LSTM (Long Short Term Memory)-based land-gas coupling data mixed assimilation method for introducing time-space lag effect

Abstract

The invention discloses a land-gas coupling data mixed assimilation method for introducing a time-space lag effect based on LSTM (Long Short Term Memory). The method comprises the following steps: generating set members by using RANDOMCV; a group of ensemble forecasting fields is obtained by using short-time ensemble forecasting, expanded land surface variable ensemble disturbance is calculated, and flow dependent background error information of land surface information is obtained; constructing a long-short memory method model; training a weight coefficient of a time-space lag term by taking a preprocessed satellite and conventional data in a past period of time as input; based on the analysis increment of mixed assimilation, adding the trained time-space lag term into the analysis increment; and assimilating satellites or conventional data by using a hybrid assimilation method of coupling land surface information control variables. According to the method, the influence of land surface information control variables is introduced, the LSTM model is used for training to realize the space-time lag effect brought by the difference between land and air, and the mode variable information is updated and optimized, so that a better analysis field is obtained.

Description

Technical Field

[0001] This invention relates to a method for mixing and assimilating land-atmosphere coupled data based on the spatiotemporal lag effect introduced by LSTM, belonging to the field of numerical simulation and data assimilation technology. Background Technology

[0002] In modern numerical weather prediction and climate forecasting, land surface information, as a key element of land-atmosphere interaction, directly influences the evolution of land surface energy, moisture exchange, and boundary layer processes. Land surface information such as humidity, temperature, and soil moisture plays a crucial role in the formation and development of weather systems. For example, land surface temperature affects the intensity of surface evaporation and latent heat flux, and is a key factor in triggering convection and precipitation; soil moisture and temperature affect surface sensible heat flux and local atmospheric stability, further regulating atmospheric circulation and the development of weather systems. Furthermore, due to the influence of various environmental factors, such as vegetation cover, soil condition, and seasonal differences, land surface information changes more slowly than atmospheric information. Data observed by satellites and other sources do not truly reflect the actual state of land surface information; there is a spatiotemporal lag effect between land and atmosphere. Traditional control variable settings (such as temperature, humidity, and wind field) typically do not consider the interaction between land surface and atmospheric information, weakening the influence of land surface information in the assimilation system. Meanwhile, traditional assimilation systems do not adequately consider the characteristics of land surface information during the assimilation of satellite and other data, often ignoring the spatiotemporal lag of land-atmosphere data. This often results in inaccurate feedback of true land surface information during the assimilation process, leading to final results that differ significantly from actual observations. Furthermore, in handling spatiotemporal lag effects, traditional methods often rely on fixed lag windows or static background error covariance (such as 3D-Var), failing to adapt to seasonal changes or variations in surface conditions. This causes adjustments to the spatiotemporal lag of land-atmosphere data to deviate from the actual lag, making it difficult to positively improve forecasts and potentially even having negative consequences. Summary of the Invention

[0003] Objective: To address the problems and shortcomings of existing technologies, this invention provides a hybrid assimilation method that couples land surface information control variables while incorporating the spatiotemporal lag characteristics of land and atmosphere based on Long Short-Term Memory (LSTM) model training. By extending the land surface information control variables using an ensemble / variational hybrid assimilation method and considering the slower change of land surface information compared to the atmosphere, the spatiotemporal lag effect is added to the land surface analysis increment and integrated into the assimilation system. This strengthens the coupling between land surface and atmospheric information and accurately reflects their influence, improving weather system analysis and forecasting, and enhancing the accuracy of numerical weather prediction and climate forecasting.

[0004] Technical Solution: A land-atmosphere coupled data assimilation method based on LSTM incorporating spatiotemporal lag effects. This method, combined with extended land surface information control variables and considering the spatiotemporal lag effects during LSTM model training, fully considers the spatiotemporal differences and interactions between land and atmosphere data, improving the analysis and forecasting of land surface and atmospheric information in the model. The method includes the following: S1, use RANDOMCV random perturbation to generate set members, add land surface information of set perturbation part as extended control variable to the hybrid assimilation system to obtain extended set perturbation variable; S2, using short-time ensemble forecasts to obtain a set of ensemble forecast fields, by calculating the ensemble mean and standard deviation of each variable in the ensemble members, and then calculating the extended land surface variable ensemble perturbation, the flow-dependent background error information of the land surface information is obtained; S3, Construct an LSTM model. The LSTM model includes the input sequence, a weight training module, a standard gating equation, and a spatiotemporal convolution module. S4, based on the introduced land surface information control variables, constructs a hybrid assimilation analysis increment that considers the spatiotemporal lag effect, so that the spatiotemporal state differences between land and air and the influence of the flow dependence background error information of land surface information can be fully considered; based on the newly constructed hybrid assimilation method that couples land surface information control variables and spatiotemporal lag effects, data assimilation of satellite or conventional data is carried out.

