Ocean salinity data correction method based on space-time gradient adaptive filtering

Abstract

The invention relates to the field of ocean observation data processing, and discloses an ocean salinity data correction method based on space-time gradient adaptive filtering, which comprises the following steps of: calculating a motion compensation salinity gradient and a motion compensation temperature gradient; constructing a double-gradient driving filtering window, introducing a gradient sensitive weight distribution strategy, and carrying out adaptive dynamic filtering on the salinity data; generating an initial hyper-parameter combination of the GBDT model; inputting salinity data subjected to adaptive dynamic filtering into the model, optimizing an initial hyper-parameter combination of the model by using a grey wolf optimization algorithm, obtaining an optimal hyper-parameter combination through a dynamic search strategy, and obtaining a final salinity correction value at the same time; and carrying out index verification on the final salinity correction value, and when the salinity does not reach the standard, correcting a model hyper-parameter or a self-adaptive dynamic filtering parameter until the salinity reaches the standard. According to the method disclosed by the invention, the physical abrupt change signal of the spring layer region is reserved to the greatest extent while the noise is suppressed, and the accuracy of salinity profile correction is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of ocean observation data processing, and particularly relates to a method for correcting ocean salinity data based on spatio-temporal gradient adaptive filtering. Background Art

[0002] As the core device of the global ocean real-time observation network, the Argo buoy is equipped with a CTD sensor, which vertically ascends and descends in the ocean along a preset trajectory, collects data at intervals and transmits it to the ground station via satellite. The CTD sensor measures the seawater conductivity, temperature and pressure, and uses the conductivity data to invert the salinity value through a formula equation.

[0003] Salinity data, as a basic physical quantity of the ocean environment, has irreplaceable scientific value. Salinity data is a core parameter for studying ocean dynamics, a key indicator of climate change, and a necessary factor affecting the ecosystem and fishery resources. However, due to factors such as the hardware problems of the sensor itself and the interference of the seawater environment, there is an error between the salinity sequence data measured by the sensor and the true seawater salinity. Therefore, it is necessary to correct the salinity data to meet the high-precision data requirements of ocean scientific research and engineering applications.

[0004] Traditionally, static filtering methods are used to correct salinity data. Fixed-window median filtering is adopted, and the median value within a fixed range (such as 5 - 15 points) is taken for each data point, and noise is suppressed through statistical filtering. However, the fixed window cannot match the stratification intensity difference, and an inappropriate filtering window size will also cause feature loss. At the same time, previous correction methods have all ignored that when the sensor ascends and descends, the salinity data of the same water mass will be misjudged as a spatial gradient due to the time sampling deviation. If the pseudo-gradient is not eliminated, the salinity change caused by movement will be misjudged as the true ocean stratification. Summary of the Invention

[0005] To solve the above technical problems, the present invention provides a method for correcting ocean salinity data based on spatio-temporal gradient adaptive filtering, so as to achieve the purpose of realizing high-precision dynamic correction of ocean salinity data.

[0006] To achieve the above purpose, the technical solution of the present invention is as follows: A method for correcting ocean salinity data based on spatio-temporal gradient adaptive filtering, comprising the following steps: Step 1, calculate the time gradient and spatial gradient of the salinity data and temperature data collected by the CTD sensor on the Argo buoy, and compensate for the pseudo-gradient caused by the movement of the sensor in combination with the actual vertical movement speed to obtain the motion-compensated salinity gradient and motion-compensated temperature gradient; Step 2: Construct a dual-gradient-driven filtering window using motion-compensated salinity gradient and motion-compensated temperature gradient, and introduce a gradient-sensitive weight assignment strategy to adaptively and dynamically filter the salinity data; Step 3: Map the motion-compensated salinity gradient as a control parameter to the learning rate, decision tree depth, and number of decision trees of the GBDT model to generate an initial hyperparameter combination of the GBDT model; Step 4: Input the adaptively and dynamically filtered salinity data into the GBDT model, and use the grey wolf optimization algorithm to optimize the initial hyperparameter combination of the model. In each iteration, calculate the physical constraint loss function using the predicted salinity correction value output by the model, evaluate the hyperparameter performance according to the physical constraint loss function, and obtain the optimal hyperparameter combination of the model through a dynamic search strategy, and at the same time obtain the final salinity correction value; Step 5: Verify the index of the final salinity correction value. When the standard is not met, correct the hyperparameters of the GBDT model or the adaptive dynamic filtering parameters until the index reaches the standard.

