Remote sensing rainfall data deviation correction method based on multivariable joint distribution and LSTM

Abstract

The invention provides a remote sensing rainfall data deviation correction method based on multivariate joint distribution and LSTM, and relates to the technical field of satellite remote sensing rainfall deviation correction. The remote sensing rainfall data deviation correction method based on multivariable joint distribution and LSTM comprises the following steps: preprocessing a CN05.1 data set and CMIP6 day-by-day rainfall data; building a multivariable joint distribution model on the basis of R-vine Copula; a VCCP method is adopted to carry out change point test on the multivariate joint distribution model; and adopting a long short-term memory recurrent neural network LSTM machine for learning, and respectively correcting remote sensing rainfall data before and after a mutation point in the multivariable joint distribution model. According to the method, the remote sensing rainfall data before and after sudden change is corrected by using the LSTM model, the accuracy of the corrected rainfall data is obviously improved, and the influence of sudden change on the rainfall time sequence can be well solved, so that the effect of correcting the remote sensing rainfall deviation data difference is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of satellite remote sensing precipitation bias correction, and in particular to a remote sensing precipitation data bias correction method based on multivariate joint distribution and LSTM. Background Art

[0002] Precipitation is a basic component of the Earth's water cycle and has important meteorological, climatological and hydrological significance. Ground-based meteorological stations are limited by factors such as coverage, spatial resolution and data delay. Remote sensing inversion of precipitation has the characteristics of high temporal and spatial resolution, wide coverage and continuous spatial estimation, which makes up for the shortcomings of conventional meteorological observation stations. However, satellite remote sensing products have strong uncertainties. Therefore, it is a challenging scientific research goal to reasonably, objectively and accurately correct remote sensing inversion of precipitation.

[0003] The existing remote sensing precipitation data bias correction methods can be roughly divided into two categories: the first category is based on statistical methods, such as proportional bias correction method, regression analysis method and spatial interpolation method; the second category is based on machine learning and data fusion methods, such as multi-scale fusion method, downscaling method, etc. In recent years, modern machine learning technology has been used to improve the accuracy of remote sensing data through complex algorithms and models, as well as the support of many relevant meteorological elements, providing a new idea for the bias correction of remote sensing precipitation.

[0004] It is worth noting that the change process of precipitation is random and discontinuous, and is jointly affected by multiple meteorological factors (such as temperature, relative humidity and wind speed); the relationship between meteorological elements is very complex, and the change of rainfall is not only driven by one or even several meteorological elements, but also forms an intricate feedback mechanism between multiple meteorological elements and precipitation. In addition, as the length of the time series increases, precipitation data may also have certain fluctuations or mutations. How to describe the driving relationship between relevant meteorological factors and remote sensing precipitation data, and how to consider the impact of "mutations" on the accuracy correction of hydrological time series are the key technical problems solved by this application. Therefore, technicians in this field provide a remote sensing precipitation data bias correction method based on multivariate joint distribution and LSTM to solve the problems raised in the above background technology. Summary of the invention

[0005] 1. Technical issues to be solved

[0006] In view of the shortcomings of the prior art, the present invention provides a remote sensing precipitation data bias correction method based on multivariate joint distribution and LSTM, which solves the problems of how to describe the driving relationship between relevant meteorological factors and remote sensing precipitation data, and how to consider the impact of "mutations" on the accuracy correction of hydrological time series.

[0007] (II) Technical solution

[0008] To achieve the above objectives, the present invention is implemented through the following technical scheme: a remote sensing precipitation data deviation correction method based on multivariate joint distribution and LSTM, including: preprocessing of CN05.1 data set and CMIP6 daily precipitation data; building a multivariate joint distribution model based on R-vine Copula; using the VCCP method to perform a change point test on the multivariate joint distribution model; and using long short-term memory recurrent neural network LSTM machine learning to respectively correct the remote sensing precipitation data before and after the mutation point in the multivariate joint distribution model.

