A method for improving water storage resolution and accuracy based on spatial downscaling using machine learning

Abstract

The present invention discloses a method and system for improving water storage resolution and accuracy based on machine learning spatial downscaling, the method comprising: establishing an initial downscaling model between abnormal terrestrial water storage changes and multiple hydrological variables based on a modeling window; using N different machine learning algorithms to learn the initial downscaling model, and combining M different optimization selectors to screen out the optimal terrestrial water storage change abnormal solution model and the optimal modeling window; constructing the optimal downscaling model according to the optimal terrestrial water storage change solution function and the optimal modeling window; outputting the terrestrial water storage change abnormal value according to the optimal downscaling model. The present invention optimizes TWSA products with low spatial resolution (such as 1° and 0.5°). In order to obtain TWSA products with high spatial resolution (such as 0.25°, 0.1° and 0.05°), the water storage resolution and accuracy are improved.

Description

Technical Field

[0001] The present invention belongs to the cross-technical fields of satellite gravity, hydrology, etc., and in particular relates to a method for improving the resolution and accuracy of water reserves based on machine learning spatial downscaling. Background Art

[0002] Water resources are one of the indispensable materials for human survival and social development. Groundwater, as an important component of freshwater resources, is an important source of water for industrial production, agricultural irrigation and domestic use. Affected by factors such as climate change, over-exploitation of groundwater, and lack of effective management, groundwater in many areas has been seriously depleted. Therefore, accurately estimating the changes in regional groundwater reserves is of great significance for effectively managing water resources and ensuring food security. Traditional methods such as establishing a large-scale groundwater monitoring network require a lot of manpower and material resources. In addition, sparse and spatially unevenly distributed groundwater observation data make it difficult to continuously estimate the changes in groundwater reserves in large areas. Therefore, some scholars have begun to explore research methods that can analyze changes in groundwater reserves with higher temporal and spatial resolution, such as hydrological models and satellite remote sensing.

[0003] The GRACE satellite, launched successfully in 2002, provides a new way to continuously monitor the changes in terrestrial water storage over a large area. Early studies have shown that over large areas (~20,000 km 2 ), the GRACE satellite observation data can be used to estimate the change of terrestrial water storage with an accuracy of 1 to 1.5 cm in equivalent water column height. At present, GRACE products have important applications in studying global terrestrial water storage changes, drought and flood disasters, glacier melting, earthquakes and sea level changes. The terrestrial water storage inverted by GRACE data is the sum of hydrological variables such as soil water, groundwater, and snow water. According to the terrestrial water storage balance equation, the corresponding variables are deducted from the terrestrial water storage to obtain the change of regional groundwater storage (GWS). Some scholars have used GRACE observation data combined with hydrological models to carry out a series of research work in typical areas where terrestrial water and groundwater storage changes are drastic, such as the Amazon Basin, the Yangtze River Basin, northwest India, the North China Plain, and the Central Valley of California. However, due to the limitation of coarse spatial resolution, GRACE products cannot meet the analysis and research of terrestrial water storage changes in small areas. Therefore, it is of great research significance to obtain high-resolution GRACE grid products for the change of terrestrial water and groundwater storage in small areas. Summary of the invention

[0004] The technology of the present invention solves the problem: overcome the shortcomings of the prior art, provide a method for improving the resolution and accuracy of water reserves based on machine learning spatial downscaling, optimize the TWSA products with low spatial resolution (such as 1° and 0.5°), and obtain TWSA products with high spatial resolution (such as 0.25°, 0.1° and 0.05°), thereby improving the resolution and accuracy of water reserves.

[0005] In order to solve the above technical problems, the present invention discloses a method for improving water reserve resolution and accuracy based on machine learning spatial downscaling, comprising:

[0006] Based on the modeling window, an initial downscaling model between the anomaly of terrestrial water storage changes and multiple hydrological variables is established;

[0007] N different machine learning algorithms are used to learn the initial downscaling model, and combined with M different optimization selectors, the optimal ground water storage change anomaly solution model and the optimal modeling window are screened;

[0008] According to the optimal terrestrial water storage change solution function and the optimal modeling window, the optimal downscaling model is constructed;

[0009] According to the optimal downscaling model, the abnormal values ​​of terrestrial water storage changes are output.

[0010] In the above method of improving water storage resolution and accuracy based on machine learning spatial downscaling, the initial downscaling model is a low spatial resolution model, which is expressed as follows:

[0011]

[0012] Among them, TWSA i represents the low spatial resolution groundwater storage anomaly of the i-th month, i = 1, 2, 3, ...; v ij represents the hydrological variable of the jth type of low spatial resolution in the i-th month, j = 1, 2, 3, ...; w n represents the modeling window, n = 3, 5, 7, 9; f i Represents the low spatial resolution groundwater storage anomaly solution model for the ith month.

[0013] In the above-mentioned method for improving the resolution and accuracy of water storage based on machine learning spatial downscaling, w3, w5, w7 and w9 represent the modeling windows of 3×3, 5×5, 7×7 and 9×9 grids, respectively.

