A sea surface temperature prediction method and device based on multi-source ocean parameter fusion

Abstract

The present invention provides a method and device for predicting sea surface temperature based on multi-source ocean parameter fusion, which belongs to the field of marine environment data processing technology. The method includes: obtaining initial ocean data from multiple platforms according to download requirements, such as obtaining different types of data from satellite remote sensing, etc., obtaining reduced dimension data through preprocessing and dimensionality reduction processing, and then obtaining fused data through weight setting and fusion processing, and then constructing a prediction model with an LSTM model, training with fused data and weights, and finally inputting the data to be predicted to obtain a prediction result. The present invention performs operations such as preprocessing, dimensionality reduction, and fusion on data from different sources, which can reduce the impact of quality problems of a single data source on the prediction results, enhance the stability and reliability of the data, and through multi-source ocean parameter fusion and refined weight setting strategies, combined with the LSTM model, it is possible to capture more intrinsic connections and potential laws between data, and improve the prediction accuracy of sea surface temperature compared with a single data source and traditional models.

Description

Technical Field

[0001] The present invention mainly relates to the technical field of marine environment data processing, and in particular to a sea surface temperature prediction method and device based on multi-source ocean parameter fusion. Background Art

[0002] At present, sea surface temperature prediction mainly relies on a single monitoring data source and traditional prediction models, such as satellite remote sensing data and buoy monitoring data. These data sources have the following limitations. Insufficient temporal and spatial resolution. A single data source is difficult to provide data with high temporal and spatial resolution, resulting in inaccurate prediction results. Unstable data quality. Due to factors such as equipment failure and environmental interference, the data quality of a single data source may be unstable, affecting the reliability of the prediction. Traditional prediction models also have the following limitations when processing complex and changeable marine environmental data. The ability to handle complex and changeable environments is limited. Traditional models often find it difficult to capture the complex dynamic changes and nonlinear relationships in the marine environment. Lacking the ability to mine deep-level features of data, traditional models usually cannot effectively extract and be based on deep-level features in the data, limiting the improvement of prediction accuracy. In summary, the existing sea surface temperature prediction methods have significant deficiencies in both data quality and model performance, and a more accurate and reliable prediction system is urgently needed. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a sea surface temperature prediction method and device based on multi-source ocean parameter fusion in view of the deficiencies in the prior art.

[0004] The technical solution of the present invention to solve the above technical problem is as follows: A sea surface temperature prediction method based on multi-source ocean parameter fusion comprises the following steps:

[0005] S1. Acquire different types of ocean initial data from multiple public data platforms that provide different types of ocean data according to the set download requirements, and perform data preprocessing on different types of ocean initial data respectively;

[0006] S2, performing dimensionality reduction processing on each pre-processed ocean initial data to obtain a plurality of ocean dimensionality reduction data;

[0007] S3, setting weights for each ocean dimension reduction data, and fusing the ocean dimension reduction data based on the weight parameters of each ocean dimension reduction data to obtain ocean fusion data;

[0008] S4, constructing a marine environment parameter prediction model based on the LSTM model, and training the marine environment parameter prediction model through the marine fusion data and the corresponding weights;

[0009] S5. Input the ocean data to be predicted into the trained ocean environment parameter prediction model, and output the sea surface temperature prediction result through the trained ocean environment parameter prediction model.

[0010] Another technical solution of the present invention to solve the above technical problem is as follows: A sea surface temperature prediction device based on multi-source ocean parameter fusion, comprising:

[0011] A data preprocessing module is used to obtain different types of ocean initial data from multiple public data platforms that provide different types of ocean data according to the set download requirements, and perform data preprocessing on different types of ocean initial data respectively;

[0012] A data dimension reduction module is used to perform dimension reduction processing on each pre-processed ocean initial data to obtain multiple ocean dimension reduction data;

[0013] A data fusion module is used to set weights for each ocean dimension reduction data, and to fuse the ocean dimension reduction data based on the weight parameters of each ocean dimension reduction data to obtain ocean fusion data;

[0014] A model training module is used to construct a marine environment parameter prediction model based on the LSTM model, and train the marine environment parameter prediction model through the marine fusion data and the corresponding weights;

[0015] The prediction output module is used to input the ocean data to be predicted into the trained ocean environment parameter prediction model, and output the sea surface temperature prediction result through the trained ocean environment parameter prediction model.

[0016] Another technical solution of the present invention to solve the above-mentioned technical problem is as follows: a sea surface temperature prediction device based on multi-source ocean parameter fusion, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the sea surface temperature prediction method based on multi-source ocean parameter fusion as described above is implemented.

