Ocean three-dimensional temperature field reconstruction method based on Stacking ensemble learning

Abstract

The invention relates to an ocean three-dimensional temperature field reconstruction method based on Stacking ensemble learning, and belongs to the field of ocean data processing and climate change research. According to the method, a high-precision temperature field reconstruction model is constructed by fusing multi-source sea surface observation data and historical thermohaline observation data. The method comprises the following specific steps: firstly, carrying out data preprocessing so as to integrate observation data from different sources; then, multiple base models (such as random forest, linear regression and the like) are adopted to carry out Stacking integration, and an Optuna hyper-parameter optimization technology is utilized to improve the model performance; and finally, carrying out system evaluation on the model, and verifying the reconstruction precision and robustness of the model at different depths. By means of the method, the reconstruction precision of the ocean three-dimensional temperature field can be remarkably improved, and important data support is provided for ocean dynamics, climate change evaluation and environment monitoring.

Description

Technical Field

[0001] The present invention relates to a three-dimensional ocean temperature field reconstruction method based on Stacking ensemble learning. Specifically, it aims to construct a high-precision temperature field reconstruction model using multi-source sea surface observation data and historical temperature and salinity data. Background Art

[0002] The three-dimensional ocean temperature field is a crucial foundation for studying ocean dynamics, climate change, and ecosystems. Accurately capturing the ocean temperature field not only facilitates the analysis and reconstruction of ocean circulation, energy, and material transport, but is also crucial for climate change assessments and ecosystem health monitoring. With global warming and intensified human activities, the impacts of ocean temperature changes on the environment and climate are becoming increasingly apparent. Therefore, establishing high-precision ocean temperature field models has become increasingly important.

[0003] Currently, ocean temperature observations rely primarily on a variety of methods, including satellite remote sensing technology, ship-borne observation equipment (such as CTD), and Argo floats. These observation methods each have their own advantages and disadvantages, but they all share some shortcomings. For example, satellite remote sensing can provide temperature information over a wide range, but is limited in spatial resolution and depth distribution; ship-borne equipment provides high-precision data in local areas, but its coverage is limited and its cost is high; and although Argo floats can obtain deep-layer temperature data, its distribution is uneven and its update frequency is low. These traditional observation methods are deficient in spatial coverage and temporal resolution, making it difficult to meet the requirements of high spatiotemporal resolution, resulting in an inability to fully capture the complex changes in the ocean temperature field.

[0004] In recent years, machine learning techniques have demonstrated significant potential in reconstructing ocean temperature fields. By leveraging historical observational data, machine learning models can learn the complex nonlinear characteristics of ocean temperature fields and effectively reconstruct them. However, single models often have limitations when dealing with complex nonlinear problems, such as overfitting and poor generalization. This makes it difficult to achieve ideal accuracy when reconstructing using a single model. Summary of the Invention

[0005] The purpose of this paper is to provide a three-dimensional ocean temperature field reconstruction method based on stacking ensemble learning to overcome the shortcomings of existing methods. This method aims to improve the accuracy and reliability of ocean temperature field reconstruction by integrating multi-source data and advanced machine learning techniques.

[0006] Ensemble learning methods are becoming an important means of improving reconstruction accuracy. By combining the reconstruction results of multiple models, ensemble learning can effectively reduce model variance and improve reconstruction stability and accuracy. Stacking ensemble learning strategies, in particular, further enhance the model's expressive power and reconstruction accuracy by inputting the outputs of multiple base learners into a meta-learner as new features.

[0007] This paper aims to improve the accuracy of reconstructing the three-dimensional ocean temperature field through a stacking ensemble learning strategy, integrating the advantages of multiple models. This method not only overcomes the shortcomings of traditional observation methods but also fully utilizes existing multi-source data to provide more accurate and reliable temperature field information for marine scientific research, thereby promoting the development of climate change research, marine ecological protection, and resource management.