[0014] In the above-mentioned method of improving water storage resolution and accuracy by spatial downscaling based on machine learning, N different machine learning algorithms include: RF machine learning algorithm, ETR machine learning algorithm, ABR machine learning algorithm and GBR machine learning algorithm.

[0015] In the above-mentioned method for improving water storage resolution and accuracy based on machine learning spatial downscaling, M different optimization selectors include: four optimization selectors with root mean square error RMSE, mean absolute error MAE, Nash efficiency coefficient NSE and correlation coefficient CC as evaluation indicators.

[0016] In the above-mentioned method of improving water storage resolution and accuracy based on machine learning spatial downscaling, the optimal ground water storage change anomaly solution model f is screened and obtained through the following formulas (2) and (3): opt and the optimal modeling window w opt :

[0017]

[0018]

[0019] Among them, f RF 、f ETR 、f ABR and f GBR They represent the basic models using RF, ETR, ABR and GBR machine learning algorithms as f in formula (1);

[0020] The meanings of formulas (2) and (3) are as follows: RF 、f ETR 、f ABR and f GBR As the basic model of f in formula (1), and respectively substitute into modeling windows w3, w5, w7 and w9 of different sizes, the four evaluation indicators of RMSE, MAE, NSE and CC are used to evaluate the performance of different f models. RF 、f ETR 、f ABR and f GBR The combined output results of modeling windows w3, w5, w7 and w9 of different sizes are comprehensively evaluated to screen out the f with the best downscaling result accuracy. opt and w opt .

[0021] In the above-mentioned method of improving the resolution and accuracy of water reserves based on spatial downscaling of machine learning, the smaller the values ​​of RMSE and MAE, the higher the accuracy of the model; the NSE value ranges from negative infinity to 1, and the closer the NSE is to 1, the better the quality of the model and the higher the model credibility; the CC value ranges from -1 to 1, and the closer the CC is to 1, the better the quality of the model and the higher the model credibility.

[0022] In the above-mentioned method of improving water storage resolution and accuracy based on machine learning spatial downscaling, the optimal downscaling model is a high spatial resolution model, which is expressed as follows:

[0023]

[0024] in, Indicates the abnormal change of groundwater storage at high spatial resolution, Represents the j-th type of high spatial resolution hydrological variable.

[0025] In the above method of improving water storage resolution and accuracy based on machine learning spatial downscaling, it also includes: The groundwater storage change with high spatial resolution is obtained by solving the following formula (5):

[0026]

[0027] in, Represents soil water storage changes at high spatial resolution.

[0028] Accordingly, the present invention also discloses a system for improving water storage resolution and accuracy based on machine learning spatial downscaling, comprising:

[0029] The initial model building module is used to build an initial downscaling model between terrestrial water storage change anomalies and multiple hydrological variables based on the modeling window;

[0030] A screening module is used to learn the initial downscaling model using N different machine learning algorithms, and combine M different optimization selectors to screen out the optimal groundwater storage change anomaly solution model and the optimal modeling window;

[0031] The optimal model building module is used to build the optimal downscaling model according to the optimal terrestrial water storage change solution function and the optimal modeling window;

[0032] The result output module is used to output the abnormal value of terrestrial water storage change according to the optimal downscaling model.

[0033] The present invention has the following advantages:

[0034] The present invention discloses a method for improving the resolution and accuracy of water reserves based on spatial downscaling through machine learning. In the spatial dimension, a plurality of machine learning algorithms are combined with modeling windows of different sizes. The root mean square error, mean absolute error, Nash efficiency coefficient and correlation coefficient are used as indicators to analyze the consistency of the change of terrestrial water reserves before and after downscaling. The TWSA product with low spatial resolution is converted into a TWSA product with high spatial resolution, thereby improving the resolution and accuracy of water reserves and providing a reference for regional groundwater management. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1is a flowchart of the steps of a method for improving water storage resolution and accuracy based on machine learning spatial downscaling in an embodiment of the present invention;

[0036] Figure 2 is a schematic diagram of a modeling window of different sizes in an embodiment of the present invention;

[0037] Figure 3 is the location and digital elevation model of Guantao County;

[0038] Figure 4 It is the flow chart of downscaling data processing;

[0039] Figure 5 It is a schematic diagram of the results of RMSE, MAE, NSE and CC of different combined downscaling models;

[0040] Figure 6 Schematic diagram of the spatial distribution of the signal change trend before and after TWSA downscaling; where 6(a) represents the original signal of CSR-M06, 6(b) represents the signal of 6(a) resampled to 0.5°, 6(c) to 6(f) represent the results of downscaling the signal of 6(b) using four machine learning algorithms (without adding residuals), and 6(g) to 6(j) represent the downscaling results of the signals of 6(c) to (f) after adding residuals;

[0041] Figure 7 This is a statistical diagram of the accuracy of the signal spatial distribution between the 0.25° downscaling result and the CSR-M06 model;

[0042] Figure 8 It is a schematic diagram of the downscaling results of different modeling windows and the time series of the CSR-M06 model in Guantao County;

[0043] Fig. 9 It is a schematic diagram of the regression analysis of the downscaling results of different modeling windows and the time series of land water storage changes in Guantao County inverted by the CSR-M06 model; among them, 9(a) represents WS3, 9(b) represents WS5, 9(c) represents WS7, and 9(d) represents WS9;

[0044] Fig.10 is the spatial distribution of TWSA in Guantao County before and after downscaling; 10(a) represents before downscaling, and 10(b) represents after downscaling.

