A method and system for assimilating non-gaussian distribution data of meteorological satellite hyperspectral errors based on quantum computing

Abstract

The application discloses a method and system for assimilating meteorological satellite hyperspectral error non-Gaussian distribution data based on quantum computing, which performs generalized quality control of meteorological satellite hyperspectral data considering outliers on the observed brightness temperature of a satellite hyperspectral infrared detector channel and equivalent data thereof, including optimal selection of the channel considering quantum information entropy, deviation correction based on artificial intelligence generalized integrated generative deep learning, and cloud detection and removal of cloud view point data based on the minimum residual method, thereby obtaining quality control data. Quantum non-Gaussian increment four-dimensional variation assimilation models compatible with error Gaussian distribution and non-Gaussian distribution are solved based on the quality control data, and analysis field data are obtained. The method solves the problem that the existing technology is limited in the assimilation of hyperspectral infrared channel brightness temperature with error non-Gaussian distribution (or non-Gaussian characteristics are obvious), and meanwhile, the solving speed of the method is fast, and the obtained analysis field data have good precision and calculation timeliness.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of atmospheric science, and particularly relates to a method and system for assimilating meteorological satellite hyperspectral error non-Gaussian distribution data based on quantum computing. BACKGROUND

[0002] Data assimilation is a key component in the field of atmospheric science, which enables the combination of observation data and numerical weather prediction models. The quality of numerical weather prediction largely depends on the accuracy of the initial conditions provided by the assimilation system. In numerical weather prediction operational systems, the assimilation of satellite data, especially hyperspectral infrared data, accounts for a high proportion. Data or data assimilation is to effectively integrate all available information with the background field to estimate the atmospheric state at a certain time as accurately as possible, i.e. the analysis field. The analysis field obtained by data assimilation can provide higher quality initial values for numerical weather prediction.

[0003] The internationally accepted classical variational data assimilation requires that the error (the error belongs to the mathematical category, and is called bias in the field of meteorological satellites) obeys Gaussian distribution. The variational assimilation requires that the error obeys Gaussian distribution, and for satellite detector data assimilation, it is required that the channel brightness temperature bias satisfies Gaussian distribution. The so-called brightness temperature bias is the difference between the actual observed brightness temperature of the satellite channel and the simulated brightness temperature. The simulated brightness temperature is obtained by simulating and interpolating the background field based on the spectral coefficient of the satellite detector. Through theoretical analysis and a large number of experimental researches, scholars at home and abroad have shown that the error obeys Gaussian distribution is only a theoretical assumption, and in practice, many data errors present non-Gaussian distribution, so there is a certain accuracy problem in the analysis field obtained by using the classical variational data assimilation, which affects the subsequent numerical prediction results. For example, the non-Gaussianity of the hyperspectral infrared water vapor channel brightness temperature data of meteorological satellites is relatively strong. If the international classical variational method is used to assimilate the hyperspectral infrared water vapor channel brightness temperature, the hyperspectral infrared water vapor channel brightness temperature with relatively strong non-Gaussianity is not used well or is limited in use, and many data will be excluded by the quality control before the variational assimilation, resulting in the loss of information brought by many "available" data, which affects the numerical prediction results of weather wet processes (such as heavy rain and typhoon). SUMMARY

[0004] To solve the problems in the prior art, the application provides a method and system for assimilating meteorological satellite hyperspectral error non-Gaussian distribution data based on quantum computing.

[0005] The technical scheme of the application is as follows:

[0006] A method for assimilating meteorological satellite hyperspectral error non-Gaussian distribution data based on quantum computing, the method comprising:

[0007] The channel simulated brightness temperature is subjected to generalized quality control of meteorological satellite hyperspectral data considering outliers, to form quality control data, and the generalized quality control includes channel optimal selection considering quantum information entropy, bias correction based on artificial intelligence generalized ensemble generative deep learning, and cloud detection removal based on the minimum residual method.

[0008] The channel simulated brightness temperature is subjected to generalized quality control of meteorological satellite hyperspectral data considering outliers, to form quality control data, and the generalized quality control includes channel optimal selection considering quantum information entropy, bias correction based on artificial intelligence generalized ensemble generative deep learning, and cloud detection removal based on the minimum residual method.

[0009] A quantum non-Gaussian increment four-dimensional variational assimilation model compatible with Gaussian distribution and non-Gaussian distribution of errors is constructed, and a dynamically adjustable weight contribution rate factor is introduced in the observation term of the cost function of the quantum non-Gaussian increment four-dimensional variational assimilation model.

[0010] Based on the quality control data, the quantum non-Gaussian increment four-dimensional variational assimilation model is solved to obtain analysis field data.

[0011] Further, the specific method of the channel optimal selection considering quantum information entropy includes:

[0012] 1) Remove satellite channels whose channel simulated brightness temperature error absolute value exceeds 10K; the channel simulated brightness temperature error is defined as the difference between the channel observed brightness temperature and the channel simulated brightness temperature.

[0013] 2) Based on the peak value layer of the satellite channel weight function, that is, the different atmospheric layer information detected by different satellite channels, the satellite channels are preliminarily selected to obtain a preliminary selected channel combination.

[0014] 3) Optimal channel selection considering quantum information entropy is performed on the preliminary selected channel combination.

[0015] Further, the channel observed brightness temperature is obtained by nonlinear brightness temperature reconstruction based on compressed sensing on meteorological satellite hyperspectral infrared channel observed brightness temperature data, and the specific method is as follows:

[0016] The meteorological satellite hyperspectral infrared channel observed brightness temperature data is subjected to sparse representation in the wavelet domain space, and the original brightness temperature signal is reconstructed by decoding, and the specific expression is as follows:

[0017]

[0018] Where x true is the reconstructed original brightness temperature signal; y signal is the observed compressed brightness temperature signal, i.e. the received brightness temperature signal; H CS is a structure operator; R signal is the brightness temperature signal error; W CS represents a transformation matrix; λCS This represents the regularization parameter.

[0019] Furthermore, the specific method for selecting the optimal channel by considering quantum information entropy from the initially selected channel combinations includes:

[0020] 1) Normalize the observed brightness temperature sequence data corresponding to each channel of the hyperspectral infrared detector in the initially selected channel combination; for the observed brightness temperature sequence data X of the i-th channel in the channel combination... i,entropy ={x i,entropy Let Y(j), j=1,2,...,N}, and let Y be the normalized observed brightness temperature sequence data of the i-th channel. i,entropy ={y i,entropy (j), j=1,2,...,N};where, x i,entropy (j) represents the j-th observed brightness temperature sequence data of the i-th channel; j is the sequence number of the observed brightness temperature sequence data, and N is the total amount of observed brightness temperature sequence data; y i,entropy (j) represents the normalized brightness temperature sequence data of the j-th observed channel;

[0021] 2) Perform quantum space reconstruction on the normalized sequence data of each channel to obtain the quantum space matrix corresponding to each channel; for the normalized observed brightness temperature sequence data Y of the i-th channel... i,entropy Its corresponding quantum space matrix The expression is as follows:

[0022]

[0023] In the formula, Let l be the l-th reconstructed component in the quantum space matrix, where l = 1, 2, ..., K, and K is the total number of reconstructed components. m is the embedding dimension; For time windows;

[0024] 3) Quantize each reconstructed component in the quantum space matrix of each channel and calculate the quantum information entropy of each channel; for the quantum space matrix... Each reconstructed component is quantized as follows:

[0025]

[0026] In the formula, |q b > indicates reconstructed components The q-th state vector; ω l,k Representing the state vector |q b The probability magnitude of >, k = 1, 2, ..., n, where n is the number of state vectors, n = 2 m, m is the amount of constituent qubits;

[0027] The probability P of the state vector appearing i,k As the probability of the event, the quantum space matrix The quantum information entropy of Wherein, k=1, 2,..., n;

[0028] 4) Based on the quantum information entropy and the variational assimilation background error covariance matrix of each channel, the quantum mutual information of each channel is calculated, a multi-dimensional feature space composed of multiple groups of channels is formed, and the satellite channel corresponding to the maximum quantum mutual information is searched in the multi-dimensional feature space and used as the optimal channel.

