A vertical gradient error detection method for temperature and salinity data considering correlation

Abstract

The present invention discloses a method for detecting vertical gradient errors of temperature and salinity data taking into account correlation. The method proposes a method for constructing a feature matrix using multivariate sequence data to coordinate the correlation of multi-section and multi-sequence data. The basic method is to calculate the inner product of potential temperature and potential density of different depth layers in the multiple profiles. Secondly, for multi-sequence temperature and salinity data, if the correlation between them is not considered, in the gradient error judgment work, not only is it necessary to perform buoyancy frequency calculation on each profile, which will increase the calculation time and resource cost, but noise may also occur due to failure to consider the influence of the surrounding environment. Moreover, within the same profile, since only the potential temperature and potential density of an adjacent depth layer are considered, the consideration of the correlation between the longitudinal depths is also lacking. Based on this, the present invention utilizes a convolutional neural network and intercepts the depth layer by sliding forward a certain step size, so as to judge the gradient of the current depth layer of the feature matrix.

Description

Technical Field

[0001] The invention relates to a method for detecting vertical gradient errors of temperature and salinity data taking correlation into consideration. Background Art

[0002] In ocean hydrological information, seawater temperature and salinity are key parameters for a comprehensive analysis of ocean conditions. They are often used to calculate the fine structure of seawater layers, such as mesoscale eddies and internal wave formation, and other physical oceanographic phenomena. Understanding these ocean phenomena will contribute to the development of ocean-related industries. The primary method for detecting seawater temperature and salinity observation data and inverting its temperature and salinity structure is to collect historical global ocean temperature and salinity observation data (such as Argo floats), then perform quality control on the observation data, construct parameterized models, and horizontally grid the model coefficients.

[0003] After constructing a parameterized model for global ocean observation data, gridded temperature and salinity data are required for statistical analysis. Using the Cressman stepwise correction scheme, a first guess field for the gridded temperature and salinity profile parameterization model coefficients is constructed. Then, using the gradient-dependent scaling formula, a scale-optimal interpolation model is constructed. Once the model is established, a temperature and salinity parameterization model is constructed at the horizontal grid points to obtain a profile parameterization model and a preliminary temperature and salinity gridded dataset. This preliminary temperature and salinity gridded dataset is then corrected for horizontal gradients using the nonlinear least squares method. Following these corrections, the surface temperature and salinity fields are constructed using the temperature parameterization model, the maximum angle method, and the Akiwa extrapolation method, respectively, to obtain preliminary gridded ocean temperature and salinity statistical products.

[0004] Anomaly Detection in Temperature and Salinity Data: Complexity and variability are a major characteristic of the marine environment. From historical data collection to grid generation, it is difficult to ensure that temperature and salinity data within a profile are free of deviations. Therefore, anomaly detection and quality control of temperature and salinity data are essential throughout the entire background field development process.

[0005] From the perspective of gradient, if we want to find the depth layer where the vertical gradient of the temperature-salinity profile is abnormal, the commonly used method is based on the calculation of buoyancy frequency. The main process is to first calculate the error between salinity and temperature at adjacent depth layers in the profile, thereby obtaining the potential temperature and density, and then calculate the Brunt-Vaisala frequency of the temperature profile along the depth direction, that is, the buoyancy frequency. Generally speaking, the lower limit of the buoyancy frequency is set to 1.5×10 -7 If the value is lower than this, it can be identified as a gradient anomaly. This solution mainly analyzes the characteristics of the hydrophysical structure. However, in the actual hydrological environment, the profiles between adjacent sea areas are also correlated. At the same time, the different causes of the anomaly will also introduce noise into the determination of the standard buoyancy frequency anomaly. Summary of the Invention

[0006] In order to solve the problems existing in the above-mentioned prior art, the present invention proposes a vertical gradient error detection method for temperature-salinity data considering correlation. The method takes into account the correlation between the gradients of temperature-salinity data of multiple profiles. It not only has better accuracy in the detection of gradient error in the sea area, but also has a certain degree of robustness because the characteristic matrix greatly reduces the influence of ocean environmental noise.

