A method for super-resolution network spatial downscaling of GRACE data by fusing geographic information and environmental variables

Abstract

The present application relates to satellite hydrological data processing technical field, specifically disclose a kind of GRACE data super-resolution network spatial downscaling method of fusion geographic information and environmental variable, comprising: S1, the original resolution GRACE data and original GLDAS data of research area are obtained, and these data are preprocessed;S2, the data obtained in step S1 is divided into training set and test set, and the training set is used to train GRACE data spatial downscaling model, to obtain the discriminator and generator after training;S3, the low-resolution data of GRACE in test set and the high-resolution environmental variable corresponding to the data at the moment are input into the generator trained in step S2, and finally the high-resolution GRACE image after downscaling is obtained.The present application can be used not only for regional scale groundwater storage dynamic monitoring, flood risk assessment, but also can be applied to agricultural drought monitoring, ecological hydrological process simulation and other scenes.

Description

Technical Field

[0001] This invention relates to the field of satellite hydrological data processing technology, specifically to a spatial downscaling method for GRACE data super-resolution networks that integrates geographic information and environmental variables. Background Technology

[0002] Currently, GRACE data downscaling methods are mainly divided into two categories: statistical downscaling and physical downscaling. Physical downscaling integrates GRACE data into physical process-based surface, hydrological, and groundwater models, enabling higher spatial resolution simulations of GRACE data. While these methods possess clear physical interpretation logic, they require substantial data and computational power, and are also constrained by model parameter settings and process expressions. In contrast, statistical downscaling requires less data and a simpler framework. It achieves scale transformation by establishing statistical relationships between low-resolution GRACE data and high-resolution auxiliary variables (such as precipitation and soil moisture). Within the statistical downscaling framework, a linear model can be used to establish the relationship between GRACE data and auxiliary variables at a low-resolution grid. This relationship is then mapped to a high-resolution grid, thus transforming the GRACE data to obtain high-resolution GRACE data.

[0003] In the field of statistical downscaling of GRACE data, Geographically Weighted Regression (GWR) has been widely used due to its core characteristic of considering the spatial non-stationarity of variables. This model, through the calculation of a spatial weight matrix, assigns higher weights to sample points closer to the target point, constructing a regression model within a local area. This approach effectively captures spatial heterogeneity characteristics and can better reflect the spatial variation trends of GWSA (Groundwater Storage Anomaly). Several performance-optimized variants of the GWR model exist, each improving its applicability in different ways. For example, the SGWR (Geographically Weighted Similarity Regression) model introduces a new weight matrix that combines spatial weights with attribute similarity weights. Applying this new weight matrix to the geographically weighted regression model framework effectively improves model performance. In the field of image super-resolution, many models can also be applied to GRACE data downscaling. For instance, SRCNN can directly learn the end-to-end mapping between low-resolution and high-resolution images, completing the image super-resolution task with a simple structure. SRGAN's core advantage lies in its definition of a novel perceptual loss function, which better aligns with human visual perception of similarity. Through adversarial training, SRGAN can generate high-frequency details. VIMM, as the foundational framework in DVMSR, aims to leverage the powerful long-range modeling capabilities and computational efficiency of the Mamba model to overcome the limitations of convolutional neural networks (CNNs) in long-range dependency modeling within existing efficient super-resolution (ESR) methods. These various super-resolution methods offer new ideas for GRACE data downscaling; however, simple super-resolution methods cannot consider spatial heterogeneity and spatial correlation, nor can they utilize more relevant data to achieve GRACE data downscaling. Summary of the Invention

[0004] This invention aims to overcome the shortcomings of existing technologies and provide a GRACE data downscaling method that integrates geographic information and environmental variables, combines the local spatial correlation characteristics of GWR, and introduces the Mamba attention mechanism. It can be used not only for regional-scale dynamic monitoring of groundwater storage and flood risk assessment, but also for applications such as agricultural drought monitoring and eco-hydrological process simulation.