[0029] Further, the specific method of bias correction based on artificial intelligence generalized ensemble generative deep learning includes:

[0030] 1) Select the bias correction predictor;

[0031] 2) For all field point observation brightness temperature equivalent data of each channel of the meteorological satellite hyperspectral infrared detector that needs to be corrected, a deep model generated based on the combination of convolutional neural network and deep neural network model is constructed as the basic model of ensemble learning, and generalized weighted ensemble machine learning is performed to find the optimal or suboptimal ensemble weight of integrating the prediction results of the basic model, and the objective minimum function of the weighted ensemble machine learning is:

[0032]

[0033] Wherein, n is the total number of samples; is the actual value of the value to be corrected i. is the prediction value of the basic model j to the value to be corrected i, and the value of the prediction value is obtained based on the selected predictor as the input quantity of the basic model; is the integrated weight corresponding to the basic model j.

[0034] Further, the specific method of cloud detection and removal based on the least residual method includes:

[0035] 1) According to the radiation transfer theory, the simulated radiation value of the meteorological satellite hyperspectral infrared detector channel is mathematically modeled, and the simulated radiation value of the satellite channel g is Indicated as: Wherein, N e Indicates the effective cloud amount of the field point; Indicates the clear sky radiation; Indicates the cloudy radiation simulation when the cloud top pressure is p c,g

[0036] ​2) Calculate the deviation of the observed radiation value and the simulated radiation value of the hyperspectral infrared detector channel, the observed radiation value of channel g and the simulated radiation value The deviation δ g is expressed as:

[0037] 3) Based on the n channel combinations obtained by the optimal channel selection of quantum information entropy, the minimum residual method objective function is minimized to obtain the effective cloud amount N e and the effective cloud top pressure p c of the field point, and the minimum residual method objective function is defined as follows:

[0038]

[0039] 4) Based on the effective cloud amount N e and the set threshold, the cloud detection of the field point is carried out, and for the field point with the effective cloud amount N e greater than the set threshold, all brightness temperature data of all channel combinations on the field point are removed.

[0040] Further, the cost function of the quantum non-Gaussian increment four-dimensional variational assimilation model is:

[0041]

[0042] w(r i )=(1 / r i )·(dρ(r i ) / dr i )

[0043] In the formula, δx0 is the analysis increment, x0 is the control variable, and the solved x0 is the analysis field; is the state vector predicted by the numerical prediction model; the superscript f indicates prediction; B is the background error covariance matrix; R i is the observation error covariance matrix; the superscript T is the matrix transpose; the superscript -1 is the matrix inverse; d i is the brightness temperature increment, y i represents the i-th observation value, i=1, 2, …, N, and N is the total number of observation values; represents the brightness temperature value of the i-th observation value simulated based on the model space through the fast radiation transfer model, that is, the equivalent data of the i-th observation value, H i is an observation operator for mapping the model integral M i,0 (·) solution from the model space to the observation space; M i,0(x0) is a non-linear numerical prediction model that integrates control variables x0 from initial conditions and initial time t = 0 to t = i, t is a time marker; matrix H i and M i,0 represent the linear representation of the tangent mode of H i and M i,0 respectively; r i is the brightness temperature increment amplitude value, r i = d i / σ i , σ i is the observation error of channel i; w(r i ) is the weight contribution rate factor; ρ(r i ) is the M-estimate cost function.

[0044] Further, the specific method for solving the quantum non-Gaussian increment four-dimensional variational assimilation model comprises:

[0045] 1) An increment four-dimensional non-Gaussian quantum data assimilation method driven by quantum annealing is adopted, real number representation is performed through Q qubits, and the analysis increment δx0 is quantized through a mapping matrix G:

[0046]

[0047] g T = [-2 Q-1 , 2 Q-2 , 2 Q-3 , …, 2 1 , 2 0 ]

[0048]

[0049] In the formula, β is an adjustable scaling parameter; b is a binary vector whose elements are 0 or 1, and 0 is a vector whose elements are all 0; g is the quantization of a real number; Q is a natural number;

[0050] 2) Based on the quantized analysis increment, the cost function of the quantum non-Gaussian increment four-dimensional variational assimilation model is solved in quantum space:

[0051]

[0052] In the formula,

[0053]

[0054] In the formula, Ham represents the Hamiltonian; w(·) represents the weight function of the constructed M-estimate new norm; A and u are as follows:

[0055] 3) The cost function is minimized by a quantum annealing method Obtaining the best analysis field.

[0056] Further, it also includes the evaluation of the analysis field data on the accuracy, timeliness and the analysis field on the improvement of the wet process numerical weather prediction accuracy.

[0057] A system for assimilating meteorological satellite hyperspectral error non-Gaussian distribution data based on quantum computing, the system comprising a first data processing module, a second data processing module, a third data processing module, and a fourth data processing module.

[0058] The first data processing module is configured to simulate the radiative transfer model of the meteorological satellite hyperspectral channel data after the background field data interpolation processing, and obtain the corresponding satellite hyperspectral channel observation brightness temperature equivalent data as the channel simulated brightness temperature.

[0059] The second data processing module is configured to perform generalized quality control of the meteorological satellite hyperspectral data on the channel simulated brightness temperature to form quality control data, wherein the generalized quality control includes channel optimal selection considering quantum information entropy, bias correction based on artificial intelligence generalized integrated generative deep learning, and cloud detection removal based on the minimum residual method.

[0060] The third data processing module is configured to construct a quantum non-Gaussian increment four-dimensional variational assimilation model compatible with error Gaussian distribution and non-Gaussian distribution, wherein a dynamically adjustable weight contribution rate factor is introduced in the observation term of the cost function of the quantum non-Gaussian increment four-dimensional variational assimilation model.

[0061] The fourth data processing module is configured to solve the quantum non-Gaussian increment four-dimensional variational assimilation model based on the quality control data to obtain the analysis field data.

