GRACE underground water reserve change downscaling method based on hierarchical learning

Abstract

The invention relates to the technical field of groundwater reserve detection, in particular to a GRACE groundwater reserve change downscaling method based on hierarchical learning. The method comprises the following steps: acquiring multi-source satellite data, underground water actual measurement data and geological auxiliary data; constructing a double-branch feature encoder, and respectively extracting shallow underground water features and deep underground water features according to multi-source satellite data and underground water measured data; the shallow underground water features and the deep underground water features are fused, channel splicing is carried out on the fused features and geological auxiliary data with the same resolution, and spatial downscaling is carried out through a multi-layer perceptron (MLP) network; and setting a physical constraint for a high-resolution prediction result output by the multi-layer perceptron MLP network, and obtaining a high-resolution groundwater reserve change diagram through iterative optimization. And by means of hierarchical learning, cross-modal fusion and the like, the GRACE underground water reserve change downscaling precision and the model generalization ability are improved.

Description

Technical Field

[0001] The present application relates to the technical field of groundwater reserve detection, and more specifically, to a GRACE groundwater reserve change downscaling method based on hierarchical learning. Background Art

[0002] Groundwater, as a vital component of water resources, has a crucial impact on human production and life, as well as the ecological environment. However, due to its underground distribution, accurate measurement and analysis of groundwater reserves has always been a challenge. While GRACE satellites, a type of satellite remote sensing technology, can indirectly reflect changes in terrestrial water reserves by measuring variations in the Earth's gravity field, they struggle to provide high-resolution information on groundwater reserves at small scales. Therefore, an effective method for processing GRACE data is needed to accurately analyze groundwater reserves. Summary of the Invention

[0003] In view of this, the present application provides a GRACE groundwater storage change downscaling method based on hierarchical learning to solve the problem in the existing technology that GRACE data is difficult to provide groundwater storage change information with high spatial resolution on a small scale.

[0004] The technical solutions provided in this application are as follows:

[0005] A hierarchical learning-based downscaling method for GRACE groundwater storage changes includes:

[0006] Acquire multi-source satellite data, groundwater measured data and geological auxiliary data;

[0007] Constructing a dual-branch feature encoder to extract shallow groundwater features and deep groundwater features respectively based on the multi-source satellite data and the groundwater measured data;

[0008] Fusing the shallow groundwater features with the deep groundwater features, and performing channel splicing on the fused features with the geological auxiliary data of the same resolution, and performing spatial downscaling through a multi-layer perceptron (MLP) network;

[0009] Physical constraints are set for the high-resolution prediction results output by the multi-layer perceptron (MLP) network, and a high-resolution groundwater storage change map is obtained through iterative optimization.

[0010] In one possible implementation, the multi-source satellite data includes GRACE satellite water storage anomaly data and vertical deformation data, the groundwater measured data includes groundwater storage change observation data, and the geological auxiliary data includes lithology maps and terrain slope data.

[0011] In one possible implementation, after acquiring the multi-source satellite data, groundwater measured data, and geological auxiliary data, the method further includes preprocessing the multi-source satellite data, groundwater measured data, and geological auxiliary data, including:

[0012] Applying VMD-PCA joint denoising to the multi-source satellite data to eliminate stripe noise and high-frequency errors;

[0013] The groundwater measured data were normalized to [-1, 1] using the Z-score standardization method for geological partitioning. The groundwater measured data were converted to standard normal distribution data with a mean of 0 and a standard deviation of 1, expressed as:

[0014]

[0015] Where z is the standardized value, x is the original data value, μ is the mean of the original data, and σ is the standard deviation of the original data;

[0016] The geological auxiliary data were resampled to a 0.1° grid, aligned with the target resolution.