[0017] The beneficial effects of the present invention are as follows: preprocessing, dimensionality reduction, fusion and other operations on data from different sources can reduce the impact of quality problems of a single data source on the prediction results, enhance the stability and reliability of the data, and through multi-source ocean parameter fusion and refined weight setting strategies, combined with the LSTM model, more intrinsic connections and potential laws between data can be captured, which improves the prediction accuracy of sea surface temperature compared with a single data source and traditional models. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A schematic diagram of a flow chart of a sea surface temperature prediction method provided by an embodiment of the present invention;

[0019] Figure 2 Another schematic diagram of a flow chart of a method for predicting sea surface temperature provided by an embodiment of the present invention;

[0020] Figure 3 A block diagram of the functional modules of the sea surface temperature prediction device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0021] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0022] In view of the shortcomings of the prior art, the present invention aims to provide a high-precision sea surface temperature prediction method and device that integrates multi-source marine environmental parameters and a refined weight adjustment strategy, so as to overcome the problem of insufficient sea surface temperature prediction accuracy in the prior art, improve the accuracy and reliability of the prediction, and provide strong technical support for the fields of marine scientific research, environmental protection, fishery resource management, etc. The present invention is described in detail below through specific embodiments.

[0023] Example 1: Figure 1 As shown, an embodiment of the present invention provides a sea surface temperature prediction method based on multi-source ocean parameter fusion, comprising the following steps:

[0024] S1. Acquire different types of ocean initial data from multiple public data platforms that provide different types of ocean data according to the set download requirements, and perform data preprocessing on different types of ocean initial data respectively;

[0025] S2, performing dimensionality reduction processing on each pre-processed ocean initial data to obtain a plurality of ocean dimensionality reduction data;

[0026] S3, setting weights for each ocean dimension reduction data, and fusing the ocean dimension reduction data based on the weight parameters of each ocean dimension reduction data to obtain ocean fusion data;

[0027] S4, constructing a marine environment parameter prediction model based on the LSTM model, and training the marine environment parameter prediction model through the marine fusion data and the corresponding weights;

[0028] S5. Input the ocean data to be predicted into the trained ocean environment parameter prediction model, and output the sea surface temperature prediction result through the trained ocean environment parameter prediction model.

[0029] In the above embodiments, preprocessing, dimensionality reduction, fusion and other operations are performed on data from different sources, which can reduce the impact of quality problems of a single data source (such as insufficient spatiotemporal resolution, unstable data, etc.) on the prediction results, enhance the stability and reliability of the data, and through multi-source ocean parameter fusion and refined weight setting strategies, combined with the LSTM model, it is possible to capture more intrinsic connections and potential laws between data, and improve the prediction accuracy of sea surface temperature compared to a single data source and traditional models.

[0030] Preferably, in step S1, obtaining different types of ocean initial data from a plurality of public data platforms providing different types of ocean data according to the set download requirements includes:

[0031] Acquire sea surface temperature data from the satellite remote sensing database according to the set time node, data resolution and spatial resolution;

[0032] Obtain seawater salinity data from the global ocean observation network according to the set time node and data resolution;

[0033] The chlorophyll concentration data are obtained from the European geosynchronous orbit satellite dataset according to the set time node and data resolution.

[0034] For example, during the data collection process:

[0035] For the satellite remote sensing database, the time node is set to the past three years, the data resolution is once a day, and the spatial resolution is 1 km × 1 km.

[0036] The sea surface temperature data of the sea area in the past three years were obtained from the satellite remote sensing database: The sea surface temperature data of the sea area in the past three years were obtained from public ocean remote sensing data platforms such as NASA's MODIS dataset, with a data resolution of once a day and a spatial resolution of 1 km × 1 km.

[0037] For the Global Ocean Observing Network, the time node is set to the past three years and the data resolution is every 12 hours.

[0038] The seawater salinity data for the past three years were obtained from public data platforms such as the Global Ocean Observing System (GOOS): specifically from the COOPS buoy data system of the National Oceanic and Atmospheric Administration (NOAA), where the buoy records data every 12 hours.

[0039] For the European geosynchronous orbit satellite dataset, the time node is set to the past 3 years and the data resolution is once a day.

[0040] The chlorophyll concentration data for the past three years were obtained from data sets released by organizations such as the International Oceanographic Commission (IOC): specifically, the data set comes from the European Geosynchronous Satellite (MERIS), with a data frequency of once a day, covering the entire sea area.

