Groundwater storage component separation method, computer device, and readable storage medium

Abstract

The application discloses a groundwater storage component separation method, a computer device and a readable storage medium, and belongs to the technical field of artificial intelligence and hydrological geographic information. The method comprises the following steps: acquiring multi-source basic data of a target area; constructing a shallow groundwater storage training label by using a spatial division method with a regional weight in combination with a water yield parameter; performing spatiotemporal consistency preprocessing on the multi-source data; constructing a three-dimensional spatiotemporal input tensor by using a sliding window method in combination with a historical storage lag feature channel; inputting a three-dimensional convolutional neural network with a spatial attention mechanism to output a shallow groundwater storage prediction result; and separating deep groundwater storage changes based on a land water storage water balance equation. The application breaks through the dependence of component separation on high-density monitoring wells, overcomes the insufficient coupling of small sample training and spatiotemporal features, and realizes accurate splitting of shallow and deep groundwater.

Description

Technical Field

[0001] This application belongs to the interdisciplinary field of artificial intelligence and hydrological geographic information, specifically relating to a method for separating groundwater storage components, computer equipment, and a readable storage medium. Background Technology

[0002] Groundwater, as a crucial component of the water resource system, plays a vital role in ecological maintenance, agricultural irrigation, and domestic water use. Affected by both climate change and increasingly intense human activities, regional groundwater resources are showing a significant downward trend, leading to a series of ecological and environmental problems such as continuous decline in groundwater levels, land subsidence, and wetland degradation. Therefore, how to conduct refined monitoring of the spatiotemporal changes in regional groundwater storage (GWS) has become a critical scientific issue urgently needing to be addressed in the field of hydrology and water resources. In recent years, the development of gravity satellite observation technology has provided a new means for dynamic monitoring of regional groundwater storage. Based on the total change data of land water storage retrieved by gravity satellites, after deducting non-groundwater components such as soil water, snow water equivalent, and canopy water, the changes in total regional groundwater storage can be retrieved. This overcomes the spatial scale limitations of traditional ground observations and provides a direct and reliable data foundation for large-scale groundwater dynamics research.

[0003] However, existing groundwater storage monitoring technologies based on gravity satellite observations still have the following shortcomings: First, the groundwater storage change signals obtained by gravity satellite inversion are superimposed responses of deep and shallow groundwater, and it is difficult to directly separate the components of deep and shallow groundwater. Existing component separation methods generally rely on high-density well observation data, which is difficult to effectively promote and apply in areas with scarce well observation data. Second, the spatiotemporal changes in groundwater storage have strong lag response and spatial heterogeneity. Existing recurrent neural network models based on long short-term memory networks or deep learning models based on two-dimensional convolutional neural networks are difficult to simultaneously capture long-term lag replenishment effects of more than six months and spatial correlation features between different grids in the region, resulting in limited prediction accuracy. Third, the time span of gravity satellite observation data is relatively limited. Under small sample time-series conditions, the training difficulty of deep learning models is relatively large and the generalization ability is relatively weak, making it difficult to guarantee the accuracy and reliability of groundwater component separation results. Summary of the Invention

[0004] The purpose of this application is to provide a method, computer equipment, and readable storage medium for separating groundwater storage components. This method can effectively overcome the long-term lag and spatial heterogeneity of hydrological systems, solve the technical problems of existing technologies such as reliance on high-density monitoring data, difficulty in training with small samples, and insufficient coupling of spatiotemporal features, and achieve accurate separation of shallow and deep groundwater storage components, providing reliable technical support for large-scale dynamic monitoring and sustainable management of groundwater.

[0005] To solve the above-mentioned technical problems, this application is implemented as follows: In a first aspect, embodiments of this application provide a method for separating components of groundwater reserves, the method comprising: Acquire multi-source basic data for the target area; wherein, the multi-source basic data includes: total change data of land water storage retrieved by gravity satellite, hydrological variable data output by land surface hydrological model, meteorological and remote sensing driving factor data, and measured water level data of shallow groundwater monitoring wells; Based on the spatial distribution of the monitoring wells, shallow groundwater storage training labels are calculated using a spatial partitioning method with regional weights combined with the water yield parameters of the monitoring wells. The multi-source basic data is subjected to spatiotemporal consistency preprocessing to obtain processed data that matches the spatiotemporal benchmark of the total change data of land water storage; wherein, the preprocessing includes spatial resampling, temporal alignment, outlier calculation of the benchmark period, and standardization processing; Based on the processed data, a spatiotemporal feature sample set is constructed using the sliding window method; each sample uses historical shallow groundwater storage data as a lag feature channel, which is then concatenated with multiple meteorological driving factor channels to form a three-dimensional spatiotemporal input tensor. The three-dimensional spatiotemporal input tensor is input into a three-dimensional convolutional neural network with a spatial attention mechanism. Spatiotemporal features are extracted through three-dimensional convolutional layers. The spatial attention mechanism is used to dynamically assign weights to the extracted spatiotemporal features and output the shallow groundwater storage prediction results. Based on the accuracy index of the shallow groundwater storage prediction results, the hyperparameters of the three-dimensional convolutional neural network are optimized to obtain the optimized shallow groundwater storage prediction results. Using the land water storage balance equation as a physical constraint, the changes in deep groundwater storage are obtained by deducting non-groundwater hydrological components and the optimized shallow groundwater storage prediction results from the total land water storage change data, so as to complete the component separation of shallow groundwater and deep groundwater.