[0005] Furthermore, in S1, land surface information from the ensemble perturbation component is added to the hybrid assimilation system, expanding the control variables of the ensemble flow-dependent component. The expanded ensemble perturbation variables... for: (1) For atmospheric variable set perturbation; To perturb the extended set of land surface variables.

[0006] Furthermore, in S2, based on the random perturbation-generated ensemble members, a set of ensemble forecast fields is obtained using short-term ensemble forecasts; the ensemble perturbation information of the variables is obtained by calculating the ensemble mean and standard deviation of the variables of each ensemble member in the ensemble forecast field; then, the extended ensemble perturbation information of the land surface variables is calculated; by calculating the mean and standard deviation of each variable among the members, the dispersion of the land surface variables in the ensemble members is quantitatively characterized, which is used to directly quantify the uncertainty of the surface variables at different spatial locations and times; based on the calculation results of the mean and standard deviation, the flow-dependent background error information is obtained, which can dynamically reflect the spatiotemporal variation characteristics of the land surface variable error with the evolution of the weather system; the formula for calculating the mean and standard deviation of the land surface information is as follows: (2) (3) (4) in This represents the average value of the land surface variable. Let K represent the land surface variable of the k-th member, where K is the number of members in the set; For the perturbation of the land surface variable set, This represents the set disturbance of land surface variables generated by the k-th set member; stdv is the standard deviation of the land surface variables.

[0007] Furthermore, in S3, before calculating the spatiotemporal lag term, the LSTM model first needs to use assimilated data from a past period as the input sequence to train the weights and bias coefficients for the time and space lags. The input sequence expression is as follows: (5) Given the input sequence, t is the target assimilation time. and represents the analytical increments of the atmosphere and land surface, respectively, and M represents the assimilation of the past M times.

[0008] The weight training module primarily utilizes forward propagation during training. It first initializes the initial weights and bias coefficients using Xavier / Glorot, then uses past results with lag terms to perform short-term forecasts. The forecast results are compared with actual observations, and the loss is calculated. Afterward, based on the loss, BPTT is used to calculate the gradients of all weights and bias coefficients, and SGD is used to update and optimize the weights and bias coefficients. The entire process is an iterative optimization process. Weight training needs to be completed before calculating the time lag term at the current assimilation time. After training the weights and bias coefficients, the land-atmosphere data of the current assimilation time window is input into the input sequence. The input sequence is then passed to the standard gating equations for calculation, and finally substituted into the spatiotemporal convolution module, which is as follows: (6) (7) (8) (9) (10) in , , , , , These are all weights and bias coefficients trained during the training process. Let h(t) be the time weight, and h(t) be the hidden state in the standard gating equation. The contribution of different assimilation moments in the past to the time lag effect. The term represents the time lag, and L represents the total time step length. For grid points exist The increase in land surface area at any given moment The weights for spatial offset, For the propagation delay of the offset, For spatial lag, () represents the position offset ,Delay The land surface increment over time, where N is the spatial neighborhood.

[0009] Furthermore, the analytical increment for land surface variables is defined in the hybrid assimilation system as follows: (11) This represents the increment of land surface information resulting from the static background error covariance in 3DVar. Since land surface information is not introduced as a control variable in 3DVar, the static portion of the land surface information increment... Set to 0, where K is the number of members in the set; The increment of land surface information variables brought about by the flow-dependent error background information provided to the set; where This represents the control variable vector for the Kth set member, where the control variables include atmospheric variables and land surface information variables; It is a spacetime lag term, and is a time lag term. and spatial lag term The sum takes into account the lag effect of the land surface relative to the atmosphere.

[0010] In a mixed assimilation system, the analytical increment of atmospheric variables is defined as: (12) in It is the increase in atmospheric information brought about by the covariance of the 3DVar static background error. This represents the increase in atmospheric information brought about by the set. This represents the set disturbance of atmospheric variables generated by the k-th set member; In assimilating satellite or conventional data, a method is developed that considers the spatiotemporal lag effect between land and atmosphere while adding perturbations to the land surface variable set. This method is based on an LSTM model and trained to derive a systematic method for calculating time and space lag terms. By incorporating the spatiotemporal lag terms into the land surface analysis increment, an analysis increment that considers both land-atmosphere interaction and the relative lag effect of the land surface on the atmosphere is obtained.