[0007] In the above solution, in Step 1, the temporal gradient and spatial gradient of the salinity data and temperature data include the salinity temporal gradient, temperature temporal gradient, salinity spatial gradient, and temperature spatial gradient, and the specific calculations are as follows: Salinity temporal gradient calculation formula: ; where, represents the th sampling point, and respectively represent the times of the th and th sampling points, and respectively represent the salinity measurement values at time points and [[ID=�5]] ;<^ Temperature temporal gradient calculation formula: ; where, and respectively represent the temperature measurement values at time points and ; Salinity spatial gradient calculation formula: ; where, and respectively represent the pressure measurement values of the th and th sampling points, and respectively represent the pressure points and the salinity measurement value at represents the density of seawater in the current layer, represents the acceleration due to gravity of seawater; Temperature spatial gradient calculation formula: ; Wherein, and respectively represent the temperature measurement values at pressure points and ;

[0008] In the above solution, in step 1, the motion-compensated salinity gradient is calculated as follows: ; Wherein, represents the salinity time gradient, represents the salinity spatial gradient, represents the vertical motion speed of the sensor in seawater: ; Wherein, and respectively represent the pressure measurement values at the (i + 1)-th and (i - 1)-th sampling points, and respectively represent the time at the (i + 1)-th and (i - 1)-th sampling points, represents the density of seawater in the current layer, represents the acceleration due to gravity of seawater; The motion-compensated temperature gradient is calculated as follows: ; Wherein, represents the temperature time gradient, represents the temperature spatial gradient.

[0009] In the above solution, in step 2, a dual-gradient-driven filtering window is constructed as follows: ; Wherein, represents the size of the filtering window, represents the minimum value of the window, represents the maximum value of the window, represents the salinity gradient weight, represents the temperature gradient weight, represents the normalized value of the salinity gradient, represents the normalized value of the temperature gradient; ; ; Among them, represents the motion-compensated salinity gradient, represents the motion-compensated temperature gradient, , respectively represent the maximum and minimum values of the absolute value of the motion-compensated salinity gradient on the seawater profile, , respectively represent the maximum and minimum values of the absolute value of the motion-compensated temperature gradient on the seawater profile.

[0010] In the above solution, in step 2, the calculation formula for adaptive dynamic filtering of salinity data is as follows: ; Among them, represents the filtered salinity data, represents the original salinity data, represents the window within the salinity data, represents the salinity gradient sensitivity weight; ; Among them, represents the rate of change of the control weight with the gradient, represents the motion-compensated salinity gradient.

[0011] In the above solution, step 3 is specifically as follows: (1) The learning rate mapping formula is as follows: ; Among them, represents the learning rate of the model, represents the maximum learning rate, represents the attenuation coefficient, represents the motion-compensated salinity gradient, represents the maximum value of the absolute value of the motion-compensated salinity gradient on the seawater profile; (2) The decision tree depth mapping formula is as follows: ; Among them, represents the decision tree depth; (3) The decision tree quantity mapping formula is as follows: ; Among them, represents the number of decision trees, represents the minimum number of decision trees, represents the complexity scaling coefficient, represents the number of current profile stratification interfaces.

[0012] In the above solution, step 4 is specifically as follows: (1)Initialization of the wolf pack: Use the initial hyperparameter combination generated in step 3 as the initial solution of the leading wolf, and the remaining individuals are distributed in the neighborhood of the leading wolf's solution. The formula is as follows: ; Where, represents the candidate hyperparameter vector, represents the initial solution vector of the leading wolf, represents the basic neighborhood range vector, corresponding to the perturbation ranges of the learning rate, decision tree depth, and number of decision trees respectively, represents the hyperbolic tangent function, which is used to map the temperature gradient to the neighborhood scaling factor, represents the motion-compensated temperature gradient, represents the maximum value of the absolute value of the motion-compensated temperature gradient on the seawater section; (2)Iterative optimization: For each candidate hyperparameter vector in each generation of the wolf pack, perform the following operations: Use to train a temporary GBDT model, with the input being the filtered salinity data , and the output being the temporary predicted salinity correction value . Calculate the physical constraint loss function; use the physical constraint loss function to evaluate the performance of the hyperparameters, so as to find the parameter combination with the smallest salinity error and the most physically consistent gradient and density in each iteration: ; (3)Dynamic search: If the current layer's motion-compensated salinity gradient , is the standard deviation of the motion-compensated salinity gradient, and it is determined as a strong gradient area. At this time, trigger the dynamic search strategy and update the convergence factor: ; Where, represents the updated convergence factor, represents the initial convergence factor, represents the current iteration number, represents the maximum iteration number; Calculation of the position update coefficient: ; Where, , represents the random perturbation coefficient between, represents the convergence factor; and represent the position update coefficients, which are used to control the search range and adjust the influence degree of the optimal solution on the current solution; Scaled position update formula: ; where denotes the candidate parameter vector at the -th iteration, denotes the scaling factor, denotes the alpha solution at the -th iteration, i.e., the current optimal parameter vector; (4) Termination and output: When the maximum number of iterations is reached or the change in the alpha loss value is for 5 consecutive times, stop the iteration and output the optimal hyperparameters , use these hyperparameters to train the final GBDT model and output the final salinity correction value: ; where denotes the final salinity correction value, denotes the predicted correction value of the -th decision tree, denotes the learning rate, denotes the number of decision trees, denotes the filtered salinity value.

