A method for classifying ocean element profiles based on a clustering analysis algorithm

Abstract

The application discloses a kind of ocean element profile classification method based on clustering analysis algorithm, belong to marine data processing technical field, for ocean element profile classification, including obtaining original ocean element profile sample and pre-processing and standardization;Covariance matrix is constructed to standardization profile, and principal component is selected using cumulative variance contribution rate, and feature vector is constructed;Expectation maximization algorithm is used to estimate Gaussian mixture model parameter;Classification result is verified based on bayesian information criterion and physical interpretability;Ocean structure partition or water mass identification is carried out based on classification result.The application combines principal component analysis with Gaussian mixture model, significantly reduces the calculation complexity while retaining the main vertical variation information of the profile, does not need artificial preset label, has strong adaptability, interpretability and generalizability, can objectively and efficiently complete ocean element profile automatic classification, and is suitable for water mass identification, acoustic environment assessment and other various marine applications.

Description

Technical Field

[0001] This invention discloses a method for classifying marine element profiles based on clustering analysis algorithms, belonging to the field of marine data processing technology. Background Technology

[0002] Ocean profiles are a fundamental data format describing the vertical structure of ocean water, recording the variations of physical quantities such as temperature, salinity, density, and sound speed with depth. With the continuous development of global ocean observation systems, the sources of ocean profile data are becoming increasingly abundant.

[0003] Since its deployment in 2000, the Argo global ocean observation network has accumulated millions of temperature and salinity profiles. Shipborne CTD observations, ocean gliders, and underwater moorings are continuously expanding the spatiotemporal coverage of observations. Drop-in detection equipment such as XBT and XCTD have provided a wealth of temperature and salinity profile data. Faced with the massive and ever-growing amount of profile data, traditional manual classification methods are no longer sufficient, and there is an urgent need to develop efficient and objective automatic classification technologies.

[0004] Existing ocean profile classification methods mostly rely on empirical thresholds, manual judgment, or classification methods based on a small number of feature parameters. While these methods have a certain degree of physical interpretability, they often suffer from problems such as strong subjectivity, limited regional adaptability, difficulty in handling large-scale data, and difficulty in quantifying classification uncertainty. Summary of the Invention

[0005] The purpose of this invention is to provide a marine feature profile classification method based on clustering analysis algorithm, so as to solve the problems in the existing technology that marine profile classification relies on human experience, has high computational complexity, is difficult to characterize classification uncertainty, and is difficult to adapt to large-scale data processing.

[0006] A method for classifying marine feature profiles based on clustering analysis algorithms, comprising: S1. Determine the sea area to be classified and obtain the original marine element profile samples. Perform quality control, missing value processing and vertical interpolation on the profile samples to obtain the preprocessed profile dataset. S2. For each profile sample in the preprocessed profile dataset, perform standardization processing on each pressure layer dimension to obtain standardized profile samples, and construct a standardized profile dataset. S3. Construct a covariance matrix based on the standardized profile dataset, solve for the eigenvalues ​​and eigenvectors, select principal components according to the cumulative variance contribution rate, project the standardized profile samples onto the feature space spanned by the principal components, obtain the feature vector of each profile sample, and construct the feature dataset. S4. Set the range of candidate values ​​for the number of classes in the Gaussian mixture model, and use the expectation-maximization algorithm on the feature dataset to estimate the model parameters to obtain the trained Gaussian mixture model. S5. Based on the Bayesian information criterion and physical interpretability verification, determine the optimal number of categories, and calculate the posterior probability vector and category label for each profile sample according to the Gaussian mixture model corresponding to the optimal number of categories. S6. Output the category label and posterior probability vector for each profile sample, output the mean profile, covariance structure, sample number, spatial distribution and temporal distribution of each category, and perform marine structure partitioning or water mass identification for the sea area to be classified.