[0017] In one possible implementation, the dual-branch feature encoder includes a satellite data processing branch and a groundwater data processing branch;

[0018] The satellite data processing branch is a three-layer depth-wise separable convolution, each layer is followed by batch normalization BathNorm and activation function LeakyReLU, which is used to output shallow groundwater characteristic maps based on the multi-source satellite data to characterize the regional water storage trend;

[0019] The groundwater data processing branch includes a 4-layer residual convolution module, which contains jump connections and preset activation functions, and is used to extract deep groundwater characteristics based on the measured groundwater data and characterize the connectivity and permeability characteristics of the aquifer.

[0020] One possible implementation method is to integrate the shallow groundwater characteristics with the deep groundwater characteristics, including:

[0021] According to the implicit network, the cross-attention mechanism is used to calculate the cosine similarity M of the shallow groundwater characteristics and the deep groundwater characteristics, which is expressed as:

[0022]

[0023] Where, F shallow Indicates the characteristics of shallow groundwater, F deep Indicates deep groundwater characteristics;

[0024] Retain feature pairs with cosine similarity M greater than or equal to 0.7 and filter low-correlation noise;

[0025] Dynamic weight allocation is performed on the shallow groundwater characteristics and the deep groundwater characteristics, and fusion weights α and β corresponding to the shallow groundwater characteristics and the deep groundwater characteristics are generated by the gated recurrent unit GRU, and fusion features are output according to the fusion weights.

[0026] In one possible implementation, the spatial downscaling using a multi-layer perceptron (MLP) network includes:

[0027] After channel splicing of the fused features and the geological auxiliary data of the same resolution, the fused features are input into the multi-layer perceptron MLP network to generate a high-resolution feature map through nonlinear mapping;

[0028] Perform bilinear interpolation upsampling on the low-resolution groundwater measured data to obtain the initial interpolation result;

[0029] Based on the geological auxiliary data, generating a pixel-by-pixel weight map W;

[0030] According to the pixel-by-pixel weight map W, the high-resolution feature map and the initial interpolation result are dynamically spatially weighted fused to obtain a high-resolution prediction result, which is expressed as:

[0031] S high =W⊙Upsample(S low )+(1-W)☉MLP(F fusion )

[0032] Where S high is the high-resolution prediction result, Upsample(Slow) is the initial interpolation result obtained by bilinear interpolation upsampling of the low-resolution groundwater measured data, and ☉ represents the element-by-element multiplication operation of two matrices or tensors of the same dimension.

[0033] In one possible implementation, the multi-layer perceptron (MLP) network adopts a parallel branch structure, including a global branch and a local branch.

[0034] The global branch is a two-layer MLP with 256 neurons, which is used to extract large-scale trend characteristics of the hydrological field;

[0035] The local branch is a four-layer MLP with 512 neurons, which is used to capture high-frequency details and abnormal point features;

[0036] Based on the signal-to-noise ratio (SNR) of the data input into the multi-layer perceptron MLP network, the gating weights g1 and g2 are generated and dynamic gating fusion is performed. The fusion branch output is expressed as:

[0037] Fout =g1F global +g2F local

[0038] Where, F global represents the global feature, F local Represents local features;

[0039] The hidden layer uses the CReLU function to enhance nonlinearity, and the output layer uses the Tanh function to limit the range of predicted values.

[0040] In one possible implementation, generating a pixel-by-pixel weight map W based on the geological auxiliary data includes:

[0041] The geological auxiliary data is input into a two-layer 1×1 convolutional network, and a spatial feature map Y is output to analyze the spatial correlation between the lithology map and the terrain slope in the geological auxiliary data;

[0042] The spatial feature map Y is normalized and weighted by the Sigmoid function to generate a pixel-by-pixel weight map W, which is expressed as:

[0043] W=σ(Conv 1*1 (X aux ))

[0044] Where σ is the Sigmoid function, X aux It is the result of splicing fusion features and geological auxiliary data.

[0045] In one possible implementation, after generating the pixel-by-pixel weight map W, the method further includes:

[0046] An exponential decay is applied to the weight of the lithologic boundary region to enhance the boundary, which is expressed as:

[0047]

[0048] Where γ is the attenuation coefficient, which is used to control the interpolation contribution weight of the boundary area, and ▽Lith o represents the lithologic gradient.

