Terrestrial Water Storage Change Grid Data Processing Method and Storage Medium

Abstract

The present invention discloses a method for processing grid data of terrestrial water storage change and a storage medium, belonging to the field of data processing for prediction purposes. To solve the problem of processing accuracy loss during the downscaling process of grid data of terrestrial water storage change, the present invention first obtains two types of hydrometeorological grid data of the study area and multi-source original grid data of terrestrial water storage change; uses the generalized triangular hat method and the least squares method to fuse the multi-source original data to obtain fused data; for the grid points in the study area, the elements of the first hydrometeorological grid data are removed one by one, input into the LSTM prediction model, and simulated by the residual correction method to screen out the key hydrometeorological elements whose influence on accuracy exceeds the threshold; the grid point sub-regions are divided according to the key elements, and each sub-region is trained based on the LSTM prediction model and the residual correction method to obtain the optimal model and the corresponding data subset; the data of the months containing extreme points are corrected using the optimal model to obtain the final result. The present invention is used in a remote sensing system and has the advantage of high accuracy.

Description

Technical Field

[0001] The present invention relates to the field of data processing for prediction purposes, and particularly to a method for processing grid data of terrestrial water storage change and a storage medium. Background Art

[0002] For the prediction research on terrestrial water storage change, according to different spatial scales, it can be divided into small and medium-scale regions and large-scale regions. Small and medium-scale regions cover basins, provincial or sub-national regions, etc. The prediction work on terrestrial water storage change in these regions focuses on the high spatio-temporal resolution water storage changes in local regions such as the Yangtze River Basin and the Sichuan Basin. By capturing hydrological dynamics on short time scales (from days to months), such as floods and drought events caused by extreme rainfall, it provides key basis for regional water resources management, ecological environment protection, and disaster warning. Large-scale regions include continents, national and even global scales. At this scale, the prediction of subtle changes in terrestrial water storage is mainly to reveal the long-term water storage change trends at the global or continental level, such as inter-annual changes and the evolution trend over decades. At the same time, it deeply analyzes the impacts of climate change and human activities (such as agricultural irrigation, reservoir operation, etc.) on water resources, so as to support the formulation of global water security policies.

[0003] In order to convert grid data of terrestrial water storage change with low spatial resolution into grid data of terrestrial water storage change with high spatial resolution (i.e., spatial downscaling), the current downscaling methods mainly include: (1) statistical downscaling method, (2) physical model downscaling method, (3) data assimilation method. The statistical downscaling method relies on the statistical relationship between GRACE data and high-resolution hydrometeorological data, and uses means such as regression analysis or machine learning to improve the resolution. However, such methods usually lack a physical constraint mechanism, resulting in poor generalization ability. The physical model downscaling method combines a land surface model or numerical simulation, and uses a high-resolution model as prior knowledge to improve the spatial resolution. However, the existing physical models have problems of uncertainty and spatial heterogeneity. The data assimilation method uses a filtering method to fuse GRACE data with a high-resolution hydrological model to enhance the spatial resolution, but its filtering calculation cost is high, and it is extremely sensitive to initial conditions and parameters.

[0004] Therefore, the field first needs to find a simple spatial downscaling method that can process complex multi-source data, accurately capture the long-term dependencies of different time series, have excellent time series prediction ability, and can automatically adjust the learning content and efficiently predict.

[0005] Secondly, traditional spatial downscaling methods generally face the problem of spatial heterogeneity. When dividing grid points, existing methods usually simply divide them according to administrative regions. For example, the Yangtze River is divided into the upper reaches, middle reaches, and lower reaches of the Yangtze River. However, the impacts of different regions on terrestrial water storage are different. If the same prediction model is used, the situation of all regions cannot be accurately considered. Therefore, the simple grid point division method leads to poor accuracy in spatial downscaling processing.

[0006] Finally, existing technologies have always ignored the impact of extreme points caused by sudden mutations and non-linear changes in grid data of terrestrial water storage changes on the accuracy of spatial downscaling. Therefore, it becomes more urgent to solve the problem of insufficient processing ability of spatial downscaling methods for non-linear relationships.