[0011] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the land-atmosphere coupled data mixing and assimilation method based on LSTM introducing spatiotemporal lag effects as described above.

[0012] A computer-readable storage medium storing a computer program that performs the land-atmosphere coupled data assimilation method based on LSTM-introduced spatiotemporal lag effects as described above.

[0013] Beneficial effects: In the process of hybrid assimilation, after generating ensemble members using RANDOMCV, this invention first utilizes the land-atmosphere time lag characteristic to perform ensemble forecasts for the land surface and atmosphere by adopting the same start time but different ensemble forecast durations. Then, it introduces land surface information control variables and combines them with LSTM model training to fully consider the spatiotemporal lag of land and atmosphere. This aims to correct the shortcomings of satellite and other data in accurately reflecting land surface information, update and optimize model variables, and thus obtain a better analytical field. Attached Figure Description

[0014] Figure 1 This is a flowchart of the method according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the training and optimization of coefficients such as weights and biases under the LSTM model in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the calculation of the spatiotemporal lag term at the assimilation time using an LSTM model in an embodiment of the present invention. Detailed Implementation

[0015] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0016] like Figure 1 As shown, this embodiment of the invention provides a method for hybrid assimilation of satellite or conventional data based on LSTM model training to calculate spatiotemporal lag effects and couple land surface information control variables. The aim is to utilize the LSTM model to achieve more flexible and efficient calculation of spatiotemporal lags and to incorporate land-atmosphere interactions into the hybrid assimilation process. The specific steps are as follows: S1 uses RANDOMCV to generate set members and couples land surface information control variables.

[0017] Traditional control variables only include atmospheric variables such as wind speed, temperature, and humidity, with little attention paid to land surface variables. Adding global control variables necessitates calculating the static background error covariance of the corresponding extended control variables. Therefore, the control variables for the ensemble flow-dependent part are extended, adding land surface information to the ensemble forecast perturbation field. This method effectively incorporates land surface information into the control variable space, allowing for a more comprehensive consideration of land-atmosphere interactions within the assimilation system. After generating ensemble members using RANDOMCV, this extended ensemble perturbation is integrated into the ensemble members. This better reflects the potential impact of land surface information on atmospheric analysis increments and model variable updates during subsequent assimilation, thereby enhancing the assimilation system's ability to capture complex feedback mechanisms between land and atmosphere. The ensemble perturbation with extended land surface information is as follows: For atmospheric variable set perturbation; To perturb the extended set of land surface variables.

[0018] S2 utilizes ensemble members generated by RANDOMCV for short-term ensemble forecasting, and obtains flow-dependent background error information by calculating the ensemble average and disturbance of ensemble forecast members incorporating land surface information.

[0019] In a hybrid assimilation system coupling land surface information control variables, ensemble members are generated using the RANDOMCV method. Based on this, short-term ensemble forecasts are used to update the ensemble members, and variables related to land surface information are incorporated into the calculation of the ensemble mean and standard deviation for comprehensive calculation. This allows for the calculation of extended land surface variable ensemble perturbations, yielding flow-dependent background error information for land surface variables, thus reflecting the magnitude and variation characteristics of land surface information errors in both spatial and temporal dimensions. The perturbation information includes both atmospheric and land surface variable ensemble perturbations, reflecting the mutual influence between variables when calculating the ensemble mean and standard deviation of the ensemble members. In this way, the model can comprehensively consider the interaction between atmospheric and land surface information, capturing the feedback mechanism between the atmosphere and land surface, and demonstrating their synergistic effect in the assimilation system. The formulas for calculating the land surface information mean and standard deviation are as follows: in This represents the average value of the land surface variable. Let K represent the land surface variable of the k-th member, where K is the number of members in the set; For the perturbation of the land surface variable set, This represents the set disturbance of land surface variables generated by the k-th set member; stdv is the standard deviation of the land surface variables.

