Deep learning based dynamic simulation and prediction system for agricultural non-point source pollution

Abstract

The application discloses a kind of based on deep learning agricultural non-point source pollution dynamic simulation and prediction system, it is related to agricultural ecological environment protection technical field, including: data acquisition and pre-processing module, for obtaining the original ground surface pixel matrix of target area and soil basic sampling sequence, and pre-processing generation three-dimensional root system damping tensor;Obtain rainfall amount numerical sequence and fertilization amount numerical sequence, and pre-processing obtains initial ground surface pollution concentration tensor.The application simulates the micro process that pollutant must overcome root interception and physical pore resistance to continue to infiltrate in real soil by reforming the interlayer feature transmission mechanism of three-dimensional convolutional neural network, effectively avoids that model appears seriously distorted prediction result when facing complex and changeable actual agricultural environment, improves the accuracy of prediction result.

Description

Technical Field

[0001] This invention relates to the field of agricultural ecological environment protection technology, specifically to a dynamic simulation and prediction system for agricultural non-point source pollution based on deep learning. Background Technology

[0002] Agricultural non-point source pollution is a significant factor contributing to groundwater ecological degradation. Accurately simulating the vertical infiltration of pollutants such as fertilizers into the groundwater level is crucial for agricultural environmental protection. In recent years, with the increasing prevalence of artificial intelligence technology, deep learning has been widely applied to the processing of environmental hydrological data and the prediction of pollution trends.

[0003] Existing intelligent prediction schemes mainly rely on a purely data-driven approach. These methods typically input large-scale environmental parameters directly into a standard neural network, relying entirely on the mathematical nonlinear mapping of the hidden layers within the network to fit the final groundwater pollution concentration.

[0004] However, the real agricultural geological environment exhibits strong spatial heterogeneity. The downward transport of pollutants in the soil is not a simple numerical fit, but is strictly limited by the physical resistance of soil pores at different depths and the objective interception by crop roots. Current neural networks generally employ conventional mathematical activation mechanisms when transmitting deep features, and this purely mathematical approach directly strips away the aforementioned real physical constraints. Even more critically, under real-world conditions such as continuous rainfall, the soil's ability to retain pollutants gradually approaches saturation over time. Because existing systems completely lack a dynamic perception and constraint mechanism for this physical accumulation saturation effect in their low-level feature calculations, once the surface environment undergoes drastic changes, their purely mathematical mapping will completely deviate from the real laws governing pollutants overcoming physical interception. Ultimately, this leads to severe distortions in the predictions of existing models when facing complex and ever-changing real agricultural environments, making it impossible to accurately reconstruct the true spatiotemporal evolution trajectory of pollutants crossing the soil profile and entering deep groundwater. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a deep learning-based dynamic simulation and prediction system for agricultural non-point source pollution.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] A deep learning-based dynamic simulation and prediction system for agricultural non-point source pollution includes:

[0008] The data acquisition and preprocessing module is used to acquire the original surface pixel matrix and soil basic sampling sequence of the target area, and perform preprocessing to generate a three-dimensional root damping tensor; acquire the rainfall numerical sequence and fertilizer application numerical sequence, and perform preprocessing to obtain the initial surface pollution concentration tensor.

[0009] The network inference input module is used to input the initial surface pollution concentration tensor into a three-dimensional convolutional neural network for infiltration inference inference. The three-dimensional convolutional neural network contains multiple feature extraction layers that map different soil physical depths.

[0010] The inter-layer feature correction and propagation module is used to perform the following correction and propagation steps for each feature extraction layer during the inter-layer feature propagation process in a 3D convolutional neural network:

[0011] Perform spatial convolution on the input tensor of the current layer to obtain the initial percolation tensor of the current layer;

[0012] Extract the cut-off threshold corresponding to the current layer depth from the three-dimensional root damping tensor, and determine whether the feature value of each node in the initial infiltration tensor of the current layer is greater than the cut-off threshold;

[0013] If so, extract the physical attenuation coefficient corresponding to the current layer depth from the three-dimensional root damping tensor, and correct the eigenvalue of the node based on the physical attenuation coefficient to obtain the corrected eigenvalue.

[0014] If not, then set the eigenvalue of that node to zero;

[0015] After updating the feature values ​​of all nodes, the data is reorganized into a sub-layer transfer tensor and input into the next feature extraction layer.

[0016] The inverse mapping result generation module is used to iteratively execute the correction and propagation step, traverse all feature extraction layers, extract the target feature map output from the bottom layer, and generate a deep groundwater pollution concentration distribution map through inverse mapping.

[0017] A deep learning-based method for dynamic simulation and prediction of agricultural non-point source pollution includes the following steps:

[0018] The original surface pixel matrix and soil basic sampling sequence of the target area are obtained and preprocessed to generate a three-dimensional root damping tensor; the rainfall numerical sequence and fertilizer application numerical sequence are obtained and preprocessed to obtain the initial surface pollution concentration tensor.