[0007] S1 includes determining the sea area to be classified, obtaining raw marine element profile samples within a predetermined time range, and performing quality control to obtain cleaned profile samples. The cleaned profile samples are vertically interpolated to a uniform pressure layer or depth layer to obtain interpolated profile samples, and a preprocessed profile dataset is constructed.

[0008] S2 includes calculating the mean of the current dimension across all profile samples for each pressure layer dimension of each profile sample; Calculate the standard deviation of the current dimension across all profile samples; Standardize each dimension of each profile sample by subtracting the mean of the current dimension and then dividing by the standard deviation of the current dimension to obtain standardized profile samples and construct a standardized profile dataset.

[0009] S3 includes constructing the covariance matrix of the standardized profile samples, solving for the eigenvalues ​​and eigenvectors, and calculating the preform based on the eigenvalues. The cumulative variance contribution rate of each principal component is used. If the cumulative variance contribution rate is greater than or equal to a preset variance contribution rate threshold, the first [component] is selected. Each principal component projects the standardized profile sample onto the front. The principal components are formed We obtain the feature vector of each profile sample in the 3D feature space and construct the feature dataset.

[0010] S4 includes, S4.1, setting the number of classes in the Gaussian mixture model. The range of candidate values; The feature vector of each profile sample in the feature dataset Considered to be from The probability density function of a mixture of observations distributed in a multidimensional Gaussian distribution. for: ; In the formula, For a given category The weight, , For the first The mean vector of the class. For the first The covariance matrix of the class, It is a multidimensional Gaussian probability density function. for The index; S4 includes S4.2, which uses the expectation-maximization algorithm to iteratively estimate the parameters for each category. ; After each iteration, the log-likelihood function is calculated; The iteration terminates when the log-likelihood function converges or reaches the preset number of iterations, resulting in a trained Gaussian mixture model.

[0011] S5 includes, S5.1, the number of categories within the candidate value range. Establish a Gaussian mixture model and calculate the Bayesian information criterion score. : ; ; In the formula, For the number of categories The log-likelihood value of the model at that time. The number of samples participating in the evaluation. This represents the number of independent parameters in the model.

[0012] S5 includes S5.2, which defines the statistical criterion as selecting the option that optimizes the Bayesian information criterion score or places it within the optimal range. The value is used as the number of candidate categories in the statistics; Perform physical interpretability verification on the classification results corresponding to the number of candidate categories. If the physical interpretability verification is passed, determine the final optimal number of categories. ; according to For each profile sample's feature vector, calculate the posterior probability of belonging to each category to obtain the posterior probability vector of the current profile sample. Assign the current profile sample to the category with the highest posterior probability to obtain the category label of the current profile sample.

[0013] S6 includes, S6.1, outputting the category label for each profile sample; Output the posterior probability vector of each profile sample across all classes.

[0014] S6 includes, in S6.2, for each category, extracting interpolated profile samples belonging to the current category based on the category affiliation label, including: The mean value is calculated based on the standard pressure layer to obtain the mean value profile for each category; Calculate the covariance matrix between each profile sample and the mean profile within the same category to obtain the covariance structure; Count the number of profile samples included in each category; Geographic distribution maps and time series maps of profile samples of each category were drawn to obtain spatial and temporal distribution characteristics; S6 includes S6.3, which outputs the mean profile, covariance structure, number of profile samples, spatial distribution characteristics, and temporal distribution characteristics for each category.

[0015] S6 includes S6.4, which performs marine structure partitioning or water mass identification on the sea area to be classified based on the category attribution label, posterior probability vector and mean profile, covariance structure, number of profile samples included, spatial distribution characteristics and temporal distribution characteristics of each category.