[0049] In one possible implementation, physical constraints are set for the high-resolution prediction results output by the multi-layer perceptron (MLP) network, and a high-resolution groundwater storage change map is obtained through iterative optimization, including:

[0050] Calculate the permeability coefficient K and hydraulic gradient based on the high-resolution prediction results Used to verify Darcy's law residuals;

[0051] The Darcy's law residual is spliced ​​with the geological auxiliary data, and inputted into a multi-layer perceptron (MLP) network to generate a correction value, thereby updating the high-resolution prediction result;

[0052] Repeat the above steps until the Darcy's law residual is lower than a preset threshold, and generate the high-resolution groundwater storage change map.

[0053] Compared with the existing technology, the technical solution provided by this application has the following beneficial effects:

[0054] This application proposes a GRACE groundwater storage change downscaling method based on hierarchical learning, which not only optimizes the inversion accuracy of GRACE data, but also achieves more efficient downscaling processing. The VMD-PCA joint denoising technology is used to effectively remove noise, and combined with the Z-score normalization method, the data from different sources have unified statistical characteristics, which is convenient for subsequent comparison and fusion. In addition, by constructing a dual-branch CNN encoder, deep water and shallow water features are extracted respectively, which further improves the accuracy of the downscaling analysis. At the same time, this application combines GRACE satellite data with groundwater measured data, and uses a hierarchical learning strategy to dynamically fuse dual-branch features, enhance cross-modal correlation, and realize cross-scale mapping from low resolution to 0.1° through a multi-branch structure (global + local), which significantly improves the accuracy and reliability of groundwater feature extraction. Finally, by using the interpolation multi-scale global response optimization feature, the groundwater characteristics of different scales can be more comprehensively considered, thereby greatly improving the generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a flowchart of a GRACE groundwater storage change downscaling method based on hierarchical learning provided in Example 1 of the present application.

[0056] Figure 2 This is a flowchart of a GRACE groundwater storage change downscaling method based on hierarchical learning provided in Example 2 of this application.

[0057] Figure 3 A flowchart of constructing a dual-branch CNN encoder to extract deep water and shallow water features is provided in Example 2 of the present application.

[0058] Figure 4 This is a flowchart of a method for generating a pixel-by-pixel weight map W based on geological auxiliary data provided in Example 2 of the present application. DETAILED DESCRIPTION

[0059] The following will combine the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0060] Example 1

[0061] See also Figure 1 , is a flowchart of a GRACE groundwater storage change downscaling method based on hierarchical learning provided in Example 1 of this application. Figure 1 As shown in , the specific implementation steps of the above method include:

[0062] Step 101: Acquire multi-source satellite data, groundwater measured data, and geological auxiliary data.

[0063] Among them, the above-mentioned multi-source satellite data include GRACE satellite water storage anomaly data and vertical deformation data, the above-mentioned groundwater measured data include groundwater storage change observation data, and the above-mentioned geological auxiliary data include lithology maps and terrain slope data.

[0064] Step 102: construct a dual-branch feature encoder to extract shallow groundwater features and deep groundwater features based on the multi-source satellite data and the groundwater measured data.

[0065] In an embodiment of the present application, a dual-branch feature encoder based on a convolutional neural network (CNN) is constructed, specifically including a satellite data processing branch and a groundwater data processing branch. The satellite data processing branch is used to extract shallow groundwater features through spatiotemporal convolution, and the groundwater data processing branch is used to extract deep groundwater features through depth-separable convolution. In a specific application scenario, the multi-source satellite data and groundwater measured data collected in step 101 are respectively input into their respective CNN branches to capture multi-scale global responses to extract low-level features of shallow groundwater and high-level features of deep groundwater.

[0066] Step 103: Fuse the shallow groundwater features and the deep groundwater features, perform channel splicing on the fused features and the geological auxiliary data of the same resolution, and perform spatial downscaling through a multi-layer perceptron (MLP) network.