[0019] The initial surface pollution concentration tensor is input into a three-dimensional convolutional neural network for infiltration simulation. The three-dimensional convolutional neural network contains multiple feature extraction layers that map different soil physical depths.

[0020] In the inter-layer feature propagation process of a 3D convolutional neural network, the following correction propagation step is performed for each feature extraction layer:

[0021] Perform spatial convolution on the input tensor of the current layer to obtain the initial percolation tensor of the current layer;

[0022] Extract the cut-off threshold corresponding to the current layer depth from the three-dimensional root damping tensor, and determine whether the feature value of each node in the initial infiltration tensor of the current layer is greater than the cut-off threshold;

[0023] If so, extract the physical attenuation coefficient corresponding to the current layer depth from the three-dimensional root damping tensor, and correct the eigenvalue of the node based on the physical attenuation coefficient to obtain the corrected eigenvalue.

[0024] If not, then set the eigenvalue of that node to zero;

[0025] After updating the feature values ​​of all nodes, the data is reorganized into a sub-layer transfer tensor and input into the next feature extraction layer.

[0026] After iterating through all feature extraction layers in a loop, the target feature map of the lowest output layer is extracted and inversely mapped to generate a deep groundwater pollution concentration distribution map.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0028] This invention modifies the interlayer feature transfer mechanism of a three-dimensional convolutional neural network. It abandons the traditional pure mathematical activation function and introduces threshold determination and physical attenuation coefficient correction into the interlayer transfer process. This simulates the microscopic process by which pollutants in real soil must overcome root interception and physical pore resistance to continue infiltration, effectively improving the physical interpretability and inference accuracy of the model in agricultural environmental prediction. By calculating the ratio of the total physical depth to the number of network feature extraction layers, the equivalent step size of the physical infiltration depth corresponding to a single-layer network is established. The system binds the forward propagation process of the three-dimensional convolutional neural network to the vertical infiltration process of pollutants in the three-dimensional space of the soil, achieving synchronous inference between the feature space and the real physical space, and reconstructing the evolution trajectory of deep groundwater pollution. This effectively avoids severe distortion of prediction results when the model faces complex and ever-changing real agricultural environments, improving the accuracy of prediction results. Attached Figure Description

[0029] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0030] Figure 1 This is a system diagram of the present invention;

[0031] Figure 2 This is a flowchart of the present invention. Detailed Implementation

[0032] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0033] Application Overview:

[0034] In the field of agricultural ecological environmental protection, especially in the dynamic simulation of non-point source pollution infiltrating into groundwater, the accurate prediction of deep pollution concentration is regarded as a key indicator for measuring the reliability of environmental assessment models. This high-precision infiltration simulation is essentially a process of material transport and layer-by-layer damping attenuation at the hydrophysical level. That is, driven by the initial action of surface rainfall and fertilization, the three-dimensional pores of the soil are used as physical channels to transfer pollutants downward with water. Along the way, pollutants are physically hindered and intercepted by the volume of crop roots, thus forming a concentration distribution with specific dynamic characteristics in soil profiles at different depths.

[0035] However, existing prediction methods based on conventional deep learning lack a constraint mechanism to ensure consistency between the feature transmission between network layers and the actual physical infiltration mechanism. This results in an inability to accurately reproduce the physical interception and dynamic saturation breakthrough phenomena in complex geological environments. Physical interception manifests as a retention threshold created by the local root density during pollutant infiltration. Conventional network activation functions based on pure mathematics cannot identify this natural barrier. Dynamic saturation breakthrough, on the other hand, is characterized by a sharp decline in the soil's ability to buffer pollutants after the retention capacity of the soil pores reaches its limit, causing a sudden increase in infiltration and the failure of physical attenuation. Consequently, a strict constraint correspondence cannot be established between the feature mapping in the network and the actual physical infiltration displacement, leading to misjudgments or fuzzy inferences about the pollution destination under extreme conditions. This, in turn, affects the accurate assessment of the spatiotemporal evolution characteristics of pollution concentration and the targeted guidance for agricultural governance.

[0036] For example, in a simulation scenario of continuous heavy rainfall causing agricultural non-point source pollution, the prediction system, using a conventional 3D convolutional network, can only capture the diffusion trend of the initial surface concentration towards deeper layers over time, but cannot distinguish whether it encounters a dense crop root damping field at a specific depth. Furthermore, when the soil becomes saturated with water and pollutants due to prolonged rainfall, the system only records the appearance of a smooth decrease in feature values ​​between layers, failing to detect the failure of local soil interception thresholds and the resulting abnormal surge in penetration. Specifically, the system misclassifies physical interception zones as unobstructed infiltration channels, giving excessively high estimates of deep pollution, or incorrectly classifies explosive penetration after dynamic saturation as normal decay. This leads to the model continuously outputting prediction results that violate hydrodynamic principles, failing to form an infiltration evolution trajectory that conforms to real physical laws.

