Underground water level observation method and system fusing multi-dimensional sensing information

Abstract

The invention discloses an underground water level observation method and system fusing multi-dimensional sensing information, and the method comprises the steps: firstly obtaining a coarse-resolution dynamic remote sensing image and a high-resolution static geographic data set, and constructing multi-dimensional input through space-time reference alignment and multi-scale feature stacking; on the basis, a subgrid heterogeneity feature extraction mechanism is utilized to explicitly quantify terrain and geological complexity in a coarse grid, and a physical constraint relation of high-resolution geographic features to low-resolution remote sensing signals is established through an adaptive modulation module, so that the model can simulate a nonlinear modulation effect of a local environment to hydrological signals. And finally, performing end-to-end nonlinear mapping and super-resolution reconstruction on the fused multi-dimensional features by using a U-Net architecture comprising an encoder, a decoder and jump connection, and recovering fine underground water level spatial distribution details while keeping a remote sensing macroscopic trend.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of groundwater monitoring and hydrogeological survey, and particularly relates to a groundwater level observation method and system fusing multi-dimensional sensing information. BACKGROUND

[0002] As a key strategic natural resource, the spatiotemporal dynamic monitoring of groundwater level is of great importance for agricultural precision irrigation, aquifer sustainable management, and ecological environment protection. Traditional groundwater observation mainly relies on surface monitoring well networks. However, the construction and maintenance of monitoring wells are costly, and the spatial distribution of monitoring wells is often sparse and uneven, making it difficult to obtain large-scale, continuous, and high-resolution water level information. Although modern remote sensing technologies such as gravity satellites can provide large-scale terrestrial water storage change information, they have the potential to make up for the shortcomings of surface observation. However, the spatial resolution of the original data is usually rough (kilometer level), which is significantly different from the fine scale (hundred-meter level) required for actual hydrogeological applications, making it difficult to directly guide fine water resource management.

[0003] To solve this problem of spatial scale mismatch, existing technical solutions usually use geostatistical methods such as co-Kriging or geographically weighted regression to try to downscale the coarse resolution remote sensing data to high resolution. However, these traditional methods have inherent limitations when dealing with large scale differences. They are usually based on the simplified assumptions of spatial stationarity and linear correlation, ignoring the complex and nonlinear physical modulation relationship between remote sensing signals and groundwater level. In fact, the distribution of groundwater level is strongly influenced by various static geographic elements such as topography, soil type, and hydrogeological structure, and this influence has great spatial heterogeneity. Due to the lack of effective integration mechanism for these key physical processes and static geographic prior knowledge, existing statistical models can only perform simple mathematical interpolation. This makes it difficult for the model to capture local microscopic hydrological characteristics, especially in vast areas where there is a lack of ground monitoring well constraints. The generated groundwater level map is not only low in accuracy, but also lacks physical consistency, making it difficult to truly reflect the complex structure of the groundwater flow field.

[0004] Therefore, an optimized groundwater level observation scheme is expected. SUMMARY

[0005] To solve the above technical problems, the present application is proposed. The embodiments of the present application provide a groundwater level observation method and system fusing multi-dimensional sensing information.

[0006] According to one aspect of the present application, a groundwater level observation method fusing multi-dimensional sensing information is provided, which comprises: obtaining a coarse resolution remote sensing image and a static geographic data set; Coarse-resolution remote sensing images and static geographic datasets are preprocessed and standardized to obtain standardized remote sensing images and standardized static images. Features are stacked between standardized remote sensing images and standardized static images to obtain multi-dimensional sensor information fusion features; Multidimensional sensor information fusion features are input into the trained U-Net model to obtain a normalized high-resolution prediction map; Based on the normalization parameters, the normalized high-resolution prediction map is denormalized to obtain the final high-resolution groundwater level map.

[0007] According to another aspect of this application, a groundwater level observation system integrating multi-dimensional sensing information is provided, comprising: The data acquisition module is used to acquire coarse-resolution remote sensing images and static geographic datasets; The preprocessing and standardization module is used to preprocess and standardize coarse-resolution remote sensing images and static geographic datasets to obtain standardized remote sensing images and standardized static images. The feature stacking module is used to stack features of standardized remote sensing images and standardized static images to obtain multi-dimensional sensor information fusion features. The high-resolution prediction map acquisition module is used to input multi-dimensional sensor information fusion features into the trained U-Net model to obtain a normalized high-resolution prediction map. The denormalization module is used to denormalize the normalized high-resolution prediction map based on the normalization parameters to obtain the final high-resolution groundwater level map.

[0008] Compared to existing technologies, this approach stacks coarse-resolution dynamic remote sensing data with high-resolution static geographic data using multi-dimensional features, constructing an input feature cube that incorporates spatiotemporal dynamics and physical environmental constraints. Utilizing the encoder-decoder structure and skip connection mechanism of the U-Net deep neural network, it automatically extracts cross-scale spatial correlation features during end-to-end training and learns the complex nonlinear laws governing the modulation of remote sensing signals by the local geographic environment. By introducing static geographic information as a physical prior constraint, the model can accurately infer the spatial distribution details of groundwater levels in areas without ground monitoring wells, based on topography and geological texture. This effectively overcomes the local prediction distortion problems caused by scale mismatch and linear assumptions in traditional statistical methods, achieving precise downscaling reconstruction from kilometer-level remote sensing data to hundred-meter-level high-resolution groundwater level maps, significantly improving the physical consistency and spatial precision of the observation results. Attached Figure Description

[0009] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 This is a flowchart of a groundwater level observation method that integrates multi-dimensional sensor information according to an embodiment of this application; Figure 2 This is a schematic diagram of the data flow of the groundwater level observation method that integrates multi-dimensional sensor information according to an embodiment of this application; Figure 3 This is a flowchart illustrating the preprocessing and standardization of coarse-resolution remote sensing images and static geographic datasets to obtain standardized remote sensing images and standardized static images, according to the groundwater level observation method that integrates multi-dimensional sensor information according to embodiments of this application. Figure 4 This is a flowchart illustrating the spatiotemporal benchmark alignment and rasterization of coarse-resolution remote sensing images and static geographic datasets to obtain remote sensing raster data and static raster data, based on coarse-resolution grid definitions and high-resolution grid definitions, according to the groundwater level observation method that integrates multi-dimensional sensing information according to embodiments of this application. Figure 5 This is a flowchart illustrating the groundwater level observation method for fusing multidimensional sensor information according to an embodiment of this application, which involves stacking features of standardized remote sensing images and standardized static images to obtain multidimensional sensor information fusion features. Figure 6 This is a block diagram of a groundwater level observation system that integrates multi-dimensional sensor information according to an embodiment of this application. Detailed Implementation