[0008] This method is primarily used in fields such as ocean data processing, climate change research, and marine environmental monitoring. By integrating data from satellite remote sensing, buoy monitoring, and other observational methods, the present invention effectively captures the spatiotemporal variations in ocean temperature, providing a scientific basis for a deeper understanding of ocean dynamics, reconstructing climate change, and assessing the marine ecosystem. This technology not only improves the accuracy and robustness of temperature field reconstruction but also provides researchers in related fields with new data support and analytical tools, promoting the advancement of marine scientific research.

[0009] The present invention adopts the following technical solution: a method for reconstructing the three-dimensional ocean temperature field based on stacking ensemble learning, comprising the following steps:

[0010] S1. Obtain observation data collected at a daily frequency, including multi-source sea surface observation data and historical temperature and salinity observation data;

[0011] S2. Preprocessing the acquired observation data to form pattern data;

[0012] S3. The temperature field reconstruction model based on the stcaking ensemble learning of the pattern data is trained separately according to samples of different depths. The base learner adopts one of the following: random forest, linear regression, K nearest neighbor, decision tree, gradient boosting tree, and multilayer perceptron. The meta-learner adopts random forest.

[0013] S4. Reconstruct the temperature field based on samples at different depths, adjust the parameters through the Optuna hyperparameter optimization framework, and obtain the temperature field reconstruction model after training and parameter adjustment.

[0014] S5, applying the temperature field reconstruction model obtained in step S4 to each depth layer to invert and obtain a three-dimensional temperature field;

[0015] S6. Evaluate the temperature field obtained at each depth layer to confirm the performance of the temperature field reconstruction model;

[0016] S7. Superimpose the temperature fields of all depth layers obtained in step S5 to obtain a comprehensive three-dimensional temperature field.

[0017] In step S1, the observation data is collected at a frequency of not less than once a day to be suitable for the observation of dynamic ocean processes;

[0018] The multi-source sea surface observation data include sea surface temperature data obtained by satellite remote sensing, temperature and salinity profile data measured by Argo floats, as well as sea surface height, sea surface temperature, and sea surface wind field data; the historical temperature and salinity observation data include temperature and salinity profile data measured by Argo floats in the past period.

[0019] In step S2, preprocessing includes:

[0020] Data cleaning: For multi-source sea surface observation data and historical temperature and salinity observation data, extreme values ​​or outliers in the data are identified and deleted by setting thresholds to reduce the impact of noise on model training. Missing values ​​are linearly interpolated, and the time characteristics of historical temperature and salinity observation data are trigonometrically converted to ensure their periodic characteristics.

[0021] Perform data matching: Align the cleaned multi-source sea surface observation data and historical temperature and salinity observation data in time and space. Specifically, find the multi-source sea surface observation data with the corresponding time and space based on the time information of year, month, and day and the spatial information of longitude and latitude contained in the existing historical temperature and salinity observation data at different depths. Integrate these data into a sample, where the temperature data in the historical temperature and salinity data at the corresponding depth is used as the label and the remaining data is used as the feature.

[0022] Normalization processing: Use the maximum-minimum normalization method to normalize the matched sample data.

[0023] In step S3, the temperature field reconstruction model based on the Stcaking ensemble learning of the pattern data is trained separately according to samples of different depths, as follows:

[0024] For each depth sample, multiple base learners are selected. In the Stacking ensemble, multiple base learners f1, f2, ...f are first trained. K , and obtain its reconstructed output Y K =f K (X);

[0025] Construct a meta-learner that minimizes the loss function:

[0026]

[0027] in, N is the number of samples, i is the sample number, K is the number of base learners, Y i Represents the historical temperature observation data of the i-th sample; weight ω k Update using gradient descent:

[0028] η is the learning rate;

[0029] During the training process, five-fold cross-validation is used to evaluate the performance of each base learner. The output of the base learner is input into the meta-learner as a new feature. Multiple base learners and a meta-learner constitute a temperature field reconstruction model, which is used to reconstruct the temperature field within a specific time period.

[0030] In step S4, the Optuna hyperparameter optimization framework is used to adjust the hyperparameters of the temperature field reconstruction model.

