Time downscaling method for improving abnormal time resolution of land water reserves

Abstract

The invention discloses a time downscaling method for improving the abnormal time resolution of land water reserves, and belongs to the cross technical field of hydrology, computer science and the like. According to the method, multiple driving variables which directly or indirectly influence land water reserve abnormity are collected to be used for driving an artificial intelligence model to construct a nonlinear relation between multiple variables and a GRACE product, then physical constraint and sliding window filtering are introduced to correct a model result, and monthly-scale GRACE data are improved to a higher time resolution. According to the method, the problem of low time resolution of a product with abnormal land water reserves inversed by GRACE is solved, physical constraints on the monthly scale are provided for the land water reserves after final time downscaling, so that data after time downscaling depicts finer daily scale details while keeping information of monthly scale data; and meanwhile, the filtering smoothing of the final sliding window effectively avoids the occurrence of extreme conditions at individual moments.

Description

Technical Field

[0001] The invention belongs to the cross-technical fields of hydrology, computer science and the like, and relates to a time downscaling method for improving the time resolution of terrestrial water storage anomalies. Background Art

[0002] Since its launch in 2002, the GRACE gravity satellite has developed into an indispensable tool for monitoring changes in water resources over large areas and long time series, providing a solid data foundation for global water cycle research and climate change, especially in quantifying water cycle components, monitoring groundwater, and monitoring extreme climate events such as droughts and floods. Therefore, after the first generation of GRACE mission ended in 2017, in order to continue to inherit the advantages of the GRACE gravity satellite mission, NASA and the German Space Research Center successfully launched its follow-up mission GRACE-FO in 2018.

[0003] However, due to the limitations of the satellite's own design, the acquired terrestrial water storage anomaly data (TWSA) faces multiple defects. First, the spatial resolution is relatively coarse, with a limited resolution of about 300 km; second, there are about 1 year of continuous missing values ​​between the GRACE mission and the GRACE-FO mission; third, the temporal resolution of TWSA is relatively coarse, on a monthly scale. These three limitations have brought certain bottlenecks to water resources-related research in many cases. In recent years, there have been many research inventions on improving spatial resolution and filling in the missing values ​​of the two missions, and they have achieved certain results; however, the research inventions on improving the temporal resolution of TWSA derived from GRACE are slightly insufficient and still need further research.

[0004] To address the limitation of GRACE's low temporal resolution, some research inventions have attempted to obtain more precise temporal water storage information from the original GRACE mission observation data (GRACE L1 B), which mainly include GPS data, accelerometer data, and K-wave ranging data. In 2009, the Kalman filter method was first used to try to decompose daily-scale data information from GRACE L1 B data. This method introduces prior information about the temporal behavior of the gravity field into the analysis process, so the temporal resolution can be enhanced without losing spatial resolution, but this method depends largely on the selected model assumptions and parameters. Of course, there are also some attempts to build statistical models. In 2019, the relationship between precipitation, temperature and terrestrial water storage was constructed, and the monthly-scale data information was decomposed into daily scales. This method has also been applied to the Haihe River Basin in China, and relatively ideal results have been achieved. However, the statistical relationship between terrestrial water storage anomalies and temperature and precipitation variables is not necessarily linear, and is easily affected by various factors. Obviously, this method also has certain limitations. Summary of the invention

[0005] In view of the problems in the prior art, the present invention provides a time downscaling method for improving the time resolution of terrestrial water storage anomalies.

[0006] A temporal downscaling method for improving the temporal resolution of terrestrial water storage anomalies includes the following steps: collecting a variety of driving variables that directly or indirectly affect terrestrial water storage anomalies, using them to drive an artificial intelligence model to construct a nonlinear relationship between multiple variables and GRACE products, and then introducing physical constraints and sliding window filtering to correct the model results, thereby improving the monthly scale GRACE data to a higher temporal resolution.

[0007] A time downscaling method for improving the temporal resolution of terrestrial water storage anomalies further comprises the following steps:

[0008] Step 1: Collect the dataset.

[0009] Step 2: Data preprocessing.

[0010] Step 3: Build an AI-based time downscaling model.