[0007] The present invention can be implemented through the following technical solutions:

[0008] A method for detecting vertical gradient errors of temperature and salinity data considering correlation specifically comprises the following steps:

[0009] 1) Select a sea area and calculate the potential temperature and density of a certain section and its adjacent sections;

[0010] 2) Calculate the characteristic matrix for the bit density of all profiles in the sea area;

[0011] 3) Slide the window downward along the cross section to generate h feature matrices;

[0012] 4) The feature matrix is ​​used as input data, embedded into the RNN model, and the feature maps are aggregated to form the reconstructed feature matrix;

[0013] 5) Use the loss function to calculate the error between the reconstructed matrix and the original feature matrix;

[0014] 6) Determine whether the loss function value exceeds the threshold. If so, mark the depth layer as "there is gradient error". If not, mark the depth layer as "there is no gradient error".

[0015] Furthermore, the step 1) is specifically as follows:

[0016] The vertical gradient correction work carried out on gridded data still depends on the physical properties of seawater. Therefore, it needs to be determined based on the variable of buoyancy frequency. The calculation formula of buoyancy frequency involves two quantities: the potential density and depth of the selected area. The calculation of potential density is derived from potential temperature:

[0017] The calculation formula of potential temperature is as follows:

[0018] t=T×1.00024

[0019] h=P r -P

[0020] xk=h×Γ(S,t,P)

[0021] t=t+0.5×xk

[0022] q=xk

[0023] p=P+0.5×h

[0024] xk=h×Γ(S,t,p)

[0025] t=t+0.29289322×(xk-q)

[0026] q=0.58578644×xk+0.121320344×q

[0027] xk=h×Γ(S,t,p)

[0028] t=t+1.707106781×(xk-q)

[0029] q=3.414213562×xk-4.121320344×q

[0030] p=p+0.5×h

[0031] xk=h×Γ(S,t,p)

[0032] θ(S, T, P, P r )={t+(xk-2.0×q)÷6.0}×0.99976 (1)

[0033] Where Γ(S,t,p) is defined as follows:

[0034] Γ(S,t,p)=a0+a1t+a2t 2 +a3t 3 +(b0+b1t)(S-35)+{c0+c1t+c2t 2 +c3t 3 +(d0+d1t)(S-35)}p+(e0+e1t+e2t 2 )p 2

[0035] The parameters in the above formula are set as follows:

[0036] a0=3.5803×10 -5 a1=8.5258×10 -6 a²=-6.8360×10 -8 a3=6.6228×10 -10

[0037] b0=1.8932×10 -6 b1=-4.2393×10 -8

[0038] c0=1.8741×10 -8 c1=-6.7795×10 -10c2=8.7330×10 -12 c3=-5.4481×10 -14

[0039] d0=-1.1351×10 -10 d1=2.7759×10 -12

[0040] e0=-4.6206×10 -13 e1=1.8676×10 -14 e²=-2.1687×10 -16

[0041] After obtaining the potential temperature θ, the potential density ρ is solved using the potential temperature. The solution formula is as follows:

[0042]

[0043] t=θ×1.00024

[0044] p=P÷10

[0045] ρ(S,t,0)=ρ w +(b0+b1t+b2t 2 +b3t 3 +b4t 4 )S+(c0+c1t+c2t 2 )S 3 / 2 +d0S 2

[0046] ρ w =a0+a1t+a2t 2 +a3t 3 +a4t 4 +a5t 5

[0047] K(S,t,p)=K(S,t,0)+Ap+Bp 2

[0048] K(S,t,0)=K w +(f0+f1t+f2t 2 +f3t 3 )S+(g0+g1t+g2t 2 )S 3 / 2

[0049] A=A w +(i0+i1t+i2t 2 )S+j0S 3 / 2

[0050] B=B w+(m0+m1t+m2t 2 )S

[0051] K w =e0+e1t+e2t 2 +e3t 3 +e4t 4

[0052] A w =h0+h1t+h2t 2 +h3t 3

[0053] B w =k0+k1t+k2t 2

[0054] The parameter coefficients involved in the above formula are set as follows:

[0055] a0=999.842594 a1=6.793952×10 -2 a²=-9.095290×10 -3 a3=1.001685×10 -4 a4=-1.120083×10 -6 a5=6.536332×10 -9

[0056] b0=8.24493×10 -1 b1=-4.0899×10 -4 b2=7.6438×10 -5 b3=-8.2467×10 -7 b4=5.3875×10 -9

