An underground water level height prediction method and system based on a multi-feature map network

Abstract

The present invention discloses a method and system for predicting the groundwater level height based on a multi-feature map network. The method includes: selecting groundwater level monitoring points at multiple different locations as spatial nodes for constructing a graph structure data set, obtaining the feature information of each node, and constructing the feature vectors of the nodes; weighing the correlation relationship between two nodes based on the feature vectors, constructing edge structure information for describing the connectivity of the nodes, and using the reciprocal of the geographical distance between two connected nodes as the edge connection weight of the nodes; constructing a graph structure data set based on the above information; improving the GCN to obtain a groundwater level prediction model; training the groundwater level prediction model based on the graph structure data set; and realizing the prediction of the groundwater level at the spatial position based on the trained groundwater level prediction model. The present invention can realize the efficient and accurate prediction of the groundwater level at unknown locations.

Description

Technical Field

[0001] The present invention relates to the technical field of hydrogeographic intelligent decision-making, and particularly relates to a method and system for predicting the groundwater level height based on a multi-feature map network. Background Art

[0002] Surface water and groundwater resources are important components of a complex ecological water cycle system. The main water sources within the water resources system are atmospheric water, surface water, groundwater, and water transferred from outside the system. Various water sources are mutually transformed under certain conditions. For example, rainfall infiltration and irrigation can supplement soil water, and when the soil water is saturated, it continues to infiltrate to form groundwater. Among them, in irrigation-based agricultural regions, long-term diversion and irrigation through river channels and poor drainage due to flood irrigation have led to a high groundwater level. Coupled with the usually strong surface evaporation in irrigation-based agricultural areas, the degree of soil salinization has deepened, and the ecological environment has gradually deteriorated. As surface water resources become scarcer, the river runoff has greatly decreased, and the irrigation water diversion volume has also decreased year by year, which can no longer meet the current agricultural irrigation needs and has severely restricted agricultural development.

[0003] The transformation relationship between surface water and groundwater is very complex. The change of the groundwater level is affected not only by the surface irrigation volume but also related to the topographic and geomorphic conditions. Studying the recharge and discharge relationship between surface water and groundwater in different spatial geographical positions can effectively alleviate the soil salinization problem caused by the unreasonable utilization of water resources and provide accurate guidance for agricultural production. The existing research mainly uses hydrogeological professional knowledge to study the evolution relationship between surface water and groundwater, and establishes traditional hydrogeological models and mathematical models to simulate the water resource scheduling process within the system. Although this method has a strong theoretical basis, for a complex water resource system, it cannot analyze the precise evolution of water resources from a data perspective.

[0004] Therefore, an intelligent solution that comprehensively considers data characteristics and principle characteristics is needed. Summary of the Invention

[0005] The present invention provides a method and system for predicting the groundwater level height based on a multi-feature map network to solve the technical problems of the lack of quantification and accuracy in the existing dynamic evolution process of surface water and groundwater.

[0006] To solve the above technical problems, the present invention provides the following technical solutions:

[0007] On the one hand, the present invention provides a method for predicting the groundwater level height based on a multi-feature map network, and the method for predicting the groundwater level height based on a multi-feature map network includes:

[0008] Select groundwater level monitoring points at multiple different locations as spatial nodes for constructing a graph-structured dataset, obtain the feature information of each node, and construct the feature vectors of each node based on the feature information;

[0009] Based on the feature vectors, weigh the correlation relationship between two nodes, construct edge structure information for describing node connectivity, and use the reciprocal of the geographical distance between two connected nodes as the edge connection weight of the nodes;

[0010] Use the feature information of the nodes and the edge structure information as sample features, and use the depth interval category to which the groundwater level burial depth corresponding to the nodes belongs as the sample label to construct a graph-structured dataset;

[0011] Improve the Graph Convolutional Network (GCN) to obtain a groundwater level prediction model; and train the groundwater level prediction model based on the graph-structured dataset;

[0012] Based on the trained groundwater level prediction model, realize the prediction of the groundwater level at the spatial location.

[0013] Furthermore, the feature information includes surface water resource information and geospatial information;

[0014] The surface water resource information includes the irrigation volume, drainage volume, and total precipitation in the preset period in the area where the node is located; the geospatial information includes: aeration zone lithology, landform type, permeability K1 zoning, land use type, total dissolved solids, permeability coefficient KCP, and the distance between the node and the branch canal.

[0015] Furthermore, constructing the feature vectors of each node based on the feature information includes:

[0016] Perform vectorization processing on the feature information to obtain the feature vectors of each node; among them, the method of performing vectorization processing on the feature information is one-hot encoding and normalization. For non-quantitative information, perform one-hot encoding first and then normalization processing. For quantitative information, directly perform normalization processing.