[0031] In step S5, the temperature field reconstruction model obtained in step S4 is applied to each depth layer to invert and obtain a three-dimensional temperature field, including the following steps:

[0032] The multi-source sea surface observation data at different depths in the measured area are input into the trained temperature field reconstruction model to reconstruct the temperatures at different depths.

[0033] In step S6, the temperature field obtained at each depth layer is evaluated, and the root mean square error is calculated to quantitatively analyze the performance of the temperature field reconstruction model.

[0034] In step S7, the temperature fields of all depth layers obtained in step S5 are superimposed to obtain a comprehensive three-dimensional temperature field.

[0035] The ocean 3D temperature field reconstruction system based on Stacking ensemble learning includes:

[0036] The data preprocessing module is used to obtain observation data collected at a daily frequency, including multi-source sea surface observation data and historical temperature and salinity observation data; preprocess the acquired observation data to form model data;

[0037] The temperature field reconstruction model training module is used to train the temperature field reconstruction model based on the stcaking ensemble learning of the pattern data according to samples of different depths. The base learner uses one of the following: random forest, linear regression, K-nearest neighbor, decision tree, gradient boosting tree, and multi-layer perceptron. The meta-learner uses random forest. The temperature field reconstruction model based on samples of different depths is adjusted through the Optuna hyperparameter optimization framework to obtain the trained and adjusted temperature field reconstruction model.

[0038] The model performance evaluation and inversion module is used to apply the obtained temperature field reconstruction model to each depth layer to invert the three-dimensional temperature field; to evaluate the temperature field obtained at each depth layer to confirm the performance of the temperature field reconstruction model; and to superimpose the temperature fields of all depth layers obtained to obtain a comprehensive three-dimensional temperature field.

[0039] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the method for reconstructing the three-dimensional ocean temperature field based on stacking ensemble learning is implemented.

[0040] Preferably, in step S1, the frequency of multi-source sea surface observation data and historical temperature and salinity observation data is no less than once a day, so as to be suitable for reconstructing the three-dimensional temperature field;

[0041] Preferably, in steps S3 and S4, multiple base learners and meta-learners are combined to improve the reconstruction effect by optimizing the loss function. To enhance the generalization ability of the model, the Optuna framework is used to optimize hyperparameters to ensure that the model performs more stably on unseen datasets.

[0042] Preferably, in steps S5, S6, and S7, the trained stacking model is applied to different depth layers to invert the three-dimensional temperature field, and the model performance is evaluated by calculating metrics such as the root mean square error (RMSE). The temperature field results of all depth layers are superimposed to form a comprehensive three-dimensional temperature field that accurately reflects the characteristics of ocean temperature distribution and provides important data support for ocean dynamics research and climate change assessment.

[0043] The present invention has the following beneficial effects and advantages:

[0044] 1. Improved Accuracy: Utilizing a stacking ensemble learning strategy, by integrating the strengths of multiple machine learning models, the accuracy of the reconstruction of the 3D ocean temperature field is significantly improved. Ensemble learning can effectively reduce the bias and variance of a single model, thereby providing a more accurate and reliable temperature field reconstruction.

[0045] 2. By combining multi-source observational data (such as satellite remote sensing data and Argo float data), this method can monitor and capture rapid changes in ocean temperature fields in real time, particularly with significantly improved resolution of sub-mesoscale and high-frequency processes. This is crucial for understanding ocean dynamics and climate change, helping scientists better analyze changes in the marine environment.

[0046] 3. Reduced Costs: Compared to traditional observation methods (such as ship-borne equipment and fixed monitoring stations), the proposed method offers significant cost advantages. By leveraging existing multi-source data and advanced machine learning techniques, it enables high-temporal and spatial resolution temperature observations at a lower cost. This makes long-term, continuous ocean temperature monitoring more feasible, meeting the demands of modern climate research.

[0047] 4. Enhanced model generalization: By leveraging the Optuna hyperparameter optimization framework, this paper effectively tunes model hyperparameters, improving generalization and ensuring robust performance on unseen datasets. This optimization process reduces the risk of overfitting, making the model more reliable in practical applications.