[0041] More detailed data such as seabed temperature and water flow speed can be obtained from underwater sensors, but due to the limited number and location of sensors, the data is relatively sparse.

[0042] In the above embodiment, there are abundant data sources, and different types of ocean data (sea surface temperature, seawater salinity, chlorophyll concentration, etc.) are obtained from multiple public data platforms. The acquisition method, time node, resolution, etc. of each data are described in detail, so that the data sources are diversified and cover various aspects of marine environmental information, providing a richer data foundation for subsequent accurate predictions.

[0043] like Figure 2 As shown, the data preparation process is introduced below.

[0044] The different types of ocean initial data include sea surface temperature data, seawater salinity data and chlorophyll concentration data;

[0045] In step S1, data preprocessing is performed on different types of ocean initial data respectively, including:

[0046] For missing values ​​or outliers in the sea surface temperature data, supplementing the missing values ​​based on an interpolation method and removing the outliers based on a mean filling method;

[0047] For missing values ​​or outliers in the seawater salinity data, supplementing the missing values ​​based on an interpolation method, and removing the outliers based on a mean filling method;

[0048] For missing values ​​or outliers in the chlorophyll concentration data, the missing values ​​are supplemented based on the interpolation method, and the outliers are removed based on the Z-score method.

[0049] For example, remove missing or outliers in sea surface temperature data (i.e., satellite remote sensing data) due to cloud cover or instrument failure: use existing data cleaning algorithms such as interpolation or mean filling to remove abnormal data in satellite remote sensing data.

[0050] Remove abnormal records in seawater salinity data (i.e., buoy data) caused by equipment failure or data transmission errors: Use existing data cleaning algorithms such as interpolation or mean filling to remove abnormal records in buoy data.

[0051] For chlorophyll concentration data (i.e., ship observation data), due to the low sampling frequency, linear interpolation was first used to supplement missing values, and then the Z-score method was used to detect and remove outliers, while obviously unreasonable observations were removed.

[0052] The difference method mentioned above uses linear interpolation to supplement missing values. The specific steps are as follows:

[0053] a. Determine the nearest valid data point before and after the missing value.

[0054] b. Calculate the linear relationship between these two valid data points.

[0055] c. Fill in missing values ​​based on linear relationships.

[0056] The specific steps of the Z-score method mentioned above are as follows:

[0057] a. Calculate the mean and standard deviation of all data points.

[0058] b. Calculate the Z score for each data point using the formula: Where x is the data point, μ is the mean, and σ is the standard deviation.

[0059] c. Remove data points with an absolute value of Z score greater than 3 and consider them as outliers.

[0060] In the above embodiments, the data quality is optimized, and the missing values ​​and outliers in the initial data of different types of oceans are processed by interpolation, mean filling, Z-score method, etc., which effectively removes the errors and unreasonable parts in the data, improves the accuracy and completeness of the data, and thus improves the data quality of the entire prediction process.

[0061] like Figure 2 As shown in the figure, the process of multi-source data fusion is introduced below.

[0062] In step S2, each pre-processed ocean initial data is subjected to dimensionality reduction processing to obtain a plurality of ocean dimensionality reduction data, including:

[0063] Each pre-processed ocean initial data is normalized based on the data normalization formula, and the normalization formula is:

[0064]

[0065] Among them, X new is the normalized value, X is one of the data in the preprocessed ocean initial data, X min and X max are the minimum and maximum values ​​in the initial ocean data after preprocessing;

[0066] Based on the PCA principal component analysis method, the dimensionality reduction processing of each ocean data after normalization is carried out, specifically:

[0067] Constructing an initial data matrix based on each ocean data after normalization, and calculating the mean and standard deviation corresponding to each ocean data in the initial data matrix;

[0068] Construct a covariance matrix based on the mean and standard deviation corresponding to each ocean data, and calculate the eigenvalue and eigenvector of the covariance matrix;

[0069] The first k principal components are selected according to the size of the eigenvalue, where k is a positive integer greater than 0, and the selected k principal components are projected onto the selected eigenvector to obtain the reduced-dimensional ocean data corresponding to the k principal components.

[0070] It will be appreciated that all data are either dimensioned in degrees Celsius (for temperature data) or dimensionless (for ratio data such as salinity, chlorophyll concentration).

[0071] Using the normalization method of min-max scaling, all data is scaled to a range of 0 to 1 for subsequent processing and analysis. Data Scaling: Data scaling refers to converting data to a specific range, usually between 0 and 1. This scaling helps eliminate dimensional differences between different features so that data can be compared and processed on the same scale. This is very important for many machine learning algorithms because it can speed up convergence and improve model performance. By scaling the data to the same scale, the model is better able to capture patterns and relationships between the data, thereby improving prediction accuracy and stability.