[0006] In a second aspect, embodiments of this application provide a computer device including a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the method described in the first aspect.

[0007] Thirdly, embodiments of this application provide a computer-readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method described in the first aspect.

[0008] Fourthly, embodiments of this application also provide a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described in the first aspect.

[0009] In summary, this application has the following beneficial effects: First, this application constructs a shallow groundwater storage training label system with a clear physical mechanism and strict observation constraints by integrating multi-source basic data, including total change data of land water storage retrieved by gravity satellite, hydrological variable data output by land surface hydrological models, and data from shallow groundwater monitoring wells. This breaks through the strong dependence of traditional methods on high-density groundwater observation well data, enabling the component separation method to still be stably applied in areas where water well observation data is relatively sparse. Second, this application addresses the problem of high training difficulty and weak generalization ability of deep learning models under small sample time series background. It adopts the sliding window method to expand the spatiotemporal samples and injects historical shallow groundwater storage data as the lag feature channel into the model input. It works in conjunction with the meteorological driving factor channel, so that the model can fully learn the hydrological integral characteristics and lag response law of the groundwater system with a limited sample size, which significantly improves the accuracy and stability of shallow groundwater storage prediction. Third, this application uses a three-dimensional convolutional neural network with spatial attention mechanism to simultaneously extract grid association features in the spatial dimension and intertemporal lag features in the temporal dimension, and dynamically assigns weights to the core hydrological change area. This overcomes the technical defects of traditional two-dimensional convolutional neural networks and recurrent neural networks, which are difficult to couple spatiotemporal features and tend to converge to the regional average value. It can accurately capture the dynamic change process of shallow groundwater storage peaks and troughs. Fourth, this application uses the land water storage balance equation as a physical constraint, organically combines gravity satellite observation data, land surface hydrological model output and high-precision shallow groundwater storage prediction results, and sequentially deducts soil water, snow water equivalent, canopy water and shallow groundwater storage changes from the total land water storage changes, and finally separates the deep groundwater storage changes, realizing the accurate separation of the two components of shallow groundwater and deep groundwater. Fifth, the groundwater storage component separation method provided in this application has good portability and scalability, and can provide reliable technical support for the fine monitoring of groundwater resources in large-scale areas, the sustainable development and utilization management of groundwater, and the assessment of water resource security under the influence of climate change. Attached Figure Description

[0010] Figure 1 This is a schematic flowchart of a method for separating groundwater storage components provided in some embodiments of this application; Figure 2 These are internal structural diagrams of a computer device provided in some embodiments of this application. Detailed Implementation

[0011] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0012] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0013] In one exemplary embodiment, this application proposes a method for separating groundwater storage components. The method for separating groundwater storage components provided in this application will be described in detail below with reference to the accompanying drawings and specific embodiments and application scenarios.

[0014] Reference Figure 1 The method includes steps 102-114. Wherein: Step 102: Obtain multi-source basic data for the target area; wherein, the multi-source basic data includes: total change data of land water storage retrieved by gravity satellite, hydrological variable data output by land surface hydrological model, meteorological and remote sensing driving factor data, and measured water level data of shallow groundwater monitoring wells.

[0015] Step 104: Based on the spatial distribution of the monitoring wells, the shallow groundwater storage training label is calculated by using a spatial partitioning method with regional weights and combining the water yield parameters of the monitoring wells.

[0016] Step 106: Perform spatiotemporal consistency preprocessing on the multi-source basic data to obtain processed data that matches the spatiotemporal benchmark of the total change data of land water storage; wherein, the preprocessing includes spatial resampling, temporal alignment, outlier calculation of the benchmark period, and standardization processing.

[0017] Step 108: Based on the processed data, a spatiotemporal feature sample set is constructed using the sliding window method; wherein, each sample uses historical shallow groundwater storage data as a lag feature channel and is spliced ​​with multiple meteorological driving factor channels to form a three-dimensional spatiotemporal input tensor.

[0018] Step 110: Input the three-dimensional spatiotemporal input tensor into a three-dimensional convolutional neural network with a spatial attention mechanism, extract spatiotemporal features through the three-dimensional convolutional layer, dynamically assign weights to the extracted spatiotemporal features through the spatial attention mechanism, and output the shallow groundwater storage prediction result.

