A method for downscaling groundwater storage deviation based on fusion model
By integrating multiple machine learning and deep learning models, combining linear and nonlinear fusion models, the problem of inconsistent resolution of GRACE satellite data and GLDAS data is solved, and the reduction accuracy of groundwater storage deviation is improved, and the groundwater with higher accuracy is provided The reserve data set provides an effective reference for groundwater resource management.
Patent Information
- Application Number
- CN202211271859.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-10-18
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2042-10-18
AI Technical Summary
In the prior art, the resolution of GRACE satellite data and GLDAS data is inconsistent, resulting in low downscale accuracy of groundwater storage deviations, and a single type of machine learning model or deep learning model has poor causal effect under complex surface conditions.
Using the groundwater reserve deviation reduction method based on the fusion model, the fusion model of a variety of machine learning and deep learning models is constructed by resampling and driving variables through bilinear interpolation, using Nash efficiency coefficient, Pearson correlation coefficient, consistency index, and Kling–Gupta efficiency coefficient evaluates the model accuracy, combines linear and nonlinear fusion models for training, unify data resolution, and improves downscale accuracy.
It significantly improves the downscale accuracy of groundwater storage deviation, provides a groundwater storage data set with higher spatial accuracy and reliability, and provides a reference for the planning and utilization of groundwater resources at regional or basin scales.
Smart Images

Figure CN115617935B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the cross-technical fields of hydrology, image enhancement, etc., and particularly relates to a groundwater storage deviation downscaling method based on a fusion model. Background Art
[0002] Groundwater is the main source of irrigation water and human life water in arid and semi-arid areas. Under the background of global extreme weather and increasing population, groundwater extraction has surged, and the problems caused by groundwater depletion have attracted more and more attention. Therefore, accurate monitoring of groundwater changes is of great significance to regional groundwater resource management. Traditional groundwater network monitoring consumes a lot of manpower and material resources and is difficult to apply to larger time and space scales. The Gravity Measurement and Climate Monitoring (GRACE) satellite launched in March 2002 can provide information on changes in the earth's gravity field with high temporal and spatial accuracy. Therefore, a large number of studies are based on water balance theory, combining GRACE satellite data and Global Land Assimilation System (GLDAS) data to invert and reconstruct groundwater storage deviations. Since the actual spatial resolutions of GRACE data after processing by the spherical harmonic coefficient method and the point mass method are 1°×1° and 0.5°×0.5°, respectively, in order to maintain the consistency of the resolution of the two types of data, most studies resampled the GLDAS data with a resolution of 0.25°×0.25° to 0.5°×0.5° or 1°×1°, resulting in the loss of many important spatial information in GLDAS.
[0003] In order to solve the problem of inconsistent resolution between GRACE satellite data and GLDAS data, scholars have proposed many downscaling methods, which are usually divided into dynamical downscaling methods and statistical downscaling methods. The dynamical downscaling method has a more solid physical basis, but it consumes a lot of computing resources. When the spatial and temporal scale is large or the resolution is increased, the computing power of the computer is extremely high. Therefore, the statistical downscaling method has a wider application. Among the many statistical downscaling methods, machine learning models and deep learning models can better handle the nonlinear relationship between driving variables and target variables. However, from the existing research, when the underlying surface conditions are complex, the causal relationship captured by a single type of machine learning model or deep learning model is different and limited in scope, so the global effect of model downscaling is still poor. Therefore, it is of great significance to propose an effective multi-model fusion method to improve the accuracy of GRACE satellite data, invert and reconstruct groundwater storage deviations. Summary of the invention
[0004] In order to overcome the problems of the prior art, the present invention proposes a groundwater storage deviation downscaling method based on a fusion model. By establishing a downscaling model of groundwater volume deviation, a set of groundwater storage deviation data sets with higher spatial accuracy and more reliability is obtained, which can provide a reference for the planning and utilization of groundwater resources at the regional or basin scale.