[0072] The principal component analysis (PCA) method is used to reduce the dimension of the preprocessed multi-source data and extract the main features of the data.

[0073] PCA achieves dimensionality reduction by calculating the covariance matrix of the data and finding the eigenvector corresponding to its largest eigenvalue.

[0074] The reduced-dimensional data are fused according to a certain proportion or weight to form a comprehensive data set containing rich information.

[0075] The data dimension reduction processing process of this embodiment is described below through a specific parameter example.

[0076] The principal component analysis (PCA) method is used to reduce the dimension of the preprocessed multi-source data and extract the main features in the data. PCA achieves dimensionality reduction by calculating the covariance matrix of the data and finding the eigenvector corresponding to its maximum eigenvalue. PCA is an existing technology widely used in data dimensionality reduction and feature extraction.

[0077] Assume that the original data has 8 dimensions (including sea surface temperature, seawater salinity, chlorophyll concentration, etc.). After PCA dimension reduction, the first 3 or 4 principal components (i.e., k values) are retained. These principal components can explain most of the information in the original data. The first 3 or 4 principal components are selected to retain as much information as possible while reducing the data dimension.

[0078] The specific steps of dimensionality reduction include:

[0079] Normalize the data: Make sure each feature has the same dimension.

[0080] Assume that there are three data sets of ocean parameters in this example: Sea Surface Temperature (SST), Seawater Salinity (SSS), and Chlorophyll Concentration (Chl_a). To simplify the example, only these three features are shown here.

[0081] The dataset X is as follows:

[0082]

[0083] Calculate the mean (μ) and standard deviation (σ) of each feature.

[0084] Normalize data:

[0085]

[0086] Calculate the covariance matrix: The covariance matrix reflects the correlation between different features.

[0087]

[0088] Solve for the eigenvalues ​​and eigenvectors of the covariance matrix:

[0089] Eigenvalue λ1=1.712, eigenvector

[0090] Eigenvalue λ2=0.284, eigenvector

[0091] Eigenvalue λ3=0.004, eigenvector

[0092] Select the first k principal components: Select the most important eigenvectors based on the size of the eigenvalues. Suppose the first two principal components are selected.

[0093] Project the data onto the selected eigenvectors:

[0094]

[0095] In the above embodiment, data redundancy and noise are reduced: the normalized ocean data is subjected to dimensionality reduction processing based on the PCA principal component analysis method, which reduces the data dimension while retaining the main features of the data, reduces data redundancy, helps to reduce the amount of calculation and speed up the convergence speed during subsequent model training, and can eliminate the interference of noise on the prediction results to a certain extent.

[0096] Unify data scale: Data normalization converts data of different dimensions and ranges to a range of 0 to 1, eliminating the dimensional differences between different features and allowing data to be compared and processed on the same scale, which helps the model better capture patterns and relationships between data, thereby improving prediction accuracy and stability.

[0097] Preferably, in step S3, weights are set for each ocean dimension reduction data, and the ocean dimension reduction data are fused based on the weight parameters of each ocean dimension reduction data to obtain ocean fused data, including:

[0098] The weights of the ocean dimension reduction data are initialized according to the set random numbers. It should be understood that an initial weight is assigned to each ocean environment parameter. In this embodiment, due to the lack of prior knowledge, a random initialization method is used to assign a random number between 0 and 1 as the initial weight to each parameter.

[0099] Based on the Bayesian optimization algorithm, the objective function is defined for each initialized ocean dimension reduction data, and the objective function is:

[0100]

[0101] Among them, y i Reduce the dimension of ocean data. is the predicted value corresponding to the ocean dimension reduction data, and n is the number of samples.

[0102] Construct an initial weight combination, initialize the initial weight combination to equal weight, and calculate the corresponding prediction value based on the initialized weight combination:

[0103]

[0104] Based on the calculated predicted value, calculate the objective function value:

[0105]

[0106] A probability model is constructed based on GPR Gaussian process regression, the objective function value and the weight relationship are fitted based on the calculated predicted value, and the next weight combination is selected based on the probability model and the fitted weight and objective function value relationship to minimize the objective function, and the process is repeated until a predetermined number of iterations is reached or a stop condition is met, thereby obtaining weight parameters for each ocean dimension reduction data;

[0107] Based on the weighted average method and the weight parameters of each ocean dimensionality reduction data, the ocean dimensionality reduction data are fused to obtain the ocean fusion data.

