Sea surface parameter sound velocity profile inversion method based on double-layer self-organizing neural network

Abstract

The invention discloses a sea surface parameter sound velocity profile inversion method based on a double-layer self-organizing neural network, and the method comprises the steps: S1, collecting remote sensing parameters of a sea surface parameter inversion sound velocity profile, constructing training samples, and carrying out the clustering of the training samples through the self-organizing neural network; s2, generalizing the self-organizing neural network according to the clustered training samples; s3, neurons generated by generalization of the self-organizing neural network are matched with incomplete neurons, and the most matched neurons corresponding to the sound velocity profile to be inversed are determined; and S4, the sound velocity profile coefficient in the most matched neuron is taken out, orthogonal empirical function reconstruction is carried out, and a sound velocity profile inversion result is obtained. According to the method, the self-organizing neural network with a double-layer structure is constructed, clustering and generalization functions are realized at the same time, the technical bottleneck of grid division can be effectively solved in the field of sea surface parameter inversion of the sound velocity profile, and the inversion precision is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of sound velocity profile measurement, and particularly relates to a method for inverting the sound velocity profile of sea surface parameters based on a double-layer self-organizing neural network. Background Art

[0002] The sound velocity profile is an important parameter affecting ocean sound propagation. An accurate sound velocity profile is very important input information for the application of sonar systems. Since the sound velocity profile has strong spatio-temporal perturbation characteristics, it is of great significance to obtain the sound velocity distribution over a large range.

[0003] Traditional methods for measuring the sound velocity profile mainly use sound velocity profile instruments or temperature-salinity-depth instruments for in-situ measurement. However, this method is time-consuming and laborious, and it is almost impossible to measure the sound velocity distribution over a large range considering the cost. Acoustic sound velocity profile inversion is also a common method. However, since the acoustic signal contains the average sound velocity profile information along its propagation path, which is equivalent to an integrating probe, the result of acoustic inversion is often the average sound velocity profile along the acoustic signal propagation path, and it is often difficult to meet the requirements of actual sonar system applications in terms of spatial resolution.

[0004] Since satellite remote sensing is currently the only observation platform covering the globe, sea surface remote sensing data can provide high-resolution global observation data. Therefore, since the 1990s, methods for inverting the underwater sound velocity profile based on remote sensing data have been gradually developed. The most classic methods include the sEOF-r method widely used by the US military, that is, assuming that there is a near-linear relationship between remote sensing parameters and underwater parameters, and obtaining the coefficients of these relationships through fitting a large number of samples, and then determining the relationship between sea surface parameters and underwater parameters for inversion. In recent years, with the development of machine learning algorithms, people have realized that linear fitting does not fully conform to objective physical laws. Therefore, various machine learning algorithms have been used to explore the objective laws between sea surface parameters and underwater parameters, and a variety of algorithms such as random forest, self-organizing neural network, and long short-term memory network have been developed. These artificial intelligence algorithms have broken through the limitation of the fixed relationship formula of the sEOF-r method and greatly improved the accuracy of sound velocity profile inversion.

[0005] Although the current methods have achieved great success, all methods inevitably have a technical bottleneck:

[0006] Since the basis of these methods is to explore the relationship between sea surface parameters and seabed parameters, and the core of the method requires the statistical laws of the calculation regions to be consistent. To ensure this, almost all methods first divide the inversion region into grids of 1° or 2°, and default that the statistical laws within the grids are consistent. This approach has obvious drawbacks:

[0007] (1) The ocean is a complex perturbation system, and its statistical time does not solely depend on spatial location. It also includes various factors such as seasons, ocean currents, and monsoons. Therefore, simply dividing the space into grids makes it difficult to ensure the consistency of data statistical laws within the grids.

