Hybrid layer deep inversion method based on CLR-MLDNet
By constructing the CLR-MLDNet model and combining multi-source ocean observation data with a deep learning network, the problems of insufficient spatiotemporal feature capture and model stability in hybrid layer depth inversion were solved, and high-precision hybrid layer depth inversion was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA ACADEMY OF SPACE TECHNOLOGY
- Filing Date
- 2026-01-30
- Publication Date
- 2026-06-05
AI Technical Summary
Existing hybrid layer depth inversion methods suffer from insufficient capture of spatiotemporal features, poor physical consistency, and insufficient model stability, resulting in unstable inversion results and limited accuracy.
A hybrid layer depth inversion method based on CLR-MLDNet is adopted. By fusing convolutional neural network (CNN), long short-term memory network (LSTM) and residual network (ResNet), a hybrid layer depth reconstruction network is constructed. Multi-source ocean observation data are used for model training and validation, and spatiotemporal information is combined to invert the hybrid layer depth.
It improves the ability to extract spatiotemporal features, enhances the dynamic sensitivity and stability of the model, and improves the inversion accuracy and computational efficiency, making it suitable for mixed layer depth inversion tasks in global or regional sea areas.
Smart Images
Figure CN122153384A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a hybrid layer depth inversion method based on CLR-MLDNet, belonging to the fields of marine exploration technology and underwater parameter inversion technology. Specifically, it relates to a hybrid layer depth inversion method based on multi-source marine observation data and numerical model output, and particularly to a method for hybrid layer depth inversion using a CLR-MLDNet model that integrates convolutional neural network (CNN), long short-term memory network (LSTM) and residual network (ResNet). Background Technology
[0002] The mixing layer depth is a crucial parameter characterizing the thermodynamic and dynamic processes of the upper ocean, holding significant scientific importance for air-sea interactions, energy exchange, and climate prediction. In particular, the mixing layer depth is critical for studying El Niño / La Niña phenomena. As a core indicator reflecting the thermodynamic and dynamic state of the upper ocean, it forms a vital foundation for understanding, monitoring, and researching El Niño / La Niña events, and its variation characteristics provide key references for related research.
[0003] Currently, methods for obtaining the depth of the mixed layer mainly fall into four categories: 1. **Measured Profile Method:** This method uses equipment such as CTD or Argo buoys to obtain temperature-salinity profiles, defining and calculating the mixed layer depth based on temperature or density thresholds (e.g., thresholding, curvature, optimal linear fitting). This method offers high accuracy, but its quality is difficult to guarantee due to limitations in spatial coverage and temporal sampling. 2. **Reanalysis Product Estimation Method:** This method relies on numerical models and data assimilation techniques to estimate the mixed layer depth using reanalysis data. However, this method is affected by model resolution and parameterization schemes, resulting in significant local errors and uncertainties in the results. 3. **Traditional Empirical Statistical and Machine Learning Methods:** This method constructs models such as linear regression, random forests, and support vector machines based on measured data, and achieves mixed layer depth fitting through parameter tuning. These methods only consider static feature associations, making it difficult to capture spatiotemporal dynamic changes, leading to unstable inversion results and poor physical consistency. 4. **Single Artificial Intelligence Methods:** This method uses deep learning models such as convolutional neural networks and long short-term memory networks to establish a nonlinear mapping between multi-source features and the mixed layer depth. However, a single model struggles to simultaneously consider spatial structure and temporal evolution characteristics, limiting inversion accuracy and resulting in significant errors.
[0004] In summary, the measured profile method is limited by spatial coverage and temporal sampling, making it difficult to guarantee quality. The reanalysis product estimation method is affected by model resolution and parameterization scheme, resulting in significant local errors and uncertainties in the results. Traditional empirical statistics and machine learning methods only consider static feature correlations and are difficult to capture spatiotemporal dynamic changes, leading to unstable inversion results and poor physical consistency. A single model of a single artificial intelligence method is difficult to take into account both spatial structure and temporal evolution characteristics, limiting inversion accuracy and resulting in significant errors. Summary of the Invention
[0005] The technical problem solved by this invention is to overcome the shortcomings of the prior art and provide a hybrid layer depth inversion method based on CLR-MLDNet, which solves the problems of insufficient spatiotemporal feature capture, poor physical consistency and insufficient model stability of the existing hybrid layer depth inversion methods.
