Shallow groundwater storage variable signal separation method based on spatio-temporal graph convolution network
Patent Information
- Application Number
- CN202611472205.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-21
- Publication Date
- 2026-10-30
AI Technical Summary
该方法虽可拆分出多组时序与空间分量，但分解结果完全依赖数学统计规则，未结合地下水实测数据校验与真实水文过程机理约束，导致分离得到的信号缺乏明确的水文物理意义
（1）本发明针对GRACE/GRACE-FO观测结果包含不同深度地下水储量变化叠加的问题，引入潜水含水层地下水位观测作为监督约束，结合时空图卷积网络综合利用地下水储变量、辅助环境变量及空间关联信息，实现浅层地下水储变量信号的有效分离。
Smart Images
Figure CN123012544A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the interdisciplinary fields of satellite remote sensing hydrological monitoring, quantitative hydrogeological inversion, and spatiotemporal deep learning, specifically a method for separating shallow groundwater storage variable signals based on spatiotemporal graph convolutional networks. Background Technology
[0002] Shallow groundwater is a core water source for maintaining regional vegetation growth and supporting ecological restoration. Accurately characterizing the spatiotemporal evolution of shallow groundwater is of significant scientific value for quantifying groundwater consumption characteristics during ecological restoration and revealing the coupling mechanism between vegetation recovery and groundwater dynamics. The GRACE / GRACE-FO gravity satellite can capture temporal variations in the Earth's gravity field and retrieve regional-scale, long-term, spatially continuous groundwater storage anomaly (GWSA) data, making it a core technology for large-scale groundwater storage dynamic monitoring. However, the groundwater storage change signals retrieved by this satellite are a comprehensive response result of the dynamic superposition of shallow and deep groundwater, and cannot be directly separated to obtain independent shallow groundwater change information.
[0003] On the one hand, the inherent observation characteristics of the GRACE / GRACE-FO satellite have limitations. Its inversion results can only reflect the overall changes of the superposition of deep and shallow groundwater, and cannot specifically characterize the dynamic characteristics of shallow groundwater that are closely related to key eco-hydrological processes such as precipitation recharge, vegetation transpiration, and soil evaporation. This makes it difficult to meet the refined needs of eco-hydrological research.
[0004] On the other hand, traditional blind source decomposition methods rely on mathematical criteria such as statistical independence, non-Gaussianity, and orthogonality of signals as the core basis for decomposition, achieving signal separation without the constraints of measured data and prior hydrological knowledge. Although this method can decompose multiple sets of temporal and spatial components, the decomposition results depend entirely on mathematical statistical rules, without combining verification with measured groundwater data and constraints from real hydrological process mechanisms, resulting in the separated signals lacking clear hydrological and physical meaning. Furthermore, various hydrological components, such as shallow groundwater, deep groundwater, lateral runoff recharge, artificial groundwater extraction, and model residuals, exhibit significant spatiotemporal superposition effects. Relying solely on blind source signal decomposition easily leads to signal aliasing, ambiguous spatiotemporal scales, and non-unique physical interpretations, ultimately resulting in low accuracy and poor stability in identifying changes in shallow groundwater storage.
[0005] To address the aforementioned technical problems of traditional methods, this paper proposes a method for separating shallow groundwater storage change signals based on spatiotemporal graph convolutional networks, aiming to achieve accurate and reliable inversion of shallow groundwater dynamics. Summary of the Invention
[0006] 1. The technical problem to be solved by the present invention
[0007] The purpose of this invention is to address the challenge of accurately separating shallow groundwater storage variable signals at the regional scale in GRACE / GRACE-FO groundwater storage variable inversion, where shallow and deep storage changes are superimposed. Existing blind source signal decomposition methods struggle to accurately separate these regional-scale shallow groundwater storage variable signals. This invention proposes a shallow groundwater storage variable signal separation method based on a spatiotemporal graph convolutional network. This method comprehensively utilizes GRACE / GRACE-FO gravity satellite, meteorological and vegetation factors, surface factors, and well observation data to construct a shallow groundwater storage variable separation model with spatial correlation and time series learning capabilities, thereby achieving the separation of shallow groundwater storage change signals at the regional scale.
[0008] 2. Technical Solution To achieve the above objectives, the present invention provides the following technical solution: a method for separating shallow groundwater storage variable signals based on spatiotemporal graph convolutional networks, comprising the following steps: S1. Multi-source data construction and random forest downscaling: Multi-source remote sensing data are fused and the random forest algorithm is used to perform spatial downscaling of groundwater storage variables to obtain a high-resolution downscaled groundwater storage variable sequence. S2, K-means clustering spatial partitioning: Based on downscaled groundwater storage variables, temporal features are extracted and regional groundwater response partitioning is completed using K-means clustering; S3. Construction of the Spatiotemporal Graph Convolutional Network Model STGCN: A weighted spatial graph structure is constructed using raster pixels in each partition as graph nodes to build a spatiotemporal graph convolutional network model that integrates temporal convolution and graph convolution. S4. Multi-constraint loss function training model: shallow groundwater monitoring labels are constructed using measured water levels in shallow wells and scale correction is completed. A multi-constraint composite loss function is introduced to iteratively train the spatiotemporal graph convolutional network model. S5. Separation of shallow groundwater storage variables: Input the complete time-series high-resolution groundwater storage variables into the trained model, output the shallow groundwater storage variable signal, and solve the difference to obtain the deep groundwater remaining storage variable signal.