[0020] S3, constructing an LSTM model, primarily used to train the weights and biases for spatiotemporal lag effects. The loss function incorporates spatiotemporal lag effects; after adding a spatiotemporal lag term to the land surface variable, the model undergoes mixing and assimilation, and the spatiotemporal weights and biases are indirectly constrained based on the short-term forecast errors. See [link to specific workflow] for details. Figure 2 The training steps include: (1) First, using the Xavier / Glorot initialization method, initial values ​​of all weight coefficients are given, with the initial value of the bias coefficient set to 0; (2) Inputting past time series data and the initialized coefficients into the standard gating equation to calculate and obtain the hidden state h(t); (3) Substituting the hidden state h(t) into the time and space weight formulas to obtain the results; (4) Substituting the time and space weights into the time and space lag expressions to obtain the final time and space lag terms; (5) Using the past assimilation results with time and space lag terms after training to perform short-term forecasts and forecast the results. Compare the reported results with the observation data at the corresponding time, and construct a loss function S with a time lag effect; (6) Examine the change of the loss function. When the loss function decreases by a small amount to a certain extent, jump out of the training loop and take the final generated result as the best weight and bias. Otherwise, proceed to the next step normally; (7) Calculate the gradient of the weight and bias coefficients using BPTT based on the loss function; (8) After calculating the gradient, update each weight and bias coefficient using SGD and resubmit it into the training and optimization of the weight and bias coefficients in the LSTM model (steps (2)-(8) above).

[0021] The expressions for time weights, time lag terms, spatial weights and biases, and spatial lag terms during training are as follows: in, , , , , , These are all weights and bias coefficients trained during the training process. Let h(t) be the time weight, and h(t) be the hidden state in the standard gating equation. The contribution of different assimilation moments in the past to the time lag effect. The term represents the time lag, and L represents the total time step length. For grid points exist The increase in land surface area at any given moment The weights for spatial offset, For the propagation delay of the offset, For spatial lag, () represents the position offset ,Delay The land surface increment over time, where N is the spatial neighborhood.

[0022] The loss function expression considering the spatiotemporal lag effect during training is as follows: Where n represents the total number of samples, For short-time forecasts with time t as the initial time Subsequent land surface variables, For the corresponding The actual observed land surface variables at any given time.

[0023] S4, After training is complete, in the Hybrid En3Dvar hybrid assimilation system, an analysis increment considering the time lag effect is defined based on the introduced land surface information control variable: This represents the increment of land surface information resulting from the static background error covariance in 3DVar. Since land surface information is not introduced as a control variable in 3DVar, the static portion of the land surface information increment... Set to 0, where K is the number of members in the set; The increment of land surface information variables brought about by the flow-dependent error background information provided to the set; where This represents the control variable vector for the Kth set member, where the control variables include atmospheric variables and land surface information variables; It is a spacetime lag term, and is a time lag term. and spatial lag term The sum takes into account the lag effect of the land surface relative to the atmosphere.

[0024] In a mixed assimilation system, the analytical increment of atmospheric variables is defined as: in It is the increase in atmospheric information brought about by the covariance of the 3DVar static background error. This represents the increase in atmospheric information brought about by the set. This represents the set disturbance of atmospheric variables generated by the k-th set member; because Because of disturbances in the ensemble of land surface variables, the analytical increments of atmospheric variables are also affected by land surface information. Changes in land surface information are... Adjusting the analytical increments of atmospheric variables affects the analytical results of model variables, allowing for the updating and optimization of these variables. This not only reflects the coupling relationship between the atmosphere and land surface in the assimilation system but also the impact of land surface information on atmospheric analysis results, thus ensuring that the assimilation system can more comprehensively capture the feedback mechanism between the land surface and the atmosphere.

[0025] S5 is a hybrid assimilation system based on an LSTM model, incorporating spatiotemporal lags and coupled with land surface information. It assimilates data within the current assimilation window. The operation process is as follows: Figure 3 As shown, the data from the assimilation window region is input as a time series into the standard gating equation to calculate the hidden state h(t). After calculating h(t), it is substituted into the time weight and spatial weight to obtain the time and spatial weights at the assimilation time. Finally, the two weights and the spatial bias are substituted into the time and spatial lag expressions to obtain the analysis increment with time and space lag terms at the assimilation time.

[0026] This novel assimilation method effectively reduces forecast errors caused by spatiotemporal biases in land-atmosphere data while reflecting the interaction between land and atmosphere. The inclusion of land surface information control variables not only strengthens the adjustment of atmospheric variables but also effectively captures the mutual influence between the surface and the atmosphere, fully representing the interaction between land and atmosphere. Furthermore, the spatiotemporal lag terms obtained from LSTM model training have a more flexible dynamic change capability compared to traditional spatiotemporal lag calculations. They can adjust weights and biases in real time based on factors such as seasonality and changes in surface conditions, and can also consider spatial elements such as topographic features, fully leveraging their unique advantages of high scalability, multiple tools, ease of modification, and long memory duration.