[0019] Step 112: Based on the accuracy index of the shallow groundwater storage prediction result, perform hyperparameter optimization on the three-dimensional convolutional neural network to obtain the optimized shallow groundwater storage prediction result.

[0020] Step 114: Using the land water storage balance equation as a physical constraint, the changes in deep groundwater storage are obtained by deducting non-groundwater hydrological components and the optimized shallow groundwater storage prediction results from the total land water storage change data, so as to complete the component separation of shallow groundwater and deep groundwater.

[0021] In this embodiment, the target area can be an administrative region, river basin, or hydrological unit with a clear geographical boundary. For example, the target area is the Beijing-Tianjin-Hebei region, which has a latitude and longitude range of approximately 36.0°N to 43.0°N and 113.0°E to 120.0°E.

[0022] The specific sources and connotations of the aforementioned multi-source basic data can be explained as follows: The total change data of land water storage retrieved by the gravity satellite is, for example, the total change data of land water storage obtained by the Gravity Recovery and Climate Experiment (GRACE) or GRACE-FO gravity satellite. It can be selected from the GRACE RL06 CSR mascon water storage dataset released by the Space Research Center of the University of Texas, or other versions of gravity satellite products released by institutions such as NASA's Jet Propulsion Laboratory and the Potsdam Geosciences Research Centre in Germany. The spatial resolution of the total change data of land water storage is 0.25° and the temporal resolution is 1 month.

[0023] The hydrological variable data output by the land surface hydrological model, exemplarily, uses monthly hydrological variable data output by the GLDAS Noah land surface hydrological model, including soil water changes at different depth levels, specifically including soil water changes at four depths: 0 to 10 cm, 10 to 40 cm, 40 to 100 cm, and 100 to 200 cm. The total soil water change is obtained by summing these data. This also includes changes in snow water equivalent. With changes in canopy water The spatial resolution is exemplarily uniformly set to 0.25°.

[0024] The meteorological and remote sensing driving factor data, exemplarily including at least one of monthly precipitation, air temperature, potential evapotranspiration, surface temperature, and normalized vegetation index, can be obtained from publicly available meteorological reanalysis products and remote sensing data platforms, with a spatial resolution exemplarily uniform at 0.25°.

[0025] The measured water level data, exemplarily, are monthly measured water level data from monitoring wells of effective shallow groundwater within the study area, used as a calibration benchmark for model training.

[0026] It should be noted that the above selection of data sources and spatial resolution is merely illustrative and is not intended to limit the scope of this application. Those skilled in the art may select other equivalent gravity satellite products, land surface hydrological model products, or meteorological remote sensing data products according to actual needs when implementing this application.

[0027] This embodiment integrates multi-source basic data from gravity satellite observations, land surface hydrological model simulations, and ground monitoring well measurements. It fully leverages the complementarity of different data sources in terms of spatial scale, physical mechanisms, and observational accuracy, providing a data foundation with clear physical mechanisms and strict observational constraints for the subsequent separation of shallow and deep groundwater storage components. This fundamentally overcomes the technical bottleneck of traditional groundwater monitoring methods that rely solely on a single observation method, resulting in limited data sources and insufficient spatial coverage. Furthermore, the systematic integration of multi-source basic data provides a multi-dimensional, multi-physical quantity input feature and label system for the training and validation of deep learning models, effectively improving the accuracy and robustness of subsequent prediction models.

[0028] In some embodiments, the region-weighted spatial partitioning method is the Thiessen polygon method; the step of calculating shallow groundwater storage training labels based on the spatial distribution of the monitoring wells, using the region-weighted spatial partitioning method combined with the water yield parameters of the monitoring wells, includes: The area of ​​the Thiessen polygon corresponding to each monitoring well is determined based on the Thiessen polygon method.

[0029] Based on the area and water yield parameters of the monitoring well, the measured water level change values ​​of the monitoring well are converted into gridded shallow groundwater storage training labels.

[0030] The calculation formula (1) for the gridded shallow groundwater storage training labels is as follows: in, For the first Measured water level changes of each monitoring well; For the first The area of ​​a Thiessen polygon. For the first The water supply parameters of the region corresponding to each Thiessen polygon; This represents the total area of ​​the grid.

[0031] It should be noted that the core principle of the Thiessen polygon method is based on the spatial partitioning and interpolation of discrete points. Its essence is to divide the study area into several non-overlapping, continuously covering polygons. Each polygon contains only one discrete observation point, i.e., one monitoring well, and the distance from any position within the polygon to that observation point is less than the distance to any other observation point.

[0032] For the set of monitoring well points on the plane ,point Thiessen polygon Defined as: in, Point to monitoring well The Euclidean distance.

