A deep learning-based spatial identification method for groundwater phosphorus-enriched areas

Abstract

This invention discloses a deep learning-based method for spatially identifying groundwater phosphorus-enriched areas. The method includes the following steps: S1. Constructing a carbon-iron-phosphorus coupling knowledge graph; S2. Expanding the carbon-iron-phosphorus coupling knowledge graph; S3. Initializing the improved Warhawk optimizer population position using a high-dimensional feature set as input, and outputting a set of feature weight vectors; S4. Constructing a saliency-guided multi-scale graph neural network model structure; S5. Generating a voxel-level phosphorus enrichment risk field; and S6. Generating land use restriction zoning, governance priority ranking, and warning time curves. Using a significance scoring mechanism, the method automatically highlights the nodes and paths that contribute most to phosphorus enrichment risk, significantly improving model interpretability and risk mechanism transparency.

Description

Technical Field

[0001] The present invention relates to the field of spatial recognition technology, and in particular to a method for spatially identifying groundwater phosphorus-enriched areas based on deep learning. Background Art

[0002] With the continuous development of geoscience big data, environmental monitoring, and artificial intelligence technologies, risk assessment of sedimentary facies-controlled phosphorus-enriched areas has gradually become an important research direction in the fields of water environment governance and land space management. As a key limiting factor in water eutrophication, phosphorus migration, activation, and re-release at the sediment-water interface are controlled by the nonlinear coupling of sedimentary facies, geochemical state, and hydrological dynamics. Existing methods for phosphorus risk identification and spatial distribution analysis often rely on two-dimensional plane interpolation, empirical statistical regression, or simple mechanistic models. These methods generally have the following limitations:

[0003] On the one hand, existing risk zone identification methods are mostly based on limited sampling point data and single-scale variables, and can only qualitatively delineate high-risk areas on coarse-resolution two-dimensional grids or profiles. These methods fail to reflect the complex coupling relationships between sediment microfacies, multi-source geological and geochemical indicators such as FeOOH crystallinity, and multi-scale environmental factors of paleohydrological variation. This results in low spatial accuracy in high-phosphorus risk areas and insufficient mechanistic explanatory power. On the other hand, traditional factor screening and parameter weighting methods, which mostly rely on manual experience or linear correlation analysis, are unable to adapt to the dynamic changes in the main controlling factors of phosphorus activation and are prone to missing the interactive effects of microfacies combinations and hydrological scenarios that play a key role in phosphorus release.

[0004] In addition, the technologies currently used for risk visualization in phosphorus-rich areas are mostly based on post-monitoring or static two-dimensional chart displays, lacking the ability to perform high-resolution dynamic simulations of the evolution of phosphorus risks in three-dimensional space as hydrological conditions evolve. This makes it difficult to meet the forward-looking early warning needs of river basin management, land control, and especially groundwater management for future trends in the spread of phosphorus pollution.

[0005] Therefore, a technical approach is urgently needed to improve the risk identification, dynamic management and scientific governance capabilities of sedimentary phase-controlled phosphorus-enriched areas. Summary of the Invention

[0006] One purpose of the present invention is to propose a deep learning-based method for spatial identification of groundwater phosphorus-enriched areas. Under the significance scoring mechanism, the present invention can automatically highlight the nodes and paths that contribute most to the risk of phosphorus enrichment, significantly improving the model interpretability and risk mechanism transparency.

[0007] A method for spatially identifying groundwater phosphorus-enriched areas based on deep learning according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect multi-source geoscience datasets and input them into a graph database by mapping entities, attributes, and relationships to construct a carbon-iron-phosphorus coupling knowledge graph.

[0009] S2. Expand the carbon-iron-phosphorus coupling knowledge graph, extract node attribute vectors, edge attribute vectors, and contextual relationship vectors, and concatenate them into a high-dimensional feature set;

[0010] S3. Initialize the improved Warhawk optimizer population position using the high-dimensional feature set as input, perform a three-stage formation search process in the search space, and output a set of feature weight vectors;

[0011] S4. Write the feature weight vector set into the initial node attributes of the saliency-guided multi-scale graph neural network to construct the saliency-guided multi-scale graph neural network model structure;

[0012] S5. Train a saliency-guided multi-scale graph neural network model to obtain a node saliency score matrix and an embedding vector for the entire sedimentary sequence graph. Using this matrix and the embedding vector, calculate voxel-level phosphorus enrichment probability values ​​on a 90-meter resolution regular grid to generate a voxel-level phosphorus enrichment risk field.

[0013] S6. Load the voxel-level phosphorus enrichment risk field into the 3D geographic information system rendering engine to generate a 3D risk visualization model. Construct a hydrological scenario simulation dataset and couple it with the 3D risk visualization model to output a dynamic risk prediction sequence. After threshold comparison and trend judgment, generate land use restriction zoning, governance priority ranking, and warning time curve.

[0014] Optionally, the S1 includes the following steps:

[0015] S11. Collect borehole samples from the Quaternary lacustrine deposits in the target area to obtain a borehole sample collection;

[0016] S12. Acquire sedimentary profile remote sensing image data and obtain a remote sensing image grid set;

[0017] S13. Collect monitoring records of hydrological stations in the target area to obtain a hydrological station data set;

[0018] S14. Obtaining geochemical experimental analysis results corresponding to the drill hole samples to obtain a geochemical analysis data set;

[0019] S15. Align the borehole sample collection, remote sensing image grid collection, hydrological station data collection, and geochemical analysis data collection to a unified geographic coordinate reference and a unified time reference to form a cross-source aligned geoscience multi-source dataset;

[0020] S16. Input the multi-source geoscience dataset into the graph database system and construct a carbon-iron-phosphorus coupling knowledge graph based on entity-attribute-relationship structure mapping. :

[0021] ;

[0022] in, Represents a set of entity nodes including sediment organic matter type, organic matter transformation products, FeOOH crystal state, microfacies type, and hydrodynamic boundary. Represents the multi-source coupling mechanism edge set between entities, including biological pump action, reduction and dissolution pathway, and sedimentary phase-controlled transport relationship edge types. Represents a collection of entity node attributes, Represents a set of edge attributes;

[0023] Coupling carbon-iron-phosphorus knowledge graph Layered slicing is performed according to time slices and spatial blocks.