[0013] In the above solution, in step 4, the physical constraint loss function formula is as follows: ; where denotes the salinity absolute error, denotes the density penalty term, denotes the gradient similarity, and denote the weight coefficients; Salinity absolute error: ; where N denotes the total amount of data, denotes the reference salinity value of the data, denotes the predicted salinity correction value; Gradient similarity: ; where denotes the predicted correction data motion compensation salinity gradient of the -th data point, denotes the reference motion compensation salinity gradient of the -th data point; Density penalty term: ; where denotes the The salinity correction value is temporarily predicted for each data point, indicating the reference salinity value of the th data point,

[0014] In the above scheme, in step 5, the index verification of the final salinity correction value includes statistical accuracy verification, specifically as follows: When the average absolute error of salinity exceeds the threshold, the hyperparameters of the GBDT model are adjusted as follows: Dynamic adjustment of model complexity: ; ; where, is the corrected tree depth, is the corrected number of trees, represents the original decision tree depth search range, represents the original decision tree number search range, represents the model complexity correction coefficient; ; where, represents the current average absolute error of salinity, represents the statistical accuracy threshold, represents the absolute value of the motion-compensated salinity gradient, represents the maximum value of the absolute value of the motion-compensated salinity gradient on the seawater section.

[0015] In the above scheme, in step 5, the index verification of the final salinity correction value includes feature retention degree verification, specifically as follows: The gradient retention rate is used to measure the retention degree of the salinity gradient feature during the correction process. The calculation formula of the gradient retention rate is as follows: ; where GPR is the gradient retention rate, represents the motion-compensated salinity gradient after correction for the i-th data point, represents the salinity gradient of the motion-compensated reference data for the i-th data point, and N represents the total number of data points in the salinity profile; When the gradient retention rate is less than the threshold , the adaptive dynamic filtering parameters are corrected: Adjustment of the maximum filtering window: ; Adjustment of the salinity gradient sensitive weight: ; where, denotes the corrected maximum filtering window, denotes the maximum filtering window, denotes the salinity gradient sensitive weight, denotes the corrected salinity gradient sensitive weight, denotes the adjustment coefficient; ; wherein, denotes the gradient retention rate threshold.

[0016] Through the above technical solutions, an ocean salinity data correction method based on spatio-temporal gradient adaptive filtering provided by the present invention has the following beneficial effects: 1. Through the motion compensation gradient in step 1 and the gradient similarity constraint in step 4 of the present invention, the sensor motion artifacts are distinguished from the real ocean stratification, and the problem that the water mass time deviation caused by the Argo float ascending and descending sampling is misjudged as the spatial gradient is solved. This method effectively strips the motion-induced false gradient components, making the spatial distribution representation of the ocean stratification characteristics closer to the real physical scenario, and significantly improving the ability of the salinity profile data to depict the ocean dynamic process; 2. Through the double-gradient-driven filtering window in step 2 and the closed-loop correction of the feature retention degree in step 5 of the present invention, the real stratification features are retained in the thermohaline layer region, and the problem that the traditional fixed-window filtering method causes over-smoothing of the thermocline and loss of features is solved. This method maximally retains the physical mutation signals in the thermocline region while suppressing noise, improving the accuracy of the ocean stratification structure analysis based on salinity data; 3. Through step 3 of the present invention, the motion compensation gradient is mapped to the hyperparameters of the GBDT model, and through step 4 of the grey wolf optimization algorithm for hyperparameter tuning, the problem that the traditional machine learning method depends too much on experience for parameter tuning is solved. This method automatically associates the physical gradient features with the algorithm parameters, realizes the adaptive optimization of the model hyperparameters, significantly enhances the robustness of the salinity correction algorithm in different ocean environments, and greatly reduces the manual tuning cost in engineering applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art.

