High-spatial-resolution underground water reserve anomaly simulation method and system

By averaging GRACE data and integrating multi-source auxiliary variable data sets, combining machine learning models and recursive feature elimination, the problem of low spatial resolution of GRACE data is solved, and the monitoring of high-spatial resolution groundwater reserve anomalies is achieved, improving the improvement of the monitoring of high-spatial resolution groundwater reserves The systematic and generalizable study was carried out.

CN118940607APending Publication Date: 2024-11-12HOHAI UNIV

Patent Information

Application Number
CN202410927556.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-11
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

When using GRACE satellite data to monitor groundwater reserve anomalies, the spatial resolution is low, making it difficult to capture the dynamic changes of groundwater at small scales. The model construction has strong regional characteristics and lacks systematic and generalizable technical methods. .

Method used

By averaging GRACE data from multiple institutions, integrating multi-source auxiliary variable data sets, building a high-spatial resolution auxiliary variable data set with a unified time scale, and filtering advantageous variables with machine learning models and recursive feature elimination (RFE) Combined, establish a downscale model of land water reserves, and use the interpretable machine learning algorithm SHAP to analyze model performance and variable interactions to calculate high-spatial resolution groundwater reserve anomaly data.

Benefits of technology

It improves the application ability of GRACE data in regional-scale groundwater research, reduces the uncertainty brought about by single-institution products, enhances the generalizability of the method, and provides a more refined and interpretable analysis of changes in groundwater reserves.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118940607A_ABST
    Figure CN118940607A_ABST
Patent Text Reader

Abstract

The invention discloses a high-spatial-resolution underground water reserve anomaly simulation method and system, and the method comprises the steps: carrying out the averaging processing of GRACE data, integrating a multi-source auxiliary variable data set, and constructing a high-spatial-resolution auxiliary variable data set and a low-spatial-resolution auxiliary variable data set of a unified time scale; based on the low-spatial-resolution auxiliary variable data set and the averaged GRACE data, training a machine learning model, and coupling a recursive feature elimination method to obtain a land water reserve downscaling model and an auxiliary variable subset; based on the auxiliary variable subset, inputting corresponding auxiliary variables in the high-spatial-resolution auxiliary variable data set into a land water reserve downscaling model to obtain downscaled high-spatial-resolution groundwater reserve abnormal data; and according to the downscaled high-spatial-resolution land water reserve abnormal data and the surface water reserve component in the global daily value meteorological data set, calculating to obtain high-spatial-resolution groundwater reserve abnormal data.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a method and system for simulating groundwater reserve anomaly with high spatial resolution, and belongs to the cross-technical fields of groundwater hydrology, remote sensing science, computer science, etc. Background Art

[0002] Terrestrial Water Storage (TWS) refers to the sum of all forms of surface and groundwater, including ice and snow, soil water, groundwater, surface water and vegetation water. Terrestrial water storage anomaly data TWSA refers to the difference between the instantaneous value and the baseline value of terrestrial water storage TWS, which is a key variable in the study of water cycle.

[0003] As an outstanding tool for monitoring the Earth's gravity field, the GRACE satellite provides valuable scientific data in the field of groundwater monitoring. However, the limitations of its spatial resolution have, to a certain extent, restricted its application in groundwater research. Due to the relatively low spatial resolution of GRACE, it is difficult to capture small-scale groundwater dynamic changes, which poses an important challenge in certain studies that require detailed monitoring of local areas. This means that GRACE's monitoring capabilities are subject to certain limitations in study areas at the hundred-kilometer level and smaller. To overcome this limitation, researchers began to explore and apply data downscaling methods to extract more information about small-scale groundwater changes from GRACE data.

[0004] Data downscaling is a statistical or modeling method used to convert observational data at a larger spatial scale into information at a smaller scale. In the context of GRACE satellite groundwater monitoring, downscaling methods can help better understand and explain the gravity changes observed by GRACE, and then infer small-scale groundwater storage changes. Statistical downscaling methods usually rely on time series analysis and statistical models, mainly by establishing mathematical statistical relationships between ground observation data, climate model data and GRACE data, and after verifying the rationality of this relationship, it is applied to GRACE data to achieve the downscaling process. The emergence of machine learning technology has greatly promoted the development of statistical downscaling technology, giving researchers more effective tools to process complex earth science data, mainly because machine learning methods can establish nonlinear relationships between input parameters and targets. Using machine learning models such as random forests, XGBoost, and support vector machines, the complex nonlinear relationship between high spatial resolution data such as ground observations or climate models and low-resolution GRACE data can be effectively captured. Combining interpretable machine learning algorithms to explain the physical mechanism behind machine learning models, analyzing the contribution of each variable to model performance and the interaction between variables is a current hot topic. The interpretable machine learning algorithm SHAP (Shapley Additive Explanations) was used to explain the downscaling model trained on data with multiple auxiliary variables and reveal the main factors affecting water storage changes. Its main goal is to provide explanations for the feature importance of the model and promote the understanding of the decision-making process behind the model prediction.

[0005] Although there are more and more studies on downscaling of GRACE data, the model construction in most regions has strong regional characteristics and lacks systematic and generalizable technical methods. Therefore, it is necessary to explore a high spatial resolution groundwater storage anomaly simulation method and system to provide new research ideas for the application of GRACE data in regional-scale groundwater storage. Summary of the invention

[0006] In view of at least one of the above technical problems, the present invention proposes a high spatial resolution groundwater reserve anomaly simulation method and system, which is suitable for high spatial resolution groundwater reserve anomaly simulation in various regions.