[0027] Obviously, those skilled in the art should understand that the steps of the land-atmosphere coupled data assimilation method based on LSTM introducing spatiotemporal lag effects in the above-described embodiments of the present invention can be implemented using general-purpose computing devices. These steps can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, which can then be stored in a storage device for execution by the computing device. In some cases, the steps shown or described can be executed in a different order than those described herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the embodiments of the present invention are not limited to any specific hardware and software combination. Furthermore, without departing from the core ideas and technical effects of the present invention, any structural modifications, adjustments, or replacements to the data input type, optimization method, spatiotemporal weights, lag terms, and the construction of backpropagation and updates, based on specific application conditions and performance requirements, should be considered equivalent implementations of the present invention.

Claims

1. A method for mixing and assimilating land-atmosphere coupled data based on the spatiotemporal lag effect introduced by LSTM, characterized in that, Includes the following steps: S1, use RANDOMCV random perturbation to generate set members, add land surface information of set perturbation part as extended control variable to the hybrid assimilation system to obtain extended set perturbation variable; S2, using short-time ensemble forecasts to obtain a set of ensemble forecast fields, by calculating the ensemble mean and standard deviation of each variable in the ensemble members, and then calculating the extended land surface variable ensemble perturbation, the flow-dependent background error information of the land surface information is obtained; S3, Construct an LSTM model. The LSTM model includes an input sequence, a weight training module, a standard gating equation, and a spatiotemporal convolution module. S4, based on the introduced land surface information control variables, constructs a hybrid assimilation analysis increment that considers the spatiotemporal lag effect, so that the spatiotemporal state differences between land and air and the influence of the flow dependence background error information of land surface information can be fully considered; based on the newly constructed hybrid assimilation method that couples land surface information control variables and spatiotemporal lag effects, data assimilation of satellite or conventional data is carried out.

2. The land-atmosphere coupled data assimilation method based on the spatiotemporal lag effect introduced by LSTM according to claim 1, characterized in that, In S1, land surface information from the ensemble perturbation component is added to the hybrid assimilation system, expanding the control variables of the ensemble flow-dependent component. The expanded ensemble perturbation variables... for: （1） For atmospheric variable set perturbation; To perturb the extended set of land surface variables.

3. The land-atmosphere coupled data assimilation method based on the spatiotemporal lag effect introduced by LSTM according to claim 1, characterized in that, In step S2, based on the random perturbation generating ensemble members, a set of ensemble forecast fields is obtained using short-term ensemble forecasts. The ensemble perturbation information of the variables is obtained by calculating the ensemble mean and standard deviation of the variables of each ensemble member in the ensemble forecast field. Then, the extended ensemble perturbation information of the land surface variables is calculated. By calculating the mean and standard deviation of each variable among the members, the dispersion of the land surface variables in the ensemble members is quantitatively characterized, which is used to directly quantify the uncertainty of the surface variables at different spatial locations and times. Based on the calculation results of the mean and standard deviation, the flow-dependent background error information is obtained, which can dynamically reflect the spatiotemporal variation characteristics of the land surface variable error with the evolution of the weather system. The formulas for calculating the mean and standard deviation of land surface information are as follows: （2） （3） （4） in This represents the average value of the land surface variable. Let K represent the land surface variable of the k-th member, where K is the number of members in the set; For the perturbation of the land surface variable set, This represents the set disturbance of land surface variables generated by the k-th set member; stdv is the standard deviation of the land surface variables.