[0037] If the above problems are not addressed, the simulation prediction system will continue to lose its objective early warning capability for changes in the real groundwater environment. Specifically, the failure to effectively simulate physical interception will cause the network to over-amplify the permeability of deep soil, leading to deviations in the deduced pollution transmission paths from the principles of porosity and root damping, thereby weakening the model's assessment of the natural buffering effect of non-point source pollution. At the same time, the lack of a mechanism to correct for the dynamic saturation breakthrough problem will cause serious distortion of predicted concentrations under extreme weather conditions, making it impossible for deep groundwater pollution distribution maps to reflect sudden ecological risk characteristics, ultimately causing the system to lose its due dynamic capture capability. As a result, the inaccuracy of prediction feedback will systematically hinder environmental protection departments from obtaining core data on non-point source pollution diffusion, affecting the achievement of macro-governance goals such as groundwater resource protection.

[0038] like Figure 1-2 As shown, a deep learning-based dynamic simulation and prediction system for agricultural non-point source pollution includes:

[0039] The data acquisition and preprocessing module is used to acquire the original surface pixel matrix and soil basic sampling sequence of the target area, and perform preprocessing to generate a three-dimensional root damping tensor; acquire the rainfall numerical sequence and fertilizer application numerical sequence, and perform preprocessing to obtain the initial surface pollution concentration tensor.

[0040] Specifically, the system first reconstructs the spatialization and tensor quantization of multi-source heterogeneous agricultural environmental data through data acquisition and preprocessing modules. For surface and geological data, the system acquires the original surface pixel matrix and basic soil sampling sequence of the target area. By identifying the canopy distribution of surface crops and combining it with their downward growth physical characteristics, while integrating the porosity distribution of soil sampling points in three-dimensional space, the system preprocesses this data to generate a three-dimensional root damping tensor. This tensor physically characterizes the combined hindering effect of crop roots and soil matrix on hydrological infiltration within the three-dimensional underground space of the target area.

[0041] Simultaneously, based on meteorological and agricultural behavior data, the system acquires numerical sequences of rainfall and fertilizer application. By unifying the time base and performing dissolution transformation, the system transforms the discrete rainfall and fertilizer application data into an initial surface pollution concentration tensor with geospatial attributes, which serves as the initial driving force for subsequent network inference.

[0042] The network inference input module is used to input the initial surface pollution concentration tensor into a three-dimensional convolutional neural network for infiltration inference inference. The three-dimensional convolutional neural network contains multiple feature extraction layers that map different soil physical depths.

[0043] Specifically, after data preprocessing, the network inference input module inputs the initial surface pollution concentration tensor into a specially customized three-dimensional convolutional neural network for infiltration inference simulation. Unlike conventional neural networks, this three-dimensional convolutional neural network contains multiple feature extraction layers, and these feature extraction layers are logically and strictly mapped to different soil physical depths. As the network propagates deeper into the network, it essentially represents the layer-by-layer infiltration and displacement of surface pollutants with water flow in the soil physical space from top to bottom, thus establishing a synchronous mapping between the network hierarchy and the actual geological depth.

[0044] The inter-layer feature correction and propagation module is used to perform the following correction and propagation steps for each feature extraction layer during the inter-layer feature propagation process in a 3D convolutional neural network:

[0045] Perform spatial convolution on the input tensor of the current layer to obtain the initial percolation tensor of the current layer;

[0046] Extract the cut-off threshold corresponding to the current layer depth from the three-dimensional root damping tensor, and determine whether the feature value of each node in the initial infiltration tensor of the current layer is greater than the cut-off threshold;

[0047] If so, extract the physical attenuation coefficient corresponding to the current layer depth from the three-dimensional root damping tensor, and correct the eigenvalue of the node based on the physical attenuation coefficient to obtain the corrected eigenvalue.

[0048] If not, then set the eigenvalue of that node to zero;

[0049] After updating the feature values ​​of all nodes, the data is reorganized into a sub-layer transfer tensor and input into the next feature extraction layer.

[0050] Specifically, in the inter-layer feature transfer process of the 3D convolutional neural network, the system performs physically constrained correction transfer on each feature extraction layer through the inter-layer feature correction transfer module. The specific steps are as follows:

[0051] Preliminary infiltration simulation: First, spatial convolution operation is performed on the input tensor of the current layer. Local features are extracted using the convolution kernel to obtain the preliminary infiltration tensor of the current layer, which represents the preliminary diffusion state of pollutants in the physical space of the layer.

[0052] Retention Threshold Determination: Discarding the physically meaningless mathematical activation functions of traditional neural networks, the system extracts a retention threshold corresponding to the physical depth of the current feature extraction layer from a pre-constructed three-dimensional root damping tensor. Subsequently, it determines whether the pollution kinetic energy characteristic value of each node (representing a local spatial point) in the preliminary infiltration tensor exceeds this retention threshold, thereby simulating the natural phenomenon in real environments where pollution concentration must reach a certain level to overcome local root and capillary barriers.

