A deep learning-based assimilation method for sparse ocean observation data
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HARBIN ENG UNIV
- Filing Date
- 2022-11-24
- Publication Date
- 2026-07-24
AI Technical Summary
Existing data assimilation methods struggle to effectively address nonlinear issues in the context of sparse observation data, resulting in poor assimilation performance.
A fully connected deep neural network model is adopted. By constructing a training dataset and a training label dataset, the deep learning network model is trained. The Adam gradient descent method and the MSEloss loss function are used to achieve nonlinear mapping and diffuse sparse observation information to unobserved locations.
It improves the utilization value and efficiency of sparse observation data, enhances the assimilation effect, and yields more accurate analytical fields, supporting precise marine environmental monitoring and forecasting.
Smart Images
Figure CN115758898B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of marine observation technology, specifically relating to a deep learning-based method for assimilating sparse marine observation data. Background Technology
[0002] The ocean has a significant impact on human maritime activities, marine engineering, aquaculture, fishing, weather, and climate. Timely monitoring of marine environmental parameters is crucial for social production, daily life, and national defense. Instrumental observation is the primary means of monitoring marine environmental parameters; however, the amount of data obtained from instrumental observations is generally extremely limited. To effectively monitor the marine environment, sparse observation data assimilation techniques are needed. These techniques can diffuse limited observational information to areas where there is no existing observational data, thereby improving the utilization value and efficiency of instrumental observations.
[0003] Existing data assimilation methods place high demands on the quantity and spatiotemporal distribution of observation data. When observation data is sparse, it exhibits strong nonlinear characteristics, making it difficult for existing data assimilation methods to achieve good assimilation results. Summary of the Invention
[0004] To address the problem of poor assimilation of sparse observation data, this invention proposes a deep learning-based method for assimilation of sparse ocean observation data. This method effectively solves the nonlinearity problem in computation by learning from historical data to find the optimal parameters to achieve nonlinear mapping.
[0005] The specific technical solution of the deep learning-based ocean sparse observation data assimilation method is as follows:
[0006] Step 1: Obtain background field data and instrument observation data of marine environmental parameters in the area to be processed, and construct a training dataset by combining the marine history reanalysis data of the area;
[0007] The training dataset includes a training sample dataset and a training label dataset, which are as follows:
[0008] The location information of each sparse observation point in the instrument observation data is extracted. For the same sparse observation point location, the parameter value of that location in the historical reanalysis data at different times is subtracted from the parameter value of that location in the background field data. The differences at all sparse observation point locations are used to construct a training sample dataset S = {S1, S2, ..., S...}. n};
[0009] By subtracting the parameter values of all historical reanalysis data at corresponding time points from the parameter values of the background field data, a training label dataset L = {L1, L2, ..., L...} is constructed. n}
[0010] Step 2: Construct a deep learning network model and train it using the training dataset to obtain a sparse observation data assimilation computation model;
[0011] The deep learning network model adopts a fully connected deep neural network architecture, including an input layer, hidden layers, and an output layer. The input layer includes batch normalization and activation functions, and the number of neurons in the input layer is equal to the number of all observation points. The hidden layer includes batch normalization and activation functions, and the number of layers and neurons in the hidden layer is determined according to the grid size. The number of neurons in the output layer is equal to the number of grid points in the background field.
[0012] The specific steps for training a deep learning network model are as follows:
[0013] Step 2.1: Initialize the hyperparameters of the deep learning network model based on the size of the training sample dataset and computing resources;
[0014] Model hyperparameters include learning rate and number of training iterations.
[0015] Step 2.2: The training sample dataset S = {S1, S2, ..., S...} is... n The data in} is used as input to the deep learning network model. Adam gradient descent is used as the optimizer, MSEloss loss function is used as the cost function, ReLU is used as the activation function, and the model's hyperparameters are trained step by step using an adaptive learning rate method. Through multiple iterations, the optimal hyperparameters are obtained when the cost function no longer decreases. The deep learning network model corresponding to the optimal hyperparameters is used as the sparse observation data assimilation calculation model.
[0016] Step 3: Input the sparse observation data of the ocean area to be processed into the assimilation calculation model to obtain the analysis field parameter values after fusing the observation information, thereby spreading the information of each observation point in the background field to the location without observation information;
[0017] The steps for calculating the field parameter values are as follows:
[0018] Step 3.1: Subtract the data values of each sparse observation point from the corresponding background field data parameter values to obtain the test sample dataset T = {T1, T2, ..., T...} n};
[0019] Step 3.2: Generate the test sample dataset T = {T1, T2, ..., T} n The data is input into the assimilation calculation model to obtain the observation increments for the entire field, and then smoothed.