[0008] (2) The selection of the grid size is sometimes very difficult. When the grid is too small, the number of samples within the grid is relatively small, making it difficult for relevant methods to explore the relationship between sea surface parameters and underwater sound speed profiles based on statistical principles. When the grid is too large, due to temporal and spatial variations, the consistency of the statistical laws of sea surface parameters - sound speed profiles within the grid deteriorates, affecting the accuracy of inversion. Summary of the Invention

[0009] In view of the above deficiencies in the prior art, the sea surface parameter sound speed profile inversion method based on a double - layer self - organizing neural network provided by the present invention solves the problems that in the existing relevant methods, simply dividing the grid in space makes it difficult to ensure the consistency of data statistical laws within the grid, and it is difficult to determine the size of the grid division, thereby affecting the accuracy of sound speed profile inversion.

[0010] To achieve the above - mentioned invention objective, the technical solution adopted by the present invention is: a sea surface parameter sound speed profile inversion method based on a double - layer self - organizing neural network, including the following steps:

[0011] S1. Collect remote sensing parameters for inverting the sound speed profile of sea surface parameters, construct training samples, and cluster the training samples through a self - organizing neural network;

[0012] The number of neurons in the self - organizing neural network is the number of clustering targets, and each training sample is clustered into different neurons;

[0013] S2. Generalize the self - organizing neural network according to the clustered training samples;

[0014] S3. Match the neurons generated by generalizing the self - organizing neural network with incomplete neurons to determine the most - matched neuron corresponding to the sound speed profile to be inverted;

[0015] The incomplete neuron is a neuron formed by taking the sea surface height anomaly value and the sea surface temperature anomaly value as the solution conditions;

[0016] S4. Take out the sound speed profile coefficients in the most - matched neuron and perform orthogonal empirical function reconstruction to obtain the sound speed profile inversion result.

[0017] Further, the step S1 includes the following sub - steps:

[0018] S11. Collect remote sensing parameters for inverting the sound speed profile of sea surface parameters, including sea surface height anomaly values, sea surface temperature anomaly values, and historical samples of sound speed profiles;

[0019] S12. Take the average of the historical samples of the sound speed profile, decompose all sound speed profiles through orthogonal empirical functions, and then represent each sound speed profile in a reduced dimension using sound speed profile coefficients;

[0020] S13. Use the sEOF-r method to perform linear regression on the sea surface height anomaly value, sea surface temperature anomaly value, and multiple groups of sound speed profile coefficients to obtain a set of linear relationship coefficients between sea surface parameters and sound speed profile coefficients;

[0021] S14. Calculate the physical driving sound speed profile parameters under the linear relationship based on the sea surface height anomaly value, sea surface temperature anomaly value, and linear relationship coefficients;

[0022] S15. Combine the sea surface height anomaly value, sea surface temperature anomaly value, reduced dimension parameter value, and physical driving sound speed profile parameters of a group of sound speed profiles to form an element in the training sample, and perform normalization processing on it, and then construct the training sample.

[0023] Further, in the step S12, the sound speed profile c s (z) after dimension reduction is:

[0024]

[0025] In the formula, c 0 (z) represents the background profile obtained by taking the average of the historical samples of the sound speed profile, represents the orthogonal empirical function during orthogonal empirical decomposition after taking the average of all historical samples of the sound speed profile, z represents the sampling point of the profile in depth, a n represents the sound speed profile coefficient, the subscript n represents the ordinal number of the sound speed profile coefficient, n = 1 to N, N = 5.

[0026] Further, in the step S13, the sound speed profile coefficient parameter a n is expressed as:

[0027] a n = A 1 + a 2 × sla + A 3 × ssta + A 4 × sla × ssta;

[0028] In the formula, sla represents the sea surface height anomaly value, ssta represents the sea surface temperature anomaly value, A 1 , A 2 , A 3 and A 4 represent the linear relationship coefficients.