[0011] Step 4: Based on the daily-scale driving data obtained from the data preprocessing in the second step, the nonlinear time downscaling model on the monthly scale constructed in the third step is driven, and then a preliminary downscaling result with high temporal resolution is obtained.

[0012] Step 5: Reduce the preliminary time downscaling results to the time resolution before downscaling, and then calculate the deviation between the results after time downscaling and the results before time downscaling.

[0013] Step 6: Assign a reasonable weight to each fine time scale result.

[0014] Step 7: Multiply the specific gravity by the deviation to obtain the compensation value of the data at each fine time scale. The calculation steps are as follows:

[0015] b day =days·b month ·W day

[0016] Among them, days represents the number of days in the month, b day Indicates the compensation value of TWSA on a daily scale.

[0017] Step 8: Add the preliminary time downscaling result and the compensation value in step 7 to obtain the final time downscaling product.

[0018]

[0019] in, Represents the final time downscaling product after the preliminary time downscaling result is corrected based on the physical constraints.

[0020] Step nine: Smooth the final time downscaling result to obtain the final time downscaling smoothed result.

[0021] In recent years, the rise of artificial intelligence (AI) algorithms has brought about earth-shaking changes in various fields, especially in the construction of statistical relationships between multiple variables. However, the combination with GRACE temporal downscaling has yet to be explored. The present invention collects a variety of driving variables that directly or indirectly affect terrestrial water storage anomalies, which are used to drive the artificial intelligence model to construct a nonlinear relationship between multiple variables and GRACE products, and then introduces physical constraints and sliding window filtering to correct the model results, thereby improving the monthly GRACE data to a higher temporal resolution.

[0022] The advantages of the present invention are: it solves the problem of low temporal resolution of the terrestrial water storage anomaly product inverted by GRACE, and the nonlinear time downscaling method based on the physically constrained artificial intelligence algorithm constructed by the present invention provides physical constraints on the monthly scale for the terrestrial water storage anomaly after the final time downscaling, so that the data after time downscaling can depict finer daily scale details while retaining the information of the monthly scale data, and the final sliding window filtering and smoothing effectively avoids the occurrence of extreme situations at individual moments. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0024] Figure 1 , location distribution of example areas.

[0025] Figure 2 , time series of terrestrial water storage anomalies at the original monthly scale (region-01).

[0026] Figure 3 , Preliminary monthly-scale terrestrial water storage anomaly time series (region-01) obtained based on the time downscaling model.

[0027] Figure 4 , the final monthly-scale terrestrial water storage anomaly time series (region-01) based on the physically constrained temporal downscaling model.

[0028] Figure 5 , the final monthly scale terrestrial water storage anomaly time series after sliding window smoothing (region-01).

[0029] Figure 6 , Taylor diagram of performance indicators of monthly-scale terrestrial water storage anomaly after downscaling compared with that before downscaling (Region-01).

[0030] Figure 7 , Preliminary daily-scale terrestrial water storage anomaly time series (region-01) obtained based on the time downscaling model.

[0031] Figure 8 , the final daily-scale terrestrial water storage anomaly time series (region-01) based on the physically constrained temporal downscaling model.

[0032] Fig. 9 , The final daily-scale terrestrial water storage anomaly time series after sliding window smoothing (region-01).

[0033] Fig.10 , time series of terrestrial water storage anomalies at the original monthly scale (region-02).

[0034] Fig.11 , Preliminary monthly-scale terrestrial water storage anomaly time series (region-02) obtained based on the time downscaling model.

[0035] Fig.12 , the final monthly-scale terrestrial water storage anomaly time series (region-02) based on the physically constrained temporal downscaling model.

[0036] Fig.13 , the final monthly scale terrestrial water storage anomaly time series after sliding window smoothing (region-02).

[0037] Fig.14 , Taylor diagram of performance indicators of monthly-scale terrestrial water storage anomaly after downscaling compared with that before downscaling (Region-02).

[0038] Fig.15 , Preliminary daily-scale terrestrial water storage anomaly time series (region-02) obtained based on the time downscaling model.