[0007] Technical solution: To solve the above technical problems, the technical solution adopted by the present invention is:

[0008] In a first aspect, a method for simulating groundwater storage anomalies with high spatial resolution comprises:

[0009] Step 1, averaging the GRACE data to obtain averaged GRACE data; wherein the GRACE data includes original GRACE mascon products from multiple institutions;

[0010] Step 2: Integrate multi-source auxiliary variable datasets and construct high spatial resolution auxiliary variable datasets and low spatial resolution auxiliary variable datasets of the same time scale;

[0011] Step 3: Based on the low spatial resolution auxiliary variable data set and the averaged GRACE data, the machine learning model is trained and coupled with the recursive feature elimination method (RFE) to obtain the terrestrial water storage downscaling model and the auxiliary variable subset;

[0012] Step 4: Based on the auxiliary variable subset, the corresponding auxiliary variables in the high spatial resolution auxiliary variable data set are input into the terrestrial water storage downscaling model to obtain downscaled high spatial resolution groundwater storage anomaly data;

[0013] Step 5: Calculate the high spatial resolution groundwater storage anomaly data based on the downscaled high spatial resolution terrestrial water storage anomaly data and the surface water storage component in the global daily meteorological dataset GLDAS to obtain the high spatial resolution groundwater storage anomaly data.

[0014] In some embodiments, in step 1, averaging the GRACE data includes:

[0015] The GRACE data comes from the original data of the three main GRACE mascon publishing institutions: Jet Propulsion Laboratory JPL (Jet Propulsion Laboratory), Center for Space Research CSR (Center for Space Research) of the University of Texas, and Goddard Space Flight Center GSFC (Goddard Space Flight Center). The GRACE data are averaged and unified with a spatial resolution of 0.5°×0.5°.

[0016] In some embodiments, the step 2, integrating multi-source auxiliary variable datasets and constructing a high spatial resolution auxiliary variable dataset and a low spatial resolution auxiliary variable dataset of a unified time scale, comprises:

[0017] According to the integrated multi-source auxiliary variable dataset, the lowest resolution among the auxiliary variables is used as the benchmark to unify them as a high spatial resolution auxiliary variable dataset;

[0018] The high spatial resolution auxiliary variable dataset is resampled and the spatial resolution is unified to be consistent with the GRACE data after averaging to obtain a low spatial resolution auxiliary variable dataset.

[0019] In some embodiments, the multi-source auxiliary variable data set is selected from the Palmer Drought Index, actual evapotranspiration, potential evapotranspiration, accumulated rainfall, downward shortwave radiation from the surface, maximum and minimum temperatures, wind speed, and vapor pressure among the 14 hydrological and meteorological variables in the TerraClimate data set, soil moisture, plant canopy water, and surface runoff surface water storage anomaly components from the global daily meteorological data set GLDAS, and digital elevation model DEM data and spatial coordinate information from NOAA.

[0020] In some embodiments, the spatial resolution of the high spatial resolution auxiliary variable dataset is 4 km×4 km, and the spatial resolution of the low spatial resolution auxiliary variable dataset is 0.5°×0.5°.

[0021] In some embodiments, the auxiliary variable subset includes Palmer Drought Index (PDSI), actual evapotranspiration (AET), vapor pressure (VAP), surface runoff (RO), surface downward shortwave radiation (SARD), wind speed (VS), vapor pressure deficit (VPD), soil moisture (SOIL), accumulated rainfall (PR), potential evapotranspiration (PET).

[0022] In some embodiments, the step 3, based on the low spatial resolution auxiliary variable data set and the averaged GRACE data, trains the machine learning model and couples the recursive feature elimination method RFE to obtain the terrestrial water storage downscaling model and the auxiliary variable subset, including:

[0023] The samples in the low spatial resolution auxiliary variable dataset are used as the input of the machine learning model, and the averaged GRACE data are used as the output of the machine learning model. The machine learning model is trained, and the recursive feature elimination method (RFE) is coupled during the training process of the machine learning model to screen the dominant variable combination from the low spatial resolution auxiliary variable dataset to obtain the terrestrial water storage downscaling model and the input auxiliary variable subset.

[0024] Furthermore, in some embodiments, the recursive feature elimination method (RFE) is coupled during the machine learning model training process to screen the dominant variable combination from the low spatial resolution auxiliary variable data set to obtain the terrestrial water storage downscaling model and the input auxiliary variable subset, including:

[0025] All auxiliary variables in the low spatial resolution auxiliary variable data set are added to the candidate feature set; a machine learning model is constructed by training the candidate feature set, and the performance of the machine learning model is evaluated;

[0026] Use a machine learning model to evaluate the importance of each feature in the candidate feature set, remove the feature with the lowest importance from the candidate feature set, and update the candidate feature set;

[0027] Iteratively perform the following loop steps until the number of features in the candidate feature set reaches a preset minimum value; wherein the loop steps include: using the updated candidate feature set to train and reconstruct the machine learning model, and performing performance evaluation on the reconstructed machine learning model; using the reconstructed machine learning model to evaluate the importance of each feature in the candidate feature set; deleting the feature with the lowest importance from the candidate feature set, and updating the candidate feature set;

[0028] According to the performance evaluation results of all constructed and reconstructed machine learning models, the machine learning model with the best performance evaluation result is used as the terrestrial water storage downscaling model; the candidate feature set used to train the machine learning model with the best performance evaluation result is used as the auxiliary variable subset.