[0029] Further, the step S2 includes the following sub-steps:

[0030] S21. Select the nearest historical sample point from the clustered training samples according to the longitude and latitude position of the sound speed profile to be inverted;

[0031] S22. Input all the training samples in the cluster to which the historical sample point belongs into the self-organizing neural network again for generalization training;

[0032] Among them, the elements in the training samples input into the self-organizing network for generalization training include sea surface height anomaly value, sea surface temperature anomaly value and sound speed profile coefficient.

[0033] Further, in the step S3, the Euclidean distance between the generalized generated neuron and the incomplete neuron is calculated to match the most matching neuron, and the generalized generated neuron with the shortest Euclidean distance is used as the most matching neuron;

[0034] The Euclidean distance is expressed as:

[0035]

[0036] In the formula, E(p) represents the Euclidean distance between the p-th neuron generated by generalization and the incomplete neuron, represents the covariance matrix of the complete neuron and the incomplete neuron, x i represents the i-th element on the incomplete neuron, represents the i-th element of the p-th neuron on the generalization network, avail represents the subset of elements existing on the neuron, missing represents the subset of elements missing on the neuron, i represents the element ordinal number on the complete neuron, and j represents the element ordinal number on the incomplete neuron.

[0037] Further, in the step S4, the sound speed profile inversion result c s (z)' is expressed as:

[0038]

[0039] In the formula, c 0c (z) represents the average profile of all sound speed profiles in the cluster category to which the sound speed profile to be inverted belongs, represents the orthogonal empirical function obtained by performing orthogonal empirical decomposition on all sound speed profiles in the cluster category to which the sound speed profile to be inverted belongs, a n ' is the sound speed profile coefficient extracted from the most matching neuron corresponding to the cluster category to which the sound speed profile to be inverted belongs, and the subscript n represents the sound speed profile coefficient ordinal number, n = 1 to N, N = 5.

[0040] The beneficial effects of the present invention are:

[0041] The present invention designs a method for retrieving the sound speed profile using sea surface remote sensing parameters based on a double-layer self-organizing neural network by ingeniously utilizing the characteristics of the self-organizing neural network and simultaneously reducing and increasing the dimension (generalizing) of the samples. The advantages are as follows:

[0042] (1) By setting the parameters of the sound speed profile through clustering and physical driving, the training samples are classified according to the perturbation law, breaking through the limitations of the conventional grid method, breaking through the limitations of time and space, clustering the profiles with the same perturbation law, ensuring the consistency of the perturbation law of the training samples in the subsequent sound speed profile inversion, and improving the inversion accuracy.

[0043] (2) This method uses only one neural network algorithm to achieve clustering and generalization, avoiding the conflicts of data input and output caused by multiple algorithms, and is easier for users to understand and get started.

[0044] (3) The biggest feature of the traditional self-organizing neural network is its single-layer structure, and its function is relatively single. The present invention constructs a double-layer self-organizing neural network to simultaneously achieve the functions of clustering and generalization, which can effectively solve the technical bottleneck of grid division in the field of retrieving the sound speed profile from sea surface parameters and improve the inversion accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a flowchart of the method for retrieving the sound speed profile of sea surface parameters based on a double-layer self-organizing neural network provided by the present invention.

[0046] Figure 2 It is a schematic diagram showing the change of error with depth provided by the present invention.

[0047] Figure 3 It is a schematic diagram showing the change of error with samples provided by the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0048] The following describes the specific embodiments of the present invention to facilitate those skilled in the art to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions created using the concept of the present invention are within the scope of protection.

[0049] The embodiment of the present invention provides a method for retrieving the sound speed profile of sea surface parameters based on a double-layer self-organizing neural network, as Figure 1 shown, including the following steps:

[0050] S1. Collect the remote sensing parameters for retrieving the sound speed profile of sea surface parameters, construct training samples, and cluster the training samples through a self-organizing neural network;

[0051] The number of neurons in the self-organizing neural network is the number of clustering targets, and each training sample is clustered into different neurons;

[0052] S2. Generalize the self-organizing neural network according to the clustered training samples;

[0053] S3. Match the neurons generated by generalizing the self-organizing neural network with the incomplete neurons to determine the most matching neuron corresponding to the sound speed profile to be inverted;

[0054] The incomplete neuron is a neuron formed by using the sea surface height anomaly value and the sea surface temperature anomaly value as the solution conditions;

[0055] S4. Take out the sound speed profile coefficients in the most matching neuron and perform orthogonal empirical function reconstruction to obtain the sound speed profile inversion result.