[0009] Preferably, the preprocessing of the CN05.1 dataset and the CMIP6 daily precipitation data comprises the following steps:

[0010] S1. Select some meteorological factor datasets from the CN05.1 dataset and select precipitation data under multiple climate models from CMIP6; for the multiple climate models selected in CMIP6, use Multi-ModelEnsemble (MME) to obtain the average precipitation value at each time step;

[0011] S2. Unify the grid resolution of the selected data to 0.25°×0.25°; use the meteorological data analysis software Climate Data Operator CDO based on the Linux operating system to perform bilinear interpolation processing and analysis on the selected data, and unify its grid resolution to 0.25°×0.25°;

[0012] Among them, the calculation formula of the two-line interpolation method is as follows:

[0013]

[0014] Where (x1, y1), (x2, y2) are the coordinates of the grid points;

[0015] S3. Spatially average the CN05.1 dataset and the CMIP6 daily precipitation data.

[0016] Preferably, the multivariate joint distribution model is constructed based on R-vine Copula, and the R-vine Copula model design includes the following steps:

[0017] S1. Preferred selection of Vine tree structure. Any dimensional joint distribution function can be decomposed into a marginal distribution function and a Copula function. R-vine becomes the structural selection range of Vine Copula function due to its structural flexibility. The R-vine is defined by the following conditions:

[0018] f(x1,...,x p)=f(x1)·f(x2|x1)·...·f(x p |x1,...,x p-1 )

[0019] S1.1. If T = {T1, T2, ..., T n-1}, T1 is a node with N1={1,…,n} and an edge with E n of trees;

[0020] S1.2. For i=2,…,n-1, T i is to satisfy N1=E i-1 ;

[0021] S1.3. For i=2,…,n-1, {a,b}∈E i , a={a1,a2},b={b1,b2}#(a,b)=1,# indicates the cardinality of the set;

[0022] Therefore, an n-dimensional R-vine structure can be represented by n-1 trees T. j There are n-1+j nodes and nj edges, which means there are nj Copula density functions; the tree T j The edges that share a node in will become nodes and are represented by the tree T j+1 The other edges in are connected;

[0023] The d-dimensional R-vine is defined by the following conditions:

[0024] f(x1,...,x d )=f(x1)·f(x2|x1)·...·f(x d |x1,...,x d-1 )

[0025] d-dimensional R-vine density function:

[0026]

[0027] Where: f i Represents x i The probability density of F(x i |x j+1 ,...,x i-1 ) is its univariate conditional distribution function, c ij|(j+1)...(i-1) is a given subset (x j+1 ,...,x i-1 ), with the two-dimensional variable (x i and i )’s density of the conditional vine copula function associated with the bivariate conditional distribution of ;

[0028] S2. Selection of Copula function: The minimum value of two model selection criteria, Akaike Information Criterion AIC and Bayesian Information Criterion BIC, is selected to determine the Copula function between variables. When the two criteria select different Copula functions, the final decision is mainly made by AIC. The best Pair-Copula function is selected from basic binary functions such as Gauss Copula, Student-t Copula, Clayton Copula, Gumbel Copula, and Frank Copula, including the following calculation formula:

[0029]

[0030] in, is the estimated likelihood, m is the number of adjustable parameters in the Copula function, and n is the number of data points;

[0031] S3. Parameter estimation of Copula function, estimate the parameters in Copula function through MLE maximum likelihood estimation;

[0032] Any two random variables u 1, The joint distribution of u2 can be represented by their respective marginal distributions and the Copula function C that represents the correlation between the two. Then the density function of the joint distribution function is:

[0033] h(u1,u2,)=(F1(u1,a1),F2(u2,a2),δ)f1(u1,a1)f2(u2,a2)

[0034] The maximum likelihood function for an observed sample is:

[0035]

[0036] The maximum likelihood estimate It is a differential equation The solution;

[0037] S4. The multivariate joint distribution model between meteorological factors is constructed by R-vine Copula, and the Maximum Spanning Tree and MST algorithms are used to participate in model construction. The specific steps are as follows:

[0038] S4.1. In the processed sample data set, calculate the Kendall-τ rank correlation coefficient for any random variable combination of each meteorological factor;

[0039] S4.2. Generate the first tree Tree 1 of the R-vine structure by the MST algorithm, that is, maximize the sum of the absolute values ​​of all Kendall-τ correlation coefficients in the tree;

[0040] S4.3. For each edge on the tree, select the optimal binary Copula function for the edge using the binary Copula function selection methods AIC and BIC, and calculate the corresponding conditional distribution function;

[0041] S4.4. Under the defined conditions, find all conditionally related random variable combinations in each layer, and calculate the corresponding Kendall-τ rank correlation coefficient based on the conditional distribution function calculated in step tree S4.3;

[0042] S4.5. Based on the Kendall-τ rank correlation coefficient obtained in S4.4, repeat steps S4.2 to S4.4 until the last tree of the R-vine structure.