[0108] It should be understood that the main process of using the Bayesian optimization algorithm to fine-tune the weight of the fused data is:

[0109] First, define an objective function, such as prediction error (such as mean square error MSE), as the optimization target. Then, the optimal weight combination is selected by iteration. Each iteration will update the probability model based on the information of the current optimal solution, thereby guiding the direction of the next round of weight selection. After multiple iterations, the Bayesian optimization algorithm will output a set of optimal weight combinations, which will be used for subsequent prediction model construction.

[0110] The following is an example of a specific parameter to illustrate the process of weight adjustment.

[0111] Define the objective function: Use mean square error (MSE) as the objective function.

[0112] Initialize Bayesian optimization:

[0113] Select the initial weight combination. For example, the initial weights are equal weights:

[0114] w0=[0.33,0.33,0.34].

[0115] Iteration process:

[0116] Calculate the predicted value under the current weight:

[0117]

[0118] Calculate the objective function value (MSE):

[0119]

[0120] Update the probability model: Use Gaussian process regression (GPR) or other probability models to fit the relationship between weights and objective function values.

[0121] Select next weight combination: Based on the probability model, select the next weight combination to minimize the objective function.

[0122] Repeat the above steps until the predetermined number of iterations is reached or the stopping condition is met.

[0123] Output the optimal weight combination:

[0124] After multiple iterations, the Bayesian optimization algorithm will output a set of optimal weight combinations. For example, the optimal weight combination is:

[0125] w * =[0.45,0.30,0.25].

[0126] In the above embodiment, the weight parameters can be accurately determined, and the Bayesian optimization algorithm is used to set the weights of the ocean dimensionality reduction data. By defining the objective function (such as MSE), the optimal weight combination is iteratively selected, and the relationship between the data characteristics and the prediction target is fully considered. In this way, each ocean environmental parameter is given a more reasonable weight in the fusion process, which improves the effectiveness of the fused data and further improves the accuracy of the prediction model.

[0127] like Figure 2 As shown in the figure, the process of building and training the prediction model is introduced below.

[0128] The prediction model is built based on the long short-term memory network (LSTM). The LSTM network can capture the long-term dependencies in time series data and is suitable for the prediction of marine environmental parameters.

[0129] In this embodiment, the input layer of the LSTM network includes the fused data and the determined weights, and the output layer is the predicted sea surface temperature value.

[0130] In step S4, the ocean environment parameter prediction model is trained by the ocean fusion data and the corresponding weights, and the main process is:

[0131] The ocean fusion data and the corresponding weights are arranged into a time series according to the acquisition time nodes, and input into the ocean environment parameter prediction model for training. The ocean environment parameter prediction model is overfitted during the training process by adding regularization terms or based on the dropout algorithm, and finally the sea surface temperature prediction result is output.

[0132] It should be understood that the input and output data of the LSTM network are understood as:

[0133] The fused data and the determined weights are used as the input of the LSTM network. In order to construct the time series input, the individual data can be arranged into a sequence of multiple time steps according to the time nodes of the data collection. The output of the LSTM network is the predicted sea surface temperature value, which can be the predicted value after one or more time steps.

[0134] For model training:

[0135] The network parameters are adjusted by the back propagation algorithm to minimize the prediction error. During the training process, regularization terms can be added or techniques such as dropout can be used to prevent overfitting. The generalization ability and prediction accuracy of the model can be evaluated by methods such as cross-validation. In this embodiment, K-fold cross-validation can be used to divide the data set into K subsets, which are used in turn as validation sets for model evaluation.

[0136] Adjusting network parameters and minimizing prediction errors through the back-propagation algorithm is a standard practice in deep learning and is part of the state of the art.

[0137] Add regularization or use dropout:

[0138] Regularization term: Overfitting can be prevented by adding L1 or L2 regularization terms to the loss function. The specific steps are as follows:

[0139] L2 regularization (weight decay):

[0140]

[0141] Among them, λ is the regularization coefficient, w i is the network weight.

[0142] Dropout: Randomly discard some neurons during training to reduce the co-adaptability between neurons. The specific steps are as follows:

[0143] In each layer of neurons, some neurons are randomly selected with a certain probability p and their outputs are set to zero.

[0144] For example, setting p = 0.5 means that the neurons in each layer have a 50% probability of being discarded.