[0007] Prior Art 1: Lilu Cui, Yu Li, et al. Assessing the impact of 2022extreme drought on the Yangtze River basin using downscaled GRACE / GRACE-FOdata obtained by partitioned random forest algorithm. Vol. 3 Issue 3, 2024, International Journal of Remote Sensing.

[0008] Prior Art 1 proposed a partitioned random forest downscaling method, which can effectively improve the accuracy compared with the traditional global random forest downscaling method. However, for the above technical problems, no more effective solutions were proposed in Prior Art 1. Summary of the Invention

[0009] To alleviate or partially alleviate the above technical problems, the solution of the present invention is as follows:

[0010] A method for processing grid data of terrestrial water storage changes, comprising:

[0011] Step S1: Obtain first hydrometeorological grid data, second hydrometeorological grid data, and original grid data of terrestrial water storage changes from multiple sources within the research area;

[0012] Step S2: Use the generalized triangular cap method and the least squares method to fuse the original grid data of terrestrial water storage changes from multiple sources to obtain the fused grid data of terrestrial water storage changes;

[0013] Step S3: For any grid point in the study area, after excluding each hydro-meteorological element in the first hydro-meteorological grid data one by one, input it into the LSTM prediction model, and obtain the simulated terrestrial water storage change grid data through the residual correction method. Retain the hydro-meteorological elements whose influence on the accuracy of the simulated terrestrial water storage change grid data is greater than the preset threshold as the key hydro-meteorological elements;

[0014] Step S4: Divide the grid points with exactly the same key hydro-meteorological elements into one sub-region; for each sub-region, train and obtain the optimal LSTM prediction model based on the LSTM prediction model and the residual correction method, and use the optimal LSTM prediction model to obtain the terrestrial water storage change grid data for each sub-region, and form a subset of the first terrestrial water storage change grid data;

[0015] Step S5: For each sub-region and the months containing extreme points in the fused terrestrial water storage change grid data, use the optimal LSTM prediction model to obtain the terrestrial water storage change grid correction data for the month, and use the terrestrial water storage change grid correction data for the month to replace the first terrestrial water storage change grid data for the month to obtain the final terrestrial water storage change grid data.

[0016] Further, the spatial resolution of the first hydro-meteorological grid data is less than the spatial resolution of the second hydro-meteorological grid data.

[0017] Further, the spatial resolution of the original terrestrial water storage change grid data from multiple sources is lower than the spatial resolution of the final terrestrial water storage change grid data and lower than the spatial resolution of the first terrestrial water storage change grid data.

[0018] Further, the residual correction method includes the following steps:

[0019] After taking the difference between the predicted terrestrial water storage change grid data output by the LSTM prediction model and the fused terrestrial water storage change grid data, interpolate to obtain the second residual grid data as the correction data.

[0020] Further, the residual correction method specifically includes the following steps:

[0021] Take the first hydro-meteorological grid data and the fused terrestrial water storage change grid data as the inputs of the LSTM prediction model to obtain the predicted terrestrial water storage change grid data;

[0022] Take the difference between the predicted terrestrial water storage change grid data and the fused terrestrial water storage change grid data to obtain the first residual grid data, and then use the Kriging interpolation method to interpolate the first residual grid data into the second residual grid data;

[0023] Use the second hydro-meteorological grid data as the input of the LSTM prediction model to obtain the second terrestrial water storage change grid data, and add the second terrestrial water storage change grid data to the second residual grid data to obtain the target terrestrial water storage change grid data.

[0024] Furthermore, when obtaining the fused terrestrial water storage change grid data, the optimal weight for data fusion is obtained by the least squares method.

[0025] Furthermore, for any hydro-meteorological grid data input into the LSTM prediction model, standardization processing is first performed.

[0026] Furthermore, the neurons in the LSTM prediction model use the Sigmoid activation function.

[0027] Furthermore, for different sub-regions, the applied optimal LSTM prediction models are different.

[0028] On the other hand, the present invention also discloses a storage medium storing computer programs / instructions, which when executed or compiled and executed by a processor, implement the steps of the terrestrial water storage change grid data processing method described in any one of the above.