[0017] Furthermore, weighing the correlation relationship between two nodes based on the feature vectors and constructing edge structure information for describing node connectivity includes:

[0018] Calculate the cosine similarity, Pearson correlation coefficient, and Euclidean distance between the feature vectors of all nodes;

[0019] Construct a one-way connected relationship graph for all nodes, retain the connected edges that meet the preset conditions, and regard the nodes corresponding to the connected edges that do not meet the preset conditions as having no connection relationship; wherein, the preset conditions are that the cosine similarity between feature vectors is greater than 0.7, the Pearson correlation coefficient is greater than 0.8, and the Euclidean distance is less than 1.

[0020] Statistically filter the retained connected relationships to obtain edge structure information for describing node connectivity.

[0021] Further, the depth interval categories include 0 - 1m, 1 - 2m, 2 - 3m, 3 - 4m, 4 - 5m, and above 5m.

[0022] Further, the improvement of GCN includes:

[0023] Add a weight matrix to the first and last graph convolution operations in the GCN network. Specifically, the implementation is as follows: The spatial nodes introduce the edge connection weight as the aggregation coefficient during the stage of aggregating the features of neighbor nodes.

[0024] Further, the improvement of GCN also includes:

[0025] Add a graph attention convolution module after the first graph convolution layer of the GCN network. The input feature of the graph attention convolution module is the feature output after aggregating neighbor nodes by the first graph convolution operation. The output feature of the graph attention convolution module is the aggregated feature after automatically learning and updating the weights between nodes and their neighbor nodes. After obtaining the feature output of the graph attention convolution module, it is fused with the output feature of the first graph convolution layer, and the fused feature is input into the last graph convolution layer.

[0026] Further, after predicting the groundwater level at the spatial position, the method further includes:

[0027] Evaluate the effectiveness of the prediction result of the groundwater level prediction model.

[0028] Further, the evaluation of the effectiveness of the prediction result of the groundwater level prediction model includes:

[0029] Compare the categories of the prediction result and the true result, and evaluate the effectiveness of the model through the prediction accuracy.

[0030] On the other hand, the present invention also provides a groundwater level height prediction system based on a multi - feature graph network. The groundwater level height prediction system based on a multi - feature graph network includes:

[0031] A graph structure dataset construction module for:

[0032] Select groundwater level monitoring points at multiple different locations as spatial nodes for constructing a graph structure dataset, obtain the feature information of each node, and respectively construct the feature vectors of each node based on the feature information;

[0033] Based on the feature vectors, weigh the correlation relationship between two nodes, construct edge structure information for describing node connectivity, and use the reciprocal of the geographical distance between two connected nodes as the edge connection weight of the node;

[0034] Use the feature information of the node and the edge structure information as sample features, and use the depth interval category to which the groundwater level depth corresponding to the node belongs as the sample label to construct a graph structure dataset;

[0035] The groundwater level prediction model construction and training module is used to improve the Graph Convolutional Network (GCN) to obtain a groundwater level prediction model; and train the groundwater level prediction model based on the graph structure dataset constructed by the graph structure dataset construction module;

[0036] The groundwater level prediction module is used to realize the prediction of the groundwater level at the spatial position based on the groundwater level prediction model trained by the groundwater level prediction model construction and training module.

[0037] On the other hand, the present invention also provides an electronic device, which includes a processor and a memory; wherein, at least one instruction is stored in the memory, and the instruction is loaded and executed by the processor to implement the above method.

[0038] On another hand, the present invention also provides a computer-readable storage medium, in which at least one instruction is stored, and the instruction is loaded and executed by the processor to implement the above method.

[0039] The beneficial effects brought by the technical solution provided by the present invention at least include:

[0040] The groundwater level height prediction method provided by the present invention is a method for predicting the groundwater level height through surface water resources and geographical space attributes. A graph structure dataset is constructed with spatial position nodes as units, the distance weight between nodes is added, and the attention mechanism and feature fusion mechanism are introduced, so that the method can obtain better feature expressions of nodes. At the same time, the method can train the model in a semi-supervised learning manner when the groundwater levels of some nodes are unknown, so as to realize the accurate prediction of the groundwater level height at the spatial position. The present invention uses data-driven artificial intelligence technology to study the evolution relationship between surface water and groundwater, and solves the problems of inability to quantify and insufficient accuracy existing in the existing surface water and groundwater evolution methods under complex conditions. Brief Description of the Drawings

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the accompanying drawings required for the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0042] Figure 1 It is a schematic execution flowchart of the underground water level height prediction method based on the multi-feature map network provided by the embodiments of the present invention;

[0043] Figure 2 It is a schematic diagram of the spatial node feature descriptor provided by the embodiments of the present invention;

[0044] Figure 3 It is a schematic calculation flowchart of the attention factor provided by the embodiments of the present invention;

[0045] Figure 4 It is a schematic diagram of the improved GCN network structure provided by the embodiments of the present invention;

[0046] Figure 5 It is a schematic diagram of the confusion matrix of the prediction results of the spatial nodes in the test set provided by the embodiments of the present invention;

[0047] Figure 6 It is a schematic diagram of the structure of the underground water level height prediction system based on the multi-feature map network provided by the embodiments of the present invention;

[0048] Figure 7 It is a block diagram of an electronic device applying the method of the present invention. Specific Embodiments

[0049] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will further describe the embodiments of the present invention in detail with reference to the accompanying drawings.