[0067] In this embodiment, a cross-attention mechanism is used to implicitly fuse shallow and deep groundwater features based on an implicit network. Hydrophysical constraints are embedded in the fusion process. A multi-layer perceptron (MLP) module is used to achieve spatial downscaling.

[0068] Step 104: Set physical constraints for the high-resolution prediction results output by the multi-layer perceptron (MLP) network, and obtain a high-resolution groundwater storage change map through iterative optimization.

[0069] Specifically, the embodiment of the present application only generates high-resolution prediction results (S high ) can be used to calculate the permeability coefficient K and hydraulic gradient based on the result. The Darcy's law residual is then verified. A feedback correction mechanism is then implemented, with the residual R fed back to the multi-layer perceptron (MLP) network for iterative correction, forming a closed-loop optimization that outputs the final high-resolution groundwater storage change map.

[0070] Compared with the prior art, the technical solution provided in Example 1 of the present application has the following beneficial effects:

[0071] Example 2

[0072] See also Figure 2 , is a flowchart of a GRACE groundwater storage change downscaling method based on hierarchical learning provided in Example 2 of this application. Figure 2 As shown in , the specific implementation steps of the above method include:

[0073] Step 201: Acquire multi-source satellite data, groundwater measured data, and high-resolution auxiliary data.

[0074] The multi-source satellite data include monthly-scale GRACE water storage anomaly data and vertical deformation data. The spatial resolution of the GRACE monthly-scale water storage anomaly data is 0.5°, and the resolution of the vertical deformation data is 100m. The groundwater measured data include GRACE terrestrial water storage change data and regional groundwater storage change observation data with a resolution of 0.25° (approximately 25km). The high-resolution auxiliary data include 30m resolution lithology maps (geological classification) and 90m resolution SRTM terrain slope data.

[0075] In this embodiment, GRACE data were obtained from the National Aeronautics and Space Administration (NASA) Earth Data Center or the German Research Center for Geosciences (GFZ). GRACE data has a monthly temporal resolution and reflects changes in the Earth's gravity field. GRACE / GRACE-FO satellite gravity field data (0.5° spatial resolution) and 0.25° groundwater data output by a ground-based hydrological model were simultaneously acquired.

[0076] Step 202: pre-process the acquired multi-source satellite data, groundwater measured data and high-resolution auxiliary data.

[0077] Specifically, the present embodiment applies VMD-PCA combined denoising to multi-source satellite data to eliminate banding noise and high-frequency errors, thereby improving the signal-to-noise ratio. The groundwater measured data is normalized to [-1, 1] using the Z-score standardization method for geological partitioning. The groundwater measured data is converted to standard normal distribution data with a mean of 0 and a standard deviation of 1, expressed as:

[0078]

[0079] Where z is the normalized value, x is the raw data value, μ is the mean of the raw data, and σ is the standard deviation of the raw data. This formula can be used to convert measured groundwater data into standard normal distribution data, matching their magnitude and distribution with GRACE data, facilitating subsequent data fusion and processing.

[0080] For high-resolution auxiliary data, the lithology map and terrain slope were uniformly resampled to a 0.1° grid (approximately 11 km, assuming the WG84 coordinate system) to align with the target resolution.

[0081] Step 203: construct a dual-branch feature encoder based on a convolutional neural network (CNN), including a satellite data processing branch for extracting shallow groundwater and a groundwater data processing branch, so as to extract shallow groundwater characteristics and deep groundwater characteristics.

[0082] In the embodiment of the present application, a dual-branch CNN encoder is constructed to extract the characteristics of deep water and shallow water, thereby performing downscaling analysis. Figure 3 , is a flowchart of constructing a dual-branch CNN encoder to extract deep water and shallow water features provided in Example 2 of this application. Figure 3 As shown in , the above method specifically includes:

[0083] Step 2031: Construct a CNN-based satellite data processing branch to extract shallow groundwater characteristics through spatiotemporal convolution.