[0020] Step 3.3: Add the smoothed full-field observation increment to the background field data parameter values to obtain the analysis field parameter values:
[0021] A = E + B
[0022] Where A represents the analysis field parameter value, E represents the observation increment of the entire field, and B represents the background field parameter value.
[0023] The beneficial effects of this invention are:
[0024] This invention utilizes the nonlinear mapping capabilities of deep learning networks to design a deep learning-based method for assimilating sparse ocean observation data. This method can diffuse limited observational information to locations without existing observational data, improving the utilization value and efficiency of instrumental observations, and enhancing the algorithm's versatility and assimilation effect. The analysis field calculated based on this model more closely reflects real ocean conditions, which is beneficial for subsequent more accurate marine environmental monitoring, ocean numerical forecasting, underwater acoustic communication, and sonar detection. Attached Figure Description
[0025] Figure 1 This is a flowchart of the steps of the deep learning-based ocean sparse observation data assimilation method of the present invention;
[0026] Figure 2 This is a schematic diagram of the distribution of observation points in an embodiment of the present invention;
[0027] Figure 3 This is a schematic diagram of the deep learning network structure in an embodiment of the present invention. Detailed Implementation
[0028] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0029] In traditional methods for assimilating sparse ocean observation data, the correlation between grid points is represented by the background error covariance. However, estimating the background error covariance in traditional assimilation methods is challenging, often relying on simple statistical analysis for approximation. This leads to information loss as observational information is extended from the observation point to other grid points. Furthermore, simple statistical analysis cannot address nonlinear issues, which are crucial for assimilation in practical engineering applications. Therefore, this invention first constructs a training dataset using ocean history reanalysis data and then trains a fully connected deep neural network model using this dataset to obtain a sparse observation data assimilation calculation model. This model achieves a more accurate assimilation of sparse observation data, resolving the nonlinear relationship problem between grid points in traditional assimilation methods.
[0030] This invention relates to a deep learning-based method for assimilating sparse ocean observation data, such as... Figure 1 As shown, it includes the following steps:
[0031] Step 1: For the area to be processed in the ocean, construct a training dataset based on the ocean history reanalysis data of the area to obtain the correlation variation law between the parameters of the observation point location and the parameters of other grid points;
[0032] The specific process is as follows:
[0033] Step 1.1: Obtain background field data of marine environmental parameters for the area to be processed;
[0034] Step 1.2: Obtain marine historical reanalysis data for the area to be processed;
[0035] Step 1.3: Acquire instrument observation data for the area to be processed;
[0036] Step 1.4: Extract the location information of each sparse observation point from the instrument observation data. For the same sparse observation point location, calculate the difference between the parameter value of that location in the historical reanalysis data at different times and the parameter value of that location in the background field data at that time. Construct a training sample dataset S = {S1, S2, ..., S...} for all sparse observation point locations. n};
[0037] Step 1.5: Subtract the parameter values of all historical reanalysis data at the corresponding time points from the parameter values of the background field data to construct the training label dataset L = {L1, L2, ..., L...} n}
[0038] Step 1.6: Training sample dataset S = {S1, S2, ..., S...} n} and the training label dataset L = {L1, L2, ..., L n Together, they constitute the training dataset.
[0039] Step 2: Construct a deep learning network model and utilize the training sample dataset S = {S1, S2, ..., S...} n} and the training label dataset L = {L1, L2, ..., L n The deep learning network model is trained to obtain a sparse observation data assimilation computation model;
[0040] The deep learning network model adopts a fully connected deep neural network architecture, including an input layer, hidden layers, and an output layer. The input layer includes batch normalization and activation functions, and the number of neurons in the input layer is equal to the number of all observation points. The hidden layer includes batch normalization and activation functions, and the number of layers and neurons in the hidden layer is determined according to the grid size. The number of neurons in the output layer is equal to the number of grid points in the background field.
[0041] Using the training sample dataset S = {S1, S2, ..., S...} n} and the training label dataset L = {L1, L2, ..., L n The deep learning network model is trained as follows:
[0042] Step 2.1: Based on the training sample dataset S = {S1, S2, ..., S...} n} and the training label dataset L = {L1, L2, ..., L n Initialize the hyperparameters of the deep learning network model, including the learning rate and the number of training iterations; taking a 128×128 grid as an example, the learning rate can be set to 0.001 and the number of training iterations can be set to 1000.