[0039] Fig.16 , the final daily-scale terrestrial water storage anomaly time series (region-02) based on the physically constrained temporal downscaling model.

[0040] Fig.17 , the final daily-scale terrestrial water storage anomaly time series after sliding window smoothing (region-02).

[0041] Fig.18 , the time downscaling flow chart of the present invention. DETAILED DESCRIPTION

[0042] 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.

[0043] Example 1: Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 , Figure 7 , Figure 8 , Fig. 9 , Fig.10 , Fig.11 , Fig.12 , Fig.13 , Fig.14 , Fig.15 , Fig.16 , Fig.17 and Fig.18 As shown in the figure, a time downscaling method for improving the temporal resolution of terrestrial water storage anomalies is used. With the help of a high-performance artificial intelligence algorithm, the temporal resolution of the terrestrial water storage anomaly products inverted by GRACE is improved while ensuring the physical information before and after downscaling.

[0044] The implementation of the present invention provides a time downscaling method for improving the temporal resolution of terrestrial water storage anomalies. The method can simultaneously integrate the robustness of machine learning and the constraints of physical constraints, thereby achieving an improvement in the temporal resolution of GRACE inversion terrestrial water storage anomaly products.

[0045] A time downscaling method for improving the temporal resolution of terrestrial water storage anomalies comprises the following steps:

[0046] Step 1: Collect data sets. The collected variables need to have a certain correlation with the target variables, and these variables need to have a higher temporal resolution. For example, when the target variable is on a monthly scale, the collected data sets are on a daily or hourly scale with a higher temporal resolution. Rainfall, runoff, evapotranspiration, multi-layer soil water storage, snow water equivalent height, vegetation canopy water equivalent height and corresponding time variables (year, month, day) related to terrestrial water storage anomalies were collected mainly from the Global Land Data Assimilation System (GLDAS) project. It is worth noting that the collected GLDAS products are on a 3-hour scale.

[0047] Step 2: Data preprocessing: The collected driving data set is converted into two time scales. The first scale must be consistent with the time scale of the target variable. Here, since the time scale of the terrestrial water storage anomaly (TWSA) inverted by GRACE is monthly, the driving data set must also be converted to monthly scale. The second scale is the time scale of the final demand, that is, the target time scale of downscaling, and the monthly time series data is downscaled to daily scale.

[0048] Step 3: At the monthly time resolution, based on the artificial intelligence (AI) algorithm, an AI-based time downscaling model is constructed, where the input variables are: rainfall, runoff, evapotranspiration, multi-layer soil water storage, snow water equivalent water height, vegetation canopy water equivalent water height and the corresponding time variables (year, month, day); target variable: GRACE inverted land water storage anomaly, the formula is as follows:

[0049]

[0050] Among them, X month represents all the collected and preprocessed monthly driving variables, mainly rainfall, runoff, evaporation, soil water, etc. represents the monthly-scale TWSA product obtained based on the GRACE gravity satellite mission, F AI(*) indicates a nonlinear time downscaling model built based on AI algorithms (machine learning or deep learning algorithms).

[0051] Step 4: The relationship between the driving variable and the target variable at the monthly scale can be well captured by the time downscaling model based on the AI ​​algorithm, and is also applicable at higher time resolutions. The nonlinear time downscaling model at the monthly scale constructed in the third step is driven by the daily-scale driving data obtained based on the data preprocessing in the second step, and then the preliminary high-time resolution downscaling results (taking the daily scale as an example) are obtained.

[0052]

[0053] Among them, X day represents all the collected and preprocessed daily-scale driving variables, mainly rainfall, runoff, evapotranspiration, multi-layer soil water storage, snow water equivalent water height, vegetation canopy water equivalent water height and the corresponding time variables (year, month, day), F AI (*) represents the nonlinear time downscaling model built based on AI algorithm (machine learning or deep learning algorithm) (obtained through step 3), Represents the preliminary time downscaling product obtained by driving the nonlinear time downscaling model based on high temporal resolution driving variables.