[0016] Compared with existing technologies, this invention has the following advantages: By combining principal component analysis with Gaussian mixture models, this invention significantly reduces computational complexity while preserving the main vertical variation information of oceanographic feature profiles; it requires no manual pre-setting of labels, is entirely data-driven, and has strong adaptability and objectivity; the method can handle multiple elements such as temperature, salinity, and sound speed, is compatible with multiple data sources such as Argo buoys, ship-based CTDs, gliders, and model outputs, and can quantify classification uncertainty through posterior probability; the final output results, such as the mean profile, covariance structure, and spatial distribution, have good physical interpretability and can be widely applied in fields such as ocean water mass identification, shelf structure classification, acoustic environment assessment, and marine environmental monitoring. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0019] A method for classifying marine feature profiles based on clustering analysis algorithms, comprising: S1. Determine the sea area to be classified and obtain the original marine element profile samples. Perform quality control, missing value processing and vertical interpolation on the profile samples to obtain the preprocessed profile dataset. S2. For each profile sample in the preprocessed profile dataset, perform standardization processing on each pressure layer dimension to obtain standardized profile samples, and construct a standardized profile dataset. S3. Construct a covariance matrix based on the standardized profile dataset, solve for the eigenvalues ​​and eigenvectors, select principal components according to the cumulative variance contribution rate, project the standardized profile samples onto the feature space spanned by the principal components, obtain the feature vector of each profile sample, and construct the feature dataset. S4. Set the range of candidate values ​​for the number of classes in the Gaussian mixture model, and use the expectation-maximization algorithm on the feature dataset to estimate the model parameters to obtain the trained Gaussian mixture model. S5. Based on the Bayesian information criterion and physical interpretability verification, determine the optimal number of categories, and calculate the posterior probability vector and category label for each profile sample according to the Gaussian mixture model corresponding to the optimal number of categories. S6. Output the category label and posterior probability vector for each profile sample, output the mean profile, covariance structure, sample number, spatial distribution and temporal distribution of each category, and perform marine structure partitioning or water mass identification for the sea area to be classified.

[0020] S1 includes determining the sea area to be classified, obtaining raw marine element profile samples within a predetermined time range, and performing quality control to obtain cleaned profile samples. The cleaned profile samples are vertically interpolated to a uniform pressure layer or depth layer to obtain interpolated profile samples, and a preprocessed profile dataset is constructed.

[0021] S2 includes calculating the mean of the current dimension across all profile samples for each pressure layer dimension of each profile sample; Calculate the standard deviation of the current dimension across all profile samples; Standardize each dimension of each profile sample by subtracting the mean of the current dimension and then dividing by the standard deviation of the current dimension to obtain standardized profile samples and construct a standardized profile dataset.

[0022] S3 includes constructing the covariance matrix of the standardized profile samples, solving for the eigenvalues ​​and eigenvectors, and calculating the preform based on the eigenvalues. The cumulative variance contribution rate of each principal component is used. If the cumulative variance contribution rate is greater than or equal to a preset variance contribution rate threshold, the first [component] is selected. Each principal component projects the standardized profile sample onto the front. The principal components are formed We obtain the feature vector of each profile sample in the 3D feature space and construct the feature dataset.

[0023] S4 includes, S4.1, setting the number of classes in the Gaussian mixture model. The range of candidate values; The feature vector of each profile sample in the feature dataset Considered to be from The probability density function of a mixture of observations distributed in a multidimensional Gaussian distribution. for: ; In the formula, For a given category The weight, , For the first The mean vector of the class. For the first The covariance matrix of the class, It is a multidimensional Gaussian probability density function. for The index; S4 includes S4.2, which uses the expectation-maximization algorithm to iteratively estimate the parameters for each category. ; After each iteration, the log-likelihood function is calculated; The iteration terminates when the log-likelihood function converges or reaches the preset number of iterations, resulting in a trained Gaussian mixture model.

[0024] S5 includes, S5.1, the number of categories within the candidate value range. Establish a Gaussian mixture model and calculate the Bayesian information criterion score. : ; ; In the formula, For the number of categories The log-likelihood value of the model at that time. The number of samples participating in the evaluation. This represents the number of independent parameters in the model.