[0029] In some embodiments, the machine learning model is selected from at least one of a random forest (RF) model, a distributed gradient boosting library XGBoost (XGB) model, and a lightGBM (LGBM) model.

[0030] In some embodiments, the step 5, calculating the high spatial resolution groundwater reserve anomaly data based on the downscaled high spatial resolution groundwater reserve anomaly data and the surface water reserve component in the global daily meteorological dataset GLDAS, comprises:

[0031] The surface water storage components in the Global Daily Meteorological Dataset (GLDAS) include soil moisture, plant canopy water, surface runoff, and melted snow water;

[0032] High spatial resolution groundwater storage anomaly data are calculated based on high spatial resolution groundwater storage anomaly data, soil moisture, plant canopy water, surface runoff and melted snow water;

[0033]

[0034] Among them, GWSA represents high spatial resolution groundwater storage anomaly data, and TWSA represents high spatial resolution terrestrial water storage anomaly data. Represents soil moisture, Represents plant canopy water, represents surface runoff, Represents melted snow water.

[0035] In some embodiments, the method further comprises: step 6, using a machine learning model interpretation algorithm SHAP to analyze the contribution of each auxiliary variable to the performance of the terrestrial water storage downscaling model and the interaction between the variables;

[0036] The mathematical expression of SHAP value is:

[0037]

[0038] Among them, S i For sample x i Output of the terrestrial water storage downscaling model; i represents the i-th sample, and j represents the j-th feature of each sample; represents the average output of all samples, which serves as the benchmark for the evaluation of terrestrial water storage downscaling models; n represents the total number of auxiliary variables.

[0039] In a second aspect, a high spatial resolution groundwater reserve anomaly simulation system is provided, comprising:

[0040] A GRACE data processing module, used for performing averaging processing on the GRACE data to obtain averaged GRACE data; wherein the GRACE data includes original GRACE mascon products from multiple institutions;

[0041] Auxiliary variable data processing module, used to integrate multi-source auxiliary variable data sets and construct high spatial resolution auxiliary variable data sets and low spatial resolution auxiliary variable data sets of unified time scale;

[0042] A dominant variable screening module is used to train a machine learning model based on the low spatial resolution auxiliary variable data set and the averaged GRACE data and couple the recursive feature elimination method RFE to obtain a terrestrial water storage downscaling model and an auxiliary variable subset;

[0043] A terrestrial water storage downscaling module is used to input the corresponding auxiliary variables in the high spatial resolution auxiliary variable data set into the terrestrial water storage downscaling model based on the auxiliary variable subset to obtain downscaled high spatial resolution groundwater storage anomaly data;

[0044] The groundwater storage downscaling module is used to calculate the high spatial resolution groundwater storage anomaly data based on the downscaled high spatial resolution groundwater storage anomaly data and the surface water storage component in the global daily meteorological dataset GLDAS.

[0045] Beneficial effects: The present invention averages the GRACE mascon products from multiple institutions, reducing the uncertainty of a single institution's products due to different processing methods. It combines multiple machine learning models and couples the RFE algorithm to screen advantageous auxiliary variable combinations, establishes a statistical relationship between GRACE data and multi-source auxiliary variable data sets, and solves the problem of poor generalizability of methods caused by differences in auxiliary variables in different study areas in existing studies. It also combines interpretable machine learning algorithms to explain the physical mechanisms behind machine learning models, analyze the contribution of each variable to model performance and the interaction between variables, and provides new research ideas for GRACE groundwater research at the regional scale. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0047] Figure 1 The figure is a flow chart of a method according to an embodiment of the present invention.

[0048] Figure 2 It is a performance graph of multiple machine learning models in one embodiment of the present invention and a TWSA trend spatial distribution graph of the best model before and after downscaling.

[0049] Figure 3 It is a correlation coefficient diagram of groundwater level and groundwater reserve changes before and after downscaling corresponding to a diving station and a pressurized water station in one embodiment of the present invention, as well as a spatial location distribution diagram of unrelated stations.

[0050] Figure 4 This is a spatiotemporal distribution diagram of GWSA in the study area from April 2002 to December 2022 in one embodiment of the present invention.

[0051] Figure 5 This is a SHAP result diagram of the terrestrial water storage downscaling model in one embodiment of the present invention. DETAILED DESCRIPTION

[0052] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0053] Embodiment 1: A method for simulating groundwater reserves anomaly with high spatial resolution comprises:

[0054] Step 1, averaging the GRACE data to obtain averaged GRACE data; wherein the GRACE data includes original GRACE mascon products from multiple institutions;

[0055] In this embodiment, the GRACE data comes from the original data of the three main GRACE mascon publishing institutions, namely, the Jet Propulsion Laboratory JPL (Jet Propulsion Laboratory), the Center for Space Research CSR (Center for Space Research) of the University of Texas, and the Goddard Space Flight Center GSFC (Goddard Space Flight Center). The GRACE data are averaged and the spatial resolution is unified to 0.5°×0.5°.

