Multi-machine learning model GRACE satellite inversion groundwater reserve change downscaling method fusing environmental factors
By integrating multiple machine learning models and ridge regression fusion models that incorporate environmental factors, and optimizing model weights and structure, the problem of insufficient resolution in groundwater storage change retrieval by GRACE satellite was solved, enabling the generation of high-precision groundwater storage change data and supporting refined water resource management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NORTH CHINA UNIV OF WATER RESOURCES & ELECTRIC POWER
- Filing Date
- 2026-01-15
- Publication Date
- 2026-04-28
AI Technical Summary
In existing technologies, the spatial resolution of groundwater storage changes retrieved by GRACE satellite is insufficient, making it impossible to generate high-precision groundwater storage change data. Furthermore, the differences between various machine learning models under different environments are not fully utilized, resulting in insufficient optimization of model parameters.
We employ multiple machine learning models that integrate environmental factors, dynamically allocate weights through a ridge regression fusion model, and optimize model parameters and structure by combining fold cross-validation and hydrological models. We use drought index and population size as environmental factors to generate high-resolution groundwater storage change data.
It improves the spatial resolution and data accuracy of groundwater storage changes, generates reliable high-resolution groundwater storage change data, and supports refined water resource management.
Smart Images
Figure CN121936288A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of GRACE gravity satellite inversion of groundwater storage changes, and in particular to a GRACE satellite inversion groundwater storage change downscaling method that integrates multiple machine learning models with environmental factors. Background Technology
[0002] In downscaling methods for groundwater storage changes retrieved from GRACE satellite data, the spatial resolution obtained using spherical harmonic coefficients and the Mascon method is too low, hindering more refined water resource studies in the research area. Therefore, downscaling has become an important method to obtain high-precision groundwater storage change data. Currently, a framework that effectively leverages the advantages of multiple machine learning models has not yet been established for downscaling research on GRACE satellite data. When using multiple machine learning models, fixed weight allocation is often adopted, failing to consider their differences under different environments and ignoring the impact of environmental factors on the models. The inability to optimize model parameters and structure based on data feedback significantly limits the generation of high-reliability, high-resolution groundwater storage change data. Summary of the Invention
[0003] The purpose of this invention is to provide a downscaling method for groundwater storage changes retrieved from the GRACE satellite by integrating multiple machine learning models based on environmental factors, thereby generating reliable high-resolution data on groundwater storage changes.
[0004] To achieve the above objectives, the present invention provides the following solution:
[0005] A downscaling method for groundwater storage changes retrieved via the GRACE satellite using multiple machine learning models that integrate environmental factors, including:
[0006] Collect GRACE satellite data and retrieve changes in terrestrial water storage based on the GRACE satellite data.
[0007] Based on the changes in terrestrial water storage, the changes in groundwater storage are obtained;
[0008] Multiple machine learning models are used to downscale the changes in groundwater storage to obtain high-resolution values of groundwater storage changes. The weights of the various machine learning models are obtained by analyzing a ridge regression fusion model that incorporates environmental factors.
[0009] Optionally, obtaining the changes in groundwater storage based on the changes in terrestrial water storage includes: obtaining the changes in terrestrial water storage obtained by inversion from GRACE satellite data and the changes in surface water storage obtained by the GLDAS assimilation model, and then obtaining the changes in groundwater storage through a water balance formula.
[0010] Optionally, during the training of various machine learning models, folded cross-validation is used to divide the training data into K subsets and perform K iterations of training and validation.
[0011] Optionally, the weights of various machine learning models can be obtained by using a ridge regression fusion model that incorporates environmental factors, including:
[0012] Multiple machine learning models are used to predict the training samples, and the predicted value corresponding to each model is obtained.
[0013] The environmental factors are fused with the predicted values and used as input to the ridge regression fusion model for weight analysis, thereby solving for the weights of various machine learning models.
[0014] Optionally, the ridge regression fusion model is:
[0015] ;
[0016] in, For the final downscaling data results, For the first The weights of a machine learning model and Drought index and population size for the first Weight adjustment coefficients for data in a machine learning model It is a drought index. It refers to population size. It is the first The predicted values of a machine learning model, where M is the number of base models.
[0017] Optionally, solving for the weights of various machine learning models includes minimizing the following loss function that includes an L2 regularization term:
[0018] ;
[0019] in This represents the true value observed by GRACE. It is a regularization parameter. These are the parameters in the prediction model.