[0053] Diversion Correction Update: If the feature value of a node is greater than the interception threshold, it indicates that the pollutant has successfully overcome the physical barrier at that depth. At this time, the system extracts the physical attenuation coefficient for the corresponding depth, performs attenuation calculation on the feature value, and obtains the corrected feature value that actually passes through the pore.

[0054] If the characteristic value of a node is not greater than the retention threshold, it indicates that the pollutants are completely retained by the soil or roots at that depth, and the system will force the characteristic value of that node to be set to zero.

[0055] Sub-layer transfer: After the feature values ​​of all nodes in the current layer are updated according to the above physical rules, the system reorganizes them into a sub-layer transfer tensor and inputs it into the next feature extraction layer to continue the infiltration deduction at the next depth.

[0056] The inverse mapping result generation module is used to iteratively execute the correction and propagation step, traverse all feature extraction layers, extract the target feature map output from the bottom layer, and generate a deep groundwater pollution concentration distribution map through inverse mapping.

[0057] Specifically, the above correction and propagation steps are executed iteratively until all feature extraction layers in the network are traversed, i.e., the groundwater layer at the target depth is extrapolated. The inverse mapping result generation module extracts the target feature map from the bottom layer. This map contains deep pollution feature information after layer-by-layer physical damping calculation. Finally, the system uses preset transformation rules to inverse map the dimensionless feature map into data containing real geographic coordinates and chemical concentration dimensions, ultimately rendering an intuitive deep groundwater pollution concentration distribution map, providing a scientific basis for agricultural environmental protection decisions.

[0058] The core innovation of the aforementioned technology lies in the modification of the inter-layer feature transfer mechanism of the 3D convolutional neural network. It abandons the traditional pure mathematical activation function and introduces the interception threshold determination and physical attenuation coefficient correction into the inter-layer transfer process. This simulates the microscopic process by which pollutants in real soil must overcome root interception and physical pore resistance to continue infiltration, effectively improving the physical interpretability and extrapolation accuracy of the model in agricultural environmental prediction. By calculating the ratio of the total physical depth to the number of network feature extraction layers, the equivalent step size of the physical infiltration depth corresponding to a single-layer network is established. The system binds the forward propagation process of the 3D convolutional neural network to the vertical infiltration process of pollutants in the three-dimensional space of the soil, achieving synchronous extrapolation between the feature space and the real physical space, and reconstructing the evolution trajectory of deep groundwater pollution. This effectively avoids severe distortion of prediction results when the model faces complex and ever-changing real agricultural environments, improving the accuracy of prediction results.

[0059] Traditional deep learning prediction models cannot directly handle natural environmental data with physical spatial heterogeneity. Furthermore, multi-source, heterogeneous input data (such as meteorological time-series data, discrete fertilization data, and soil spatial data) lack a unified formatted representation, causing neural networks to fail to perceive the actual physical barrier state between crop roots and soil pores during feature extraction. To address this, a data acquisition and preprocessing module is proposed, as follows:

[0060] Because soil sampling points obtained in the physical environment are limited by human and material resources, they are sparsely and irregularly distributed in space. Deep learning networks (such as 3D convolutional neural networks) require continuous, regular 3D tensor data without missing values ​​as input. Irregular discrete data cannot be directly input into neural networks for spatial convolution feature extraction. Therefore, this paper proposes generating a basic porosity matrix, specifically including:

[0061] Analyze the basic soil sampling sequence and extract the geospatial coordinate data and corresponding measured porosity values ​​of each physical sampling point;

[0062] The system acquires the three-dimensional geographic boundary information of the target area and the preset grid division resolution, and generates a three-dimensional spatial base grid that matches the target area. The three-dimensional spatial base grid contains multiple discrete grid nodes.

[0063] Traverse each grid node in the three-dimensional spatial base grid and calculate the spatial Euclidean distance between the current grid node and each physical sampling point;

[0064] Physical sampling points whose spatial Euclidean distance is less than a preset neighborhood threshold are selected as valid reference points. The reciprocal of the spatial Euclidean distance is used as the distance attenuation weight. The measured porosity values ​​of all valid reference points are weighted and summed to obtain the inferred porosity value of the current grid node.

[0065] The inferred porosity values ​​of all grid nodes are spliced ​​and assigned values ​​according to the spatial topology array order to generate a basic porosity matrix containing continuous spatial porosity distribution data.

[0066] Soil basic sampling sequence: refers to a series of soil sample data obtained at different spatial locations within a target area through physical sampling methods (such as using a soil corer). The data acquisition method in this sequence is as follows: using GPS equipment to record the latitude, longitude, and elevation information of the sampling points, and measuring the porosity parameters of each sample through conventional laboratory drying and weighing methods or scanning instruments.