[0057] c0=-5.72466×10 -3 c1=1.0227×10 -4 c²=-1.6546×10 -6

[0058] d0=4.8314×10 -4

[0059] e0=19652.21 e1=148.4206 e2=-2.327105 e3=1.360477×10 -2 e4=-5.155288×10 -5

[0060] f0=54.6746 f1=-0.603459 f2=1.09987×10-2 f3=-6.1670×10 -5

[0061] g0=7.944×10 -2 g1=1.6483×10 -2 g2=-5.3009×10 -4

[0062] h0=3.239908 h1=1.43713×10 -3 h2=1.16092×10 -4 h3=-5.77905×10 -7

[0063] i0=2.2838×10 -3 i1=-1.0981×10 -5 i²=-1.6078×10 -6

[0064] j0=1.91075×10 -4

[0065] k0=8.50935×10 -5 k1=-6.12293×10 -6 k²=5.2787×10 -8

[0066] m0=-9.9348×10 -7 m1=2.0816×10 -8 m²=9.1697×10 -10 .

[0067] Furthermore, the step 2) is specifically as follows:

[0068] The characteristic matrix is ​​constructed from the bit-density data of the profile to be measured and another adjacent profile as input data; assuming that the historical bit-density data of the two profiles at the depth of h layers forward are: Then, assuming the generated feature matrix M, the element in the i-th row and j-th column is calculated as:

[0069]

[0070] Among them, the denominator k is the reconstruction factor. By calculating with this scheme, all the bit-mass data profiles in the sea area to be measured will eventually generate a characteristic matrix M. The dimension of the matrix is ​​the number of profiles in the sea area, and this characteristic matrix is ​​also a matrix symmetric along the diagonal.

[0071] Furthermore, the step 3) is specifically as follows:

[0072] After calculating the feature matrix required for the current depth layer, the time window with a step size of h is pushed forward one step along the next depth layer of all profiles within the coastal area to calculate the feature matrix after the step; then, it is necessary to obtain the feature matrix of h steps, aggregate the representation of these information feature maps, and form a more refined feature map output The calculation method is based on the recurrent neural network RNN ​​with a step size of h, and the hidden state updated at each step in the RNN is required. The calculation formula is as follows:

[0073]

[0074] Where Vec represents the vectorization operation of the hidden state matrix, and χ is the reconstruction factor.

[0075] Furthermore, the step 5) is specifically as follows:

[0076] Design a loss function to compare the reconstructed feature matrix with the feature matrix obtained from the initial observation value. The calculation method is as follows:

[0077]

[0078] Furthermore, in step 6), when selecting the threshold, from the perspective of buoyancy frequency, the buoyancy frequency N 2 Whether it exceeds the set minimum value The buoyancy frequency is calculated as follows:

[0079]

[0080] Beneficial effects

[0081] 1) The present invention uses the bit density of all temperature and salinity data detection profiles in the selected sea area to construct a feature matrix, thereby coordinating the correlation of the temperature and salinity data of the water flow corresponding to adjacent profiles in the sea area, and the subsequent gradient error correction can also take into account the correlation of the adjacent profile data;

[0082] 2) Although the present invention adopts the method of moving the sliding window downward to iteratively select the feature matrix corresponding to h steps in the upper layer as historical data, not all upper temperature and salinity data are related to the vertical gradient of the current layer. Therefore, the present invention uses RNN to adaptively select the upper layer depth related to the current depth, which can reduce the influence of noise and irrelevant water flow on the reconstructed matrix, thereby improving accuracy and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 This is a flowchart of the working principle of RNN recurrent neural network;

[0084] Figure 2 This is a flow chart of the vertical gradient error detection method. DETAILED DESCRIPTION

[0085] The following describes the embodiments of the present invention through specific embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification.

[0086] refer to Figure 2 As shown, a method for detecting vertical gradient errors of temperature and salinity data taking into account correlation of the present invention comprises the following steps:

[0087] (1) First, it is necessary to calculate the position density of different depth layers of the profile. The vertical gradient correction work carried out on the grid data still depends on the physical properties of the seawater. Therefore, it is necessary to make a judgment based on the variable of buoyancy frequency. The calculation formula of buoyancy frequency involves the position density and depth of the selected area. The calculation of position density is derived from the potential temperature.