[0048] 5. Supporting Decision-Making and Research: The resulting comprehensive three-dimensional temperature field not only enables better analysis of the spatial variations in ocean temperature but also provides crucial data support for ocean dynamics research and climate change assessments. This approach provides a solid foundation for scientific research, policymaking, and resource management, promoting further development in the field of marine science.

[0049] In summary, the stacking ensemble learning-based ocean three-dimensional temperature field reconstruction method proposed in this paper can significantly improve the reconstruction accuracy and scale resolution of the temperature field, providing important data support for ocean dynamics research and climate change assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a flowchart of the reconstruction method of the present invention.

[0051] Figure 2a The reconstructed 3D temperature field based on stacking ensemble learning is compared with the temperatures at 5 meters, 150 meters, and 475 meters from the buoy observation station and GLORYS reanalysis data from January 1, 2018 to May 18, 2018.

[0052] Figure 2b The three-dimensional temperature fields of January 1, April 1, July 1, and November 1, 2018, reconstructed based on Stacking ensemble learning, with depths of 5 meters, 200 meters, 500 meters, and 1000 meters. DETAILED DESCRIPTION

[0053] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0054] like Figure 1 As shown in FIG, a method for reconstructing the three-dimensional ocean temperature field based on stacking ensemble learning includes the following steps:

[0055] S1: Acquire multi-source sea surface observation data and historical temperature and salinity observation data, which are collected at a daily frequency; S2: Preprocess the acquired observation data, including spatial matching, outlier removal, linear interpolation, matching, and data normalization, to improve data quality;

[0056] S3: Train the model based on the Stcaking ensemble learning on the pattern data. The base learners include random forest, linear regression, K-nearest neighbor, decision tree, gradient boosting, and multi-layer perceptron. The meta-learner is random forest.

[0057] S4: Use the Optuna hyperparameter optimization framework to optimize hyperparameters to improve the accuracy of model reconstruction and the generalization ability of the model;

[0058] S5: Apply the stacking model obtained in step S4 to each depth layer to invert and obtain a three-dimensional temperature field; S6: Evaluate the temperature field obtained at each depth layer and calculate the root mean square error (RMSE) to quantitatively analyze the model performance;

[0059] S7: Superimpose the temperature fields of all depth layers obtained in step S5 to obtain a comprehensive three-dimensional temperature field.

[0060] In step S1, the observation data is collected at a frequency of not less than once a day to be suitable for the observation of dynamic ocean processes. The data includes sea surface temperature data obtained by satellite remote sensing, temperature and salinity profile data measured by Argo floats, and other relevant environmental variables (such as wind speed, air pressure, etc.). By integrating data from different sources, the obtained data set is ensured to have high timeliness and spatial coverage. At the same time, attention is paid to the geographical distribution of the data to ensure uniform coverage within the study area, so as to provide a solid foundation for the subsequent temperature field reconstruction;

[0061] In step S2, data preprocessing includes setting reasonable thresholds, identifying and deleting extreme values ​​or erroneous values ​​in the data, and reducing the impact of noise on model training. In addition, data matching is performed to align data from different sources in time and space to ensure data consistency. This usually involves converting timestamps and standardizing spatial coordinates, and performing linear interpolation on missing values. Finally, normalization is used to standardize the data to the same range, thereby improving the efficiency and stability of model training and ensuring the comparability of the model between different features. Among them, the temperature data in the historical temperature and salinity data is a label used to reconstruct the temperature value, and the remaining data are all features used to reconstruct the temperature variable and build a mapping relationship between features and labels.

[0062] In step S3, the model training based on Stacking ensemble learning is performed on the pattern data, and multiple base learners are selected. In the Stacking ensemble, multiple base learners f1, f2, ...fK are first trained and their reconstructed output Y is obtained. K =f K (X). Then build a meta-learner by minimizing the loss function:

[0063]