[0033] The above definition can be intuitively understood as follows: a sphere of influence is defined for each monitoring well. Within this sphere of influence, any location is closer to that monitoring well than to other monitoring wells. Thus, the measured water level change value of each monitoring well represents the water level change of the entire polygonal area, and then a weighted average is calculated according to the proportion of each polygon's area in the total grid area.

[0034] water yield parameters Characterizing the first The change in groundwater storage corresponding to a unit water level change within a Tyson polygon area can be determined comprehensively based on hydrogeological data, aquifer lithology, and historical field test results within the study area.

[0035] For grids in the study area where no monitoring wells are distributed, Kriging interpolation is used as an example to fill in the gaps. Based on the spatial autocorrelation structure of variables, Kriging interpolation can provide the optimal linear unbiased estimate of the shallow groundwater storage changes in unobserved grids.

[0036] In this embodiment, on the one hand, by defining a clear spatial control range for each monitoring well, the actual situation of uneven spatial distribution of shallow groundwater monitoring wells in the study area can be fully respected, avoiding the problem of weight dilution in high-density observation areas and weight amplification in sparse observation areas caused by simple spatial averaging; on the other hand, by introducing a water yield parameter to physically transform water level changes into storage changes, the water level observation results of the monitoring wells are organically linked with the gridded shallow groundwater storage indicators, constructing a label system with clear physical connotations for the subsequent training of deep learning models, further improving the physical interpretability and practical application value of the model learning results.

[0037] In some embodiments, the formula (2) for calculating the outlier in the baseline period is: in, The raw data for the target time period; The average value of the corresponding data within the preset benchmark period is used; the benchmark unification of the multi-source basic data and the total change data of land water storage is achieved through the calculation of outliers in the benchmark period.

[0038] In some embodiments, the standardization process employs the Z-score standardization method, and its calculation formula (3) is as follows: in, The original data values, This represents the mean of the corresponding data sequence. The standard deviation of the corresponding data sequence is used to eliminate dimensional differences between different features through the standardization process.

[0039] Specifically, the preset benchmark period can be a time interval that covers multiple complete hydrological years and can represent the long-term average hydrological state of the study area. For example, January 2004 to December 2009 can be selected as the unified benchmark period.

[0040] For hydrological variable data output by land surface hydrological models, for example, before calculating the outliers in the baseline period, the physical units of soil water data from each layer are first uniformly converted to equivalent water height units, such as centimeters, and then the data are summed to obtain the total soil water change. Then to Changes in snow water equivalent and changes in canopy water The outliers relative to the baseline period are calculated using the formula (2) above. Similarly, the outliers relative to the baseline period are calculated using the same formula (2) for the meteorological and remote sensing driving factor data. After the calculation of the baseline period outliers, the multi-source basic data are all expressed as deviations from the average hydrological state of the same period, forming a unified physical meaning with the total change data of land water storage retrieved from gravity satellites, thus avoiding systematic biases caused by different data sources using different baseline periods.

[0041] The Z-score standardization process further maps feature data of different physical quantities and dimensions to a dimensionless space with a mean of zero and a standard deviation of one. For example, it standardizes feature data such as precipitation, air temperature, potential evapotranspiration, surface temperature and normalized vegetation index after outlier calculation, so that the gradient update direction will not be dominated by the excessively large magnitude of individual features during the training process of the deep learning model.

[0042] In this embodiment, on the one hand, by using a unified baseline period outlier calculation, the multi-source basic data from different data sources are made to have a consistent expression in terms of physical mechanism with the total change data of land water storage retrieved by gravity satellites, laying a physical foundation for subsequent component separation based on the water balance equation; on the other hand, the Z-score standardization process eliminates the dimensional differences between different features, enabling the deep learning model to learn the contribution of each feature to the change of shallow groundwater storage more evenly, significantly improving the training stability and prediction accuracy of the model.

[0043] In some embodiments, the total change data of land water storage retrieved by gravity satellite is monthly-scale total change data of land water storage; the hydrological variable data output by the land surface hydrological model includes soil water change, snow water equivalent change and canopy water change; the meteorological and remote sensing driving factor data includes at least one of precipitation, air temperature, potential evapotranspiration, surface temperature and normalized vegetation index; and the measured water level data is monthly-scale measured water level data.

[0044] The spatial resampling unifies the spatial resolution of the meteorological and remote sensing driving factor data and the hydrological variable data output by the land surface hydrological model to a preset grid resolution, which is consistent with the spatial resolution of the total land water storage change data retrieved by the gravity satellite; the time alignment unifies the multi-source basic data into a monthly time series and removes missing values.