[0056] In step S1 of the embodiment of the present invention, it includes the following sub-steps:

[0057] S11. Collect the remote sensing parameters for inverting the sound speed profile of sea surface parameters, including sea surface height anomaly value, sea surface temperature anomaly value and the historical samples of the sound speed profile;

[0058] S12. Take the average value of the historical samples of the sound speed profile and decompose all the sound speed profiles through orthogonal empirical functions, and then represent each sound speed profile in a reduced dimension with the sound speed profile coefficients;

[0059] S13. Perform linear regression on the sea surface height anomaly value, the sea surface temperature anomaly value and multiple groups of sound speed profile coefficients by using the sEOF-r method to obtain a set of linear relationship coefficients between the sea surface parameters and the sound speed profile coefficients;

[0060] S14. Calculate and obtain the physically driven sound speed profile parameters under the linear relationship according to the sea surface height anomaly value, the sea surface temperature anomaly value and the linear relationship coefficients;

[0061] S15. Combine the sea surface height anomaly value, the sea surface temperature anomaly value, the reduced dimension parameter value and the physically driven sound speed profile parameters of a group of sound speed profiles to form the elements in a training sample, and perform normalization processing on them, and then construct the training sample.

[0062] Specifically, in step S11, for the sea surface height anomaly value, select the satellite product with daily resolution and the longitude, latitude and time of the sea area sound speed profile samples corresponding one by one. A relatively mature data product is the sea surface height anomaly data of the European Copernicus; for the sea surface temperature anomaly value, select the satellite product with daily resolution and the longitude, latitude and time of the sea area sound speed profile samples corresponding one by one. A relatively mature product is the sea surface temperature anomaly data of the European Copernicus.

[0063] In step S12 of this embodiment, for the historical samples of the sound speed profile, since dimensionality reduction is required in the inverse problem of the sound speed profile, the most conventional method is used to process the sound speed profile, that is, the average value of the used sound speed profiles is taken to obtain the background profile c 0 (z), where z is the sampling point in depth. Then, all sound speed profiles are decomposed by orthogonal empirical functions. Using the classical sound speed dimensionality reduction method, each sample sound speed profile c s (z) is expressed as:

[0064]

[0065] In the formula, c 0 (z) represents the background profile obtained by taking the average value of the historical samples of the sound speed profile, represents the orthogonal empirical function during orthogonal empirical decomposition after taking the average value of all historical samples of the sound speed profile. z represents the sampling point of the profile in depth. a n represents the sound speed profile coefficient. The subscript n represents the ordinal number of the sound speed profile coefficient, n = 1 to N, and N = 5.

[0066] Among them, a n is the coefficient corresponding to the orthogonal empirical function, which can be obtained by performing regression on each sample profile using the orthogonal empirical function. Here, the most conventional method of the orthogonal empirical function is adopted, and a fifth-order orthogonal empirical function is taken. In this way, each sample can be dimensionally reduced and represented by a 1 , a 2 , a 3 , a 4 , a 5 parameters. These coefficients and the corresponding sea surface height anomaly value sla and sea surface temperature anomaly value ssta all constitute the elements of the training sample vector.