[0045] Fig.11 This is a schematic diagram of the TWSA time series in Guantao County before and after downscaling;

[0046] Fig.12 Schematic diagram of the spatial distribution of GWSA in Guantao County before and after downscaling; 12(a) represents before downscaling, and 12(b) represents after downscaling.

[0047] Fig.13 This is a schematic diagram of the time series of GWSA in Guantao County before and after downscaling;

[0048] Fig.14 This is a schematic diagram of the time series of inverted and measured groundwater changes in Guantao County. DETAILED DESCRIPTION

[0049] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments disclosed in the present invention will be further described in detail below with reference to the accompanying drawings.

[0050] The present invention discloses a method for improving the resolution and accuracy of water reserves based on spatial downscaling using machine learning. By establishing a downscaling model between low spatial resolution TWSA and hydrological variables in the spatial dimension and exploring the influence of different modeling windows on the downscaling accuracy, the TWSA products with low spatial resolution (such as 1° and 0.5°) are optimized to obtain high spatial resolution (such as 0.25°, 0.1° and 0.05°) water reserves. It should be noted that all variables with 'Λ' in the text are data with high spatial resolution.

[0051] like Figure 1 In this embodiment, the method for improving water storage resolution and accuracy based on machine learning spatial downscaling includes:

[0052] Step 101, based on the modeling window, an initial downscaling model between abnormal changes in terrestrial water storage and multiple hydrological variables is established.

[0053] In this embodiment, the initial downscaling model is a low spatial resolution model, and the expression is as follows:

[0054]

[0055] Among them, TWSA i represents the low spatial resolution groundwater storage anomaly of the i-th month, i = 1, 2, 3, ...; v ij represents the hydrological variable of the jth type of low spatial resolution in the i-th month, j = 1, 2, 3, ...; w n represents the modeling window, n = 3, 5, 7, 9; f i Represents the low spatial resolution groundwater storage anomaly solution model for the ith month.

[0056] Preferably, Figure 2 , w3, w5, w7, and w9 represent the modeling windows of 3×3, 5×5, 7×7, and 9×9 grids, respectively.

[0057] Step 102, adopting N different machine learning algorithms to learn the initial downscaling model, and combining M different optimization selectors to screen out the optimal ground water storage change anomaly solution model and the optimal modeling window.

[0058] In this embodiment, the N different machine learning algorithms may specifically refer to: RF machine learning algorithm, ETR machine learning algorithm, ABR machine learning algorithm and GBR machine learning algorithm. The M different optimization selectors may specifically refer to: four optimization selectors using root mean square error RMSE, mean absolute error MAE, Nash efficiency coefficient NSE and correlation coefficient CC as evaluation indicators.

[0059] Preferably, the optimal ground water storage abnormality solution model f can be obtained by screening through the following formulas (2) and (3): opt and the optimal modeling window w opt :

[0060]

[0061]

[0062] Among them, f RF 、f ETR 、f ABR and f GBR They represent the basic models using RF, ETR, ABR and GBR machine learning algorithms as f in formula (1).

[0063] It should be noted that the meanings of formulas (2) and (3) are as follows: f RF 、f ETR 、f ABR and f GBR As the basic model of f in formula (1), and respectively substitute into modeling windows w3, w5, w7 and w9 of different sizes, the four evaluation indicators of RMSE, MAE, NSE and CC are used to evaluate the performance of different f models. RF 、f ETR 、f ABR and f GBR The combined output results of modeling windows w3, w5, w7 and w9 of different sizes are comprehensively evaluated to screen out the f with the best downscaling result accuracy. opt and w opt .

[0064] Generally speaking, the smaller the values ​​of RMSE and MAE, the higher the accuracy of the model; the value of NSE ranges from negative infinity to 1, and the closer the NSE is to 1, the better the quality of the model and the higher the model credibility; the value of CC ranges from -1 to 1, and the closer the CC is to 1, the better the quality of the model and the higher the model credibility.

[0065] Step 103: construct an optimal downscaling model according to the optimal terrestrial water storage change solution function and the optimal modeling window.

[0066] In this embodiment, the optimal downscaling model is a high spatial resolution model, and the expression is as follows:

[0067]

[0068] in, Indicates the abnormal change of groundwater storage at high spatial resolution, Represents the j-th type of high spatial resolution hydrological variable.

[0069] Step 104: outputting the abnormal value of the change of terrestrial water storage according to the optimal downscaling model.