[0056] Specifically, the spatial resolution of the CSR product is 0.25°×0.25°, and the spatial resolution of the JPL and GSFC products is 0.5°×0.5°. Through the resampling method, the spatial resolution is unified to 0.5°×0.5°, and then the average values ​​of the three product data sets are calculated to obtain the averaged GRACE data, that is, a low spatial resolution terrestrial water storage data set is formed.

[0057] Step 2: Integrate multi-source auxiliary variable datasets and construct high spatial resolution auxiliary variable datasets and low spatial resolution auxiliary variable datasets of the same time scale.

[0058] In some embodiments, the step 2, integrating multi-source auxiliary variable datasets and constructing a high spatial resolution auxiliary variable dataset and a low spatial resolution auxiliary variable dataset of a unified time scale, comprises:

[0059] According to the integrated multi-source auxiliary variable dataset, the lowest resolution among the auxiliary variables is used as the benchmark to unify them as a high spatial resolution auxiliary variable dataset;

[0060] The high spatial resolution auxiliary variable dataset is resampled and the spatial resolution is unified to be consistent with the GRACE data after averaging to obtain a low spatial resolution auxiliary variable dataset.

[0061] In this embodiment, the multi-source auxiliary variable data set is selected from the Palmer Drought Index, actual evapotranspiration, potential evapotranspiration, accumulated rainfall, downward shortwave radiation from the surface, maximum and minimum temperatures, wind speed, and vapor pressure among the 14 hydrological and meteorological variables in the TerraClimate data set, soil moisture, plant canopy water, and surface runoff surface water storage anomaly components from the global daily meteorological data set GLDAS, and digital elevation model DEM data and spatial coordinate information from NOAA.

[0062] In some embodiments, the spatial resolution of the high spatial resolution auxiliary variable dataset is 4 km×4 km, and the spatial resolution of the low spatial resolution auxiliary variable dataset is 0.5°×0.5°.

[0063] Step 3: Based on the low spatial resolution auxiliary variable data set and the averaged GRACE data, the machine learning model is trained and coupled with the recursive feature elimination method (RFE) to obtain the terrestrial water storage downscaling model and the auxiliary variable subset;

[0064] In some embodiments, step 3 includes: using samples in the low spatial resolution auxiliary variable data set as the input of the machine learning model, using the averaged GRACE data as the output of the machine learning model, training the machine learning model, and coupling the recursive feature elimination method RFE during the machine learning model training process to screen the dominant variable combination from the low spatial resolution auxiliary variable data set to obtain the terrestrial water storage downscaling model and the input auxiliary variable subset.

[0065] Furthermore, the recursive feature elimination (RFE) method is coupled to the machine learning model training process to screen the dominant variable combination from the low spatial resolution auxiliary variable data set to obtain the terrestrial water storage downscaling model and the input auxiliary variable subset, including:

[0066] All auxiliary variables in the low spatial resolution auxiliary variable data set are added to the candidate feature set; a machine learning model is constructed by training the candidate feature set, and the performance of the machine learning model is evaluated;

[0067] Use a machine learning model to evaluate the importance of each feature in the candidate feature set, remove the feature with the lowest importance from the candidate feature set, and update the candidate feature set;

[0068] Iteratively perform the following loop steps until the number of features in the candidate feature set reaches a preset minimum value; wherein the loop steps include: using the updated candidate feature set to train and reconstruct the machine learning model, and performing performance evaluation on the reconstructed machine learning model; using the reconstructed machine learning model to evaluate the importance of each feature in the candidate feature set; deleting the feature with the lowest importance from the candidate feature set, and updating the candidate feature set;

[0069] According to the performance evaluation results of all constructed and reconstructed machine learning models, the machine learning model with the best performance evaluation result is used as the terrestrial water storage downscaling model; the candidate feature set used to train the machine learning model with the best performance evaluation result is used as the auxiliary variable subset.

[0070] In some embodiments, multiple machine learning models can be selected to perform model training simultaneously, construct a statistical relationship between terrestrial water storage data and multi-source auxiliary variable data, couple the RFE algorithm in the model to screen the dominant variable combination, and obtain the auxiliary variable subset for the best model input. In this embodiment, the machine learning model is selected from at least one of a random forest (RF) model, a distributed gradient boost library XGBoost (XGB) model, and a lightGBM (LGBM) model.

[0071] The downscaling performance of the model was evaluated using three statistical indicators: correlation coefficient (R), normalized root mean square error (NRMSE) and coefficient value (R 2 ). To elaborate, the correlation coefficient is used to analyze the association between two data sets, while NRMSE is used to quantify the difference between two variables. In addition, R 2 Represents an overall measure of the goodness of fit of the model. The mathematical formulas for these indicators are shown below.

[0072]

[0073]

[0074]

[0075] where y i TWSA represents the output of the model, o i Represents the original TWSA, represents the averaged TWSA, o max and min They represent the maximum and minimum values ​​of the original TWSA respectively, and n represents the total number of samples.

[0076] R ranges from -1 to 1. R close to 1 indicates that the model prediction results are positively correlated with the observed results, and R close to -1 indicates that the prediction results are negatively correlated with the observed results. The value of can help judge the quality of the model. Its value range is from 0 to 1. The closer the value is to 0, the worse the model fitting effect is, and the closer the value is to 1, the better the model fitting effect is. NRMSE represents the difference between the observed value and the predicted value. The closer the NRMSE is to 0, the better the fit between the predicted value and the observed value is.