[0043] Preferably, the VCCP method is used to perform change point test on the multivariate joint distribution model. In the VCCP method, binary segmentation and MOSUM are selected as two segmentation methods, and Vuong and Stationary Bootstrap are selected as two methods for inferring candidate change points; the segmentation method and the change point inference method are combined to obtain four detection methods to perform change point detection on the multivariate joint distribution model.

[0044] Preferably, the method of using long short-term memory recurrent neural network (LSTM) machine learning to correct the remote sensing precipitation data before and after the mutation point in the multivariate joint distribution model comprises the following steps:

[0045] S1. Normalize the input data. Except for precipitation data, other input data are processed by Z-Score normalization. The formula is as follows:

[0046]

[0047] Where μ is the mean of all sample data, σ is the standard deviation of all sample data;

[0048] Since the precipitation distribution is discrete and has a large span, the precipitation data is log-normalized and the conversion function is:

[0049] y=log(x+1)

[0050] S2. Train the LSTM model and preliminarily determine the model parameters. The LSTM model is constructed by an input layer, a hidden layer, and an output layer. Each LSTM unit in the hidden layer contains three control state gates: an input gate, a forget gate, and an output gate. The forward calculation process formula of LSTM is as follows:

[0051] i t =σ(W xi x t +W hi h t-1 +b i ) f t =(W xf x t +W hf h t-1 +b f ) c t =f t ☉c t-1 +i t ☉tanh(W xc x t +W hc h t-1 +b c ) o t =σ(W x0 x t +W ho h t-1 +b o )

[0052] h t = t ⊙tanh(c t )

[0053] Where i, f, and o represent the input, forget, and output gates, respectively; ⊙ represents the multiplication of corresponding elements; W and b represent the weight matrix and bias vector of the network; t represents the time step; x and h represent the input and output vectors of the LSTM hidden layer; c represents the memory unit; and σ represents the Sigmoid function, which is a nonlinear activation function in machine learning.

[0054] The meteorological factor data selected from the CN05.1 dataset is used as input data, and the CN05.1 precipitation data is used as target data; the model is trained using the gradient descent method to optimize the model parameters; the model with the smallest loss function value of the validation set is selected as the optimal model; during the model training and optimization process, the model is continuously optimized through methods such as learning rate decay and early stopping to accelerate model convergence and prevent model overfitting;

[0055] S3. Model accuracy evaluation: the precipitation data in the CN05.1 dataset is used as the observed precipitation data, and the precipitation data in the CMIP6 dataset is used as the remote sensing satellite precipitation data. The correlation coefficient (CC), mean absolute error (MAE), root mean square error (RMSE), and relative error (Bias) are used to evaluate the model effect.

[0056]

[0057]

[0058] Where n represents the total amount of corrected precipitation data; G represents observed precipitation data, S represents corrected remote sensing satellite precipitation data; G i represents the i-th value in the observed precipitation data, S i represents the i-th value in the corrected remote sensing satellite precipitation data; represents the mean of the observed precipitation data, Represents the mean of the corrected remote sensing satellite precipitation data;

[0059] S4. Using the optimal LSTM model obtained after training, the remote sensing precipitation data are corrected for the years before and after the mutation point in the joint distribution model of multiple variables, and the correction effect is evaluated using CC, MAE, RMSE, and Bias.

[0060] (III) Beneficial effects

[0061] The present invention provides a remote sensing precipitation data deviation correction method based on multivariate joint distribution and LSTM.