[0067] In step S13 of this embodiment, the drive-away driving sound speed profile parameters are used as parameters to assist in statistical feature classification in the neural network. Considering that in a large number of data studies, it is found that there is an approximate linear relationship between sea surface parameters and the underwater sound speed profile. Therefore, the most classical method, the sEOF-r method, can be used to define 5 physical drive-away driving sound speed profile parameters. In this embodiment, through multiple groups of sound speed profile coefficients, sea surface height anomaly values, and sea surface temperature anomaly values in the historical samples, a set of linear relationship coefficients between sea surface parameters and sound speed profile coefficients can be obtained, and the reconstructed sound speed profile parameters b n (that is, in the following formula, the position of a n is regarded as an unknown parameter); among them, the sound speed profile coefficient parameter a n is expressed as:

[0068] a n = A 1 + A2 × sla + A 3 × ssta + A 4 × sla × ssta;

[0069] Wherein, sla represents the sea surface height anomaly value, ssta represents the sea surface temperature anomaly value, and A 1 、A 2 、A 3 and A 4 represent the linear relationship coefficients.

[0070] In this embodiment, in the above method, the sEOF-r method is a classical solution method. For other existing methods, the inversion results have been obtained here, and the entire inversion process ends. In order to improve the accuracy of the results, the present invention only uses this result as a reference for the final result, inputs it into the neural network designed subsequently as part of the training, and compares the actual coefficients in the training samples with the sEOF-r inversion coefficients. Essentially, it classifies the sound speed profile samples by using the deviation degree of the actual parameters from the linear relationship as an index, and ensures that the profiles with consistent statistical relationships are processed together.

[0071] In step S1 of the embodiment of the present invention, when clustering the training samples through the self-organizing neural network, a large number of training samples are input into the self-organizing neural network, and the number of neurons is set to the number of categories to be classified (the number of categories needs to be determined according to the actual number of samples, and a good classification effect can be achieved by using 4 categories). By setting the number of neurons to a very small number of categories, each training sample is classified into different neurons, achieving the purpose of classifying the sound speed profile according to the law of sound speed perturbation. Thus, the first layer structure of the double-layer self-organizing neural network in the present invention is formed. Its main function is to perform clustering, classify the samples with the same inversion law, break the influence of the time and space grids of the traditional method, effectively gather the samples with the same statistical law together for processing, and improve the accuracy of the sound speed profile inversion method.

[0072] Step S2 of the embodiment of the present invention includes the following sub-steps:

[0073] S21. Select the nearest historical sample point in the clustered training samples according to the longitude and latitude position of the sound speed profile to be inverted;

[0074] S22. Input all the training samples in the cluster to which the historical sample point belongs into the self-organizing neural network again for generalization training;

[0075] Among them, the elements in the training samples input into the self-organizing network for generalization training include the sea surface height anomaly value, the sea surface temperature anomaly value, and the sound speed profile coefficient.

[0076] Specifically, in this embodiment, generalization is performed through a self-organizing neural network. The law of sound speed perturbation is generalized based on existing training samples. Through the training of the self-organizing neural network, possible perturbation situations are extended, and different neurons are generated to represent possible situations. First, the nearest historical sample point is selected from the classified samples according to the longitude and latitude positions of the profile to be solved. The cluster corresponding to this historical sample point can be considered to be of the same perturbation law as the profile to be solved. All samples in this cluster are input into the self-organizing neural network for training again, and only the elements of sla, ssta, and a are retained in the samples 1 and a 2 and a 3 and a 4 and a 5 , and the number of neurons this time is larger than the number of samples. According to a large number of experiments, three times the number of samples can achieve better results. Through generalization, each newly generated neuron represents a possible sound speed profile perturbation. Here, the second-layer structure of the double-layer self-organizing neural network is formed, and its main function is to perform generalization, invert samples with the same statistical law, which is equivalent to predicting possible situations under this law and generating various possible sound speed profile situations.