[0045] Specifically, the selection of the monthly time series should ensure coverage of a sufficient number of hydrological years to fully reflect the interannual variation characteristics of groundwater storage in the study area. The preset grid resolution is, for example, 0.25°, consistent with the spatial resolution of GRACE or GRACE-FO gravity satellite products.

[0046] The spatial resampling, for example, employs methods such as bilinear interpolation or nearest neighbor interpolation to resample meteorological and remote sensing driving factor data with inconsistent original spatial resolutions, as well as hydrological variable data output by land surface hydrological models, to a preset grid resolution, thereby achieving complete alignment of all data on the spatial reference.

[0047] The time alignment operation exemplarily performs time dimension matching on all data, unifying them into a monthly time series, and removes data from individual missing months or supplements them using reasonable interpolation methods of adjacent months to ensure that the time span of all data is completely consistent.

[0048] This embodiment integrates multi-source basic data from different data sources and observation methods under the same spatiotemporal reference by strictly aligning the spatial resolution and time scale. This avoids model training bias and component separation error caused by inconsistencies in spatiotemporal references between data sources, and provides a rigorous data foundation for subsequent physical constraints based on the water balance equation.

[0049] In some embodiments, the calculation formula (4) for the change in deep groundwater storage is: in, This refers to the total change data of the terrestrial water storage; For the aforementioned soil water changes, The change in the equivalent of snow water For the changes in canopy water; This refers to the optimized shallow groundwater storage prediction results.

[0050] The changes in deep groundwater storage are uncertain, and the sources of uncertainty include: observation errors in the total changes in land water storage data retrieved by gravity satellite, simulation errors in the hydrological variable data output by the land surface hydrological model, and residuals in the prediction results of shallow groundwater storage; the range of uncertainty is limited according to preset hydrological inversion reference data.

[0051] Specifically, based on the land water storage balance equation, non-groundwater hydrological components such as soil water, snow water equivalent, and canopy water are sequentially deducted from the total change in land water storage retrieved by gravity satellites to obtain the change in total regional groundwater storage. Then, by further deducting the optimized shallow groundwater storage prediction results, the change in deep groundwater storage can be separated, thus completing the component separation of shallow and deep groundwater.

[0052] The uncertainty in the results of deep groundwater storage changes mainly stems from three aspects: First, the total change data of land water storage retrieved by gravity satellites is affected by factors such as satellite orbit errors, atmospheric correction errors, striping errors, and signal leakage errors, resulting in inherent observation errors. Second, the hydrological variable data output by the land surface hydrological model is affected by factors such as model parameterization schemes, driving data accuracy, and grid resolution, resulting in certain simulation errors. Third, the predicted results of shallow groundwater storage themselves have certain model residuals compared to actual observations. For example, classic research results published in the field of hydrological retrieval can be referenced to comprehensively estimate the magnitude of the above three types of errors, thereby reasonably limiting the uncertainty range of the deep groundwater storage change results.

[0053] In this embodiment, on the one hand, the water balance equation of terrestrial water storage is used as a strong physical constraint, so that the separation results of deep groundwater storage changes have a clear physical mechanism explanation, avoiding the problem of fuzzy physical mechanism in pure data-driven methods; on the other hand, by systematically sorting out the sources of uncertainty and reasonably limiting the uncertainty range, the reliability and practical application value of the separation results are further enhanced, providing a reliable quantitative basis for the sustainable management of regional groundwater resources.

[0054] In some embodiments, the number of samples in the spatiotemporal feature sample set is calculated using the following formula (5): in, The number of samples, The total length of the multi-source basic data. The length of the sliding window.

[0055] In some embodiments, adjacent samples generated by the sliding window method overlap, and the spatiotemporal feature sample set is divided into a training set and a test set according to the temporal order. The temporal order of the training set is earlier than that of the test set, and a validation set is obtained by splitting the training set according to the temporal division method.

[0056] Specifically, the length L of the sliding window is 12 months for example, and the sliding step size is 1 month for example, so that each sliding can cover the complete annual hydrological cycle, while making the most of the limited time series data to build training samples.

[0057] For example, for a time series spanning 84 months from January 2018 to December 2024, a total of 73 training samples can be generated by using a 12-month sliding window with a step size of 1.

[0058] For each sample, the historical shallow groundwater storage data for the 12 consecutive months preceding the corresponding time period is used as an independent lag feature channel. This channel is then concatenated with five meteorological and remote sensing driving factor channels, including precipitation, air temperature, potential evapotranspiration, surface temperature, and normalized vegetation index, to form a three-dimensional spatiotemporal input tensor with a time step of 12, 6 channels, and a height and width equal to the number of spatial grids. The model training objective is the shallow groundwater storage label for the corresponding time period.