[0062] It has the following beneficial effects:

[0063] 1. In the present invention, multiple meteorological factors related to precipitation are fully considered, a joint distribution model of multiple meteorological factors is constructed based on Vine Copula, and the VCCP method is used to detect change points in the multivariate joint distribution model; on this basis, the time series before and after the mutation point are corrected using the LSTM long and short time neural network model respectively to improve the correction accuracy of remote sensing precipitation data. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a schematic diagram of the basic process of the present invention;

[0065] Figure 2 It is a schematic diagram showing the research area in the present invention;

[0066] Figure 3 It is a schematic diagram of various meteorological factors with a grid resolution of 0.25°×0.25° in the present invention;

[0067] Figure 4 Schematic diagram of parameters and structure of the multivariate joint distribution model in the present invention;

[0068] Figure 5 This is a schematic diagram of the change point detection results of the multivariate joint distribution model using VCCP in the present invention;

[0069] Figure 6 It is a schematic diagram for comparing the results of different correction situations in the present invention;

[0070] Figure 7 It is a schematic diagram comparing the correction errors of different correction situations in different time periods in the present invention. DETAILED DESCRIPTION

[0071] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0072] Embodiment 1:

[0073] like Figure 1-7 As shown, an embodiment of the present invention provides a remote sensing precipitation data bias correction method based on multivariate joint distribution and LSTM, including: preprocessing of the CN05.1 dataset and CMIP6 daily precipitation data from 1980 to 2021; building a multivariate joint distribution model based on R-vine Copula; using the VCCP method to perform a change point test on the multivariate joint distribution model; and using long short-term memory recurrent neural network LSTM machine learning to respectively correct the remote sensing precipitation data before and after the mutation point in the multivariate joint distribution model.

[0074] The preprocessing of CN05.1 dataset and CMIP6 daily precipitation data includes the following steps:

[0075] S1. Select some meteorological factor datasets from the CN05.1 dataset, and select precipitation data under multiple climate models from CMIP6; for the multiple climate models selected in CMIP6, select 6 meteorological factors from the CN05.1 dataset, and see Table 1 for the explanation table. For the 6 climate models selected in CMIP6, use Multi-Model Ensemble (MME) to obtain the average precipitation value for each time step. The years selected for the two sets of data are daily-scale data from 1980 to 2021:

[0076] factor Interpretation unit factor Interpretation unit Pre Precipitation mm / d Tmin Daily minimum temperature ℃ Tm Average Temperature ℃ Win Average wind speed m / s Tmax Daily maximum temperature ℃ Rhu Relative humidity ％

[0077] Table 1. Explanation of meteorological factors in the selected CN05.1 dataset

[0078] S2. Unify the grid resolution of the selected data to 0.25°×0.25°; use the meteorological data analysis software Climate Data Operator CDO based on the Linux operating system to perform bilinear interpolation processing and analysis on the selected data, and unify its grid resolution to 0.25°×0.25°;

[0079] Among them, the calculation formula of the two-line interpolation method is as follows:

[0080]

[0081] Where (x1, y1), (x2, y2) are the coordinates of the grid points;

[0082] S3. Spatially average the CN05.1 dataset and the CMIP6 daily precipitation data.

[0083] A multivariate joint distribution model is constructed based on R-vine Copula. The R-vine Copula model design includes the following steps:

[0084] S1. Preferred selection of Vine tree structure. Any dimensional joint distribution function can be decomposed into a marginal distribution function and a copula function. R-vine becomes the structural selection range of Vine Copula function due to its structural flexibility. R-vine is defined by the following conditions:

[0085] f(x1,...,x p )=f(x1)·f(x2|x1)·...·f(x p |x1,...,x p-1 )

[0086] S1.1. If T = {T1, T2, ..., T n-1}, T1 is a node with N1={1,…,n} and an edge with E n of trees;

[0087] S1.2. For i=2,…,n-1, T i is to satisfy N1= i-1 ;

[0088] S1.3. For i=2,…,n-1, {a,b}∈E i , a={a1,a2},b={b1,B2}#(a,b)=1,# indicates the cardinality of the set;

[0089] Therefore, an n-dimensional R-vine structure can be represented by n-1 trees T.j There are n-1+j nodes and nj edges, which means there are nj Copula density functions; the tree T j The edges that share a node in will become nodes and are represented by the tree T j+1 The other edges in are connected;

[0090] The d-dimensional R-vine is defined by the following conditions:

[0091] f(x1,...,x d )=f(x1)·f(x2|x1)·...·f(x d |x1,...,x d-1 )

[0092] d-dimensional R-vine density function:

[0093]

[0094] Where: f i Represents x i The probability density of F(x i |x j+1 ,...,x i-1 ) is its univariate conditional distribution function, c ij|(j+1)...(i-1) is a given subset (x j+1 ,...,x i-1 ), with the two-dimensional variable (x i and i )’s density of the conditional vine copula function associated with the bivariate conditional distribution of ;

[0095] S2. Selection of Copula function: The minimum value of two model selection criteria, Akaike Information Criterion AIC and Bayesian Information Criterion BIC, is selected to determine the Copula function between variables. When the two criteria select different Copula functions, the final decision is mainly made by AIC. The best Pair-Copula function is selected from basic binary functions such as Gauss Copula, Student-t Copula, Clayton Copula, Gumbel Copula, and Frank Copula, including the following calculation formula:

[0096]

[0097] in, is the estimated likelihood, m is the number of adjustable parameters in the Copula function, and n is the number of data points;

[0098] S3. Parameter estimation of Copula function, estimate the parameters in Copula function through MLE maximum likelihood estimation;

[0099] Any two random variables u 1, The joint distribution of u2 can be represented by their respective marginal distributions and the Copula function C that represents the correlation between the two. Then the density function of the joint distribution function is:

[0100] h(u1,u2,θ)=c(F1(u1,a1),F2(u2,a2),δ)f1(u1,a1)f2(u2,a2)

[0101] The maximum likelihood function for an observed sample is:

[0102]

[0103] The maximum likelihood estimate It is a differential equation The solution;

[0104] S4. The multivariate joint distribution model between meteorological factors is constructed by R-vine Copula, and the Maximum Spanning Tree and MST algorithms are used to participate in model construction. The specific steps are as follows:

[0105] S4.1. In the processed sample data set, calculate the Kendall-τ rank correlation coefficient for any random variable combination of each meteorological factor;

[0106] S4.2. Generate the first tree Tree 1 of the R-vine structure by the MST algorithm, that is, maximize the sum of the absolute values ​​of all Kendall-τ correlation coefficients in the tree;

[0107] S4.3. For each edge on the tree, select the optimal binary Copula function for the edge using the binary Copula function selection methods AIC and BIC, and calculate the corresponding conditional distribution function;

[0108] S4.4. Under the defined conditions, find all conditionally related random variable combinations in each layer, and calculate the corresponding Kendall-τ rank correlation coefficient based on the conditional distribution function calculated in step tree S4.3;

[0109] S4.5. Based on the Kendall-τ rank correlation coefficient obtained in S4.4, repeat steps S4.2 to S4.4 until the last tree of the R-vine structure.

[0110] The VCCP method is used to perform change point test on the multivariate joint distribution model. In the VCCP method, binary segmentation and MOSUM are selected as two segmentation methods, and Vuong and Stationary Bootstrap are selected as two methods for inferring candidate change points. The segmentation method and the change point inference method are combined to obtain four detection methods for change point detection of the multivariate joint distribution model.

[0111] Building a multivariate joint distribution model based on R-vine Copula

[0112] Step 1: Select the daily-scale data of six meteorological factors in the CN05.1 dataset from 2000 to 2018, determine the marginal distribution of each meteorological factor, the Pre variable follows the gamma distribution, the SOI follows the normal distribution, and the remaining variables follow the KDE distribution;

[0113] Step 2: Use R-vine Copula to connect the six meteorological factors and build a six-dimensional multivariate joint distribution model. Figure 4 The selection of Copula functions at different tree levels, parameters, and the structure of multivariate joint distribution models are shown;

[0114] The VCCP method is used to perform change point test on the multivariate joint distribution model: the daily scale data set from 2000 to 2018 has a total of 6935 data points. Among the four VCCP change point detection methods, the mutation point occurs around the 4000th data point, see Figure 5 , that is, the multivariate joint distribution model changed suddenly in December 2010.