[0062] Compared with the prior art, the present application has the following beneficial effects:

[0063] The application provides a method and system for assimilating high-spectral error non-Gaussian distribution data of a meteorological satellite based on quantum computing, which is characterized in that the method is based on the equivalent data of the observed brightness temperature of the satellite high-spectral infrared detector channel to perform generalized quality control of the observed brightness temperature data of the high-spectral channel, so that the quality control data are obtained, and the quality of the data variational assimilation is ensured; and the quantum non-Gaussian increment four-dimensional variational assimilation model compatible with error Gaussian distribution and non-Gaussian distribution is solved in a quantum space based on the quality control data, so that the analysis field data are obtained, the method solves the problem that the assimilation of the high-spectral infrared channel brightness temperature with error non-Gaussian distribution (or non-Gaussian characteristics) is limited in the prior art, effectively assimilates the high-spectral water vapor channel brightness temperature, and improves the numerical prediction accuracy of the wet processes such as precipitation and typhoon; meanwhile, the solving speed of the method is fast, the analysis field data obtained have good accuracy and calculation timeliness, and the accuracy of the numerical weather prediction of the wet processes can be effectively improved.

[0064] In the generalized quality control, the method adopts the channel optimal selection considering quantum information entropy based on a three-step method, the entropy reduction method commonly used in the field is to select the channel by judging the reduction amplitude of the analysis error after updating the analysis field. Different from the general entropy reduction method, the channel optimal selection of the application can accurately represent the advantages of the objective existence law based on the superposition, coherence and entanglement of the quantum state.

[0065] In the method, the channel observed brightness temperature is obtained by performing nonlinear brightness temperature reconstruction based on compressed sensing on the high-spectral infrared channel observed brightness temperature data of the meteorological satellite, which is different from the conventional principal component analysis in the field, which is difficult to remove the "stripe" noise in the reconstruction of the satellite high-spectral observed brightness temperature. The application proposes nonlinear brightness temperature reconstruction based on compressed sensing, which can effectively remove the "stripe" noise, and can obtain better reconstruction results with fewer random sampling samples by introducing sparse representation in the wavelet domain space. In the generalized quality control, the method adopts a bias correction method based on artificial intelligence generalized ensemble generative deep learning. Different from the common "offline" and "online" bias correction methods in the field, the application proposes a bias correction method based on artificial intelligence generalized ensemble generative deep learning combined with the bias characteristics of the high-spectral channel brightness temperature, generates a deep ensemble learning base model based on the combination of a convolutional neural network (CNN) and a deep neural network (DNN) model. The prediction results of the weighted integrated different base models are used for bias correction of the high-spectral channel brightness temperature.

[0066] The method of the application carries out new norm quantum non-Gaussian increment four-dimensional variational assimilation based on M-estimation. Different from the assumption that the observation error obeys Gaussian distribution in the classical variational assimilation, the application is based on the basic properties of M-estimation cost function (such as continuity, non-negativity, convexity, etc.) and the basic properties of weight function (such as continuity, monotonicity, etc.) to construct a new norm on the basis of the classical M-estimation method (such as Huber-estimation). The M-estimation new norm is coupled to the cost function of the classical variational assimilation, and a quantum non-Gaussian variational assimilation method is proposed in the quantum state space. The non-Gaussian variational assimilation solves the defect that the classical variational method can only assimilate data error obeying Gaussian distribution. The key core technology of the non-Gaussian variational assimilation is that, different from the classical variational assimilation which removes some "outliers" with larger errors before minimization iteration, the non-Gaussian variational assimilation method introduces the weight function of the new M-estimation norm as the weight contribution rate factor in the observation term of the variational assimilation cost function, reconstructs the new cost function of the variational assimilation, and uses the "outliers", but reduces the contribution of the "outliers" to the cost function. The new non-Gaussian variational assimilation method solves the problem of limited assimilation of hyperspectral infrared water vapor channel brightness temperature, and improves the accuracy of typhoon and rainstorm and other wet process weather prediction. Different from the Euclidean space metric of the classical variational method, the new metric relationship in the quantum state space can better reveal the nature that the observed brightness temperature is closer to the simulated brightness temperature obtained by the variational assimilation.

[0067] The application adopts increment four-dimensional non-Gaussian quantum data assimilation based on quantum annealing driving. The classical variational assimilation method usually adopts gradient descent method and related improved methods to solve the cost function by minimization iteration, so as to generate the optimal analysis field. The whole minimization iteration is time-consuming in the solving process, because the iteration calculation complexity required for sufficiently reducing the cost function is high, which leads to the fact that the existing data variational assimilation method needs a large amount of computing resources in the numerical weather prediction system. The application adopts the solution of the quantum non-Gaussian increment four-dimensional variational assimilation model driven by quantum annealing, which accelerates the solution of the variational assimilation cost function and reduces the consumption of computing resources. BRIEF DESCRIPTION OF DRAWINGS

[0068] Figure 1 The flowchart of the method of the application for assimilating high-spectral error non-Gaussian distribution data of meteorological satellite based on quantum computing;

[0069] Figure 2 The principle block diagram of the method of the application for assimilating high-spectral error non-Gaussian distribution data of meteorological satellite based on quantum computing. DETAILED DESCRIPTION

[0070] The present application will be further clarified by the following examples, which should be considered as merely illustrative of the present application and not limiting the scope of the application, and by reading the present application. Various modifications of the present application, in terms of the examples, will be apparent to those skilled in the art from the disclosure herein, which modifications are to be within the scope of the claims.

[0071] Embodiment one:

[0072] A method for assimilating meteorological satellite hyperspectral error non-Gaussian distribution data based on quantum computing, such as Figure 1 and Figure 2 The method comprises:

[0073] S1, interpolating the background field data matched in space and time with the meteorological satellite hyperspectral channel observation data to the field of view points of the meteorological satellite hyperspectral channel data;

[0074] Further, the meteorological satellite hyperspectral channel data in this example can be FY-4A / GIIRS data, of course, it can also be selected from Fengyun-4B hyperspectral, other types of polar orbit and geostationary satellite hyperspectral infrared detector data. FY-4A / GIIRS data is obtained from the official website of the National Satellite Meteorological Center (http: / / satellite.nsmc.org.cn) by registration. The background field data of the present application is obtained from the Final Global Data Assimilation System (FNL) of the National Centers for Environmental Prediction (NCEP), and the data is obtained by registration from the relevant official website.

[0075] The above-mentioned FY-4A / GIIRS data refers to the interferometric atmospheric vertical sounder (GIIRS) detection data carried by Fengyun-4A (FY-4A).

[0076] The interpolation in this example can be based on the FY-4A / GIIRS field of view point latitude and longitude information, and the NCEP / FNL data is interpolated to the FY-4A / GIIRS field of view point by using the bilinear interpolation method.

[0077] S2, radiative transfer model simulation is performed on the temperature profile, humidity profile and other model space data after the background field data interpolation processing, and the corresponding satellite hyperspectral channel observation brightness temperature equivalent data is obtained as the channel simulated brightness temperature;

[0078] S3, based on the channel simulation brightness temperature, considering the outlier of the hyperspectral observation brightness temperature data, the generalized quality control is formed, and the generalized quality control includes the channel optimal selection considering the quantum information entropy, the bias correction based on the artificial intelligence generalized integrated generative deep learning, and the cloud detection removing the cloud view point data based on the minimum residual method;

[0079] Further, unlike the classical variational assimilation which removes the "outliers" in the initial quality control process, the generalized quality control process of the present application retains the "outliers" for use in the subsequent data assimilation process, but reduces the contribution of the "outliers" to the cost function.