[0009] Preferably, S1 specifically includes: S1.1 Collect GRACE / GRACE-FO Total Water Storage Anomaly (TWSA), GLDAS Land Surface Model Soil Water Storage Anomaly (SMA), Snow Water Equivalent Anomaly (SWEA), and Vegetation Canopy Water Equivalent (CWSA) data. Simultaneously collect multi-source auxiliary data including precipitation, air temperature, evapotranspiration, Normalized Difference Vegetation Index (NDVI), Digital Elevation Model (DEM), Land Use Data Center (LUCC), and measured groundwater level. S1.2. By subtracting interference terms related to soil, snow cover, and vegetation water storage, the low-resolution original total groundwater storage variable GWSA is obtained; the specific expression is: ； S1.3 Construct a random forest nonlinear mapping model, and input the 1km high-resolution meteorological, vegetation and topographic auxiliary variables into the model to obtain the preliminary high-resolution groundwater storage variable fitting results; S1.4 Calculate the residuals between the low-resolution original GWSA and the low-resolution fitted values of the random forest, and spatially interpolate them to a high-resolution grid. Overlay the fitting results to obtain the downscaled groundwater storage variable DSGWSA. The calculation formula is as follows:
[0010]
[0011] In the formula, It is a random forest model. Represents a high-resolution set of auxiliary variables. This represents a low-resolution set of groundwater storage variables. This represents the residual correction term at the original 0.25° spatial scale; i Indicates spatial location, t Indicates time.
[0012] Preferably, S2 specifically includes: S2.1 Calculate the average groundwater storage variable of all pixels within the zoning region r. Calculate the relative anomaly sequence of groundwater storage variables per pixel The The specific formula is as follows:
[0013]
[0014] In the formula, This represents the average groundwater storage variable of region r at time t; Represents the set of downscaled groundwater storage variables; This indicates the number of valid pixels within region r; Let t represent the total set of groundwater storage variables at time t.
[0015] S2.2 Extracting multidimensional clustering feature vectors based on relative anomaly sequences, including trend, seasonal amplitude, seasonal phase, short-term fluctuation intensity, amplitude change trend, and pixel plane coordinates. The vector expression is:
[0016] In the formula, Represents the cluster feature vector; Indicates relative trends of change; Indicates seasonal amplitude; Indicates seasonal phases, Indicates the intensity of short-term fluctuations; Indicates the trend of amplitude change; , Represents the plane coordinates of a pixel.
[0017] S2.3. After standardizing the feature vectors, input them into the K-means clustering model, set the number of clusters K to complete the partitioning, and obtain the partition to which the i-th pixel belongs. The rationality of clustering partitions is evaluated using silhouette coefficients, and different groundwater response partitions are modeled independently; The expression is:
[0018] In the formula, k Indicates the number of clusters.
[0019] Preferably, S3 specifically includes: S3.1. Using each raster cell within the partition as a graph node, calculate the graph edge weights between nodes by combining node spatial distance, elevation difference, and land use matching weight. The weight The calculation formula is:
[0020] In the formula, Let i be the spatial distance between nodes i and j. , These are distance and elevation scale parameters, respectively. Weights are assigned to land use types; and All are represented as digital elevations; S3.2 Construct the original adjacency matrix A based on the graph edge weights, and obtain the normalized adjacency matrix using symmetric normalization. The The specific expression is:
[0021] In the formula, D is the graph node degree matrix, and its diagonal elements are the sum of the connection weights of the corresponding nodes in the adjacency matrix A; A is the adjacency matrix. This is the normalized adjacency matrix.
[0022] S3.3 Constructing a spatiotemporal graph convolutional network with input features By integrating downscaled groundwater storage variables, meteorological vegetation variables, and monthly cycle sine and cosine coding, the aforementioned The specific expression is:
[0023] In the formula, m is the month number; , , , These are respectively represented as precipitation, temperature, evapotranspiration, and normalized vegetation index; S3.4 The spatiotemporal graph convolutional network consists of a feature normalization module, a temporal convolutional layer, a graph convolutional layer, and a hidden layer decoder connected in series, taking the input features as input. With normalized adjacency matrix As input, through model parameters Forward operation outputs estimated values of shallow groundwater storage variables. The The expression is
[0024] In the formula, [] represents the mapping function of the spatiotemporal graph convolutional network.
[0025] Preferably, S4 specifically includes: S4.1 Measured water level in the submersible well Anomalies and standardization transformations are performed, and original monitoring labels are constructed by combining the standard deviation of groundwater storage variables corresponding to well points in the grid. Learnable scaling coefficients are then introduced. Bias terms Correcting dimensional differences yields the corrected supervision label. The The specific expression is:
[0026] In the formula, This represents the standard deviation of the groundwater level sequence for the j-th unconfined well. This represents the standard deviation of the grid-downscaled groundwater reservoir variable series corresponding to this well; This represents the average groundwater level during the observation period of the j-th submerged well.