[0070] Step 105, according to The groundwater storage change with high spatial resolution is obtained by solving the following formula (5):

[0071]

[0072] in, Represents soil water storage changes at high spatial resolution.

[0073] Based on the above embodiments, the verification process of the method for improving water reserve resolution and accuracy based on machine learning spatial downscaling described in the present invention is described below.

[0074] 1. Study area

[0075] The research area of ​​this paper is Guantao County and its surrounding areas, located in the southern part of Hebei Province, China, with a total area of ​​460km 2 , located at 115°06′-115°40′ east longitude and 36°27′-36°47′ north latitude. Figure 3 The terrain is high in the southwest and low in the northeast, and is generally flat. The southern part is about 43m above sea level and the northern part is about 36m above sea level. Figure 3 ). Guantao County has a typical warm temperate semi-humid continental monsoon climate, with abundant sunshine and distinct dry and wet seasons. The average annual temperature is 14°C, and the average temperatures of the coldest month (January) and the hottest month (July) are -2.5°C and 27°C, respectively. Due to the lack of rational use and effective management of water resources, there is a serious over-exploitation of groundwater in Guantao County. The depletion of groundwater has induced a series of ecological and environmental and geological disasters such as ground subsidence, ground fissures, and soil salinization, affecting the comprehensive, coordinated and sustainable development of the region's economy and society. Therefore, it is very necessary to analyze and study the changes in water resource reserves in Guantao County.

[0076] 2. Data Source

[0077] The research data used in this invention include CSR RL06 Mascon, GLDAS, MODIS, TEMP and measured groundwater level data, etc. Previous studies have shown that these climate variables are closely related to changes in terrestrial water storage.

[0078] 2.1) GRACE data

[0079] The GRACE satellite, jointly developed by the National Aeronautics and Space Administration (NASA) and the German Aerospace Center (Deutsches Zentrum für Luft-und Raumfahrt, DLR), was successfully launched in 2002 and completed its mission in October 2017. Currently, GRACE data is mainly divided into two types of products: Spherical Harmonic Coefficient (SHC) and Mass Concentration Solutions (Mascon). The high-order coefficients of the SHC product have high-frequency noise and north-south striping errors. Therefore, when it is converted into global land water storage changes, it is necessary to perform data processing such as filtering and removing striping errors. Since the Mascon product has undergone first-order correction, C 20 The preprocessing work such as term replacement and GIA correction is not required, so no post-processing is required when using it. In addition, compared with SHC products, Mascon products have added correlation constraints and have a higher signal-to-noise ratio. Therefore, the present invention will analyze the changes in regional terrestrial water storage based on the CSR-M06 model released by CSR. The spatial resolution of the CSR-M06 model is 0.25°×0.25°, the temporal resolution is monthly, and the average value from January 2004 to December 2009 has been deducted.

[0080] 2.2) GLDAS data

[0081] The Global Land Data Assimilation System (GLDAS) was jointly developed by the Goddard Space Flight Center (GSFC) and the National Centers for Environmental Prediction (NCEP). Based on advanced land surface models and data assimilation methods, GLDAS outputs global surface status and flux data using satellite and surface observation data as basic data. Currently, GLDAS products include a variety of land surface models, namely Noah, Mosaic, CLM and VIC. The present invention uses soil moisture from the Noah land surface model with a spatiotemporal resolution of 0.25°×0.25° and monthly values, including soil moisture at four different depths (i.e., 0-10cm, 10-40cm, 40-100cm and 100-200cm). Since the spatial resolution of the GLDAS model cannot meet the requirements, the data is resampled to 0.05°.

[0082] 2.3) MODIS data

[0083] The Moderate-resolution Imaging Spectroradiometer (MODIS) carried on the Terra and Aqua satellites is an important instrument in the U.S. Earth Observation System (EOS) program mainly used to observe global biological and physical processes. Based on the Penman-Monteith formula and Mu's ET improved algorithm, the MODIS-MOD16 data set uses daily meteorological reanalysis data and MODIS-derived land cover, albedo, leaf area index and enhanced vegetation index data as input variables to calculate global land surface evaporation. The data set includes global evapotranspiration (ET), latent heat flux, potential surface evapotranspiration, and potential heat flux, with a spatial resolution of 500m and a temporal resolution of 8 days (MOD16A2) and 1 year (MOD16A3). The present invention extracts the ET component from the MOD16A2 product, and averages the 8-day time-resolved product to obtain monthly time resolution data. The diurnal land surface temperature (LST_day and LST_night) data are obtained from the 6th version of the MODIS-MOD11C3 product, with a spatial resolution of 0.05°×0.05° (~5.6×5.6km) and a temporal resolution of monthly.

[0084] 2.4) TEMP data

[0085] TEMP is the monthly average temperature data of China published in the National Tibetan Plateau Science Data Center, with a spatial resolution of about 1km and a time span of 1901.01 to 2017.12. Based on the global 0.5° climate data released by CRU and the global high-resolution climate products released by WorldClim, this dataset is the downscaling result generated in China through the Delta spatial downscaling scheme. The dataset was verified using observation data from 496 independent meteorological observation stations, and the results are reliable.