[0024] Optionally, the S2 includes the following steps:

[0025] S21. Knowledge graph of carbon-iron-phosphorus coupling Expand the graph structure and transform each entity node The corresponding entity node attributes are from the node attribute collection Extract and construct a node attribute vector set , where each node attribute vector Representation node corresponding sediment characteristics;

[0026] S22. For each edge From the edge attribute set Extract the semantic attributes of the edge and construct the edge attribute vector set , where the edge attribute vector Represents a connection node With node The causal mechanism relationship attributes;

[0027] S23. Construct a contextual relationship set based on the topological path and semantic path between nodes in the graph structure, and define the nodes The context neighborhood set , encode the context information into a context relationship vector Forming a contextual relationship vector set , where the contextual relationship vector Representation node High-order path information between its semantic neighbor nodes;

[0028] S24. Set the node attribute vector , edge attribute vector set Vector collection of contextual relationships Perform splicing and fusion to build a high-dimensional feature set .

[0029] Optionally, S3 includes the following steps:

[0030] S31. Set high-dimensional features As the initial search space of the improved Warhawk optimizer, each comprehensive high-dimensional feature vector Representation node The corresponding combination of characteristic relationships between sedimentary microfacies, geochemical carriers, and hydrological environments;

[0031] S32. Based on the characteristic spatial structure of the sedimentary facies-controlled phosphorus-enriched area, the Latin hypercube sampling method is used to initialize the population in the initial search space to obtain the initial population position set. , in the initial population position set, each individual position Represents the comprehensive high-dimensional feature vector Initial weight distribution scheme;

[0032] S33. Construct a four-stage improved Warhawk optimization search strategy for sedimentary facies-controlled phosphorus enrichment feature space, including factor weight guidance, feature interaction and coordination, risk hotspot penetration, and critical path precision targeting.

[0033] S34. Update individual positions after four stages of position after each iteration , and conduct fitness evaluation to define the sedimentary facies-controlled phosphorus enrichment risk response fitness function :

[0034] ;

[0035] in, Representation node Whether it is located in the risk hotspot area penetration weight factor, Indicates the contribution weight factor of the node in the critical path, is the phosphorus enrichment risk response index of the node;

[0036] S35. When the set iteration termination condition is met, the final feature weight vector set is output .

[0037] Optionally, in the factor weight guidance stage, a feedback-reinforced factor weight update is constructed based on the phosphorus activation fitness value corresponding to the current optimal individual position:

[0038] ;

[0039] in, For the The individual position at the iteration, is the optimal individual position of the current population, is the dynamic feedback coefficient in the factor weight guidance stage, Represented by the node The characteristic contribution matrix of the phosphorus activation sensitivity evaluation feedback, is the updated position result of collaborative individual j in the factor guidance stage;

[0040] In the feature interaction and coordination stage, based on the spatial correlation and causal relationship strength between the carbon-iron-phosphorus coupling knowledge graph nodes, a spatial interaction neighborhood is established to coordinate the interaction between the perception factors, and the individual position is updated as follows:

[0041] ;

[0042] in, is the spatial action intensity coefficient of the characteristic interaction synergy stage, For individuals The set of adjacent collaborative neighborhoods, represents the factor interaction weight coefficient between adjacent individuals, is the position result of the kth individual after the collaborative interaction stage;

[0043] The risk hotspot penetration phase targets nodes with high fitness values ​​in the current sedimentary facies-controlled phosphorus-enriched risk hotspot area, simulating the penetration maneuver search of key risk nodes by a swarm of warplanes:

[0044] ;

[0045] in, is the penetration maneuver control parameter in the risk hotspot penetration phase, is the position vector of the set of high phosphorus enrichment risk nodes in the current iteration, is the penetration speed adjustment matrix calculated based on the fitness gradient of the hotspot area nodes, is the search position result of the kth individual after the breakthrough phase;

[0046] In the critical path locking stage, the sedimentary phase-controlled phosphorus enrichment risk field and the contextual relationship network are combined to identify the path combination with the node feature vector that has the greatest risk contribution, and the search is carried out through an adaptive learning mechanism:

[0047] ;

[0048] in, is the fine search step parameter in the precise locking phase, The microfacies-carrier-hydrologic critical path position with the greatest contribution identified in the iteration, is the path accuracy guidance matrix obtained based on the risk contribution score, is the position result of the individual after the final stage.

[0049] Optionally, the S4 includes the following steps:

[0050] S41. Assign each feature weight in the final output feature weight vector set to the initialization attribute of the corresponding node;

[0051] S42. Construct a saliency-guided multi-scale graph neural network model structure. The saliency-guided multi-scale graph neural network model structure consists of a micro-phase convolution layer, a sub-graph attention aggregation layer, and a full-graph mosaic output layer. It sequentially performs feature extraction and multi-scale fusion on the input node initial embedding vector.

[0052] S43. In the micro-phase convolution layer, the initial embedding vector of each node is weighted and summed with the information of all adjacent nodes according to the adjacency weight. The output features of the node in the micro-phase convolution layer are obtained through linear transformation and nonlinear activation function.

[0053] S44. In the subgraph attention aggregation layer, for each local subgraph, the attention coefficient of each subgraph node to other nodes is calculated. The micro-phase convolution output features of all subgraph nodes are weighted summed according to the attention coefficient and linearly transformed to obtain the sub-layer output features.

[0054] S45. In the full-graph mosaic output layer, the sub-layer output features of all nodes are weighted and summed according to the significance score of each node and linearly transformed to obtain the risk structure representation of the entire graph.

[0055] Optionally, the calculation of the node initial embedding vector representation is to perform an element-by-element product operation on the feature weight of each node and the comprehensive high-dimensional feature vector of the node, which is used to express the personalized feature weight distribution of the node in the risk contribution.

[0056] Optionally, the S5 includes the following steps:

[0057] S51. Using the initial node embedding vector set as input, train the saliency-guided multi-scale graph neural network model structure to obtain the saliency-guided multi-scale graph neural network model parameter set and the node saliency score matrix;

[0058] S52. Perform a forward propagation operation on all nodes in the graph based on the trained saliency-guided multi-scale graph neural network model parameter set, and obtain a mosaic output vector for the entire sedimentary sequence graph by weighted summing the sub-layer output features of all nodes according to the saliency score of the corresponding node and the linear transformation weight of the entire graph mosaic layer;

[0059] S53. Using the node significance score matrix and the full-map mosaic output vector of the sedimentary sequence as input, the full-map mosaic information is projected onto a 90-meter resolution regular grid in the target area to construct a three-dimensional regular grid set, where each unit voxel in the three-dimensional regular grid set corresponds to a spatial coordinate;

[0060] S54. For each voxel unit in the three-dimensional space regular grid set, the node significance score corresponding to the voxel unit position and the mosaic output vector of the entire sedimentation sequence are combined, and the spatial contribution factor of the node to the voxel unit is calculated through the node position and the grid space distance. The spatial contribution factors of all nodes are weightedly summed with the node significance score and the mosaic output vector of the entire sedimentation sequence to obtain the phosphorus enrichment probability value of the voxel unit.