[0018] Figure 1 is a schematic flow chart of an ocean salinity data correction method based on spatio-temporal gradient adaptive filtering disclosed in the embodiments of the present invention; Figure 2 is the effect diagram of the salinity data profile correction of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention.

[0020] The present invention provides a method for correcting ocean salinity data based on spatio-temporal gradient adaptive filtering, as Figure 1 shown, including the following steps: Step 1: First, a temperature-salinity-depth sensor (CTD) named SHY and Sea-Bird that meets international measurement standards is carried on an Argo buoy, and seawater at 19.5°N, 116.0°E in the sea area is sampled and measured at fixed intervals in the vertical direction to obtain conductivity, temperature, and pressure sequence data respectively. Among them, the sensor moves uniformly in the vertical direction at a speed of 0.1 m / s, and data is collected at each one-meter depth interval, with a time interval of 0.1 m / s.

[0021] The salinity data obtained by the SHY sensor is used as the data to be corrected, and the salinity data obtained by the Sea-Bird sensor is used as the reference data.

[0022] Calculate the time gradient and space gradient of the salinity data and temperature data collected by the CTD sensor on the Argo buoy, and compensate for the pseudo-gradient caused by the sensor movement in combination with the actual vertical movement speed to obtain the motion-compensated salinity gradient and motion-compensated temperature gradient.

[0023] The time gradient and space gradient of the salinity data and temperature data include salinity time gradient, temperature time gradient, salinity space gradient, and temperature space gradient, and the specific calculations are as follows: Salinity time gradient calculation formula: ; Among them, represents the th sampling point, and respectively represent the times of the rd and th sampling points, and respectively represent the salinity measurement values at time points and ; Temperature time gradient calculation formula: ; Among them, and respectively represent the temperature measurement values at time points and ; Salinity space gradient calculation formula: ; Among them, and respectively represent the pressure measurement values at the and the th sampling points, and respectively represent the salinity measurement values at the pressure points and ; represents the density of seawater in the current layer, represents the gravitational acceleration of seawater; Calculation formula for temperature spatial gradient: ; Among them, and respectively represent the temperature measurement values at the pressure points and ;

[0024] The motion-compensated salinity gradient is calculated as follows: ; Among them, represents the salinity temporal gradient, represents the salinity spatial gradient, represents the vertical motion speed of the sensor in seawater: ; Among them, and respectively represent the pressure measurement values at the (i + 1)-th and (i - 1)-th sampling points, and respectively represent the times at the (i + 1)-th and (i - 1)-th sampling points, represents the density of seawater in the current layer, represents the gravitational acceleration of seawater; Based on the UNESCO seawater equation of state, by coupling the density at standard atmospheric pressure and the pressure compression effect, calculate the seawater density at any pressure: ; Among them, represents the seawater density at standard atmospheric pressure, represents the seawater pressure, represents the bulk modulus of elasticity, which describes the ability of seawater to resist compression and is calculated by the UNESCO standard formula and applicable to the global ocean pressure range: ; Among them, is the bulk modulus of elasticity at standard atmospheric pressure, P represents the seawater pressure, and S represents the salinity. denotes the bulk modulus of elasticity, and A and B denote the pressure correction coefficients.

[0025] The motion-compensated temperature gradient is calculated as follows: ; where denotes the temperature-time gradient, denotes the temperature-space gradient.

[0026] Step 2: Use the motion-compensated salinity gradient and the motion-compensated temperature gradient to construct a double-gradient-driven filtering window, and introduce a gradient-sensitive weight assignment strategy to adaptively dynamically filter the salinity data.

[0027] It is obtained through historical data testing that a window width less than 5 results in noise residue, and greater than 15 results in smooth transition of the thermocline. Therefore, the initial window minimum value = 5, and the maximum value is 15. While the salinity gradient dominates the window size, the temperature gradient is introduced to compensate for the thermal expansion effect, and the double-gradient-driven filtering window of the temperature gradient and the salinity gradient is constructed: ; where denotes the filtering window size, denotes the window minimum value, denotes the window maximum value, denotes the salinity gradient weight, denotes the temperature gradient weight, denotes the normalized salinity gradient value, denotes the normalized temperature gradient value; normalizing the gradient can avoid the influence of different physical dimensions on window calculation; ; ; where denotes the motion-compensated salinity gradient, denotes the motion-compensated temperature gradient, , respectively denote the maximum and minimum values of the absolute value of the motion-compensated salinity gradient on the seawater section, , respectively denote the maximum and minimum values of the absolute value of the motion-compensated temperature gradient on the seawater section.