[0011] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0012] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0013] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0014] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0015] Existing groundwater level observation technologies, when attempting to fuse coarse-resolution remote sensing data and ground monitoring data, are often limited by linear and spatial stationarity assumptions, making it difficult to effectively handle the significant spatial scale differences between the two. This results in groundwater level maps in areas lacking ground monitoring well constraints that not only produce low local prediction accuracy but also lack physical consistency in representing complex hydrogeological environments. To address this challenge, this application proposes a groundwater level observation method that fuses multi-dimensional sensor information. Therefore, the technical solution of this application proposes a groundwater level observation method that fuses multi-dimensional sensor information, aiming to overcome the limitations of traditional downscaling methods by explicitly modeling sub-grid-level geospatial heterogeneity. Specifically, this scheme first performs rigorous spatiotemporal benchmark alignment and rasterization on coarse-resolution remote sensing images and high-resolution static geographic datasets to construct multi-scale input features. Subsequently, it innovatively introduces a sub-grid heterogeneity feature extraction mechanism to calculate the statistical variation features of high-resolution static attributes within the coarse grid. This heterogeneity feature is then used to dynamically weight the remote sensing signal through an adaptive modulation module, thereby simulating the nonlinear physical modulation effect of local environmental factors such as topography and geology on macroscopic hydrological signals. Finally, the physically enhanced fused features are input into the U-Net model. Utilizing its multi-scale skip connection architecture, while maintaining the macroscopic trend of remote sensing, it accurately reconstructs a high-resolution groundwater level distribution map with fine textures, significantly improving the physical realism and spatial detail of the prediction results.

[0016] Figure 1 This is a flowchart of a groundwater level observation method that integrates multi-dimensional sensor information according to an embodiment of this application. Figure 2 This is a schematic diagram of the data flow in the groundwater level observation method that integrates multi-dimensional sensor information according to an embodiment of this application. Figure 1 and Figure 2As shown, the groundwater level observation method according to an embodiment of this application, which integrates multi-dimensional sensor information, includes the following steps: S100, acquiring coarse-resolution remote sensing images and static geographic datasets; S200, preprocessing and standardizing the coarse-resolution remote sensing images and static geographic datasets to obtain standardized remote sensing images and standardized static images; S300, stacking features of the standardized remote sensing images and standardized static images to obtain multi-dimensional sensor information fusion features; S400, inputting the multi-dimensional sensor information fusion features into a trained U-Net model to obtain a normalized high-resolution prediction map; S500, based on the normalization parameters, performing inverse normalization on the normalized high-resolution prediction map to obtain a final high-resolution groundwater level map.

[0017] Specifically, in step S100, coarse-resolution remote sensing images and static geographic datasets are acquired. It should be noted that, because the dynamic changes in groundwater levels are controlled by both regional hydrological cycles and local geological environmental characteristics, a single-dimensional data source cannot comprehensively characterize the spatiotemporal evolution of groundwater. While coarse-resolution remote sensing data can capture large-scale time-varying signals of water storage, it lacks the spatial details needed to describe local geological structures; and while static geographic data provides detailed surface textures and media properties, it cannot reflect the hydrological state flowing over time. Based on this, the technical solution of this application first acquires coarse-resolution remote sensing images and static geographic datasets to construct a multi-source heterogeneous data foundation containing dynamic time-varying information and static spatial constraints. Through the above processing, the problem of missing information from a single data source can be effectively solved, providing complete physical feature inputs for subsequent deep learning models to mine the nonlinear correlation between surface remote sensing signals and groundwater levels.

[0018] More specifically, in a particular example of this application, the data acquisition process involves extracting specific types of Earth observation products from different specialized databases or observation archives. First, for coarse-resolution remote sensing images reflecting hydrological dynamics, corresponding time-series image files are retrieved from the satellite gravity observation data center or soil moisture observation mission database. These images encompass terrestrial water storage anomaly data obtained from gravity satellite inversion and surface soil moisture data obtained from microwave radiometer inversion, with a spatial resolution of kilometer-level or coarser grid scales, primarily used to characterize the macroscopic fluctuations in water volume within the region. Second, for static geographic datasets reflecting surface environmental characteristics, raster layers with higher spatial resolution are downloaded from geospatial data clouds or geological survey databases. These datasets contain digital elevation model data, soil texture type distribution data, and land cover type data, with a spatial resolution of hundred-meter-level or finer grid scales, used to provide key geological background information such as topographic relief and media permeability, thereby completing the initial aggregation of multidimensional input data.

[0019] Specifically, in step S200, the coarse-resolution remote sensing image and the static geographic dataset are preprocessed and standardized to obtain standardized remote sensing images and standardized static images. It should be noted that, given that the originally acquired coarse-resolution remote sensing images and static geographic datasets typically originate from different observation platforms or mapping tasks, there is a natural geometric mismatch in their spatial reference coordinate systems, data coverage, and raster grid specifications. Furthermore, data with different physical attributes differ significantly in numerical dimensions and dynamic range. This presents a dual obstacle of spatial misalignment and numerical computational instability when directly performing pixel-level multi-source information fusion. Based on this, the technical solution of this application further preprocesses and standardizes the coarse-resolution remote sensing image and the static geographic dataset to obtain standardized remote sensing images and standardized static images. This achieves accurate geometric registration of multi-source data within a unified geospatial framework and maps the physical values ​​of heterogeneous data to a unified dimensionless interval applicable to the model. Through the above processing, the spatiotemporal reference deviation and dimensional influence between multi-source data can be effectively eliminated, ensuring that the data input to the model are strictly corresponding in spatial location and conducive to algorithm convergence in numerical distribution, thereby ensuring the accurate extraction and fusion of multi-scale spatial features by the groundwater level observation model.

[0020] Figure 3 This is a flowchart illustrating the preprocessing and standardization of coarse-resolution remote sensing images and static geographic datasets to obtain standardized remote sensing images and standardized static images, according to the groundwater level observation method integrating multi-dimensional sensor information as described in this application. Figure 3 As shown, step S200 includes: S210, based on the coarse resolution grid definition and the high resolution grid definition, performing spatiotemporal reference alignment and rasterization on the coarse resolution remote sensing image and the static geographic dataset to obtain remote sensing raster data and static raster data; S220, based on the normalization parameter, performing numerical standardization on the remote sensing raster data and the static raster data to obtain standardized remote sensing image and standardized static image.