[0025] S5 includes S5.2, which defines the statistical criterion as selecting the option that optimizes the Bayesian information criterion score or places it within the optimal range. The value is used as the number of candidate categories in the statistics; Perform physical interpretability verification on the classification results corresponding to the number of candidate categories. If the physical interpretability verification is passed, determine the final optimal number of categories. ; according to For each profile sample's feature vector, calculate the posterior probability of belonging to each category to obtain the posterior probability vector of the current profile sample. Assign the current profile sample to the category with the highest posterior probability to obtain the category label of the current profile sample.

[0026] S6 includes, S6.1, outputting the category label for each profile sample; Output the posterior probability vector of each profile sample across all classes.

[0027] S6 includes, in S6.2, for each category, extracting interpolated profile samples belonging to the current category based on the category affiliation label, including: The mean value is calculated based on the standard pressure layer to obtain the mean value profile for each category; Calculate the covariance matrix between each profile sample and the mean profile within the same category to obtain the covariance structure; Count the number of profile samples included in each category; Geographic distribution maps and time series maps of profile samples of each category were drawn to obtain spatial and temporal distribution characteristics; S6 includes S6.3, which outputs the mean profile, covariance structure, number of profile samples, spatial distribution characteristics, and temporal distribution characteristics for each category.

[0028] S6 includes S6.4, which performs marine structure partitioning or water mass identification on the sea area to be classified based on the category attribution label, posterior probability vector and mean profile, covariance structure, number of profile samples included, spatial distribution characteristics and temporal distribution characteristics of each category.

[0029] The marine elements of this invention include one or more of temperature, salinity, density, sound speed, dissolved oxygen, and chlorophyll; the sound speed profile can be obtained by direct measurement from a sound speed sensor, or it can be calculated from temperature, salinity, and pressure using empirical formulas; the sources of marine element profile data include one or more of Argo buoys, shipborne CTD, ocean gliders, underwater moorings, XBT, XCTD, high-resolution ocean numerical model output, satellite inversion profiles, and ocean reanalysis data products.

[0030] The quality control of this invention includes screening profile samples marked as valid (e.g., ARGO marker 1 or 0), and simultaneously performing physical plausibility checks on key elements (temperature, salinity, pressure) in the profiles, eliminating those with obvious physical anomalies such as temperatures below -2.5℃ or above 40℃, or salinity below 0 or above 40 psu. For a small number of missing values ​​in the middle section (missing rate less than or equal to 10%, and the number of consecutively missing pressure layers not exceeding 5), a linear interpolation method based on adjacent pressure layers is used to fill in the missing values.

[0031] This invention ensures that the Bayesian information criterion score is within the optimal range. The value is used as the statistical candidate category number. The optimal interval is defined as: the set of all candidate category numbers whose BIC increment ΔBIC ≤ 2 compared to the global minimum BIC; or the region where the decreasing trend of BIC with increasing category number significantly slows down. Finally, the model with the fewest categories within this interval is selected as the optimal solution.

[0032] The cumulative variance contribution rate threshold of this invention is 99% to 99.9%; the expected maximization algorithm is used to iteratively estimate the parameters of each category, including: (a) the expectation step: based on the current parameter estimate, calculate the posterior probability of each sample belonging to each category; (b) the maximization step: update the mean vector of each category according to the posterior probability. Covariance matrix and mixed weights .

[0033] The physical interpretability verification of this invention includes any one or a combination of the following: comparing the mean profiles of each category with known water mass characteristics; comparing the spatial distribution of each category with existing artificial classification standards; analyzing whether the newly added category has independent physical meaning or whether it leads to a large amount of overlap between existing categories after increasing the number of categories; analyzing whether significantly different physical structures are merged after decreasing the number of categories.

[0034] The classification results and category feature information output by this invention include: the category affiliation label of each sample; the posterior probability vector of each sample in all categories, used to characterize the classification uncertainty; the mean profile, variance structure, number of samples, spatial distribution characteristics and temporal distribution characteristics of each category; and the marine structure zoning or water mass identification of the study area based on the classification results.