[0080] Satellite hyperspectral data includes correct data, error data and outlier data. Error data is directly removed. However, more outlier data is some "extreme" but correct observation data including important weather processes or weather events, and assimilating such data can improve the quality of the assimilation analysis field, and further improve the accuracy of numerical weather prediction (e.g. precipitation intensity and falling area, typhoon path).

[0081] S4, a quantum non-Gaussian increment four-dimensional variational assimilation model compatible with Gaussian distribution and non-Gaussian distribution of errors is constructed, and a dynamically adjustable weight contribution rate factor is introduced in the observation term of the cost function of the quantum non-Gaussian increment four-dimensional variational assimilation model;

[0082] Further, the method of the present application solves the problem that the periodic fluctuation of the hyperspectral infrared water vapor channel brightness temperature bias causes the cost function minimization iteration of the classical variational assimilation method to be difficult to converge or fail, and improves the accuracy of typhoon and rainstorm weather prediction.

[0083] S5, based on the quality control data, the quantum non-Gaussian increment four-dimensional variational assimilation model is solved to obtain the analysis field data.

[0084] Embodiment two:

[0085] The embodiment further designs on the basis of embodiment one, that is, satellite channel optimal selection as a key technology for hyperspectral atmospheric vertical probe variational assimilation or inversion can reduce the ill-posedness of variational assimilation or inversion caused by redundant information of hyperspectral probe observation, and further solve the channel "dimension disaster" problem. The present application proposes a "three-step" quantum information entropy satellite hyperspectral channel optimal selection method. The core method is to accurately express the objectively existing law according to the superposition, coherence and entanglement of quantum state in quantum theory. The specific method of the channel optimal selection considering the quantum information entropy in this example includes:

[0086] 1) Remove the satellite channel whose simulated brightness temperature error absolute value exceeds 10K; the simulated brightness temperature error is defined as the difference between the channel observed brightness temperature and the channel simulated brightness temperature.

[0087] 2) Based on the peak value of the satellite channel weighting function layer, that is, the different height atmospheric information detected by different satellite channels, the satellite channels are preliminarily selected to ensure that the selected channels can detect atmospheric information of different atmospheric height layers, and each atmospheric height layer has a channel selected, and a preliminary selected channel combination is obtained;

[0088] The weighting function (Weighting Function, abbreviated as WF) represents the contribution degree of different height atmospheres to the radiation of different channels of the satellite, and the peak position reflects the effective detection layer of the channel. The WF is defined as:

[0089]

[0090] Wherein, τ represents the channel transmittance; v represents the channel center frequency; θ represents the satellite zenith angle; p represents the air pressure; represents the partial derivative.

[0091] The channel preliminary selection based on the channel weighting function is as follows:

[0092] 2.1) Eliminate multi-peak channels. Single-peak channels are retained to ensure that the information detected by the channel is "single".

[0093] 2.2) Layered screening of channels. In the same atmospheric height layer or the same air pressure layer of the radiation transmission mode (for example, the 850 hPa air pressure layer), the channel with the largest weighting function peak layer and the steepest weighting function is preferentially selected to avoid the uncertainty of the channel detection information. The meteorological information of adjacent height layers is not introduced.

[0094] 2.3) Considering the ground emissivity and other problems, the channel with the peak value of the weighting function located on the ground is eliminated to eliminate the influence of the ground signal on the variational assimilation result.

[0095] 3) Optimal channel selection considering quantum information entropy for the preliminary selected channel combination.

[0096] Embodiment three:

[0097] The embodiment is further designed on the basis of the embodiment one, that is, because the observation data or the signal transmission process may be disturbed, the satellite observation data may contain noise, and therefore, the brightness temperature reconstruction is required when the related data is used to achieve the denoising purpose. Different from the "linear transformation" principal component analysis in the prior art, the present application proposes a "nonlinear" sampling compressive sensing (Compressive Sensing, abbreviated as CS) method for reconstructing the high-spectral infrared channel observation brightness temperature. In this example, the channel observation brightness temperature is obtained by performing nonlinear brightness temperature reconstruction based on compressive sensing on the meteorological satellite high-spectral infrared channel observation brightness temperature data.

[0098] The brightness temperature data from the hyperspectral infrared channel observations of meteorological satellites are represented sparsely in the wavelet domain. The original brightness temperature signal is then reconstructed through decoding. The specific expression is as follows:

[0099]

[0100] Where, x true The reconstructed original brightness temperature signal; y signal To observe the compressed brightness temperature signal, i.e., the received brightness temperature signal; H CS For structure operators; R signal For brightness temperature signal error; W CS Denotes the transformation matrix; λ CS This represents the regularization parameter, which is set to 0.1 in this invention.

[0101] Furthermore, the main idea of ​​nonlinear sampling compressed sensing is to compress and sample sparse signals, and after transmission, select an appropriate reconstruction method to restore the compressed signal to the original signal. This allows the sparse state vector to be reconstructed through a small number of random measurements. In practice, by introducing sparse representations in the wavelet domain, better reconstruction results can be obtained with fewer random samples.

[0102] Example 4:

[0103] This embodiment, based on Embodiment 1, is further designed in that: based on the theory of classical information entropy, that is, from the perspective of information theory, the channel selection principle should be that, given the number of selected channels, the variational assimilation of the selected channel subset (also known as the channel combination) yields a better analytical field. Channel selection stops when the magnitude of entropy change no longer increases significantly or when the number of channels in the selected channel combination reaches the initially set value (e.g., selecting 100 channels).

[0104] Due to the superposition, coherence, and entanglement of quantum states, they can accurately express objective laws, leading to significant differences between quantum theory and traditional information expression methods. "Quantum" possesses superior performance.

[0105] In this example, the specific method for selecting the optimal channel based on quantum information entropy, considering the initially selected channel combinations, includes:

[0106] 1) Normalize the observed brightness temperature sequence data for each channel of the hyperspectral infrared detector in the initially selected channel combinations; for the observed brightness temperature sequence data X of the i-th channel in the channel combination... i,entropy ={x i,entropy Let Y(j), j=1,2,...,N}, and let Y be the normalized observed brightness temperature sequence data of the i-th channel. i,entropy ={y i,entropy (j), j=1,2,...,N};where, xi,entropy (j) is the jth observation brightness temperature sequence data of the ith channel; j is the serial number mark of the observation brightness temperature sequence data, and N is the total amount of data of the observation brightness temperature sequence data; y i,entropy (j) is the jth observation brightness temperature sequence data of the ith channel after normalization;

[0107] 2) Quantum space reconstruction is respectively performed on the sequence data of each channel after normalization, so as to obtain a quantum space matrix corresponding to each channel; for the observation brightness temperature sequence data Y i,entropy of the ith channel after normalization, the corresponding quantum space matrix is expressed as follows:

[0108]

[0109] In the formula, q is the qth state vector of the lth reconstructed component in the quantum space matrix, l = 1, 2,..., K, K is the total number of reconstructed components, m is the embedding dimension; is a time window, and 6 hours are selected in the application, and 6 hours are also the time window of four-dimensional variational assimilation.