[0029] The technical solution of the present invention has one or more of the following beneficial technical effects:

[0030] (1) Independently process the grid data at the extreme values in the time series using a long short-term memory network to reduce the influence of non-linear changes on the accuracy of spatial downscaling results processing.

[0031] (2) By predicting all input hydro-meteorological elements based on the long short-term memory network, different key hydro-meteorological elements for different grid points are determined, effectively weakening the adverse influence of spatial heterogeneity on the accuracy of spatial downscaling processing.

[0032] (3) Adopt a spatial downscaling method based on the long short-term memory network, which can capture long-term dependencies in the time series, retain and output key information through gating mechanisms and memory units, and its processing method is simple and efficient, with gradient stability.

[0033] In addition, other beneficial effects of the present invention will be mentioned in the specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a flowchart of the terrestrial water storage change grid data processing method. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0035] To make the objectives, technical solutions and advantages of the present invention clearer, the following will clearly and completely describe the technical solutions in the present invention with reference to the accompanying drawings in the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present invention without creative efforts shall fall within the protection scope of the present invention.

[0036] For the convenience of clearly describing the technical solutions of the embodiments of the present invention, in the embodiments of the present invention, terms such as "first" and "second" are used to distinguish the same items or similar items with basically the same functions and effects. Those skilled in the art can understand that terms such as "first" and "second" do not limit the quantity and execution order. In addition, the "multiple" mentioned in the present invention means at least three.

[0037] Term Explanation:

[0038] Gravity Recovery And Climate Experiment (GRACE): It refers to a space geoscience mission that monitors geophysical phenomena such as global terrestrial water storage, ocean circulation, and glacier melting by measuring changes in the Earth's gravity field, providing key GRACE data for studying climate change, water resource distribution, and Earth system science. The GRACE mission measures changes in the Earth's gravity field through GRACE satellites and obtains relatively low spatial resolution terrestrial water storage change data on a global scale.

[0039] Global Land Data Assimilation System (GLDAS): The main purpose of GLDAS is to provide estimates of terrestrial surface hydrological and meteorological variables with high spatio-temporal resolution on a global scale by integrating multiple observational data and model simulation results. It combines satellite observational data, ground observational data, and the output of numerical weather prediction models, and uses data assimilation technology to optimize model parameters and initial conditions, thereby generating more accurate and comprehensive terrestrial surface state information.

[0040] Long - Short Term Memory (LSTM): It is a special recurrent neural network in the field of artificial intelligence, which is good at and usually used to process sequential data such as speech.

[0041] Three - Cornered Hat (TCH): It is a method for estimating the error variance of each data set by analyzing the mutual relationship between multiple data sets, and is used to quantify the uncertainty of multiple data sets (generally at least three).

[0042] Study area: The geographical area targeted or selected by the data processing method of the present invention.

[0043] Evapotranspiration: It refers to the process by which water on the Earth's surface enters the atmosphere through evaporation and plant transpiration. It is an important link in the water cycle. Evaporation is the process of converting liquid water into gaseous water, mainly occurring on water surfaces, soil surfaces, etc. Transpiration is the process by which plants release the water in their bodies into the atmosphere through the stomata on the leaf surface.

[0044] Hydrometeorological grid data: It is a data form that organizes and stores hydrological and meteorological element information according to a certain spatial grid and can be obtained from the meteorological bureau.

[0045] Terrestrial water storage change grid data: It is a data form used to describe the change of terrestrial water storage on the Earth's land surface over time and space, containing information on the storage changes of various terrestrial water bodies (such as rivers, lakes, reservoirs, soil moisture, groundwater, etc.). By comprehensively monitoring and analyzing the storage changes of these different water bodies, the dynamic changes of terrestrial water storage at different regions and different time scales can be reflected.

[0046] The method for processing terrestrial water storage change grid data of the present invention includes the following steps:

[0047] Step S1: Obtain the first hydrometeorological grid data, the second hydrometeorological grid data, and the original terrestrial water storage change grid data from multiple sources within the study area.