[0021] In step S210, based on the coarse-resolution grid definition and the high-resolution grid definition, the coarse-resolution remote sensing image and the static geographic dataset are spatiotemporally aligned and rasterized to obtain remote sensing raster data and static raster data. It should be noted that, given that coarse-resolution remote sensing images and static geographic datasets are often collected or stored based on different geographic coordinate reference systems, and their spatial resolution and coverage differ, this spatiotemporal inconsistency makes it impossible to directly establish a pixel-level spatial correspondence between different data sources. Therefore, the technical solution of this application further aligns and rasterizes the coarse-resolution remote sensing image and the static geographic dataset based on the coarse-resolution grid definition and the high-resolution grid definition to obtain remote sensing raster data and static raster data. This transforms all input layers to a standard target coordinate system, and the data is cropped and resampled according to preset grid parameters, thereby constructing a data matrix with strictly aligned geometric positions. Through the above processing, spatial positioning errors caused by projection differences and grid misalignment can be effectively eliminated, ensuring accurate spatial matching between dynamic remote sensing signals and static geographic attributes, laying a solid geometric foundation for subsequent multi-scale feature fusion.

[0022] Figure 4 This is a flowchart illustrating the spatiotemporal alignment and rasterization of coarse-resolution remote sensing images and static geographic datasets to obtain remote sensing raster data and static raster data, based on coarse-resolution and high-resolution grid definitions, according to an embodiment of this application for groundwater level observation methods that integrate multi-dimensional sensor information. Figure 4 As shown, step S210 includes: S211, based on the coarse-resolution grid definition and the high-resolution grid definition, performing coordinate system reprojection on the coarse-resolution remote sensing image and the static geographic dataset to obtain reprojected remote sensing data and reprojected static data; S212, based on the coarse-resolution grid definition and the high-resolution grid definition, performing spatial range cropping on the reprojected remote sensing data and the reprojected static data to obtain cropped remote sensing data and cropped static data; S213, based on the coarse-resolution grid definition and the high-resolution grid definition, generating a target network on the cropped remote sensing data and the cropped static data to obtain remote sensing raster data and static raster data.

[0023] In step S211, based on the coarse-resolution grid definition and the high-resolution grid definition, the coarse-resolution remote sensing image and the static geographic dataset are reprojected to obtain reprojected remote sensing data and reprojected static data. It should be noted that, since multi-source heterogeneous data are often released by different aerospace agencies or surveying departments, the geographic coordinate reference systems embedded in their original storage formats often differ. For example, gravity satellite remote sensing images may use a global geocentric coordinate system, while high-resolution digital elevation models may use a regional projected coordinate system. This inconsistency in spatial reference directly leads to geographic location offsets and geometric distortions when different layers are overlaid, severely hindering pixel-level feature fusion. Therefore, the technical solution of this application further reprojects the coarse-resolution remote sensing image and the static geographic dataset to obtain reprojected remote sensing data and reprojected static data based on the coarse-resolution grid definition and the high-resolution grid definition, thereby forcibly unifying the spatial reference reference of all input data. Through the above processing, spatial misalignment caused by different ellipsoid definitions or projection methods can be effectively eliminated, ensuring that geographic information stored in different data structures can be accurately mapped to the same physical spatial plane, providing a rigorous geometric premise for subsequent grid alignment.

[0024] More specifically, in a concrete example of this application, the implementation process of coordinate system reprojection follows a logical path from metadata parsing to mathematical transformation operations. First, for the input coarse-resolution remote sensing image file and static geographic dataset file, the metadata information in their file headers is parsed to accurately extract the source coordinate system parameters that identify the original spatial reference. These parameters encompass the ellipsoid model, the geodetic datum definition, and the specific projection method code. Second, based on the pre-defined coarse-resolution and high-resolution grid definitions, the unified target coordinate system parameters used in the target application scenario are retrieved. These parameters establish the standard spatial framework of the observation area. Subsequently, a geographic coordinate transformation model is constructed, and rigorous mathematical transformation operations are performed on each spatial location point in the original data. This operation process translates the latitude and longitude or original projected coordinates in the source coordinate system, or the original projected coordinates, according to the datum transformation parameters, performing necessary translations, rotations, and scaling, and remapping them to calculate coordinate values ​​in the target coordinate system. This results in the output of reprojected remote sensing data and reprojected static data with a completely unified spatial datum.

[0025] In step S212, based on the coarse-resolution grid definition and the high-resolution grid definition, the reprojected remote sensing data and the reprojected static data are spatially cropped to obtain cropped remote sensing data and cropped static data. It should be noted that since the original acquired remote sensing images and static geographic data often cover the entire globe or continental region, while actual groundwater level observation tasks focus only on specific hydrogeological units or administrative jurisdictions, retaining redundant data outside the study area not only unnecessarily consumes a large amount of computing resources and storage space, but may also introduce edge effect interference in subsequent convolution operations or interpolation processing. Therefore, the technical solution of this application further performs spatial cropping on the reprojected remote sensing data and the reprojected static data based on the coarse-resolution grid definition and the high-resolution grid definition to obtain cropped remote sensing data and cropped static data. This removes all invalid data falling outside the boundary of the target study area and enforces strict consistency in the spatial breadth of all input layers. The above processing can effectively reduce the computational load during model training and inference, and ensure that multi-source data interact only within the effective geographic spatial overlap range, thereby guaranteeing the spatial accuracy of the observation results.

[0026] More specifically, in a concrete example of this application, the spatial extent clipping process primarily relies on geometric intersection operations. First, from the pre-determined coarse-resolution and high-resolution grid definition data structures, the four boundary coordinate parameters defining the target study area—namely, minimum longitude, maximum longitude, minimum latitude, and maximum latitude—are precisely extracted to construct a closed rectangular spatial bounding box. Subsequently, this bounding box is used as a clipping mask, applied to the reprojected remote sensing data and static data, respectively. For raster data, the row and column number range of the raster matrix is ​​calculated, and rows and columns falling outside the bounding box are directly cut off. For vector data, a spatial intersection algorithm is executed to cut off and discard geometric features connected to the outside of the bounding box boundary, retaining only the geometric features and attribute information inside the bounding box. Finally, clipped remote sensing data and clipped static data with spatial extents perfectly matching the target grid definition are output.

[0027] In step S213, based on the coarse-resolution grid definition and the high-resolution grid definition, target networks are generated for the cropped remote sensing data and the cropped static data to obtain remote sensing raster data and static raster data. It should be noted that although the cropped remote sensing data and the static geographic data achieve geographical consistency, their internal data organization still retains the original physical structure. For example, the remote sensing data may have irregular pixel arrangements or non-standard spatial resolution, while the static geographic data may still exist in a vector topology. This inconsistency in the underlying data form directly hinders the deep learning model from reading and operating the matrix-based feature tensor. Therefore, the technical solution of this application further uses target networks to generate remote sensing raster data and static raster data based on the coarse-resolution grid definition and the high-resolution grid definition to perform rigorous rasterization reconstruction and resampling operations, discretizing continuous or vector geospatial information into a standard pixel matrix that meets the input requirements of the U-Net model. Through the above processing, the pixel-level correspondence between dynamic and static data at their respective target resolutions can be effectively established, and the key physical characteristics of data types with different attributes can be preserved, thereby providing structurally sound data support for constructing a multi-channel feature cube.