[0088] The calculation formula of potential temperature is as follows:

[0089] t=T×1.00024

[0090] h=P r -P

[0091] xk=h×Γ(S,t,P)

[0092] t=t+0.5×xk

[0093] q=xk

[0094] p=P+0.5×h

[0095] xk=h×Γ(S,t,p)

[0096] t=t+0.29289322×(xk-q)

[0097] q=0.58578644×xk+0.121320344×q

[0098] xk=h×Γ(S,t,p)

[0099] t=t+1.707106781×(xk-q)

[0100] q=3.414213562×xk-4.121320344×q

[0101] p=p+0.5×h

[0102] xk=h×Γ(S,t,p)

[0103] θ(S, T, P, P r )={t+(xk-2.0×q)÷6.0}×0.99976 (1)

[0104] Where Γ(S,t,p) is defined as follows:

[0105] Γ(S,t,p)=a0+a1t+a2t 2 +a3t 3 +(b0+b1t)(S-35)+{c0+c1t+c2t 2 +c3t 3 +(d0+d1t)(S-35)}p+(e0+e1t+e2t 2 )p 2

[0106] The parameters in the above formula are set as follows:

[0107] a0=3.5803×10 -5 a1=8.5258×10 -6 a²=-6.8360×10 -8 a3=6.6228×10 -10

[0108] b0=1.8932×10 -6 b1=-4.2393×10 -8

[0109] c0=1.8741×10 -8 c1=-6.7795×10 -10 c2=8.7330×10 -12 c3=-5.4481×10 -14

[0110] d0=-1.1351×10 -10 d1=2.7759×10 -12

[0111] e0=-4.6206×10 -13 e1=1.8676×10 -14 e²=-2.1687×10 -16

[0112] After obtaining the potential temperature θ, the potential density ρ is solved using the potential temperature. The solution formula is as follows:

[0113]

[0114] t=θ×1.00024

[0115] p=P÷10

[0116] ρ(S,t,0)=ρ w +(b0+b1t+b2t 2 +b3t 3 +b4t 4 )S+(c0+c1t+c2t 2 )S 3 / 2 +d0S 2

[0117] ρ w =a0+a1t+a2t 2 +a3t 3 +a4t 4 +a5t 5

[0118] K(S,t,p)=K(S,t,0)+Ap+Bp 2

[0119] K(S,t,0)=K w +(f0+f1t+f2t 2 +f3t 3 )S+(g0+g1t+g2t 2 )S 3 / 2

[0120] A=A w +(i0+i1t+i2t 2 )S+j0S 3 / 2

[0121] B=B w +(m0+m1t+m2t 2 )S

[0122] K w =e0+e1t+e2t 2 +e3t 3 +e4t 4

[0123] A w =h0+h1t+h2t 2 +h3t 3

[0124] B w =k0+k1t+k2t 2

[0125] The parameter coefficients involved in the above formula are set as follows:

[0126] <h2 style=";text-align:left;direction:ltr">a0=999.842594 a1=6.793952×10<h2 style=";text-align:left;direction:ltr"> -2 <h2 style=";text-align:left;direction:ltr"> a2 = 9.095290×10<h2 style=";text-align:left;direction:ltr"> -3 <h2 style=";text-align:left;direction:ltr"> a3 = 1.001685×10<h2 style=";text-align:left;direction:ltr"> -4 <h2 style=";text-align:left;direction:ltr"> a4 = 1.120083×10<h2 style=";text-align:left;direction:ltr"> -6 <h2 style=";text-align:left;direction:ltr"> a5 = 6.536332×10<h2 style=";text-align:left;direction:ltr"> -9 <h2 style=";text-align:left;direction:ltr"> <h2 style=";text-align:left;direction:ltr">

[0127] <h2 style=";text-align:left;direction:ltr"> b0 = 8.24493×10<h2 style=";text-align:left;direction:ltr"> -1 <h2 style=";text-align:left;direction:ltr"> b1 = -4.0899×10<h2 style=";text-align:left;direction:ltr"> -4 <h2 style=";text-align:left;direction:ltr"> b2 = 7.6438×10<h2 style=";text-align:left;direction:ltr"> -5 <h2 style=";text-align:left;direction:ltr"> b3 = -8.2467×10<h2 style=";text-align:left;direction:ltr"> -7 <h2 style=";text-align:left;direction:ltr"> b4 = 5.3875×10<h2 style=";text-align:left;direction:ltr"> -9 <h2 style=";text-align:left;direction:ltr"> <h2 style=";text-align:left;direction:ltr">

[0128] <h2 style=";text-align:left;direction:ltr"> c0 = -5.72466×10<h2 style=";text-align:left;direction:ltr"> -3 <h2 style=";text-align:left;direction:ltr"> c1 = 1.0227×10<h2 style=";text-align:left;direction:ltr"> -4 <h2 style=";text-align:left;direction:ltr"> c2 = -1.6546×10<h2 style=";text-align:left;direction:ltr"> -6 <h2 style=";text-align:left;direction:ltr"> <h2 style=";text-align:left;direction:ltr">