[0005] To achieve the above objectives, this invention provides a GRACE data super-resolution network spatial downscaling method that integrates geographic information and environmental variables, comprising:

[0006] Step S1: Obtain the original GRACE data and original GLDAS data of the study area and perform preprocessing to obtain preprocessed low-resolution GRACE data, original resolution GRACE data and preprocessed high-resolution GLDAS data.

[0007] Step S2: Divide the preprocessed data into a training set and a test set. Use the training set to train the GRACE data space downscaling model to obtain a trained discriminator and generator. The GRACE data space downscaling model includes an untrained discriminator and generator. The specific steps for model training are as follows:

[0008] Step S2.1: First, input the low-resolution GRACE data in the training set into the generator, and then calculate the MSE loss by combining the generator's output with the original resolution GRACE data in the training set to obtain the pre-trained generator.

[0009] Step S2.2: Initialize the generator using the pre-trained weight parameters to obtain the initialized generator. Then, input the low-resolution GRACE data and high-resolution GLDAS data from the training set into the initialized generator to obtain the first pseudo-super-resolution image. Then, input the first pseudo-super-resolution image into the untrained discriminator to obtain the first feature map output. Then, calculate the adversarial loss based on the first feature map and the original resolution GRACE data in the training set. Then, calculate the gradient of each parameter of the discriminator through backpropagation and adjust the discriminator parameters according to the gradient.

[0010] Step S2.3: First, input the low-resolution GRACE data and high-resolution GLDAS data from the training set into the initialized generator to obtain the second pseudo-super-resolution image; then, calculate the VGG loss based on the second pseudo-super-resolution image and the original resolution GRACE data in the training set; then, input the second pseudo-super-resolution image into the discriminator optimized in step S2.2 to obtain the second feature map and calculate the adversarial loss; then, sum the VGG loss and the adversarial loss by weight to obtain the combined loss; based on the combined loss, calculate the gradient of each parameter of the generator through backpropagation, and then optimize the parameters of the generator.

[0011] Step S2.4, repeat steps S2.2 and S2.3 until the preset stopping iteration condition is met, which means that the model training is complete;

[0012] Step S3: Input the low-resolution GRACE data from the test set and the corresponding high-resolution GLDAS data at the time point into the generator trained in step S2, and finally obtain the downscaled high-resolution GRACE image.

[0013] Furthermore, in step S1, the specific steps of data preprocessing are as follows:

[0014] Step S1.1: First, remove null values ​​from the original GRACE data and the original GLDAS data, and then unify the time series and spatial range of the two to make them correspond in time and align in space; wherein, the spatial resolution of the original GRACE data and the original GLDAS data is 0.25°×0.25°.

[0015] Step S1.2: Perform interpolation on the original resolution GRACE data obtained in step S1.1 to generate low-resolution GRACE data with a spatial resolution of 0.5°×0.5°.

[0016] Step S1.3: Input the GLDAS data obtained in step S1.1 into the feature selection layer. In this feature selection layer, a random forest model is constructed based on the preprocessed GLDAS data and the original resolution GRACE data. The contribution of each variable to the change of GRACE is evaluated by calculating the feature importance score of each variable. Then, each variable is sorted according to the importance score, and a preset number of high-resolution GLDAS data with the highest ranking is selected for subsequent operations.

[0017] Further, the generator includes a low-level feature extraction layer, a high-level feature extraction layer, an upsampling layer, a Conv+Mamba module, and a geographic activation function. The processing steps of the generator in step S2 are as follows: the preprocessed low-resolution GRACE data is used as input, and this data is first processed by the low-level feature extraction layer to extract preliminary features; then the extracted preliminary features are input to the high-level feature extraction layer to further extract high-level features; then, the high-level features are input to the upsampling layer, and the high-level features are sampled to the target resolution through the upsampling layer; at the same time, the high-resolution GLDAS data selected in the preprocessing process is input to the Conv+Mamba module, and after being processed by the geographic activation function, it is concatenated with the high-level features that have reached the target resolution; finally, the high-resolution GRACE image is obtained through convolutional kernel reconstruction.