[0028] The calculation formula for adaptively dynamically filtering the salinity data is as follows: ; where denotes the filtered salinity data, denotes the original salinity data, Represents the salinity data within the window within the window, represents the salinity gradient sensitive weight; to prevent the traditional median filtering from smoothing the true steep change characteristics of the pycnocline, within each window a balance between retaining the original data in the pycnocline area and suppressing noise in the uniform area is achieved according to the salinity gradient adaptive weight.

[0029] ; Among them, represents the rate at which the control weight changes with the gradient, represents the motion-compensated salinity gradient.

[0030] Step 3: Map the motion-compensated salinity gradient as a control parameter to the learning rate, decision tree depth, and number of decision trees of the GBDT (Gradient Boosting Decision Tree) model to generate an initial hyperparameter combination of the GBDT model.

[0031] Specifically as follows: (1) Adaptively adjust the learning rate according to the salinity gradient, using a smaller learning rate in the pycnocline area with a larger gradient and a larger learning rate in the uniform area. The learning rate mapping formula is as follows: ; Among them, represents the learning rate of the model, represents the maximum learning rate, represents the decay coefficient, represents the motion-compensated salinity gradient, represents the maximum value of the absolute value of the motion-compensated salinity gradient on the seawater profile; (2) The decision tree depth is set in layers according to the magnitude of the salinity gradient. The larger the gradient, the greater the tree depth to capture complex non-linear relationships. The decision tree depth mapping formula is as follows: ; Among them, represents the decision tree depth; (3) The number of decision trees increases with the increase in the stratification intensity and the number of stratification interfaces. The decision tree number mapping formula is as follows: ; Among them, represents the number of decision trees, represents the minimum number of decision trees, represents the complexity scaling coefficient, represents the current number of stratification interfaces of the profile.

[0032] Step 4: Input the salinity data after adaptive dynamic filtering into the GBDT model, and use the Grey Wolf Optimization algorithm to optimize the initial hyperparameter combination of the model. In each iteration, calculate the physical constraint loss function using the predicted salinity correction value output by the model, evaluate the hyperparameter performance based on the physical constraint loss function, obtain the optimal hyperparameter combination of the model through a dynamic search strategy, and simultaneously obtain the final salinity correction value.

[0033] Specifically as follows: (1) Initialization of the wolf pack: Use the initial hyperparameter combination generated in Step 3 as the initial solution of the lead wolf, and the remaining individuals are distributed in the neighborhood of the lead wolf's solution. Moreover, the greater the temperature gradient, the more concentrated the distribution of the initial solution. The formula is as follows: ; Among them, represents the candidate hyperparameter vector, represents the initial solution vector of the lead wolf, represents the basic neighborhood range vector, corresponding to the perturbation ranges of the learning rate, decision tree depth, and number of decision trees respectively, represents the hyperbolic tangent function, which is used to map the temperature gradient into the neighborhood scaling factor, represents the motion compensation temperature gradient, represents the maximum value of the absolute value of the motion compensation temperature gradient on the seawater section (2) Iterative optimization: For each candidate hyperparameter vector in each generation of the wolf pack, perform the following operations: Use to train a temporary GBDT model, with the input being the filtered salinity data , and the output being the temporary predicted salinity correction value , and calculate the physical constraint loss function; The formula for the physical constraint loss function is as follows: ; Among them, represents the salinity absolute error, represents the density penalty term, represents the gradient similarity, and represent the weight coefficients; Specifically, the salinity absolute error: ; Among them, N represents the total amount of data, represents the reference salinity value of the data, represents the predicted salinity correction value; Gradient similarity: ; in, Indicates the Data points predicted corrected data motion compensated salinity gradient, Indicates the Data points are referenced to the motion-compensated salinity gradient; Density penalty: ; in, Indicates the Temporary predicted salinity correction value for data points, Indicates the The reference salinity value for each data point is Seawater density calculated for the UNESCO seawater equation of state.