[0129] <h2 style=";text-align:left;direction:ltr"> d0 = 4.8314×10<h2 style=";text-align:left;direction:ltr"> -4 <h2 style=";text-align:left;direction:ltr"> <h2 style=";text-align:left;direction:ltr">

[0130] <h2 style=";text-align:left;direction:ltr"> e0 = 19652.21 e1 = 148.4206 e2 = - 2.327105 e3 = 1.360477×10<h2 style=";text-align:left;direction:ltr"> -2 <h2 style=";text-align:left;direction:ltr"> e4 = -5.155288×10<h2 style=";text-align:left;direction:ltr"> -5 <h2 style=";text-align:left;direction:ltr"> <h2 style=";text-align:left;direction:ltr">

[0131] <h2 style=";text-align:left;direction:ltr"> f0 = 54.6746 f1 = -0.603459 f2 = 1.09987×10<h2 style=";text-align:left;direction:ltr"> -2 <h2 style=";text-align:left;direction:ltr"> f3 = 6.1670×10<h2 style=";text-align:left;direction:ltr"> -5 <h2 style=";text-align:left;direction:ltr"> <h2 style=";text-align:left;direction:ltr">

[0132] <h2 style=";text-align:left;direction:ltr"> g0 = 7.944×10<h2 style=";text-align:left;direction:ltr"> -2 <h2 style=";text-align:left;direction:ltr"> g1 = 1.6483×10<h2 style=";text-align:left;direction:ltr"> -2 <h2 style=";text-align:left;direction:ltr"> g2 = -5.3009×10<h2 style=";text-align:left;direction:ltr"> -4 <h2 style=";text-align:left;direction:ltr"> <h2 style=";text-align:left;direction:ltr">

[0133] <h2 style=";text-align:left;direction:ltr"> h0=3.239908 h1=1.43713×10<h2 style=";text-align:left;direction:ltr"> -3 <h2 style=";text-align:left;direction:ltr"> h2 = 1.16092×10<h2 style=";text-align:left;direction:ltr"> -4 <h2 style=";text-align:left;direction:ltr"> h3 = -5.77905×10<h2 style=";text-align:left;direction:ltr"> -7 <h2 style=";text-align:left;direction:ltr"> <h2 style=";text-align:left;direction:ltr">

[0134] <h2 style=";text-align:left;direction:ltr"> i0 = 2.2838×10<h2 style=";text-align:left;direction:ltr"> -3 <h2 style=";text-align:left;direction:ltr"> i1 = -1.0981×10<h2 style=";text-align:left;direction:ltr"> -5i²=-1.6078×10 -6

[0135] j0=1.91075×10 -4

[0136] k0=8.50935×10 -5 k1=-6.12293×10 -6 k²=5.2787×10 -8

[0137] m0=-9.9348×10 -7 m1=2.0816×10 -8 m²=9.1697×10 -10

[0138] When solving the above equation, although many parameters are used, their values ​​are all known;

[0139] (2) Based on the bit density obtained from the multi-profile observation data in the surrounding area, the depth layer to be detected is advanced by h steps in a sliding window-like manner, and the bit density of the corresponding depth layer is regarded as the historical input value. Subsequently, a feature matrix is ​​generated based on the historical input values ​​of the adjacent multi-profiles. It can use the multivariate bit density dataset of a certain sea area to systematically represent the distribution of temperature and salinity data in the sea area, thereby finding the vertical gradient error in the three-dimensional sea area and taking into account the correlation of temperature and salinity data of multiple profiles.

[0140] The present invention constructs a feature matrix using the bit-encrypted data of the profile to be measured and another adjacent profile as input data. Assume that the historical bit-encrypted data of the two profiles at the depth of h layers forward are: Then, assuming the generated feature matrix M, the element in the i-th row and j-th column is calculated as:

[0141]

[0142] The denominator k is the reconstruction factor. This calculation scheme ultimately generates a characteristic matrix M for all bit-encrypted data profiles within the measured sea area. The matrix dimension is the number of profiles within the sea area, and this characteristic matrix is ​​also symmetric along the diagonal. Furthermore, for depth layers close to the sea surface, correlation gradient error detection can be achieved by further refining the depth layer in smaller depth units.