[0035] The theoretical basis of the Gaussian mixture model is the assumption that the observed data is generated by a mixture of K multidimensional Gaussian distributions. The fitting accuracy and classification effect of the model depend on the following basic conditions: (1) The data as a whole can be reasonably fitted by multiple Gaussian components, that is, the overall distribution of the data can be approximated as a weighted superposition of multiple Gaussian distributions. (2) There is a certain degree of distinguishability between the Gaussian components, that is, different categories have different mean and / or covariance structures in the feature space.

[0036] The following description, in conjunction with the accompanying drawings and embodiments, further illustrates the process of this invention. Figure 1 As shown, first, marine element profile data is input, then principal component analysis (PCA) is performed for dimensionality reduction. The dimensionality reduction results are then subjected to unsupervised clustering using a Gaussian mixture model. The clustering results are scored using the BIC statistical criterion and checked for physical interpretability. The optimal number of categories is determined, and the classification results (labels, posterior probability vectors) and physical label interpretations are output respectively.

[0037] This invention classifies Argo buoy temperature profile data for a certain sea area to identify the main vertical temperature structure types. Temperature profile data was acquired from 2001 to 2024 (i.e., a predetermined time range). The raw data includes the temperature values ​​and corresponding mass markers for each profile at each sampling depth.

[0038] Quality control of the profile data: Temperature values ​​marked as excellent are selected, and profiles containing a large amount of questionable or missing data are removed. Specifically, the preset threshold for the proportion of missing values ​​is [value missing]. The proportion of missing values ​​removed The cross-sectional samples were examined. Simultaneously, the standard pressure layer was inspected, and the proportion of missing values ​​in all samples was eliminated. The pressure layer. After the above treatment, retain to Standard pressure layers (1m interval) within the range. For the few missing values ​​in the remaining profiles, a linear interpolation method based on adjacent pressure layers is used to fill in the missing values. Specifically, for a given profile at the [missing value] ... In the case of a missing pressure layer, this profile is used in the first... The and the first The effective temperature values ​​of each pressure layer are linearly interpolated to obtain the filling value. If multiple consecutive layers are missing, linear interpolation is performed using the nearest effective values ​​at both ends of the missing segment. Each cleaned profile is represented as a 990-dimensional temperature vector. ; In the formula, , , The symbol is for transpose; after cleaning, approximately 82,668 effective profiles are obtained.

[0039] For each pressure layer dimension of all profiles, calculate the mean of that dimension across all samples. and standard deviation For each cross section, the first... Standardize each dimension: ; in, For the first Section 1 Temperature values ​​of each pressure layer These are the standardized values. After standardization, each dimension has zero mean and unit variance. The purpose of standardization is to eliminate the influence of differences in temperature value ranges between different depth layers; for example, sea surface temperature typically varies within a certain range. to Between, and The temperature variation range at depth may only be to Without standardization, surface temperature variations will dominate subsequent principal component analysis and clustering, masking differences in deeper temperature structures.

[0040] Constructing the covariance matrix of standardized profile data : ; Because the data has been standardized, Therefore This is the correlation matrix of standardized data; Eigenvalue decomposition is performed on the covariance matrix C to obtain eigenvalues ​​and their corresponding eigenvectors. The cumulative variance contribution rate is then calculated. : ; According to the cumulative variance contribution rate, The principle of selection Each of the first six principal components is then projected onto a low-dimensional feature space spanned by the first six principal components. A Gaussian mixture model is built for the low-dimensional feature data of all cross-sections. The number of classes in the Gaussian mixture model is set. The range of candidate values ​​is to For each candidate Value, representing low-dimensional feature data as Weighted sum of multiple Gaussian distributions: ; The Expectation-Maximization (EM) algorithm is used to iteratively estimate the parameters for each category. ; The expected step (E-step) includes calculating for each sample based on the current parameter estimates. Belongs to the Posterior probability of the class: ; The maximization step (M-step) involves updating the parameters of each class based on the posterior probability; log-likelihood function for: ; Alternately execute the E-step and M-step until the change in the log-likelihood function is less than a preset convergence threshold (e.g., ...). ) or reach the preset maximum number of iterations (e.g. The iteration is terminated after (number of iterations) to obtain the trained Gaussian mixture model.