[0028] More specifically, in a concrete example of this application, the implementation process of generating the target network adopts differentiated spatial interpolation and discretization strategies based on the different physical attributes of the data. First, based on the number of rows and columns and the cell size defined by the coarse-resolution grid, a target coarse grid matrix is ​​initialized in memory. For cropped remote sensing data with spatially continuous variation characteristics (such as land water storage anomalies), bilinear interpolation or cubic convolution interpolation algorithms are used to smoothly map their values ​​to the center of each cell in the coarse grid matrix, thereby generating remote sensing raster data that retains macroscopic trend characteristics. Second, a target fine grid matrix is ​​initialized based on the high-resolution grid definition. The cropped static geographic data is then classified: for continuous static variables such as digital elevation models, high-order interpolation algorithms are also applied to maintain the gradual details of terrain gradients; while for discrete classification variables such as soil type maps and land use maps, the nearest neighbor interpolation algorithm or the centroid inclusion method is forcibly used to directly assign the category codes of geometrically adjacent spatial locations to the target cells. Numerical averaging is strictly prohibited to prevent the generation of physically meaningless intermediate category values. Through the above steps, the final output consists of remote sensing raster data and static raster data that strictly match the deep learning model design in terms of spatial resolution and matrix dimension.

[0029] In step S220, based on normalization parameters, the remote sensing raster data and static raster data are numerically standardized to obtain standardized remote sensing images and standardized static images. It should be noted that because coarse-resolution remote sensing images (e.g., land water storage anomaly data retrieved from satellite gravity inversion) and static geographic datasets (e.g., digital elevation model data) originate from drastically different physical observation systems, their original values ​​differ significantly in units and orders of magnitude. For example, elevation data may fluctuate within an altitude range of several thousand meters, while gravity anomaly signals representing changes in water storage often vary only within the centimeter range. Directly inputting raw physical data with such vastly different orders of magnitude into a deep learning model can lead to difficulties in convergence during gradient descent optimization, and the neural network's weight updates will tend to focus on large numerical features while ignoring the key hydrogeological laws contained in small numerical features. Based on this, the technical solution of this application further standardizes remote sensing raster data and static raster data using normalization parameters to obtain standardized remote sensing images and standardized static images. This eliminates the dimensional influence between multi-source heterogeneous data and maps all input features to a unified numerical range (such as a closed interval between zero and one). Through the above processing, the instability of model training or prediction bias caused by differences in numerical ranges can be effectively avoided, and the distribution characteristics of input data in the inference stage can be strictly consistent with those in the training stage, thereby ensuring the balanced extraction and accurate regression of groundwater level nonlinear features by the U-Net model.

[0030] More specifically, in a concrete example of this application, the numerical standardization process follows a logical path of parameter reuse and linear transformation. First, the normalized parameter set, statistically compiled and stored during the model training phase, is retrieved. This parameter set records in detail the global maximum and global minimum values ​​of each feature channel (covering terrestrial water storage anomalies, surface soil moisture, digital elevation, and soil type coding, etc.) in historical training samples. Then, for the currently input remote sensing raster data and static raster data, the corresponding extreme value parameters are indexed from the normalized parameter set based on the physical attribute identifiers of each data channel. Next, a minimum-maximum linear transformation operation is performed on all raster pixels within each data channel. During this operation, the original physical value of each pixel is subtracted from the corresponding channel's global minimum value, and the difference is then divided by the difference between the global maximum and global minimum values ​​to calculate the standardized value of that pixel. By traversing all input raster matrices and performing the above operations, standardized remote sensing images and standardized static images with numerical ranges strictly constrained within a closed interval of zero to one are finally generated. This process not only achieves dimensionless data processing, but more importantly, it forces the inference data to follow the same statistical distribution boundary as the training data, providing standardized data input for subsequent feature stacking and efficient inference of neural networks.

[0031] Specifically, in step S300, the standardized remote sensing image and the standardized static image are stacked to obtain multi-dimensional sensor information fusion features. It should be noted that, given that when fusing kilometer-level coarse-resolution remote sensing data with hundred-meter-level high-resolution static geographic data, simply using mean downsampling or direct stitching will erase key sub-grid-level spatial heterogeneity information within the coarse grid (such as local topographic relief or soil texture mixing), and ignore the physical modulation effect of high-resolution surface features on coarse-resolution remote sensing signals (such as rainfall infiltration or gravity water storage signals), deep learning models will struggle to capture nonlinear response patterns in complex hydrogeological environments without explicit physical constraints. Based on this, the technical solution of this application further stacks features of standardized remote sensing images and standardized static images to obtain multi-dimensional sensing information fusion features. The implementation includes first extracting a heterogeneity feature map representing the degree of variation in sub-grid physical attributes, and then using this heterogeneity feature map to adaptively weightedly modulate the standardized remote sensing image to generate a modulated remote sensing image. Finally, the modulated remote sensing image, the original standardized remote sensing image, the downsampled standardized static image, and the heterogeneity feature map are deeply stitched together along the channel dimension. This constructs a high-dimensional feature cube integrating the regional average state, internal spatial complexity index, original observation signal, and enhanced signal corrected by physical laws. Through the above processing, the problem of information loss caused by resolution differences can be effectively overcome, providing explicit quantified spatial heterogeneity priors and physical interaction constraints to the deep learning model, thereby reducing the difficulty of the model resolving complex mapping relationships and improving the physical consistency and accuracy of groundwater level prediction in areas without monitoring well constraints.

[0032] Figure 5 This is a flowchart illustrating the process of stacking features from standardized remote sensing images and standardized static images to obtain multi-dimensional sensor information fusion features, according to an embodiment of this application, for groundwater level observation methods that fuse multi-dimensional sensor information. Figure 5 As shown, step S300 includes: S310, extracting subgrid heterogeneity features from the standardized remote sensing image to obtain a heterogeneity feature map; S320, performing heterogeneity adaptive modulation on the standardized remote sensing image based on the heterogeneity feature map to obtain a modulated remote sensing image; S330, constructing the multidimensional sensing information fusion feature based on the standardized remote sensing image, the standardized static image, the modulated remote sensing image, and the heterogeneity feature map.