[0054] Step 5: Since the model built based on the AI ​​algorithm is prone to underestimate high values ​​and overestimate low values, the phenomenon also exists in the nonlinear time downscaling model, which leads to inconsistency between the results after time downscaling and the results before time downscaling. Therefore, it is necessary to convert the preliminary time downscaling results to the time resolution before downscaling (i.e., monthly scale) and then find the deviation between the results after time downscaling and the results before time downscaling.

[0055]

[0056]

[0057] Among them, f agg (*) represents the average function, which is mainly based on the preliminary time downscaling results obtained based on the time downscaling model The monthly time resolution is calculated to be consistent with the GRACE inversion TWSA product, and the monthly time resolution is obtained after preliminary time downscaling. represents the original monthly-scale TWSA product inverted by GRACE, b month It represents the deviation between the original monthly-scale TWSA product of the original GRACE inversion and the monthly-scale TWSA after the preliminary time downscaling.

[0058] Step 6: After the deviation calculation is completed, the most important thing is how to reasonably distribute these deviations to each fine time scale, and impose a physical constraint on the results after the initial downscaling to ensure that the information before and after the time downscaling is consistent. The core idea is that the magnitude information before and after the time downscaling must be consistent, but at the same time, the relationship between the driving variable simulated by machine learning and the target variable must be guaranteed. To achieve this goal, it is necessary to assign a reasonable proportion to the results of each fine time scale, and the calculation of the proportion is as follows:

[0059]

[0060] in, represents the result of daily-scale TWSA regularization after preliminary time downscaling, W day It represents the proportion of the TWSA value at each daily spatial resolution to the monthly temporal resolution.

[0061] Step 7: Multiply the weight by the deviation to obtain the compensation value of the data on each fine time scale. The calculation steps are as follows:

[0062] b day =days·b month ·W day

[0063] Among them, days represents the number of days in the month, b day Indicates the compensation value of TWSA on a daily scale.

[0064] Step 8: Add the preliminary time downscaling result and the compensation value in step 7 to obtain the final time downscaling product.

[0065]

[0066] in, Represents the final time downscaling product after the preliminary time downscaling result is corrected based on the physical constraints.

[0067] Step 9: The final time downscaling result after physical constraints is easily affected by the sudden change value in a certain time period. In order to make the time downscaling result smoother, the final time downscaling result is smoothed at the end. The main idea is to use the average value of 5 days of data to represent the value of the middle day. Based on this idea, a smoothing filter with a sliding window of 5 days and a step size of 1 day is designed, and finally the final time downscaling smoothing result is obtained.

[0068] Example 2: Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 , Figure 7 , Figure 8 , Fig. 9 , Fig.10 , Fig.11 , Fig.12 , Fig.13 , Fig.14 , Fig.15 , Fig.16 , Fig.17 and Fig.18 As shown in the figure, a time downscaling method for improving the temporal resolution of terrestrial water storage anomaly (TWSA) is constructed for application in any region on the global scale. It is mainly used to target the low temporal resolution of terrestrial water storage anomaly inverted by the GRACE mission, and aims to obtain high temporal resolution products (daily scale) of terrestrial water storage anomaly TWSA, so as to improve the application potential of this product in drought or flood disaster monitoring.

[0069] The processing flow of a time downscaling method for improving the time resolution of terrestrial water storage anomalies provided by an example of the present invention is as follows: Fig.18 The specific implementation process of the embodiment of the present invention is as follows:

[0070] The first step is to collect data sets. The collected variables need to have a certain correlation with the target variables, and these variables need to have a higher temporal resolution. For example, when the target variable is on a monthly scale, the collected data sets are on a daily or hourly scale with a higher temporal resolution. The data sets mainly include rainfall, runoff, evapotranspiration, multi-layer soil water storage, snow water equivalent height, vegetation canopy water equivalent height and corresponding time variables (year, month, day) related to terrestrial water storage anomalies collected from the Global Land Data Assimilation System (GLDAS) project. It is worth noting that the collected GLDAS products are on a 3-hour scale.

[0071] Step 2: Data preprocessing: The collected driving data set is converted into two time scales. The first scale must be consistent with the time scale of the target variable. Here, since the time scale of the terrestrial water storage anomaly (TWSA) inverted by GRACE is monthly, the driving data set must also be converted to monthly scale. The second scale is the time scale of the final demand, that is, the target time scale of downscaling. Here, the monthly time series data is planned to be downscaled to daily scale.