[0005] The object of the present invention is achieved in that:
[0006] A groundwater storage deviation downscaling method based on a fusion model includes the following steps:
[0007] Step 1, driving variable resampling:
[0008] Each driving variable related to groundwater storage was resampled from a resolution of 0.25°×0.25° to 0.5°×0.5° using bilinear interpolation to obtain a data set of driving variables;
[0009] Step 2: Construct a statistical relationship model between the driving variable and the target variable:
[0010] A fusion model is used to construct a statistical relationship model between the driving variable processed in step 1 and the target variable; wherein the target variable is the GRACE terrestrial water storage deviation data obtained by the point mass method;
[0011] The construction of the fusion model includes:
[0012] S21, using a 3-fold cross validation method, using m machine learning models and n deep learning models to perform training on the data set of the driving variables obtained in step 1 respectively;
[0013] S22, Nash efficiency coefficient (NS), Pearson correlation coefficient (CC), consistency index (IG) and Kling–Gupta efficiency coefficient (KG) are used to evaluate the accuracy of each model in S21, and the expressions are:
[0014]
[0015] In the formula, o i and i are the ith observation value and the ith simulation value respectively; σ s and σ o are the standard deviations of the simulated and observed values, respectively; μ s and μ o are the means of the simulated and observed values, respectively; and are the means of the observed and simulated series, respectively;
[0016] The new array is formed by the output of the single type model with the highest ranking in the sum of the above four indicators and the output of the m+n models with the smallest absolute deviation from the original value;
[0017] S23, respectively using a linear fusion model and a nonlinear fusion model to train the new array obtained by processing in S22, and using a fusion model with the largest sum of NS, CC, IG and KG as the statistical relationship model;
[0018] Step 3, obtain high-resolution terrestrial water storage deviations:
[0019] The statistical relationship model constructed in step 2 is used to simulate the terrestrial water storage deviation with a resolution of 0.5°×0.5°, and the residual with a resolution of 0.5°×0.5° is obtained by subtracting the residual from the target variable, and the residual resolution is resampled to 0.25°×0.25° by the nearest neighbor interpolation method; and then the residual is added to the terrestrial water storage deviation with a resolution of 0.25°×0.25° simulated by the statistical relationship model constructed in step 2 to obtain the downscaled terrestrial water storage deviation;
[0020] Step 4: Verify the rationality of the downscaling results:
[0021] The surface water storage deviation is subtracted from the downscaled terrestrial water storage deviation obtained in step 3 to obtain a groundwater storage deviation of 0.25°×0.25°. Then, the rationality of the downscaling result is verified by performing a correlation analysis on the measured groundwater level in the study area.
[0022] Furthermore, the driving variables in step 1 include precipitation, surface temperature, surface runoff, groundwater runoff, actual evapotranspiration and soil moisture.
[0023] Furthermore, in step 2 S21, the machine learning model is constructed using the PyCaret package in Python, and the tune_model function is used to obtain the optimal hyperparameters of each model.
[0024] Furthermore, in step 2 S21, the deep learning model is constructed using the tsai package in Python, and hyperparameter optimization is performed using the hyperop package.
[0025] Furthermore, in step 2 S23, the super ensemble average model is selected as the linear fusion model, and the expression is:
[0026]
[0027] In the formula, (S MMSE ) t is the simulated value of the MMSE model during period t; is the average value of the multi-model simulation values during period t; w j is the weight value of each model; F j,t and are the predicted value and predicted mean of the j-th model in period t respectively; z is the number of models.
[0028] Furthermore, in step 2 S23, a random forest model is selected as the nonlinear fusion model.
[0029] Furthermore, in step 4, the surface water storage deviation includes canopy water storage deviation, snow water equivalent deviation and soil water storage deviation.
[0030] The advantages and beneficial effects of the present invention compared to the prior art are:
[0031] The groundwater reserve deviation downscaling method described in the present invention constructs a fusion model of multiple single-type machine learning models and deep learning models, and uses the Nash efficiency coefficient, Pearson correlation coefficient, consistency index and Kling–Gupta efficiency coefficient as indicators to verify the consistency of the model before and after downscaling, unifies the spatial resolution of GRACE satellite data and GLDAS data, significantly improves the downscaling accuracy of traditional single-type models for groundwater reserve deviations, and provides a certain reference for regional or basin-scale groundwater resource planning and utilization. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.
[0033] Figure 1 It is a flow chart of the groundwater storage deviation downscaling method based on the fusion model of the present invention;
[0034] Figure 2 The figure is a flow chart for building the fusion model described in the present invention. DETAILED DESCRIPTION
[0035] The following will be combined with the drawings of the present invention in the embodiments of the present invention to clearly and completely describe the technical solutions of the present invention in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0036] Example:
[0037] like Figure 1 The process shown in FIG. 1 is a process for downscaling groundwater reserves based on a fusion model, and the process is shown in FIG. This embodiment takes the study of groundwater reserves in Northwest my country as an example, and provides a groundwater reserves deviation downscaling method based on a fusion model, including the following steps:
[0038] Step 1, driving variable resampling:
[0039] The driving variables related to groundwater storage, namely precipitation, surface temperature, surface runoff, groundwater runoff, evapotranspiration, and soil moisture, were resampled from a resolution of 0.25°×0.25° to 0.5°×0.5° using the bilinear interpolation method to obtain a data set of driving variables.