[0041] right to For each candidate value, calculate the Bayesian Information Criterion (BIC) score: ; in BIC through The term imposes a penalty on model complexity; the larger the number of classes and the more parameters, the heavier the penalty, thus achieving a balance between model fit and model complexity. BIC curve analysis shows that the BIC score in The time zone tends to stabilize or reach its optimal level; Based on the combined BIC statistical evaluation and physical interpretability verification, the optimal number of categories was selected. According to the final determined number of categories, the posterior probability of each profile belonging to each category was calculated, and the profile was assigned to the category with the highest posterior probability, thus completing the classification of all profiles.

[0042] Output the following classification results and category feature information: (1) Category label for each profile, indicating the water mass type or temperature structure type to which it belongs. (2) Posterior probability vector of each profile in all categories. This probability vector reflects the determinism of the classification: when the posterior probability of a certain category is much higher than that of other categories, it indicates that the profile can be deterministically classified into that category; when the posterior probabilities of multiple categories are close, it indicates that the profile is in the water mass transition zone. (3) Mean temperature profile and standard deviation range of each category, reflecting the typical temperature vertical structure and intra-class variation of the category. (4) Based on the classification results, match each category with the known water mass name and divide the study area into marine structure zones.

[0043] In practical applications, the distribution characteristics of low-dimensional feature data should be checked before performing Gaussian mixture model clustering. If the data exhibits severe non-normality in certain dimensions, appropriate distribution adjustment transformations (such as Box-Cox transformation, logarithmic transformation, etc.) can be considered before inputting it into the Gaussian mixture model. However, in most ocean profile classification applications, the data after standardization and dimensionality reduction by principal component analysis can already satisfy the distribution assumptions of the Gaussian mixture model quite well.

[0044] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for classifying marine element profiles based on clustering analysis algorithms, characterized in that, include: S1. Determine the sea area to be classified and obtain the original marine element profile samples. Perform quality control, missing value processing and vertical interpolation on the profile samples to obtain the preprocessed profile dataset. S2. For each profile sample in the preprocessed profile dataset, perform standardization processing on each pressure layer dimension to obtain standardized profile samples, and construct a standardized profile dataset. S3. Construct a covariance matrix based on the standardized profile dataset, solve for the eigenvalues ​​and eigenvectors, select principal components according to the cumulative variance contribution rate, project the standardized profile samples onto the feature space spanned by the principal components, obtain the feature vector of each profile sample, and construct the feature dataset. S4. Set the range of candidate values ​​for the number of classes in the Gaussian mixture model, and use the expectation-maximization algorithm on the feature dataset to estimate the model parameters to obtain the trained Gaussian mixture model. S5. Based on the Bayesian information criterion and physical interpretability verification, determine the optimal number of categories, and calculate the posterior probability vector and category label for each profile sample according to the Gaussian mixture model corresponding to the optimal number of categories. S6. Output the category label and posterior probability vector for each profile sample, output the mean profile, covariance structure, sample number, spatial distribution and temporal distribution of each category, and perform marine structure partitioning or water mass identification for the sea area to be classified.

2. The method for classifying marine element profiles based on clustering analysis algorithm according to claim 1, characterized in that, S1 includes determining the sea area to be classified, obtaining raw marine element profile samples within a predetermined time range, and performing quality control to obtain cleaned profile samples. The cleaned profile samples are vertically interpolated to a uniform pressure layer or depth layer to obtain interpolated profile samples, and a preprocessed profile dataset is constructed.