[0033] In step S310, sub-grid heterogeneity features are extracted from the standardized remote sensing image to obtain a heterogeneity feature map. It should be noted that existing methods have a fundamental limitation when fusing high-resolution static geographic data with coarse-resolution dynamic remote sensing data to generate groundwater level maps. This limitation stems from the crude compression methods used when downsampling high-resolution static data, such as simple mean pooling, which completely erases the sub-grid-level spatial heterogeneity information within the coarse-grid cells. This processing method ignores the physical modulation relationship between the coarse-resolution remote sensing signal and groundwater level changes. Therefore, the technical solution of this application further extracts sub-grid heterogeneity features from the standardized remote sensing image to obtain a heterogeneity feature map, thereby compensating for the information loss caused by downsampling. This requires explicitly quantifying the degree of spatial variation of physical attributes within each coarse-resolution grid cell. Through the above processing, in the application scenario of groundwater level, this step transforms the complexity of a region, such as topographic relief and soil type mixing, into a concise numerical index, thereby generating a feature map that clearly represents the complexity of each region, thus preserving the key information that was discarded in the original mechanism for subsequent analysis.

[0034] More specifically, in a concrete example of this application, the feature extraction process captures the discrete features of local geographic texture using statistical methods. Specifically, for each geographic feature channel in the input standardized high-resolution still image, the statistical standard deviation of the pixel values ​​of all high-resolution sub-grids covered by each coarse-resolution grid is calculated within the spatial range of that grid. Subsequently, to enhance the nonlinear expressive power of the features and constrain their numerical range, the hyperbolic tangent function is used to activate this standard deviation, thereby generating a geographic heterogeneity index, which constitutes a novel, coarse-resolution heterogeneity feature map. This calculation process is rigorously defined by the following formula: in, This represents the pixel value of the k-th channel in the heterogeneity feature map at the coarse grid coordinates (i,j). It is the value of the k-th channel in the standardized remote sensing image at the high-resolution pixel p. This represents the set of all high-resolution pixels covered by the coarse grid (i,j). It is the standard deviation calculation function. It is an adjustable scaling factor specific to channel k. The hyperbolic tangent function is used. Taking a real-world groundwater hydrogeological environment as an example, suppose a coarse-resolution grid (e.g., 1km x 1km) covers a valley region with complex topography. Within this region, a high-resolution digital elevation model (DEM) shows one half as towering ridges and the other half as low-lying valleys; simultaneously, soil type data indicates that the region is a mixture of highly permeable coarse sand and highly impermeable clay. If only conventional mean downsampling is used, the elevation of this grid will be averaged to a meaningless median, and soil properties will be blurred, thus losing the two key physical facts of dramatic undulations and heterogeneous media. The method in this embodiment calculates the standard deviation of all high-resolution DEM pixels and soil-encoded pixels within the grid. Due to the large numerical fluctuations, the calculated standard deviation will be very high. After activation with the hyperbolic tangent function, the corresponding pixel value of this grid in the heterogeneity feature map will approach 1. This near-saturation value serves as a strong prior signal, clearly indicating to subsequent models that the hydrological response mechanism in this region is extremely complex and cannot be simply applied to rainfall-infiltration models for homogeneous plains. This endows the algorithm with the ability to identify and adapt to complex geological environments.

[0035] In step S320, based on the heterogeneous feature map, the standardized remote sensing image is subjected to heterogeneous adaptive modulation to obtain the modulated remote sensing image. It should be noted that, given the strong physical modulation effect of the spatial distribution of high-resolution geographic features on coarse-resolution remote sensing signals—for example, areas with dramatic topographic relief accelerate surface runoff, thus reducing groundwater recharge efficiency, while flat areas facilitate water infiltration—ignoring this physical relationship and relying solely on implicit learning via neural networks would result in insufficient information density in the model input, making it difficult to capture the true hydrological response under complex geological backgrounds. Therefore, the technical solution of this application further utilizes the heterogeneous feature map to perform heterogeneous adaptive modulation on the standardized remote sensing image to obtain the modulated remote sensing image, thereby establishing a mechanism to simulate the physical modulation effect of high-resolution geographic features on coarse-resolution remote sensing signals. Through the above processing, the remote sensing data can effectively carry the imprint of the local geographic environment, becoming more informative and physically realistic, thus producing an enhanced remote sensing feature map that has been corrected by physical laws and is closer to the actual hydrological response.

[0036] More specifically, in a concrete example of this application, this step performs heterogeneous adaptive modulation. Its core logic is to use the heterogeneous feature map generated in the previous step as a regulator to dynamically and adaptively weight the input standardized new remote sensing image. In practical implementation, a learnable gating mechanism is employed, applying heterogeneous information to the original remote sensing signal through a 1x1 convolution operation to generate the modulated remote sensing image. The modulation process is strictly defined by the following formula: in, For modulated remote sensing images, The original, standardized remote sensing image. This is a heterogeneity feature map. The symbol represents the Hadamard product, which is element-wise multiplication. The symbol represents the convolution operation, while This is a 1x1 convolutional kernel that can be learned during model training. This step allows the model to autonomously learn physical laws. Taking an alluvial fan region with complex terrain features as an example, the terrain in this region has large undulations, resulting in high elevation heterogeneity. Therefore, the calculated heterogeneous feature map... The value of this channel is relatively large. During model training, the algorithm will automatically adjust the parameters to make its corresponding... The weights should be negative. Calculated using the above formula, this will weaken the remote sensing signal of the total water storage in the area. This mathematical operation physically simulates the real hydrological process where runoff exceeds infiltration under drastically undulating terrain, meaning that most rainfall is rapidly lost and not converted into groundwater storage. Therefore, the model does not mechanically read remote sensing values, but intelligently corrects the observed signals based on the complexity of the geological environment, thus enabling inferences consistent with hydrogeological common sense even in areas lacking actual well logging data.

[0037] In step S330, the multi-dimensional sensor information fusion feature is constructed based on standardized remote sensing imagery, standardized static imagery, modulated remote sensing imagery, and heterogeneous feature maps. It should be noted that although the preceding steps generate static data representing the average state of the region, remote sensing data representing macroscopic hydrological changes, heterogeneous feature maps representing subgrid spatial complexity, and modulated remote sensing imagery containing physical interaction laws, if these features exist independently in a dispersed or unstructured form, the deep learning model will find it difficult to automatically establish deep coupling relationships between them. This could lead to the model falling into a local extremum trap, relying solely on a single strong signal while ignoring subtle physical constraints. Therefore, the technical solution of this application further constructs the multi-dimensional sensor information fusion feature based on standardized remote sensing imagery, standardized static imagery, modulated remote sensing imagery, and heterogeneous feature maps. This is used to construct an enhanced feature cube, integrating various valuable information generated in the preceding steps with the original information, providing a rich and comprehensive input for the subsequent deep learning model. Through the above processing, a well-designed and highly structured dataset is provided for the model, thereby reducing the difficulty of the model learning complex mapping relationships and ultimately improving the generation accuracy and reliability of high-resolution groundwater level maps. In particular, in areas lacking ground monitoring well constraints, the physical consistency and accuracy of the prediction results will be improved.