[0072] Step 3: At the monthly time resolution, based on the artificial intelligence (AI) algorithm, an AI-based time downscaling model is constructed, where the input variables are: rainfall, runoff, evapotranspiration, multi-layer soil water storage, snow water equivalent water height, vegetation canopy water equivalent water height and the corresponding time variables (year, month, day); target variable: GRACE inverted land water storage anomaly, the formula is as follows:

[0073]

[0074] Among them, X month represents all the collected and preprocessed monthly driving variables, mainly rainfall, runoff, evaporation, soil water, etc. represents the monthly-scale TWSA product obtained based on the GRACE gravity satellite mission, F AI (*) indicates a nonlinear time downscaling model built based on AI algorithms (machine learning or deep learning algorithms).

[0075] Step 4: The relationship between the driving variable and the target variable at the monthly scale can be well captured by the time downscaling model based on the AI ​​algorithm, and is also applicable at higher time resolutions. Therefore, the nonlinear time downscaling model at the monthly scale constructed in the third step is driven by the daily-scale driving data obtained based on the data preprocessing in the second step, and then the preliminary high-time resolution downscaling results (taking the daily scale as an example) are obtained.

[0076]

[0077] Among them, X day represents all the collected and preprocessed daily-scale driving variables, which mainly refer to rainfall, runoff, evapotranspiration, multi-layer soil water storage, snow water equivalent water height, vegetation canopy water equivalent water height and corresponding time variables (year, month, day). AI (*) represents the nonlinear time downscaling model built based on AI algorithm (machine learning or deep learning algorithm) (obtained through step 3), Represents the preliminary time downscaling product obtained by driving the nonlinear time downscaling model based on high temporal resolution driving variables.

[0078] Step 5: Since the model built based on the AI ​​algorithm is prone to underestimate high values ​​and overestimate low values, this also exists in the nonlinear time downscaling model of the present invention, which leads to inconsistency between the results after time downscaling and the results before time downscaling. Therefore, it is necessary to convert the preliminary time downscaling results to the time resolution before downscaling (i.e., monthly scale) and then find the deviation between the results after time downscaling and the results before time downscaling.

[0079]

[0080]

[0081] Among them, f agg (*) represents the average function, which is mainly based on the preliminary time downscaling results obtained based on the time downscaling model The monthly time resolution is calculated to be consistent with the GRACE inversion TWSA product, and the monthly time resolution is obtained after preliminary time downscaling. represents the original monthly-scale TWSA product inverted by GRACE, b month It represents the deviation between the original monthly-scale TWSA product of the original GRACE inversion and the monthly-scale TWSA after the preliminary time downscaling.

[0082] Step 6: After the deviation calculation is completed, the most important thing is how to reasonably distribute these deviations to each fine time scale. Here, a physical constraint is imposed on the results after the initial downscaling to ensure that the information before and after the time downscaling is consistent. The core idea is that the magnitude information before and after the time downscaling must be consistent, but at the same time, the relationship between the driving variable simulated by machine learning and the target variable must be guaranteed. To achieve this goal, a reasonable proportion needs to be assigned to the results of each fine time scale. The calculation of this proportion is as follows:

[0083]

[0084] in, represents the result of daily-scale TWSA regularization after preliminary time downscaling, W day It represents the proportion of the TWSA value at each daily spatial resolution to the monthly temporal resolution.

[0085] Step 7: Multiply the weight by the deviation to obtain the compensation value of the data on each fine time scale. The calculation steps are as follows:

[0086] b day =days·b month ·W day

[0087] Among them, days represents the number of days in the month, b day Indicates the compensation value of TWSA on a daily scale.

[0088] Step 8: Add the preliminary time downscaling result and the compensation value in step 7 to obtain the final time downscaling product.

[0089]

[0090] in, Represents the final time downscaling product after the preliminary time downscaling result is corrected based on the physical constraints.