[0050] First Embodiment

[0051] This embodiment provides a method for predicting the underground water level height based on a multi-feature map network. This method can be implemented by an electronic device, which can be a terminal or a server. The execution process of the method for predicting the underground water level height based on the multi-feature map network is as Figure 1 shown, including the following steps:

[0052] S1. Select underground water level monitoring points at multiple different locations as spatial nodes for constructing a graph structure dataset, obtain the feature information of each node, and respectively construct the feature vectors of each node based on the feature information;

[0053] S2. Weigh the correlation relationship between two nodes based on the eigenvector, construct the edge structure information for describing node connectivity, and use the reciprocal of the geographical distance between two connected nodes as the edge connection weight of the node.

[0054] S3. Use the node's feature information and edge structure information as sample features, and use the depth interval category to which the groundwater level depth corresponding to the node belongs as the sample label to construct a graph structure data set.

[0055] S4. Improve the Graph Convolutional Network (GCN) to obtain a groundwater level prediction model, and train the groundwater level prediction model based on the graph structure data set.

[0056] S5. Based on the trained groundwater level prediction model, realize the prediction of the groundwater level at the spatial location.

[0057] Among them, the feature information in S1 includes surface water resource information and geographical space information; the surface water resource information includes the irrigation volume, drainage volume, and total precipitation in the preset period of the area where the node is located; the geographical space information includes: vadose zone lithology, landform type, permeability K1 zoning, land use type name, total dissolved solids, permeability coefficient KCP, and the distance between the node and the branch canal. Based on this, the implementation process of S1 is as follows:

[0058] Obtain the surface irrigation and drainage, meteorological data, and geographical attribute information at multiple locations in the space, and construct an eigenvector for describing the surface information of the nodes at the spatial locations based on the obtained information; among them, the surface irrigation and drainage data is flow statistics, the meteorological data is height statistics, and the geographical attribute data is text information; generate the features of the spatial nodes by establishing the mapping between the groundwater monitoring points and the surface spatial location nodes, including: performing surface water volume standardization calculation on the irrigation and drainage data and meteorological data and allocating them to the grid cells where the nodes are located, and performing feature encoding on the geographical attribute information and fusing to obtain the node eigenvector. Its specific implementation process includes the following steps:

[0059] Divide the irrigation area into several sub-watersheds according to the irrigation canal system. Each irrigation sub-watershed has a clear water diversion source and water diversion data statistics. The drainage source considers the entire agricultural irrigation area as a whole and allocates the drainage data according to the corresponding sub-watershed irrigation area weight.

[0060] Divide the sub-watersheds into 100m * 100m grid cells, perform spatial superposition of the divided grid cells and the above-mentioned irrigation and drainage sub-watersheds, establish the topological relationship between the grid cells and the irrigation and drainage units, clarify the water diversion area and drainage area where each grid cell is located, and the calculation method of the water resource characteristics of the grid cells is as follows:

[0061]

[0062] Daily irrigation and drainage volume = average daily irrigation and drainage flow × 24 × 60 × 60

[0063] Daily precipitation and evaporation amount = daily precipitation and evaporation height × sub - basin area × 666.7 × 10 -4

[0064] Establish a mapping relationship between groundwater monitoring points and surface grid units to form spatial nodes. According to the coordinate positions of the spatial nodes, their geospatial attributes are statistically analyzed, including: lithology of the vadose zone, geomorphic type, permeability K1 zoning, land use type name, total dissolved solids, permeability coefficient KCP, distance between the node and the branch canal. Among them, the different permeability values of the lithology of the vadose zone include sand, sandy clay, interlayer water distribution area, upper clay and lower sandy soil, etc.; the types of geomorphic types include aeolian dunes, river - lake alluvial plains, alluvial - proluvial plains, modern proluvial fans, piedmont proluvial inclined plains, hilly terraces, etc.; the value range of permeability K1 zoning has 1 - 3, 3 - 5, 5 - 10, 10 - 20 corresponding to different permeabilities; the land use type names include corn, wheat, grape, wolfberry, rice, village, large - scale vegetable, city, abandoned land, etc. The geospatial attributes are collected through the spatial join function of ArcGis. For polygon - type layers, the polygons in which the nodes fall are selected, and for point - type and line - type layers, the points or lines closest to the nodes are selected. After collection, the geospatial attributes are fused with the water resource characteristics to construct a node feature descriptor. The spatial node feature descriptor is as Figure 2 shown.

[0065] After obtaining the feature descriptors (feature information) of each spatial node, the spatial node descriptors are vectorized. The specific vectorization method is: one - hot encoding and normalization. For non - quantitative columns, they are encoded first and then normalized, and for quantitative columns, they are only normalized, so as to obtain the feature vectors of each spatial node.

[0066] Furthermore, in this embodiment, the feature matrix of the spatial nodes is represented in the form of a two - dimensional matrix, where the row index represents the number and index of the spatial nodes, and the column index represents the attribute features of the nodes, including surface irrigation and drainage, meteorological data, and multi - dimensional topographic and geomorphic attribute data.