[0077] In some embodiments, the auxiliary variable subset includes Palmer Drought Index (PDSI), actual evapotranspiration (AET), vapor pressure (VAP), surface runoff (RO), surface downward shortwave radiation (SARD), wind speed (VS), vapor pressure deficit (VPD), soil moisture (SOIL), accumulated rainfall (PR), and potential evapotranspiration (PET).

[0078] Step 4: Based on the auxiliary variable subset, the corresponding auxiliary variables in the high spatial resolution auxiliary variable data set are input into the terrestrial water storage downscaling model to obtain downscaled high spatial resolution groundwater storage anomaly data;

[0079] In this embodiment, a variety of machine learning algorithms are used to construct a terrestrial water storage downscaling model. According to the best auxiliary variable subset screened out, the corresponding auxiliary variables in the high spatial resolution auxiliary variable data set are used as model inputs, and the corresponding samples are input into the optimal machine learning model established in the previous step to obtain the high spatial resolution GRACE data set, thereby completing the spatial downscaling of the GRACE data.

[0080] Step 5: Calculate the high spatial resolution groundwater storage anomaly data based on the downscaled high spatial resolution terrestrial water storage anomaly data and the surface water storage component in the global daily meteorological dataset GLDAS to obtain the high spatial resolution groundwater storage anomaly data.

[0081] In some embodiments, in this step, the surface water storage component in the downscaled high spatial resolution groundwater storage anomaly data is removed by using the global daily meteorological data set GLDAS to obtain high spatial resolution groundwater storage anomaly data. Auxiliary data from the GLDAS-Noah model are used, including soil moisture, plant canopy water, surface runoff and melted snow water. The monthly water storage anomalies of these variables are obtained by calculating their respective differences from the average value in the corresponding time period. The calculation method can be expressed as:

[0082]

[0083] Among them, GWSA represents high spatial resolution groundwater storage anomaly data, and TWSA represents high spatial resolution terrestrial water storage anomaly data. Represents soil moisture, Represents plant canopy water, represents surface runoff, Represents melted snow water.

[0084] In some embodiments, the method may further include step 6, using an interpretable machine learning algorithm SHAP to analyze the contribution of each variable to the model performance and the interaction between the variables.

[0085] Specifically, in this step, SHAP is used to interpret the downscaling model trained on data with multiple auxiliary variables to determine the main controlling factors affecting the changes in TWSA of high spatial resolution groundwater storage anomaly data. Its main purpose is to explain the feature importance of the model and facilitate the understanding of the decision-making process behind the model prediction. The basic idea of ​​SHAP comes from the Shapley value in cooperative game theory, which is used to measure the contribution of each feature to the model output, thereby determining the degree of influence of each feature. Compared with the feature importance indicators generated by machine learning models, SHAP provides a personalized explanation of feature importance. It calculates the contribution of auxiliary variables to the prediction of each specific sample instead of relying on a global average. The main advantage of SHAP values ​​is that they can clarify the impact of features on each sample and can identify whether it is a positive or negative impact. The mathematical expression of the SHAP value is as follows:

[0086]

[0087] Among them, S i For sample x i Output of the terrestrial water storage downscaling model; i represents the i-th sample, j represents the j-th feature of each sample; S i Represents the sample x i The output of the model is: represents the average output of all samples as the benchmark for model evaluation; n represents the total number of auxiliary variables. For example, is the impact of the first feature. If it is a positive value, it means that the feature has a positive impact on the output of the model, otherwise, it has a negative impact on the output.

[0088] Application example: This example selects a certain area as the study area to illustrate the specific implementation process of the present invention: The study area is located in eastern China, which is a typical coastal plain area, covering an area of ​​about 6,490 square kilometers, with a mild and humid climate. The area has a flat terrain, an altitude of 1 to 7 meters, and belongs to a subtropical monsoon climate. Precipitation varies significantly within the year and between years, and the peak rainfall occurs in June to September. The phreatic aquifer in the area is mainly composed of Holocene marine sediments, alluvial sediments and lake sediments, including silt, clay and sandy silt, with a thickness usually between 0.5 and 5 meters, and the depth of groundwater varies between 1.0 and 2.6 meters. The area mainly relies on the vertical infiltration of precipitation to replenish the groundwater system. The flow rate of groundwater in the aquifer is slow, and evaporation and artificial mining are the main ways of groundwater discharge. Confined water aquifers are mainly divided into three layers. There are 67 groundwater monitoring stations, including 36 confined water monitoring stations and the rest are phreatic water monitoring stations. The monthly groundwater level time series data from 2018 to 2021 are used to verify the downscaled groundwater storage. Generally speaking, the change of groundwater level should be combined with the water supply degree to estimate the change of groundwater storage. However, due to the challenges of accurately quantifying the actual distribution of water supply degree in aquifers, no such conversion is performed for qualitative research in this case. Due to the coarse monthly time resolution, there may be a time lag effect between the changes in groundwater levels. Therefore, a lagged correlation analysis is performed, and the correlation coefficient corresponding to the maximum correlation moment is used as the final result. Considering the problem of sample number when training the model, the data within the provincial administrative region where the study area is located are selected as the input sample.