[0061] The beneficial effects of the present invention are:

[0062] The present invention combines Latin hypercube sampling with a multi-stage improved Warhawk optimizer. Through four major stages, namely, factor weight guidance, feature interaction and coordination, risk hotspot penetration, and precise locking of key paths, it targets the high-dimensional and complex coupling relationship of microfacies-carrier-hydrological multi-source characteristics in sedimentary facies-controlled phosphorus-enriched areas, and realizes the global optimization and local fine analysis of the feature weight vector. It can effectively avoid the omission of key factors and the local optimal trap, and improve the accuracy of identifying the main control paths of phosphorus activation and the mechanisms of high-risk areas.

[0063] The present invention embeds the final optimized feature weights into the initial node attributes, and designs a three-layer structure of micro-phase convolution, subgraph attention aggregation and full-graph mosaicking to achieve layer-by-layer focusing and feature fusion of multi-scale spatial structures and causal paths. Under the significance scoring mechanism, it can automatically highlight the nodes and paths that contribute most to the phosphorus enrichment risk, significantly improving the model interpretability and risk mechanism transparency. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0065] Figure 1 This is a flow chart of a method for spatial identification of groundwater phosphorus-enriched areas based on deep learning proposed in the present invention. DETAILED DESCRIPTION

[0066] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0067] refer to Figure 1 A method for spatial identification of groundwater phosphorus-enriched areas based on deep learning includes the following steps:

[0068] S1. Collect borehole samples of Quaternary lacustrine deposits in the target area, remote sensing images of sedimentary profiles, hydrological station monitoring records, and geochemical experimental results to form a multi-source geoscience dataset. This dataset is then input into a graph database using entity, attribute, and relationship mapping to construct a carbon-iron-phosphorus coupled knowledge graph encompassing sediment organic matter types, FeOOH crystallinity, hydrological evolution parameters, and their semantic connections.

[0069] S2. Expand the carbon-iron-phosphorus coupling knowledge graph, extract node attribute vectors, edge attribute vectors, and contextual relationship vectors, and concatenate them into a high-dimensional feature set;

[0070] S3. Initialize the population position of the improved Warhawk Optimizer using the high-dimensional feature set as input, so that the improved Warhawk Optimizer performs a three-stage formation search process in the search space and outputs a set of feature weight vectors;

[0071] S4. Write the feature weight vector set into the initial node attributes of the saliency-guided multi-scale graph neural network to construct the saliency-guided multi-scale graph neural network model structure;

[0072] S5. Train a saliency-guided multi-scale graph neural network model to obtain a node saliency score matrix and an embedding vector for the entire sedimentary sequence graph. Using this matrix and the embedding vector, calculate voxel-level phosphorus enrichment probability values ​​on a 90-meter resolution regular grid to generate a voxel-level phosphorus enrichment risk field.

[0073] S6. Load the voxel-level phosphorus enrichment risk field into the 3D geographic information system rendering engine, generate a 3D risk visualization model through ray casting and voxel sectioning, call the hydrological scenario generator to construct a hydrological scenario simulation data set, and couple it with the 3D risk visualization model to output a dynamic risk prediction sequence. After threshold comparison and trend judgment, generate land use restriction zoning, governance priority ranking and warning time curve.

[0074] In this embodiment, S1 includes the following steps:

[0075] S11. Collect borehole samples from the Quaternary lacustrine deposits in the target area to obtain a borehole sample set. Each borehole sample in the borehole sample set includes the longitude, latitude, burial depth, particle size index, and sediment color value of the borehole sample;

[0076] S12. Acquire sedimentary profile remote sensing image data to obtain a remote sensing image grid set, wherein each remote sensing grid pixel in the remote sensing image grid set includes geographical coordinates, visible light band reflectance information, and near infrared band reflectance information of the remote sensing grid pixel;

[0077] S13. Collect monitoring records of hydrological stations in the target area to obtain a hydrological station data set, where each hydrological station record in the hydrological station data set includes the location, maximum flow, average annual flow, and water level fluctuation amplitude of the hydrological station record;

[0078] S14. Obtain geochemical experimental analysis results corresponding to the drill hole samples to obtain a geochemical analysis data set. The analysis results of each sample in the geochemical analysis data set include an organic matter type indicator, FeOOH crystallinity value, and total phosphorus content per unit mass of sediment of the sample. The organic matter type indicator includes the ratio of aliphatic to aromatic compounds.

[0079] S15. Align the borehole sample collection, remote sensing image grid collection, hydrological station data collection, and geochemical analysis data collection to a unified geographic coordinate reference and a unified time reference to form a cross-source aligned geoscience multi-source dataset;

[0080] S16. Input the multi-source geoscience dataset into the graph database system and construct a carbon-iron-phosphorus coupling knowledge graph based on entity-attribute-relationship structure mapping. :

[0081] ;

[0082] in, Represents a set of entity nodes including sediment organic matter type, organic matter transformation products, FeOOH crystal state, microfacies type, and hydrodynamic boundary. Represents the multi-source coupling mechanism edge set between entities, including biological pump action, reduction and dissolution pathway, and sedimentary phase-controlled transport relationship edge types. Represents a set of entity node attributes, corresponding to color value, crystal index, and phosphorus concentration level index. Represents a set of edge attributes, corresponding to semantic features of causal weight, occurrence frequency, and spatial scope;

[0083] Coupling carbon-iron-phosphorus knowledge graph Layered slicing is performed according to time slices and spatial blocks.