[0067] Furthermore, in this embodiment, in the above S2, the edge structure information is constructed by calculating specific indicators of the feature vectors, including three indicators: cosine similarity, Pearson correlation coefficient, and Euclidean distance. Cosine similarity is used to evaluate the similarity of two vectors by calculating the cosine value of the included angle between them, and to judge whether the two vectors generally point in the same direction. The closer the cosine value is to 1, the higher the similarity. The calculation formula is as follows:

[0068]

[0069] The Pearson correlation coefficient is used to measure whether two vectors are on the same line, to measure the linear correlation between interval variables, and is obtained by calculating the covariance and standard deviation between two samples. The closer the absolute value of the correlation coefficient is to 1, the stronger the linear correlation. The calculation formula is as follows:

[0070]

[0071] The Euclidean distance is a commonly used distance definition, which refers to the actual distance between two points in an m-dimensional space, used to calculate the distance between two vectors, and the value range is from 0 to positive infinity. Obviously, if the distance between two vector spaces is small, then the vectors are definitely more similar. The calculation formula is as follows:

[0072]

[0073] Specifically, the construction process of the above-mentioned spatial node connectivity edge structure is as follows:

[0074]

[0075] After obtaining the connectivity edge structure information, it is mapped to the geographical space according to the index, and the geographical coordinates of the spatial nodes are obtained. The geographical distances between the connected nodes are calculated respectively through the geographical coordinates, and the reciprocal of the distance is taken as the edge connection weight of the spatial nodes in the network model, and output to an n×n sparse matrix. The nodes without edge connections correspond to a distance weight of 0. In the graph convolution process, the neighbor nodes with closer distances can aggregate more information, and the neighbor nodes with farther distances contribute less feature information.

[0076] Furthermore, in this embodiment, each correlation evaluation index and the connected edge weight are represented by a two-dimensional matrix. The row and column values represent the spatial node indexes, and the matrix elements are the calculation results of the correlation evaluation index between nodes and the reciprocal of the distance; the connectivity edge structure information is represented by an n*2 matrix, and each row of the matrix is the indexes of two spatial nodes, indicating the connectivity between the two nodes.

[0077] Furthermore, in this embodiment, the depth interval categories in the above S3 include 0-1m, 1-2m, 2-3m, 3-4m, 4-5m, and above 5m, corresponding to different interval categories; the graph structure data set is divided into a training set, a validation set, and a test set; among them, the training set is used to train the groundwater level height prediction model, the validation set is used to verify the model accuracy during the training process, and the test set is used to evaluate the trained model. Another semi-supervised learning method trains the model through the nodes with known class labels to predict the spatial nodes with unknown class labels. Specifically, the implementation process of the above S3 is as follows:

[0078] Select 200 groundwater level monitoring points in the agricultural irrigation area as the spatial nodes for constructing the graph dataset of this method. According to the mapping relationship between the groundwater level monitoring points and the surface space grid, organize the feature descriptors of the spatial nodes and then construct the feature vectors. The attributes used to describe the spatial node features include surface water resource attributes and geospatial attributes. The surface water resource attributes include the total monthly irrigation, drainage, and precipitation of the spatial grid. The geospatial attributes include the vadose zone lithology, landform type, permeability K1 zoning, land use type name, total dissolved solids, permeability coefficient KCP, and the distance between the node and the branch canal. The groundwater level depth of the groundwater monitoring point is used as the class label of the spatial node, and each class contains a certain range of groundwater level depths. For example, the depths of 0-1m, 1-2m, 2-3m, 3-4m, 4-5m, and above 5m respectively correspond to different classes.

[0079] In this embodiment, the attribute data of the graph dataset is represented by a two-dimensional matrix with a dimension of 200×12. The first column is the index of the spatial node, and the last column is the class label corresponding to the groundwater level depth of the spatial node. The middle 10 columns are the feature vectors of each spatial node. The edge connection data of the graph dataset is represented as a two-dimensional matrix with a dimension of 2315×2, indicating that 200 spatial nodes have 2315 connected edges, and each row represents the indices of two mutually connected nodes.

[0080] Furthermore, in this embodiment, the improvement of the GCN in S4 above is as follows: Improve the GCN network by adding distance weights and an attention mechanism to enable it to better utilize the features of surrounding neighbors to learn the features of the node itself; including: adding a weight matrix in the first and last graph convolution operations, specifically implemented as introducing a distance factor as the aggregation coefficient in the stage of aggregating the features of neighbor nodes by the spatial node. Add a graph attention convolution module after the first graph convolution layer. The input feature of this module is the feature output after aggregating the neighbor nodes by the first graph convolution operation, and the output feature of this module is the aggregated feature after automatically learning and updating the weights between the node and its neighbor nodes. After obtaining the feature output of the graph attention convolution module, fuse it with the output feature of the first graph convolution layer, and input the fused feature into the last graph convolution layer.