[0048] In the present invention, the original terrestrial water storage change grid data from multiple sources can be the terrestrial water storage change grid data released by multiple institutions well-known in the art. The present invention preferably selects at least 5 different sources of original terrestrial water storage change grid data. The original terrestrial water storage change grid data from multiple sources can be regarded as low-spatial-resolution terrestrial water storage change grid data in the present invention. The present invention is not limited to a specific 5 different sources of original terrestrial water storage change grid data.

[0049] In the present invention, the hydrometeorological grid data includes but is not limited to the following hydrometeorological elements: precipitation, evapotranspiration, snow water equivalent, snow depth, vegetation canopy moisture, runoff, surface temperature, and soil moisture data.

[0050] The present invention involves the first hydrometeorological grid data and the second hydrometeorological grid data, where the spatial resolution of the first hydrometeorological grid data is lower than that of the second hydrometeorological grid data. That is to say, the second hydrometeorological grid data has a high spatial resolution, and the first hydrometeorological grid data has a low spatial resolution.

[0051] The meteorological grid data and the grid data of terrestrial water storage change are divided into different grid points for different positions in the study area, which belongs to the conventional processing methods well-known in the art, and the present invention will not elaborate on this.

[0052] Step S2: Using the generalized triangular cap method and the least squares method, fuse the original grid data of terrestrial water storage change from multiple sources to obtain the fused grid data of terrestrial water storage change.

[0053] This step S2 can improve the data accuracy of the grid data of terrestrial water storage change and serve as the basis for data processing in subsequent steps.

[0054] Regarding the generalized triangular cap method, at least refer to the prior art 2: Evaluating the uncertainty of water storage change in the **** region by GRACE inversion using the generalized triangular cap method. Yao Chaolong, Li Qiong, Luo Zhicai. Chinese Journal of Geophysics, 2019. 62(3): 883-897; DOI: 10.6038 / cjg2019L0454

[0055] Specifically, a feasible example of fusing the original grid data of terrestrial water storage change from multiple sources using the generalized triangular cap method and the least squares method is as follows:

[0056] (1) Obtain the original grid data of terrestrial water storage change from multiple sources with a time series nature from five different GRACE products.

[0057] (2) Calculate the covariance matrix C , for the five original grid data of terrestrial water storage change x 1, x 2, x 3, x 4, x 5, calculate the covariance between each pair of them C ij , where i, j = 1, 2, 3, 4, 5, and i ≠ j , and construct a 5×5 covariance matrix C , where the diagonal elements of the covariance matrix C C ii are the variances of the respective original grid data of terrestrial water storage change.

[0058] (3) Introduce a 5×5 noise covariance matrix R , here it is assumed that the noise is independent and identically distributed, the noise covariance matrix R is a diagonal matrix, and the diagonal elements of the noise covariance matrix R r iiIndicates the noise variance of the i th original terrestrial water storage change grid data, where i= 1, 2, 3, 4, 5.

[0059] (4)Construct the noise covariance matrix R The relationship with the covariance matrix C Using the objective function f and the corresponding constraint function g to solve the objective function f to obtain the elements of the noise covariance matrix R , where the diagonal elements of the noise covariance matrix R are r ii , used as a measure of noise variance and uncertainty, where i= 1, 2, 3, 4, 5.

[0060] (5)According to the diagonal elements R of the noise covariance matrix r ii calculate the fusion weight w i , the fusion weight w i is inversely proportional to the noise variance, and the specific calculation formula of the fusion weight w i is , where i= 1, 2, 3, 4, 5.

[0061] (6)The fusion model in the present invention is , where y is the fused result, x i is the i th original terrestrial water storage change grid data, w i is the i th fusion weight, is the error term. In the present invention, the single error , the sum of squared errors , that is, the least squares method is used here, where i= 1, 2, 3, 4, 5.

[0062] (7)Take the partial derivative of the sum of squared errors S with respect to the fusion weight w i , and set the partial derivative equal to zero to obtain a set of constraint equations. Solving this set of constraint equations can obtain the optimal weight S that minimizes the sum of squared errors , where i=1, 2, 3, 4, 5。

[0063] (8) Substitute the obtained optimal weights into the fusion model , and the fused terrestrial water storage change grid data with time series attributes can be obtained , where i= 1, 2, 3, 4, 5。

[0064] Step S3: For any grid point in the study area, after excluding each hydro-meteorological element in the first hydro-meteorological grid data one by one, input it into the LSTM prediction model, and obtain the simulated terrestrial water storage change grid data through the residual correction method. Retain the hydro-meteorological elements whose influence on the accuracy of the simulated terrestrial water storage change grid data is greater than the preset threshold as the key hydro-meteorological elements.