[0064] in, The meta-learner optimizes the weights ω k To adjust the influence of each base learner, and use the gradient descent method to update the weight: Where η is the learning rate, which is an important parameter that controls the update step of the model parameters. Ultimately, the meta-learner combines the output of the base learners to form a more accurate reconstruction. The base models include random forest, linear regression, K-nearest neighbor, decision tree, gradient boosting tree and multi-layer perceptron. These base learners are trained on the training set to capture the complex characteristics of the ocean temperature field. During the training process, five-fold cross-validation is used to evaluate the performance of each base learner to ensure its robustness on different data subsets. The meta-learner uses random forest, which is responsible for integrating the reconstruction results of each base learner. By inputting the output of the base learner as a new feature into the meta-learner, a more accurate and stable temperature field reconstruction model is constructed;

[0065] In step S4, the Optuna hyperparameter optimization framework is used to adjust the parameters to optimize the model's hyperparameters. This process involves defining a hyperparameter search space, determining the range of parameters that need to be optimized, such as the number of trees, learning rate, minimum number of sample leaves, etc., and automatically exploring the optimal parameter combination using algorithms such as Bayesian optimization. Through multiple experiments, the impact of each set of hyperparameter settings on model performance is evaluated, and cross-validation is used to ensure the effectiveness of each set of parameters to improve the model's generalization ability and reconstruction accuracy. This optimization process ensures that the final model performs more robustly on unseen datasets and reduces the risk of overfitting;

[0066] In step S5, the Stacking model obtained in step S4 is applied to each depth layer to invert the three-dimensional temperature field. The specific operation is: multi-source observation data is input into the trained Stacking model, and the temperature of different depth layers is reconstructed using the model. The temperature inversion process of each depth layer is carried out independently to ensure that the model can adapt to the characteristic differences of each layer. Finally, the reconstruction results of each layer are integrated to form a complete three-dimensional temperature field, which reflects the temperature distribution characteristics of each depth layer and provides a basis for subsequent analysis and research.

[0067] In step S6, the temperature field obtained at each depth layer is evaluated, and the root mean square error (RMSE) is calculated to quantitatively analyze the model performance. The specific operation includes comparing the model reconstruction value with the actual observation value to obtain the RMSE index and evaluate the reconstruction accuracy of the model at different depth layers. In addition, other evaluation indicators such as correlation coefficient and determination coefficient can also be calculated to fully understand the performance of the model. This evaluation process not only helps to identify the advantages and disadvantages of the model, but also provides a basis for subsequent improvements, ensuring that the model can effectively reflect ocean temperature changes and enhance its reliability in practical applications;

[0068] In step S7, the temperature fields of all depth layers obtained in step S5 are superimposed to obtain a comprehensive three-dimensional temperature field. This process involves integrating the temperature data from each depth layer to form a complete three-dimensional temperature distribution model. By weighting and combining the temperature data from each layer, a more accurate representation of the temperature field can be obtained. The final comprehensive model not only enables better analysis of the spatial variation characteristics of ocean temperature, but also provides important data support for ocean dynamics research and climate change assessments, laying the foundation for more in-depth scientific research and decision-making.

[0069] Embodiment 1:

[0070] S1: Obtain multi-source sea surface observation data and historical temperature and salinity observation data for the target sea area from January 1, 2005 to December 31, 2020, collected at a daily frequency, and download buoy observation station and renewable data;

[0071] S2: Trigonometric functions are applied to the time and location inputs to extract their periodic characteristics. Before model training, the training data undergoes preprocessing, including cleaning, matching, and normalization. Data cleaning removes missing values ​​and outliers from the model input and output data. For example, by comparing outliers with climatological temperatures, data exceeding two standard deviations is removed. Matching pairs the input and output data based on spatial and temporal proximity. Finally, both the input and output data are normalized to ensure rapid convergence during the modeling process.

[0072] S3: Select base learners including random forest, linear regression, decision tree, K nearest neighbor, multilayer perceptron, gradient boosting tree and meta-learner to select random forest training data and generate model;

[0073] S4: Use the Optuna framework to optimize the model's hyperparameters to improve the model's reconstruction accuracy.

[0074] S5: Apply the stacking model obtained in step S4 to the 5-meter, 150-meter, and 475-meter intervals from January 1, 2018 to May 18, 2018, and invert the three-dimensional temperature field; Figure 2a shown.