[0084] In this embodiment, S2 includes the following steps:

[0085] S21. Knowledge graph of carbon-iron-phosphorus coupling Expand the graph structure and transform each entity node The corresponding entity node attributes are from the node attribute collection Extract and construct a node attribute vector set , where each node attribute vector Representation node corresponding sediment characteristics;

[0086] S22. For each edge From the edge attribute set Extract the semantic attributes of the edge and construct the edge attribute vector set , where the edge attribute vector Represents a connection node With node The causal mechanism relationship attributes, including causal weight, frequency of occurrence, and spatial scope;

[0087] S23. Construct a contextual relationship set based on the topological path and semantic path between nodes in the graph structure, and define the nodes The context neighborhood set , encode the context information into a context relationship vector Forming a contextual relationship vector set , where the contextual relationship vector Representation node High-order path information between its semantic neighbor nodes;

[0088] S24. Set the node attribute vector , edge attribute vector set Vector collection of contextual relationships Perform splicing and fusion to build a high-dimensional feature set :

[0089] ;

[0090] in, Representation node The comprehensive high-dimensional feature vector of Represents vector concatenation operation, Indicates the total number of nodes in the graph.

[0091] In this embodiment, S3 includes the following steps:

[0092] S31. Set high-dimensional features As the initial search space of the improved Warhawk optimizer, each comprehensive high-dimensional feature vector Representation node The corresponding combination of characteristic relationships between sedimentary microfacies, geochemical carriers, and hydrological environments;

[0093] S32. Based on the characteristic spatial structure of the sedimentary facies-controlled phosphorus-enriched area, the Latin hypercube sampling method is used to initialize the population in the initial search space to obtain the initial population position set. , in the initial population position set, each individual position Represents the comprehensive high-dimensional feature vector Initial weight distribution scheme;

[0094] S33. Construct a four-stage improved Warhawk optimization search strategy for sedimentary facies-controlled phosphorus enrichment feature space, including factor weight guidance, feature interaction and coordination, risk hotspot penetration, and critical path precision targeting.

[0095] S34. Update individual positions after four stages of position after each iteration , and conduct fitness evaluation to define the sedimentary facies-controlled phosphorus enrichment risk response fitness function :

[0096] ;

[0097] in, Representation node Whether it is located in the risk hotspot area penetration weight factor, Indicates the contribution weight factor of the node in the critical path, is the phosphorus enrichment risk response index of the node;

[0098] S35. When the set iteration termination condition is met, the final feature weight vector set is output .

[0099] In this implementation, during the factor weight guidance phase, a feedback-enhanced factor weight update is constructed based on the phosphorus activation fitness value corresponding to the current optimal individual position:

[0100] ;

[0101] in, For the The individual position at the iteration, is the optimal individual position of the current population, is the dynamic feedback coefficient in the factor weight guidance stage, Represented by the node The characteristic contribution matrix of the phosphorus activation sensitivity evaluation feedback, is the updated position result of collaborative individual j in the factor guidance stage;

[0102] The factor weight guidance phase aims to fully explore the differentiated contributions of various controlling factors to the risk of sedimentary facies-controlled phosphorus enrichment in the multi-source geological feature space. By introducing a feedback reinforcement mechanism, the phosphorus activation fitness corresponding to the previous optimal solution is dynamically applied to the weight update of the current factor, achieving adaptive adjustment of the factor weight distribution. Unlike traditional fixed weight or linear weight updates, this method explicitly embeds factor sensitivity derived from geological knowledge into the search strategy, enabling the optimization process to quickly focus on the variables with the greatest risk impact early on, improving search efficiency and the speed of convergence of the controlling path.

[0103] In the feature interaction and coordination stage, based on the spatial correlation and causal relationship strength between the carbon-iron-phosphorus coupling knowledge graph nodes, a spatial interaction neighborhood is established to coordinate the interaction between the perception factors, and the individual position is updated as follows:

[0104] ;

[0105] in, is the spatial action intensity coefficient of the characteristic interaction synergy stage, For individuals The set of adjacent collaborative neighborhoods, represents the factor interaction weight coefficient between adjacent individuals, is the position result of the kth individual after the collaborative interaction stage;

[0106] In the feature interaction and coordination stage, the synergistic effect and complex coupling between different factors in the process of simulating sedimentary phase-controlled phosphorus enrichment are studied. In the feature interaction and coordination stage, the spatial neighborhood of nodes and the causal association network are established, and the feature weights are coordinated by multiple agents according to the spatial proximity between nodes and the correlation of physical processes. Different from the independent search or simple local perturbation adopted by the existing optimization algorithms, the present invention incorporates a factor collaborative perception mechanism into the algorithm structure, so that the optimization individuals can perceive the information of other individuals in the neighborhood, realize the dynamic resonance adjustment of the factor weights, fully reflect the nonlinear superposition characteristics of the multi-factor action in the sedimentary environment, improve the modeling ability of complex phosphorus activation pathways, and effectively avoid the defect that traditional methods are prone to falling into local optimality in multi-peak search space.

[0107] The risk hotspot penetration phase targets nodes with high fitness values ​​in the current sedimentary facies-controlled phosphorus-enriched risk hotspot area, simulating the penetration maneuver search of key risk nodes by a swarm of warplanes:

[0108] ;

[0109] in, is the penetration maneuver control parameter in the risk hotspot penetration phase, is the position vector of the set of high phosphorus enrichment risk nodes in the current iteration, is the penetration speed adjustment matrix calculated based on the fitness gradient of the hotspot area nodes, is the search position result of the kth individual after the breakthrough phase;

[0110] The risk hotspot breakthrough stage solves the highly uneven risk distribution in sedimentary phase-controlled phosphorus-enriched areas and the surge in risks in key areas. During the optimization process, the risk hotspot breakthrough stage identifies high-phosphorus risk hotspot voxels or nodes in real time, actively shifts the optimization search focus to the highest-risk area, and simulates the strategy of a group of warhawks to conduct coordinated breakthroughs on target points. Different from the uniform search or focus only on the global optimum in traditional algorithms, the present invention achieves agile response and deep search of high-risk areas through dynamic hotspot locking and gradient-driven breakthrough maneuvers, effectively improving the ability to identify extremely high-risk spaces and ensuring that high-risk areas will not be diluted or missed.

[0111] In the key path locking stage, the sedimentary facies-controlled phosphorus enrichment risk field and the contextual relationship network are combined to identify the path combination with the node feature vector that has the greatest risk contribution. The search is carried out through an adaptive learning mechanism:

[0112] ;

[0113] in, is the fine search step parameter in the precise locking phase, The microfacies-carrier-hydrologic critical path position with the greatest contribution identified in the iteration, is the path accuracy guidance matrix obtained based on the risk contribution score, is the position result of the individual after the final stage.