[0145] In the above embodiment, the model can be prevented from overfitting. In the process of training the LSTM model based on ocean fusion data and weights, by adding regularization terms (such as L2 regularization) or using a dropout algorithm, the co-adaptability between neurons during model training is effectively reduced, preventing the model from over-learning the details in the training data and losing the ability to generalize to new data, thereby improving the stability and prediction accuracy of the model, so that it can better adapt to different input data for sea surface temperature prediction.

[0146] like Figure 2 As shown, finally, the prediction output and verification are shown.

[0147] Prediction output:

[0148] The data within the time period to be predicted is input into the trained LSTM prediction model to obtain the predicted sea surface temperature value.

[0149] The forecast results can be output as a time series, containing the forecast value for each time step.

[0150] Preferably, after obtaining the sea surface temperature prediction result, it can also be optimized: the sea surface temperature prediction result is smoothed and filtered to eliminate possible noise and fluctuation.

[0151] The marine environmental parameter prediction model is verified below to evaluate the prediction performance of the model.

[0152] Use the validation set data to validate the prediction model and evaluate the prediction performance of the model. Common evaluation indicators include mean square error (MSE) (i.e., formula 1), root mean square error (RMSE) (i.e., formula

[0153] Formula 2), mean absolute error (MAE) (i.e. Formula 3), etc.

[0154]

[0155]

[0156]

[0157] MSE measures the average of the squared differences between the predicted values ​​and the true values. A smaller MSE indicates a better predictive performance of the model.

[0158] RMSE is the square root of MSE, which scales the error to the unit of the original data for easier interpretation. A smaller RMSE indicates a better predictive performance of the model.

[0159] MAE measures the average absolute difference between the predicted value and the true value. A smaller MAE indicates a better prediction performance of the model.

[0160] Through these evaluation indicators, we can fully understand the predictive performance of the model and make necessary adjustments and improvements accordingly.

[0161] Tune and improve the model based on the verification results. If the prediction performance is not ideal, you can adjust the structure of the LSTM network (such as increasing the number of layers, changing the number of neurons, etc.), optimize the algorithm (such as learning rate, batch size, etc.) or re-adjust the weights.

[0162] Example 2: Figure 3 As shown, another embodiment of the present invention further provides a sea surface temperature prediction device based on multi-source ocean parameter fusion, comprising:

[0163] A data preprocessing module is used to obtain different types of ocean initial data from multiple public data platforms that provide different types of ocean data according to the set download requirements, and perform data preprocessing on different types of ocean initial data respectively;

[0164] A data dimension reduction module is used to perform dimension reduction processing on each pre-processed ocean initial data to obtain multiple ocean dimension reduction data;

[0165] A data fusion module is used to set weights for each ocean dimension reduction data, and to fuse the ocean dimension reduction data based on the weight parameters of each ocean dimension reduction data to obtain ocean fusion data;

[0166] A model training module is used to construct a marine environment parameter prediction model based on the LSTM model, and train the marine environment parameter prediction model through the marine fusion data and the corresponding weights;

[0167] The prediction output module is used to input the ocean data to be predicted into the trained ocean environment parameter prediction model, and output the sea surface temperature prediction result through the trained ocean environment parameter prediction model.

[0168] Preferably, in the data fusion module, weights are set for each ocean dimension reduction data, and the ocean dimension reduction data are fused based on the weight parameters of each ocean dimension reduction data to obtain ocean fusion data, including:

[0169] Initialize the weights of each ocean dimension reduction data according to the set random numbers;

[0170] Based on the Bayesian optimization algorithm, the objective function is defined for each initialized ocean dimension reduction data, and the objective function is:

[0171]

[0172] Among them, y i Reduce the dimension of ocean data. is the predicted value corresponding to the ocean dimension reduction data, and n is the number of samples;

[0173] Construct an initial weight combination, initialize the initial weight combination to equal weight, and calculate the corresponding prediction value based on the initialized weight combination:

[0174]

[0175] Based on the calculated predicted value, calculate the objective function value:

[0176]

[0177] A probability model is constructed based on GPR Gaussian process regression, the objective function value and the weight relationship are fitted based on the calculated predicted value, and the next weight combination is selected based on the probability model and the fitted weight and objective function value relationship to minimize the objective function, and the process is repeated until a predetermined number of iterations is reached or a stop condition is met, thereby obtaining weight parameters for each ocean dimension reduction data;

[0178] Based on the weighted average method and the weight parameters of each ocean dimensionality reduction data, the ocean dimensionality reduction data are fused to obtain the ocean fusion data.