[0110] 3) Quantumization processing is performed on each reconstructed component in the quantum space matrix of each channel, and quantum information entropy of each channel is calculated; for each reconstructed component in the quantum space matrix , quantumization processing is performed as follows:

[0111]

[0112] In the formula, q b > represents the qth state vector of the reconstructed component ; ω l,k represents the probability amplitude value of the state vector q b >, k = 1, 2,..., n, n is the number of state vectors, n = 2 m , and m is the quantum bit.

[0113] The probability P i,k of the appearance of the state vector is taken as the probability of the occurrence of an event, so as to obtain the quantum information entropy of the quantum space matrix . k = 1, 2,..., n.

[0114] 4) Based on the quantum information entropy of each channel and the variational assimilation background error covariance matrix, quantum mutual information of each channel is calculated, a multi-dimensional feature space composed of multiple groups of channels is formed, and a satellite channel corresponding to the maximum quantum mutual information is searched in the multi-dimensional feature space and taken as an optimal channel.

[0115] Further, the quantum information entropy of the background error covariance matrix B after the data transformation and variational assimilation is S(σ B ) = -tr(σ B logσ B ), that is, the trace of σ B logσ B . σ B represents the density matrix of |B>. Assuming that an arbitrary channel brightness temperature sequence data (for example, T ) is denoted as c, and is converted into a quantum state denoted as |c>. The quantum information entropy of |c> is obtained as S(σ c ) = -tr(σ c logσ c ). σ c represents the density matrix of |c>. Further, the density matrix ρ cB of a composite quantum composed of |c> and |B> is obtained. In the quantum space, the quantum mutual information of the channel observed brightness temperature sequence data |c> of the hyperspectral infrared detector and the variational assimilation background error covariance matrix |B> is:

[0116]

[0117] The channel optimal selection problem of the hyperspectral infrared detector based on the quantum information entropy is converted into a problem of finding a best or optimal feature subspace. In the original m-dimensional feature space composed of m channels, a process of searching for an optimal d-dimensional feature set |c> * , that is,

[0118] S(|c> * :|B) = max i S(|c>:|B).

[0119] Embodiment Five

[0120] The embodiment is further designed on the basis of the embodiment one, and the channel brightness temperature deviation of the satellite hyperspectral infrared detector contains various systematic errors. The effect of the deviation correction determines the quality of the analysis field after the assimilation of the satellite data, and influences the prediction accuracy of the final numerical weather prediction model, such as the precipitation area and rain intensity, and the typhoon path. The satellite observation data itself contains errors due to the influence of the instrument sensitivity, calibration, and the like, so the deviation correction must be performed before the assimilation of the satellite hyperspectral infrared detector channel brightness temperature, and the systematic errors mentioned above are corrected or reduced.

[0121] The international classic "off-line" method of air mass bias correction (Bias Correction, abbreviated as BC) is based on a set of prediction factors for correction. If the brightness temperature deviation of each satellite channel j is B j , the following relationship is obtained:

[0122]

[0123] wherein the coefficient A ji and C j are obtained by least square fitting method from a large number of samples.

[0124] The present application proposes a bias correction method of artificial intelligence generalized ensemble generated deep learning, which is different from the least square fitting method used in the "offline" bias correction. The present application rewrites the above formula. Assuming that the brightness temperature bias of a certain channel is y BC , and the corresponding prediction factor is x BC , they can be represented as the following relationship:

[0125] y BC = f(x BC ) + v

[0126] wherein f represents forward mapping; f is a linear combination in the "offline" bias correction method; v represents a constant. y BC represents the dependent variable; x BC represents the independent variable.

[0127] The specific method of bias correction based on artificial intelligence generalized ensemble generated deep learning in this example includes:

[0128] 1) Selecting the prediction factor for bias correction, the prediction factor includes the model background field 1000-300 hPa thickness, 200-50 hPa thickness, model surface temperature, total water vapor, longitude information and latitude information of the field of view point, and can also include the surface wind field, topographic features, tropopause height, cloud liquid water content and temperature decrement rate square term as the "prediction factor".

[0129] 2) For all field of view points of each channel of the meteorological satellite high-spectral infrared detector to be corrected, a deep model generated based on the combination of convolutional neural network and deep neural network model is constructed as the basic model of ensemble learning, and generalized weighted ensemble machine learning is performed to find the optimal or suboptimal ensemble weight of the prediction result of the integrated basic model, and the objective minimization function of the weighted ensemble machine learning is:

[0130]

[0131] wherein n is the total number of samples; is the actual value of the value to be corrected i. is the prediction value of the basic model j to the value to be corrected i, and the value of the prediction value is obtained based on the selected prediction factor as the input of the basic model. The method of predicting the value to be corrected by the basic model through the selected prediction factor is the prior art, which will not be described here; Integrated weight corresponding to the base model j.

[0132] Embodiment six

[0133] The embodiment is further designed on the basis of embodiment one, and the specific method of cloud detection and elimination of all channel brightness temperature observation values of the cloud field point based on the minimum residual method in the embodiment includes the following steps.

[0134] 1) According to the radiation transfer theory, the simulated radiation value of the meteorological satellite high-spectral infrared detector channel is mathematically modeled, and the simulated radiation value of the satellite channel g is expressed as: Wherein, N e represents the effective cloud amount of the field point; represents the clear sky radiation; represents the cloud top pressure p c,g The channel radiation value can be directly converted into a brightness temperature value by using the Planck function, and this method is prior art and will not be described here.

[0135] 2) Calculate the deviation of the observed radiation value and the simulated radiation value of the high-spectral infrared detector channel, and the deviation δ g of the observed radiation value and the simulated radiation value of the channel g is expressed as:

[0136] 3) Based on the n-channel combination obtained by the optimal channel selection of quantum information entropy, the minimum residual method objective function is solved to obtain the effective cloud amount N e and the effective cloud top pressure p c of the field point, and the effective cloud top pressure p c is obtained by comprehensively combining the cloud top pressure of the n-channel combination, w g is the comprehensive proportion. The minimum residual method objective function is defined as follows:

[0137]

[0138] 4) Based on the effective cloud amount N e and the set threshold value (the set threshold value is generally 0.1), the cloud detection of the field point is performed, and for the field point with the effective cloud amount N e greater than the set threshold value, all brightness temperature data of the field point on all channel combinations are eliminated. The present application only assimilates the channel combination brightness temperature data of the retained field point.

[0139] Further, the minimum solution of the effective cloud amount N e and the effective cloud top pressure p c is performed in two steps, and the specific steps are as follows: ​

[0140] First step: according to the mode pressure layer of the radiation transmission mode adopted, the effective cloud top pressure is given and substituted into the minimum residual method objective function to obtain the effective cloud amount N after minimization calculation e :

[0141]

[0142] Second step: the effective cloud amount N e is substituted into the minimum residual method objective function to obtain the effective cloud top pressure p c after minimization calculation, and p c is the cloud top pressure synthesized by n channel combinations, that is, the effective cloud top pressure value of a certain field of view point set finally.