[0081] Specifically, in this embodiment, the implementation process of improving the GCN is as follows:

[0082] First of all, it should be noted that the GCN was proposed to solve the problem that non-regularized data cannot share convolutional kernels during convolution. Traditional convolution uses local perception regions and shared weights, and can well extract the spatial features of images. The graph structure does not have the translational invariance of pictures, so the traditional convolution method is not applicable to the graph structure. GCN is a first-order local approximation of spectral graph convolution. It aggregates neighborhood information through convolutional layers. By stacking several convolutional layers, information transfer of multi-order neighborhoods can be achieved. In each layer, the adjacency matrix A and the feature matrix H are multiplied to obtain the summary of the features of each vertex's neighbors, and then multiplied by a trainable parameter matrix W. The adjacency matrix is normalized by the degree matrix to prevent large differences in the feature distributions between vertices with high degrees and vertices with low degrees.

[0083] Based on the above, the first improvement of this embodiment is to add edge connection weights during the graph convolution process. The weights are represented by the reciprocal of the geospatial distance. Its function is that each node not only considers the degree of neighbor nodes when aggregating neighbor node information, but also considers the distance between neighbor nodes and this node. It can be realized that neighbor nodes farther away have a smaller contribution to this node, and neighbor nodes closer have a greater contribution to this node. The calculation process of GCN after adding distance weights is as follows, where H (l+1) represents the feature representation of the central node after one convolution, obtained by summing the node adjacency matrix and the identity matrix I, represents the degree matrix obtained by summing by row according to W (l) represents the trainable parameter matrix, Dist is the introduced distance weight matrix. Since the diagonal elements of Dist are all 0, it is necessary to sum with the identity matrix I to ensure that the features of the central node itself are also considered in the process of aggregating neighbor node features.

[0084]

[0085] The second improvement of this embodiment to GCN is to add a graph attention module. The calculation method of the attention factor in the attention module is as Figure 3 shown. The network input is the graph structure dataset constructed above. Graph represents the constructed graph structure data, X represents the adjacency matrix corresponding to the graph nodes. Solid dots indicate that there are edge connection relationships for the corresponding nodes, and hollow dots indicate that there are no edge connection relationships between nodes; after the first graph convolution layer, the attention module is added, and then the output features of the attention module and the output features of the graph convolution layer are fused. This improvement can enable spatial nodes to more accurately utilize the information of surrounding neighbor nodes and improve the prediction accuracy. The overall structure of the improved GCN model network is as Figure 4 shown.

[0086] Specifically, in this embodiment, after the spatial node features and their edge connection information are input into the first graph convolutional layer, the graph node structure remains unchanged, and each node generates a feature representation with a length of 32. In this process, the dimension of the parameter matrix W is 10×32, corresponding to the input feature dimension and the output feature dimension respectively. The features output by the first graph convolutional layer are processed by the ReLU activation function, which can accelerate model training and overcome the problem of gradient disappearance. Then, dropout processing is performed to prevent model overfitting.

[0087] The node features after dropout processing of the first graph convolutional layer are input into the graph attention module. Since GCN cannot assign different weights according to the importance of neighbor nodes, although the reciprocal of the spatial distance is added as the weight of different nodes in the first convolutional layer, the weight information cannot be learned and updated during network training. However, the graph attention layer can autonomously learn and update the weight coefficients of nodes through training, and can better learn the dependence relationship between global features. The calculation method of the attention factor is as follows, where represents a feed-forward neural network, and its parameters can be updated through training. W represents the parameter matrix, represents the feature representation of the node, represents the total number of neighbor nodes of the central node. After multiplying and concatenating the features, they are non-linearized by LeakyReLU and normalized by Softmax to obtain the attention coefficients.

[0088]

[0089] After each node in the graph attention layer aggregates the features of surrounding neighbor nodes through the attention factor, it also generates a node feature representation with a dimension of 32. In this embodiment, the features output by the graph attention module are fused with the features output by the first graph convolutional layer, and the dimension of the fused features is n×64, where n represents the number of spatial nodes. This method can obtain an enhanced representation of the spatial node features. The fused enhanced features are then input into a graph convolutional layer, and the output result is normalized by the log_softmax exponential function to map the output to the range of 0-1. Then, the loss during model training is calculated by NLLLoss, and the network model parameters are corrected through the backpropagation algorithm.

[0090] When training the groundwater level height prediction model, it is necessary to first read the spatial node feature matrix and the edge connection matrix into memory, perform one-hot encoding on the node categories, construct the adjacency matrix and degree matrix of the edges, sum the adjacency matrix and its transposed matrix to convert the directed graph into an undirected graph, and perform normalization operations on the feature matrix and the degree matrix. When dividing the training set, validation set, and test set, randomly shuffle the spatial nodes and divide them in a ratio of 3:1:2. Input the training set into the model for training, and verify the effectiveness of the model through the validation set in each training epoch. The test set is used to test the prediction accuracy of the model after all training epochs are completed, and calculate the loss of the training set to update the model parameters through the backpropagation algorithm. The losses of the validation set and the test set are not used for model optimization. Specifically, the training process of the above groundwater level height prediction model is as follows:

[0091]

[0092]

[0093] Repeat the above steps to train the improved GCN network until the loss stabilizes and no longer decreases, the model converges, and the optimal training parameters are obtained.