[0084] Specifically, the satellite data processing branch employs a 3D convolutional architecture, comprising three layers of depthwise separable convolution. The input dimensions are time × latitude × longitude, with a convolution kernel size of 3*5*5. Each layer of the three depthwise separable convolutions is followed by batch normalization (BathNorm) and a LeakyReLU activation function. This extracts seasonal fluctuations in shallow groundwater levels from the multi-source satellite data and outputs a shallow groundwater characteristic map to characterize regional water storage trends.

[0085] Step 2032: Construct a CNN-based groundwater data processing branch to extract deep groundwater features through depth-separable convolution.

[0086] The above-mentioned groundwater data processing branch adopts a depth-separable convolution structure with an input dimension of channel × latitude × longitude, with the number of channels ≥ 5, including well water level, lithology code, permeability coefficient, terrain gradient, and pumping volume. The convolution kernel size is 3*3, which is used to capture the geological structure correlation characteristics of deep groundwater. In the embodiment of the present application, the groundwater data branch structure is a 4-layer residual convolution module, each residual convolution module contains two 3*3 convolutions and jump connections, and the activation function is CReLU. The output of the groundwater data branch structure is a deep groundwater characteristic map (resolution 0.1°, number of channels 128) to characterize the connectivity and permeability characteristics of the aquifer.

[0087] Step 2033: Input the preprocessed multi-source satellite data and groundwater measured data into their respective CNN branches to capture multi-scale global responses to extract low-level features of shallow groundwater and high-level features of deep groundwater.

[0088] Step 204: Based on the implicit network, a cross-attention mechanism is used to calculate the cosine similarity M between the shallow groundwater feature and the deep groundwater feature, which is expressed as:

[0089]

[0090] Where, F shallow Indicates the characteristics of shallow groundwater, F deep represents the deep groundwater characteristics, and M represents the cosine similarity between the shallow groundwater characteristics at the i-th position and the deep groundwater characteristics at the j-th position.

[0091] Step 205: Dynamically assign weights to the shallow groundwater features and the deep groundwater features, and generate fusion weights α and β corresponding to the shallow groundwater features and the deep groundwater features through the gated recurrent unit GRU to generate fusion features.

[0092] Among them, the above fusion weights α and β are expressed as:

[0093] α, β = GRU(F shallow , F deep )

[0094] The output fusion feature is expressed as:

[0095] F fusion =αF shallow +βF deep

[0096] In the embodiment of the present application, feature pairs with similarity greater than or equal to 0.7 are retained and low-correlation noise is filtered. The GRU network is used to generate shallow weight α=0.6 and deep weight β=0.4, and the fusion feature F is output. fusion , expressed as:

[0097] F fusion =0.6F shallow +0.4F deep

[0098] Furthermore, the above fusion feature F fusion Perform bilinear interpolation upsampling to the target resolution.

[0099] Step 206: perform channel splicing on the fused features and the geological auxiliary data of the same resolution, input the fused features into the multi-layer perceptron (MLP) network, and generate a high-resolution feature map through nonlinear mapping.

[0100] Specifically, the fusion feature F output by the implicit retrieval network is received fusion , the above fusion feature F fusion The high-resolution geological auxiliary data, including lithologic maps and terrain slope data, were loaded into the MLP network. Specifically, the MLP network input consisted of fused features and lithologic maps / terrain slopes (resolution ≤ 0.1°). The fused features had 192 channels, while the lithologic maps / terrain slopes had 2 channels.

[0101] The MLP network receives the fused feature map output by the implicit retrieval network, performs MLP multimodal mapping, concatenates the fused feature map with the high-resolution auxiliary data, and inputs the multi-layer perceptron MLP network to generate a high-resolution feature map through nonlinear mapping, which is expressed as:

[0102]

[0103] in, Indicates channel splicing, X aux For auxiliary data.

[0104] The multi-layer perceptron (MLP) network uses a lightweight parallel branching structure. The global branch consists of a two-layer MLP with 256 neurons, which is used to extract large-scale trend characteristics of the hydrological field. The local branch consists of a four-layer MLP with 512 neurons, which is used to capture high-frequency details and outlier features.

