(z), and the discriminator D and the generator G of the generative adversarial network are trained by the observation data and the generated virtual observation data G(z), until the output result of the discriminator D is D(G(z))=1;

[0071] The training process of the generative adversarial network includes the following steps:

[0072] S41: The data error w is updated iteratively according to the input through the generative network of the generator G i,t Generate virtual observation data samples and input them into the discriminator D, and use the discriminant network of the discriminator D to make judgments;

[0073] S42: The discriminant network classifies the input virtual observation data samples and real data, and calculates the classification error;

[0074] S43: updating the parameters of the discriminant network according to the classification error to improve its classification accuracy;

[0075] S44: the generating network updates its own parameters according to the feedback information of the discriminating network to generate more realistic virtual observation data samples;

[0076] S45: Repeat the above steps S41 to S44 until a preset number of training rounds or convergence condition is reached;

[0077] The training process of the generator G includes the following steps:

[0078] S411: Obtain data samples of iteratively updated data errors;

[0079] S412: transforming the data sample of the data error;

[0080] S413: Determine the classification result of the discriminator D and classify the classification result as "true" or "false", wherein the output of the discriminator D is a binary variable D(x)=1 or D(x)=0; when the output D(x)=1, it indicates that the classification result of the discriminator D is true; when the output D(x)=0, it indicates that the classification result of the discriminator D is false;

[0081] S414: Calculate the loss of the discriminant network according to the classification result of the discriminator D, wherein the loss of the discriminant network can be calculated by a loss function, and the expression is:

[0082]

[0083] In the formula, J (D) (θ (D) ,θ (G) ) represents the loss of the discriminant network; E is the expected probability; x~P represents the distribution of x satisfying P;

[0084] S415: Back propagate through the discriminator D and the generator G to obtain the gradient of the discriminant network;

[0085] S416: Use the gradient of the discriminant network to change the generation network parameters of the generator G;

[0086] The training process of the discriminator D includes the following steps:

[0087] S421: Classify the input real data and the virtual observation data samples generated by the generator G;

[0088] S422: The loss function of the discriminant network of the discriminator D penalizes the mistakes it makes;

[0089] S423: Through back propagation, the discriminator D updates the weights of its discriminant network;

[0090] The discriminator D is used to discriminate the authenticity of the input data, that is, to distinguish the real data from the virtual observation data generated by the generator G; wherein, the input parameter of the discriminator D is set to be x, and the output parameter is D(x), wherein D(x) represents the probability that x is the real data, and if the output parameter D(x)=1, it means that the input parameter x is 100% real data, and if the output parameter D(x)=0, it means that the input parameter x cannot be the real data;

[0091] It should be noted that during the training phase, the generative adversarial network uses two data sources: real data, that is, sampled observation data, which is an example that the generator G tries to replicate; and generated data from the generator G; real data is used as positive examples in training, while the generator G examples are used as negative examples. During the training process of the discriminator D, the generator G remains unchanged: when it generates data for the discriminator D to train, the weights of its network do not change. When the training starts, the generator G generates random and obviously fake data, and the discriminator D can easily judge the data as fake data. As the training progresses, the generator G gets closer and closer to producing outputs that can deceive the discriminator D until it starts to classify fake data as real data.

[0092] Step 5: Use the trained generator G model to synthesize several virtual observation data samples, where the sample size of virtual observation data and the sample size of real data satisfy the following relationship:

[0093]

[0094] In the formula, Q i It is represented as the data sample size of the i-th set of observed data; Q v It is represented by the sample size of virtual observation data; Q maxIt is expressed as the maximum data sample size in the N set of observed data; ε is a constant coefficient not less than 1;

[0095] Step 6: Repeat steps S2-S4 until the assimilation performance evaluation index reaches the set expected value, where the assimilation performance evaluation index includes the root mean square error RMSE of the observed data and the ensemble mean RMSE of the predicted value of the observed data at time t. t , where the calculation formula for the root mean square error RMSE of the observed data is:

[0096]

[0097] RMSE of the ensemble mean of the predicted values ​​of the observed data t The calculation formula is:

[0098]

[0099] Where L is the length of assimilation time; X(t) is the variable function of the real state at time t;

[0100] It should be noted that when the calculation result of the assimilation performance evaluation index meets the set expected value, it means that the time series observation data of marine environmental elements over the time length L has been assimilated. If the time length L ≤ the total sampling time of the observation data, it is necessary to continue to assimilate the remaining observation data. If the time length L > the total sampling time of the observation data, it means that all the sampled observation data have been assimilated.