[0006] The technical solution of the present invention is: Firstly, a hybrid layer depth inversion method based on CLR-MLDNet, comprising: We collected multi-source sea surface observation data, climatological temperature data, and historical temperature observation profile data at a daily observation frequency. The sea surface observation data included sea surface temperature, sea surface salinity, sea surface wind speed, and sea surface height. We also calculated the mixed layer depth for each historical temperature observation profile. Each historical temperature observation profile is matched one-to-one with the corresponding multi-source sea surface observation data and climatological temperature data. Each historical temperature observation profile corresponds to a sample, which includes six parameters: sea surface temperature, sea surface salinity, sea surface height, sea surface wind speed, climatological temperature data, and mixed layer depth. All features of the sample are normalized. The ocean hybrid layer deep reconstruction network CLR-MLDNet model is constructed based on the convolutional neural network models CNN, LSTM, and ResNet framework. All samples were divided into training, testing and validation sets according to a preset ratio. Five features, namely sea surface temperature, sea surface salinity, sea surface height, sea surface wind speed and climatological temperature, were selected as inputs to the CLR-MLDNet model. The mixed layer depth data were used as labels and outputs of the CLR-MLDNet model. The CLR-MLDNet model was trained using the training set until the model converged or the maximum number of iterations was reached. Using the trained CLR-MLDNet model, the parameters of the CLR-MLDNet model are adjusted using the test set until the accuracy of the CLR-MLDNet model meets the preset requirements. The accuracy of the CLR-MLDNet model that meets the preset requirements is evaluated using the validation set. The accuracy of the CLR-MLDNet model inverting the hybrid layer is obtained. If the accuracy meets the preset accuracy requirements, the CLR-MLDNet model that meets the preset requirements is used to invert the sea surface temperature, sea surface salinity, sea surface height, sea surface wind speed and climatological temperature data of the actual scene. Otherwise, the model is retrained, tested and validated.
[0007] Furthermore, the sea surface observation data and historical temperature observation profile data are collected at a frequency of no less than once per day.
[0008] Furthermore, the calculation of the mixed layer depth value for each historical temperature observation profile includes: for each historical temperature observation profile data, linear interpolation to a resolution of 1m, and setting a temperature threshold Δ. T = 0.2℃, the temperature at a depth of 10m is denoted as T 10 Calculate the temperature at a depth of 10m or less. T and T 10 The difference Δ T 10 Arranged in ascending order of water depth, select the first Δ T 10 >Δ T The depth is denoted as the hybrid layer depth value.
[0009] Furthermore, the step of matching the spatiotemporal information of each historical temperature observation profile with the corresponding sea surface observation data and climatological temperature data one by one includes: matching historical temperature observation profile data with sea surface observation data and climatological temperature data using spatiotemporal information; based on the principle of spatiotemporal proximity, first finding the sea surface observation data of the same date for each historical temperature observation profile, and then using spatial distance to find the nearest grid as its corresponding sea surface observation data.
[0010] Furthermore, the CLR-MLDNet model includes m sets of CLR structural blocks and n fully connected layers connected in series. Each CLR structural block includes a series of convolutional layers, LSTM layers, and residual blocks. The ReLU function is used as the activation function between the layers in the CLR structural block, and a max pooling layer is added after the convolutional layer for dimensionality reduction and compression.
[0011] Furthermore, the values of m and n are adjusted using a grid search method, including: determining the range of values for m and n, building different models based on different m and n; then training each model using the same training strategy, and calculating its RMSE value on the test set as an evaluation index of model performance; finally, selecting the parameter combination with the smallest RMSE value on the validation set as the model parameters.
[0012] Furthermore, if any parameter in each of the samples is empty, then that sample is deleted.
[0013] Furthermore, the preset ratio is 5:3:2.
[0014] In a second aspect, a computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the CLR-MLDNet-based hybrid layer depth inversion method.
[0015] Thirdly, a hybrid layer depth inversion device based on CLR-MLDNet includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the hybrid layer depth inversion method based on CLR-MLDNet.