[0086] 2.5) Measured groundwater level data

[0087] In order to verify the high-resolution GWSA downscaling results, the present invention collected a total of 21 groundwater well observation data in Guantao County (from the groundwater monitoring management department of Guantao County), including shallow groundwater wells and deep groundwater wells. The spatial distribution of the points is shown in the figure. Figure 2 The black dots in the middle are shown. The groundwater level change can be obtained by subtracting the groundwater depth from the elevation of the groundwater well. The water level change value of each well is deducted from its own average value and compared with the groundwater storage change inverted by GRACE.

[0088] 3. Machine Learning Algorithms

[0089] Ensemble learning algorithm is a new research trend in machine learning. By integrating the training results of multiple weak learners to form a strong learner, the generalization ability of the model can be improved, thereby improving the prediction accuracy of the model. Compared with a single learning model, the ensemble learning model has higher accuracy, robustness and flexibility. Ensemble learning algorithms mainly include bagging and boosting. The idea of ​​the bagging algorithm is to extract data from the training set according to a specified proportion to form a sub-training set, then process the sub-training set separately, and finally average the results obtained from all sub-training sets as the final prediction result of the model. Its representative algorithms are Random Forest (RF) and Extra Trees Regressor (ETR) algorithms. The RF algorithm is a superior ensemble machine learning method first proposed by Breiman in 2001. It consists of a variety of classification and regression decision trees and is a popular tool for data classification and regression tasks. The randomness in the RF algorithm is mainly manifested in two aspects: one is to randomly select an equal amount of data with replacement from the original training data as training samples, and the other is to randomly select a part of the features to establish the decision tree when establishing the decision tree. These two kinds of randomness make the correlation between individual decision trees small, thereby improving the accuracy of the model. Previous studies have shown that random forests built based on these two kinds of randomness are very robust. The ETR algorithm is a relatively new machine learning algorithm developed on the basis of the random forest algorithm. Compared with the RF algorithm, the ETR algorithm completely randomly splits the descriptors at the nodes, and each tree is trained using the entire data set.

[0090] Unlike bagging, boosting is an iterative algorithm that trains the same training set multiple times. The weight of the sample is adjusted according to the result of the previous training in the next training. After meeting certain indicators, the final training result is output. The Adaptive Boosting Regressor (ABR) algorithm is a representative algorithm of boosting. At the beginning of training, the ABR algorithm first assigns equal weight to each sample, iteratively calculates the sample, and the next round of iteration uses the error of the weak learner in the previous round to update the weight value of the sample. Training terminates after reaching the set number of iterations or meeting the set accuracy requirements. The Gradient Boosting Regressor (GBR) algorithm is also designed based on the boosting idea and is often compared with the ABR algorithm. Different from the iteration conditions of the ABR algorithm, the GBR algorithm rebuilds the model in the gradient descent direction of the loss function of the previous iteration. The loss function can be used to evaluate the performance of the model. The smaller the loss function, the better the model performance. The performance of the model can be continuously improved by letting the loss function descend along the gradient direction. The relevant model is obtained by training the existing data set through the ensemble learning algorithm, and the trained model can be used to judge and predict new data. Since the ensemble learning algorithm has a very strong processing ability for nonlinear problems, it is widely used in the field of earth science and has achieved good results. In the downscaling study of GRACE data, RF and GBR algorithms have been widely used, but there are fewer discussions about ETR and ABR algorithms. The present invention will use RF, ETR, ABR and GBR algorithms to downscale the GRACE grid products and perform accuracy analysis. These four machine learning algorithms can all be called in Python's Scikit-learn function library.

[0091] 4. Data processing flow

[0092] The data processing flow of the present invention includes setting a modeling window, optimizing a modeling window and a machine learning algorithm, obtaining high-resolution TWSA and GWSA products, and verifying the downscaling results using measured data from groundwater wells. The specific downscaling process is as follows: Figure 4 shown. Figure 4 The dashed arrows in the middle represent the input data, the dotted arrows represent the output downscaling model and downscaling results, and the green dash-dot arrows represent the input validation data.

[0093] (1) Set up the modeling window ( Figure 2). In order to analyze the impact of the modeling window on the downscaling results, with Guantao County as the center and 0.5° grid as the basis, windows with sizes of 3×3, 5×5, 7×7 and 9×9 grids were established, and they are called Window Size 3 (WS3), Window Size 5 (WS5), Window Size 7 (WS7) and Window Size 9 (WS9). The green grid in the figure is the total study area, and the blue grid is the training data used in each modeling window.

[0094] (2) Due to the inconsistent temporal and spatial resolutions of the research data, the data needs to be preprocessed. The spatial resolution of TWSA is divided into 0.5° and 0.25°; the spatial resolutions of ET, SM, LST_day, LST_night and TEMP are divided into 0.5°, 0.25° and 0.05°. Except for the measured groundwater data, the temporal resolution of other research data is unified as monthly values.