[0034] The physical constraint loss function is used to evaluate the hyperparameter performance, so that each iteration finds the parameter combination with the minimum salinity error and the gradient and density that best conform to the physical laws: ; (3) Dynamic search: If the current layer motion compensates the salinity gradient , The standard deviation of the motion-compensated salinity gradient is determined to be a strong gradient area. At this time, the traditional GWO is prone to fall into the local optimum, triggering the dynamic search strategy and updating the convergence factor: ; in, represents the updated convergence factor, represents the initial convergence factor, Indicates the current iteration number, Indicates the maximum number of iterations; Position update coefficient calculation: ; in, , express The random disturbance coefficient, represents the convergence factor; and Represents the position update coefficient, which is used to control the search range and adjust the influence of the optimal solution on the current solution; The formula for updating the position after scaling is: ; in, Indicates the Iteration candidate parameter vector, represents the scaling factor, Indicates the The leading wolf solution of the current iteration, i.e., the current optimal parameter vector; (4)Termination and output: When the maximum number of iterations is reached or the change in the leading wolf loss value has been continuous for 5 times Stop the iteration and output the optimal hyperparameters , use these hyperparameters to train the final GBDT model and output the final salinity correction value: ; Among them, represents the final salinity correction value, represents the predicted correction value of the th decision tree, represents the learning rate, represents the number of decision trees, represents the salinity value after filtering.

[0035] Step 5: Verify the indicators of the final salinity correction value. When the standard is not met, correct the hyperparameters of the GBDT model or the adaptive dynamic filtering parameters until the indicators meet the standard.

[0036] Verifying the indicators of the final salinity correction value includes statistical accuracy verification, specifically as follows: When the mean absolute error of salinity exceeds the threshold, adjust the hyperparameters of the GBDT model as follows: Dynamic adjustment of model complexity: ; ; Among them, is the corrected tree depth, is the corrected number of trees, represents the original decision tree depth search range, represents the original decision tree number search range, represents the model complexity correction coefficient; ; Among them, represents the current mean absolute error of salinity, represents the statistical accuracy threshold, such as 0.02 PSU; represents the absolute value of the motion-compensated salinity gradient, represents the maximum value of the absolute value of the motion-compensated salinity gradient on the seawater section.

[0037] Verifying the indicators of the final salinity correction value includes feature retention degree verification, specifically as follows: The gradient retention rate is used to measure the retention degree of the salinity gradient feature during the correction process. The calculation formula of the gradient retention rate is as follows: ; where GPR is the gradient retention rate, represents the salinity gradient after motion compensation correction for the i-th data point, represents the salinity gradient of the motion compensation reference data for the i-th data point, and N represents the total number of data points in the salinity profile; When the gradient retention rate is less than the threshold ( %), the adaptive dynamic filtering parameters are corrected: Adjustment of the maximum filtering window: ; Adjustment of the salinity gradient sensitive weight: ; where represents the corrected maximum filtering window, represents the maximum filtering window, represents the salinity gradient sensitive weight, represents the corrected salinity gradient sensitive weight, represents the adjustment coefficient; ; where represents the gradient retention rate threshold.

[0038] After correction, the salinity correction effect of the embodiments of the present invention is shown in Figure 2 . It can be seen that the method of the present invention makes the corrected salinity curve closely fit the reference data curve, the steep gradient in the pycnocline area is retained, and the noise in the uniform area is effectively suppressed, fully demonstrating the high-precision correction ability of the present invention in complex marine environments.

[0039] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An ocean salinity data correction method based on spatio-temporal gradient adaptive filtering, characterized in that, It includes the following steps: Step 1: Calculate the temporal gradient and spatial gradient of the salinity data and temperature data collected by the CTD sensor on the Argo float, and compensate for the pseudo-gradient caused by the sensor movement in combination with the actual vertical movement speed to obtain the motion-compensated salinity gradient and motion-compensated temperature gradient; Step 2: Use the motion-compensated salinity gradient and motion-compensated temperature gradient to construct a dual-gradient-driven filtering window, and introduce a gradient-sensitive weight assignment strategy to perform adaptive dynamic filtering on the salinity data; Step 3: Map the motion-compensated salinity gradient as a control parameter to the learning rate, decision tree depth, and number of decision trees of the GBDT model to generate an initial hyperparameter combination of the GBDT model; Step 4: Input the salinity data after adaptive dynamic filtering into the GBDT model, use the grey wolf optimization algorithm to optimize the initial hyperparameter combination of the model. In each iteration, calculate the physical constraint loss function using the predicted salinity correction value output by the model, evaluate the performance of the hyperparameters according to the physical constraint loss function, obtain the optimal hyperparameter combination of the model through a dynamic search strategy, and at the same time obtain the final salinity correction value; Step 5: Verify the indicators of the final salinity correction value. When the standard is not met, correct the hyperparameters of the GBDT model or the parameters of the adaptive dynamic filtering until the standard is met.