[0027] S4.2 Construct a composite multi-constraint loss function Loss, which includes a label fitting term, a time-series trend consistency term, a time smoothing term, a frequency domain suppression term, and a scale parameter regularization term: The expression for the composite multi-constraint loss function Loss is:
[0028] In the formula, This is a variable signal for deep residual groundwater storage; This represents the estimated shallow groundwater storage variables output by the spatiotemporal graph convolutional network. This represents a monitoring label for shallow groundwater reservoir variables constructed from water levels observed in submerged wells; Time difference series representing estimates of shallow groundwater storage variables; Time-difference sequences representing deep residual groundwater reservoir variable signals; Represents the frequency domain weighting function; , , These represent the weight coefficients of the trend consistency term, time smoothing term, and frequency domain constraint term, respectively. It is represented as the weight of the regularization term of the scaling parameter.
[0029] S4.3 Perform power spectrum analysis on the difference series of regional average groundwater storage variables to identify seasonal cycles. Interannual cycle Calculate the frequency domain cutoff period Cutoff frequency number Construct a frequency domain weighting function M is the total timing length; the frequency domain cutoff period Cutoff frequency number Frequency domain weighting function The expression is:
[0030]
[0031]
[0032] S4.4 Dynamically update the weights of each loss term based on the NTK gradient norm of the neural tangent kernel. Update the formula:
[0033] In the formula Fit the loss gradient to the label. The gradient of the loss term to be adjusted. To minimize the constant and prevent the denominator from being zero, For weighted amplitude limiting operators; S4.5. Employ gradient descent method to jointly optimize the parameters of the spatiotemporal graph convolutional network. Scale correction parameters , The model convergence training is completed.
[0034] Preferably, S5 specifically includes: S5.1. Input the full-time high-resolution DSGWSA and the corresponding auxiliary variables into the training spatiotemporal graph convolutional network to output the whole area grid-by-grid shallow groundwater storage variable SGWSA. S5.2. The deep residual groundwater storage variable DRGWSA is obtained through differential calculation. The calculation formula is as follows:
[0035] As a preferred option, an accuracy verification step is also included: Data from CSRRL06Mascon0.25°GRACE / GRACE-FO, GLDASNoah land surface hydrological data, 1km monthly meteorological vegetation data, 30m DEM and land use data, and monthly groundwater level monitoring data from regional groundwater wells were collected as input data sources. The well point data were divided into training set and independent validation set in a 7:3 ratio. The Pearson correlation coefficient R and Nash efficiency coefficient NSE were used to evaluate the matching accuracy between the separated shallow groundwater storage variables and the measured groundwater level.
[0036] As a preferred approach, K-means segmentation and regional modeling are used to eliminate interference from dynamic differences in groundwater in different hydrological units; spatial weighted map structure is integrated with topographic and land use constraints to achieve grid-based spatial hydrological correlation modeling; measured unconfined well monitoring labels and adaptive frequency domain constraints are used to achieve physically interpretable separation of shallow and deep groundwater signals, thus solving the defects of signal aliasing and unclear hydrological and physical meaning in traditional blind source decomposition methods.
[0037] Compared with existing technologies, the shallow groundwater storage variable signal separation method based on spatiotemporal graph convolutional networks provided by this invention has the following beneficial effects: (1) In view of the problem that the GRACE / GRACE-FO observation results contain superimposed changes in groundwater storage at different depths, this invention introduces groundwater level observation of unconfined aquifers as a supervisory constraint, and combines spatiotemporal graph convolutional network to comprehensively utilize groundwater storage variables, auxiliary environmental variables and spatial correlation information to achieve effective separation of shallow groundwater storage variable signals.
[0038] (2) To address the issue of spatial heterogeneity in the variation characteristics of groundwater storage variables in different regions, the K-means spatial partitioning method is used to identify the temporal variation differences of groundwater storage variables within a region. Furthermore, a graph structure with spatial correlation expression capability is constructed by combining a spatiotemporal graph convolutional network, enabling the model to perform differentiated learning for the variation characteristics of groundwater storage variables in different regions and reducing the impact of signal mixing in different regions on the separation results of shallow groundwater storage variables.
[0039] (3) Learnable scale correction parameters are introduced to realize the conversion between supervision labels and groundwater storage variables; at the same time, adaptive frequency domain constraints and multi-constraint loss functions are combined to constrain the model optimization process from both frequency domain characteristics and time series change characteristics, thereby improving the stability and physical rationality of the model separation results. Attached Figure Description
[0040] Figure 1 This is a schematic diagram of the process for separating shallow groundwater storage variable signals based on spatiotemporal graph convolutional networks according to the present invention. Figure 2This is a schematic diagram of the model structure of the present invention; Figure 3 This is a schematic diagram comparing the time series of the original GWSA and the separated shallow and deep GWSA of this invention; Detailed Implementation 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0041] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.
[0042] Example 1, please refer to Figures 1 to 2 As shown: To address the problems mentioned in the technical solutions, this application provides a method for separating shallow groundwater storage variable signals based on spatiotemporal graph convolutional networks, as detailed below: Part 1: Multi-Source Data Construction and Random Forest Downscaling First, obtain the GRACE / GRACE-FO total land water storage anomaly TWSA and the soil water storage anomaly output by the hydrological model. Abnormal snow water equivalent ( ), vegetation canopy water equivalent ( ), and precipitation ( ), temperature ( ), evaporation ( ), vegetation index ( ), elevation ( ), land use type ( Data such as the measured groundwater level H were obtained by subtracting from TWSA. , and The total groundwater storage variable GWSA at time t is obtained, and the calculation process is shown below: .