[0038] More specifically, in a concrete example of this application, the feature construction process performs a rigorous tensor stitching operation. Specifically, the modulated remote sensing image, the original normalized remote sensing image, the normalized static image after conventional downsampling, and the heterogeneous feature map are stitched along the channel dimension of the data. This operation aligns and stacks spatial information with different physical properties, sources, and levels of abstraction at each corresponding grid location, thereby forming a multi-channel feature matrix. This stitching operation is rigorously defined by the following mathematical formula: in, It is a feature of multi-dimensional sensor information fusion. This represents a splicing operation along the channel dimension. This is the downsampling function. This step provides the model with a three-dimensional, multi-angle view at every spatial location: the average condition of the existing region (from...). It also has its own internal complexity (from...) ), including both original satellite observation signals (from There are also signals modulated by local physical laws (from...). In this way, the constructed enhanced input features This allows for a more comprehensive and insightful description of the observed area to be provided to deep learning models. Taking a groundwater monitoring scenario covering the boundary between a river alluvial plain and a bedrock mountainous area as an example, the constructed fusion feature cube contains four sets of key values ​​at each 1km x 1km grid point: the first set of data (static imagery) tells the model that the average elevation of the grid is decreasing; the second set of data (raw remote sensing) shows an abnormal reduction in the overall gravity field of the area, indicating water loss; the third set of data (heterogeneity features) shows high values ​​at the boundary, suggesting that the geological structure here is extremely fragmented and heterogeneous; the fourth set of data (modulated remote sensing) enhances or suppresses the gravity signal based on the aforementioned heterogeneity. When the model processes this feature cube, it no longer simply sees a decrease in water volume, but understands, in the context, that in a geologically fragmented piedmont area (high heterogeneity), although the gravity signal is weakened, the actual drop in groundwater level may be much greater than inferred from the gravity signal alone due to the presence of rapid lateral runoff (implied by the modulated signal). This deep information fusion enables the model to distinguish between a drop in actual water level caused by pumping and a signal attenuation caused by topographic factors, thereby accurately outputting a groundwater level distribution map of the complex area.

[0039] Specifically, in step S400, the multidimensional sensor information fusion features are input into the trained U-Net model to obtain a normalized high-resolution prediction map. It should be noted that due to the extremely complex nonlinear coupling relationship between groundwater level and multidimensional sensor information (such as gravity remote sensing signals, topography, soil texture, and their spatial heterogeneity), and the transition from coarse-resolution macroscopic observation to high-resolution microscopic distribution is a typical ill-conditioned inversion problem, traditional linear regression models or geostatistical interpolation methods are difficult to accurately analyze this cross-scale physical mapping mechanism without sufficient ground control point constraints. Based on this, the technical solution of this application further inputs the multidimensional sensor information fusion features into the trained U-Net model to obtain a normalized high-resolution prediction map, thereby utilizing the powerful feature extraction and nonlinear fitting capabilities of deep convolutional neural networks to perform end-to-end downscaling reconstruction and parameter inversion on the input enhanced feature cube. Through the above processing, the pre-trained weight parameters inside the model can be effectively activated, accurately capturing and recovering the spatial details and gradient changes of the groundwater level at the sub-grid scale, thereby outputting a normalized groundwater level distribution map that is consistent with the static geographic data in spatial resolution and driven by remote sensing data in physical trends.

[0040] More specifically, in a concrete example of this application, the prediction process performs a forward propagation operation based on a deep neural network, encompassing the entire chain from feature encoding compression to spatial detail reconstruction. First, the multidimensional sensor information fusion features (including batch size, number of channels, height, and width) in the form of a four-dimensional tensor constructed in the preceding steps are fed as input layer data to the encoder path of the U-Net architecture. In the encoder stage, the data flows through a series of stacked convolutional and pooling layers, with the model extracting abstract features from low-level texture to high-level semantics layer by layer, while progressively compressing the spatial size of the feature maps to expand the receptive field, thereby capturing regional trends in groundwater storage changes. Subsequently, in the decoder reconstruction stage, the spatial resolution of the feature maps is gradually restored through upsampling operations, and a skip connection mechanism is used to directly concatenate the high-resolution spatial detail features (such as boundary information determined by terrain) preserved at the corresponding encoder layer into the decoding path along the channel dimension. Finally, after layer-by-layer fusion and convolution processing by the decoder, the output layer maps and regresses the multi-dimensional depth features into a two-dimensional matrix through a single-channel convolution kernel, generating a normalized high-resolution prediction map with a numerical range strictly limited to zero to one. Each pixel value in this map represents the predicted groundwater depth of the corresponding geographical location in the normalized space.

[0041] Specifically, based on the technical solution of this application embodiment, the training of the U-Net model is a supervised learning-based iterative parameter optimization process. Its core objective is to establish a complex nonlinear mapping relationship between multi-dimensional sensor information fusion features and high-resolution groundwater level distribution. Specifically, the training dataset consists of pairs of input features and target ground values. The input is not the original data, but an enhanced feature cube constructed through previous steps. This cube deeply integrates physically modulated remote sensing images, original standardized remote sensing images, downsampled static geographic images, and heterogeneous feature maps representing subgrid spatial complexity in the channel dimension. The target ground value is a high-resolution groundwater level reference map generated using geostatistical methods such as Kriging interpolation based on sparsely distributed measured water level data from monitoring wells. Although this reference map exhibits a certain smoothing effect in areas without wells, it provides crucial global trend guidance and strong local well location constraints for the model, enabling the model to perform feature learning in a supervised environment.

[0042] During the forward propagation phase of training, multidimensional feature tensors are fed into the U-Net network architecture. The data flow first passes through the encoder path, extracting multi-scale spatial features from texture details to semantic abstractions layer by layer through stacked convolution and pooling operations, while compressing the spatial size to capture regional hydrogeological trends. Notably, the heterogeneity adaptive modulation mechanism proposed in this application includes a key learnable parameter, the 1x1 convolution kernel $W$, used to control the weights of heterogeneous features. During training, this parameter participates in the forward computation as part of the computation graph. Through continuous trial and error and adjustment, the model automatically learns how to enhance or suppress remote sensing signals under different geological heterogeneity conditions. Subsequently, in the decoder path, the feature maps are upsampled to gradually restore spatial resolution, and the high-frequency boundary information retained in the encoder is directly fused to the reconstruction layer using a skip connection mechanism. Finally, a normalized high-resolution prediction map is generated through the output layer.