[0077] In step S3 of the embodiment of the present invention, the Euclidean distance between the neurons generated by generalization and the incomplete neurons is calculated to match the most matching neuron, and the neuron generated by generalization with the shortest Euclidean distance is used as the most matching neuron; specifically, sla and ssta as the solution conditions are combined into an incomplete neuron with only the elements sla and ssta, and are matched with the neurons generated by generalization. Among them, the Euclidean distance is expressed as:

[0078]

[0079] In the formula, E(p) represents the Euclidean distance between the p-th neuron generated by generalization and the incomplete neuron, represents the covariance matrix of the complete neuron and the incomplete neuron, and x i represents the i-th element on the incomplete neuron, represents the i-th element of the p-th neuron on the generalization network, avail represents the subset of elements existing on the neuron, missing represents the subset of elements missing on the neuron, i represents the element ordinal number on the complete neuron, and j represents the element ordinal number on the incomplete neuron.

[0080] The Euclidean distance between each neuron generated by generalization and the input incomplete neuron can be calculated through the above formula, and the neuron with the shortest distance is the most matching neuron.

[0081] In step S4 of the embodiment of the present invention, the sound velocity profile coefficient elements in the most matching neuron are taken out, that is, a in the most matching neuron 1 、a 2 、a 3 、a 4 、a 5 as the final inversion result; with the help of the most conventional orthogonal empirical function reconstruction, the sound velocity profile inversion result c s (z)' is expressed as:

[0082]

[0083] In the formula, c 0c (z) represents the average profile of all sound velocity profiles in the clustering category to which the sound velocity profile to be inverted belongs, represents the orthogonal empirical function obtained by orthogonal empirical decomposition of all sound velocity profiles in the clustering category to which the sound velocity profile to be inverted belongs, a n ' is the sound velocity profile coefficient in the most matching neuron corresponding to the clustering category of the sound velocity profile to be inverted extracted, and the subscript n represents the ordinal number of the sound velocity profile coefficient, n = 1 to N, N = 5.

[0084] In the embodiment of the present invention, a verification example of the above method is provided. In this embodiment, the argo buoy profiles in a certain sea area from 2007 to April 2022 are selected for testing, among which the historical samples from 2007 to 2016 are used for training, and the rest are used for verification. The experiment shows that the accuracy of using the self-organizing neural network iterative algorithm is significantly higher than the method of simply using the self-organizing neural network for inversion. The average error results of the inversion at each depth are as Figure 2 and Figure 3 .

[0085] In the present invention, specific embodiments are used to elaborate on the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present invention.

[0086] Those of ordinary skill in the art will realize that the embodiments described here are for helping the reader understand the principle of the present invention, and it should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those of ordinary skill in the art can make various other specific deformations and combinations that do not deviate from the essence of the present invention according to the technical revelations disclosed in the present invention, and these deformations and combinations are still within the protection scope of the present invention.

Claims

1. A sea surface parameter sound velocity profile inversion method based on a double-layer self-organizing neural network, characterized in that: The following steps are involved: S1, collect remote sensing parameters of sea surface parameter inversion sound velocity profile, construct training samples, and cluster the training samples through self-organizing neural network; The number of neurons in the self-organizing neural network is the clustering target number, and each training sample is clustered into a different neuron; S2, generalize the self-organizing neural network based on the clustered training samples; S3, matching the neurons generated by the generalization of the self-organizing neural network with the incomplete neurons, and determining the best matching neurons corresponding to the sound velocity profile to be inverted; The incomplete neuron is a neuron composed of the abnormal value of sea surface height and the abnormal value of sea surface temperature as solution conditions; S4. Take out the sound velocity profile coefficient in the most matching neuron, and reconstruct it by orthogonal empirical function to obtain the sound velocity profile inversion result.