[0020] Optionally, after obtaining high-resolution groundwater storage change values, the following can be included:
[0021] The high-resolution groundwater storage change values were verified. Based on the verification results, the parameters and structure of the ridge regression fusion model were optimized. When there were abnormally abrupt grid data in the ridge regression fusion model, the ridge regression fusion model was re-corrected by adding groundwater well data to the weights. If groundwater well data existed in the grid data, the groundwater well data was identified. If groundwater well data did not exist in the grid data, the grid data was corrected using the inverse distance weighted interpolation method.
[0022] The beneficial effects of this invention are as follows: This invention can invert the changes in regional terrestrial water storage based on GRACE satellite data; based on the changes in terrestrial water storage, it solves for changes in groundwater storage by using existing surface water storage changes; it utilizes multiple machine learning models to downscale the changes in groundwater storage, employing folded cross-validation during model training to ensure the objectivity of model optimization and data accuracy; it uses a ridge regression fusion model, whose unique regularization mechanism overcomes the collinearity interference of multiple machine learning models, and incorporates environmental factors, dynamically allocating model weights to the fusion model of influencing factors. This forms a method of complementary advantages among multiple machine learning models that integrate environmental factors, generating high-precision groundwater storage change data; the downscaled high-resolution product is validated using independent observation data and hydrological model simulation results; based on the validation results, the model parameters and structure are optimized; groundwater well data is used to intervene at anomalous abrupt change points, ultimately generating reliable high-resolution groundwater storage change data. This invention achieves the goal of improving spatial resolution accuracy by fusing multiple machine learning models and optimizing model parameters and structure. Attached Figure Description
[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0024] Figure 1 This is a flowchart illustrating a method for downscaling groundwater storage changes using the GRACE satellite inversion, which incorporates multiple machine learning models based on environmental factors, according to an embodiment of the present invention. Detailed Implementation
[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0026] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0027] like Figure 1 As shown, this embodiment provides a downscaling method for groundwater storage changes retrieved by the GRACE satellite using multiple machine learning models that integrate environmental factors, including:
[0028] Collect GRACE satellite data and retrieve changes in terrestrial water storage based on the GRACE satellite data.
[0029] Based on the changes in terrestrial water storage, the changes in groundwater storage are obtained;
[0030] Multiple machine learning models are used to downscale the changes in groundwater storage to obtain high-resolution values of groundwater storage changes. The weights of the various machine learning models are obtained by analyzing a ridge regression fusion model that incorporates environmental factors.
[0031] Furthermore, based on the changes in terrestrial water storage, the changes in groundwater storage are obtained by: obtaining the changes in terrestrial water storage from GRACE satellite data inversion and the changes in surface water storage from the GLDAS assimilation model, and then obtaining the changes in groundwater storage through a water balance formula.
[0032] Furthermore, the changes in groundwater storage in the inverted area include:
[0033] Collect GRACE satellite data, high-precision downscaling factor data (precipitation data, evapotranspiration data, normalized vegetation index, etc.) and hydrological model data (NOAH, VIC, CLM, MOSAIC in the GLDAS global land surface data assimilation model), and retrieve the changes in regional groundwater storage based on existing GRACE satellite data;
[0034] For example, changes in terrestrial water storage will cause changes in the Earth's gravitational field. These changes can be converted into equivalent changes in water height, thus revealing the changes in terrestrial water storage. The formula for retrieving changes in terrestrial water storage is as follows:
[0035] ;
[0036] in For the equivalent water column height, The change in surface density. The density of water (take 1000) ), For the Earth's radius, Earth's average density (taken as 5517) ), Indicates the first The number of load Loves in the order. For regularized Legendre functions, and For perturbation potential coefficients, and For latitude and longitude.
[0037] Changes in terrestrial water storage can be derived from GRACE satellite data, including changes in surface water and groundwater storage. The surface water storage change data is derived from the GLDAS assimilation model, including soil water content, canopy water, and snow water equivalent data. The changes in terrestrial water storage derived from GRACE satellite data and the surface water storage changes obtained from the GLDAS assimilation model are further compared using a water balance formula to derive groundwater storage changes. The formula is as follows:
[0038] ;
[0039] in For changes in groundwater storage, For changes in snow water equivalent, For changes in soil water storage, This represents changes in canopy water storage.