[0115] Long short-term memory recurrent neural network (LSTM) machine learning is used to correct the remote sensing precipitation data before and after the mutation point in the multivariate joint distribution model, including the following steps:

[0116] S1. Normalize the input data. Except for precipitation data, other input data are processed by Z-Score normalization. The formula is as follows:

[0117]

[0118] Where μ is the mean of all sample data, σ is the standard deviation of all sample data;

[0119] Since the precipitation distribution is discrete and has a large span, the precipitation data is log-normalized and the conversion function is:

[0120] y=log(x+1)

[0121] S2. Train the LSTM model and preliminarily determine the model parameters. The LSTM model consists of an input layer, a hidden layer, and an output layer. Each LSTM unit in the hidden layer contains three control state gates: input gate, forget gate, and output gate. The forward calculation process formula of LSTM is as follows

[0122] i t =σ(W xi x t +W hi h t-1 +b i )

[0123] f t =(W xf x t +W hf h t-1 +b f ) c t =f t ☉c t-1 +i t ☉tanh(W xc x t +W hc h t-1 +b c ) o t =σ(W x0 x t +W ho h t-1 +b o )

[0124] h t = t ⊙tanh(c t )

[0125] Where i, f, and o represent the input, forget, and output gates, respectively; ⊙ represents the multiplication of corresponding elements; W and b represent the weight matrix and bias vector of the network; t represents the time step; x and h represent the input and output vectors of the LSTM hidden layer; c represents the memory unit; and σ represents the Sigmoid function, which is a nonlinear activation function in machine learning.

[0126] The meteorological factor data selected from the CN05.1 dataset is used as input data, and the CN05.1 precipitation data is used as target data; the model is trained using the gradient descent method to optimize the model parameters; the model with the smallest loss function value of the validation set is selected as the optimal model; during the model training and optimization process, the model is continuously optimized through methods such as learning rate decay and early stopping to accelerate model convergence and prevent model overfitting;

[0127] S3. Model accuracy evaluation. The precipitation data in the CN05.1 dataset is used as the observed precipitation data, and the precipitation data in the CMIP6 dataset is used as the remote sensing satellite precipitation data. The correlation coefficient (CC), mean absolute error (MAE), root mean square coefficient (RMSE), and relative error (Bias) are used to evaluate the model effect. The optimal model parameters are shown in Table 2, and the model accuracy evaluation is shown in Table 3.

[0128]

[0129]

[0130]

[0131] Table 2, Optimal LSTM model parameters

[0132] parameter uncorrected After correction CC 0.48 0.54 MAE 0.25 0.20 RMSE 0.38 0.29 Bias 54.23％ 23.81％

[0133] Table 3, Model accuracy evaluation table

[0134] Where n represents the total amount of corrected precipitation data; G represents observed precipitation data, S represents corrected remote sensing satellite precipitation data; G i represents the i-th value in the observed precipitation data, S i represents the i-th value in the corrected remote sensing satellite precipitation data; represents the mean of the observed precipitation data, Represents the mean of the corrected remote sensing satellite precipitation data;

[0135] S4. The optimal LSTM model obtained after training is used to calibrate the remote sensing precipitation data for the years before and after the mutation point in the joint distribution model of multiple variables, and the optimal LSTM model obtained after training is used to calibrate the remote sensing precipitation data from 2000 to 2018. Since the multivariate joint distribution model mutated in December 2010, the trained LSTM model is used to calibrate the remote sensing precipitation data from 2000 to 2010 and from 2010 to 2018, and the correction effect is evaluated using CC, MAE, RMSE, and Bias.

[0136] Figure 6 The results of uncorrected, conventionally corrected, and corrected before and after the mutation point based on the multivariate joint distribution model are shown respectively. Figure 7 The error distribution under different correction conditions was selected from January to March 2001 and April to June 2006 before the mutation point, and from July to September 2011 and October to December 2016 after the mutation point. It can be seen that the conventional LSTM model has a certain effect on the bias correction of remote sensing precipitation data, and the bias correction effect before and after the mutation point is better than the conventional correction.

[0137] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. Remote sensing precipitation data bias correction method based on multivariate joint distribution and LSTM, Features: include: Preprocessing of CN05.1 dataset and CMIP6 daily precipitation data; A multivariate joint distribution model was constructed based on R-vine Copula. The VCCP method was used to perform change point test on the multivariate joint distribution model. Long short-term memory recurrent neural network (LSTM) machine learning was used to calibrate the remote sensing precipitation data before and after the mutation point in the multivariate joint distribution model.