[0018] Furthermore, the low-level feature extraction layer consists of a 9×9 convolutional kernel with 64 output channels and a corresponding activation function; the high-level feature extraction layer consists of n residual blocks. Within each residual block, the input first passes through a 3×3 convolutional kernel with 64 output channels, followed by a batch normalization layer using PReLU as the activation function. The activated feature map then enters a second convolutional kernel with the same configuration, followed by another batch normalization. After all residual blocks, there is a convolutional kernel and normalization layer with the same configuration. Skip connections are used between each residual block and after the high-level feature extraction layer for element-wise summation; the upsampling layer consists of a bilinear interpolation module, a 3×3 convolutional kernel, and a PReLU activation function.

[0019] Furthermore, the geographic activation function processing procedure is as follows: First, each target point i in the high-resolution GLDAS data is determined, and then the actual spatial distance between point i and other points j is calculated using the Haversine formula. Next, bandwidth h is introduced as the core adjustment parameter, and a Gaussian kernel function is used to transform the spatial distance into a spatial distance weight with decay characteristics. ;Specifically:

[0020]

[0021]

[0022] in, and Let i and j be the latitudes of points i and j, respectively. and The longitudes of points i and j are respectively. The radius is the Earth's radius.

[0023] Furthermore, the discriminator contains eight convolutional layers, and as the number of layers increases, the number of output channels of the 3×3 convolutional layer increases from 64 to 512, doubling each time; whenever the number of features doubles, the network reduces the image size by using convolutional layers with a stride of 2; the final 512 feature maps are followed by two dense layers and a final sigmoid activation function to obtain the probability of sample classification.

[0024] Further, in step S2.1, the expression for calculating the MSE loss is:

[0025]

[0026] in, For MSE loss, and These represent the width and height of the input low-resolution GRACE data, respectively. This is the magnification factor. To preprocess the original resolution GRACE data in coordinates The value at that location, Super-resolution images generated by the generator network in coordinates The value at that location.

[0027] Further, in step S2.2, the expression for calculating the VGG loss is:

[0028]

[0029] in, For VGG loss, and These are the width and height of the input feature map, respectively. The output obtained by inputting the original resolution GRACE data into VGG is in coordinates. The value at that location, The super-resolution image generated by the generator network is input into VGG, and the output is obtained in coordinates. The value at that location.

[0030] Furthermore, in step S2.2, the resistance loss is... The calculation expression is:

[0031]

[0032] in, For low-resolution images, The super-resolution image output by the generator. It is a discriminator.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] (1) Geographic information extraction and application: This invention uses a geographic activation function in the generator, and adjusts the influence of surrounding sample points on sample points by calculating geographic weights, thereby controlling the activation degree of the activation function.

[0035] (2) Incorporation of environmental variables: Based on the traditional super-resolution model, this invention first calculates feature importance scores through a random forest model to select variables that have stronger explanatory power for GRACE. Then, the selected environmental variables are input into the generator to improve the physical consistency and environmental responsiveness of the downscaling results.

[0036] (3) Fusion of local and global feature information: This invention combines the local feature extraction capability of convolutional networks with the global dependency modeling capability of Mamba to achieve the coordinated expression of spatial details and overall structure, thereby enhancing the accuracy and spatial consistency of GRACE data downscaling results.

[0037] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0038] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings:

[0039] Figure 1 This is a flowchart illustrating a GRACE data downscaling method based on super-resolution networks that integrates geographic information and environmental variables, according to the present invention.

[0040] Figure 2 This is a flowchart of data preprocessing in an embodiment of the present invention;

[0041] Figure 3 This is a flowchart of the generator training process in an embodiment of the present invention;

[0042] Figure 4 This is a flowchart of the training process of the discriminator in an embodiment of the present invention;

[0043] Figure 5 This is a flowchart of downscaling prediction in an embodiment of the present invention;

[0044] Figure 6 This is a schematic diagram of the network structure of the generator in an embodiment of the present invention;

[0045] Figure 7 This is a schematic diagram of the network structure of the discriminator in an embodiment of the present invention;