[0059] Because there is an 11-month overlap between adjacent samples, to avoid data leakage from the validation and test sets to the training set, the sample order is not randomly shuffled. Instead, a strict temporal partitioning method is used to divide the sample set: samples at the beginning of the time sequence are used as the training set, samples at the end of the time sequence are used as the test set, and a portion of the training set is partitioned according to the same temporal partitioning method as the validation set. For example, the first 50 samples in the time sequence can be used as the training set, the last 11 samples as the test set, and a portion of the training set can be partitioned according to time sequence as the validation set, thereby ensuring the objectivity and authenticity of the model evaluation results.

[0060] It should be noted that the groundwater system is essentially a hydrological integral system. The change in shallow groundwater storage in the current period is not determined by the current meteorological driving factors, but is the result of the combined effect of the historical groundwater state and the current recharge and consumption process. That is, the change in shallow groundwater storage in the current period is equal to the superposition of the historical shallow groundwater storage state and the current precipitation infiltration recharge, minus the consumption caused by current evapotranspiration and human extraction.

[0061] Taking the Beijing-Tianjin-Hebei region as an example, precipitation recharges shallow groundwater through soil infiltration, which usually has a lag response period of three to twelve months or even longer. If the model only inputs meteorological factors within the current period or short window period without introducing historical groundwater storage information, it is difficult to obtain the initial hydrological conditions necessary for prediction. The model can only blindly fit the absolute water level changes and is difficult to conform to the actual physical laws of groundwater movement.

[0062] Therefore, based on the construction of the sliding window sample, this implementation plan further introduces a hydrological lag feature injection mechanism: the historical shallow groundwater storage data for the 12 consecutive months preceding the corresponding period of each prediction sample is used as an independent lag feature channel for model input. The 12-month lag window is consistent with the length of the sliding window. On the one hand, it covers the main lag response cycle of precipitation infiltration into shallow groundwater in the Beijing-Tianjin-Hebei region. On the other hand, it forms a clear physical division of labor with meteorological driving factors—meteorological factors are used to characterize the hydrological driving factors in the current period and within the window period, while the lag shallow groundwater storage channel is used to characterize the integral state and historical storage and discharge effects of the groundwater system. This allows the model to make corrected predictions based on historical water levels and current meteorological conditions, rather than blindly fitting absolute values.

[0063] In this embodiment, on the one hand, the sliding window method effectively increases the number of training samples in the context of small time series, overcoming the lack of sample size in the original monthly scale data; on the other hand, the hydrological integral characteristics and historical storage and discharge status of the groundwater system are explicitly introduced through the hysteresis feature injection mechanism, so that the input feature system of the model is highly consistent with the physical mechanism of groundwater movement, significantly improving the model's fitting ability and generalization ability to the process of shallow groundwater storage change.

[0064] In some embodiments, the weight calculation formula (6) of the spatial attention mechanism is: in, The spatiotemporal features output by the three-dimensional convolutional layer; This represents the Sigmoid activation function; This represents a convolution operation with a filter size of 7×7; and These respectively represent the spatiotemporal features. The average pooling and max pooling operations performed; A spatial attention weight map generated for the spatial attention mechanism.

[0065] In some embodiments, the three-dimensional convolutional layer of the three-dimensional convolutional neural network simultaneously extracts grid association features in the spatial dimension and intertemporal lag features in the temporal dimension to form the initial high-dimensional spatiotemporal features; the spatial attention mechanism dynamically assigns weights to the initial high-dimensional spatiotemporal features, strengthens the feature weights of the core hydrological change area, obtains optimized spatiotemporal features, and outputs the shallow groundwater storage prediction result based on the optimized spatiotemporal features.

[0066] Specifically, the three-dimensional convolutional layer of the three-dimensional convolutional neural network exemplarily adopts a three-dimensional convolutional kernel with three directions: time dimension, height dimension and width dimension, to perform sliding scan on the input three-dimensional spatiotemporal input tensor. In the same convolution operation, it simultaneously captures the spatial positional association information between adjacent grids and the intertemporal dependency information between different time steps, thereby forming an initial high-dimensional spatiotemporal feature with spatiotemporal coupling characteristics.

[0067] The spatial attention mechanism, exemplarily, introduces a convolutional block attention module after the 3D convolutional layer. The initial high-dimensional spatiotemporal features F are subjected to average pooling and max pooling operations along the channel dimension, respectively, to obtain two sets of spatial feature descriptors that retain spatial dimensional information. These two sets of spatial feature descriptors are then concatenated along the channel dimension, and a spatial attention weight map is calculated through a 7×7 convolution operation and a sigmoid activation function. The spatial attention weight map is multiplied element-wise with the initial high-dimensional spatiotemporal features to obtain the optimized spatiotemporal features after being weighted by the spatial attention mechanism.

[0068] It should be noted that the spatial attention mechanism in this application is not a simple transplant of image recognition components in the general field of computer vision, but a domain-adapted design for the spatial heterogeneity characteristics of regional groundwater storage.