3. The method for classifying marine element profiles based on clustering analysis algorithm according to claim 2, characterized in that, S2 includes calculating the mean of the current dimension across all profile samples for each pressure layer dimension of each profile sample; Calculate the standard deviation of the current dimension across all profile samples; Standardize each dimension of each profile sample by subtracting the mean of the current dimension and then dividing by the standard deviation of the current dimension to obtain standardized profile samples and construct a standardized profile dataset.

4. The method for classifying marine element profiles based on clustering analysis algorithm according to claim 3, characterized in that, S3 includes constructing the covariance matrix of the standardized profile samples, solving for the eigenvalues ​​and eigenvectors, and calculating the preform based on the eigenvalues. The cumulative variance contribution rate of each principal component is used. If the cumulative variance contribution rate is greater than or equal to a preset variance contribution rate threshold, the first [component] is selected. Each principal component projects the standardized profile sample onto the front. The principal components are formed We obtain the feature vector of each profile sample in the 3D feature space and construct the feature dataset.

5. The method for classifying marine element profiles based on clustering analysis algorithm according to claim 4, characterized in that, S4 includes, S4.1, setting the number of classes in the Gaussian mixture model. The range of candidate values; The feature vector of each profile sample in the feature dataset Considered to be from The probability density function of a mixture of observations distributed in a multidimensional Gaussian distribution. for: ； In the formula, For a given category The weight, , For the first The mean vector of the class. For the first The covariance matrix of the class, It is a multidimensional Gaussian probability density function. for The index; S4 includes S4.2, which uses the expectation-maximization algorithm to iteratively estimate the parameters for each category. ; After each iteration, the log-likelihood function is calculated; The iteration terminates when the log-likelihood function converges or reaches the preset number of iterations, resulting in a trained Gaussian mixture model.

6. The method for classifying marine element profiles based on clustering analysis algorithm according to claim 5, characterized in that, S5 includes, S5.1, the number of categories within the candidate value range. Establish a Gaussian mixture model and calculate the Bayesian information criterion score. : ； ； In the formula, For the number of categories The log-likelihood value of the model at that time. The number of samples participating in the evaluation. This represents the number of independent parameters in the model.

7. The method for classifying marine element profiles based on clustering analysis algorithm according to claim 6, characterized in that, S5 includes S5.2, which defines the statistical criterion as selecting the option that optimizes the Bayesian information criterion score or places it within the optimal range. The value is used as the number of candidate categories in the statistics; Perform physical interpretability verification on the classification results corresponding to the number of candidate categories. If the physical interpretability verification is passed, determine the final optimal number of categories. ; according to For each profile sample's feature vector, calculate the posterior probability of belonging to each category to obtain the posterior probability vector of the current profile sample. Assign the current profile sample to the category with the highest posterior probability to obtain the category label of the current profile sample.

8. The method for classifying marine element profiles based on clustering analysis algorithm according to claim 7, characterized in that, S6 includes, S6.1, outputting the category label for each profile sample; Output the posterior probability vector of each profile sample across all classes.

9. A method for classifying marine element profiles based on clustering analysis algorithm according to claim 8, characterized in that, S6 includes, in S6.2, for each category, extracting interpolated profile samples belonging to the current category based on the category affiliation label, including: The mean value is calculated based on the standard pressure layer to obtain the mean value profile for each category; Calculate the covariance matrix between each profile sample and the mean profile within the same category to obtain the covariance structure; Count the number of profile samples included in each category; Geographic distribution maps and time series maps of profile samples of each category were drawn to obtain spatial and temporal distribution characteristics; S6 includes S6.3, which outputs the mean profile, covariance structure, number of profile samples, spatial distribution characteristics, and temporal distribution characteristics for each category.

10. A method for classifying marine element profiles based on clustering analysis algorithm according to claim 9, characterized in that, S6 includes S6.4, which performs marine structure partitioning or water mass identification of the sea area to be classified based on the category attribution label, posterior probability vector and mean profile, covariance structure, number of profile samples included, spatial distribution characteristics and temporal distribution characteristics of each category.

Images