[0101] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A method for assimilating marine environmental information data based on a deep learning model, characterized in that: The following steps are involved: S1: Obtaining time series observation data of any marine environmental element in the observation area in the spatial distribution field, wherein the marine environmental element includes temperature, salinity, and ocean current velocity; S2: Classify the observed data into N sets according to the sampling time, and define is the state analysis value of the i-th set at time t, is the state prediction value of the i-th set at time t+1. The acquired time series observation data is expressed using the collective Kalman filter formula, and the state analysis value is iteratively updated. The collective Kalman filter formula is expressed as follows: Where M t,t+1 is the linear prediction factor from time t to time t+1; w i,t is the model error at time t, and w i,t ~N(0,Q t ), Q t is the covariance matrix of the set; is the predicted value of the state of the i-th set at time t; is the state analysis value of the i-th set at time t; α is the current state parameter; f is the predicted state parameter; S3: Iteratively updated data error w i,t As the input of the generator G, the obtained observation data is used as the real data to build a generative adversarial network, and the generator G is used to generate virtual observation data G(z). The discriminator D and the generator G of the generative adversarial network are trained by the observation data and the generated virtual observation data G(z) until the output result of the discriminator D is D(G(z))=1; S4: Use the trained generator G model to synthesize several virtual observation data samples; S5: Repeat steps S2-S4 until the assimilation performance evaluation index reaches the set expected value, where the assimilation performance evaluation index includes the root mean square error RMSE of the observed data and the ensemble mean RMSE of the predicted value of the observed data at time t. t ; The state analysis value iterative update process expression is: In formula (1), is the state analysis value of the i-th set at time t+1; K t+1 is the gain matrix; is the observed data at time t+1; v i,t is the observation error at time t, and v i,t ~N(0,Q t );H t+1 is the observation factor at time t+1; In formula (2), is the mean value of the state analysis value at time t+1; N is the total amount of data; In formula (3), is the analysis field error variance matrix; H is the observation factor; R t is the measurement noise covariance matrix at time t; In formula (4), is the mean of the state prediction values ​​of the i-th set at time t; In formula (7), is the mean of the state prediction values ​​at time t+1.

2. The method for assimilating marine environmental information data based on a deep learning model according to claim 1, characterized in that: The training process of the generative adversarial network includes the following steps: S31: The data error w is updated iteratively according to the input through the generative network of the generator G i,t Generate virtual observation data samples and input them into the discriminator D, and use the discriminant network of the discriminator D to make judgments; S32: The discriminant network classifies the input virtual observation data samples and real data, and calculates the classification error; S33: updating the parameters of the discriminant network according to the classification error to improve its classification accuracy; S34: the generating network updates its own parameters according to the feedback information of the discriminating network to generate more realistic virtual observation data samples; S35: Repeat the above steps S31 to S34 until a preset number of training rounds or convergence condition is reached.

3. The method for assimilating marine environmental information data based on a deep learning model according to claim 1, characterized in that: The training process of the generator G includes the following steps: S311: Obtain data samples of iteratively updated data errors; S312: transforming the data sample of the data error; S313: Determine the classification result of the discriminator D and classify the classification result as "true" or "false", wherein the output of the discriminator D is a binary variable D(x)=1 or D(x)=0; when the output D(x)=1, it indicates that the classification result of the discriminator D is true; when the output D(x)=0, it indicates that the classification result of the discriminator D is false; S314: Calculate the loss of the discriminant network according to the classification result of the discriminator D; S315: Back propagate through the discriminator D and the generator G to obtain the gradient of the discriminant network; S316: Use the gradient of the discriminant network to change the generation network parameters of the generator G.