[0114] In the critical path precise locking stage, an adaptive fine search mechanism based on risk contribution and contextual relationship network is introduced to address the link complexity and spatial heterogeneity of the main control mechanism in the sedimentary facies-controlled phosphorus enrichment system. In the critical path precise locking stage, the node significance score, risk spatiotemporal evolution trend and overall contribution weight of the path on the characteristic path are considered on the basis of the breakthrough results of the previous stage, and the microfacies-carrier-hydrological path that contributes most to risk prediction is dynamically locked and precisely adjusted. The critical path precise locking stage embodies the integration of causal inference and mechanism explanation capabilities, ensuring that the final output feature weight scheme is not only optimal but also highly interpretable.

[0115] In this embodiment, S4 includes the following steps:

[0116] S41. Assign each feature weight in the final output feature weight vector set to the initialization attribute of the corresponding node;

[0117] S42. Construct a saliency-guided multi-scale graph neural network model structure. The saliency-guided multi-scale graph neural network model structure consists of a micro-phase convolution layer, a sub-graph attention aggregation layer, and a full-graph mosaic output layer. The model sequentially extracts features and performs multi-scale fusion on the input node initial embedding vector. The micro-phase convolution layer is used to extract local spatial structural features between nodes, and the sub-graph attention aggregation layer is used to focus on structural correlation features in key areas.

[0118] S43. In the microfacies convolution layer, the initial embedding vector of each node is weighted and summed with the information of all adjacent nodes according to the adjacency weight. The output features of the node in the microfacies convolution layer are obtained through linear transformation and nonlinear activation function. The output features of the node in the microfacies convolution layer are used to express the local spatial structural characteristics between sedimentary microfacies.

[0119] S44. In the subgraph attention aggregation layer, for each local subgraph, the attention coefficient of each subgraph node to other nodes is calculated. Based on the attention coefficient, the micro-phase convolution output features of all subgraph nodes are weighted summed and linearly transformed to obtain the sub-layer output features. The sub-layer output features are used to express the key features of the local deposition control area.

[0120] S45. In the full-graph mosaic output layer, the sub-layer output features of all nodes are weighted and summed according to the significance score of each node and linearly transformed to obtain the risk structure representation of the entire graph. The significance score is derived from the comprehensive weight of the node in terms of critical path contribution and risk penetration sensitivity.

[0121] In this embodiment, the calculation of the node initial embedding vector representation is to perform an element-by-element product operation on the feature weight of each node and the comprehensive high-dimensional feature vector of the node. The node initial embedding vector representation is used to express the personalized feature weight distribution of the node in the risk contribution.

[0122] In this embodiment, S5 includes the following steps:

[0123] S51. Using the initial node embedding vector set as input, the saliency-guided multi-scale graph neural network model structure is trained. The training objective is to minimize the node-level risk prediction error loss function. The saliency-guided multi-scale graph neural network model parameter set and the node saliency score matrix are obtained. The node saliency score matrix is ​​used to measure the comprehensive contribution of each node in phosphorus enrichment risk prediction.

[0124] In the present invention, the construction of the loss function for minimizing the node-level risk prediction error is mainly based on the idea of ​​supervised learning to compare the true label of the phosphorus enrichment risk of each node with the risk prediction probability output by the significance-guided multi-scale graph neural network model. By defining commonly used loss functions such as weighted cross entropy or mean square error, the deviation between the true label of the node and the predicted value is quantified.

[0125] In this implementation, when weighted cross entropy is used, differentiated weights are assigned to nodes of different risk levels to enhance the accuracy of distinguishing high-risk nodes. The optimization goal is to adjust the network parameters to minimize the overall loss between the predicted outputs of all nodes and their actual risk levels.

[0126] S52. Based on the trained saliency-guided multi-scale graph neural network model parameter set, a forward propagation operation is performed on all nodes in the graph. The chimeric output vector of the entire sedimentary sequence graph is obtained by weighted summing the output features of the sub-layers of all nodes according to the saliency scores of the corresponding nodes and the linear transformation weights of the full-graph chimeric layer. The chimeric output vector of the entire sedimentary sequence graph is used to express the global chimeric expression of the risk of the entire sedimentary sequence under the spatial pattern of phosphorus enrichment.

[0127] S53. Using the node significance score matrix and the full-map mosaic output vector of the sedimentary sequence as input, the full-map mosaic information is projected onto a 90-meter resolution regular grid in the target area to construct a 3D regular grid set. Each unit voxel in the 3D regular grid set corresponds to a spatial coordinate, which is derived from the 3D grid coordinate set after spatial projection.

[0128] S54. For each voxel unit in the three-dimensional space regular grid set, the node significance score corresponding to the voxel unit position and the mosaic output vector of the entire sedimentation sequence are combined, and the spatial contribution factor of the node to the voxel unit is calculated through the node position and the grid space distance. The spatial contribution factors of all nodes are weightedly summed with the node significance score and the mosaic output vector of the entire sedimentation sequence to obtain the phosphorus enrichment probability value of the voxel unit. The phosphorus enrichment probability value of the voxel unit is used to represent the phosphorus risk probability density at the current position.

[0129] In this embodiment, S6 includes the following steps:

[0130] S61. Load the voxel-level phosphorus enrichment risk field as input into a 3D geographic information system rendering engine, perform 3D spatial mapping on the voxel-level phosphorus enrichment risk field, and generate a 3D risk voxel layer with a 90-meter resolution;

[0131] S62. Perform multi-directional slicing and spatial navigation of the 3D risk voxel layer in a 3D GIS rendering engine to obtain spatial structural views of the risk in different cross-sectional directions.

[0132] S63. Generate a hydrological scenario simulation dataset, which records the dynamic changes of hydrological variables in the form of time series;

[0133] S64. Dynamically couple the hydrological scenario simulation dataset with the 3D risk voxel layer. Input the hydrological variables at each moment into the 3D risk voxel model, driving the 3D risk distribution to change over time and hydrological context, and outputting a dynamic risk prediction sequence.

[0134] S65. Perform threshold comparison and trend analysis on the dynamic risk prediction sequence, partition the three-dimensional space according to the phosphorus enrichment risk probability density threshold, and form a land use restriction zoning layer. Define the corresponding spatial area as a land use restriction zone when the dynamic risk prediction value is greater than the phosphorus enrichment risk probability density threshold;

[0135] S66. For each moment and spatial unit, dynamic risk prediction values ​​are ranked according to risk level and its temporal evolution trend, and governance priorities are ranked for each spatial unit to form a governance priority ranking list.