[0040] Furthermore, during the training process of various machine learning models, folded cross-validation is used to divide the training data into K subsets and perform K iterations of training and validation.
[0041] Specifically, during model training, a machine learning model is used to perform downscaling learning based on the scaling factor data. Cross-validation is used to ensure the objectivity of model optimization and the accuracy of data;
[0042] Machine learning models include random forest models, support vector machine models, and multiple linear regression models. These models are used to downscale changes in groundwater storage. To avoid overfitting and performance distortion when downscaling the same dataset using machine learning models, a [specific approach is needed]. Folded cross-validation, this strategy divides the training data into... A subset, to perform This process involves multiple iterations of training and validation to ensure that each sample is used for both training and validation, thereby making full use of limited data and obtaining a more robust assessment of the model's generalization ability. Finally, by integrating the outputs of these cross-validated and optimized models, a reliable high-resolution groundwater storage change data product is generated. This involves processing the data... In each iteration China: The first Fold as the validation set, and the remaining The data were folded and merged to form the training set.
[0043] Furthermore, a ridge regression fusion model incorporating environmental factors is used to analyze and obtain the weights of various machine learning models, including:
[0044] Multiple machine learning models are used to predict the training samples, and the predicted value corresponding to each model is obtained.
[0045] The environmental factors are fused with the predicted values and used as input to the ridge regression fusion model for weight analysis, thereby solving for the weights of various machine learning models.
[0046] Specifically, a ridge regression fusion model is employed, which overcomes the collinearity interference of multiple machine learning models through its unique regularization mechanism. Environmental factors are incorporated, and the influencing factors are fused into the model, with model weights dynamically allocated. This approach, which integrates multiple machine learning models with environmental factors, leverages their complementary strengths to generate high-precision data on groundwater storage changes.
[0047] Obtain the prediction results of the base models. Using multiple pre-trained base machine learning models, predict the training samples to obtain the prediction value for each model. ,in The number of base models.
[0048] Environmental factors are introduced as weighting modifiers. Environmental factors with high spatial resolution closely related to groundwater changes, particularly drought indices and population size, are collected. These environmental factors are then standardized at the pixel scale. This is based on relevant regional data, including but not limited to regional arid zones, regional drought indices, regional population size, and its distribution.
[0049] A ridge regression fusion model is constructed. Predictions from multiple base models are used as initial features, and environmental factors are fused with the predictions of each base model as input to the ridge regression. This model aims to learn an optimal combination of weights that is not only a function of the predictions from various machine learning models but also a function of environmental conditions.
[0050] The ridge regression fusion model incorporating environmental factors was used to perform weight analysis on multiple machine learning models, and the final result is as follows:
[0051] ;
[0052] in, For the final downscaling data results, For the first The weights of a machine learning model and Drought index and population size for the first Weight adjustment coefficients for data in a machine learning model It is a drought index. It refers to population size. It is the first The predicted value of a machine learning model.
[0053] Solve for the parameters in the prediction model , , This is achieved by minimizing the following loss function that includes an L2 regularization term:
[0054] ;
[0055] in This represents the true value observed by GRACE. It is a regularization parameter, which can be determined through cross-validation.
[0056] Furthermore, after obtaining high-resolution values of groundwater storage changes, the following are included:
[0057] The high-resolution groundwater storage change values were verified. Based on the verification results, the parameters and structure of the ridge regression fusion model were optimized. When there were abnormally abrupt grid data in the ridge regression fusion model, the ridge regression fusion model was re-corrected by adding groundwater well data to the weights. If groundwater well data existed in the grid data, the groundwater well data was identified. If groundwater well data did not exist in the grid data, the grid data was corrected using the inverse distance weighted interpolation method.
[0058] Specifically, the downscaled high-resolution product is validated using independent observation data and hydrological model simulation results. Based on the validation results, the model parameters and structure are optimized. Groundwater well data are used to intervene at anomalous abrupt change points, ultimately generating reliable high-resolution data on groundwater storage changes.
[0059] The accuracy of the results generated by the new model is analyzed by comparing them with the observation data and the simulation results of the hydrological model. The reliability of the results generated by the new model is determined by relying on the root mean square error (RMSE) and the correlation coefficient (r).