[0065] In the present invention, for any grid point, all the key hydro-meteorological elements constitute the hydro-meteorological element combination of the any grid point.

[0066] Exemplarily, in the present invention, if there are 8 hydro-meteorological elements in the hydro-meteorological grid data, after excluding one hydro-meteorological element each time, there are still 7 hydro-meteorological elements left. For example, after excluding precipitation once, the hydro-meteorological elements still include: evapotranspiration, snow water equivalent, snow depth, vegetation canopy water, runoff, surface temperature, and soil moisture data. And the next time evapotranspiration is excluded, the hydro-meteorological elements still include: precipitation, snow water equivalent, snow depth, vegetation canopy water, runoff, surface temperature, and soil moisture data.

[0067] Since after excluding the hydro-meteorological elements in the hydro-meteorological grid data one by one and obtaining the simulated terrestrial water storage change grid data through LSTM and the residual correction method, if a certain hydro-meteorological element is a key hydro-meteorological element of the grid point and is excluded during a certain process, the data accuracy between the obtained simulated terrestrial water storage change grid data and the previously obtained simulated terrestrial water storage change grid data will exceed the preset threshold; while if a certain hydro-meteorological element is not a key hydro-meteorological element of the grid point and is excluded during a certain process, the data accuracy between the obtained simulated terrestrial water storage change grid data and the previously obtained simulated terrestrial water storage change grid data will not exceed the preset threshold. Therefore, in this way, it can be identified whether it is a key hydro-meteorological element of the grid point. In other words, if a certain meteorological element is a key hydro-meteorological element of the grid point and is excluded during a certain process, the accuracy of the obtained simulated terrestrial water storage change grid data will have a relatively significant fluctuation, thereby determining that the excluded hydro-meteorological element belongs to the key hydro-meteorological element.

[0068] Exemplarily, the accuracy fluctuation value can be obtained by comparing with the simulated terrestrial water storage change grid data before excluding a certain hydro-meteorological element. Further, the accuracy fluctuation value can be measured by the root mean square error.

[0069] In addition, the specific implementation manners of the above residual correction method and the LSTM prediction model will be given later in the present invention.

[0070] Step S4: Divide the grid points with exactly the same key hydro-meteorological elements into a sub-region; for each sub-region, train and obtain an optimal LSTM prediction model based on the LSTM prediction model and the residual correction method, and use the optimal LSTM prediction model to obtain the terrestrial water storage change grid data for each sub-region, and form a subset of the first terrestrial water storage change grid data.

[0071] In other words, by combining the terrestrial water storage change grid data of all sub-regions obtained according to the optimal LSTM prediction model, the first terrestrial water storage change grid data can be obtained. Compared with the original terrestrial water storage change grid data from multiple sources, the first terrestrial water storage change grid data is high-resolution terrestrial water storage change grid data.

[0072] In the present invention, for different sub-regions, the corresponding and applied optimal LSTM prediction models are different.

[0073] Further, in step S4, the fused terrestrial water storage change grid data for 12 consecutive months can be uniformly processed to obtain the downscaled first terrestrial water storage change grid data.

[0074] The inventor found that, similar to the common practice in the prior art, if only directly using the obtained first terrestrial water storage change grid data as the finally downscaled data, the influence of extreme points generated by mutation anomalies and non-linear changes in the terrestrial water storage change grid data on the spatial downscaling accuracy will also be ignored.

[0075] Further, step S5: For each sub-region and the months containing extreme points in the fused terrestrial water storage change grid data, use the optimal LSTM prediction model to obtain the terrestrial water storage change grid correction data for the month, and use the terrestrial water storage change grid correction data for the month to replace the first terrestrial water storage change grid data for the month to obtain the final terrestrial water storage change grid data.