[0043] Due to the low spatial resolution of the GRACE / GRACE-FO products, a random forest model is used. Spatial downscaling of groundwater storage variables. At a low-resolution scale, the value of the i-th pixel at time t is used. and As an auxiliary variable, it was established to correlate with low-resolution groundwater storage variables. The nonlinear mapping relationship between them. Subsequently, a high-resolution set of auxiliary variables... Input the trained random forest model Preliminary downscaling results were obtained. To maintain consistency between the downscaling results and the original GRACE / GRACE-FO regional total constraints, the original groundwater storage variables at the low-resolution scale were... The residuals are obtained by subtracting the fitted values from the random forest values. These residuals are then spatially interpolated to a high-resolution scale to obtain the downscaled groundwater storage anomaly. ),in, DSGWSA The specific process is as follows:
[0044]
[0045] In the formula, It is a random forest model. Represents a high-resolution set of auxiliary variables. This represents a low-resolution set of groundwater storage variables. This represents the residual correction term at the original 0.25° spatial scale; i Indicates spatial location, t Indicates time.
[0046] Part Two: Spatial Partitioning of K-means Clustering To reduce the impact of differences in groundwater variation characteristics across different regions on shallow signal separation results, K-means clustering spatial partitioning is performed based on downscaled groundwater storage variables. First, the average groundwater storage variable of region r at time t is calculated. ,in This represents the number of valid cells within region r. Then, for each cell... minus To obtain the relative anomaly sequence The specific process is as follows:
[0047]
[0048] In the formula, This represents the average groundwater storage variable of region r at time t; Represents the set of downscaled groundwater storage variables; This indicates the number of valid pixels within region r; Let t represent the total set of groundwater storage variables at time t.
[0049] Subsequently, based on Extract the clustering feature vector of the i-th pixel. . Including relative trends Seasonal amplitude Seasonal phases Short-term volatility Amplitude variation trend and pixel space coordinates and .in, The specific calculation process is as follows:
[0050] In the formula, Represents the cluster feature vector; Indicates relative trends of change; Indicates seasonal amplitude; Indicates seasonal phases, Indicates the intensity of short-term fluctuations; Indicates the trend of amplitude change; , Represents the plane coordinates of a pixel.
[0051] For all After standardization, the data is input into the K-means model, and the number of clusters K is set to obtain the partition to which the i-th pixel belongs. The specific calculation process is as follows:
[0052] In the formula, k Indicates the number of clusters.
[0053] The rationality of the partitioning was verified by calculating the silhouette coefficients using different clustering methods. Finally, the partitioning results were used for subsequent STGCN partitioning modeling.
[0054] Part 3: Construction of Spatiotemporal Graph Convolutional Network Model The spatiotemporal graph convolutional network model structure used in this invention is shown in the attached figure. Figure 2 As shown, the model consists of an input feature module, a feature standardization module, a temporal convolutional layer, a graph convolutional layer, hidden layers and a decoder, a shallow label construction and scale correction module, and a model output module. Specifically, the input features, after standardization, enter the temporal convolutional layer to extract temporal dynamic features; the graph structure construction results enter the graph convolutional layer to extract spatial correlation features; and the groundwater level in the unconfined well, after shallow label construction and scale correction, provides supervised constraints for the shallow groundwater storage variable output by the model.
[0055] This invention constructs a spatiotemporal graph convolutional network model within each K-means partition. Each pixel within a partition is considered a graph node, and the spatial distance between node i and node j is... Digital Elevation and Weight of the same land use type Construct graph edge weights together .in The specific calculation process is as follows:
[0056] in, This represents a spatial distance scale parameter. This represents the elevation difference scale parameter. The map structure is used to constrain information transfer between spatially adjacent pixels with similar surface and topographic conditions. After constructing the adjacency matrix A, a symmetric normalization method is used to obtain the normalized adjacency matrix. As shown below:
[0057] In the formula, D is the graph node degree matrix, and its diagonal elements are the sum of the connection weights of the corresponding nodes in the adjacency matrix A; A is the adjacency matrix. This is the normalized adjacency matrix.
[0058] Input features of spatiotemporal graph convolutional networks Including the i-th pixel at time t , , , , and the seasonal periodic term constructed from month m. and The specific expression is as follows:
[0059] in, The model provides constraints on total storage changes, with meteorological and vegetation variables reflecting the response of shallow groundwater to precipitation recharge, evapotranspiration, and vegetation water consumption. A monthly periodic term is used to express seasonal variation characteristics. and As input, through model parameters Perform training and output the estimated value of the shallow groundwater storage variable for the i-th pixel at time t. The specific expression is as follows:
[0060] In the formula, [] represents the mapping function of the spatiotemporal graph convolutional network.