2. The marine salinity data correction method based on spatio-temporal gradient adaptive filtering according to claim 1, wherein In Step 1, the temporal gradient and spatial gradient of the salinity data and temperature data include the salinity temporal gradient, temperature temporal gradient, salinity spatial gradient, and temperature spatial gradient, and the specific calculations are as follows: Salinity temporal gradient calculation formula: ； Among them, represents the th sampling point, and respectively represent the time of the th and the th sampling points, and respectively represent the salinity measurement values at the time points and [[ID= Temperature temporal gradient calculation formula: ； Among them, and respectively represent the temperature measurement values at time points and ; Salinity spatial gradient calculation formula: ； wherein, and respectively represent the pressure measurement values at the th and th sampling points, and respectively represent the salinity measurement values at the pressure points and , represents the density of the seawater in the current layer, represents the acceleration of gravity of the seawater; Temperature spatial gradient calculation formula: ； Among them, and respectively represent the temperature measurement values at the pressure points and and 3. A method for correcting ocean salinity data based on spatio-temporal gradient adaptive filtering according to claim 1, characterized in that, In Step 1, the motion-compensated salinity gradient is calculated as follows: ； Among them, represents the salinity time gradient, represents the salinity space gradient, represents the vertical movement speed of the sensor in seawater: ； Among them, and represent the pressure measurement values at the (i + 1)-th and (i - 1)-th sampling points respectively, and represent the times at the (i + 1)-th and (i - 1)-th sampling points respectively, represents the density of seawater at the current layer, represents the gravitational acceleration of seawater; The motion-compensated temperature gradient is calculated as follows: ； Among them, represents the temperature-time gradient, represents the temperature-space gradient.

4. A method for correcting ocean salinity data based on spatio-temporal gradient adaptive filtering according to claim 1, characterized in that, In Step 2, the dual-gradient-driven filtering window is constructed as follows: ； Among them, represents the filtering window size, represents the minimum value of the window, represents the maximum value of the window, represents the salinity gradient weight, represents the temperature gradient weight, represents the salinity gradient normalization value, represents the temperature gradient normalization value; ； ； Among them, represents the motion-compensated salinity gradient, represents the motion-compensated temperature gradient, , respectively represent the maximum and minimum values of the absolute value of the motion-compensated salinity gradient on the seawater section, , respectively represent the maximum and minimum values of the absolute value of the motion-compensated temperature gradient on the seawater section.

5. A method for correcting ocean salinity data based on spatio-temporal gradient adaptive filtering according to claim 1, characterized in that In Step 2, the calculation formula for performing adaptive dynamic filtering on the salinity data is as follows: ； Among them, represents the salinity data after filtering, represents the original salinity data, represents the window salinity data within, represents the salinity gradient sensitivity weight; ； Among them, represents the rate at which the control weight changes with the gradient, represents the motion-compensated salinity gradient.

6. The marine salinity data correction method based on spatio-temporal gradient adaptive filtering according to claim 1, characterized in that Step 3 is specifically as follows: (1) The learning rate mapping formula is as follows: ； Among them, represents the learning rate of the model, represents the maximum learning rate, represents the attenuation coefficient, represents the motion-compensated salinity gradient, represents the maximum value of the absolute value of the motion-compensated salinity gradient on the seawater section; (2) The decision tree depth mapping formula is as follows: ； Among them, represents the depth of the decision tree; (3) The number of decision trees mapping formula is as follows: ； Among them, represents the number of decision trees, represents the minimum number of decision trees, represents the complexity scaling factor, represents the number of current profile stratification interfaces.