[0076] The final terrestrial water storage change grid data is high-spatial-resolution terrestrial water storage change grid data, which is the processing result ultimately desired in the present invention.

[0077] In the present invention, the specific implementation method of the residual improvement method is as follows:

[0078] (1) Use the first hydro-meteorological grid data and the fused terrestrial water storage change grid data as the input of the LSTM prediction model to obtain the predicted terrestrial water storage change grid data.

[0079] (2) Subtract the predicted terrestrial water storage change grid data from the fused terrestrial water storage change grid data to obtain the first residual grid data, and then use the Kriging interpolation method to interpolate the first residual grid data into the second residual grid data.

[0080] In comparison, the spatial resolutions of the second residual grid data and the predicted terrestrial water storage change grid data in the present invention both belong to low-resolution grid data, and the second residual grid data belongs to high-resolution grid data.

[0081] (3) Use the second hydro-meteorological grid data as the input of the LSTM prediction model to obtain the second terrestrial water storage change grid data, and add the second terrestrial water storage change grid data to the second residual grid data to obtain the target terrestrial water storage change grid data.

[0082] In the present invention, a feasible embodiment of the Kriging interpolation method is as follows:

[0083] (1) Determine the longitude coordinates and latitude coordinates of the known low-spatial-resolution terrestrial water storage change grid data within the study area.

[0084] (2) Calculate the variogram to describe the spatial variability of the regionalized variable, where the variogram can be expressed as:

[0085] ;

[0086] where the variogram value represents the number of sample pairs with a spacing of h , N ([[]] h ) represents the number of sample pairs with a spacing of h , and are the observed values at the known location points and the location point respectively, i is the serial number value.

[0087] (3) Determine the interpolation points, set as a regular grid of 0.1°×0.1°, where "°" represents the unit in the coordinate system composed of longitude and latitude.

[0088] (4)For each interpolation point, based on the spatial positions of the known points and the variogram , establish the Kriging equations. Specifically, the Kriging equations are:

[0089]

[0090] where is the weight coefficient of the j th known position point , is the Lagrange multiplier, and are both known position points, is the interpolation point, and are respectively the variogram values between the known position point and the known position point , and the variogram value between the known position point and the interpolation point . By solving the above Kriging equations, the weight coefficient of any known position point is obtained, where i , j are both serial numbers, n is a positive integer.

[0091] (5)According to the calculated weight coefficients of each known position point, perform weighted summation on the observed values of the known position points to obtain the estimated value of the interpolation point. The interpolation formula is: , where is the estimated value of the interpolation point , is the observed value of the j th known position point , is the j th known position point 's weight coefficient.

[0092] In addition, as a specific embodiment, the implementation scheme of the LSTM prediction model in the present invention can be realized by the following method:

[0093] (1)Determine the high-spatial-resolution, monthly-scale hydrometeorological grid data such as precipitation, evapotranspiration, snow water equivalent, snow depth, vegetation canopy water, runoff, surface temperature, and soil moisture in the study area, and standardize them using the Z-score function. The standardization formula is , where x is the hydrometeorological grid data, is the mean value, is the standard deviation, It is the standardized hydro-meteorological grid data. The standardized hydro-meteorological grid data is used as the input of the LSTM prediction model, and the grid data of the change in terrestrial water storage at a high resolution and monthly scale is used as the output of the LSTM prediction model.

[0094] (2)Taking an example area, the LSTM prediction model in the present invention includes 1 input layer, 3 hidden layers and 1 output layer, and there are 128 neurons (or cells) in the hidden layer.

[0095] (3)In the network layer of the LSTM prediction model, the learning mechanism is controlled by a forget gate, an input gate, a memory unit and an output gate.

[0096] In the present invention, the forget gate determines how much information in the cell state vector t- at the previous time step C t-1 needs to be forgotten. The formula is: , where f t is the activation value, is the Sigmoid activation function, W f is the weight matrix of the forget gate, h t-1 represents the hidden state vector at the previous time step t- 1, x t represents the input vector at time step t , b f is the bias vector of the forget gate, h t-1 , x t means concatenating h t-1 and x t into a vector, t -1 and t are both time steps.