[0094] Furthermore, in this embodiment, the implementation process of the above S5 is as follows:

[0095]

[0096] Furthermore, after implementing the prediction of the groundwater level at the spatial location, the method further includes:

[0097] Evaluate the effectiveness of the prediction results of the groundwater level prediction model. Specifically, in this embodiment, the effectiveness evaluation index of the groundwater level height prediction model is accuracy. Accuracy refers to the proportion of correctly predicted spatial node categories in the total number of nodes in the test set, and the calculation method of accuracy is as follows:

[0098]

[0099] In this embodiment, two datasets with different surface water characteristics and groundwater level embeddings in the wet season and dry season are respectively constructed, and the groundwater level height prediction model is evaluated according to the above evaluation index, and the evaluation results are shown in Table 1. In order to explore the performance of the model, different training batches and learning rate sizes are used to evaluate the effectiveness of the model.

[0100] Table 1 Evaluation Results of the Groundwater Level Height Prediction Model

[0101]

[0102]

[0103] In Table 1 above, Epoch is the number of iterations when training the model using the training set. Different numbers of training iterations have an impact on the prediction effectiveness of the model. Learning rate represents the influence amplitude on parameter changes during the error backpropagation in the training process. The magnitude of the learning rate also affects the prediction effect of the model. As can be seen from the above table, under the conditions of 1500 iterations and a learning rate of 0.005, the groundwater level height prediction model has good performance in both the wet season and the dry season. The confusion matrix of the prediction results of the spatial nodes is as Figure 5 shown.

[0104] In summary, this embodiment provides a method for predicting the groundwater level height through surface water resources and geospatial attributes. By constructing a graph structure dataset with spatial position nodes as units, adding distance weights between nodes, and introducing an attention mechanism and a feature fusion mechanism, the method can obtain better feature expressions of nodes. At the same time, the method of this embodiment can train the model through semi-supervised learning when the groundwater levels of some nodes are unknown, so as to achieve accurate prediction of the groundwater level height at spatial positions, and solve the problems of inability to quantify and insufficient accuracy existing in the existing surface water and groundwater evolution methods.

[0105] Second Embodiment

[0106] This embodiment provides a groundwater level height prediction system based on a multi-feature graph network. The structure of the groundwater level height prediction system based on the multi-feature graph network is as Figure 6 shown, and includes the following modules: [[ID=I7]]

[0107] Graph structure dataset construction module, used for:

[0108] Selecting groundwater level monitoring points at multiple different locations as spatial nodes for constructing the graph structure dataset, obtaining the feature information of each node, and respectively constructing feature vectors of each node based on the feature information;

[0109] Based on the feature vectors, weighing the correlation relationship between two nodes, constructing edge structure information for describing node connectivity, and using the reciprocal of the geographical distance between two connected nodes as the edge connection weight of the node;

[0110] Using the feature information and edge structure information of the nodes as sample features, and using the depth interval category to which the groundwater level burial depth corresponding to the node belongs as the sample label, constructing a graph structure dataset;

[0111] The groundwater level prediction model construction and training module is used to improve the Graph Convolutional Network (GCN) to obtain a groundwater level prediction model; and train the groundwater level prediction model based on the graph structure dataset constructed by the graph structure dataset construction module;

[0112] The groundwater level prediction module is used to realize the prediction of the groundwater level at the spatial position based on the groundwater level prediction model trained by the groundwater level prediction model construction and training module;

[0113] The model evaluation module is used to evaluate the effectiveness of the prediction results of the groundwater level prediction model.

[0114] The groundwater level height prediction system based on the multi-feature graph network in this embodiment corresponds to the groundwater level height prediction method based on the multi-feature graph network in the above first embodiment; among them, the functions realized by each functional module in the groundwater level height prediction system based on the multi-feature graph network correspond one by one to each process step in the groundwater level height prediction method based on the multi-feature graph network; therefore, it will not be elaborated here.

[0115] The third embodiment

[0116] This embodiment provides an electronic device, which includes a processor and a memory; wherein, at least one instruction is stored in the memory, and the instruction is loaded and executed by the processor to implement the method of the first embodiment.

[0117] This electronic device may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPUs) and one or more memories. Among them, at least one instruction is stored in the memory, and the instruction is loaded and executed by the processor to implement the above method.

[0118] Specifically, as Figure 7 shown, this electronic device may include a processor (CPU) 701, which can perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) 702 or the program loaded from the storage section 708 into the random access memory (RAM) 703. The processor may include a general microprocessor (such as a CPU), an instruction set processor, and / or a related chipset, and / or a dedicated microprocessor (such as an application-specific integrated circuit (ASIC)), etc. The processor may also include on-board memory for caching purposes.