[0043] Step 2.2: Generate the training sample dataset S = {S1, S2, ..., S...} n The data in the dataset is used as input to the deep learning network model. Adam gradient descent is used as the optimizer, MSEloss as the cost function, and ReLU as the activation function. An adaptive learning rate is used to progressively train each hyperparameter. Through multiple iterations, the optimal hyperparameters are obtained when the cost function no longer decreases. The deep learning network model with the optimal hyperparameters is then used as the assimilation model for sparse observation data. Generally, after the loss stops decreasing, approximately 500 more training iterations are needed to reach the optimal parameters. Increasing the number of training iterations makes the model output smoother.
[0044] Because all grid points in a fully connected deep neural network are interconnected, correlations between different grid points can be established through training on historical data, and nonlinear relationships between grid points can be further established through activation functions. Therefore, intelligent assimilation based on fully connected deep neural networks can achieve results surpassing traditional assimilation methods. Furthermore, the advantage of using fully connected deep neural networks lies in the model's adaptability to assimilation tasks with different observation distributions. For fully connected deep neural networks, all observation data enters the hidden layers in a flattened form from the input layer, enabling the model's adaptive function.
[0045] Step 3: Input the sparse observation data of the selected ocean region to be processed into the assimilation calculation model, and use the assimilation calculation model to obtain the analysis field after fusing the observation information.
[0046] The process of calculating and analyzing the field using the assimilation computational model includes:
[0047] Step 3.1: Subtract the data values of each sparse observation point from the corresponding background field parameter values to obtain the test sample dataset T = {T1, T2, ..., T...} n};
[0048] Step 3.2: Generate the test sample dataset T = {T1, T2, ..., T} n The input is fed into the assimilation calculation model to obtain the observation increment for the entire field;
[0049] Step 3.3: Use the CV2 tool to smooth the observation increments across the entire field.
[0050] Step 3.4: Add the smoothed full-field observation increment to the background field parameter value to obtain the analysis field parameter value:
[0051] A = E + B
[0052] Where A represents the analysis field parameter value, E represents the observation increment of the entire field, and B represents the background field parameter value.
[0053] Example
[0054] Taking the assimilation process of ocean temperature in the South China Sea on September 16, 2016 as an example, four Argo observation data were assimilated into the temperature background field. Among them, the Argo observation data came from the Hangzhou Global Ocean Argo System Field Scientific Observation and Research Station, and the ocean temperature background field and ocean history reanalysis data came from HYCOM reanalysis data products.
[0055] The specific implementation steps of the deep learning-based ocean sparse observation data assimilation method in this embodiment are as follows:
[0056] Step 1: Construct a training dataset based on ocean history reanalysis data. The specific process is as follows:
[0057] Step 1.1: Obtain the background temperature field data of the South China Sea on September 16, 2016;
[0058] Step 1.2: Obtain historical reanalysis data of temperature in the South China Sea using HYCOM;
[0059] Step 1.3: Obtain Argo observation data for the South China Sea around September 16, 2016;
[0060] Step 1.4: Extract the location information of each sparse observation point in the Argo observation data, such as... Figure 2 As shown, the parameter values of each sparse observation point location in the historical reanalysis data of the South China Sea at different times are subtracted from the corresponding parameter values of the sparse observation point locations in the background field data of September 16, 2016, to construct a training sample dataset S = {S1, S2, ..., S...}. n};
[0061] Step 1.5: Subtract the parameter values of the historical reanalysis data at the corresponding time point from the parameter values of the background field data to construct the training label dataset L = {L1, L2, ..., L...} n}
[0062] Step 1.6: Training sample dataset S = {S1, S2, ..., S...} n} and the training label dataset L = {L1, L2, ..., L n Together, they constitute the training dataset.
[0063] Step 2: Construct a deep learning network model;
[0064] The deep learning network model adopts a fully connected deep neural network architecture, including an input layer, hidden layers, and an output layer, as shown in the network structure below. Figure 3 As shown;
[0065] The input layer includes batch normalization and ReLU activation function, and the number of neurons in the input layer is equal to the number of all sparse observation points. The hidden layer also includes batch normalization and ReLU activation function. In this embodiment, the hidden layer has 8 layers, and each layer has 64 neurons. The number of neurons in the output layer is equal to the number of background field grid points.
[0066] Step 3: Using the training sample dataset S = {S1, S2, ..., S...} n} and the training label dataset L = {L1, L2, ..., L n The deep learning network model is trained to obtain a sparse observation data assimilation computation model, as follows:
[0067] Step 3.1: Based on the training dataset S = {S1, S2, ..., S...} n} and the labeled dataset L = {L1, L2, ..., L n The hyperparameters of the deep learning network model are determined; in this embodiment, the learning rate is set to 0.001 and the number of training iterations is 1000.