[0097] In the present invention, the input gate is used to update the decision and create candidate values. The update decision formula is , where i t represents the update degree of each cell state element, is the Sigmoid activation function, W i is the weight matrix of the input gate, h t-1 represents the hidden state vector at the previous time step t- 1, xt denotes the input vector at time step t , b i is the bias vector of the input gate, h t-1 , x t represents the vector formed by concatenating h t-1 and x t , where i is the sequence number value, t -1 and t are both time steps.

[0098] Furthermore, the formula for creating candidate values in the present invention is: , where the current cell state vector C t refers to the cell state vector at time step t , tanh is the hyperbolic tangent activation function, W C is the weight matrix for creating candidate values, h t-1 represents the hidden state vector at the previous time step t- 1, x t denotes the input vector at time step t , b C is the corresponding bias vector, h t-1 , x t represents the vector formed by concatenating h t-1 and x t . t -1 and t are both time steps. The memory unit updates the current cell state vector C t according to the outputs of the input gate and the forget gate, and the update formula is , where denotes element-wise multiplication, t -1 and t are both time steps, f t is the activation value, i t represents the update degree of each cell state element, C t-1 is the cell state vector at the previous time step t -1.

[0099] In the present invention, the output gate determines how much information in the current cell state vector C t needs to be output to the hidden state output vector at time step t , including the output decision vector and the hidden state output vector. The formula for the output decision is h t , where is the output decision vector, is the Sigmoid activation function, is the weight matrix of the output gate, W o represents the hidden state vector at the previous time step h t-1 1, t- is the input vector at time step x t , t is the bias vector of the output gate, b o is h t-1 , x t represents the vector formed by concatenating h t-1 and x t , t -1 and t are both time steps; in addition, the formula for the hidden state output vector is , where h t is the hidden state output vector, is the output decision vector, represents element-wise multiplication, tanh is the hyperbolic tangent activation function, C t is the current cell state vector.

[0100] (4) Train the LSTM prediction model with the normalized hydro-meteorological grid data of the input, and then use the second hydro-meteorological grid data with high spatial resolution as the input of the LSTM prediction model, and output the grid prediction data of the change in terrestrial water storage. More specific implementation methods and details of the LSTM prediction model itself are common knowledge in the art and will not be elaborated in the present invention.

[0101] In addition, the present invention also discloses a storage medium storing a computer program / instructions, which when executed or compiled and executed by a processor, implements the steps of the method for processing grid data of the change in terrestrial water storage as described in any one of the above.

[0102] ​

[0103] Table 1: Comparison of Related Indicators between the Present Invention and Other Solutions

[0104]

[0105] Table 1 shows the comparison of related indicators between an actual embodiment of the present invention and other solutions. Table 1 as a whole reflects the accuracy of different downscaling results. Among them, the closer the correlation coefficient is to 1, the better; the closer the Nash coefficient is to 1, the better; the closer the consistency coefficient is to 1, the better; the smaller the root mean square error, the better; and the smaller the mean absolute error, the better, where mm is the unit of millimeter. It is worth mentioning that the present invention has also been compared with the solution based only on LSTM.

[0106] Compared with the LSTM solution, the correlation coefficient, Nash coefficient, consistency coefficient, root mean square error, and mean absolute error of the present invention have increased by 1.01%, 4.04%, 3.09%, 67.75%, and 71.23% respectively. Compared with the gradient boosting solution, the correlation coefficient, Nash coefficient, consistency coefficient, root mean square error, and mean absolute error of the present invention have increased by 51.51%, 74.75%, 156.41%, 91.26%, and 91.90% respectively. And compared with the random forest solution, the correlation coefficient, Nash coefficient, consistency coefficient, root mean square error, and mean absolute error of the present invention have increased by 2.04%, 19.28%, 12.36%, 81.95%, and 83.52% respectively.

[0107] In summary, the present invention not only comprehensively leads in multiple indicators compared with the traditional solutions, but also achieves a relatively high improvement compared with the solution that only relies on LSTM, especially in terms of the root mean square error and mean absolute error. In other words, compared with the solution that only relies on LSTM, the sub-region partitioning method and the method for correcting the data of relevant months where extreme points are located of the present invention have effectively improved the prediction accuracy of different spatial positions and the accuracy of downscaling results, making the present invention overall superior to other solutions.