2. According to the remote sensing precipitation data bias correction method based on multivariate joint distribution and LSTM according to claim 1, Features: The preprocessing of the CN05.1 dataset and the CMIP6 daily precipitation data includes the following steps: S1. Select some meteorological factor datasets from the CN05.1 dataset and select precipitation data under multiple climate models from CMIP6; for the multiple climate models selected in CMIP6, use the Multi-Model Ensemble (MME) to obtain the average precipitation value at each time step; S2. Unify the grid resolution of the selected data to 0.25°×0.25°; use the meteorological data analysis software Climate Data Operator (CDO) based on the Linux operating system to perform bilinear interpolation processing and analysis on the selected data, and unify its grid resolution to 0.25°×0.25°; Among them, the calculation formula of the two-line interpolation method is as follows: In the formula, (x 1 ,y 1 ),(x 2 ,y 2 ) are the coordinates of the grid points; S3. Spatially average the CN05.1 dataset and the CMIP6 daily precipitation data.

3. According to the remote sensing precipitation data bias correction method based on multivariate joint distribution and LSTM according to claim 1, Features: The multivariate joint distribution model is constructed based on R-vine Copula, and the R-vine Copula model design includes the following steps: S1. Preferred selection of Vine tree structure. Any dimensional joint distribution function can be decomposed into a marginal distribution function and a Copula function. R-vine becomes the structural selection range of Vine Copula function due to its structural flexibility. The R-vine is defined by the following conditions: f(x 1 ,...,x p )=f(x 1 )·f(x 2 |x 1 )·…·f(x p |x 1 ,...,x p-1 ) S1.

1. If T = {T 1 , T 2 ,…,T n-1 }, T 1 The node is N 1 = {1, ..., n}, with E as the edge 1 of trees; S1.

2. For i=2,…,n-1,T i is to satisfy N 1 =E i-1 ; S1.

3. For i=2,…,n-1, {a, b}∈E i , a={a 1 , a 2 }, b={b 1 , b 2 }#(a, b)=1, # represents the cardinality of the set; Therefore, an n-dimensional R-vine structure can be represented by n-1 trees (T), and the tree Tj has n-1+j nodes and nj edges, which means there are nj Copula density functions; the tree T j The edges that share a node in will become nodes and are represented by the tree T j+1 The other edges in are connected; The d-dimensional R-vine is defined by the following conditions: f(x 1 ,...,x d )=f(x 1 )·f(x 2 |x 1 )·…·f(x d |x 1 ,...,x d-1 ) d-dimensional R-vine density function: Where: f i Represents x i The probability density of F(x i |x j+1 , ..., x i-1 ) is its univariate conditional distribution function, c ij|(j+1)…(i-1) is a given subset (x j+1 , ..., x i-1 ), with the two-dimensional variable (x i and i )’s binary conditional distribution associated with the density of the conditional vineCopula function; S2. Selection of Copula function. The minimum value of two model selection criteria, Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC), is selected to determine the Copula function between variables. When the two criteria select different Copula functions, the final decision is mainly made by AIC. The best Pair-Copula function is selected from basic binary functions such as Gauss Copula, Student-t Copula, Clayton Copula, Gumbel Copula, and Frank Copula, including the following calculation formula: in, is the estimated likelihood, m is the number of adjustable parameters in the Copula function, and n is the number of data points; S3. Parameter estimation of Copula function, estimate the parameters in Copula function through MLE (maximum likelihood estimation); Any two random variables u 1, u 2 The joint distribution of can be represented by their respective marginal distributions and the Copula function C that represents the correlation between the two. Then the density function of the joint distribution function is: h(u i ,u 2 ,θ)=c(F 1 (u 1 ,a 1 ),F 2 (u 2 ,a 2 ),δ)f 1 (u 1 ,a 1 )f 2 (u 2 ,a 2 ) The maximum likelihood function for an observed sample is: The maximum likelihood estimate It is a differential equation The solution; S4. The multivariate joint distribution model between meteorological factors was constructed by R-vine Copula, and the Maximum Spanning Tree (MST) algorithm was used to participate in the model construction. The specific steps are as follows: S4.