[0069] Taking the Beijing-Tianjin-Hebei study area as an example, the spatial distribution of shallow groundwater storage changes in the study area is extremely uneven: the Yanshan Mountains and the Taihang Mountains piedmont area are affected by relatively low groundwater extraction intensity and bedrock conditions, resulting in small interannual fluctuations in shallow groundwater storage, which belong to the background stable area; while the concentrated agricultural irrigation area and urban water supply area in the central North China Plain have drastic interannual fluctuations in shallow groundwater storage, which are the core sensitive areas controlling the trend of changes in the total groundwater storage in the region.

[0070] Meanwhile, the spatial distribution of shallow groundwater monitoring wells is relatively sparse and uneven. Under the preset grid resolution, a few high-fluctuation grids contribute far more to the change in the total groundwater storage in the region than a large number of low-fluctuation background grids.

[0071] Traditional two-dimensional convolutional neural networks and time-series prediction models based on recurrent neural networks assign equal weights to all spatial grids, making it easy for the prediction results to converge to the regional average value, and difficult to fit the peak and trough variation curves that are close to the actual observations.

[0072] To address the aforementioned issues, this implementation plan utilizes a spatial attention mechanism to automatically identify core spatial regions causing significant changes in shallow groundwater reserves during model backpropagation training. It dynamically increases the feature weights of sensitive grids such as agricultural over-extraction areas and high-intensity mining areas, while simultaneously decreasing the weights of low-fluctuation background areas such as stable piedmont zones. Essentially, it focuses on the main hydrological control area in the spatial dimension, answering the hydrological question of "where is the main control area for changes in shallow groundwater reserves," rather than focusing on "which pixel block to focus on" in image recognition. The 3D convolution is responsible for extracting spatiotemporal coupling features, while the spatial attention mechanism is responsible for selecting the main control grids that contribute most to the prediction in the spatial dimension. Their synergistic effect enables the model to capture the dynamic changes in groundwater reserves in the study area during key transitional stages, avoiding mediocre linear prediction results.

[0073] This embodiment, through the organic combination of three-dimensional convolutional layers and spatial attention mechanisms, enables the model to simultaneously extract spatiotemporal coupling features and focus on core hydrological change areas. This overcomes the technical shortcomings of existing models based on two-dimensional convolutional neural networks or recurrent neural networks, which are unable to capture the change process of the spatially heterogeneous main control area, and significantly improves the prediction accuracy of shallow groundwater storage.

[0074] In some embodiments, the Adam optimizer is used to train the three-dimensional convolutional neural network; an early stopping strategy is introduced during the training process, and the training is automatically terminated and the optimal model weights are saved when the validation set loss does not decrease for a preset number of consecutive rounds; mean square error and coefficient of determination are used as the accuracy indicators, and the learning rate and training rounds are dynamically adjusted according to the accuracy indicators.

[0075] Specifically, when training the three-dimensional convolutional neural network using the Adam optimizer, the initial learning rate is set to an example of [missing information]. The batch size is 16, and the maximum number of training epochs is set to 100. An early stopping strategy is also introduced, with the early stopping patience value set to 30 epochs for example. When the validation set loss does not decrease for 30 consecutive epochs, training is automatically terminated and the model weights corresponding to the optimal validation set loss are saved to prevent model overfitting.

[0076] During training, the mean squared error and coefficient of determination on the validation set are used as the main accuracy evaluation indicators, and the mean absolute error can be used as a reference. Hyperparameters such as learning rate and training rounds are dynamically adjusted according to changes in the indicators.

[0077] For example, in a typical training process, the model can achieve a mean absolute error of about 0.26, a mean square error of about 0.12, and a coefficient of determination of about 0.88 on a completely unseen test set. The prediction curve is highly consistent with the shallow groundwater storage calibration label constructed by the Thiessen polygon method in terms of overall trend and key peak and trough locations.

[0078] This embodiment achieves stable training and effective convergence of the model with a limited sample size by combining the Adam optimizer with an early stopping strategy, avoiding overfitting caused by too many training rounds. At the same time, the hyperparameter dynamic adjustment strategy based on accuracy index can continuously optimize the model's generalization performance according to the actual data characteristics of the study area, providing high-precision shallow prediction results for subsequent deep groundwater storage separation based on the water balance equation.

[0079] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0080] In one exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 2As shown, the computer device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When the computer program is executed by the processor, it implements a method for separating components of groundwater reserves. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0081] Those skilled in the art will understand that Figure 2 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0082] In one embodiment, a computer-readable storage medium is provided, on which a program or instructions are stored, which, when executed by a processor, implement the steps in the above-described method embodiments.