[0067] Spatial Euclidean distance: refers to the absolute straight-line distance between two spatial coordinate points in a three-dimensional Cartesian coordinate system, used to measure the spatial proximity between grid nodes and physical sampling points.

[0068] Distance attenuation weight: In spatial interpolation calculations, based on the geographical principle that "the closer the distance, the stronger the correlation", variables that are closer to the sampling points are assigned a higher weight coefficient, while variables that are farther away are assigned a lower weight coefficient.

[0069] Theoretical porosity values ​​refer to the theoretical porosity estimates assigned to discrete grid nodes that have not undergone actual physical sampling, after calculation by an algorithm.

[0070] Spatial topology array order: refers to the index order in which three-dimensional mesh nodes are arranged in computer memory or matrix data structure according to a specific dimension of spatial coordinates (such as first the X-axis, then the Y-axis, and finally the Z-axis).

[0071] Step 1: Parse sampling data: Read the pre-acquired basic soil sampling sequence and extract from it... Data from the physical sampling point. For the Extract geospatial coordinate data from each physical sampling point. and the corresponding measured porosity values .

[0072] Step 2: Generate the 3D spatial base mesh: Obtain the 3D geographic boundary information of the target area, i.e., the maximum and minimum values ​​of longitude, latitude, and depth. Set a preset mesh resolution according to the required computational accuracy (e.g., the spatial step size is set to a fixed value between one and ten meters, the specific value depending on the size of the target area and computing resources). Divide the 3D geographic boundary at equal intervals according to this resolution to generate a 3D spatial base mesh containing multiple discrete mesh nodes. Let the coordinates of any current mesh node be... .

[0073] Step 3: Calculate the spatial Euclidean distance: Traverse each grid node in the 3D spatial base grid and calculate the distance for each grid node. With the physical sampling points Spatial Euclidean distance between The calculation formula is as follows:

[0074] ;

[0075] in: Indicates the relationship between the current grid node and the first Spatial Euclidean distance between physical sampling points; These represent the three-dimensional coordinate components of the current mesh node; They represent the first The three-dimensional coordinate components of each physical sampling point are obtained. By acquiring the actual straight-line physical distance between two points in three-dimensional space, an objective geometric quantitative indicator is provided for subsequent evaluation of the influence of the sampling points on unknown nodes.

[0076] Step 4: Filter valid reference points and calculate weights: Set a preset neighborhood threshold. (This threshold is typically set to two to five times the grid resolution to ensure sufficient reference points around each grid node.) Spatial Euclidean distance Less than the preset neighborhood threshold The physical sampling points are determined to be valid reference points.

[0077] For each selected valid reference point, the reciprocal of the spatial Euclidean distance is used as the distance decay weight. The calculation formula is as follows: ;

[0078] in: Indicates the first Distance decay weight of each effective reference point relative to the current grid node; This represents the spatial Euclidean distance between corresponding points. The inverse distance function is used to calculate the weights to implement the principle of local similarity of geospatial variables, ensuring that the closer a sampling point is to a grid node, the more dominant its measured porosity becomes in the final estimation, thus improving the accuracy of local extrapolation.

[0079] Step 5: Weighted summation to obtain the projected porosity value: Using the measured porosity values ​​of all valid reference points and their corresponding distance attenuation weights, perform a weighted summation to obtain the projected porosity value of the current grid node. The calculation formula is as follows:

[0080] ;

[0081] in: This represents the inferred porosity value of the current grid node; Indicates satisfaction The total number of valid reference points for the conditions; Indicates the first Distance decay weights for each effective reference point; Indicates the first The measured porosity values ​​at several effective reference points were used. By integrating the data characteristics of multiple neighboring sampling points, abrupt changes in spatial variables were smoothed, generating reasonable estimates that conform to the natural geological distribution patterns.

[0082] Step 6: Splicing and Assigning Values ​​to Generate a Matrix: This step involves combining the calculated porosity values ​​of all grid nodes. The coordinates are written into the corresponding index positions of the three-dimensional array in ascending order (longitude first, then latitude, then depth) to generate the basic porosity matrix.

[0083] By overcoming the limitation of uneven spatial distribution in the original sampling data, discrete, finite-point data was successfully transformed into a structured, continuous basic porosity matrix. This matrix directly characterizes the continuous variation of subsurface three-dimensional porosity in the target area, meets the stringent requirements of convolutional neural networks for input data format, and provides standardized physical environment parameters for subsequent infiltration simulations.