[0046] Figure 8 This is a schematic diagram of the network structure of the residual block in an embodiment of the present invention;

[0047] Figure 9 This is the original GRACE data diagram in the embodiment of the present invention;

[0048] Figure 10 This is a preprocessed low-resolution GRACE data image in an embodiment of the present invention;

[0049] Figure 11 This is a high-resolution GRACE data map output by the downscaling model in this embodiment of the invention;

[0050] Figure 12 This is an average surface temperature map from high-resolution GLDAS data in an embodiment of the present invention;

[0051] Figure 13 This is a subsurface runoff map from high-resolution GLDAS data in an embodiment of the present invention;

[0052] Figure 14 This is a rainfall rate map from high-resolution GLDAS data in an embodiment of the present invention;

[0053] Figure 15 This is a soil temperature map from high-resolution GLDAS data in an embodiment of the present invention;

[0054] Figure 16 This is a temperature map from high-resolution GLDAS data in an embodiment of the present invention. Detailed Implementation

[0055] The present invention will now be described in detail with reference to the embodiments shown in the accompanying drawings. However, it should be noted that these embodiments are not intended to limit the present invention. Equivalent transformations or substitutions in function, method, or structure made by those skilled in the art based on these embodiments are all within the scope of protection of the present invention.

[0056] Please see Figures 1 to 5 This embodiment provides a GRACE data downscaling method based on super-resolution networks that fuses geographic information and environmental variables, specifically including the following steps:

[0057] Step S1: First, acquire the original GRACE data and original GLDAS data of the study area. Then, preprocess these data to obtain preprocessed low-resolution GRACE data with a spatial resolution of 0.5°×0.5°, original-resolution GRACE data with a spatial resolution of 0.25°×0.25°, and high-resolution GLDAS data. It should be noted that the resolution of the preprocessed high-resolution GLDAS data is the same as that of the original GLDAS data, both being 0.25°×0.25°. The specific steps of data preprocessing in this step are as follows:

[0058] Step S1.1: First, remove null values ​​from the acquired raw GRACE and GLDAS data to eliminate invalid raster points and avoid interference from missing data in subsequent feature calculations. Then, unify the temporal and spatial ranges of these two types of data, ensuring temporal correspondence and raster alignment. Removing null values ​​eliminates the impact of invalid or missing data on model training and analysis results, while raster alignment ensures precise correspondence between GRACE and GLDAS data in spatial resolution and geographic coordinates, thereby achieving effective matching and comparison at the pixel level.

[0059] Step S1.2: Perform interpolation on the original resolution GRACE data with a spatial resolution of 0.25°×0.25° obtained by preprocessing in step S1.1 to generate low resolution GRACE data with a spatial resolution of 0.5°×0.5°.

[0060] Step S1.3: Input the GLDAS data obtained in step S1.1 into the feature selection layer. The GLDAS data contains numerous environmental variables, such as soil moisture, snow water equivalent, and canopy water interception. In this feature selection layer, a random forest model is constructed based on the preprocessed GLDAS data and the original resolution GRACE data. The contribution of each variable in the GLDAS data to GRACE change is evaluated by calculating the feature importance score. Finally, the variables are ranked according to their importance scores, and the top five environmental variables are selected for subsequent operations. Among them, the environmental variables with higher rankings have a stronger explanatory power for GRACE data. This step is performed in parallel with step S1.2.

[0061] Step S2: Divide the preprocessed data obtained in Step S1 into a training set and a test set in a 4:1 ratio. That is, 80% of the preprocessed original resolution GRACE data, low-resolution GRACE data, and high-resolution GLDAS data (environmental variables) is used as the training set, and the remaining 20% ​​is used as the test set. The GRACE data spatial downscaling model is trained using the training set, ultimately obtaining the trained discriminator and generator. This GRACE data spatial downscaling model includes an untrained generator and an untrained discriminator. The model training in Step S2 specifically includes the following steps:

[0062] Step S2.1: First, input the low-resolution GRACE data from the training set into the generator. Calculate the MSE loss by comparing the generator's output with the original resolution GRACE data to obtain the pre-trained generator. The expression for calculating the MSE loss is as follows:

[0063]

[0064] in, For MSE loss, and These represent the width and height of the input low-resolution GRACE data, respectively. This is the magnification factor. To preprocess the original resolution GRACE data in coordinates The value at that location, Super-resolution images generated by the generator network in coordinates The value at that location.