4. The land-atmosphere coupled data assimilation method based on the spatiotemporal lag effect introduced by LSTM according to claim 1, characterized in that, The LSTM model is used to train the weights and bias coefficients of the spatiotemporal lag effect. The spatiotemporal lag effect is considered in the construction of the loss function. After adding the spatiotemporal lag term to the land surface variable, the spatiotemporal weights and biases are indirectly constrained by the forecast error of the short-term forecast after mixing and assimilation. The training steps include: (1) First, use the Xavier / Glorot initialization method to give the initial value of the coefficients, where the initial value of the bias coefficient is set to 0; (2) Input the past time series data and the initialized coefficients into the standard gating equation to calculate and obtain the hidden state h(t); (3) Substitute the hidden state h(t) into the time and space weight formula to obtain the results of both; (4) Substitute the time and space weights into the time and space lag expression to obtain the final time and space lag terms; (5) Use the past assimilation results with time and space lag terms after training to make short-term forecasts and compare the forecast results with the observation data at the corresponding time to construct a loss function S with time and space lag effect; (6) Check the change of the loss function. When the loss function decreases to the required level, the training loop will be exited and the final generated result will be used as the best weight and bias; (7) BPTT is used to calculate the gradient of the weight and bias coefficients based on the loss function; (8) After calculating the gradient, SGD is used to update each weight and bias coefficient and re-substitute it into the LSTM model to execute steps (2)-(8).

5. The land-atmosphere coupled data assimilation method based on the spatiotemporal lag effect introduced by LSTM according to claim 1, characterized in that, In S3, before calculating the spatiotemporal lag term, the LSTM model first needs to use assimilated data from a past period as the input sequence to train the weights and bias coefficients for the time and space lags. The input sequence expression is as follows: （5） Given the input sequence, t is the target assimilation time. and represents the analytical increments of the atmosphere and land surface, respectively, and M represents the assimilation of the past M times.

6. The land-atmosphere coupled data assimilation method based on the spatiotemporal lag effect introduced by LSTM according to claim 1, characterized in that, The weight training module utilizes forward propagation during training. First, it initializes the initial weights and bias coefficients using Xavier / Glorot. Then, it substitutes past time-series data and performs forward propagation to calculate the first spatiotemporal lag term. This lag term is then added to the incremental analysis of land surface variables for mixing and assimilation, and short-term forecasts are generated. Next, it uses past results with lag terms to generate short-term forecasts, compares the forecast results with actual observations, and calculates the loss. Afterward, based on the loss, it uses BPTT to calculate the gradient of all weights and bias coefficients and uses SGD to update and optimize the weights and bias coefficients. Weight training must be completed before calculating the time lag term at the current assimilation time. After training the weights and bias coefficients, the land-atmosphere data of the current assimilation time window is input into the input sequence. The input sequence is then passed to the standard gating equations for calculation. Finally, the results are substituted into the spatiotemporal convolution module, which is as follows: （6） （7） （8） （9） （10） in , , , , , These are all weights and bias coefficients trained during the training process. Let h(t) be the time weight, and h(t) be the hidden state in the standard gating equation. The contribution of different assimilation moments in the past to the time lag effect. The term represents the time lag, and L represents the total time step length. For grid points exist The increase in land surface area at any given moment The weights for spatial offset, For the propagation delay of the offset, For spatial lag, () represents the position offset ,Delay The land surface increment over time, where N is the spatial neighborhood.

7. The land-atmosphere coupled data assimilation method based on the spatiotemporal lag effect introduced by LSTM according to claim 4, characterized in that, The loss function expression considering the spatiotemporal lag effect during training is as follows: （11） Where n represents the total number of samples, For short-time forecasts with time t as the initial time Subsequent land surface variables, For the corresponding The actual observed land surface variables at any given time.

8. The land-atmosphere coupled data assimilation method based on the spatiotemporal lag effect introduced by LSTM according to claim 1, characterized in that, In a hybrid assimilation system, the analytical increment for land surface variables is defined as: （12） This represents the increment of land surface information resulting from the static background error covariance in 3DVar. Since land surface information is not introduced as a control variable in 3DVar, the static portion of the land surface information increment... Set to 0, where K is the number of members in the set; The increment of land surface information variables brought about by the flow-dependent error background information provided to the set; where This represents the control variable vector for the Kth set member, where the control variables include atmospheric variables and land surface information variables; It is a spacetime lag term, and is a time lag term. and spatial lag term The sum takes into account the hysteresis effect of the land surface relative to the atmosphere; In a mixed assimilation system, the analytical increment of atmospheric variables is defined as: （13） in It is the increase in atmospheric information brought about by the covariance of the 3DVar static background error. This represents the increase in atmospheric information brought about by the set. This represents the set disturbance of atmospheric variables generated by the k-th set member.

9. A computer device, characterized in that: The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the land-atmosphere coupled data mixing and assimilation method based on the spatiotemporal lag effect introduced by LSTM as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program that executes a land-atmosphere coupled data assimilation method based on LSTM-introduced spatiotemporal lag effects as described in any one of claims 1-8.

Images