[0016] The advantages of this invention compared to the prior art are: (1) Enhanced spatiotemporal feature extraction capability: The combination of convolutional neural network and long short-term memory network structure can capture time series features and spatial distribution features simultaneously, which has higher dynamic sensitivity than traditional machine learning models.
[0017] (2) Improved stability of deep features: The introduction of residual modules effectively alleviates the gradient vanishing problem and improves the model training stability and generalization performance.
[0018] (3) High computational efficiency and strong generalizability: The model structure of this invention is modular, and the method is applicable to mixed layer depth inversion tasks at different scales in global or regional sea areas. Attached Figure Description
[0019] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a flowchart of the reconstruction method of the present invention; Figure 2 This is a model structure diagram of the hybrid layer depth inversion method based on CLR-MLDNet; Figure 3 This is a diagram of the model structure in the embodiment; Figure 4 This is a comparison between the hybrid layer depth inversion results and the actual values. Detailed Implementation
[0020] To better understand the above technical solutions, the technical solutions of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features therein are detailed descriptions of the technical solutions of the present invention, rather than limitations thereof. Where there is no conflict, the embodiments of the present invention and the technical features therein can be combined with each other. The following, in conjunction with the accompanying drawings, provides a further detailed description of the hybrid layer depth inversion method based on CLR-MLDNet provided by the embodiments of the present invention.
[0021] In the solutions provided in the embodiments of the present invention, such as Figure 1As shown, a hybrid layer depth inversion method based on multi-model fusion includes the following steps: S1: Collect multi-source sea surface observation data, climatological temperature data (CNT), and historical temperature observation profile data at a daily observation frequency. The sea surface observation data includes sea surface temperature (SST), sea surface salinity (SSS), sea surface wind speed (WSPD), and sea surface height (SSH). Calculate the mixed layer depth (MLD) for each historical temperature observation profile. S2: Match the spatiotemporal information of each historical temperature observation profile with the corresponding sea surface observation data and climatological temperature data (CNT). Each historical temperature observation profile corresponds to a sample, which includes six parameters: sea surface temperature (SST), sea surface salinity (SSS), sea surface height (SSH), sea surface wind speed (WSPD), climatological temperature data (CNT), and mixed layer depth (MLD). All features of the sample are normalized. S3: Construct the CLR-MLDNet model, a deep reconstruction network for ocean hybrid layers, based on convolutional neural network models CNN, LSTM, and the ResNet framework. Figure 2 ; S4: Divide all samples into training set, test set and validation set according to the preset ratio, and select five features, namely sea surface temperature (SST), sea surface salinity (SSS), sea surface height (SSH), sea surface wind speed (WSPD) and climatological temperature data (CNT), as input to the CLR-MLDNet model. Mixed layer depth (MLD) data is used as the label and is used as the output of the CLR-MLDNet model. The CLR-MLDNet model is trained using the training set until the model converges or reaches the maximum number of iterations. S5: Using the trained CLR-MLDNet model, use the test set to adjust the parameters of the CLR-MLDNet model until the accuracy of the CLR-MLDNet model reaches the optimal level, and obtain the optimal CLR-MLDNet model. S6: Use the validation set to evaluate the accuracy of the optimal CLR-MLDNet model and obtain the accuracy of the CLR-MLDNet model inverting the hybrid layer. If the accuracy meets the preset accuracy requirements, then use the CLR-MLDNet model that meets the preset requirements to invert the sea surface temperature, sea surface salinity, sea surface height, sea surface wind speed and climatological temperature data of the actual scene. Otherwise, retrain, test and validate.