[0084] Example 1: Assume the preset grid resolution of a local subspace of the target region is 1 meter. The coordinates of an unknown grid node need to be calculated. The derived porosity value. Preset neighborhood threshold. The distance was set to 3 meters. By reading the soil baseline sampling sequence, it was determined that there were two physical sampling points around this node, with the following parameters:

[0085] Sampling point 1: coordinates Measured porosity values ;

[0086] Sampling point 2: coordinates Measured porosity values ;

[0087] Calculate spatial Euclidean distance:

[0088] Distance between the current node and sampling point 1: (rice);

[0089] Distance between the current node and sampling point 2: (rice);

[0090] Filtering valid reference points:

[0091] because Sampling point 1 is a valid reference point;

[0092] because Sampling point 2 is a valid reference point;

[0093] Calculate the distance decay weight:

[0094] Weight of sampling point 1: ;

[0095] Weight of sampling point 2: ;

[0096] Calculation and derivation of porosity values:

[0097] ;

[0098] Matrix assignment: assign the value of the node Assign the corresponding coordinates to the basic porosity matrix The location.

[0099] In simulations of non-point source pollution infiltration, the actual infiltration resistance is jointly determined by the bio-interception of plant roots and the physical confinement of soil pores. Traditional deep learning models lack a mechanism to uniformly quantify and map these two heterogeneous environmental factors into a three-dimensional network structure. Therefore, this paper proposes generating a three-dimensional root damping tensor, specifically including:

[0100] Convert the original surface pixel matrix into a grayscale matrix;

[0101] By using a preset edge detection operator to perform local gradient calculation on the gray value matrix, the set of gradient coordinates representing the crop canopy boundary is extracted;

[0102] The pixel nodes within the gradient coordinate set are assigned the first identifier value, and the pixel nodes outside the boundary are assigned the second identifier value to construct a two-dimensional canopy contour tensor.

[0103] Obtain the vertical downward growth step length of the root system corresponding to the crop variety in the target area. Based on the two-dimensional canopy contour tensor, perform multi-layer tensor replication and layer-by-layer element value reduction calculation along the depth direction according to the vertical downward growth step length.

[0104] The tensors after multi-level decreasing calculation are stacked in three dimensions to generate a three-dimensional root system occupancy tensor whose elemental values ​​represent the root growth volume within the three-dimensional spatial coordinate points.

[0105] The soil basic sampling sequence is mapped into a grid according to spatial coordinates to generate a basic porosity matrix;

[0106] The three-dimensional root system occupancy tensor is fused with the basic porosity matrix to generate the corresponding interception threshold and physical attenuation coefficient, thus constructing the three-dimensional root system damping tensor.

[0107] Raw surface pixel matrix: refers to color digital image matrix data containing the surface cover of the target area, obtained through drone aerial photography or remote sensing satellite equipment.

[0108] Edge detection operators: mathematical algorithms used in image processing (such as the Sobel operator) to locate the contour boundaries of crops on the ground by calculating the spatial rate of change (gradient) of the brightness of image pixels.

[0109] Two-dimensional canopy profile tensor: A two-dimensional numerical matrix whose elements consist of zeros and ones, used to represent the covered and uncovered areas of a crop in a two-dimensional plane.

[0110] Root vertical growth step length: refers to a fixed physical distance unit (such as 0.1 meters or 0.2 meters) set according to the biological characteristics of specific crop varieties (such as corn and wheat) to divide the root distribution layers in the depth direction.

[0111] 3D root system occupancy tensor: A 3D spatial matrix whose element values ​​represent the proportion of soil physical volume occupied by plant roots at a specific 3D spatial coordinate point.

[0112] Retention threshold: refers to the upper limit of the physical retention capacity of the root system and soil pores at a certain depth for infiltrating pollutants. When the pollution concentration is lower than this value, the pollutants are retained and cannot continue to infiltrate.

[0113] Physical attenuation coefficient: refers to the proportional constant of the physical attenuation of pollutant concentration as it passes through specific pores and root barriers.

[0114] Step 1: Image Grayscale Conversion: Obtain the raw surface pixel matrix captured by the drone. The values ​​of the red, green, and blue color channels in this matrix are weighted and summed according to a standard ratio to convert it into a single-channel grayscale matrix. The calculation formula is as follows:

[0115] ;

[0116] in: Representing coordinates The grayscale value at that location; , , These represent the red, green, and blue channel pixel values ​​at these coordinates, respectively. By removing the interference of color information on boundary extraction while preserving the brightness variation characteristics of the image, the data dimensionality of subsequent matrix operations is reduced.

[0117] Step 2: Edge Detection and Gradient Calculation: The local gradient of the grayscale matrix is ​​calculated using the pre-defined Sobel edge detection operator. The horizontal gradient is calculated separately. and vertical gradient Then, the gradient magnitude is obtained, and the pixel coordinates with magnitudes greater than a preset gradient threshold are extracted to form a set of gradient coordinates representing the crop canopy boundary. .