[0119] In addition, the device may further include an input / output (I / O) interface 705, and the input / output (I / O) interface 705 is also connected to the bus 704. Moreover, it may further include one or more of the following components connected to the I / O interface 705: an input part 706 including a keyboard, a mouse, etc.; an output part 707 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc. and a speaker, etc.; a storage part 708 including a hard disk, etc.; and a communication part 709 including a network interface card such as a LAN card, a modem, etc. The communication part 709 performs communication processing via a network such as the Internet. The drive 710 is also connected to the I / O interface 705 as needed. A removable medium 711, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 710 as needed so that a computer program read from it can be installed into the storage part 708 as needed.

[0120] Fourth Embodiment

[0121] This embodiment provides a computer-readable storage medium storing at least one instruction, which is loaded and executed by a processor to implement the method of the first embodiment. Among them, the computer-readable storage medium may be a ROM, a random access memory, a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc. The instructions stored therein can be loaded and executed by the processor in the terminal to perform the above method.

[0122] In addition, it should be noted that the present invention can be provided as a method, a device, or a computer program product. Therefore, the embodiments of the present invention can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the embodiments of the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code.

[0123] The embodiments of the present invention are described with reference to the flowcharts and / or block diagrams of methods, terminal devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, and the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, an embedded processor, or other programmable data processing terminal devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal devices generate a device for implementing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0124] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing terminal device to work in a specific manner, such that the instructions stored in the computer-readable memory produce a manufacture including an instruction device that implements the functions specified in one process or a plurality of processes and / or one block or a plurality of blocks. These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device, such that a series of operation steps are executed on the computer or other programmable terminal device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable terminal device provide steps for implementing the functions specified in one process or a plurality of processes and / or one block or a plurality of blocks. Figure 1 one process or a plurality of processes and / or Figure 1 one block or a plurality of blocks Figure 1 one process or a plurality of processes and / or Figure 1 one block or a plurality of blocks

[0125] It should also be noted that in this document, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or terminal device. Without further limitation, an element defined by the statement "comprising one..." does not exclude the presence of additional identical elements in the process, method, article or terminal device comprising the said element.

[0126] Finally, it should be noted that the above is the preferred embodiment of the present invention. It should be pointed out that although the preferred embodiments of the present invention have been described, for those skilled in the art of this technology, once the basic creative concept of the present invention is known, several improvements and refinements can be made without departing from the principles described in the present invention, and these improvements and refinements should also be regarded as the protection scope of the present invention. Therefore, the appended claims are intended to be construed as including the preferred embodiments and all changes and modifications falling within the scope of the embodiments of the present invention.

Claims

1. A method for predicting the groundwater level height based on a multi-feature map network, characterized in that Including: Select groundwater level monitoring points at multiple different locations as spatial nodes for constructing a graph structure dataset, obtain the feature information of each node, and respectively construct the feature vectors of each node based on the feature information; Based on the feature vectors, weigh the correlation relationship between two nodes, construct edge structure information for describing node connectivity, and use the reciprocal of the geographical distance between two connected nodes as the edge connection weight of the node; Use the feature information of the nodes and the edge structure information as sample features, and use the depth interval category to which the groundwater level depth corresponding to the node belongs as the sample label to construct a graph structure dataset; Improve the Graph Convolutional Network (GCN) to obtain a groundwater level prediction model; and train the groundwater level prediction model based on the graph structure dataset; Based on the trained groundwater level prediction model, realize the prediction of the groundwater level at the spatial position; The feature information includes surface water resource information and geospatial information; The surface water resource information includes the irrigation volume, drainage volume, and total precipitation within a preset period in the area where the node is located; The geospatial information includes: vadose zone lithology, landform type, permeability K1 zoning, land use type, total dissolved solids, permeability coefficient KCP, and the distance between the node and the branch canal; the process of constructing the feature vector is as follows: divide the irrigation area into several sub-basins according to the irrigation canal system, divide the sub-basins into grid cells of 100m * 100m, and the calculation method of the water resource characteristics of the grid cells is as follows: Daily irrigation and drainage volume = average daily irrigation and drainage flow × 24 × 60 × 60 Daily precipitation evaporation = daily precipitation evaporation height × sub - basin area × 666.7 × 10 -4 Establish a mapping relationship between the groundwater monitoring points and the surface grid cells to form spatial nodes; count the geospatial information according to the spatial node coordinate positions; fuse the geospatial information and the water resource characteristics to construct a node feature descriptor; perform vectorization processing on the spatial node descriptor to obtain the feature vector of the spatial node; Construct edge structure information for describing node connectivity, including: Calculate the cosine similarity, Pearson correlation coefficient, and Euclidean distance between the feature vectors of all nodes; Construct a one-way connectivity relationship graph of all nodes, and retain the connected edges that simultaneously satisfy that the cosine similarity between the feature vectors is greater than 0.7, the Pearson correlation coefficient is greater than 0.8, and the Euclidean distance is less than 1; Count the retained connectivity relationships after screening to obtain edge structure information for describing node connectivity; Improve the GCN network, including: Add a weight matrix in the first and last graph convolution operations of the GCN network. Specifically, it is implemented as: the edge connection weight is introduced as an aggregation coefficient when the spatial node aggregates the feature of the neighbor node; Add a graph attention convolution module after the first graph convolution layer of the GCN network. The input feature of the graph attention convolution module is the feature output after aggregating the neighbor nodes by the first graph convolution operation, and the output feature is the aggregated feature after automatically learning and updating the weights between the node and its neighbor nodes; after obtaining the feature output of the graph attention convolution module, fuse it with the output feature of the first graph convolution layer, and input the fused feature into the last graph convolution layer.