[0022] Example: S1. Collect multi-source observational sea surface temperature (SST), sea surface salinity (SSS), sea surface height (SSH), sea surface wind speed (WSPD), and climatological temperature (CNT) data at a daily observation frequency, and calculate mixed layer depth (MLD) data using the threshold method. The observation data area in this embodiment is located in a certain sea area, and the time range is from January 1, 2005 to December 31, 2020. The SST data, SSS data, SSH data and CNT data in this embodiment are obtained from the CTD observation data of the World Ocean Database (WOD). The WSPD data are obtained from the observation data of the meteorological station in the area where the above CTD observation data is located in a certain energy meteorological big data platform. The MLD data is based on the threshold method. For each temperature profile in the acquired CTD observation data, linear interpolation is performed to a resolution of 1m. The temperature threshold ΔT = 0.2℃ is set. The temperature at a water depth of 10m is recorded as T10. The difference ΔT10 between the temperature T below a water depth of 10m and T10 is calculated. The values are arranged in ascending order of water depth, and the first depth where ΔT10 > ΔT is selected as MLD. S2. Match the spatiotemporal information of each profile with the corresponding sea surface observations and climatological temperatures one by one, with each profile corresponding to one sample; delete samples containing null values, and normalize the sample data using the maximum-minimum normalization method to obtain the dataset, which has a total of 57,071 samples. S3. Divide all samples into training set, test set and validation set according to the ratio of 5:3:2, and select SST, SSS, SSH, WSPD and CNT data as 5 features as input to CLR-MLDNet model, and MLD data as labels as output of CLR-MLDNet model. S4. Based on the CNN, LSTM, and ResNet frameworks, a CLR-MLDNet model was constructed using a sequential concatenation approach. The model was then trained using a training dataset. The parameters m and n were adjusted using a grid search method. Specifically, based on the characteristics of the dataset, the range of m was selected as {1, 2, 3, 4}, and the range of n as {1, 2, 3}, resulting in 12 possible model structures. The distribution of neurons in each layer followed an expansion-contraction structure, meaning the number of neurons first increased and then decreased with increasing depth. All epochs were set to 2000, and the Adam optimizer was used to train each model. Finally, m=2 and n=1 were determined as the optimal parameter combination. The resulting CLR-MLDNet model structure is shown below. Figure 3As shown, the data passes through the input layer, then sequentially through a Conv layer with 128 neurons, a max pooling layer, an LSTM layer with 256 LSTM modules, a ResNet layer with 512 neurons, another Conv layer with 256 neurons, a max pooling layer, an LSTM layer with 128 LSTM modules, and a ResNet layer with 64 neurons. After flattening, the data passes through a fully connected layer with 32 neurons before being output. S5. Using the trained CLR-MLDNet model and the test dataset, perform model parameter tuning using the network search method until the model accuracy reaches its optimal level, thus obtaining the optimal CLR-MLDNet model. Specifically, select the batch_size range as {32, 64, 128, 256, 512, 1024} and the lr range as {10-2, 10-3, 10-4}, forming 18 parameter combinations. Set the epoch to 2000 for each combination. Use the Adam optimizer to train the CLR-MLDNet model obtained in step S4. Finally, determine that batch_size=512 and lr=10-4 are the optimal parameter combinations, thus obtaining the optimal CLR-MLDNet model. S6. Evaluate the accuracy of the optimal model using the validation dataset, calculate RMSE and R², and base the predictions on the validation set as follows: Figure 4 As shown.
[0023] The effects of the present invention will be described below with reference to the accompanying drawings. To visually demonstrate the accuracy of the model calculation results, a scatter plot is drawn with the actual values on the x-axis and the predicted values on the y-axis, resulting in... Figure 4 The results shown indicate that, based on the model calculations, the RMSE value is 26.34 and the R² value is 0.74. The scattered points in the figure are mainly distributed around the line y = x. The results show that the CLR-MLDNet model in this invention integrates the spatial awareness of the CNN model, the temporal memory of the LSTM model, and the deep feature stability of the ResNet model. It can efficiently capture the spatiotemporal nonlinear coupling relationship of the hybrid layer depth and has a certain suitability for hybrid layer depth calculation, thus possessing high application value and promising prospects for promotion.
[0024] This invention provides a computer-readable storage medium storing computer instructions that, when executed on a computer, cause the computer to perform... Figure 1 The method described.
[0025] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.
[0026] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0027] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0028] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0029] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.
[0030] The contents not described in detail in this specification are common knowledge to those skilled in the art.