[0043] During the error calculation and parameter update phase, the predicted image output by the model is compared pixel-by-pixel with the corresponding ground truth value. Mean squared error is typically used as the loss function to quantify the deviation between the two. Based on the calculated loss value, the gradient of the loss function with respect to all learnable parameters in the network is calculated using the backpropagation algorithm. These parameters not only cover the tens of thousands of convolutional kernel weights and bias terms within the U-Net network, but also explicitly include the weight parameters in the heterogeneous modulation module. Subsequently, the optimizer updates all parameters synchronously along the error descent direction based on this gradient information, and can set a learning rate decay strategy to ensure the stability of model convergence. After multiple cycles of forward prediction-error evaluation-gradient backpropagation-weight update iterations, the model gradually converges, ultimately learning how to use static geographic prior knowledge to physically correct dynamic remote sensing signals, and accurately reconstruct the corrected multi-source information into a high-precision spatial distribution of groundwater level.

[0044] Specifically, in step S500, the normalized high-resolution prediction map is inversely normalized based on the normalization parameters to obtain the final high-resolution groundwater level map. It should be noted that, to ensure numerical stability and convergence efficiency during gradient descent optimization, the prediction results generated by the output layer of deep convolutional neural networks are typically strictly constrained to a specific dimensionless numerical range (e.g., a closed interval between zero and one). This results in the model's directly output values ​​representing only relative intensity and failing to intuitively reflect the true physical depth of groundwater (e.g., meters). Therefore, the technical solution of this application further inversely normalizes the normalized high-resolution prediction map based on the normalization parameters to obtain the final high-resolution groundwater level map, thereby performing an inverse mapping transformation of the numerical domain and restoring the relative values ​​output by the model to absolute water level values ​​with definite physical units. Through the above processing, the physical attributes and dimensional meaning of the observed data can be effectively restored, outputting spatial distribution results that conform to hydrogeological industry standards, thus directly supporting quantitative assessment of groundwater resources, delineation of over-extraction areas, and refined management decisions.

[0045] More specifically, in a concrete example of this application, the inverse normalization process strictly follows the inverse operation logic of linear transformation, aiming to achieve accurate regression from the abstract feature space to the physical observation space. First, the normalization parameters of the groundwater level target variable, statistically determined and locked during the data preparation phase before model training, are retrieved from persistent storage. These parameters precisely record the global minimum and global maximum values ​​of the measured groundwater level data in the historical training sample set, thus establishing the bijective transformation boundary between the physical values ​​and the normalized values. Subsequently, for each independent raster cell in the normalized high-resolution prediction map output by the U-Net model, pixel-level inverse linear calculations are performed. This calculation involves multiplying the normalized predicted value of the current cell by the difference between the global maximum and global minimum values ​​(i.e., the physical range), and adding the resulting product to the global minimum value to calculate the true physical water level value corresponding to that location. Finally, all image data that has been numerically restored is reorganized and formatted according to the spatial coordinate reference system defined by the high-resolution grid to generate a final high-resolution groundwater level map with complete georeferenced information. This layer accurately quantifies the groundwater depth readings under each subdivided grid in the study area.

[0046] In summary, the groundwater level observation method integrating multi-dimensional sensor information according to the embodiments of this application is explained. It stacks coarse-resolution dynamic remote sensing data with high-resolution static geographic data to construct an input feature cube that includes spatiotemporal dynamics and physical environmental constraints. Utilizing the encoder-decoder structure and skip connection mechanism of the U-Net deep neural network, it automatically extracts cross-scale spatial correlation features during end-to-end training and learns the complex nonlinear laws governing the modulation of remote sensing signals by the local geographic environment. By introducing static geographic information as a physical prior constraint, the model can accurately infer the spatial distribution details of groundwater levels in areas without ground monitoring wells based on topography and geological texture. This effectively overcomes the local prediction distortion problems caused by scale mismatch and linear assumptions in traditional statistical methods, achieving accurate downscaling reconstruction from kilometer-level remote sensing data to hundred-meter-level high-resolution groundwater level maps, significantly improving the physical consistency and spatial precision of the observation results.

[0047] Furthermore, a groundwater level observation system that integrates multi-dimensional sensor information is also provided.

[0048] Figure 6 This is a block diagram of a groundwater level monitoring system that integrates multi-dimensional sensor information according to an embodiment of this application. Figure 6 As shown, the groundwater level observation system 100 according to an embodiment of this application, which integrates multi-dimensional sensor information, includes: a data acquisition module 110, used to acquire coarse-resolution remote sensing images and static geographic datasets; The preprocessing and standardization module 120 is used to preprocess and standardize coarse-resolution remote sensing images and static geographic datasets to obtain standardized remote sensing images and standardized static images. The feature stacking module 130 is used to stack features of standardized remote sensing images and standardized static images to obtain multi-dimensional sensing information fusion features. The high-resolution prediction map acquisition module 140 is used to input multi-dimensional sensor information fusion features into the trained U-Net model to obtain a normalized high-resolution prediction map. The denormalization module 150 is used to denormalize the normalized high-resolution prediction map based on the normalization parameters to obtain the final high-resolution groundwater level map.

[0049] As described above, the groundwater level observation system 100 integrating multi-dimensional sensor information according to the embodiments of this application can be implemented in various types of computing devices or control units. For example, it can be implemented in high-performance servers deployed in hydrological monitoring centers, data processing nodes running on cloud computing platforms, or professional graphics processing units integrated into geographic information system workstations. In one possible implementation, the groundwater level observation system 100 integrating multi-dimensional sensor information according to the embodiments of this application can be integrated into the computing device as a software module and / or hardware module. For example, the groundwater level observation system 100 integrating multi-dimensional sensor information can be a resident data processing service in the operating system of the computing device. This software module is configured to perform spatiotemporal benchmark alignment of coarse-resolution remote sensing and static geographic data, heterogeneous feature extraction based on sub-grid statistics, adaptive modulation of remote sensing signals guided by physical priors, and cross-scale water level extrapolation based on the U-Net model. Alternatively, it can be a dedicated regional water resources remote sensing monitoring algorithm program developed for the computing device. Of course, the groundwater level observation system 100 that integrates multi-dimensional sensing information can also be one of the many hardware modules of the computing device or control unit, or it can be embedded in the graphics processing unit or tensor processing unit to accelerate the deep convolution operation and high-dimensional feature matrix processing in parallel, or it can be used as a hydrogeological parameter inversion acceleration card for a specific application.