[0091] Step 9: The final time downscaling result after physical constraints is easily affected by the sudden change value of a certain time period. In order to make the time downscaling result smoother, the final time downscaling result is smoothed at the end. The idea is to use the average value of 5 days of data to represent the value of the middle day. Based on this idea, a smoothing filter with a sliding window of 5 days and a step size of 1 day is designed, and finally the final time downscaling smoothing result is obtained.

[0092] Step 10: Result analysis, such as Figure 1 Region-01 is the study area selected for this example. Figure 2 represents the TWSA time series of the region before temporal downscaling; Figure 3 , Figure 4 and Figure 5 The figures respectively represent the preliminary results, final results and the comparison between the final smoothed results and the original results based on the nonlinear time downscaling model at the monthly scale; Figure 6 It is the Taylor diagram of the performance index of the monthly land water storage anomaly after downscaling compared with that before downscaling; Figure 7 , Figure 8 and Fig. 9 The figure shows the comparison of the preliminary results, final results and final smoothed results based on the nonlinear time downscaling model at the daily scale. These results show that, whether on the daily scale or the monthly scale, there is a certain deviation between the preliminary results obtained based on the nonlinear time downscaling model and the original results, and this deviation has been greatly improved after adding physical constraints. The final smoothing based on sliding window filtering further optimizes the time downscaling results, so that the results retain the original monthly scale information while describing the daily scale change information with a more reasonable time distribution.

[0093] A time downscaling method to improve the temporal resolution of terrestrial water storage anomaly (TWSA) is constructed for application in any region on the global scale. It is mainly used to improve the temporal resolution of terrestrial water storage anomaly inverted by the GRACE mission. The purpose is to obtain high temporal resolution products (daily scale) of terrestrial water storage anomaly TWSA, so as to improve the application potential of this product in drought or flood disaster monitoring.

[0094] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A temporal downscaling method to improve the temporal resolution of terrestrial water storage anomalies, It is characterized in that The method includes the following steps: collecting a variety of driving variables that directly or indirectly affect terrestrial water storage anomalies, using them to drive the artificial intelligence model to build a nonlinear relationship between multiple variables and GRACE products, and then introducing physical constraints and sliding window filtering to correct the model results, thereby improving the monthly GRACE data to a higher temporal resolution.

2. A time downscaling method for improving the temporal resolution of terrestrial water storage anomalies according to claim 1, It is characterized in that It also contains the following steps: Step 1: Collect the dataset. Step 2: Data preprocessing, Step 3: Build an AI-based time downscaling model. Step 4: Based on the daily driving data obtained from the data preprocessing in the second step, the nonlinear time downscaling model on the monthly scale constructed in the third step is driven, and then a preliminary high-temporal-resolution downscaling result is obtained. Step 5: Reduce the preliminary time downscaling results to the time resolution before downscaling, and then find the deviation between the results after time downscaling and the results before time downscaling. Step 6: Assign a reasonable weight to each fine time scale result. Step 7: Multiply the specific gravity by the deviation to obtain the compensation value of the data at each fine time scale. The calculation steps are as follows: b day =days·b month ·W day Among them, days represents the number of days in the month, b day represents the compensation value of TWSA on a daily scale, Step 8: Add the preliminary time downscaling result and the compensation value in step 7 to obtain the final time downscaling product. in, represents the final time downscaling product after the preliminary time downscaling result is corrected based on physical constraints, Step nine: Smooth the final time downscaling result to obtain the final time downscaling smoothed result.

3. A time downscaling method for improving the temporal resolution of terrestrial water storage anomalies according to claim 2, It is characterized in that Collecting the data set also includes the following steps: collecting rainfall, runoff, evapotranspiration, multi-layer soil water storage, snow water equivalent height, vegetation canopy water equivalent height and the corresponding time variables year, month and day related to terrestrial water storage anomalies from the Global Land Data Assimilation System (GLDAS) project.

4. A time downscaling method for improving the temporal resolution of terrestrial water storage anomalies according to claim 2, It is characterized in that Data preprocessing,also contains the following steps: convert the collected driving data set into two time scales, the first scale must be consistent with the time scale of the target variable, and the driving data set is a monthly scale; The second scale is the time scale of final demand, that is, the target time scale of downscaling, which is to downscale the monthly time series data to the daily scale.