[0061] To make the model output With shallow groundwater observation constraints, groundwater level is introduced into unconfined wells. Construct shallow monitoring labels. For the j-th unconfined well, first calculate its average water level. and water level standard deviation Then, combined with the standard deviation of the downscaled groundwater storage variable series of the pixel where the well is located... Converting water level changes into shallow groundwater storage variables for monitoring and labeling. Considering the water level in the submersible well The shallow groundwater storage variables differ in dimensions, amplitude, and baseline, so a learnable scaling coefficient is further set. and learnable bias terms ,right The correction is performed to obtain the corrected supervision label. The specific expression is as follows:
[0062] In the formula, This represents the standard deviation of the groundwater level sequence for the j-th unconfined well. This represents the standard deviation of the grid-downscaled groundwater reservoir variable series corresponding to this well; This represents the average groundwater level during the observation period of the j-th submerged well.
[0063] During the training, With STGCN model parameters They jointly optimize to adaptively correct the scale difference between water level labels and shallow reservoir signals.
[0064] Through the aforementioned spatial graph structure constraints, total reserves input constraints, unmanned well monitoring constraints, and scale correction constraints, the model is able to learn the spatial continuous distribution and temporal dynamic changes of shallow groundwater storage variables under the constraints of the GRACE / GRACE-FO total groundwater storage variables.
[0065] Part 4: Model Training and Separation of Shallow Groundwater Storage Variables To improve the stability and physical plausibility of the model's separation results, multiple loss functions were constructed. These functions simultaneously considered factors such as subsurface well label fitting, consistency of trends, temporal smoothness, frequency domain constraints, and scale parameter regularization. The shallow label fitting term was used to constrain the estimated shallow groundwater reservoir variables output by the model. Approaching the calibrated submersible well monitoring label The trend consistency term is used to constrain the difference between adjacent months in the model output. Difference from supervised label The change direction remains consistent; the time smoothing term is used to suppress unreasonable abrupt changes in shallow groundwater storage variables; the frequency domain constraint term is used to suppress high-frequency fluctuations in deep residual signals, ensuring that shallow signals primarily characterize groundwater changes that respond quickly to precipitation, evapotranspiration, and vegetation changes; the scale parameter regularization term is used to avoid learnable scale coefficients. and bias terms The deviation is too large. The expression for the composite multi-constraint loss function Loss is as follows:
[0066] In the formula, This is a variable signal for deep residual groundwater storage; This represents the estimated shallow groundwater storage variables output by the spatiotemporal graph convolutional network. This represents a monitoring label for shallow groundwater reservoir variables constructed from water levels observed in submerged wells; Time difference series representing estimates of shallow groundwater storage variables; Time-difference sequences representing deep residual groundwater reservoir variable signals; Represents the frequency domain weighting function; , , These represent the weight coefficients of the trend consistency term, time smoothing term, and frequency domain constraint term, respectively. It is represented as the weight of the regularization term of the scaling parameter.
[0067] Among them, deep signals Downscaled total groundwater storage variables and estimated shallow groundwater storage variables from the model output Subtraction yields, In the frequency domain constraint, a Fast Fourier Transform is first performed on the difference sequence of the total groundwater storage variable with regional averages to identify seasonal and interannual peaks and determine the frequency domain cutoff parameter. Subsequently, ... Difference sequences Perform a Fast Fourier Transform (FFT) and utilize a frequency domain weighting function. Constraining the high-frequency components allows the deep signals to primarily retain relatively slow-changing storage changes, while the shallow signals better characterize groundwater dynamics that respond more rapidly to precipitation, evapotranspiration, and vegetation changes.
[0068] The present invention further considers the average groundwater storage variable of each partition z. Constructing difference sequences Power spectrum analysis was performed on the data. Seasonal peak periods were identified based on the power spectrum. and interannual peak cycle Calculate the frequency domain cutoff period and frequency domain cutoff frequency number Based on this, a frequency domain weighting function is constructed. The specific expression is as follows:
[0069] During training, an adaptive weight adjustment strategy based on gradient norm or neural tangent kernel is adopted to dynamically adjust the loss term. weight An update will be performed. Among other things, Indicates the label fit term on the model parameters gradient, This indicates the effect of the loss term to be adjusted on the model parameters. gradient, To prevent extremely small constants with a denominator of zero, This means limiting the weights to a preset range, as shown in the following expression:
[0070] In the formula Fit the loss gradient to the label. The gradient of the loss term to be adjusted. To minimize the constant and prevent the denominator from being zero, For weighted amplitude limiting operators; After the model training is complete, the complete time series will be... The shallow groundwater storage anomaly (SGWSA) was obtained by inputting auxiliary variables into the trained STGCN model, thus obtaining the shallow groundwater storage anomaly for the entire study period. Subsequently, the shallow groundwater storage anomaly was transferred from... After subtracting from the data, the Deep Residual Groundwater Storage Anomaly (DRGWSA) is obtained, and its specific expression is shown below.
[0071] DRGWSA is not equivalent to directly observed changes in deep aquifer reserves, but rather a residual reserve signal relative to the shallow component.