[0179] Preferably, in the model training module, the ocean environment parameter prediction model is trained by using the ocean fusion data and the weights corresponding thereto, including:

[0180] The ocean fusion data and the corresponding weights are arranged into a time series according to the acquisition time nodes, and input into the ocean environment parameter prediction model for training. The ocean environment parameter prediction model is overfitted during the training process by adding regularization terms or based on the dropout algorithm, and finally the sea surface temperature prediction result is output.

[0181] The advantages of the present invention over the prior art are: preprocessing, dimension reduction, fusion and other operations on data from different sources can reduce the impact of single data source quality problems (such as insufficient spatiotemporal resolution, unstable data, etc.) on the prediction results, enhance the stability and reliability of the data, and through multi-source ocean parameter fusion and refined weight setting strategy, combined with the LSTM model, it can capture more intrinsic connections and potential laws between data, and improve the prediction accuracy of sea surface temperature compared to single data source and traditional models. In addition, the present invention is not only applicable to the prediction of sea surface temperature, but can also be extended to the prediction of other marine environmental parameters, providing more extensive technical support for the fields of marine scientific research, environmental protection, and fishery resource management.

[0182] Embodiment 3: Another embodiment of the present invention also provides a sea surface temperature prediction device based on multi-source ocean parameter fusion, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the sea surface temperature prediction method based on multi-source ocean parameter fusion as described above is implemented.

[0183] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described devices and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0184] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0185] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the embodiments of the present invention.

[0186] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A sea surface temperature prediction method based on multi-source ocean parameter fusion, characterized in that: The steps include: S1. Acquire different types of ocean initial data from multiple public data platforms that provide different types of ocean data according to the set download requirements, and perform data preprocessing on different types of ocean initial data respectively; S2, performing dimensionality reduction processing on each pre-processed ocean initial data to obtain a plurality of ocean dimensionality reduction data; S3, setting weights for each ocean dimension reduction data, and fusing the ocean dimension reduction data based on the weight parameters of each ocean dimension reduction data to obtain ocean fusion data; S4, constructing a marine environment parameter prediction model based on the LSTM model, and training the marine environment parameter prediction model through the marine fusion data and the corresponding weights; S5. Input the ocean data to be predicted into the trained ocean environment parameter prediction model, and output the sea surface temperature prediction result through the trained ocean environment parameter prediction model.

2. The sea surface temperature prediction method according to claim 1, characterized in that: In step S1, different types of ocean initial data are obtained from multiple public data platforms providing different types of ocean data according to the set download requirements, including: Acquire sea surface temperature data from the satellite remote sensing database according to the set time node, data resolution and spatial resolution; Obtain seawater salinity data from the global ocean observation network according to the set time node and data resolution; The chlorophyll concentration data are obtained from the European geosynchronous orbit satellite dataset according to the set time node and data resolution.

3. The sea surface temperature prediction method according to claim 1, characterized in that: The different types of ocean initial data include sea surface temperature data, seawater salinity data and chlorophyll concentration data; In step S1, data preprocessing is performed on different types of ocean initial data respectively, including: For missing values ​​or outliers in the sea surface temperature data, supplementing the missing values ​​based on an interpolation method and removing the outliers based on a mean filling method; For missing values ​​or outliers in the seawater salinity data, supplementing the missing values ​​based on an interpolation method, and removing the outliers based on a mean filling method; For missing values ​​or outliers in the chlorophyll concentration data, the missing values ​​are supplemented based on the interpolation method, and the outliers are removed based on the Z-score method.

4. The sea surface temperature prediction method according to claim 2, characterized in that: In step S2, each pre-processed ocean initial data is subjected to dimensionality reduction processing to obtain a plurality of ocean dimensionality reduction data, including: Each pre-processed ocean initial data is normalized based on the data normalization formula, and the normalization formula is: Among them, X new is the normalized value, X is one of the data in the preprocessed ocean initial data, X min and X max are the minimum and maximum values ​​in the initial ocean data after preprocessing; Based on the PCA principal component analysis method, the dimensionality reduction processing of each ocean data after normalization is carried out, specifically: Constructing an initial data matrix based on each ocean data after normalization, and calculating the mean and standard deviation corresponding to each ocean data in the initial data matrix; Construct a covariance matrix based on the mean and standard deviation corresponding to each ocean data, and calculate the eigenvalue and eigenvector of the covariance matrix; The first k principal components are selected according to the size of the eigenvalue, where k is a positive integer greater than 0, and the selected k principal components are projected onto the selected eigenvector to obtain the reduced-dimensional ocean data corresponding to the k principal components.