[0050] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A groundwater level observation method that fuses multi-dimensional sensing information, characterized by, The method comprises the following steps: obtaining coarse resolution remote sensing images and static geographic data sets; preprocessing and standardizing the coarse resolution remote sensing images and the static geographic data sets to obtain standardized remote sensing images and standardized static images; stacking features of the standardized remote sensing images and the standardized static images to obtain multi-dimensional sensor information fusion features; inputting the multi-dimensional sensor information fusion features into a trained U-Net model to obtain a normalized high-resolution prediction map; based on the normalization parameters, the normalized high-resolution prediction map is de-normalized to obtain a final high-resolution groundwater level map.

2. The method of claim 1, wherein the fusion of multi-dimensional sensing information of groundwater level observation is characterized by, The preprocessing and standardization of the coarse resolution remote sensing images and the static geographic data sets to obtain the standardized remote sensing images and the standardized static images comprise: based on the coarse resolution grid definition and the high resolution grid definition, the coarse resolution remote sensing images and the static geographic data sets are aligned in time and space and rasterized to obtain remote sensing raster data and static raster data; based on the normalization parameters, the remote sensing raster data and the static raster data are numerically standardized to obtain the standardized remote sensing images and the standardized static images.

3. The method of claim 2, wherein the fusion of multi-dimensional sensor information is performed by a neural network. Based on the coarse resolution grid definition and the high resolution grid definition, the coarse resolution remote sensing images and the static geographic data sets are aligned in time and space and rasterized to obtain remote sensing raster data and static raster data, comprising: based on the coarse resolution grid definition and the high resolution grid definition, the coarse resolution remote sensing images and the static geographic data sets are re-projected in the coordinate system to obtain re-projected remote sensing data and re-projected static data; based on the coarse resolution grid definition and the high resolution grid definition, the re-projected remote sensing data and the re-projected static data are spatially cropped to obtain cropped remote sensing data and cropped static data; based on the coarse resolution grid definition and the high resolution grid definition, the cropped remote sensing data and the cropped static data are target network generated to obtain remote sensing raster data and static raster data.

4. The method of claim 1, wherein the fusion of multi-dimensional sensing information of groundwater level observation is characterized by, The feature stacking of the standardized remote sensing images and the standardized static images to obtain multi-dimensional sensor information fusion features comprises: extracting sub-grid heterogeneity features from the standardized remote sensing images to obtain a heterogeneity feature map; based on the heterogeneity feature map, the standardized remote sensing images are adaptively modulated based on the heterogeneity to obtain modulated remote sensing images; based on the standardized remote sensing images, the standardized static images, the modulated remote sensing images and the heterogeneity feature map, the multi-dimensional sensor information fusion features are constructed.

5. The method of claim 4, wherein the fusion of multi-dimensional sensor information is performed by a neural network. The sub-grid heterogeneity feature extraction from the standardized remote sensing images to obtain the heterogeneity feature map comprises: the sub-grid heterogeneity feature extraction from the standardized remote sensing images is performed according to the following formula: wherein, represents the pixel value of the kth channel of the heterogeneous feature map at the coarse grid coordinate (i,j), is the value of the kth channel of the normalized remote sensing image at the high-resolution pixel point p, represents the set of all high-resolution pixel points covered by the coarse grid (i,j), is the standard deviation calculation function, is a tunable scaling factor specific to the channel k, is the hyperbolic tangent function.

6. The method of claim 4, wherein the fusion of multi-dimensional sensing information of groundwater level observation is characterized by, based on the heterogeneity feature map, the standardized remote sensing images are adaptively modulated based on the heterogeneity to obtain modulated remote sensing images, comprising: the standardized remote sensing images are adaptively modulated based on the heterogeneity according to the following formula: wherein, is a post-modulation remote sensing image, is an original normalized remote sensing image, is a heterogeneity feature map, denotes a Hadamard product, represents a convolution operation, is a 1x1 convolution kernel that is learnable during model training.

7. The method of claim 4, wherein the fusion of multi-dimensional sensor information is performed by a neural network. based on the standardized remote sensing images, the standardized static images, the modulated remote sensing images and the heterogeneity feature map, the multi-dimensional sensor information fusion features are constructed, comprising: the multi-dimensional sensor information fusion features are constructed according to the following formula: wherein, is a multi-dimensional sensor information fusion feature, represents a concatenation operation along the channel dimension, is a down-sampling function.

8. A groundwater level observation system that fuses multi-dimensional sensing information, characterized by, The method comprises the following steps: The data acquisition module is configured to acquire the coarse-resolution remote sensing image and the static geographic data set. The preprocessing and standardization module is configured to preprocess and standardize the coarse-resolution remote sensing image and the static geographic data set to obtain a standardized remote sensing image and a standardized static image. The feature stacking module is configured to stack features of the standardized remote sensing image and the standardized static image to obtain multi-dimensional sensor information fusion features. The high-resolution prediction map acquisition module is configured to input the multi-dimensional sensor information fusion features into a trained U-Net model to obtain a normalized high-resolution prediction map. The inverse normalization module is configured to perform inverse normalization on the normalized high-resolution prediction map based on a normalization parameter to obtain a final high-resolution groundwater level map.

9. The groundwater water level observation system that fuses multi-dimensional sensing information according to claim 8, characterized by, The preprocessing and standardization module comprises: The spatio-temporal reference alignment and rasterization unit is configured to perform spatio-temporal reference alignment and rasterization on the coarse-resolution remote sensing image and the static geographic data set based on a coarse-resolution grid definition and a high-resolution grid definition to obtain remote sensing raster data and static raster data. The numerical standardization unit is configured to perform numerical standardization on the remote sensing raster data and the static raster data based on a normalization parameter to obtain the standardized remote sensing image and the standardized static image.

10. The groundwater water level observation system that fuses multi-dimensional sensing information according to claim 8, characterized by, The feature stacking module comprises: The sub-grid heterogeneity feature extraction unit is configured to extract sub-grid heterogeneity features from the standardized remote sensing image to obtain a heterogeneity feature map. The heterogeneity adaptive modulation unit is configured to perform heterogeneity adaptive modulation on the standardized remote sensing image based on the heterogeneity feature map to obtain a modulated remote sensing image. The multi-dimensional sensor information fusion feature construction module is configured to construct the multi-dimensional sensor information fusion features based on the standardized remote sensing image, the standardized static image, the modulated remote sensing image, and the heterogeneity feature map.

Images