[0136] S67. Count the high-risk occurrence time series of all spatial units within the simulation period and output the warning time curve for each land use unit.

[0137] Example 1: In response to the need for phosphorus enrichment risk management in a delta lake area in the east, a scientific research team cooperated with local water conservancy and environmental protection departments to carry out a one-year field data collection and intelligent risk simulation and prediction research based on the present invention, in order to provide a scientific decision-making basis for regional land space planning, agricultural non-point source pollution control and water environment management.

[0138] The Lake A basin is located in a plain lake area with a total area of ​​approximately 320 km². It has a dense water network, crisscrossing rivers and lakes, and significant seasonal fluctuations. In recent years, due to the multiple pressures of agricultural fertilization, urban non-point source input and extreme climate events, high phosphorus risks have frequently occurred in local bays and waterway mouths, and algae outbreaks and sediment phosphorus re-release are prominent problems in some areas. The monitoring data of the basin management department over the years show that the total phosphorus content in the sediments of the main lake area of ​​Lake A fluctuates between 0.571.98 mg / g, and can reach up to 2.34 mg / g in some bay areas. The maximum seasonal variation is nearly 60%. Traditional two-dimensional grid interpolation and empirical risk zoning methods can only achieve risk static layers at a resolution of 1 km×1 km. It is difficult to accurately identify future high-risk micro-areas for complex hydrodynamics and heterogeneous sedimentary environments, and the ability to trace and explain the main controlling mechanism of phosphorus release is insufficient, making it difficult to accurately implement management measures.

[0139] The team drilled 62 sedimentation boreholes throughout the Lake A basin and its main inflow channels, collecting sediment profiles at 0.60-cm depths at each site. Sampling was performed at 5-cm intervals, yielding a total of 685 borehole samples. Each sample was geochemically analyzed for TOC (total organic carbon), FeOOH crystallinity, particle size distribution, total phosphorus content, and aliphatic / aromatic ratio. Data from hydrological stations in the basin were also collected, including annual average flow, maximum flow, flood season water level fluctuations, duration of extreme flooding, concurrent flow velocity, COD, and NH4-N cofactors at 14 stations. Using drone remote sensing and satellite imagery, 0.5-meter resolution visible and near-infrared imagery was obtained for the entire area.

[0140] All borehole samples, remote sensing pixels, hydrological station data, and geochemical analysis results were spatially and temporally aligned according to a unified geographic coordinate system and observation time point, forming a complete geoscience multi-source dataset. Statistics show that sediment organic matter type indicators range from 0.89 to 2.57 (aliphatic / aromatic ratio), the FeOOH crystallinity index ranges from 16.37 to 4.5, the maximum hydrological extreme flow rate is 4.12 × 10³ m³ / s, and the average sediment phosphorus content is 1.22 mg / g. Some watershed areas exhibit distinct spatially isolated high-value islands.

[0141] Using the method presented in this paper, a coupled carbon-iron-phosphorus knowledge graph was constructed, comprising nodes for "sediment organic matter type," "FeOOH crystallinity," and "paleohydrologic evolution type." These nodes are connected by multiple edges related to "biological pumping," "sediment resuspension," and "Fe-P substitution." After the graph structure was expanded, node attribute vectors (phosphorus content, FeOOH index, particle size), edge attributes (edge ​​type, spatial weight), and contextual relationship vectors were extracted and assembled into a high-dimensional feature set.

[0142] During the intelligent optimization phase, an improved Warhawk optimizer was used to adaptively search for dominant control factor weights in a high-dimensional feature space, fully simulating the dominant control path penetration and coordination mechanisms within complex lake sedimentary environments. Through a four-stage optimization strategy, the factor combination with the highest risk for phosphorus enrichment was ultimately identified: microfacies fine-grained mud, high FeOOH, and summer high water level fluctuations. The weighted contribution of the path explained 32.4% of the total risk, a significant improvement over traditional principal component analysis methods.

[0143] The optimized feature weight vector is written into the saliency-guided multi-scale graph neural network. The initial embedding vector of each node is generated by element-by-element multiplication of the feature weight and the high-dimensional feature. The model structure adopts three layers of micro-facies convolution, two layers of sub-graph attention aggregation and one layer of full-graph mosaic output, which can automatically focus on key micro-areas and important factor paths at different spatial and semantic scales. The model training set uses historical monitoring data and field profile samples, totaling 900 samples, and the validation set uses new deployment points and independent hydrological periods, totaling 230 samples. During the training process, the judgment of phosphorus enrichment risk level is used as the supervision target, and weighted cross-entropy loss is used for optimization. The average consistency coefficient between the final node significance score and the expert experience weight is 0.93, while that of the traditional GNN method is 0.76.

[0144] In the spatial mapping stage, the significance score and the full-map mosaic output were projected onto a 90m×90m×2m three-dimensional regular grid in the Lake A area, with a grid number of 18,432 voxel units. According to the hydrological scenario during the flood season, the system can dynamically output the spatiotemporal evolution trend of high-risk voxels in the next one to five years. The results showed that the number of voxels in the high-risk area in summer identified by the method of the present invention was 1,560, and the actual overlap rate with the algal bloom outbreak and phosphorus-exceeding areas measured in the previous year was 86.5%, while the overlap rate of the traditional two-dimensional interpolation method was 63.8%. The accuracy of this method in early warning of potential high-risk areas in 2025-2026 was 82.4%, an increase of 23.7% compared with the traditional method.

[0145] At the same time, in terms of decision-making output, the system automatically delineated land use restriction zones based on a voxel risk probability density threshold of 0.7. That year, the number of voxels in the East Bay area of ​​Lake A with restricted land use was 410, and their spatial distribution was 91% consistent with the local ecological protection red line. Based on the interannual trend of voxel risk values, the top five priority areas for governance were generated: the northern part of the East Bay area, the north shore delta, the southwest of the main lake, the lake entrance, and the central sediment dam. The priority governance recommendations were fully consistent with the on-site expert assessments. The early warning time curve shows that between 2024 and 2028, the high-risk threshold for the central dam area will first be exceeded in April 2025, and in the northern part of the East Bay area in August 2026. This represents an advance warning time of approximately 14 months earlier than existing empirical methods.