[0065] Step S2.2: Initialize the generator using pre-trained weight parameters to obtain the initialized generator; then input the low-resolution GRACE data and high-resolution GLDAS data from the training set into the initialized generator to obtain the first pseudo-super-resolution image; then input the first pseudo-super-resolution image into the untrained discriminator to obtain the first feature map output; then calculate the adversarial loss using the first feature map and the original resolution GRACE data from the training set, and then calculate the gradients of each parameter of the discriminator through backpropagation, and adjust the discriminator parameters according to the gradients. The adversarial loss... The calculation expression is as follows:

[0066]

[0067] in, For low-resolution images, The super-resolution image output by the generator. It is a discriminator.

[0068] Step S2.3: Input the low-resolution GRACE data and high-resolution GLDAS data from the training set into the initialized generator to obtain the second pseudo-super-resolution image; then calculate the VGG loss using the second pseudo-super-resolution image and the original resolution GRACE data from the training set; then input the second pseudo-super-resolution image into the discriminator optimized in step S2.2 to obtain the second feature map; then calculate the adversarial loss using the second feature map; finally, sum the VGG loss and the adversarial loss in a weighted manner to obtain the combined loss; based on this combined loss, calculate the gradient through backpropagation to optimize the generator parameters. In this step, the expression for the VGG loss function is:

[0069]

[0070] in, For VGG loss, and These are the width and height of the input feature map, respectively. The output obtained by inputting the original resolution GRACE data into VGG is in coordinates. The value at that location, The super-resolution image generated by the generator network is input into VGG, and the output is obtained in coordinates. The value at that point. Update the generator weights through backpropagation.

[0071] Step S2.4, repeat steps S2.2 and S2.3 until the preset stopping iteration condition is met, which means that the model training is complete.

[0072] In the GRACE data space downscaling model, the core function of the generator is to generate high-resolution images. For example... Figure 6 As shown, the generator includes a low-level feature extraction layer, a high-level feature extraction layer, an upsampling layer, a Conv+Mamba module, and a geographic activation function. Combined with... Figure 3 As shown, the generator's processing steps are as follows:

[0073] First, the preprocessed low-resolution GRACE data is input into a low-level feature extraction layer for preliminary feature extraction. This layer consists of a 9×9 convolutional kernel with 64 output channels and a corresponding activation function. This layer can quickly process low-resolution images, capturing and extracting preliminary features by increasing the number of channels. After the preliminary features are generated, they are then input into higher-level feature extraction layers to further mine deeper information, such as... Figure 7 As shown, the high-level feature extraction layer consists of n residual blocks. Within each residual block, the input first passes through a 3×3 convolutional kernel with 64 output channels, followed by a batch normalization layer using PReLU as the activation function. The activated feature map then enters a second convolutional kernel with the same configuration, followed by another batch normalization. After all residual blocks, there is a convolutional kernel and normalization layer with the same configuration. Skip connections are used between residual blocks and after the high-level feature extraction layer for element-wise summation. Using skip connections between multiple residual blocks and after the high-level feature extraction layer reduces the burden on the network architecture to model identity mappings. After high-level feature extraction, the extracted high-level features are input to the upsampling layer, which samples them to the target resolution. The upsampling layer consists of a bilinear interpolation module, a 3×3 convolutional kernel, and a PReLU activation function. The bilinear interpolation module is used to improve image resolution, and the 3×3 convolutional kernel can perform detail restoration and supplement high-frequency details. Meanwhile, the high-resolution GLDAS data (high-resolution environmental variables) selected during the preprocessing process is input into the Conv+Mamba module. In this module, the input features first pass through two cascaded ViMM (Vision MambaModule) modules, which utilize the powerful modeling capabilities of the Mamba state space model to capture long-range dependencies and global context information. Subsequently, local spatial features are extracted through convolutional layers, thus taking into account both global and local receptive fields, and finally outputting the features of each high-resolution environmental variable.