[0118] Step 3: Constructing a 2D canopy contour tensor: Using a polygon filling algorithm, the gradient coordinate set... The inner pixel nodes are assigned the first identifier value (1), and the outer pixels are assigned the second identifier value (0), thus constructing a two-dimensional canopy contour tensor. The calculation formula is as follows:

[0119] ;

[0120] ;

[0121] in: Represents the two-dimensional canopy profile tensor in coordinates The numerical value of the label at the location; Represents the set of gradient coordinates The enclosed internal region. By converting complex image pixels into a binary matrix, the computer can clearly distinguish between the root-covered initial surface area and the bare soil area without root cover.

[0122] Step 4: Generation of 3D Root System Occupancy Tensor: Obtain the vertical downward growth step length of the root system of the target crop variety. .by Based on this, the data is replicated layer by layer downwards along the depth direction, and the numerical value is decreased according to the natural law of exponential decay of plant root systems with depth, generating a three-dimensional root space tensor. The calculation formula is as follows:

[0123] ;

[0124] in: Representing coordinates The root growth volume at that location; The hierarchical index number representing the depth direction; Indicates the vertical downward growth step size; This represents the root volume depth decay constant (typically ranging from 0.1 to 0.5 depending on the crop type). , which is a natural constant. The natural exponential decay formula is used to simulate the three-dimensional distribution characteristics of real plant roots in soil, expanding the two-dimensional cover data of the plane into a three-dimensional spatial barrier parameter that conforms to biological laws.

[0125] Step 5: Generate the 3D root damping tensor through fusion calculation: Call the basic porosity matrix generated in the previous steps (its node values ​​are...) The three-dimensional root system occupancy tensor and the basic porosity matrix are fused element-wise according to their corresponding coordinates, and the retention threshold and physical attenuation coefficient are calculated respectively. Retention threshold calculation formula:

[0126] ;

[0127] in: Indicates the threshold for rejection; This indicates the root system interception weight (value range: 0.3 to 0.7). This represents the porosity retention weight (ranging from 0.05 to 0.2). The larger the root volume and the smaller the porosity, the stronger the retention capacity. This formula quantifies the blocking threshold resulting from the combined effects of biological and geological factors.

[0128] Formula for calculating physical attenuation coefficient:

[0129] ;

[0130] in: Indicates the physical attenuation coefficient; This represents the matrix permeability conversion constant (ranging from 0.5 to 0.9). By converting microscopic resistance into attenuation multipliers that can be directly accessed by the subsequent neural network, the higher the porosity and the smaller the root volume, the higher the attenuation coefficient (i.e., the smoother the flow).

[0131] By spatially aligning and mathematically fusing surface image features with subsurface soil sampling data, a three-dimensional root damping tensor was constructed. This provides standardized physical constraint parameters for simulating the spatiotemporal evolution of pollutants under complex geological and biological synergistic resistance, thus improving the accuracy of the simulation.

[0132] Example 2: Continuing from the previous example, the coordinates of the grid nodes in the target area are... Its extrapolated porosity value has been calculated to be... Suppose that the surface coordinates are determined through edge detection in drone images. Located inside the crop canopy outline, i.e., the value of the two-dimensional canopy outline tensor. Set the vertical downward growth step length of crop roots. Meters, root volume depth attenuation constant Root system weighting Pore ​​retention weight matrix permeability transformation constant .

[0133] Calculate the 3D root system place value: Current node depth level index :

[0134] ;

[0135] This indicates that at this depth, the root system occupies approximately [a certain area / area]. The reference volume.

[0136] Calculate the retention threshold:

[0137] ;

[0138] Calculate the physical attenuation coefficient:

[0139] .

[0140] Non-point source pollution is driven by both agricultural fertilization and meteorological rainfall, but these two types of data exhibit significant heterogeneity in their spatiotemporal distribution. Fertilization data is mostly spatially averaged and exhibits low temporal frequency, while rainfall data fluctuates frequently over time. If the raw data is directly input into the network, the model cannot learn the prerequisite that "no rainfall means no infiltration, and no fertilization means no source." Therefore, this paper proposes obtaining an initial surface pollution concentration tensor, specifically including:

[0141] Both the rainfall and fertilizer application data sequences were resampled to the same target timestamp.

[0142] The resampled fertilizer application numerical sequence is multiplied with the preset dissolution conversion constant, and the result is multiplied and weighted by the resampled rainfall numerical sequence to generate the initial single-point pollution characteristic value.

[0143] The initial single-point pollution characteristic values ​​are extended and assigned according to the two-dimensional geospatial grid of the target area to generate the initial surface pollution concentration tensor.

[0144] Rainfall numerical series: refers to a set of surface precipitation depth data that is continuously recorded at specific time intervals through local meteorological station monitoring equipment or meteorological satellite interfaces.

[0145] Fertilizer application numerical sequence: refers to the set of chemical fertilizer quality data applied to a unit area of ​​farmland at a specific point in time, obtained by consulting agricultural production management ledgers or IoT device records of the target area.