[0075] S6: Calculate the root mean square error and RMSE of the reconstructed temperature field and the buoy observation station and reproducible data respectively;

[0076] S7: The temperature fields at three depths obtained in S5 are superimposed to obtain a three-dimensional temperature field.

[0077] Through the above steps, the reconstruction of the three-dimensional ocean temperature field can be effectively achieved, and the accuracy and reliability of marine environmental data analysis can be improved.

[0078] Example 2:

[0079] The effects of the present invention will be described below with reference to the accompanying examples. The temperature results at 5 meters, 200 meters, 500 meters and 1000 meters on January 1, April 1, July 1 and November 1, 2018, reconstructed using the trained reconstruction model and sea surface observations, are as follows: Figure 2b As shown. Figure 2b It can be clearly seen that between 128°E and 150°

[0080] E. In the Kuroshio Extension region between 25°N and 40°N, temperatures at different depths on January 1, April 1, July 1, and November 1, 2018, exhibited a pattern consistent with the regional temperature characteristics. Influenced by the Kuroshio Current, the overall temperature distribution in this region decreases from low latitudes to high latitudes. Temperatures at a depth of 5 meters are significantly affected by solar radiation and seasonal variations, resulting in significant seasonal fluctuations. Summer highs can reach around 27.54°C, while winter lows are below 12.55°C. A distinct trend of warmer temperatures in the south and colder temperatures in the north persists, suggesting that the transport of warm water from the Kuroshio Current is more pronounced at lower latitudes. Temperatures at a depth of 200 meters are lower than those at 5 meters, and the seasonal warming and cooling is relatively minor, but the pattern of warmer temperatures in the south and colder temperatures in the north persists. At a depth of 500 meters, seawater temperatures no longer vary seasonally, remaining between 3.56°C and 9.56°C across the region, maintaining a high temperature pattern in the south and low temperature pattern in the north. The water temperature at a depth of 1000 meters is the lowest and most stable, maintaining between 0.56℃ and 3.56℃ throughout the year, and is almost unaffected by seasonal and surface temperature changes.

Claims

1. A three-dimensional ocean temperature field reconstruction method based on stacking ensemble learning, characterized by: The steps include: S1. Obtain observation data collected at a daily frequency, including multi-source sea surface observation data and historical temperature and salinity observation data; S2. Preprocessing the acquired observation data to form pattern data; S3. The temperature field reconstruction model based on the stcaking ensemble learning of the pattern data is trained separately according to samples of different depths. The base learner adopts one of the following: random forest, linear regression, K nearest neighbor, decision tree, gradient boosting tree, and multilayer perceptron. The meta-learner adopts random forest. S4. Reconstruct the temperature field based on samples at different depths, adjust the parameters through the Optuna hyperparameter optimization framework, and obtain the temperature field reconstruction model after training and parameter adjustment. S5, applying the temperature field reconstruction model obtained in step S4 to each depth layer to invert and obtain a three-dimensional temperature field; S6. Evaluate the temperature field obtained at each depth layer to confirm the performance of the temperature field reconstruction model; S7. Superimpose the temperature fields of all depth layers obtained in step S5 to obtain a comprehensive three-dimensional temperature field.

2. The method for reconstructing ocean three-dimensional temperature field based on stacking ensemble learning according to claim 1 is characterized in that: In step S1, the observation data is collected at a frequency of not less than once a day to be suitable for the observation of dynamic ocean processes; The multi-source sea surface observation data include sea surface temperature data obtained by satellite remote sensing, temperature and salinity profile data measured by Argo floats, as well as sea surface height, sea surface temperature, and sea surface wind field data; the historical temperature and salinity observation data include temperature and salinity profile data measured by Argo floats in the past period.