[0146] Comparative data further demonstrate the beneficial effects of the present invention. In terms of screening the main controlling factors of phosphorus risk, the traditional correlation coefficient screening method was used, and the cumulative contribution of the top five factors was 48.7%, while the screening results of the present invention were 78.3%. Under three-dimensional high-resolution prediction, the spatial resolution of voxel risk of the present invention was increased to 90m, while the traditional method was only 1000m, and the spatial coverage accuracy was improved by more than 10 times. The overall recall rate of the model in predicting seasonal high-phosphorus areas in 2024 reached 0.87, while the traditional method was 0.65; the prediction accuracy of dynamic risk areas in the next three years was 0.82, while the traditional method was 0.59.

[0147] The specific training data is shown in Table 1 below:

[0148] Table 1 Specific training data

[0149]

[0150] In summary, this Example 1 truly demonstrates that in the A Lake Basin, a typical sedimentary phase-controlled phosphorus-enriched area, the three-dimensional intelligent risk prediction process of the present invention is used to greatly improve the scientificity and effectiveness of the main control mechanism identification, risk spatial positioning and dynamic governance decision-making, from high-dimensional heterogeneous data fusion, factor weight intelligent optimization, mechanism-explainable risk inference to three-dimensional spatial dynamic risk visualization and early warning output.

[0151] The present invention combines Latin hypercube sampling with a multi-stage improved Warhawk optimizer. Through four major stages, namely, factor weight guidance, feature interaction and coordination, risk hotspot penetration, and precise locking of key paths, it targets the high-dimensional and complex coupling relationship of microfacies-carrier-hydrological multi-source characteristics in sedimentary facies-controlled phosphorus-enriched areas, and realizes the global optimization and local fine analysis of the feature weight vector. It can effectively avoid the omission of key factors and the local optimal trap, and improve the accuracy of identifying the main control paths of phosphorus activation and the mechanisms of high-risk areas.

[0152] The present invention embeds the final optimized feature weights into the initial node attributes, and designs a three-layer structure of micro-phase convolution, subgraph attention aggregation and full-graph mosaicking to achieve layer-by-layer focusing and feature fusion of multi-scale spatial structures and causal paths. Under the significance scoring mechanism, it can automatically highlight the nodes and paths that contribute most to the phosphorus enrichment risk, significantly improving the model interpretability and risk mechanism transparency.

[0153] The present invention realizes the seamless coupling of the saliency-guided multi-scale graph neural network output and the high-resolution three-dimensional geographic information system, projects the node saliency score and the full-map risk vector to a 90-meter resolution three-dimensional voxel grid, supports voxel-level risk probability dynamic simulation, interactive ray casting visualization and scenario-driven multi-time series risk deduction, and combines hydrological scenario simulation to output dynamic risk spatiotemporal evolution sequence, and automatically generates land use restriction zoning, governance priority ranking and warning time curve.

[0154] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for spatial identification of groundwater phosphorus-enriched areas based on deep learning, characterized in that: The steps include: S1. Collect multi-source geoscience datasets and input them into a graph database by mapping entities, attributes, and relationships to construct a carbon-iron-phosphorus coupling knowledge graph. S2. Expand the carbon-iron-phosphorus coupling knowledge graph, extract node attribute vectors, edge attribute vectors, and contextual relationship vectors, and concatenate them into a high-dimensional feature set; S3. Initialize the population position of the improved Warhawk optimizer using the high-dimensional feature set as input, perform a four-stage formation search process in the search space, and output a set of feature weight vectors; S4. Write the feature weight vector set into the initial node attributes of the saliency-guided multi-scale graph neural network to construct the saliency-guided multi-scale graph neural network model structure; S5. Train a saliency-guided multi-scale graph neural network model to obtain a node saliency score matrix and an embedding vector for the entire sedimentary sequence graph. Using this matrix and the embedding vector, calculate voxel-level phosphorus enrichment probability values ​​on a 90-meter resolution regular grid to generate a voxel-level phosphorus enrichment risk field. S6. Load the voxel-level phosphorus enrichment risk field into a 3D geographic information system rendering engine to generate a 3D risk visualization model. Construct a hydrological scenario simulation dataset and couple it with the 3D risk visualization model to output a dynamic risk prediction sequence. Through threshold comparison and trend analysis, generate land use restriction zoning, governance priority ranking, and warning time curves. The S3 includes the following steps: S31. Set the high-dimensional feature set F high As the initial search space of the improved Warhawk optimizer, each comprehensive high-dimensional feature vector f i Represents node v i The corresponding combination of characteristic relationships between sedimentary microfacies, geochemical carriers, and hydrological environments; S32. Based on the characteristic spatial structure of the sedimentary facies-controlled phosphorus-enriched area, the Latin hypercube sampling method is used to initialize the population in the initial search space to obtain the initial population position set P (0) , in the initial population position set, each individual position Represents the comprehensive high-dimensional feature vector f i Initial weight distribution scheme; S33. Construct a four-stage improved Warhawk optimization search strategy for sedimentary facies-controlled phosphorus enrichment feature space, including factor weight guidance, feature interaction and coordination, risk hotspot penetration, and critical path precision targeting. S34. Update individual positions after four stages of position after each iteration And conduct fitness evaluation to define the sedimentary facies-controlled phosphorus enrichment risk response fitness function in, Represents node v i Whether it is located in the risk hotspot area penetration weight factor, Represents the contribution weight factor of the node in the critical path, ρ i is the phosphorus enrichment risk response index of the node; S35. When the set iteration termination condition is met, the final feature weight vector set W is output * ; The S4 comprises the following steps: S41. Assign each feature weight in the final output feature weight vector set to the initialization attribute of the corresponding node; S42. Construct a saliency-guided multi-scale graph neural network model structure. The saliency-guided multi-scale graph neural network model structure consists of a micro-phase convolution layer, a sub-graph attention aggregation layer, and a full-graph mosaic output layer. It sequentially performs feature extraction and multi-scale fusion on the input node initial embedding vector. S43. In the micro-phase convolution layer, the initial embedding vector of each node is weighted and summed with the information of all adjacent nodes according to the adjacency weight. The output features of the node in the micro-phase convolution layer are obtained through linear transformation and nonlinear activation function. S44. In the subgraph attention aggregation layer, for each local subgraph, the attention coefficient of each subgraph node to other nodes is calculated. The micro-phase convolution output features of all subgraph nodes are weighted summed according to the attention coefficient and linearly transformed to obtain the sub-layer output features. S45. In the full-graph mosaic output layer, the sub-layer output features of all nodes are weighted and summed according to the significance score of each node and linearly transformed to obtain the risk structure representation of the entire graph.