[0108] To better illustrate the present invention, numerous specific details are given in the above specific implementation manners. Those skilled in the art should understand that the present invention can also be implemented without some specific details. In some instances, methods, means, elements, and circuits well-known to those skilled in the art are not described in detail to highlight the gist of the present invention.

[0109] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, and all should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claimed rights.

Claims

1. A method for processing grid data of terrestrial water storage change, characterized in that Including: Step S1: Obtain the first hydrometeorological grid data, the second hydrometeorological grid data, and the original terrestrial water storage change grid data from multiple sources within the study area; Step S2: Use the generalized triangular cap method and the least squares method to fuse the original terrestrial water storage change grid data from multiple sources to obtain the fused terrestrial water storage change grid data; Step S3: For any grid point in the study area, after excluding each hydrometeorological element in the first hydrometeorological grid data one by one, input it into the LSTM prediction model, and obtain the simulated terrestrial water storage change grid data through the residual correction method, and retain the hydrometeorological elements whose influence on the accuracy of the simulated terrestrial water storage change grid data is greater than the preset threshold as the key hydrometeorological elements; Step S4: Divide the grid points with exactly the same key hydrometeorological elements into a sub-region; for each sub-region, train and obtain the optimal LSTM prediction model based on the LSTM prediction model and the residual correction method, and use the optimal LSTM prediction model to obtain the terrestrial water storage change grid data for each sub-region, and form a subset of the first terrestrial water storage change grid data; Step S5: For each sub-region and the months containing extreme points in the fused terrestrial water storage change grid data, use the optimal LSTM prediction model to obtain the terrestrial water storage change grid correction data for the month, and use the terrestrial water storage change grid correction data for the month to replace the first terrestrial water storage change grid data for the month to obtain the final terrestrial water storage change grid data; The residual correction method includes the following steps: After subtracting the terrestrial water storage change grid prediction data output by the LSTM prediction model from the fused terrestrial water storage change grid data, interpolate to obtain the second residual grid data as the correction data; Use the first hydrometeorological grid data and the fused terrestrial water storage change grid data as the input of the LSTM prediction model to obtain the terrestrial water storage change grid prediction data; Subtract the terrestrial water storage change grid prediction data from the fused terrestrial water storage change grid data to obtain the first residual grid data, and then use the Kriging interpolation method to interpolate the first residual grid data into the second residual grid data; Use the second hydrometeorological grid data as the input of the LSTM prediction model to obtain the second terrestrial water storage change grid data, and add the second terrestrial water storage change grid data to the second residual grid data to obtain the target terrestrial water storage change grid data.

2. The method for processing the terrestrial water storage change grid data according to claim 1, wherein: The spatial resolution of the first hydrometeorological grid data is less than the spatial resolution of the second hydrometeorological grid data.

3. The method for processing the terrestrial water storage change grid data according to claim 2, wherein: The spatial resolution of the original terrestrial water storage change grid data from multiple sources is lower than the spatial resolution of the final terrestrial water storage change grid data and lower than the spatial resolution of the first terrestrial water storage change grid data.

4. The method for processing grid data of terrestrial water storage change according to claim 3, wherein: When obtaining the fused grid data of terrestrial water storage change, the optimal weight for the fused data is obtained by the least squares method.

5. The method for processing grid data of terrestrial water storage change according to claim 4, wherein: For any hydrometeorological grid data input into the LSTM prediction model, standardization processing is first performed.

6. The method for processing grid data of terrestrial water storage change according to claim 5, wherein: The neurons in the LSTM prediction model use the Sigmoid activation function.

7. The method for processing grid data of terrestrial water storage change according to claim 6, wherein: For different sub-regions, the applied optimal LSTM prediction models are different.

8. A storage medium storing a computer program / instructions, characterized in that: When the computer program / instructions are executed or compiled and executed by a processor, the steps of the method for processing grid data of terrestrial water storage change according to any one of claims 1-7 are implemented.

Images