[0068] Step 3.2: The training dataset S = {S1, S2, ..., S...} is... n The data in} is used as input to the deep learning network model. Adam gradient descent is used as the optimizer and MSEloss loss function is used as the cost function. The hyperparameters are trained step by step using an adaptive learning rate method. Through multiple iterations, the optimal hyperparameters are obtained when the cost function no longer decreases. The deep learning network model with the optimal hyperparameters is used as the sparse observation data assimilation calculation model.
[0069] Step 4: Input the sparse observation data of the South China Sea on September 16, 2016 into the assimilation calculation model, and use the assimilation calculation model to obtain the analysis field after fusing the observation information.
[0070] The calculation process is as follows:
[0071] Step 4.1: Subtract the data values of each sparse observation point from the corresponding background field parameter values to obtain the test sample dataset T = {T1, T2, ..., T...} n};
[0072] Step 4.2: Generate the test sample dataset T = {T1, T2, ..., T} n The input is fed into the assimilation calculation model to obtain the observation increment for the entire field;
[0073] Step 4.3: Use the CV2 tool to smooth the observation increments across the entire field. The kernel function used is GaussianBlur, with a kernel size of 5×5 and a standard deviation of 0.9.
[0074] Step 4.4: Add the smoothed full-field observation increment to the background field parameter value to obtain the analysis field parameter value:
[0075] A = E + B
[0076] Where A represents the analysis field parameter value, E represents the observation increment of the entire field, and B represents the background field parameter value.
[0077] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A deep learning-based method for assimilating sparse ocean observation data, characterized in that, The specific technical solution is as follows: Step 1: Obtain background field data and instrument observation data of marine environmental parameters in the area to be processed, and construct a training sample dataset by combining the marine history reanalysis data of the area; Specifically: Extract the locations of sparse observation points from the instrument observation data. For the same sparse observation point location, calculate the difference between the parameter value at that location in the historical reanalysis data at different times and the parameter value at that location in the background field data. Construct a training sample dataset S={S1,S2,...,S...} using the differences at all sparse observation point locations. n }; Step 2: Construct a deep learning network model and train it using the training dataset to obtain a sparse observation data assimilation computation model; The deep learning network model adopts a fully connected deep neural network architecture, including an input layer, hidden layers, and an output layer. The input layer includes batch normalization and activation functions, and the number of neurons in the input layer is equal to the number of all sparse observation points. The hidden layer includes batch normalization and activation functions, and the number of layers and neurons in the hidden layer is determined according to the grid size. The number of neurons in the output layer is equal to the number of grid points in the background field. The specific steps for training a deep learning network model are as follows: Step 2.1: Initialize the hyperparameters of the deep learning network model based on the size of the training sample dataset and computing resources; Step 2.2: The training sample dataset S = {S1, S2, ..., S...} n The data in} is used as input to the deep learning network model. Adam gradient descent is used as the optimizer, MSEloss loss function is used as the cost function, ReLU is used as the activation function, and each hyperparameter is trained step by step using an adaptive learning rate method. Through multiple iterations, the optimal hyperparameters are obtained when the cost function no longer decreases. The deep learning network model corresponding to the optimal hyperparameters is used as the sparse observation data assimilation calculation model. Step 3: Input the sparse observation data of the ocean area to be processed into the assimilation calculation model to obtain the analysis field parameter values after fusing the observation information, thereby spreading the information of each observation point in the background field to the location without observation information; The steps for calculating the field parameter values are as follows: Step 3.1: Subtract the data values of each sparse observation point from the corresponding background field data parameter values to obtain the test sample dataset T={T1,T2,...,T...} n }; Step 3.2: Generate the test sample dataset T={T1,T2,...,T...} n The data is input into the assimilation calculation model to obtain the observation increments for the entire field, and then smoothed. Step 3.3: Add the smoothed full-field observation increment to the background field data parameter values to obtain the analysis field parameter values: A = E + B Where A represents the analysis field parameter value, E represents the observation increment of the entire field, and B represents the background field parameter value.
2. The deep learning-based method for assimilation of sparse ocean observation data according to claim 1, characterized in that, The hyperparameters of the deep learning network model include the learning rate and the number of training iterations.
3. The deep learning-based method for assimilation of sparse ocean observation data according to claim 1, characterized in that, The smoothing operation on the observation increments across the entire field was performed using the CV2 tool.