[0074] The selected high-resolution environmental variables are processed by a geographic activation function and then matched and stitched with the high-level features that have reached the target resolution. Finally, the stitched features are integrated and reconstructed by a convolutional kernel with a size of 9×9 and 3 output channels to obtain a high-resolution GRACE image.

[0075] Subsequently, when calculating geographic weights using the coordinate information of high-resolution environmental variables, each target point i with the characteristics of the high-resolution environmental variables is first determined, and then the actual spatial distance between target point i and other target points j is calculated using the Haversine formula. Next, bandwidth is introduced. As the core adjustment parameter (bandwidth set to 50 km), the spatial distance is transformed into a spatial distance weight with attenuation characteristics using a Gaussian kernel function. For n locations on a raster map, a corresponding geographic weight needs to be calculated for each location, ultimately forming n geographic weight matrices. The closer two points are, the larger their weight values, and vice versa. This accurately characterizes the strength of geographic spatial connections, providing quantitative support for subsequent spatial analyses (such as spatial interpolation and regional correlation calculations). Specifically, spatial distance... Spatial distance weights The calculation formulas are as follows:

[0076]

[0077]

[0078] in, and Let i and j be the latitudes of points i and j, respectively. and The longitudes of points i and j are respectively. The radius is the Earth's radius.

[0079] For each target point i in a high-resolution environmental variable feature, its variable value is obtained by multiplying the variable values ​​of surrounding points by the corresponding elements of the geographic weight matrix of target point i and then summing the results. After processing with geographic weights, the result is input into an activation function to obtain the high-level features of the high-resolution environmental variables. This operation can control the activation level of each point through geographic weights, allowing these features to fully consider spatial autocorrelation characteristics, that is, the feature value of each pixel or spatial location will be affected by the correlation of features in its surrounding area.

[0080] The discriminator's main function is to guide the generator to learn spatial structure features that more closely approximate the actual Earth's mass distribution by comparing the generated results with the real high-resolution GRACE field. For example... Figure 8As shown, the discriminator consists of two feature extraction layers. Each feature extraction layer includes a 1×1 convolutional kernel, a normalization layer, and an activation function. In the discriminator, the input image is processed through successive convolutional layers, batch normalization (BN) layers, and LeakyReLU activation function layers to extract hierarchical features. These features are then nonlinearly mapped and reduced in dimensionality through two fully connected layers, and finally, a true / false probability map is obtained through the Sigmoid function.

[0081] In the training process of Generative Adversarial Networks (GANs), adversarial loss is a core metric characterizing the adversarial game relationship between the discriminator and the generator. For the discriminator, its adversarial loss... This is used to measure the discriminator's ability to distinguish between real high-resolution images and pseudo-high-resolution images generated by the generator, while introducing a gradient penalty mechanism to improve training stability.

[0082] After training the discriminator for several epochs, its parameters are fixed and no longer updated. At this point, only the generator network is optimized. The original resolution GRACE data from the training set is input into the generator to obtain high-resolution images. The parameters in the generator are adjusted by calculating the loss function and backpropagation. Perceptual loss is used in the generator, i.e., high-level features are obtained by inputting the images into VGG, and the generator is trained by minimizing feature differences.

[0083] Step S3: Input the low-resolution GRACE data in the test set and the high-resolution environmental variables at the corresponding time points into the generator trained in step S2 to obtain the downscaled high-resolution GRACE image. Compare and analyze the result with the original resolution GRACE data in the test set. Evaluate the result by calculating indicators such as root mean square error to determine the performance of the model.

[0084] Data sources and preparation for this embodiment: GRACE Mascon data from 2002 to 2024 were selected as the original GRACE data, and Global Land Data Assimilation System (GLDAS) data from the same period were selected as the original GLDAS data (high-resolution environmental auxiliary data); the spatial resolution of both the original GRACE and original GLDAS data was 0.25° × 0.25°. The GLDAS data used included 10 environmental variables such as snow water equivalent, soil evaporation, snowfall rate, and groundwater runoff.