2. A method for spatial identification of groundwater phosphorus-enriched areas based on deep learning according to claim 1, characterized in that: Said S1 comprises the following steps: S11. Collect borehole samples from the Quaternary lacustrine deposits in the target area to obtain a borehole sample collection; S12. Acquire sedimentary profile remote sensing image data and obtain a remote sensing image grid set; S13. Collect monitoring records of hydrological stations in the target area to obtain a hydrological station data set; S14. Obtaining geochemical experimental analysis results corresponding to the drill hole samples to obtain a geochemical analysis data set; S15. Align the borehole sample collection, remote sensing image grid collection, hydrological station data collection, and geochemical analysis data collection to a unified geographic coordinate reference and a unified time reference to form a cross-source aligned geoscience multi-source dataset; S16. Input the multi-source geoscience dataset into the graph database system and construct the carbon-iron-phosphorus coupling knowledge graph G based on the entity-attribute-relationship structure mapping. C-F-P =(V,E,A V ,A E ), where V represents the entity node set including sediment organic matter type, organic matter transformation products, FeOOH crystal state, microfacies type, and hydrodynamic boundary; E represents the multi-source coupling mechanism edge set between entities, including the biological pump effect, reduction and dissolution phosphorus pathway, and sedimentary phase-controlled transport relationship edge type; A V Represents the attribute set of entity node, A E Represents a set of edge attributes; Coupling the carbon-iron-phosphorus knowledge graph G C-F-P Layered slicing is performed according to time slices and spatial blocks.

3. The method for spatial identification of groundwater phosphorus-enriched areas based on deep learning according to claim 1, characterized in that: The S2 comprises the following steps: S21. Knowledge graph of carbon-iron-phosphorus coupling G C-F-P Expand the graph structure and convert each entity node v i ∈V corresponding entity node attributes from the node attribute set A V Extract and construct a node attribute vector set Among them, each node attribute vector Represents node v i corresponding sediment characteristics; S22. For each edge e i,j ∈E from the edge attribute set A E Extract the semantic attributes of the edge and construct the edge attribute vector set Among them, the edge attribute vector Represents the connection node v i With node v j The causal mechanism relationship attributes; S23. Construct a contextual relationship set based on the topological path and semantic path between nodes in the graph structure, and define node v i The context neighborhood set N(v i ), encode the context information into a context relationship vector Forming a contextual relationship vector set Among them, the context relationship vector Represents node v i High-order path information between its semantic neighbor nodes; S24. Set the node attribute vector X V , edge attribute vector set X E and context relationship vector set X C Perform splicing and fusion to construct a high-dimensional feature set F high .

4. The method for spatial identification of groundwater phosphorus-enriched areas based on deep learning according to claim 1, characterized in that: In the factor weight guidance stage, the feedback-reinforced factor weight update is constructed based on the phosphorus activation fitness value corresponding to the current optimal individual position: in, is the individual position at the tth iteration, g (t) is the optimal individual position of the current population, α (t) is the dynamic feedback coefficient in the factor weight guidance stage, Represented by node f i The characteristic contribution matrix of the phosphorus activation sensitivity evaluation feedback, is the updated position result of collaborative individual j in the factor guidance stage; In the feature interaction and coordination stage, based on the spatial correlation and causal relationship strength between the carbon-iron-phosphorus coupling knowledge graph nodes, a spatial interaction neighborhood is established to coordinate the interaction between the perception factors, and the individual position is updated as follows: Among them, β (t) is the spatial action intensity coefficient of the characteristic interaction synergy stage, N p (k) is the collaborative neighborhood set adjacent to individual k, represents the factor interaction weight coefficient between adjacent individuals, is the position result of the kth individual after the collaborative interaction stage; The risk hotspot penetration phase targets nodes with high fitness values ​​in the current sedimentary facies-controlled phosphorus-enriched risk hotspot area, simulating the penetration maneuver search of key risk nodes by a swarm of warplanes: Among them, γ (t) is the penetration maneuver control parameter in the risk hotspot penetration phase, is the position vector of the set of high phosphorus enrichment risk nodes in the current iteration, is the penetration speed adjustment matrix calculated based on the fitness gradient of the hotspot area nodes, is the search position result of the kth individual after the breakthrough phase; In the critical path precise locking stage, the sedimentary phase-controlled phosphorus enrichment risk field and the contextual relationship network are combined to identify the path combination with the node feature vector that has the greatest risk contribution, and the search is carried out through an adaptive learning mechanism: Among them, η (t) is the fine search step parameter in the precise locking phase, The microfacies-carrier-hydrologic critical path position with the greatest contribution identified in the iteration, is the path accuracy guidance matrix obtained based on the risk contribution score, is the position result of the individual after the final stage.

5. The method for spatial identification of groundwater phosphorus-enriched areas based on deep learning according to claim 1, characterized in that: The calculation of the node initial embedding vector representation is to perform an element-by-element product operation on the feature weight of each node and the comprehensive high-dimensional feature vector of the node, which is used to express the personalized feature weight distribution of the node in the risk contribution.

6. The method for spatial identification of groundwater phosphorus-enriched areas based on deep learning according to claim 1, characterized in that: The S5 comprises the following steps: S51. Using the initial node embedding vector set as input, train the saliency-guided multi-scale graph neural network model structure to obtain the saliency-guided multi-scale graph neural network model parameter set and the node saliency score matrix; S52. Perform a forward propagation operation on all nodes in the graph based on the trained saliency-guided multi-scale graph neural network model parameter set, and obtain a mosaic output vector for the entire sedimentary sequence graph by weighted summing the sub-layer output features of all nodes according to the saliency score of the corresponding node and the linear transformation weight of the entire graph mosaic layer; S53. Using the node significance score matrix and the full-map mosaic output vector of the sedimentary sequence as input, the full-map mosaic information is projected onto a 90-meter resolution regular grid in the target area to construct a three-dimensional regular grid set, where each unit voxel in the three-dimensional regular grid set corresponds to a spatial coordinate; S54. For each voxel unit in the three-dimensional space regular grid set, the node significance score corresponding to the voxel unit position and the mosaic output vector of the entire sedimentation sequence are combined, and the spatial contribution factor of the node to the voxel unit is calculated through the node position and the grid space distance. The spatial contribution factors of all nodes are weightedly summed with the node significance score and the mosaic output vector of the entire sedimentation sequence to obtain the phosphorus enrichment probability value of the voxel unit.

Images