[0095] (3) A 0.5° regression model between TWSA and hydrological variables was established monthly by combining RF, ETR, GBR and ABR algorithms (e.g., WS3+RF algorithm, referred to as the combined model) through the WS3, WS5, WS7 and WS9 modeling windows. The hydrological variables of 0.25° were input into the regression model of the corresponding month, and the 0.25° TWSA downscaling results were output. Then, the RMSE, MAE, NSE and CC between the 0.25° TWSA downscaled products and the CSR-M06 model were compared and analyzed in terms of spatiotemporal signals to determine the optimal combined model.

[0096] (4) Based on the optimal combination model obtained, the TWSA change with a spatial resolution of 0.05° is obtained by inputting the hydrological variables at 0.05°. According to the terrestrial water storage balance equation, the SM variables are deducted from the TWSA to obtain the high-resolution GWSA change. Finally, the measured groundwater well observation data in Guantao County are used to compare and verify the high-resolution groundwater storage changes after downscaling.

[0097] 5. Evaluation Metrics

[0098] The present invention evaluates the downscaling accuracy of the above design scheme through four statistical indicators, namely, root mean square error (RMSE), mean absolute error (MAE), Nash Sutcliffe Efficiency coefficient (NSE) and correlation coefficient (CC). The calculation formulas of each indicator are as follows:

[0099]

[0100]

[0101]

[0102]

[0103] Among them, Y represents the observed value, X represents the predicted value, and Represent the average values ​​of Y and X respectively, and n represents the number of data sets. For model building, the higher the correlation coefficients NSE and CC between the observed and predicted values, the better the model accuracy. The smaller the RMSE and MAE, the closer the predicted value is to the observed value, and the higher the accuracy of the model.

[0104] 6. Verification

[0105] 6.1) Performance

[0106] The regression relationship between TWSA and hydrological variables was established monthly using the WS3, WS5, WS7, and WS9 modeling windows in combination with RF, ETR, ABR, and GBR algorithms, and the 0.5° TWSA data were downscaled to 0.25°. Figure 5 The average values ​​of RMSE, MAE, NSE and CC of the validation set of the combined downscaling models for each month are given. Overall, the results of each combined downscaling model show good accuracy, with RMSE and MAE values ​​ranging from 6.77 to 13.62 mm and 3.36 to 10.21 mm, respectively, while NSE and CC are not less than 0.51 and 0.76. In WS3, WS5, WS7 and WS9, the RMSE of the ETR algorithm is 6.77 mm, 9.94 mm, 9.56 mm and 10.90 mm (MAE is 3.36 mm, 4.62 mm, 4.32 mm and 4.85 mm, respectively), which are better than the other three algorithms, and in each modeling window, the NSE and CC of the ETR algorithm have the maximum value. The above discussion shows that the ETR algorithm is better than the RF, ABR and GBR algorithms in model accuracy statistics.

[0107] When the combined downscaling model is used to downscale the TWSA grid data, there is a difference between the predicted data and the original data, which is called the residual. Previous studies have shown that the residual between the low-resolution TWSA original data and the predicted data is interpolated to high resolution through the Kriging interpolation method, and then added to the high-resolution downscaling result, which can improve the accuracy of the downscaling model to a certain extent. Therefore, it is necessary to further evaluate each combined downscaling model. Figure 6The spatial distribution of the TWSA change trend signal downscaled from 0.5° to 0.25° from January 2003 to December 2016 is given (taking the downscaling results of WS9 as an example). Figure 6 (a) shows the changing trend of terrestrial water storage inverted by the CSR-M06 model, with a spatial resolution of 0.25°; Figure 6 (b) indicates Figure 6 (a) The trend of terrestrial water storage resampled to 0.5°; Figure 6 (c)-(j) represent the downscaling results of different combination models. Overall, the spatial distribution of the downscaled signal is consistent with the original signal, showing that TWSA decreases more severely in the southwest and decreases more slowly in the northeast. After adding the residual, the spatial distribution of the downscaled signal of each combination model is closer to the original signal ( Figure 6 (g)-(j)). The downscaled signals in the WS3, WS5, and WS7 modeling windows all have similar performances.

[0108] In order to further analyze the accuracy of each combined downscaling model, Figure 7 The agreement between the downscaled signals of each model and the CSR-M06 model signal in spatial distribution is statistically analyzed, including the situation before and after adding residuals. Figure 7 It can be seen that after adding residuals, the downscaling accuracy of each combined model is improved. The RMSE value decreases from 1.69-2.55 mm / yr to 1.36-1.89 mm / yr; the MAE value decreases from 1.19-2.10 mm / yr to 1.03-1.53 ​​mm / yr; the NSE value increases from 0.28-0.69 to 0.51-0.87; and the CC value increases from 0.73-0.88 to 0.75-0.94. The accuracy of RMSE, MAE, NSE and CC increases by 20%-26%, 13%-27%, 20%-82% and 3%-7% respectively. After adding residuals, the downscaling accuracy of the RF algorithm in each modeling window is excellent. For example, in WS5, the RMSE and MAE of the RF algorithm are both the smallest, 1.36 mm / yr and 1.12 mm / yr respectively, while the NSE and CC are both the largest, 0.82 and 0.93 respectively. It can be seen that the downscaled signal obtained by using the RF algorithm to downscale the GRACE grid data is closer to the original signal in terms of spatial distribution.