[0060] ;
[0061] ;
[0062] When the overall result error is large, adjust the downscaling factor and weights. and the weighting adjustment coefficients of environmental factors and This reduces data errors in the new model. By performing targeted iterative optimization of the model parameters and structure, a self-improving analytical loop is formed.
[0063] When the model corresponding to formula (3) of the ridge regression fusion model that integrates environmental factors has a certain abnormal mutation in grid data, groundwater well data is used for intervention. By adding groundwater well data to the weights, the model is re-corrected to reach a suitable range. The correction result is as follows:
[0064] ;
[0065] For groundwater well data for the first Weight adjustment coefficients for data in a machine learning model It's data from groundwater wells.
[0066] The modified regularization loss function is:
[0067] (8);
[0068] Because groundwater well data is point data and its distribution is relatively scattered, while GRACE data is grid data, which is area data, two situations arise: groundwater well data exists within the grid data, and groundwater well data does not exist within the grid data.
[0069] If the grid contains groundwater well data, then the groundwater well data is directly determined:
[0070] ;
[0071] This represents the average value of groundwater well data within the grid data. The mean value of groundwater well data within the study area. The standard deviation of groundwater well data within the study area.
[0072] If no groundwater well data is found in the grid data, the grid data is corrected using inverse distance weighted interpolation.
[0073] ;
[0074] This represents the number of groundwater wells located around the grid data. For the first Data from groundwater wells, For the first Distance of groundwater well data This represents the distance attenuation coefficient.
[0075] Groundwater well data is used to eliminate abnormal and abrupt data in the new model, i.e., the model corresponding to formula (7). The parameter is determined and then integrated into the final model formula (formula (7) and formula (8)) to complete the repair of the data in that area.
[0076] By verifying and correcting the data, the generation of abnormal and sudden grid data is effectively reduced. This enables the model to ultimately form a set of groundwater storage change data products with multi-source verification, algorithm optimization, and high spatial resolution and high accuracy, providing reliable data support for refined water resource management.
[0077] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A downscaling method for groundwater storage changes retrieved from the GRACE satellite using multiple machine learning models that integrate environmental factors, characterized in that... include: Collect GRACE satellite data and retrieve changes in terrestrial water storage based on the GRACE satellite data. Based on the changes in terrestrial water storage, the changes in groundwater storage are obtained; Multiple machine learning models are used to downscale the changes in groundwater storage to obtain high-resolution values of groundwater storage changes. The weights of the various machine learning models are obtained by analyzing a ridge regression fusion model that incorporates environmental factors.
2. The method according to claim 1, characterized in that, Based on the changes in terrestrial water storage, the changes in groundwater storage are obtained by: obtaining the changes in terrestrial water storage from GRACE satellite data inversion and the changes in surface water storage from the GLDAS assimilation model, and then obtaining the changes in groundwater storage through the water balance formula.
3. The method according to claim 1, characterized in that, In the training process of various machine learning models, folded cross-validation is used to divide the training data into K subsets and perform K iterations of training and validation.
4. The method according to claim 1, characterized in that, The weights of various machine learning models are obtained by using a ridge regression fusion model that incorporates environmental factors, including: Multiple machine learning models are used to predict the training samples, and the predicted value corresponding to each model is obtained. The environmental factors are fused with the predicted values and used as input to the ridge regression fusion model for weight analysis, thereby solving for the weights of various machine learning models.
5. The method according to claim 1, characterized in that, The ridge regression fusion model is as follows: ； in, For the final downscaling data results, For the first The weights of a machine learning model and Drought index and population size for the first Weight adjustment coefficients for data in a machine learning model It is a drought index. It refers to population size. It is the first The predicted values of a machine learning model, where M is the number of base models.
6. The method according to claim 5, characterized in that, Solving for the weights of various machine learning models involves minimizing the following loss function, which includes an L2 regularization term: ； in, This represents the true value observed by GRACE. It is a regularization parameter. These are the parameters in the prediction model.
7. The method according to claim 1, characterized in that, After obtaining high-resolution groundwater storage change values, the following are included: The high-resolution groundwater storage change values were verified. Based on the verification results, the parameters and structure of the ridge regression fusion model were optimized. When there were abnormally abrupt grid data in the ridge regression fusion model, the ridge regression fusion model was re-corrected by adding groundwater well data to the weights. If groundwater well data existed in the grid data, the groundwater well data was identified. If groundwater well data did not exist in the grid data, the grid data was corrected using the inverse distance weighted interpolation method.