4. The method for assimilating marine environmental information data based on a deep learning model according to claim 1, characterized in that: The training process of the discriminator D includes the following steps: S321: Classify the input real data and the virtual observation data samples generated by the generator G; S322: The loss function of the discriminant network of the discriminator D penalizes the mistakes it makes; S323: Through back propagation, the discriminator D updates the weights of its discriminant network; The discriminator D is used to discriminate the authenticity of the input data, that is, to distinguish between real data and virtual observation data generated by the generator G; wherein, the input parameter of the discriminator D is set to be x, and the output parameter is D(x), wherein D(x) represents the probability that x is real data, if the output parameter D(x) = 1, it means that the input parameter x is 100% real data, if the output parameter D(x) = 0, it means that the input parameter x cannot be real data.

5. The method for assimilating marine environmental information data based on a deep learning model according to claim 1, characterized in that: The calculation formula of the root mean square error RMSE of the observed data is: RMSE of the ensemble mean of the predicted values ​​of the observed data t The calculation formula is: Where L is the length of assimilation time; X(t) is the variable function of the real state at time t; is the predicted value of the state of the i-th set at time t-1.

6. The method for assimilating marine environmental information data based on a deep learning model according to claim 1, characterized in that: The sample size of virtual observation data and the sample size of real data satisfy the following relationship: In the formula, Q i It is represented as the data sample size of the i-th set of observed data; Q v It is represented by the sample size of virtual observation data; Q max It is expressed as the maximum data sample size in the N set of observed data; ε is a constant coefficient not less than 1.

7. The marine environment information data assimilation system based on deep learning model is characterized by: It includes data collection module, data classification module, data processing module, generative adversarial network construction module, data iteration module and performance determination module; The data acquisition module is used to obtain time series observation data of any marine environmental element in the observation area in the spatial distribution field, wherein the marine environmental element includes temperature, salinity, and current velocity; The data classification module is used to classify the observed data into N sets according to the sampling time, defining is the state analysis value of the i-th set at time t, is the predicted value of the state of the i-th set at time t+1; The data processing module is used to express the acquired time series observation data using a collective Kalman filter formula and iteratively update the state analysis value, wherein the collective Kalman filter formula expression is: Where M t,t+1 is the linear prediction factor from time t to time t+1; w i,t is the model error at time t, and w i,t ~N(0,Q t ), Q t is the covariance matrix of the set; is the predicted value of the state of the i-th set at time t; is the state analysis value of the i-th set at time t; α is the current state parameter; f is the predicted state parameter; the iterative update process expression is: In formula (1), is the state analysis value of the i-th set at time t+1; K t+1 is the gain matrix; is the observed data at time t+1; v i,t is the observation error at time t, and v i,t ~N(0,Q t );H t+1 is the observation factor at time t+1; In formula (2), is the mean value of the state analysis value at time t+1; N is the total amount of data; In formula (3), is the analysis field error variance matrix; H is the observation factor; R t is the measurement noise covariance matrix at time t; In formula (4), is the mean of the state prediction values ​​of the i-th set at time t; In formula (7), is the mean value of the state prediction at time t+1; The generative adversarial network building module is used to iteratively update the data error w i,t As the input of the generator G, the obtained observation data is used as the real data to build a generative adversarial network, and the generator G is used to generate virtual observation data G(z). The discriminator D and the generator G of the generative adversarial network are trained by the observation data and the generated virtual observation data G(z). When the output result of the discriminator D is D(G(z))=1, a number of virtual observation data samples are synthesized using the trained generator G model; The data iteration module is used to express the acquired virtual observation data samples using a collective Kalman filter formula and iteratively update the state analysis value; The performance judgment module is used to evaluate the data assimilation process and determine whether the data assimilation process continues to be iteratively updated according to the set expected value of the assimilation performance evaluation index, wherein the assimilation performance evaluation index includes the root mean square error RMSE of the observed data and the ensemble mean RMSE of the predicted value of the observed data at time t. t , the calculation formulas are: Where L is the length of assimilation time; X(t) is the variable function of the real state at time t; is the predicted value of the state of the i-th set at time t-1.

8. The marine environment information data assimilation system based on deep learning model according to claim 7 is characterized in that: The system comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the steps of the method according to any one of claims 1 to 6 are implemented when the processor executes the program.

Images