3. The method for reconstructing ocean three-dimensional temperature field based on stacking ensemble learning according to claim 1 is characterized in that: In step S2, preprocessing includes: Data cleaning: For multi-source sea surface observation data and historical temperature and salinity observation data, extreme values ​​or outliers in the data are identified and deleted by setting thresholds to reduce the impact of noise on model training. Missing values ​​are linearly interpolated, and the time characteristics of historical temperature and salinity observation data are trigonometrically converted to ensure their periodic characteristics. Perform data matching: Align the cleaned multi-source sea surface observation data and historical temperature and salinity observation data in time and space. Specifically, find the multi-source sea surface observation data with the corresponding time and space based on the time information of year, month, and day and the spatial information of longitude and latitude contained in the existing historical temperature and salinity observation data at different depths. Integrate these data into a sample, where the temperature data in the historical temperature and salinity data at the corresponding depth is used as the label and the remaining data is used as the feature. Normalization processing: Use the maximum-minimum normalization method to normalize the matched sample data.

4. The method for reconstructing ocean three-dimensional temperature field based on stacking ensemble learning according to claim 1 is characterized in that: In step S3, the temperature field reconstruction model based on the Stcaking ensemble learning of the pattern data is trained separately according to samples of different depths, as follows: For each depth sample, multiple base learners are selected. In the Stacking ensemble, multiple base learners f1, f2, ...f are first trained. K , and obtain its reconstructed output Y K =f K (X); Construct a meta-learner that minimizes the loss function: in, N is the number of samples, i is the sample number, K is the number of base learners, Y i Represents the historical temperature observation data of the i-th sample; weight ω k Update using gradient descent: η is the learning rate; During the training process, five-fold cross-validation is used to evaluate the performance of each base learner. The output of the base learner is input into the meta-learner as a new feature. Multiple base learners and a meta-learner constitute a temperature field reconstruction model, which is used to reconstruct the temperature field within a specific time period.

5. The method for reconstructing ocean three-dimensional temperature field based on stacking ensemble learning according to claim 1 is characterized in that: In step S4, the Optuna hyperparameter optimization framework is used to adjust the hyperparameters of the temperature field reconstruction model.

6. The method for reconstructing three-dimensional ocean temperature field based on stacking ensemble learning according to claim 1 is characterized in that: In step S5, the temperature field reconstruction model obtained in step S4 is applied to each depth layer to invert and obtain a three-dimensional temperature field, including the following steps: The multi-source sea surface observation data at different depths in the measured area are input into the trained temperature field reconstruction model to reconstruct the temperatures at different depths.

7. The method for reconstructing ocean three-dimensional temperature field based on stacking ensemble learning according to claim 1 is characterized in that: In step S6, the temperature field obtained at each depth layer is evaluated, and the root mean square error is calculated to quantitatively analyze the performance of the temperature field reconstruction model.

8. The method for reconstructing three-dimensional ocean temperature field based on stacking ensemble learning according to claim 1 is characterized in that: In step S7, the temperature fields of all depth layers obtained in step S5 are superimposed to obtain a comprehensive three-dimensional temperature field.

9. The ocean three-dimensional temperature field reconstruction system based on stacking ensemble learning is characterized by: include: The data preprocessing module is used to obtain observation data collected at a daily frequency, including multi-source sea surface observation data and historical temperature and salinity observation data; Preprocessing the acquired observation data to form pattern data; The temperature field reconstruction model training module is used to train the temperature field reconstruction model based on the stcaking ensemble learning of the pattern data according to samples of different depths. The base learner uses one of the following: random forest, linear regression, K-nearest neighbor, decision tree, gradient boosting tree, and multi-layer perceptron. The meta-learner uses random forest. The temperature field reconstruction model based on samples of different depths is adjusted through the Optuna hyperparameter optimization framework to obtain the trained and adjusted temperature field reconstruction model. The model performance evaluation and inversion module is used to apply the obtained temperature field reconstruction model to each depth layer to invert the three-dimensional temperature field; to evaluate the temperature field obtained at each depth layer to confirm the performance of the temperature field reconstruction model; and to superimpose the temperature fields of all depth layers obtained to obtain a comprehensive three-dimensional temperature field.

10. A computer-readable storage medium, characterized in that The storage medium stores a computer program, and when the computer program is executed by the processor, the method for reconstructing the three-dimensional ocean temperature field based on stacking ensemble learning according to any one of claims 1 to 8 is implemented.

Images