[0089] The GRACE data are averaged using the original data products from three institutions: JPL (Jet Propulsion Laboratory), CSR (Center for Space Research), and GSFC (Goddard Space Flight Center). The multi-source auxiliary variable dataset uses 14 hydrological and meteorological variables from the TerraClimate dataset, DEM data from NOAA, and spatial coordinate information. The original spatial resolution of the data is 4 km × 4 km, which is consistent with the spatial resolution of the GRACE dataset after resampling, which is 0.5° × 0.5°.

[0090] Figure 2 It is a performance graph of multiple machine learning models and a spatial distribution graph of TWSA trends of the best model before and after downscaling. Three common machine learning models, Random Forest (RF), XGBoost (XGB) and lightGBM (LGBM), were selected for model construction. Figure 2(a) shows the changes in NRMSE of the three models as the number of selected features increases. It can be seen that with the addition of auxiliary variables, the performance of the three models is gradually improving, and the best model is the state of RF when 10 auxiliary variables are input. The auxiliary variable subset is Palmer Drought Index (PDSI), actual evapotranspiration (AET), vapor pressure (VAP), surface runoff (RO), surface downward shortwave radiation (SARD), wind speed (VS), vapor pressure deficit (VPD), soil moisture (SOIL), accumulated rainfall (PR), potential evapotranspiration (PET).

[0091] Figure 2 (b), (c), and (d) show the scatter plots of the observed and simulated values ​​of the three models at their respective optimal parameters. It can be seen that RF is better than the other two models in terms of the three performance parameters. Therefore, the optimal parameter model of RF is used as the basis of the downscaling model in the future. Figure 2 (e) and (f) in the middle show the changing trends of terrestrial water storage in the provincial administrative region where the study area is located from April 2002 to December 2022 before and after downscaling, respectively. It can be seen that the results after downscaling can achieve a detailed characterization of the original data.

[0092] Figure 3 The following is the correlation coefficient diagram of groundwater level and groundwater storage changes before and after downscaling for the phreatic station and the confined water station, as well as the spatial location distribution diagram of unrelated stations. Generally speaking, when the correlation coefficient of two variables is less than 0.3, it can be considered that there is no correlation between them. Figure 3 As can be seen in (a) and (b), the correlation coefficients between groundwater level and groundwater storage at most sites have been improved after downscaling, including 45.16% of phreatic water sites and 69.44% of confined water sites. For phreatic water sites, the correlation coefficient increased by an average of 15.94%, and six sites transitioned from an uncorrelated state to a weakly correlated state. The correlation coefficients of confined water sites increased by an average of 22.86%, and the correlation coefficients of four sites transitioned from uncorrelated to weakly correlated. Figure 3 (c) shows the spatial distribution of unrelated sites. This situation may be due to the following two reasons: first, the change in groundwater reserves at the site is mainly affected by the change in water volume in other aquifers. For example, in the northwestern mountainous area, all phreatic water sites are in an unrelated state, while confined water sites are in a related state. Second, the water exchange between aquifers may cause a significant change in the water level of a certain aquifer, but the overall water reserves have not changed significantly, resulting in the lack of correlation between groundwater level and groundwater reserves.

[0093] Figure 4This is the spatiotemporal distribution map of GWSA in the study area from April 2002 to December 2022. Figure 4 In (a) and (b), we can see that the trend of groundwater storage changes before and after downscaling is well consistent, and the results after downscaling can more clearly show the spatial differences in groundwater storage changes. Figure 4 In the middle (c), we can see that due to the impact of groundwater exploitation activities, groundwater reserves showed a downward trend from 2002 to 2011. After groundwater exploitation stopped, groundwater reserves began to recover year by year.

[0094] Figure 5 This is the SHAP result of the downscaling model, showing the SHAP values ​​of 10 selected features, among which PDSI (Palmer Drought Index), AET (Actual Evapotranspiration) and VAP (Vapor Pressure) are the most influential factors. PDSI is a meteorological indicator that measures the drought or humidity of a region by comparing the current hydrological conditions with the long-term average. PDSI is positively correlated with TWSA. AET and VAP represent actual evapotranspiration and vapor pressure, respectively. In a humid environment, an increase in temperature will increase evaporation and vapor pressure, while in a cold climate, vapor pressure decreases. Figure 5 (d) shows the SHAP dependence plot between AET and VAP, where it can be observed that the two variables are indeed positively correlated. However, the scatter plot shows that the samples have a vertical offset, which is the result of the effects caused by the interaction between multiple variables. This effect is particularly evident in the remaining auxiliary variables. Abundant precipitation (PR) intensifies runoff (RO) and soil moisture (SOIL), increasing TWS. However, TWS is affected by multiple factors. Vapor pressure deficit (VPD) is negatively correlated with TWS due to increased air dryness and enhanced evaporation. This interaction highlights the complexity of water balance. Potential evapotranspiration (PET) contrasts with actual evapotranspiration (AET) in its effects on TWS, which is due to the historical overexploitation of groundwater in the study area. The decline in groundwater levels enhances soil evaporation capacity, resulting in a negative correlation between PET and TWS. Downward shortwave radiation (SRAD) and wind speed (VS) are generally positively correlated with TWSA. However, some results deviate from this trend, and higher wind speeds may indicate precipitation events, increasing TWS, while in the case of lower wind speeds, they intensify evaporation and reduce TWS. Likewise, high SRAD values ​​indicate increased solar radiation, which promotes evaporation and reduces TWS. Figure 5 (e) is the interaction of SHAP values, reflecting the significant degree of characteristic interaction. These relationships are affected by natural processes and human activities and are crucial for comprehensive hydrological assessment.