[0143] Example Seven

[0144] This embodiment is further designed on the basis of example one, that is, the basic assumption in the classical four-dimensional variational method is to require the distribution of observation errors and control variables to obey Gaussian distribution. The Bayesian theorem can be used to derive the cost function of the variable with Gaussian distribution. The total cost function of the classical four-dimensional variational assimilation method is defined as follows:

[0145]

[0146] The cost functions of the background term and the observation term are defined as follows:

[0147]

[0148] In the formula, "background" represents the cost function of the background term; "observation" represents the cost function of the observation term; matrix B represents the error covariance matrix of the background (including different meteorological variables such as temperature and humidity). R i represents the observation error covariance matrix (the observation error here is the channel observation brightness temperature error of the satellite high-spectral infrared detector). Superscripts -1 and T represent matrix inversion and transposition mark respectively. x0 is the control variable, and the solved x0 is the analysis field; x b,0 represents the background initial condition, including different meteorological variables such as temperature, humidity, wind field and surface air temperature; y i represents the i-th observation value, i = 1, 2, …, N, and N is the total number of observation values; H i is the observation operator for mapping the model integral M i,0 (·) from the model space to the observation space; M i,0 (x0) is the nonlinear numerical prediction model for integrating from the initial condition and the control variable x0 at the initial time t = 0 to t = i, and t is the time mark;

[0149] Further, the classical four-dimensional variational assimilation total cost function is changed to a meteorological element increment form, defined as follows:

[0150]

[0151] where δx0is the analysis increment, x0is the control variable; is the state vector predicted by the numerical prediction model; the superscript f indicates the prediction; B is the background error covariance matrix; R i is the observation error covariance matrix; d i is the brightness temperature increment, y i denotes the i-th observation; denotes the brightness temperature value of the i-th observation based on the model space simulated by the fast radiative transfer model, i.e., the equivalent data of the i-th observation, H i is the observation operator that maps the model integral M i,0 (·) from the model space to the observation space; M i,0 (x0) is the non-linear numerical prediction model that integrates from the initial condition and the control variable x0at the initial time t = 0 to t = i, t being the time index; the matrix H i and M i,0 denote the linear representation of the tangent model of H i and M i,0 respectively.

[0152] The cost function J is a quadratic cost function in the increment four-dimensional space δx0space.

[0153] By Taylor expansion and approximation, the calculation of δx0simplifies the following problem:

[0154] Aδx0= q

[0155] where, denotes the Hessian matrix in the increment four-dimensional variational assimilation cost function.

[0156] In practical atmospheric science numerical weather prediction problems, the dimension of the A matrix is usually more than 10 8 ~ 10 9 , making it difficult or impossible to calculate the inverse of the matrix. However, using an iterative optimization process can avoid direct calculation of the inverse of the matrix. The conjugate gradient descent algorithm is an optimization algorithm suitable for quadratic cost functions. The classical increment four-dimensional variational method reduces the cost function iteration cost by using a gradient-based quasi-Newton method to update the analysis increment δx0.

[0157] The quantum non-Gaussian increment four-dimensional variational assimilation can simultaneously consider the characteristics of the deviation Gaussian and non-Gaussian, and the quantum calculation is adopted, the speed of solving the cost function is accelerated, and the calculation time is reduced. The cost function of the quantum non-Gaussian increment four-dimensional variational assimilation model in the example is as follows:

[0158]

[0159] w(r i )=(1 / r i )·(dρ(r i ) / dr i )

[0160] In the formula, δx0 is an analysis increment, x0 is a control variable; is a state vector predicted by a numerical prediction model; the superscript f indicates prediction; B is a background error covariance matrix; R i is an observation error covariance matrix; the superscript T is a matrix transpose; the superscript -1 is a matrix inverse; d i is a brightness temperature increment, y i indicates the ith observation value, i=1, 2, …, N, and N is the total number of observation values; indicates the brightness temperature value of the ith observation value simulated based on the model space through the fast radiative transfer model, that is, the equivalent data of the ith observation value, also called the channel simulated brightness temperature; H i is an observation operator for mapping the model integral M i,0 (·) from the model space to the observation space; for satellite data, H i is also called a fast radiative transfer model. H i achieves the projection mapping or conversion of the model space variables (such as different meteorological variables such as air temperature) to the observation space (such as the channel observed brightness temperature of the satellite high-spectral infrared detector); M i,0 (x0) is a nonlinear numerical prediction model for integrating the control variable x0 from the initial condition and the initial time t=0 to t=i, t is a time mark; the matrices H i and M i,0 respectively indicate the linear representation of the tangent model of H i and M i,0 ; r i is a brightness temperature increment amplitude value, r i =d i / σ i , σ i is the observation error of the channel i; w(r i ) is a weight contribution rate factor; and ρ(r i ) is an M-estimate cost function.

[0161] Further, the core of the M-estimation new norm is the construction of the cost function. The construction of the new M-estimation cost function needs to meet the following 9 basic properties. At the same time, the weight function of the M-estimation cost function needs to meet the following 5 basic properties, see Table 1. Such properties serve as the basis for constructing the new norm of the M-estimation method.

[0162] Table 1 Basic properties of M-estimation cost function and weight function

[0163]

[0164]

[0165] According to the basic properties of Table 1, by mathematical equations and integrals, etc., the process of constructing the new norm of M-estimation by the application is as follows:

[0166]

[0167] Wherein, the superscript'indicates the derivative mark.

[0168] Integrate both sides of the above formula respectively, and according to its basic properties (see Table 1), then:

[0169]

[0170] It should be noted that in the construction of the new norm of the M-estimation method, the "adjustment scale" needs to be statistically and analyzed according to the actual situation.

[0171] In order to facilitate the calculation by using the quantum model, the cost function of the M-estimation quantum non-Gaussian increment four-dimensional variational assimilation method is approximated. The specific expression form is as follows:

[0172]

[0173] Wherein,

[0174]

[0175] Wherein, The tilde represents the linear approximation value. After derivation and analysis, the optimization only operates on the analysis increment, as follows:

[0176]

[0177] Example eight:

[0178] The embodiment is further designed on the basis of embodiment seven, that is, compared with the classical incremental four-dimensional variational assimilation method which needs to calculate the cost function and its gradient, the quantum annealing based on quantum calculation only needs to calculate the cost function. However, the quantum annealing cost function is represented by binary variables (i.e. 0 or 1). The basic principle of the quantum annealing algorithm is to search the solution space through the quantum superposition and quantum entanglement of quantum bits. In quantum calculation, a quantum bit can be in a superposition state of multiple states, and multiple possible solutions can be processed at the same time. Quantum entanglement is the interaction between quantum bits, which can make the optimization of searching the solution space more efficient. Quantum annealing algorithm uses quantum tunneling effect to make quantum have the ability to penetrate the potential barrier higher than its own energy, so that the algorithm can escape from the local extremum and approach the global optimum with a higher probability.