2. The sea surface parameter sound velocity profile inversion method based on a double-layer self-organizing neural network according to claim 1 is characterized in that: The step S1 comprises the following sub-steps: S11. Collect remote sensing parameters for inverting sound velocity profiles from sea surface parameters, including sea surface height anomalies, sea surface temperature anomalies and historical samples of sound velocity profiles; S12, taking the average of the historical samples of the sound velocity profile, decomposing all the sound velocity profiles by using an orthogonal empirical function, and then reducing the dimension of each sound velocity profile and expressing it by a sound velocity profile coefficient; S13, performing linear regression on the abnormal values ​​of sea surface height, abnormal values ​​of sea surface temperature and multiple groups of sound velocity profile coefficients using the sEOF-r method to obtain a set of linear relationship coefficients between sea surface parameters and sound velocity profile coefficients; S14, calculating and obtaining the physical driven sound velocity profile parameters under the linear relationship according to the sea surface height anomaly value, the sea surface temperature anomaly value and the linear relationship coefficient; S15, forming a set of sea surface height anomaly values, sea surface temperature anomaly values, dimension reduction parameter values ​​and physical driven sound speed profile parameters of the sound speed profile into elements of a training sample, and normalizing them to construct a training sample.

3. The sea surface parameter sound velocity profile inversion method based on a double-layer self-organizing neural network according to claim 2 is characterized in that: In step S12, the sound velocity profile c after dimensionality reduction is used. s (z) is: Where c0(z) represents the background profile obtained by averaging the historical samples of the sound velocity profile. represents the orthogonal empirical function when all historical samples of the sound velocity profile are averaged and then subjected to orthogonal empirical decomposition. z represents the sampling point of the profile at depth. a n It represents the sound speed profile coefficient. The subscript n represents the ordinal number of the sound speed profile coefficient, n=1~N, N=5.

4. The method for inverting the sea surface parameter sound velocity profile based on a double-layer self-organizing neural network according to claim 2 is characterized in that: In step S13, the sound velocity profile coefficient parameter a n It is expressed as: a n =A1+A2×sla+A3×ssta+A4×sla×ssta; Where sla represents the sea surface height anomaly, ssta represents the sea surface temperature anomaly, and A1, A2, A3 and A4 represent the linear relationship coefficients.

5. The sea surface parameter sound velocity profile inversion method based on a double-layer self-organizing neural network according to claim 2 is characterized in that: The step S2 comprises the following sub-steps: S21, selecting the nearest historical sample point from the clustered training samples according to the latitude and longitude position of the sound velocity profile to be inverted; S22, inputting all training samples in the cluster to which the historical sample points belong into the self-organizing neural network again for generalization training; The elements of the training samples input into the self-organizing network for generalization training include sea surface height anomalies, sea surface temperature anomalies and sound speed profile coefficients.

6. The method for inverting the sea surface parameter sound velocity profile based on a double-layer self-organizing neural network according to claim 1 is characterized in that: In the step S3, the best matching neuron is matched by calculating the Euclidean distance between the generalized neuron and the incomplete neuron, and the generalized neuron with the shortest Euclidean distance is used as the best matching neuron; The Euclidean distance is expressed as: Where E(p) represents the Euclidean distance between the p-th neuron generated by generalization and the incomplete neuron. represents the covariance matrix of complete neurons and incomplete neurons, x i represents the i-th element on the incomplete neuron, represents the i-th element of the p-th neuron on the generalization network, avail represents the element subset available on the neuron, missing represents the element subset missing on the neuron, i represents the element ordinal number on the complete neuron, and j represents the element ordinal number on the incomplete neuron.

7. The method for inverting the sea surface parameter sound velocity profile based on a double-layer self-organizing neural network according to claim 2 is characterized in that: In step S4, the sound velocity profile inversion result c s (z)′ is expressed as: In the formula, c 0c (z) represents the average profile of all sound velocity profiles in the cluster category to which the sound velocity profile to be inverted belongs, represents the orthogonal empirical function obtained by orthogonal empirical decomposition of all sound velocity profiles in the cluster category to which the sound velocity profile to be inverted belongs, a n ′ is the sound speed profile coefficient in the most matching neuron corresponding to the cluster category to which the extracted sound speed profile to be inverted belongs, and the subscript n represents the ordinal number of the sound speed profile coefficient, n=1~N, N=5.

Images