[0072] Example 2: Based on Embodiment 1, but with some differences, the method for separating shallow groundwater storage variable signals based on spatiotemporal graph convolutional networks proposed in this invention will be described below with reference to specific examples and accompanying drawings. The specific content is as follows: This invention proposes a method for separating shallow groundwater storage variables based on spatiotemporal graph convolutional networks. This method comprehensively utilizes GRACE / GRACE-FO gravity satellite data, meteorological vegetation factors, DEM data, land use type data, and well observation data to construct a shallow groundwater storage variable separation model with spatial correlation and time series learning capabilities, thereby achieving regional-scale separation of shallow groundwater storage change signals. Taking two example areas in Inner Mongolia, downscaled GWSA, precipitation, NDVI, temperature, evapotranspiration, DEM, land use type data, and groundwater level data are used as input data to separate the total GWSA into shallow and deep groundwater storage variables. The method specifically includes three parts: data processing, model training, and result validation.
[0073] The first part is data processing.
[0074] First, the original GWSA was calculated based on the GRACE / GRACE-FO terrestrial water storage anomaly, and spatial downscaling was performed using random forest. Subsequently, temporal matching and standardization were performed on various monthly-scale data. K-means partitioning was performed based on the variation characteristics of pixels relative to the regional average GWSA. Example 1 was divided into two partitions; Example 2 was divided into three partitions. The unconfined groundwater level, after anomaly reduction, standardization, and scale transformation, was used as the shallow GWSA supervision label. 70% of the well points within each partition were used for model training, and the remaining 30% were used for independent validation.
[0075] (1) GRACE / GRACE-FO gravity satellite data: The CSRRL06Mascon monthly groundwater storage anomaly data released by the Space Research Center at the University of Texas, with a spatial resolution of 0.25° and a time range from April 2002 to December 2024, is used to provide information on regional groundwater storage changes. The data is from the Space Research Center at the University of Texas (website: [website address missing]). https: / / www2.csr.utexas.edu / grace / RL06_ mascons.html).
[0076] (2) GLDASNoah land surface model data: Data from the NASA-released GLDASNoah v2.1 land surface model, including land surface hydrological variables such as soil moisture content, canopy water, and snow cover equivalent, were used. The spatial resolution was 0.25°, the temporal resolution was monthly, and the time range was consistent with the study period. The data came from NASA's Goddard Earth Science Data and Information Service (website: [website address missing]). (https: / / www.earthdata.nasa.gov / centers / gesdisc-daac).
[0077] (3) Precipitation, temperature, and evapotranspiration data: Precipitation, temperature, and evapotranspiration data with a spatial resolution of 1 km and a temporal resolution of monthly scale were used as input variables to construct the groundwater storage variable downscaling model. Data were sourced from the National Tibetan Plateau Scientific Data Center (website: [website address missing]). https: / / www.tpdc.ac.cn / home).
[0078] (4) DEM: Digital elevation model data with a spatial resolution of 30m was used to characterize the topographic conditions of the study area and as an input auxiliary variable for the model. The data was sourced from the Geospatial Data Cloud (website: https: / / www.gscloud.cn / ).
[0079] (5) Land use types: Land use classification data with a spatial resolution of 30m was used to describe the land cover characteristics of the study area and to participate in the construction of auxiliary variables for the model. The data came from the Wuhan University public dataset (website: [website address missing]). (https: / / doi.org / 10.5281 / zenodo.5816591).
[0080] (6) Groundwater level data: Monthly groundwater level monitoring well data of unconfined aquifers in two example areas were collected, including the spatial location of the monitoring wells, the observation time and the corresponding groundwater level value, which were used to construct a supervision label for shallow groundwater storage variables and to carry out model training and verification.
[0081] (7) GWSA downscaling: The random forest method was used, with precipitation, temperature and NDVI as auxiliary variables, to establish a nonlinear relationship between the original GWSA and high-resolution environmental factors, and obtain high-resolution GWSA. This provides input data for subsequent K-means partitioning and STGCN signal separation.
[0082] (8) K-means partitioning: The K-means algorithm was used to cluster the GWSA sequences after removing the regional average change. Instance 1 was divided into two partitions (relatively stable region and relatively decreasing region), and instance 2 was divided into three partitions (relatively stable region, moderate response region, and accelerated decreasing region). The partitioning results were used to construct and train the STGCN model respectively.
[0083] The second part is model training.
[0084] STGCN models were trained for two instance regions separately. The models used effective raster pixels as graph nodes and determined node connection weights by integrating spatial distance, DEM elevation difference, and land use similarity. Model inputs included GWSA, precipitation, NDVI, temperature, evapotranspiration, and monthly sine and cosine codes. The loss function included Huber fitting loss, trend consistency loss, temporal smoothing constraint, FFT frequency domain constraint, and label scaling correction term. The weights of the frequency domain constraint and temporal smoothing constraint were dynamically adjusted using NTK gradient information; specific parameters are shown in Table 1.
[0085] Table 1 Model Parameter Settings
[0086] The third part is accuracy verification.
[0087] The shallow GWSA obtained by separating the original GWSA from the model was compared and validated using the Pearson correlation coefficient (R) and Nash efficiency coefficient (NSE). (See Table 2 for details.) Table 2 Accuracy Comparison
[0088] The results in Table 2 show that the consistency between shallow GWSA and observations from unconfined wells is significantly higher than that of the original GWSA (Table 2). In the independent validation set, the average correlation coefficient of Example 1 increased from 0.084 to 0.590, and the average NSE increased from 0.076 to 0.420; the average correlation coefficient of Example 2 increased from 0.486 to 0.682, and the NSE increased from 0.341 to 0.531.