2. The groundwater level height prediction method based on a multi-feature map network according to claim 1, wherein Respectively construct the feature vectors of each node based on the said feature information, including: Perform vectorization processing on the feature information to obtain the feature vectors of each node; among them, the method for performing vectorization processing on the feature information is one-hot encoding and normalization. For non-quantitative information, one-hot encoding is first performed, and then normalization processing is performed. For quantitative information, direct normalization processing is performed.

3. The groundwater level height prediction method based on a multi-feature map network according to claim 1, characterized in that, The depth interval categories include 0-1m, 1-2m, 2-3m, 3-4m, 4-5m, and above 5m.

4. The groundwater level height prediction method based on a multi-feature map network according to claim 1, wherein, After realizing the prediction of the groundwater level at the spatial position, the method further includes: Perform effectiveness evaluation on the prediction results of the groundwater level prediction model.

5. The groundwater level height prediction method based on a multi-feature map network according to claim 4, characterized in that The performing effectiveness evaluation on the prediction results of the groundwater level prediction model includes: Compare the categories of the prediction results and the true results, and evaluate the effectiveness of the model through the prediction accuracy.

6. An underground water level height prediction system based on a multi-feature map network, characterized in that, Include: Graph structure dataset construction module, used for: Select groundwater level monitoring points at multiple different positions as spatial nodes for constructing the graph structure dataset, obtain the feature information of each node, and respectively construct the feature vectors of each node based on the feature information; Weigh the correlation relationship between two nodes based on the feature vectors, construct edge structure information for describing node connectivity, and use the reciprocal of the geographical distance between two connected nodes as the edge connection weight of the node; Use the feature information and edge structure information of the node as sample features, and use the depth interval category to which the groundwater depth corresponding to the node belongs as the sample label to construct the graph structure dataset; Groundwater height prediction model construction and training module, used to improve the Graph Convolutional Network (GCN) to obtain the groundwater level prediction model; and train the groundwater level prediction model based on the graph structure dataset constructed by the graph structure dataset construction module; Groundwater height prediction module, used to realize the prediction of the groundwater level at the spatial position based on the groundwater level prediction model trained by the groundwater height prediction model construction and training module; The feature information includes surface water resource information and geospatial information; The surface water resource information includes the irrigation volume, drainage volume, and total precipitation within a preset period in the area where the node is located; The geospatial information includes: vadose zone lithology, landform type, permeability K1 zoning, land use type, total dissolved solids, permeability coefficient KCP, and the distance between the node and the branch canal; the process of constructing the feature vector is as follows: divide the irrigation area into several sub-watersheds according to the irrigation canal system, divide the sub-watersheds into grid cells of 100m * 100m, and the calculation method of the water resource characteristics of the grid cells is as follows: Daily irrigation and drainage volume = average daily irrigation and drainage flow × 24 × 60 × 60 Daily precipitation evaporation = daily precipitation evaporation height × sub - watershed area × 666.7 × 10 -4 Establish a mapping relationship between the groundwater monitoring point and the surface grid cell to form a spatial node; count its geospatial information according to the spatial node coordinate position; fuse the geospatial information and water resource characteristics to construct a node feature descriptor; perform vectorization processing on the spatial node descriptor to obtain the feature vector of the spatial node; Construct edge structure information for describing node connectivity, including: Calculate the cosine similarity, Pearson correlation coefficient, and Euclidean distance between the feature vectors of all nodes; Construct a unidirectional connected relationship graph for all nodes, and retain the connected edges that simultaneously satisfy the cosine similarity between eigenvectors greater than 0.7, the Pearson correlation coefficient greater than 0.8, and the Euclidean distance less than 1; Statistically filter the retained connected relationships to obtain edge structure information for describing node connectivity; Improve the GCN network, including: Add a weight matrix to the first and last graph convolution operations of the GCN network. Specifically, the implementation is as follows: When spatial nodes aggregate neighbor node features, the edge connection weight is introduced as the aggregation coefficient; Add a graph attention convolution module after the first graph convolution layer of the GCN network. The input feature of the graph attention convolution module is the feature output after aggregating neighbor nodes by the first graph convolution operation, and the output feature is the aggregated feature after automatically learning and updating the weights between nodes and their neighbor nodes. After obtaining the feature output of the graph attention convolution module, it is fused with the output feature of the first graph convolution layer, and the fused feature is input into the last graph convolution layer.

Images