[0179] The specific method for solving the quantum non-Gaussian incremental four-dimensional variational assimilation model in this example includes:

[0180] 1) The incremental four-dimensional non-Gaussian quantum data assimilation method driven by quantum annealing is adopted, real number representation is performed through Q quantum bits, and the analysis increment δx0 is quantized through the mapping matrix G:

[0181]

[0182] g T =[-2 Q-1 ,2 Q-2 ,2 Q-3 ,...,2 1 ,2 0 ]

[0183]

[0184] In the formula, β is a adjustable scaling parameter, and the value is 0.2; b is a binary vector whose elements are 0 or 1, and 0 is a vector whose elements are all 0; g is the quantization of real number; Q is a natural number;

[0185] 2) Based on the quantized analysis increment, the cost function of the quantum non-Gaussian incremental four-dimensional variational assimilation model is solved in quantum space:

[0186]

[0187] In the formula,

[0188]

[0189] In the formula, Ham represents Hamiltonian; w(·) represents the weight function of the constructed M-estimation new norm; A and u are as follows:

[0190] 3) Further variable transformation or conversion After that, the cost function is minimized by a quantum annealing method to obtain the optimal analysis field.

[0191] Quantum Annealing (QA) is an optimization algorithm based on the principles of quantum mechanics, aiming to solve complex optimization problems by simulating the annealing process of quantum systems. Its core idea is to use quantum tunneling effects and quantum superposition states to make the system quickly cross high energy barriers, so as to find the global optimal solution of the problem.

[0192] Basic principle: Quantum annealing algorithm simulates the annealing process in solid physics, by gradually changing the Hamiltonian of the system, slowly evolving the initial state to the target state, so as to find the optimal solution of the problem. Unlike classical simulated annealing algorithm, quantum annealing uses quantum tunneling effect, allowing the system to directly cross high energy barriers, avoiding the local optimal solution that may be encountered in traditional annealing methods.

[0193] Further, unlike existing variational assimilation methods, which reduce the cost function based on gradient optimization to obtain the best analysis field. Due to the large number of iterations required to sufficiently reduce the cost function in actual numerical weather prediction, data variational assimilation methods require a large amount of computing resources in numerical weather prediction model systems, and are also "time-consuming". In view of this, due to quantum effects (such as tunneling, superposition and entanglement), the solution based on quantum annealing driven by the present application can accelerate the solution of the variational assimilation cost function.

[0194] Embodiment nine:

[0195] This embodiment is further designed on the basis of embodiment one, that is, in this example, it also includes evaluating the accuracy of the analysis field data, the solution timeliness and the accuracy of the analysis field in improving the numerical weather prediction of the wet process.

[0196] Embodiment ten:

[0197] The system for assimilating meteorological satellite hyperspectral error non-Gaussian distribution data based on quantum computing according to the present application comprises a first data processing module, a second data processing module, a third data processing module and a fourth data processing module.

[0198] The first data processing module is used for radiative transfer model simulation on the meteorological satellite hyperspectral channel data after background field data interpolation processing, to obtain corresponding satellite hyperspectral channel observation brightness temperature equivalent data as channel simulated brightness temperature.

[0199] The second data processing module is configured to perform generalized quality control of outliers on the channel analog brightness temperature by taking into account the meteorological satellite hyperspectral data, to form quality control data, and the generalized quality control includes channel optimal selection taking into account quantum information entropy, bias correction based on artificial intelligence generalized ensemble generative deep learning, and cloud detection elimination based on the minimum residual method.

[0200] The third data processing module is configured to construct a quantum non-Gaussian increment four-dimensional variational assimilation model compatible with Gaussian distribution and non-Gaussian distribution of errors, and a weight contribution rate factor that can be dynamically adjusted is introduced in an observation term in a cost function of the quantum non-Gaussian increment four-dimensional variational assimilation model.

[0201] The fourth data processing module is configured to solve the quantum non-Gaussian increment four-dimensional variational assimilation model based on the quality control data, to obtain analysis field data.

[0202] Embodiment eleven:

[0203] An electronic device includes a memory and a processor, the memory stores a computer program, and the processor is configured to call and run the computer program stored in the memory to execute the method of any one of the above embodiments.

[0204] A computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of any one of the above embodiments.

[0205] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for assimilating non-Gaussian distribution data of high spectral errors of meteorological satellites based on quantum computing, characterized in that, The method includes: Radiative transfer model simulation was performed on the meteorological satellite hyperspectral channel data after background field data interpolation to obtain the corresponding satellite hyperspectral channel observation brightness temperature equivalent data as channel simulation brightness temperature; The simulated brightness temperature of the channel is subjected to generalized quality control considering outliers from meteorological satellite hyperspectral data to form quality control data. The generalized quality control includes optimal channel selection considering quantum information entropy, bias correction based on artificial intelligence generalized ensemble generative deep learning, and cloud detection and removal based on the minimum residue method. A quantum non-Gaussian incremental four-dimensional variational assimilation model compatible with both Gaussian and non-Gaussian error distributions is constructed. The observation term in the cost function of the quantum non-Gaussian incremental four-dimensional variational assimilation model is introduced with a dynamically adjustable weight contribution rate factor. Based on the quality control data, the quantum non-Gaussian incremental four-dimensional variational assimilation model is solved to obtain the analysis field data.

2. The method for assimilating non-Gaussian distribution data of high spectral error of meteorological satellite based on quantum computing according to claim 1, characterized in that, The specific methods for optimal channel selection considering quantum information entropy include: 1) Eliminate satellite channels with a simulated brightness temperature error exceeding 10K in the radiative transfer mode channel; the simulated brightness temperature error is defined as the difference between the observed brightness temperature of the channel and the simulated brightness temperature of the channel. 2) Based on the peak layer of the satellite channel weighting function, that is, the atmospheric information at different altitudes detected by different satellite channels, the satellite channels are initially selected to obtain the preliminary channel combinations; 3) The optimal channel selection is considered based on quantum information entropy for the initially selected channel combinations.

3. The method for assimilating non-Gaussian distributed high spectral error meteorological satellite data based on quantum computing according to claim 2, characterized in that, The observed brightness temperature of the channel is obtained by nonlinear brightness temperature reconstruction based on compressed sensing from the hyperspectral infrared channel observation data of meteorological satellites. The specific method is as follows: The brightness temperature data from the hyperspectral infrared channel observations of meteorological satellites are represented sparsely in the wavelet domain. The original brightness temperature signal is then reconstructed through decoding. The specific expression is as follows: where x true is the reconstructed original brightness temperature signal; y signal is the observed compressed brightness temperature signal, i.e. the received brightness temperature signal; H CS is the structure operator; R signal is the brightness temperature signal error; W CS denotes the transformation matrix; λ CS denotes the regularization parameter.