[0089] like Figure 3 As shown, the original GWSA and the separated shallow and deep GWSA time series are presented in two embodiments. The results indicate that the original GWSA is a composite signal resulting from the superposition of groundwater changes at different depths, making it difficult to directly characterize shallow groundwater changes. After separation using this method, the shallow GWSA exhibits more obvious seasonal fluctuations and short-term response characteristics, while the deep GWSA mainly shows long-term trends and low-frequency changes. This demonstrates that this method can effectively extract groundwater storage signals consistent with shallow unconfined groundwater dynamics.
[0090] Please refer to the above work process. Figures 1 to 3 It should be noted that the term "comprising" or any other variation thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0091] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for separating shallow groundwater storage variable signals based on spatiotemporal graph convolutional networks, characterized in that, Includes the following steps: S1. Multi-source data construction and random forest downscaling: Multi-source remote sensing data are fused and the random forest algorithm is used to perform spatial downscaling of groundwater storage variables to obtain a high-resolution downscaled groundwater storage variable sequence. S2, K-means clustering spatial partitioning: Based on downscaled groundwater storage variables, temporal features are extracted and regional groundwater response partitioning is completed using K-means clustering; S3. Construction of the Spatiotemporal Graph Convolutional Network Model STGCN: A weighted spatial graph structure is constructed using raster pixels in each partition as graph nodes to build a spatiotemporal graph convolutional network model that integrates temporal convolution and graph convolution. S4. Multi-constraint loss function training model: shallow groundwater monitoring labels are constructed using measured water levels in shallow wells and scale correction is completed. A multi-constraint composite loss function is introduced to iteratively train the spatiotemporal graph convolutional network model. S5. Separation of shallow groundwater storage variables: Input the complete time-series high-resolution groundwater storage variables into the trained model, output the shallow groundwater storage variable signal, and solve the difference to obtain the deep groundwater remaining storage variable signal.
2. The method for separating shallow groundwater storage variable signals based on spatiotemporal graph convolutional networks according to claim 1, characterized in that, S1 specifically includes: S1.1 Collect GRACE / GRACE-FO Total Water Storage Anomaly (TWSA), GLDAS Land Surface Model Soil Water Storage Anomaly (SMA), Snow Water Equivalent Anomaly (SWEA), and Vegetation Canopy Water Equivalent (CWSA) data. Simultaneously collect multi-source auxiliary data including precipitation, air temperature, evapotranspiration, Normalized Difference Vegetation Index (NDVI), Digital Elevation Model (DEM), Land Use Data Center (LUCC), and measured groundwater level. S1.
2. By subtracting interference terms related to soil, snow cover, and vegetation water storage, the low-resolution original total groundwater storage variable GWSA is obtained; the specific expression is: ； S1.3 Construct a random forest nonlinear mapping model, and input the 1km high-resolution meteorological, vegetation and topographic auxiliary variables into the model to obtain the preliminary high-resolution groundwater storage variable fitting results; S1.4 Calculate the residuals between the low-resolution original GWSA and the low-resolution fitted values of the random forest, and spatially interpolate them to a high-resolution grid. Stack the fitting results to obtain the downscaled groundwater storage variable. DSGWSA The calculation formula is: In the formula, It is a random forest model. Represents a high-resolution set of auxiliary variables. This represents a low-resolution set of groundwater storage variables. This represents the residual correction term at the original 0.25° spatial scale; i Indicates spatial location, t Indicates time.
3. The method for separating shallow groundwater storage variable signals based on spatiotemporal graph convolutional networks according to claim 1, characterized in that, S2 specifically includes: S2.1 Calculate the average groundwater storage variable of all pixels within the zoning region r. Calculate the relative anomaly sequence of groundwater storage variables per pixel The The specific formula is as follows: In the formula, This represents the average groundwater storage variable of region r at time t; Represents the set of downscaled groundwater storage variables; This indicates the number of valid pixels within region r; Represents the set of total groundwater storage variables at time t; S2.2 Extracting multidimensional clustering feature vectors based on relative anomaly sequences, including trend, seasonal amplitude, seasonal phase, short-term fluctuation intensity, amplitude change trend, and pixel plane coordinates. The vector expression is: In the formula, Represents the cluster feature vector; Indicates relative trends of change; Indicates seasonal amplitude; Indicates seasonal phases, Indicates the intensity of short-term fluctuations; Indicates the trend of amplitude change; , Represents the plane coordinates of a pixel; S2.
3. After standardizing the feature vectors, input them into the K-means clustering model, set the number of clusters K to complete the partitioning, and obtain the partition to which the i-th pixel belongs. The rationality of clustering partitions is evaluated using silhouette coefficients, and different groundwater response partitions are modeled independently; The expression is: In the formula, k Indicates the number of clusters.
4. The method for separating shallow groundwater storage variable signals based on spatiotemporal graph convolutional networks according to claim 1, characterized in that, S3 specifically includes: S3.