Claims
1. A hybrid layer depth inversion method based on CLR-MLDNet, characterized in that, include: We collected multi-source sea surface observation data, climatological temperature data, and historical temperature observation profile data at a daily observation frequency. The sea surface observation data included sea surface temperature, sea surface salinity, sea surface wind speed, and sea surface height. We also calculated the mixed layer depth for each historical temperature observation profile. Each historical temperature observation profile is matched one-to-one with the corresponding multi-source sea surface observation data and climatological temperature data. Each historical temperature observation profile corresponds to a sample, which includes six parameters: sea surface temperature, sea surface salinity, sea surface height, sea surface wind speed, climatological temperature data, and mixed layer depth. All features of the sample are normalized. The ocean hybrid layer deep reconstruction network CLR-MLDNet model is constructed based on the convolutional neural network models CNN, LSTM, and ResNet framework. All samples were divided into training, testing and validation sets according to a preset ratio. Five features, namely sea surface temperature, sea surface salinity, sea surface height, sea surface wind speed and climatological temperature, were selected as inputs to the CLR-MLDNet model. The mixed layer depth data were used as labels and outputs of the CLR-MLDNet model. The CLR-MLDNet model was trained using the training set until the model converged or the maximum number of iterations was reached. Using the trained CLR-MLDNet model, the parameters of the CLR-MLDNet model are adjusted using the test set until the accuracy of the CLR-MLDNet model meets the preset requirements. The accuracy of the CLR-MLDNet model that meets the preset requirements is evaluated using the validation set. The accuracy of the CLR-MLDNet model inverting the hybrid layer is obtained. If the accuracy meets the preset accuracy requirements, the CLR-MLDNet model that meets the preset requirements is used to invert the sea surface temperature, sea surface salinity, sea surface height, sea surface wind speed and climatological temperature data of the actual scene. Otherwise, the model is retrained, tested and validated.
2. The hybrid layer depth inversion method based on CLR-MLDNet according to claim 1, characterized in that, The sea surface observation data and historical temperature observation profile data are collected at a frequency of no less than once per day.
3. The hybrid layer depth inversion method based on CLR-MLDNet according to claim 1, characterized in that, The calculation of the mixed layer depth value for each historical temperature observation profile includes: for each historical temperature observation profile, linear interpolation to a resolution of 1m is performed, and a temperature threshold Δ is set. T = 0.2℃, the temperature at a depth of 10m is denoted as T 10 Calculate the temperature at a depth of 10m or less. T and T 10 The difference Δ T 10 Arranged in ascending order of water depth, select the first Δ T 10 >Δ T The depth is denoted as the hybrid layer depth value.
4. The hybrid layer depth inversion method based on CLR-MLDNet according to claim 1, characterized in that, The process of matching historical temperature observation profile data with corresponding sea surface observation data and climatological temperature data one by one according to the spatiotemporal information of each historical temperature observation profile data includes: matching historical temperature observation profile data with sea surface observation data and climatological temperature data using spatiotemporal information; firstly, finding the sea surface observation data of the same date for each historical temperature observation profile according to the principle of spatiotemporal proximity; and then finding the nearest grid as its corresponding sea surface observation data using spatial distance.
5. The hybrid layer depth inversion method based on CLR-MLDNet according to claim 1, characterized in that, The CLR-MLDNet model consists of m sets of CLR structural blocks and n fully connected layers connected in series. Each CLR structural block includes a series of convolutional layers, LSTM layers, and residual blocks. The ReLU function is used as the activation function between the layers in the CLR structural block, and a max pooling layer is added after the convolutional layer for dimensionality reduction and compression.
6. The hybrid layer depth inversion method based on CLR-MLDNet according to claim 5, characterized in that, The values of m and n are adjusted using a grid search method, including: determining the range of values for m and n, building different models based on different m and n; then training each model using the same training strategy and calculating its RMSE value on the test set as an evaluation index of model performance; finally, selecting the parameter combination with the smallest RMSE value on the validation set as the model parameters.
7. The hybrid layer depth inversion method based on CLR-MLDNet according to claim 1, characterized in that, If any of the parameters in a given sample is empty, then that sample is deleted.
8. The hybrid layer depth inversion method based on CLR-MLDNet according to claim 1, characterized in that, The preset ratio is 5:3:
2.
9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 8.
10. A hybrid layer depth inversion device based on CLR-MLDNet, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 8.