[0040] The data sources and spatial accuracy of the driving variables in step 1 are shown in Table 1.
[0041] Table 1 Data sources and specific information
[0042]
[0043] Step 2: Construct a statistical relationship model between the driving variable and the target variable:
[0044] The fusion model is used to construct a statistical relationship model between the driving variable processed in step 1 and the target variable, wherein the target variable is the GRACE terrestrial water storage deviation (GRACE-TWSA) data obtained by the Jet Propulsion Laboratory (JPL) using the point mass method. The specific data comes from the article "Improved methods for observing Earth's time variable mass distribution with GRACE using spherical cap mascons" (doi: 10.1016 / j.geophys.res.solidearth.2015) published by Watkins et al. in the journal "J.Geophys.Res.SolidEarth" (doi: 10.1002 / 2014JB011547 ) at:
[0045] https: / / podaac.jpl.nasa.gov / dataset / TELLUS_GRAC-GRFO_MASCON_GRID_RL06_V2).
[0046] Compared with the spherical harmonic coefficient method, the point mass method can minimize the leakage error of GRACE data without post-processing. Therefore, in order to reduce the uncertainty of GRACE data, this application selects the data obtained by the point mass method as the target variable.
[0047] like Figure 2 As shown, the construction of the fusion model includes the following three steps:
[0048] S21, using the 3-fold cross-validation method (i.e., using 2 / 3 of the pixels as the training set and 1 / 3 of the pixels as the validation set each time), m machine learning models and n deep learning models are trained on the data set of driving variables with a resolution of 0.5°×0.5° obtained in step 1.
[0049] This embodiment includes 22 machine learning models and 16 deep learning models, and each model is shown in Table 2 and Table 3 below.
[0050] Table 2 22 machine learning models
[0051]
[0052]
[0053] Table 3 16 deep learning models
[0054]
[0055] The above machine learning models are constructed using the pycaret package in Python, and the tune_model function is used to obtain the optimal hyperparameters of each model.
[0056] The deep learning model is constructed using the tsai package in Python, which is based on Pytorch and Fastai and uses the latest deep learning technology. It can be used for time series classification, regression, and prediction. Due to the large number of models and samples, the range of neural network hyperparameters is given in advance, and the hyperparameter optimization is performed using the hyperop package. The types and ranges of the hyperparameters of each model are shown in Tables 4 and 5 below.
[0057] Table 4 Common hyperparameters of deep learning models and their value ranges
[0058]
[0059] Table 5. Relevant hyperparameters unique to some deep models and their value ranges
[0060]
[0061] S22, Nash efficiency coefficient (NS), Pearson correlation coefficient (CC), consistency index (IG) and Kling–Gupta efficiency coefficient (KG) are used to evaluate the accuracy of each model in S21, and the expressions are:
[0062]
[0063] In the formula, o i and i are the ith observation value and the ith simulation value respectively; σ s and σ o are the standard deviations of the simulated and observed values, respectively; μ s and μ o are the means of the simulated and observed values, respectively; and are the means of the observed and simulated series, respectively.
[0064] The closer the above indicators are to 1, the higher the model accuracy. Therefore, the output of the single type model with the highest ranking in the sum of the four indicators and the output of the m+n models with the smallest absolute deviation from the original value form a new array.
[0065] Specifically, to keep the code simple, this embodiment selects the top five models and the output results with the smallest absolute deviation from the measured values among the m+n models (a total of 38 models) to form a new array.
[0066] The following Tables 6 and 7 respectively list the cross-validation results of the correlation coefficient (CC), consistency index (IG), Nash efficiency coefficient (NS) and Kling–Gupta efficiency coefficient (KG) of the 16 deep learning models and 22 machine learning models involved in this embodiment. Compared with the machine learning model, the mean of the evaluation index of the deep learning model is relatively large, and it shows stronger learning ability when simulating TWSA. At the same time, the standard deviation of each indicator of the deep learning model is smaller than that of the machine learning model, indicating that each deep learning model has a more similar learning effect.