[0085] like Figure 9 As shown, this embodiment obtains the original GRACE data of the Yellow River Basin, and after preprocessing, interpolates to obtain the following: Figure 10 The low-resolution GRACE data shown is used as the test dataset input into the model, and the output is as follows. Figure 11 The high-resolution GRACE data shown is from [source missing]. Figures 9 to 11The comparison shows that the high-resolution GRACE data inherits the spatial distribution trend of the original GRACE data, and highly overlaps with the original data in terms of geographical trends and evolution gradients, proving the accuracy of the downscaling process. From a numerical distribution perspective, the core numerical range of the high-resolution GRACE remains consistent with the original GRACE, meaning that while the model recovers the masked local extrema, it does not deviate from the overall numerical constraints of the original GRACE data. Furthermore, the fine textures generated by the model are not created out of thin air, but are restored based on the original GRACE data and through input environmental variables. Figures 12 to 16 These are schematic diagrams of five variables in high-resolution GLDAS data: average surface temperature, groundwater runoff, rainfall rate, soil temperature, and air temperature.

[0086] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A spatial downscaling method for GRACE data super-resolution networks that integrates geographic information and environmental variables, characterized in that, include: Step S1: Obtain the original GRACE data and original GLDAS data of the study area and perform preprocessing to obtain preprocessed low-resolution GRACE data, original resolution GRACE data and preprocessed high-resolution GLDAS data. Step S2: Divide the preprocessed data into a training set and a test set. Use the training set to train the GRACE data space downscaling model to obtain a trained discriminator and generator. The GRACE data space downscaling model includes an untrained discriminator and generator. The specific steps for model training are as follows: Step S2.1: First, input the low-resolution GRACE data in the training set into the generator, and then calculate the MSE loss by combining the generator's output with the original resolution GRACE data in the training set to obtain the pre-trained generator. Step S2.2: Initialize the generator using the pre-trained weight parameters to obtain the initialized generator; then input the low-resolution GRACE data and high-resolution GLDAS data from the training set into the initialized generator to obtain the first pseudo-super-resolution image; then input the first pseudo-super-resolution image into the untrained discriminator to obtain the first feature map output; then calculate the adversarial loss based on the first feature map and the original resolution GRACE data in the training set, then calculate the gradient of each parameter of the discriminator through backpropagation, and adjust the discriminator parameters according to the gradient; Step S2.3: First, input the low-resolution GRACE data and high-resolution GLDAS data from the training set into the initialized generator to obtain the second pseudo-super-resolution image; then, calculate the VGG loss based on the second pseudo-super-resolution image and the original resolution GRACE data in the training set; then, input the second pseudo-super-resolution image into the discriminator optimized in step S2.2 to obtain the second feature map and calculate the adversarial loss; then, sum the VGG loss and the adversarial loss by weight to obtain the combined loss; based on the combined loss, calculate the gradient of each parameter of the generator through backpropagation, and then optimize the parameters of the generator. Step S2.4, repeat steps S2.2 and S2.3 until the preset stopping iteration condition is met, which means that the model training is complete; Step S3: Input the low-resolution GRACE data from the test set and the corresponding high-resolution GLDAS data at the time point into the generator trained in step S2, and finally obtain the downscaled high-resolution GRACE image.

2. The GRACE data super-resolution network spatial downscaling method according to claim 1, characterized in that, In step S1, the specific steps of data preprocessing are as follows: Step S1.1: First, remove null values ​​from the original GRACE data and the original GLDAS data, and then unify the time series and spatial range of the two to make them correspond in time and align in space; wherein, the spatial resolution of the original GRACE data and the original GLDAS data is 0.25°×0.25°. Step S1.2: Perform interpolation on the original resolution GRACE data obtained in step S1.1 to generate low-resolution GRACE data with a spatial resolution of 0.5°×0.5°. Step S1.3: Input the GLDAS data obtained in step S1.1 into the feature selection layer. In this feature selection layer, a random forest model is constructed based on the preprocessed GLDAS data and the original resolution GRACE data. The contribution of each variable to the change of GRACE is evaluated by calculating the feature importance score of each variable. Then, each variable is sorted according to the importance score, and a preset number of high-resolution GLDAS data with the highest ranking is selected for subsequent operations.