[0083] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0084] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0085] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0086] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for separating components of groundwater reserves, characterized in that, The method for separating components of groundwater reserves includes: Acquire multi-source basic data for the target area; wherein, the multi-source basic data includes: total change data of land water storage retrieved by gravity satellite, hydrological variable data output by land surface hydrological model, meteorological and remote sensing driving factor data, and measured water level data of shallow groundwater monitoring wells; Based on the spatial distribution of the monitoring wells, shallow groundwater storage training labels are calculated using a spatial partitioning method with regional weights combined with the water yield parameters of the monitoring wells. The multi-source basic data is subjected to spatiotemporal consistency preprocessing to obtain processed data that matches the spatiotemporal benchmark of the total change data of land water storage; wherein, the preprocessing includes spatial resampling, temporal alignment, outlier calculation of the benchmark period, and standardization processing; Based on the processed data, a spatiotemporal feature sample set is constructed using the sliding window method; each sample uses historical shallow groundwater storage data as a lag feature channel, which is then concatenated with multiple meteorological driving factor channels to form a three-dimensional spatiotemporal input tensor. The three-dimensional spatiotemporal input tensor is input into a three-dimensional convolutional neural network with a spatial attention mechanism. Spatiotemporal features are extracted through three-dimensional convolutional layers. The spatial attention mechanism is used to dynamically assign weights to the extracted spatiotemporal features and output the shallow groundwater storage prediction results. Based on the accuracy index of the shallow groundwater storage prediction results, the hyperparameters of the three-dimensional convolutional neural network are optimized to obtain the optimized shallow groundwater storage prediction results. Using the land water storage balance equation as a physical constraint, the changes in deep groundwater storage are obtained by deducting non-groundwater hydrological components and the optimized shallow groundwater storage prediction results from the total land water storage change data, so as to complete the component separation of shallow groundwater and deep groundwater.

2. The method for separating groundwater storage components according to claim 1, characterized in that, The weighted spatial partitioning method is the Thiessen polygon method; the shallow groundwater storage training labels are calculated based on the spatial distribution of the monitoring wells, using the weighted spatial partitioning method combined with the water yield parameters of the monitoring wells, including: The area of ​​the Thiessen polygon corresponding to each monitoring well is determined based on the Thiessen polygon method. Based on the area and water yield parameters of the monitoring well, the measured water level change values ​​of the monitoring well are converted into gridded shallow groundwater storage training labels. The calculation formula for the gridded shallow groundwater storage training labels is as follows: in, For the first Measured water level changes of each monitoring well; For the first The area of ​​a Thiessen polygon. For the first The water supply parameters of the region corresponding to each Thiessen polygon; This represents the total area of ​​the grid.

3. The method for separating groundwater storage components according to claim 1, characterized in that, The formula for calculating the outliers in the baseline period is as follows: in, The raw data for the target time period; This is the multi-year average of the corresponding data within the preset benchmark period.

4. The method for separating groundwater storage components according to claim 1, characterized in that, The total land water storage change data retrieved by the gravity satellite is monthly-scale total land water storage change data; the hydrological variable data output by the land surface hydrological model includes soil water change, snow water equivalent change and canopy water change; the meteorological and remote sensing driving factor data includes at least one of precipitation, air temperature, potential evapotranspiration, surface temperature and normalized vegetation index; the measured water level data is monthly-scale measured water level data.

5. The method for separating groundwater storage components according to claim 4, characterized in that, The formula for calculating the change in deep groundwater storage is as follows: in, This refers to the total change data of the terrestrial water storage; For the aforementioned soil water changes, The change in the equivalent of snow water For the changes in canopy water; This refers to the optimized shallow groundwater storage prediction results.

6. The method for separating groundwater storage components according to claim 5, characterized in that, The changes in deep groundwater storage are uncertain, and the sources of uncertainty include: observation errors in the total changes in land water storage data retrieved by gravity satellite, simulation errors in the hydrological variable data output by the land surface hydrological model, and residuals in the prediction results of shallow groundwater storage; the range of uncertainty is limited according to preset hydrological inversion reference data.

7. The method for separating groundwater storage components according to claim 1, characterized in that, The number of samples in the spatiotemporal feature sample set is calculated using the following formula: in, The number of samples, The total length of the multi-source basic data. The length of the sliding window.

8. The method for separating groundwater storage components according to claim 1, characterized in that, The weight calculation formula for the spatial attention mechanism is as follows: in, The spatiotemporal features output by the three-dimensional convolutional layer; This represents the Sigmoid activation function; This represents a convolution operation with a filter size of 7×7; and These respectively represent the spatiotemporal features. The average pooling and max pooling operations performed; A spatial attention weight map generated for the spatial attention mechanism.

9. A computer device, characterized in that, It includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the groundwater storage component separation method as described in any one of claims 1-8.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the groundwater storage component separation method as described in any one of claims 1-8.

Images