[0105] Dynamic gate fusion generates gating weights g1 and g2 based on the signal-to-noise ratio (SNR) of the input data, and the fusion branch outputs:

[0106] F out =g1F global +g2F local

[0107] Where, F global represents the global feature, F local Represents local features.

[0108] The hidden layer uses the CReLU function to enhance nonlinearity, and the output layer uses the Tanh function to limit the range of predicted values.

[0109] Step 207: Perform bilinear interpolation upsampling on the low-resolution groundwater measured data to obtain an initial interpolation result.

[0110] Step 208: Generate a pixel-by-pixel weight map W based on the above-mentioned geological auxiliary data.

[0111] Specifically, see Figure 4 , is a flow chart of a method for generating a pixel-by-pixel weight map W based on geological auxiliary data provided in Example 2 of this application. Figure 4 As shown in , the above method specifically includes:

[0112] Step 2081: Input the lithology map and terrain slope into a two-layer 1×1 convolutional network for auxiliary feature encoding, and output a spatial feature map Y.

[0113] Step 2082: Normalize the weights using the Sigmoid function, expressed as:

[0114] W=σ(M)

[0115] Where σ is the Sigmoid function, which ensures that the weight value is in the range [0,1].

[0116] Step 2083: Apply exponential decay to the weight of the lithologic boundary region (gradient value ≥ preset threshold) to enhance the boundary, which is expressed as:

[0117]

[0118] Where γ is the attenuation coefficient, which is used to control the interpolation contribution weight of the boundary area.

[0119] Step 2084: The dynamic spatial weights are fused to generate a weight map. The spatial correlation between the lithology map and the terrain slope is analyzed through the spatial attention mechanism (1×1 convolution layer) to generate a pixel-by-pixel weight map W. The formula is:

[0120] W=σ(Conv 1*1 (X aux ))

[0121] Where σ is the Sigmoid function, X aux It is the result of auxiliary feature stitching.

[0122] Step 209: Based on the pixel-by-pixel weight map W, the high-resolution feature map and the initial interpolation result are dynamically spatially weighted to obtain a high-resolution prediction result, thereby adaptively enhancing the detail reconstruction capability of complex areas (such as lithologic boundaries and terrain mutation areas).

[0123] Among them, the above high-resolution prediction results are expressed as:

[0124] S high =W⊙Upsample(S low )+(1-W)☉MLP(F fusion )

[0125] Where S high is the high-resolution prediction result, Upsample(Slow) is the initial interpolation result obtained by bilinear upsampling of the low-resolution groundwater measured data, and ☉ represents the element-by-element multiplication of two matrices or tensors of the same dimension.

[0126] Step 210: Based on the high-resolution prediction result S high Perform physical constraint iterative optimization to calculate the permeability coefficient K and hydraulic gradient To verify Darcy's law residuals.

[0127] The Darcy's law residual is expressed as:

[0128]

[0129] Where N represents the total number of high-resolution grid cells involved in the calculation, that is, the number of spatial locations corresponding to a 0.1° resolution, and q i is the actual groundwater flow vector at the ith grid cell, in m 3 / s.

[0130] Step 211: concatenate the Darcy's law residual with the geological auxiliary data, input the result into a multi-layer perceptron (MLP) network to generate a correction value, and update the high-resolution prediction result.

[0131] Step 212: Repeat steps 210-211 until the Darcy's law residual is lower than a preset threshold, thereby generating the high-resolution groundwater storage change map.

[0132] This application is after the MLP module. Only after the MLP generates high-resolution prediction results, can the permeability coefficient K and hydraulic gradient be calculated based on the above high-resolution prediction results. Then verify the Darcy law residual. Then the feedback correction mechanism is implemented to convert the residual R (0) Combine with auxiliary data, input into MLP to generate correction value ΔS, and update Shigh (1) , the residual is reduced, and the iteration continues until the conditions are met and the optimization is terminated. The final output is a 0.1° (about 11km) resolution groundwater storage change map.