4. The method for assimilating non-Gaussian distributed high spectral error meteorological satellite data based on quantum computing according to claim 2, characterized in that, The specific method for selecting the optimal channel by considering quantum information entropy in the initially selected channel combinations includes: 1) normalizing the observed brightness temperature sequence data corresponding to each channel of the hyperspectral infrared detector in the preliminary screened channel combination; for the observed brightness temperature sequence data X = {x(j), j = 1, 2,..., N} of the i th channel in the channel combination, the corresponding normalized observed brightness temperature sequence data of the i th channel is denoted as Y = {y(j), j = 1, 2,..., N}; wherein x(j) is the j th observed brightness temperature sequence data of the i th channel; j is the serial number mark of the observed brightness temperature sequence data, and N is the total amount of data of the observed brightness temperature sequence data; y(j) is the j th observed brightness temperature sequence data of the i th channel after normalization; i,entropy i,entropy i,entropy i,entropy i,entropy i,entropy ​​​​​​ 2) quantum space reconstruction is performed on each channel of the normalized sequence data respectively to obtain a quantum space matrix corresponding to each channel; for the i-th channel of the normalized observation brightness temperature sequence data Y i,entropy , the expression of the corresponding quantum space matrix is as follows: In the formula, is the lth reconstruction component in the quantum space matrix, l = 1, 2, …, K, K is the total number of reconstruction components, m is the embedding dimension; is the time window; 3) quantumizing each reconstructed component in the quantum space matrix of each channel, and calculating the quantum information entropy of each channel; for each reconstructed component in the quantum space matrix quantumizing is as follows: where |q b > denotes the reconstructed component of the qth state vector; ω l,k denotes the probability amplitude of the state vector |q b >, k = 1, 2,..., n, n is the number of state vectors, n = 2 m , m is the number of constituent qubits; The probability P of the state vector appearing i,k The quantum information entropy of the quantum space matrix wherein, ​ 4) Based on the quantum information entropy and variational assimilation background error covariance matrix of each channel, calculate the quantum mutual information of each channel to form a multidimensional feature space composed of multiple channels. Search for the satellite channel with the maximum quantum mutual information in the multidimensional feature space and select it as the optimal channel.

5. The method for assimilating non-Gaussian distributed high spectral error meteorological satellite data based on quantum computing according to claim 2, characterized in that, The specific methods for bias correction based on generalized ensemble generative deep learning in artificial intelligence include: 1) Select the forecast factors for bias correction; 2) For the equivalent brightness temperature data of all field-of-view points of each channel of the meteorological satellite hyperspectral infrared detector that needs correction, a deep model based on a combination of convolutional neural networks and deep neural networks is constructed as the base model for ensemble learning. Generalized weighted ensemble machine learning is then performed to find the optimal or suboptimal ensemble weights that integrate the prediction results of the base model. The objective minimization function of the weighted ensemble machine learning is: wherein n is the total number of samples; is the actual value of the to-be-corrected value i; is the predicted value of the to-be-corrected value i by the base model j, and the predicted value is obtained based on the selected prediction factor as an input quantity of the base model; is the integrated weight corresponding to the base model j.

6. The method for assimilating non-Gaussian distributed high spectral error meteorological satellite data based on quantum computing according to claim 3, characterized in that, The specific method for cloud detection and removal based on the minimum residue method includes: 1) According to the radiation transfer theory, the mathematical modeling of the simulated radiation value of the meteorological satellite hyperspectral infrared detector channel is carried out, and the simulated radiation value of the satellite channel g is represented as: wherein N e represents the effective cloud cover of the field point; represents the clear sky radiation; represents the cloudy radiation simulation when the cloud top pressure is p c,g ; 2) Calculate the bias of the observed and simulated radiance values for the hyperspectral infrared detector channel, the observed radiance value for channel g and the simulated radiance value δ g is expressed as: 3) Based on the n channel combination obtained by the optimal channel selection of quantum information entropy, the minimum residual method objective function is solved to obtain the effective cloud cover N of the field point e and the effective cloud top pressure p c The minimum residual method objective function is defined as follows: 4) Based on the effective cloud cover N e Cloud detection at the field of view point with a set threshold, for effective cloud cover N e For field of view points with effective cloud cover N greater than the set threshold, all brightness temperature data from all channel combinations at that field of view point are rejected.

7. The method for assimilating non-Gaussian distributed high spectral error meteorological satellite data based on quantum computing according to claim 1, characterized in that, The cost function of the quantum non-Gaussian incremental four-dimensional variational assimilation model is: w(r i ) = (1 / r i ) · (dρ(r i ) / dr i ) where δx0is the analysis increment, x0is the control variable, and the solved x0is the analysis field; is the state vector predicted by the numerical prediction model; the superscript f denotes prediction; B is the background error covariance matrix; R i is the observation error covariance matrix; the superscript T denotes matrix transpose; the superscript -1 denotes matrix inverse; d i is the brightness temperature increment, y i denotes the ith observation, i = 1, 2, …, N, and N is the total number of observations; denotes the brightness temperature value simulated by the fast radiative transfer model based on the model space for the ith observation, that is, the equivalent data of the ith observation, H i is the observation operator that maps the model integral M i,0 (·) solution from the model space to the observation space; M i,0 (x0) is a nonlinear numerical forecast model that integrates the control variables x0 from initial conditions and initial time t = 0 to t = i, t being a time index; the matrix H i and M i,0 are linear representations of the tangent model of H i and M i,0 , respectively; r i is the brightness temperature increment amplitude value, r i = d i / σ i , σ i is the observation error of channel i; w(r i ) is the weight contribution rate factor; p(r i ) is the M-estimate cost function.

8. The method for assimilating non-Gaussian distributed high spectral error meteorological satellite data based on quantum computing according to claim 7, characterized in that, The specific method for solving the quantum non-Gaussian incremental four-dimensional variational assimilation model includes: 1) Adopting quantum annealing driven incremental four-dimensional non-Gaussian quantum data assimilation method, real number representation is carried out through Q quantum bits, and the analysis increment δx0 is quantized through the mapping matrix G: g T ＝[-2 Q-1 ,2 Q-2 ,2 Q-3 ,...,2 1 ,2 0 ] In the formula, β is an adjustable scaling parameter; b is a binary vector whose elements are 0 or 1, and 0 is a vector whose elements are all 0; g is the quantization of a real number; Q is a natural number; 2) Based on the quantized analysis increment, the cost function of the quantum non-Gaussian incremental four-dimensional variational assimilation model is solved in the quantum space: In the formulae, In the formula, Ham represents the Hamiltonian; w(·) represents the weight function of the constructed M-estimation new norm; A and u are as follows: 3) Minimize by quantum annealing method Get the best analysis field.

9. The method for assimilating non-Gaussian distributed meteorological satellite hyperspectral errors based on quantum computing according to claim 1, characterized in that, It also includes the evaluation of the accuracy, timeliness and analysis field on the improvement of the accuracy of the numerical weather prediction of the wet process.

10. A system for assimilating non-Gaussian distributed data of meteorological satellite hyperspectral errors based on quantum computing, characterized in that, The system comprises a first data processing module, a second data processing module, a third data processing module and a fourth data processing module. The first data processing module is configured to perform radiative transfer model simulation on the meteorological satellite hyperspectral channel data after background field data interpolation processing, to obtain corresponding satellite hyperspectral channel observation brightness temperature equivalent data as channel simulated brightness temperature. The second data processing module is configured to perform generalized quality control of meteorological satellite hyperspectral data considering outliers on the channel simulated brightness temperature to form quality control data, wherein the generalized quality control comprises channel optimal selection considering quantum information entropy, bias correction based on artificial intelligence generalized ensemble generative deep learning and cloud detection removal based on the minimum residual method. The third data processing module is configured to construct a quantum non-Gaussian incremental four-dimensional variational assimilation model compatible with Gaussian distribution and non-Gaussian distribution, wherein a dynamically adjustable weight contribution rate factor is introduced in the observation term of the cost function of the quantum non-Gaussian incremental four-dimensional variational assimilation model. The fourth data processing module is configured to solve the quantum non-Gaussian incremental four-dimensional variational assimilation model based on the quality control data to obtain analysis field data.

Images