1. Using each raster cell within the partition as a graph node, calculate the graph edge weights between nodes by combining node spatial distance, elevation difference, and land use matching weight. The weight The calculation formula is: In the formula, Let i be the spatial distance between nodes i and j. , These are distance and elevation scale parameters, respectively. Weights are assigned to land use types; and All are represented as digital elevations; S3.2 Construct the original adjacency matrix A based on the graph edge weights, and obtain the normalized adjacency matrix using symmetric normalization. The The specific expression is: In the formula, D Let be the degree matrix of the graph nodes, and let its diagonal elements be the adjacency matrix. A The sum of the connection weights of the corresponding nodes; A It is an adjacency matrix; This is the normalized adjacency matrix; S3.3 Constructing a spatiotemporal graph convolutional network with input features By integrating downscaled groundwater storage variables, meteorological vegetation variables, and monthly cycle sine and cosine coding, the aforementioned The specific expression is: In the formula, m The month number; , , , These are respectively represented as precipitation, temperature, evapotranspiration, and normalized vegetation index; S3.4 The spatiotemporal graph convolutional network consists of a feature normalization module, a temporal convolutional layer, a graph convolutional layer, and a hidden layer decoder connected in series, taking the input features as input. With normalized adjacency matrix As input, through model parameters Forward operation outputs estimated values of shallow groundwater storage variables. The The expression is: In the formula, [] represents the mapping function of the spatiotemporal graph convolutional network.
5. The method for separating shallow groundwater storage variable signals based on spatiotemporal graph convolutional networks according to claim 1, characterized in that, S4 specifically includes: S4.1 Measured water level in the submersible well Anomalies and standardization transformations are performed, and original monitoring labels are constructed by combining the standard deviation of groundwater storage variables corresponding to well points in the grid. Learnable scaling coefficients are then introduced. Bias terms Correcting dimensional differences yields the corrected supervision label. The The specific expression is: In the formula, This represents the standard deviation of the groundwater level sequence for the j-th unconfined well. This represents the standard deviation of the grid-downscaled groundwater reservoir variable series corresponding to this well; This represents the average groundwater level during the observation period of the j-th unconfined well; S4.2 Construct a composite multi-constraint loss function Loss, which includes a label fitting term, a time-series trend consistency term, a time smoothing term, a frequency domain suppression term, and a scale parameter regularization term: The expression for the composite multi-constraint loss function Loss is: In the formula, This is a variable signal for deep residual groundwater storage; This represents the estimated shallow groundwater storage variables output by the spatiotemporal graph convolutional network. This represents a monitoring label for shallow groundwater reservoir variables constructed from water levels observed in submerged wells; Time difference series representing estimates of shallow groundwater storage variables; Time-difference sequences representing deep residual groundwater reservoir variable signals; Represents the frequency domain weighting function; , , These represent the weight coefficients of the trend consistency term, time smoothing term, and frequency domain constraint term, respectively. Represented as the weight of the scaling parameter regularization term; S4.3 Perform power spectrum analysis on the difference series of regional average groundwater storage variables to identify seasonal cycles. Interannual cycle Calculate the frequency domain cutoff period Cutoff frequency number Construct a frequency domain weighting function M is the total timing length; the frequency domain cutoff period Cutoff frequency number Frequency domain weighting function The expression is: S4.4 Dynamically update the weights of each loss term based on the NTK gradient norm of the neural tangent kernel. Update the formula: In the formula Fit the loss gradient to the label. The gradient of the loss term to be adjusted. To minimize the constant and prevent the denominator from being zero, For weighted amplitude limiting operators; S4.
5. Employ gradient descent method to jointly optimize the parameters of the spatiotemporal graph convolutional network. Scale correction parameters , The model convergence training is completed.
6. The method for separating shallow groundwater storage variable signals based on spatiotemporal graph convolutional networks according to claim 1, characterized in that, S5 specifically includes: S5.
1. Input the full-time high-resolution DSGWSA and the corresponding auxiliary variables into the training spatiotemporal graph convolutional network to output the whole area grid-by-grid shallow groundwater storage variable SGWSA. S5.
2. The deep residual groundwater storage variable DRGWSA is obtained through differential calculation. The calculation formula is as follows: 。 7. The method for separating shallow groundwater storage variable signals based on spatiotemporal graph convolutional networks according to any one of claims 1-6, characterized in that, It also includes an accuracy verification step: Data from CSRRL06Mascon0.25°GRACE / GRACE-FO, GLDASNoah land surface hydrological data, 1km monthly meteorological vegetation data, 30m DEM and land use data, and monthly groundwater level monitoring data from regional groundwater wells were collected as input data sources. The well point data were divided into training set and independent validation set in a 7:3 ratio. The Pearson correlation coefficient R and Nash efficiency coefficient NSE were used to evaluate the matching accuracy between the separated shallow groundwater storage variables and the measured groundwater level.
8. The method for separating shallow groundwater storage variable signals based on spatiotemporal graph convolutional networks according to claim 1, characterized in that, K-means segmentation and regional modeling eliminates interference from dynamic differences in groundwater in different hydrological units; spatial weighted graph structure integrates topographic and land use constraints to achieve grid-based spatial hydrological correlation modeling; measured unconfined well monitoring labels + adaptive frequency domain constraints achieve physically interpretable separation of shallow and deep groundwater signals, solving the defects of signal aliasing and unclear hydrological and physical meaning in traditional blind source decomposition methods.