[0133] Compared with the prior art, the technical solution provided in Example 2 of the present application has the following beneficial effects:

[0134] This application adopts the VMD-PCA joint denoising technology to effectively remove noise and improve data quality, and combines it with the Z-score normalization method to make data from different sources have unified statistical characteristics, which is convenient for subsequent comparison and fusion. In addition, a dual-branch CNN encoder is constructed to extract deep water and shallow water features respectively, achieving more accurate downscaling analysis. At the same time, the combination of GRACE satellite data and groundwater measured data further improves the accuracy and reliability of groundwater feature extraction. The hierarchical learning mechanism can implicitly retrieve the network to dynamically fuse dual-branch features and enhance cross-modal correlation, while the multi-branch structure (global + local) realizes cross-scale mapping from low resolution to 0.1°. Finally, by interpolating multi-scale global response optimization features, groundwater characteristics of different scales can be more comprehensively considered, thereby significantly improving the generalization ability of the model. Overall, this application provides a better and more efficient method for monitoring and analyzing groundwater reserve changes while improving the accuracy of GRACE inversion of groundwater reserves.

[0135] Although the embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A GRACE groundwater storage change downscaling method based on hierarchical learning, characterized by: include: Acquire multi-source satellite data, groundwater measured data and geological auxiliary data; Constructing a dual-branch feature encoder to extract shallow groundwater features and deep groundwater features respectively based on the multi-source satellite data and the groundwater measured data; Fusing the shallow groundwater features with the deep groundwater features, and performing channel splicing on the fused features with the geological auxiliary data of the same resolution, and performing spatial downscaling through a multi-layer perceptron (MLP) network; Physical constraints are set for the high-resolution prediction results output by the multi-layer perceptron (MLP) network, and a high-resolution groundwater storage change map is obtained through iterative optimization.

2. The GRACE groundwater storage change downscaling method based on hierarchical learning according to claim 1 is characterized in that: The multi-source satellite data includes GRACE satellite water storage anomaly data and vertical deformation data, the groundwater measured data includes groundwater storage change observation data, and the geological auxiliary data includes lithology maps and terrain slope data.

3. The GRACE groundwater storage change downscaling method based on hierarchical learning according to claim 1 is characterized in that: After acquiring the multi-source satellite data, groundwater measured data, and geological auxiliary data, the method further includes preprocessing the multi-source satellite data, groundwater measured data, and geological auxiliary data, including: Applying VMD-PCA joint denoising to the multi-source satellite data to eliminate stripe noise and high-frequency errors; The groundwater measured data were normalized to [-1, 1] using the Z-score standardization method for geological partitioning. The groundwater measured data were converted to standard normal distribution data with a mean of 0 and a standard deviation of 1, expressed as: Where z is the standardized value, x is the original data value, μ is the mean of the original data, and σ is the standard deviation of the original data; The geological auxiliary data were resampled to a 0.1° grid, aligned with the target resolution.

4. The GRACE groundwater storage change downscaling method based on hierarchical learning according to claim 1 is characterized in that: The dual-branch feature encoder includes a satellite data processing branch and a groundwater data processing branch; The satellite data processing branch is a three-layer depth-wise separable convolution, each layer is followed by batch normalization BathNorm and activation function LeakyReLU, which is used to output shallow groundwater characteristic maps based on the multi-source satellite data to characterize the regional water storage trend; The groundwater data processing branch includes a 4-layer residual convolution module, which contains jump connections and preset activation functions, and is used to extract deep groundwater characteristics based on the measured groundwater data and characterize the connectivity and permeability characteristics of the aquifer.

5. The GRACE groundwater storage change downscaling method based on hierarchical learning according to claim 1 is characterized in that: The shallow groundwater characteristics and deep groundwater characteristics are integrated, including: According to the implicit network, the cross-attention mechanism is used to calculate the cosine similarity M of the shallow groundwater characteristics and the deep groundwater characteristics, which is expressed as: Where, F shallow Indicates the characteristics of shallow groundwater, F deep Indicates deep groundwater characteristics; Retain feature pairs with cosine similarity M greater than or equal to 0.7 and filter low-correlation noise; Dynamic weight allocation is performed on the shallow groundwater characteristics and the deep groundwater characteristics, and fusion weights α and β corresponding to the shallow groundwater characteristics and the deep groundwater characteristics are generated by the gated recurrent unit GRU, and fusion features are output according to the fusion weights.