[0109] 6.2) Time series analysis of terrestrial water storage changes in Guantao County before and after downscaling

[0110] According to the spatial signal analysis, the RF algorithm has better downscaling performance. The downscaling accuracy of the WS3, WS5, WS7 and WS9 modeling windows combined with the RF algorithm will be analyzed from the time dimension. Figure 8The two 0.25° grids where Guantao County is located ( Figure 6 (c) Small and medium boxes: 115.0°-115.5°E, 36.5°-36.75°N) TWSA time series, including the downscaling results of the CSR-M06 model and the RF algorithm combined with each modeling window. Figure 8 It can be seen that, except for individual time nodes (such as early 2004, early 2014, and late 2015), the downscaling results of different combination models are consistent with the time series of terrestrial water storage changes in Guantao County inverted by the CSR-M06 model. In order to analyze the accuracy of each combination model, the downscaling results of each modeling window and the time series of the CSR-M06 model are respectively regressed. Fig. 9 As shown. Fig. 9 It can be seen that the NSE and CC of the downscaling results of the four modeling windows combined with the RF algorithm are all greater than 0.98, indicating that the downscaling results of each modeling window combined with the RF algorithm have good accuracy and little difference. The RMSE, MAE, NSE and CC values ​​of WS5 are 9.67mm, 6.80mm, 0.990 and 0.997, respectively, and all indicators are better than WS3, WS7 and WS9. The statistical accuracy of the WS3 and WS7 modeling windows is second, and the statistical accuracy of the WS9 modeling window is the worst.

[0111] 7. Application

[0112] 7.1) Downscaling results at high spatial resolution

[0113] 7.11) Spatial and temporal distribution of TWSA downscaled signals

[0114] Based on the verification results of the new machine learning spatial downscaling method, it can be seen that the combined algorithm formed by WS5 and RF can obtain better downscaling results. In order to meet the needs of analyzing the changes in terrestrial water storage in Guantao County, the WS5 combined with the RF algorithm will be used to obtain TWSA products with a spatial resolution of 0.05°. Fig.10 The spatial distribution of TWSA change trends before and after downscaling from January 2003 to December 2016 is given. Within the WS5 modeling window, the spatial distribution of the two is very consistent. The terrestrial water storage decreases faster in the southwest and central regions, and decreases slower in the northeastern region. The downscaled signal can better reflect the detailed characteristics of terrestrial water storage changes. The time series of terrestrial water storage changes in Guantao County ( Fig.11 ), the NSE and CC values ​​of both are above 0.99, and decrease at the rates of -20.86 mm / yr and -21.79 mm / yr respectively.

[0115] 7.12) Spatial and temporal distribution of GWSA downscaled signals

[0116] According to the terrestrial water storage balance equation, the regional groundwater storage changes can be obtained by deducting the SM variable from the TWSA. Fig.12 The spatial distribution of GWSA obtained by inverting the CSR-M06 model and the 0.05° downscaling results after deducting the SM with the same spatial resolution is given. Fig.12 It can be seen that the spatial distribution of GWSA before and after downscaling is relatively consistent, and groundwater storage decreases faster in the western region and slower in the eastern region ( Fig.12 ). Fig.13 The time series of groundwater storage changes in Guantao County are given ( Fig.13 ). The NSE and CC of the groundwater reserves after downscaling are 0.980 and 0.994 respectively with those of the CSR-M06 inversion, which are in good consistency. The groundwater reserves change trends in Guantao County calculated by the CSR-M06 model and the 0.05° downscaling results are -14.53mm / yr and -15.46mm / yr respectively. The above discussion on the spatiotemporal signals of TWSA and GWSA before and after downscaling shows that the downscaling scheme proposed in the present invention can not only improve the spatial resolution of GRACE data, but also effectively retain the original signal and has high reliability.

[0117] 7.2) Comparative analysis of high-resolution downscaling results and measured data

[0118] In order to analyze the reliability of the high-resolution (0.05°) GWSA after downscaling, the present invention collects observation data from 21 groundwater wells in Guantao County to verify it, including 11 shallow wells (numbered: 1-11) and 10 deep wells (numbered: 12-21). The common time period of the inverted groundwater and measured groundwater data is 2012-2016. Since the water supply degree of groundwater in different regions has a large uncertainty, the present invention mainly analyzes the change trend and correlation of the inverted and measured groundwater. Fig.14 The spatial distribution of groundwater logging wells is given, where black dots and red dots represent shallow and deep groundwater wells, respectively. The time series of water level changes of some groundwater wells and the groundwater storage changes inverted from their grids are also given. Due to the irregular observation time of groundwater well data, the correlation coefficients of measured and inverted groundwater changes on the measured time scale and annual time scale are calculated, as shown in Figure 2. Fig.14 Most of the measured data and the groundwater storage changes inverted by GRACE show a downward trend ( Fig.14). The consistency between the observation data of shallow groundwater wells and the inverted groundwater changes is poor. Except for the correlation coefficients of observation wells 1, 3, 5 and 6, the correlation coefficients of other observation wells are negative. The annual correlation coefficients of the groundwater changes in deep groundwater wells and the inverted groundwater changes are all above 0.7, except for observation wells 18 and 19, which have poor correlations. By comparing the changes in groundwater reserves in observation wells of two different depths with the inverted results, it is shown that the groundwater loss in Guantao County may be mainly caused by over-exploitation of deep groundwater.