7. A method for correcting ocean salinity data based on spatio-temporal gradient adaptive filtering according to claim 1, characterized in that, Step 4 is specifically as follows: (1) Grey wolf group initialization: Use the initial hyperparameter combination generated in Step 3 as the initial solution of the alpha wolf, and the remaining individuals are distributed in the neighborhood of the alpha wolf solution, and the formula is as follows: ； Among them, represents the candidate hyperparameter vector, represents the initial solution vector of the alpha wolf, represents the basic neighborhood range vector, corresponding to the perturbation ranges of the learning rate, decision tree depth, and number of decision trees respectively, represents the hyperbolic tangent function, which is used to map the temperature gradient to the neighborhood scaling factor, represents the motion-compensated temperature gradient, represents the maximum value of the absolute value of the motion-compensated temperature gradient on the seawater section (2) Iterative optimization: For each candidate hyperparameter vector in each generation of the wolf pack , perform the following operations: Use Train a temporary GBDT model with the input being the filtered salinity data and the output being the temporary predicted salinity correction value Calculate the physical constraint loss function; evaluate the hyperparameter performance using the physical constraint loss function to find the parameter combination with the smallest salinity error and the most physically consistent gradient and density at each iteration: ; (3) Dynamic search: If the current layer has motion-compensated salinity gradient , is the standard deviation of the motion-compensated salinity gradient. If it is determined to be a strong gradient area, the dynamic search strategy is triggered at this time to update the convergence factor: ； Among them, represents the updated convergence factor, represents the initial convergence factor, represents the current iteration number, represents the maximum iteration number; Position update coefficient calculation: ； Among them, , represents the random perturbation coefficient between intervals, represents the convergence factor; and represent the position update coefficients, which are used to control the search range and adjust the influence degree of the optimal solution on the current solution; Scaled position update formula: ； Among them, represents the candidate parameter vector for the th iteration, represents the scaling factor, represents the alpha wolf solution for the th iteration, that is, the current optimal parameter vector; (4) Termination and output: When the maximum number of iterations is reached or the change amount of the alpha value of the leading wolf remains unchanged for 5 consecutive times stop the iteration and output the optimal hyperparameters Use these hyperparameters to train the final GBDT model and output the final salinity correction value: ； Among them, represents the final salinity correction value, represents the predicted correction value of the th decision tree, represents the learning rate, represents the number of decision trees, represents the salinity value after filtering.

8. A method for correcting ocean salinity data based on spatio-temporal gradient adaptive filtering according to claim 1, characterized in that, In Step 4, the physical constraint loss function formula is as follows: ； Among them, represents the absolute error of salinity, represents the density penalty term, represents the gradient similarity, and represents the weight coefficient; Salinity absolute error: ； where N represents the total amount of data, represents the reference salinity value of the data, represents the predicted salinity correction value; Gradient similarity: ； Among them, represents the predicted and corrected data motion compensation salinity gradient for the th data point, and represents the reference motion compensation salinity gradient for the Density penalty term: ； Among them, represents the temporarily predicted salinity correction value of the th data point, represents the reference salinity value of the th data point, is the seawater density calculated by the UNESCO equation of state for seawater.

9. A method for correcting ocean salinity data based on spatio-temporal gradient adaptive filtering according to claim 1, wherein, In Step 5, verifying the indicators of the final salinity correction value includes statistical accuracy verification, and specifically as follows: When the mean absolute error of salinity exceeds the threshold, adjust the hyperparameters of the GBDT model as follows: Dynamic adjustment of model complexity: ； ； Among them, is the corrected tree depth, is the corrected number of trees, represents the original decision tree depth search range, represents the original decision tree number search range, represents the model complexity correction coefficient; ； Among them, represents the current mean absolute error of salinity, represents the statistical accuracy threshold, represents the absolute value of the motion-compensated salinity gradient, represents the maximum value of the absolute value of the motion-compensated salinity gradient on the seawater profile.

10. A method for correcting ocean salinity data based on spatio-temporal gradient adaptive filtering according to claim 1, characterized in that, In Step 5, verifying the indicators of the final salinity correction value includes feature retention degree verification, and specifically as follows: The gradient retention rate is used to measure the retention degree of the salinity gradient feature during the correction process, and the gradient retention rate calculation formula is as follows: ； where GPR is the gradient retention rate, represents the salinity gradient of the motion-compensated corrected data at the i-th data point, represents the salinity gradient of the motion-compensated reference data at the i-th data point, and N represents the total number of data points in the salinity profile; When the gradient retention rate is less than the threshold the adaptive dynamic filtering parameters are corrected: Adjustment of the maximum filtering window: ； Adjustment of the salinity gradient-sensitive weight: ； Among them, represents the corrected maximum filtering window, represents the maximum filtering window, represents the salinity gradient sensitive weight, represents the corrected salinity gradient sensitive weight, represents the adjustment coefficient; ； Among them, represents the gradient retention rate threshold.

Images