[0095] Embodiment 2: Based on Embodiment 1, a high spatial resolution groundwater reserve anomaly simulation system is provided, comprising:

[0096] A GRACE data processing module, used for performing averaging processing on the GRACE data to obtain averaged GRACE data; wherein the GRACE data includes original GRACE mascon products from multiple institutions;

[0097] Auxiliary variable data processing module, used to integrate multi-source auxiliary variable data sets and construct high spatial resolution auxiliary variable data sets and low spatial resolution auxiliary variable data sets of unified time scale;

[0098] A dominant variable screening module is used to train a machine learning model based on the low spatial resolution auxiliary variable data set and the averaged GRACE data and couple the recursive feature elimination method RFE to obtain a terrestrial water storage downscaling model and an auxiliary variable subset;

[0099] A terrestrial water storage downscaling module is used to input the corresponding auxiliary variables in the high spatial resolution auxiliary variable data set into the terrestrial water storage downscaling model based on the auxiliary variable subset to obtain downscaled high spatial resolution groundwater storage anomaly data;

[0100] The groundwater storage downscaling module is used to calculate the high spatial resolution groundwater storage anomaly data based on the downscaled high spatial resolution groundwater storage anomaly data and the surface water storage component in the global daily meteorological dataset GLDAS.

[0101] In some embodiments, the method further includes: an influencing factor diagnosis module for analyzing the contribution of each auxiliary variable to the performance of the terrestrial water storage downscaling model and the interaction between variables using the machine learning model interpretation algorithm SHAP, so as to further determine the main controlling factors affecting the change of water storage.

[0102] This paper proposes a new high spatial resolution groundwater reserve anomaly simulation method and system by coupling the RFE algorithm with multiple machine learning models and combining the explainable machine learning algorithm SHAP method. Compared with the existing methods, the method is more generalizable, and the implementation process is simple and easy to understand, with strong applicability, which provides a new research idea for GRACE groundwater research at the regional scale.

[0103] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0104] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0105] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0106] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0107] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are included in the scope of the claims of the present invention to be approved.

Claims

1. A method for simulating groundwater reserves anomaly with high spatial resolution, characterized in that: The method comprises: Performing averaging processing on the GRACE data to obtain averaged GRACE data; wherein the GRACE data includes original GRACE mascon products from multiple institutions; Integrate multi-source auxiliary variable datasets and construct high spatial resolution auxiliary variable datasets and low spatial resolution auxiliary variable datasets of the same time scale; Based on the low spatial resolution auxiliary variable data set and the averaged GRACE data, a machine learning model is trained and coupled with a recursive feature elimination method to obtain a terrestrial water storage downscaling model and an auxiliary variable subset; Based on the auxiliary variable subset, the corresponding auxiliary variables in the high spatial resolution auxiliary variable data set are input into the terrestrial water storage downscaling model to obtain downscaled high spatial resolution groundwater storage anomaly data; Based on the downscaled high spatial resolution terrestrial water storage anomaly data and the surface water storage component in the global daily meteorological dataset, high spatial resolution groundwater storage anomaly data are calculated.

2. The method according to claim 1, characterized in that: Averaging of GRACE data, including: The GRACE data are raw data from three major GRACE mascon publishing institutions: Jet Propulsion Laboratory JPL, University of Texas Space Research Center CSR, and Goddard Space Flight Center GSFC. The GRACE data are averaged and unified with a spatial resolution of 0.5°×0.5°.

3. The method according to claim 1, characterized in that: Integrate multi-source auxiliary variable datasets and construct high spatial resolution auxiliary variable datasets and low spatial resolution auxiliary variable datasets of the same time scale, including: According to the integrated multi-source auxiliary variable dataset, the lowest resolution among the auxiliary variables is used as the benchmark to unify them as a high spatial resolution auxiliary variable dataset; The high spatial resolution auxiliary variable dataset is resampled and the spatial resolution is unified to be consistent with the GRACE data after averaging to obtain a low spatial resolution auxiliary variable dataset.

4. The method according to claim 1, characterized in that: The multi-source auxiliary variable dataset is selected from the 14 hydrological and meteorological variables in the TerraClimate dataset, including the Palmer Drought Index, actual evapotranspiration, potential evapotranspiration, accumulated rainfall, downward shortwave radiation from the surface, maximum and minimum temperature, wind speed, and vapor pressure; the soil moisture, plant canopy water, and surface runoff surface water storage anomaly components from the global daily meteorological dataset GLDAS; and the digital elevation model DEM data and spatial coordinate information from NOAA; and / or, the spatial resolution of the high spatial resolution auxiliary variable dataset is 4 km × 4 km, and the spatial resolution of the low spatial resolution auxiliary variable dataset is 0.5° × 0.5°; and / or, the auxiliary variable subset includes Palmer Drought Index, actual evapotranspiration, vapor pressure, surface runoff, surface downward shortwave radiation, wind speed, vapor pressure deficit, soil moisture, accumulated rainfall, potential evapotranspiration.