5. The sea surface temperature prediction method according to claim 1, characterized in that: In step S3, weights are set for each ocean dimension reduction data, and the ocean dimension reduction data are fused based on the weight parameters of each ocean dimension reduction data to obtain ocean fusion data, including: Initialize the weights of each ocean dimension reduction data according to the set random numbers; Based on the Bayesian optimization algorithm, the objective function is defined for each initialized ocean dimension reduction data, and the objective function is: Among them, y i Reduce the dimension of ocean data. is the predicted value corresponding to the ocean dimension reduction data, and n is the number of samples; Construct an initial weight combination, initialize the initial weight combination to equal weight, and calculate the corresponding prediction value based on the initialized weight combination: Based on the calculated predicted value, calculate the objective function value: A probability model is constructed based on GPR Gaussian process regression, the objective function value and the weight relationship are fitted based on the calculated predicted value, and the next weight combination is selected based on the probability model and the fitted weight and objective function value relationship to minimize the objective function, and the process is repeated until a predetermined number of iterations is reached or a stop condition is met, thereby obtaining weight parameters for each ocean dimension reduction data; Based on the weighted average method and the weight parameters of each ocean dimensionality reduction data, the ocean dimensionality reduction data are fused to obtain the ocean fusion data.

6. The sea surface temperature prediction method according to claim 1, characterized in that: In the step S4, the marine environment parameter prediction model is trained by using the marine fusion data and the weights corresponding thereto, including: The ocean fusion data and the corresponding weights are arranged into a time series according to the acquisition time nodes, and input into the ocean environment parameter prediction model for training. The ocean environment parameter prediction model is overfitted during the training process by adding regularization terms or based on the dropout algorithm, and finally the sea surface temperature prediction result is output.

7. A sea surface temperature prediction device based on multi-source ocean parameter fusion, characterized in that: include: A data preprocessing module is used to obtain different types of ocean initial data from multiple public data platforms that provide different types of ocean data according to the set download requirements, and perform data preprocessing on different types of ocean initial data respectively; A data dimension reduction module is used to perform dimension reduction processing on each pre-processed ocean initial data to obtain multiple ocean dimension reduction data; A data fusion module is used to set weights for each ocean dimension reduction data, and to fuse the ocean dimension reduction data based on the weight parameters of each ocean dimension reduction data to obtain ocean fusion data; A model training module is used to construct a marine environment parameter prediction model based on the LSTM model, and train the marine environment parameter prediction model through the marine fusion data and the corresponding weights; The prediction output module is used to input the ocean data to be predicted into the trained ocean environment parameter prediction model, and output the sea surface temperature prediction result through the trained ocean environment parameter prediction model.

8. The sea surface temperature prediction device according to claim 7, characterized in that: In the data fusion module, weights are set for each ocean dimension reduction data, and the ocean dimension reduction data are fused based on the weight parameters of each ocean dimension reduction data to obtain ocean fusion data, including: Initialize the weights of each ocean dimension reduction data according to the set random numbers; Based on the Bayesian optimization algorithm, the objective function is defined for each initialized ocean dimension reduction data, and the objective function is: Among them, y i Reduce the dimension of ocean data. is the predicted value corresponding to the ocean dimension reduction data, and n is the number of samples; Construct an initial weight combination, initialize the initial weight combination to equal weight, and calculate the corresponding prediction value based on the initialized weight combination: Based on the calculated predicted value, calculate the objective function value: A probability model is constructed based on GPR Gaussian process regression, the objective function value and the weight relationship are fitted based on the calculated predicted value, and the next weight combination is selected based on the probability model and the fitted weight and objective function value relationship to minimize the objective function, and the process is repeated until a predetermined number of iterations is reached or a stop condition is met, thereby obtaining weight parameters for each ocean dimension reduction data; Based on the weighted average method and the weight parameters of each ocean dimensionality reduction data, the ocean dimensionality reduction data are fused to obtain the ocean fusion data.

9. The sea surface temperature prediction device according to claim 7, characterized in that: In the model training module, the ocean environment parameter prediction model is trained by using the ocean fusion data and the weights corresponding thereto, including: The ocean fusion data and the corresponding weights are arranged into a time series according to the acquisition time nodes, and input into the ocean environment parameter prediction model for training. The ocean environment parameter prediction model is overfitted during the training process by adding regularization terms or based on the dropout algorithm, and finally the sea surface temperature prediction result is output.

10. A sea surface temperature prediction device based on multi-source ocean parameter fusion, 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 sea surface temperature prediction method based on multi-source ocean parameter fusion as described in any one of claims 1 to 6.

Images