5. A time downscaling method for improving the temporal resolution of terrestrial water storage anomalies according to claim 2, It is characterized in that The construction of the AI-based time downscaling model also includes the following steps: at a monthly time resolution, based on the artificial intelligence (AI) algorithm, the AI-based time downscaling model is constructed, where the input variables are: rainfall, runoff, evapotranspiration, multi-layer soil water storage, snow water equivalent water height, vegetation canopy water equivalent water height and the corresponding time variables year, month and day; target variable: GRACE inverted land water storage anomaly, the formula is as follows: Among them, X month represents all the collected and preprocessed monthly driving variables, mainly rainfall, runoff, evaporation, multi-layer soil water storage, snow water equivalent water height, vegetation canopy water equivalent water height and the corresponding time variables year, month and day, represents the monthly-scale TWSA product obtained based on the GRACE gravity satellite mission, F AI (*) indicates a nonlinear time downscaling model built based on AI algorithms (machine learning or deep learning algorithms).

6. A time downscaling method for improving the temporal resolution of terrestrial water storage anomalies according to claim 2, It is characterized in that Based on the daily-scale driving data obtained by data preprocessing in step 2, the nonlinear time downscaling model on the monthly scale constructed in step 3 is driven to obtain preliminary downscaling results with high temporal resolution. Among them, X day represents all the collected and preprocessed daily-scale driving variables, mainly rainfall, runoff, evapotranspiration, multi-layer soil water storage, snow water equivalent water height, vegetation canopy water equivalent water height and the corresponding time variables year, month, day, F AI (*) represents the nonlinear time downscaling model built based on AI algorithm (machine learning or deep learning algorithm) (obtained through step 3), Represents the preliminary time downscaling product obtained by driving the nonlinear time downscaling model based on high temporal resolution driving variables.

7. A time downscaling method for improving the temporal resolution of terrestrial water storage anomalies according to claim 2, It is characterized in that Assigning a reasonable weight to each fine time scale result also includes the following steps: reducing the preliminary time downscaling results to the time resolution before downscaling, and then finding the deviation between the time downscaling results and the time resolution before downscaling, Since the model built based on AI algorithm is prone to underestimate the high value and overestimate the low value, the phenomenon also exists in nonlinear time downscaling model, which leads to inconsistency between the results after time downscaling and the results before time downscaling. It is necessary to reduce the preliminary time downscaling results to the time resolution before downscaling, that is, monthly scale, and then find the deviation between the results after time downscaling and the results before time downscaling. Among them, f agg (*) represents the average function, which is based on the preliminary time downscaling results obtained by the time downscaling model The monthly time resolution is calculated to be consistent with the GRACE inversion TWSA product, and the monthly time resolution is obtained after preliminary time downscaling. represents the original monthly-scale TWSA product inverted by GRACE, b month represents the deviation between the original monthly-scale TWSA product of the original GRACE inversion and the monthly-scale TWSA after the initial time downscaling, After the deviation calculation is completed, these deviations are reasonably distributed to each fine time scale, imposing a physical constraint on the results after the initial downscaling to ensure that the information before and after the time downscaling is consistent. The magnitude information before and after the time downscaling must be consistent, but at the same time, the relationship between the driving variable simulated by machine learning and the target variable is guaranteed. The proportion is calculated as follows: in, represents the result of daily-scale TWSA regularization after preliminary time downscaling, W day It represents the proportion of the TWSA value at each daily spatial resolution to the monthly temporal resolution.

8. A time downscaling method for improving the temporal resolution of terrestrial water storage anomalies according to claim 2, It is characterized in that The final time downscaling result is smoothed to obtain the final time downscaling smoothed result, which also includes the following steps: the final time downscaling result after physical constraints is easily affected by the sudden change value of a certain time period. In order to make the time downscaling result smoother, the final time downscaling result is smoothed at the end. The average value of 5 days of data is used to represent the value of the middle day. A smoothing filter with a sliding window of 5 days and a step size of 1 day is designed.

Images