[0119] On the basis of the above embodiments, the present invention also discloses a system for improving water reserve resolution and accuracy based on machine learning spatial downscaling, including: an initial model construction module, used to establish an initial downscaling model between terrestrial water reserve change anomalies and multiple hydrological variables based on a modeling window; a screening module, used to use N different machine learning algorithms to learn the initial downscaling model, and combined with M different optimization selectors, screen out the optimal terrestrial water reserve change anomaly solution model and the optimal modeling window; an optimal model construction module, used to construct an optimal downscaling model based on the optimal terrestrial water reserve change solution function and the optimal modeling window; a result output module, used to output the terrestrial water reserve change anomaly value according to the optimal downscaling model.

[0120] As for the system embodiment, since it corresponds to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiment part.

[0121] Although the present invention has been disclosed as above in the form of a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art may make possible changes and modifications to the technical solution of the present invention by using the methods and technical contents disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope of the technical solution of the present invention.

[0122] The contents not described in detail in the specification of the present invention belong to the common knowledge of professionals in the field.

Claims

1. A method for improving water storage resolution and accuracy based on machine learning spatial downscaling, characterized in that: include: Based on the modeling window, an initial downscaling model between terrestrial water storage anomalies and multiple hydrological variables is established; N different machine learning algorithms are used to learn the initial downscaling model, and combined with M different optimization selectors, the optimal ground water storage change anomaly solution model and the optimal modeling window are screened; According to the optimal terrestrial water storage change solution function and the optimal modeling window, the optimal downscaling model is constructed; According to the optimal downscaling model, the abnormal values ​​of terrestrial water storage changes are output; The initial downscaling model is a low spatial resolution model, expressed as follows: Among them, TWSA i represents the low spatial resolution groundwater storage anomaly of the i-th month, i = 1, 2, 3, ...; v ij represents the hydrological variable of the jth type of low spatial resolution in the i-th month, j = 1, 2, 3, ...; w n represents the modeling window, n = 3, 5, 7, 9; f i It represents the low spatial resolution model for solving the abnormal change of groundwater storage in the ith month; w3, w5, w7, and w9 represent the modeling windows of 3×3, 5×5, 7×7, and 9×9 grids, respectively; N different machine learning algorithms, including: RF machine learning algorithm, ETR machine learning algorithm, ABR machine learning algorithm and GBR machine learning algorithm; M different optimization selectors, including: four optimization selectors using root mean square error RMSE, mean absolute error MAE, Nash efficiency coefficient NSE and correlation coefficient CC as evaluation indicators; The optimal groundwater storage abnormality solution model f is obtained by screening through the following formulas (2) and (3): opt and the optimal modeling window w opt : Among them, f RF 、f ETR 、f ABR and f GBR They represent the basic models using RF, ETR, ABR and GBR machine learning algorithms as f in formula (1); The meanings of formulas (2) and (3) are as follows: RF 、f ETR 、f ABR and f GBR As the basic model of f in formula (1), and respectively substitute into modeling windows w3, w5, w7 and w9 of different sizes, the four evaluation indicators of RMSE, MAE, NSE and CC are used to evaluate the performance of different f models. RF 、f ETR 、f ABR and f GBR The combined output results of modeling windows w3, w5, w7 and w9 of different sizes are comprehensively evaluated to screen out the f with the best downscaling result accuracy. opt and w opt .

2. The method for improving water storage resolution and accuracy based on machine learning spatial downscaling according to claim 1 is characterized in that: The smaller the values ​​of RMSE and MAE, the higher the accuracy of the model; the value of NSE ranges from negative infinity to 1, and the closer the NSE is to 1, the better the quality of the model and the higher the model credibility; the value of CC ranges from -1 to 1, and the closer the CC is to 1, the better the quality of the model and the higher the model credibility.

3. The method for improving water storage resolution and accuracy based on machine learning spatial downscaling according to claim 1 is characterized in that: The optimal downscaling model is a high spatial resolution model, expressed as follows: in, Indicates the abnormal change of groundwater storage at high spatial resolution, Represents the j-th type of high spatial resolution hydrological variable.

4. The method for improving water storage resolution and accuracy based on machine learning spatial downscaling according to claim 3 is characterized in that: Also includes: according to The groundwater storage change with high spatial resolution is obtained by solving the following formula (5): in, Represents soil water storage changes at high spatial resolution.

Images