[0067] For example, LSTM, GRU and RNN are all recurrent neural networks. In addition, different evaluation indicators have different focuses and the evaluation results vary greatly. For example, in the deep learning models, the CC and IG values of LSTM-FNC, GRU, RNN and LSTM are large, but the NS and KG values are small, indicating that although these models can accurately simulate the changing trend of TWSA, the simulation error is large.
[0068] Table 6 Cross-validation results of 22 machine learning models
[0069]
[0070] Table 7 Cross-validation results of 16 deep learning models
[0071]
[0072]
[0073] S23, respectively using a linear fusion model and a nonlinear fusion model to train the new array obtained by processing S22, and using the fusion model with the largest sum of NS, CC, IG and KG as the statistical relationship model for downscaling.
[0074] Among them, the super ensemble average (MMSE) model is used as the linear fusion model, and the expression is:
[0075]
[0076] In the formula, (S MMSE ) t is the simulated value of the MMSE model during period t; is the average value of the multi-model simulation values during period t; w j is the weight value of each model; F j,t and are the predicted value and predicted mean of the j-th model in time period t respectively; z is the number of models, which is 5 in this embodiment; and the weight of each model is obtained by minimizing the model error.
[0077] The random forest (RF) model is used as the nonlinear fusion model.
[0078] Step 3, obtain high-resolution terrestrial water storage deviations:
[0079] The statistical relationship model constructed in step 2 is used to simulate the terrestrial water storage deviation (Corse-TWSA) with a coarse resolution of 0.5°×0.5°, and the Corse-TWSA is subtracted from the GRACE terrestrial water storage deviation (GRACE-TWSA, i.e., the target variable) obtained by the point mass method to obtain a residual with a resolution of 0.5°×0.5°, and the residual resolution is resampled to 0.25°×0.25° using the nearest neighbor interpolation method; and then the residual is added to the terrestrial water storage deviation (Fine-TWSA) with a fine resolution of 0.25°×0.25° simulated by the statistical relationship model constructed in step 2 to obtain the downscaled terrestrial water storage deviation (SIM-TWSA).
[0080] As shown in Table 8, MMSE did not improve the simulation accuracy of a single model, while the nonlinear RF model significantly improved KG and NS and reduced the simulation error of a single model. Compared with the deep learning model (ResNet) and machine learning model (xgboost) with the highest accuracy, the KG and NS of the RF fusion model increased by 49%, 60% and 45%, 51%, respectively. Therefore, the nonlinear RF fusion model was used to downscale the groundwater storage deviation in Northwest China. Taking June 2005 as an example, before and after downscaling, the high values of TWSA were concentrated in Yili, Junggar Basin, northern Ningxia and northern Shaanxi in western Xinjiang, and the low values were concentrated in the border between Xinjiang and Qinghai, southern Qinghai Province and southern Shaanxi Province. At the same time, the downscaled data set can more carefully describe the distribution of TWSA.
[0081] Table 8 Cross-validation results of two fusion models
[0082]
[0083] Step 4: Verify the rationality of the downscaling results:
[0084] From the downscaled terrestrial water storage deviation (SIM-TWSA) obtained in step 3, subtract the surface water storage deviation (canopy water storage deviation, snow water equivalent deviation and soil water storage deviation) to obtain a groundwater storage deviation of 0.25°×0.25°. Since the groundwater storage deviation (GWSA) is equal to the groundwater level deviation multiplied by the soil water supply degree, and there is currently no spatial distribution data on the soil aquifer water supply degree, the rationality of the downscaling result can be verified by performing a correlation analysis on the measured groundwater level in the study area.
[0085] This embodiment obtains the groundwater reserve deviation after downscaling based on step 4. The rationality of the downscaling result is verified by performing correlation analysis on typical measured groundwater levels in the Shiyang River Basin, Guanzhong region and the middle and upper reaches of the Heihe River. When multiple observation wells are distributed on a grid point, the mean of the groundwater level deviations of the multiple wells is taken for calculation.
[0086] The Pearson correlation coefficients of each typical region are relatively high, and most grid points pass the significance test of 0.01. The correlation coefficients in Guanzhong region range from 0.08 to 0.88, with an average of 0.66; the correlation coefficients in Shiyang River Basin range from -0.04 to 0.72, with an average of 0.48; the correlation coefficients in Heihe River Basin range from 0.05 to 0.93, with an average of 0.45, indicating that the downscaled GWSA can effectively capture the changes in groundwater levels in the study area.