6. The GRACE groundwater storage change downscaling method based on hierarchical learning according to claim 1 is characterized in that: The spatial downscaling is performed using a multi-layer perceptron (MLP) network, including: After channel splicing of the fused features and the geological auxiliary data of the same resolution, the fused features are input into the multi-layer perceptron MLP network to generate a high-resolution feature map through nonlinear mapping; Perform bilinear interpolation upsampling on the low-resolution groundwater measured data to obtain the initial interpolation result; Based on the geological auxiliary data, generating a pixel-by-pixel weight map W; According to the pixel-by-pixel weight map W, the high-resolution feature map and the initial interpolation result are dynamically spatially weighted fused to obtain a high-resolution prediction result, which is expressed as: S igh =W⊙Upsample(S low )+(1-W)⊙MLP(F fusion ) Where S high is the high-resolution prediction result, Upsample(Slow) is the initial interpolation result obtained by bilinear interpolation upsampling of the low-resolution groundwater measured data, and ☉ represents the element-by-element multiplication operation of two matrices or tensors of the same dimension.

7. The GRACE groundwater storage change downscaling method based on hierarchical learning according to claim 1 is characterized in that: The multi-layer perceptron MLP network adopts a parallel branch structure, including global branches and local branches; The global branch is a two-layer MLP with 256 neurons, which is used to extract large-scale trend characteristics of the hydrological field; The local branch is a four-layer MLP with 512 neurons, which is used to capture high-frequency details and abnormal point features; Based on the signal-to-noise ratio (SNR) of the data input into the multi-layer perceptron MLP network, the gating weights g1 and g2 are generated and dynamic gating fusion is performed. The fusion branch output is expressed as: F out =g1F global +g2F local Where, F global represents the global feature, F local Represents local features; The hidden layer uses the CReLU function to enhance nonlinearity, and the output layer uses the Tanh function to limit the range of predicted values.

8. The GRACE groundwater storage change downscaling method based on hierarchical learning according to claim 6 is characterized in that: Based on the geological auxiliary data, a pixel-by-pixel weight map W is generated, including: The geological auxiliary data is input into a two-layer 1×1 convolutional network, and a spatial feature map Y is output to analyze the spatial correlation between the lithology map and the terrain slope in the geological auxiliary data; The spatial feature map Y is normalized and weighted by the Sigmoid function to generate a pixel-by-pixel weight map W, which is expressed as: W=σ(Conv 1*1 (X aux )) Where σ is the Sigmoid function, X aux It is the result of splicing fusion features and geological auxiliary data.

9. The GRACE groundwater storage change downscaling method based on hierarchical learning according to claim 8 is characterized in that: After generating the pixel-by-pixel weight map W, the method further includes: An exponential decay is applied to the weight of the lithologic boundary region to enhance the boundary, which is expressed as: Where γ is the attenuation coefficient, which is used to control the interpolation contribution weight of the boundary area, and ▽Lith o represents the lithologic gradient.

10. The GRACE groundwater storage change downscaling method based on hierarchical learning according to claim 1 is characterized in that: Physical constraints are set for the high-resolution prediction results output by the multi-layer perceptron (MLP) network, and a high-resolution groundwater storage change map is obtained through iterative optimization, including: Calculate the permeability coefficient K and hydraulic gradient based on the high-resolution prediction results Used to verify Darcy's law residuals; The Darcy's law residual is spliced ​​with the geological auxiliary data, and inputted into a multi-layer perceptron (MLP) network to generate a correction value, thereby updating the high-resolution prediction result; Repeat the above steps until the Darcy's law residual is lower than a preset threshold, and generate the high-resolution groundwater storage change map.

Images