1. In the processed sample data set, calculate the Kendall-τ rank correlation coefficient for any random variable combination of each meteorological factor; S4.

2. Generate the first tree (Tree 1) of the R-vine structure by the MST algorithm, that is, maximize the sum of the absolute values ​​of all Kendall-τ correlation coefficients in the tree; S4.

3. For each edge on the tree, select the optimal binary copula function for the edge using the binary copula function selection method (AIC and BIC), and calculate the corresponding conditional distribution function; S4.

4. Under the defined conditions, find all conditionally related random variable combinations in each layer, and calculate the corresponding Kendall-τ rank correlation coefficient based on the conditional distribution function calculated in step tree S4.3; S4.

5. Based on the Kendall-τ rank correlation coefficient obtained in S4.4, repeat steps S4.2 to S4.4 until the last tree of the R-vine structure.

4. The remote sensing precipitation data bias correction method based on multivariate joint distribution and LSTM according to claim 1, Features: The VCCP method is used to perform change point test on the multivariate joint distribution model. In the VCCP method, binary segmentation and MOSUM are selected as two segmentation methods, and Vuong and Stationary Bootstrap are selected as two methods for inferring candidate change points; the segmentation method and the change point inference method are combined to obtain four detection methods to perform change point detection on the multivariate joint distribution model.

5. The remote sensing precipitation data bias correction method based on multivariate joint distribution and LSTM according to claim 1, Features: The method of using long short-term memory recurrent neural network (LSTM) machine learning to correct remote sensing precipitation data before and after the mutation point in the multivariate joint distribution model includes the following steps: S1. Normalize the input data. Except for precipitation data, other input data are processed by Z-score normalization. The formula is as follows: Where μ is the mean of all sample data, σ is the standard deviation of all sample data; Since the precipitation distribution is discrete and has a large span, the precipitation data is log-normalized and the conversion function is: y=log(x+1) S2. Train the LSTM model and preliminarily determine the model parameters. The LSTM model is constructed by an input layer, a hidden layer, and an output layer. Each LSTM unit in the hidden layer contains three control state gates: an input gate, a forget gate, and an output gate. The forward calculation process formula of LSTM is as follows: i t =σ(W xi x t +W hi h t-1 +b i ) f t =(W xf x t +W hf h t-1 +b f ) c t =f t ⊙c t-1 +i t ⊙tanh(W xc x t +W hc h t-1 +b c ) o t =σ(W x0 x t +W ho h t-1 +b o ) h t =o t ⊙tanh(c t ) Where i, f, and o represent the input, forget, and output gates, respectively; ⊙ represents the multiplication of corresponding elements; W and b represent the weight matrix and bias vector of the network; t represents the time step; x and h represent the input and output vectors of the LSTM hidden layer; c represents the memory unit; and σ represents the Sigmoid function, which is a nonlinear activation function in machine learning. The meteorological factor data selected from the CN05.1 dataset is used as input data, and the CN05.1 precipitation data is used as target data; the model is trained using the gradient descent method to optimize the model parameters; the model with the smallest loss function value of the validation set is selected as the optimal model; during the model training and optimization process, the model is continuously optimized through methods such as learning rate decay and early stopping to accelerate model convergence and prevent model overfitting; S3. Model accuracy evaluation: the precipitation data in the CN05.1 dataset is used as the observed precipitation data, and the precipitation data in the CMIP6 dataset is used as the remote sensing satellite precipitation data. The correlation coefficient (CC), mean absolute error (MAE), root mean square error (RMSE), and relative error (Bias) are used to evaluate the model effect. Where n represents the total amount of corrected precipitation data; G represents observed precipitation data, S represents corrected remote sensing satellite precipitation data; G i represents the i-th value in the observed precipitation data, S i Represents the i-th value in the corrected remote sensing satellite precipitation data; represents the mean of the observed precipitation data, Represents the mean of the corrected remote sensing satellite precipitation data; S4. Using the optimal LSTM model obtained after training, the remote sensing precipitation data are corrected for the years before and after the mutation point in the joint distribution model of multiple variables, and the correction effect is evaluated using CC, MAE, RMSE, and Bias.

Images