[0087] Finally, it should be noted that the above is only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred arrangement scheme, a person skilled in the art should understand that the technical solution of the present invention (such as the use of various formulas, the sequence of steps, etc.) can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.
Claims
1. A groundwater storage deviation downscaling method based on a fusion model, characterized in that: The steps of the method are: Step 1, driving variable resampling: Each driving variable related to groundwater storage was resampled from a resolution of 0.25°×0.25° to 0.5°×0.5° using bilinear interpolation to obtain a data set of driving variables; The driving variables include precipitation, surface temperature, surface runoff, subsurface runoff, actual evapotranspiration, and soil moisture; Step 2: Construct a statistical relationship model between the driving variable and the target variable: A fusion model is used to construct a statistical relationship model between the driving variable processed in step 1 and the target variable, wherein the target variable is the GRACE terrestrial water storage deviation data obtained by the point mass method; The construction of the fusion model includes: S21, using a 3-fold cross-validation method, 22 machine learning models and 16 deep learning models are respectively trained on the data set of the driving variables obtained in step 1; the machine learning model is constructed using the PyCaret package in Python, and the tune_model function is used to obtain the optimal hyperparameters of each model; the deep learning model is constructed using the tsai package in Python, and the hyperparameter optimization is performed using the hyperop package; The 22 machine learning models are: extreme gradient boosting model, light gradient boosting model, Cat boosting regressor, Ada boosting regressor, support vector machine, Theilson regressor, decision tree model, Cornell bridge model, gradient boosting regression model, Bayesian bridge model, orthogonal matching pursuit model, automatic regression decision model, lasso regression model, Hooper regressor, passive invasion regressor, K-neighbor regressor, extra tree regressor, lasso minimum angle regressor, bridge regressor, random sampling consistency model, minimum angle regressor, and linear regressor; The 16 deep learning models are: residual neural network, residual convolutional neural network, gated recurrent unit-fully convolutional network, time series transformer, multilayer perceptron, full-scale 1D convolutional neural network, extreme start time series network, start time series network, recurrent neural network-fully convolutional network, temporal convolutional network, releasable convolutional neural network, long short-term memory-fully convolutional network, gated recurrent unit neural network, recurrent neural network, long short-term memory neural network, and fully convolutional network; S22, Nash efficiency coefficient (NS), Pearson correlation coefficient (CC), consistency index (IG) and Kling–Gupta efficiency coefficient (KG) are used to evaluate the accuracy of each model in S21, and the expressions are: In the formula, o i and i are the ith observation value and the ith simulation value respectively; σ s and σ o are the standard deviations of the simulated and observed values, respectively; μ s and μ o are the means of the simulated and observed values, respectively; and are the means of the observed and simulated series, respectively; The new array is composed of the outputs of the top five single-type models ranked by the sum of the above four indicators, and the output results of the 22+16 models with the smallest absolute deviation from the original value; S23, respectively using a linear fusion model and a nonlinear fusion model to train the new array obtained by processing in S22, and using a fusion model with the largest sum of NS, CC, IG and KG as the statistical relationship model; The super ensemble average model is selected as the linear fusion model, and the expression is: In the formula, (S MMSE ) t is the simulated value of the MMSE model during period t; is the average value of multi-model simulation values during period t; w j is the weight value of each model; F j,t and are the predicted value and predicted mean of the jth model in period t respectively; z is the number of models; The random forest model is selected as the nonlinear fusion model; Step 3, obtain high-resolution terrestrial water storage deviations: The statistical relationship model constructed in step 2 is used to simulate the terrestrial water storage deviation with a resolution of 0.5°×0.5°, and the deviation is subtracted from the target variable to obtain a residual with a resolution of 0.5°×0.5°, and the residual resolution is resampled to 0.25°×0.25° using the nearest neighbor interpolation method; and then the residual is added to the terrestrial water storage deviation with a resolution of 0.25°×0.25° simulated by the statistical relationship model constructed in step 2 to obtain the downscaled terrestrial water storage deviation; Step 4: Verify the rationality of the downscaling results: The surface water storage deviation is subtracted from the downscaled terrestrial water storage deviation obtained in step 3 to obtain a groundwater storage deviation of 0.25°×0.25°. Then, the rationality of the downscaling result is verified by performing a correlation analysis on the measured groundwater level in the study area. The surface water storage deviation includes canopy water storage deviation, snow water equivalent deviation and soil water storage deviation.