[0143] (3) After calculating the feature matrix required for the current depth layer, the time window with a step length of h is pushed forward one step to the next depth layer of all profiles within the coastal area, and the feature matrix after the step is calculated. Subsequently, it is necessary to obtain the feature matrix of h steps, aggregate the representation of these information feature maps, and form a more refined feature map output In the present invention The calculation method is based on the recurrent neural network RNN ​​with a step size of h, and the hidden state updated at each step in the RNN is required. Based on this, The calculation formula is as follows:

[0144]

[0145] Where Vec represents the vectorization operation of the hidden state matrix, and χ is the reconstruction factor. The calculation process is simply summarized as taking the last hidden layer, that is, the hidden layer at the current depth, as the group-level context vector, and then using the softmax function to measure the weights α in the depth layers corresponding to the previous steps. i , and then accumulate the hidden layer output of each step in the RNN model to obtain the feature map The basic working principle of RNN is shown in the figure below. Unlike traditional recurrent neural networks, the gating units used in the present invention are all two-dimensional tensors.

[0146] (4) The cyclic neural network aggregates the hidden layer in the above way to obtain Essentially, the reconstructed feature matrix is ​​obtained by combining the current depth layer with the bit density of the previous layers using a neural network. To determine whether there is a gradient error, it is necessary to compare it with the feature matrix corresponding to the current depth layer. This invention uses a loss function to compare the reconstructed feature matrix with the feature matrix obtained from the initial observation value. The calculation method is shown in the following formula:

[0147]

[0148] If the calculated loss value does not exceed the threshold, the depth area is classified as "normal gradient error", otherwise the depth layer is marked as "abnormal". When selecting the threshold, from the perspective of buoyancy frequency, the buoyancy frequency N 2 Whether it exceeds the set minimum value The buoyancy frequency is calculated as follows:

[0149]

[0150] When setting the loss function threshold of the characteristic matrix, it can also be designed according to the physical characteristics of the ocean related to buoyancy frequency.

[0151] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for detecting vertical gradient errors of temperature and salinity data considering correlation, characterized in that: The specific steps include: 1) Select a sea area and calculate the potential temperature and density of a certain section and its adjacent sections; 2) Calculate the characteristic matrix for the bit density of all profiles in the sea area; 3) Slide the window downward along the cross section to generate h feature matrices; 4) The feature matrix is ​​used as input data, embedded into the RNN model, and the feature maps are aggregated to form the reconstructed feature matrix; 5) Use the loss function to calculate the error between the reconstructed matrix and the original feature matrix; 6) Determine whether the loss function value exceeds the threshold. If so, mark the depth layer as "with gradient error". If not, mark the depth layer as "without gradient error". The step 2) is specifically as follows: The characteristic matrix is ​​constructed from the bit-density data of the profile to be measured and another adjacent profile as input data; assuming that the historical bit-density data of the two profiles at the depth of h layers forward are: Then, assuming the generated feature matrix M, the element in the i-th row and j-th column is calculated as: Among them, the denominator K is the reconstruction factor. By calculating with this scheme, all the bit-mass data profiles in the sea area to be measured will eventually generate a characteristic matrix M. The dimension of the matrix is ​​the number of profiles in the sea area, and this characteristic matrix is ​​also a matrix symmetric along the diagonal.

2. The method for detecting vertical gradient errors of temperature and salinity data considering correlation according to claim 1, characterized in that: The step 3) is specifically as follows: After calculating the feature matrix required for the current depth layer, the time window with a step size of h is pushed forward one step along the next depth layer of all profiles within the coastal area to calculate the feature matrix after the step; then, it is necessary to obtain the feature matrix of h steps, aggregate the representation of these information feature maps, and form a more refined feature map output The calculation method is based on the recurrent neural network RNN ​​with a step size of h, and the hidden state updated at each step in the RNN is required. The calculation formula is as follows: Where Vec represents the vectorization operation of the hidden state matrix, and χ is the reconstruction factor.

3. The method for detecting vertical gradient errors of temperature and salinity data considering correlation according to claim 1, characterized in that: The step 5) is specifically as follows: Design a loss function to compare the reconstructed feature matrix with the feature matrix obtained from the initial observation value. The calculation method is as follows:

4. The method for detecting vertical gradient errors of temperature and salinity data considering correlation according to claim 1, characterized in that: In step 6), when selecting the threshold, from the perspective of buoyancy frequency, the buoyancy frequency N 2 Whether it exceeds the set minimum value The buoyancy frequency is calculated as follows:

Images