3. The GRACE data super-resolution network spatial downscaling method according to claim 1, characterized in that, The generator includes a low-level feature extraction layer, a high-level feature extraction layer, an upsampling layer, a Conv+Mamba module, and a geographic activation function. The processing steps of the generator in step S2 are as follows: the preprocessed low-resolution GRACE data is used as input, and the data is first processed by the low-level feature extraction layer to extract preliminary features; then the extracted preliminary features are input to the high-level feature extraction layer to further extract high-level features. Next, the high-level features are input into the upsampling layer, which samples the high-level features to the target resolution. Meanwhile, the high-resolution GLDAS data selected during the preprocessing is input into the Conv+Mamba module. After being processed by the geographic activation function, it is stitched together with the high-level features that have reached the target resolution. Finally, the high-resolution GRACE image is obtained through convolutional kernel reconstruction.

4. The GRACE data super-resolution network spatial downscaling method according to claim 3, characterized in that, The low-level feature extraction layer consists of a 9×9 convolutional kernel with 64 output channels and a corresponding activation function. The high-level feature extraction layer consists of n residual blocks. Within each residual block, the input first passes through a 3×3 convolutional kernel with 64 output channels, followed by a batch normalization layer using PReLU as the activation function. The activated feature map then enters a second convolutional kernel with the same configuration, followed by another batch normalization. After all residual blocks, there is a convolutional kernel and normalization layer with the same configuration. Skip connections are used between each residual block and after the high-level feature extraction layer for element-wise summation. The upsampling layer consists of a bilinear interpolation module, a 3×3 convolutional kernel, and a PReLU activation function.

5. The GRACE data super-resolution network spatial downscaling method according to claim 3, characterized in that, The geographic activation function processing procedure is as follows: First, each target point i in the high-resolution GLDAS data is determined, and then the actual spatial distance between point i and other points j is calculated using the Haversine formula. Next, bandwidth h is introduced as the core adjustment parameter, and a Gaussian kernel function is used to transform the spatial distance into a spatial distance weight with decay characteristics. ;Specifically: in, and Let i and j be the latitudes of points i and j, respectively. and The longitudes of points i and j are respectively. The radius is the Earth's radius.

6. The GRACE data super-resolution network spatial downscaling method according to claim 3, characterized in that, The discriminator contains eight convolutional layers, and as the number of layers increases, the number of output channels of the 3×3 convolutional layer increases from 64 to 512, doubling each time. Whenever the number of features doubles, the network reduces the image size by using convolutional layers with a stride of 2. The final 512 feature maps are followed by two dense layers and a final sigmoid activation function to obtain the probability of sample classification.

7. The GRACE data super-resolution network spatial downscaling method according to claim 1, characterized in that, In step S2.1, the expression for calculating the MSE loss is: in, For MSE loss, and These represent the width and height of the input low-resolution GRACE data, respectively. This is the magnification factor. To preprocess the original resolution GRACE data in coordinates The value at that location, Super-resolution images generated by the generator network in coordinates The value at that location.

8. The GRACE data super-resolution network spatial downscaling method according to claim 7, characterized in that, In step S2.2, the expression for calculating the VGG loss is: in, For VGG loss, and These are the width and height of the input feature map, respectively. The output obtained by inputting the original resolution GRACE data into VGG is in coordinates. The value at that location, The super-resolution image generated by the generator network is input into VGG, and the output is obtained in coordinates. The value at that location.

9. The GRACE data super-resolution network spatial downscaling method according to claim 1, characterized in that, In step S2.2, combating loss The calculation expression is: in, For low-resolution images, The super-resolution image output by the generator. It is a discriminator.

Images