5. The method according to claim 1, characterized in that Based on the low spatial resolution auxiliary variable data set and the averaged GRACE data, the machine learning model is trained and coupled with the recursive feature elimination method to obtain the terrestrial water storage downscaling model and the auxiliary variable subset, including: The samples in the low spatial resolution auxiliary variable dataset are used as the input of the machine learning model, and the averaged GRACE data are used as the output of the machine learning model. The machine learning model is trained, and the recursive feature elimination method (RFE) is coupled during the training process of the machine learning model to screen the dominant variable combination from the low spatial resolution auxiliary variable dataset to obtain the terrestrial water storage downscaling model and the input auxiliary variable subset.

6. The method according to claim 5, characterized in that In the process of machine learning model training, the recursive feature elimination method (RFE) is coupled to screen the dominant variable combination from the low spatial resolution auxiliary variable data set to obtain the terrestrial water storage downscaling model and the input auxiliary variable subset, including: All auxiliary variables in the low spatial resolution auxiliary variable data set are added to the candidate feature set; a machine learning model is constructed by training the candidate feature set, and the performance of the machine learning model is evaluated; Use a machine learning model to evaluate the importance of each feature in the candidate feature set, remove the feature with the lowest importance from the candidate feature set, and update the candidate feature set; Iteratively perform the following loop steps until the number of features in the candidate feature set reaches a preset minimum value; wherein the loop steps include: using the updated candidate feature set to train and reconstruct the machine learning model, and performing performance evaluation on the reconstructed machine learning model; using the reconstructed machine learning model to evaluate the importance of each feature in the candidate feature set; deleting the feature with the lowest importance from the candidate feature set, and updating the candidate feature set; According to the performance evaluation results of all constructed and reconstructed machine learning models, the machine learning model with the best performance evaluation result is used as the terrestrial water storage downscaling model; the candidate feature set used to train the machine learning model with the best performance evaluation result is used as the auxiliary variable subset.

7. The method according to claim 1, characterized in that The machine learning model is selected from at least one of a random forest RF model, a distributed gradient boosting library (GGB) model and a LGBM model.

8. The method according to claim 1, characterized in that: Based on the downscaled high spatial resolution groundwater storage anomaly data and the surface water storage component in the global daily meteorological dataset GLDAS, high spatial resolution groundwater storage anomaly data are calculated, including: The surface water storage components in the Global Daily Meteorological Dataset (GLDAS) include soil moisture, plant canopy water, surface runoff, and melted snow water; High spatial resolution groundwater storage anomaly data are calculated based on high spatial resolution groundwater storage anomaly data, soil moisture, plant canopy water, surface runoff and melted snow water; , Among them, GWSA represents high spatial resolution groundwater storage anomaly data, and TWSA represents high spatial resolution terrestrial water storage anomaly data. Represents soil moisture, Represents plant canopy water, represents surface runoff, Represents melted snow water.

9. The method according to claim 1, characterized in that: It also includes: using the machine learning model interpretation algorithm SHAP to analyze the contribution of each auxiliary variable to the performance of the terrestrial water storage downscaling model and the interactions between variables; The mathematical expression of SHAP value is: , Among them, S i For sample x i Output of the terrestrial water storage downscaling model; i represents the i-th sample, and j represents the j-th feature of each sample; represents the average output of all samples, which serves as the benchmark for the evaluation of terrestrial water storage downscaling models; n represents the total number of auxiliary variables.

10. A high spatial resolution groundwater reserve anomaly simulation system, characterized in that: The system comprises: A GRACE data processing module, used for performing averaging processing on the GRACE data to obtain averaged GRACE data; wherein the GRACE data includes original GRACE mascon products from multiple institutions; Auxiliary variable data processing module, used to integrate multi-source auxiliary variable data sets and construct high spatial resolution auxiliary variable data sets and low spatial resolution auxiliary variable data sets of unified time scale; A dominant variable screening module is used to train a machine learning model based on the low spatial resolution auxiliary variable data set and the averaged GRACE data and couple the recursive feature elimination method to obtain a terrestrial water storage downscaling model and an auxiliary variable subset; A terrestrial water storage downscaling module is used to input the corresponding auxiliary variables in the high spatial resolution auxiliary variable data set into the terrestrial water storage downscaling model based on the auxiliary variable subset to obtain downscaled high spatial resolution groundwater storage anomaly data; The groundwater storage downscaling module is used to calculate the high spatial resolution groundwater storage anomaly data based on the downscaled high spatial resolution groundwater storage anomaly data and the surface water storage component in the global daily meteorological dataset.

Citation Information

Patent Citations

  • Forest penetration rainfall space map establishing method based on remote sensing and machine learning

    CN117908167A

  • Method for improving drought monitoring resolution based on random forest drought index

    CN118152929A

  • Continuous groundwater monitoring using machine learning

    US20240053508A1

Cited By

  • Underground water reserve inversion method and system based on multi-source data

    CN119740129A

  • Underground water reserve abnormity inversion product applicability evaluation method

    CN120181657A

  • A method for evaluating the applicability of groundwater storage anomaly inversion products

    CN120181657B

  • Oasis ecological hydrological simulation method, device and equipment and storage medium

    CN120874606A

  • GRACE data super-resolution network space downscaling method fusing geographic information and environment variables

    CN121564574A