[0146] Resampling: A time series data processing method used to uniformly map, interpolate, or align data sequences with different sampling frequencies or different recording time points to a specified time base.

[0147] Target timestamp: The discrete time step reference point set by the simulation system, used to synchronize various dynamic environmental variables that change over time.

[0148] Dissolution conversion constant: A physical parameter characterizing the proportion of solid chemical fertilizers that are converted into liquid free pollutants that can infiltrate with hydrology under a specific amount of rainfall.

[0149] Initial single-point pollution characteristic value: The initial effective pollutant concentration value generated by the combined effects of fertilization and rainfall at a single spatial calculation node under a specific target timestamp.

[0150] Initial surface pollution concentration tensor: a two-dimensional numerical matrix whose internal elements represent the initial pollution concentration distribution on each grid node of the entire target area at a specific timestamp, serving as the input source for subsequent three-dimensional convolutional neural network inference.

[0151] Step 1: Data Acquisition and Unified Resampling: Obtain the original rainfall data sequence through the meteorological data interface and the original fertilizer application data sequence through the agricultural ledger database. Since fertilization is typically a low-frequency discrete event (e.g., once a month), while rainfall is a high-frequency continuous event (e.g., once an hour), they need to be uniformly resampled to the same target timestamp. .

[0152] For the fertilization amount numerical sequence, a forward filling algorithm is used for resampling, that is, the fertilization amount is kept valid for a period of time after the fertilization event, until the next fertilization event or natural dissipation ends; for the rainfall amount numerical sequence, the nearest time point alignment method is used for extraction to obtain the valid value of the target timestamp.

[0153] Step 2: Calculate the initial single-point contamination feature value: Obtain the target timestamp after resampling. The calculation involves comparing the applied fertilizer amount with the rainfall data. Introducing a preset solubility conversion constant, the product of the applied fertilizer amount and this constant is first calculated to obtain the soluble potential of the fertilizer. Then, this potential is multiplied and weighted by the rainfall data to generate an initial single-point pollution characteristic value. The calculation formula is as follows:

[0154] ;

[0155] in: This represents the initial single-point contamination feature value generated under the target timestamp; This represents the amount of fertilizer applied corresponding to the current timestamp obtained after resampling. This represents the preset solubility conversion constant, which is usually set between 0.1 and 0.4 depending on the type of fertilizer. This represents the rainfall value corresponding to the current timestamp obtained through resampling. By quantifying how rainfall causes fertilizer dissolution and conversion into free pollutants, isolated agricultural behavior data is physically coupled with meteorological data to determine the strength of effective pollution sources with true downward infiltration kinetic energy.

[0156] Step 3: Expand and assign values ​​to generate a tensor: Obtain the two-dimensional canopy contour tensor constructed in the previous steps (its node values ​​are identified as follows). A value of 1 represents a crop-covered area and an area with effective agricultural fertilization, while a value of 0 represents a non-agricultural area. The calculated initial single-point pollution characteristic values ​​are then expanded and assigned values ​​according to the two-dimensional geospatial grid of the target area. The calculation formula is as follows: ;

[0157] in: Represents the initial surface pollution concentration tensor in plane coordinates The node value at that location; This represents the initial single-point pollution characteristic value calculated in the second step; This represents the numerical value of the two-dimensional canopy contour tensor at the corresponding coordinates. By assigning mask values, the concentration data of a single point is reasonably distributed to the surface space where agricultural activities actually occur, thereby generating a two-dimensional initial source input matrix with geospatial distribution characteristics, which meets the data requirements of convolutional neural networks for two-dimensional or high-dimensional tensor structures.

[0158] By transforming asynchronous, discrete, multi-source macroscopic environmental data into synchronous, continuous, and physically logical surface pollution source tensors, the feature extraction bias caused by data temporal disorder in neural networks is effectively avoided, providing an accurate, quantitative, and physically causal initial driving force for the dynamic evolution of deep groundwater pollution.

[0159] Example 3: Continuing from the previous example, the planar coordinates of the grid nodes in the target area are... The previous steps have determined that this location is within the crop canopy, i.e., the two-dimensional canopy profile tensor value. The system sets the target timestamp for the current simulation to be... (e.g., 12:00 on a certain day). Data is obtained through database resampling:

[0160] The amount of fertilizer applied corresponding to the current timestamp (Unit: kilograms per hectare).

[0161] Rainfall value corresponding to the current timestamp (Unit: millimeters).

[0162] Based on the characteristics of commonly used urea fertilizers in the local area, the preset solubility conversion constant is... .

[0163] Calculate the initial single-point pollution characteristic value: Perform product and weighted calculation based on the dissolution formula:

[0164] .

[0165] The initial magnitude of the pollution characteristic at this time step was calculated to be 60.

[0166] Extended assignment to generate tensor node values: Spatial extended assignment based on the distribution of the land surface canopy:

[0167] .

[0168] In the aforementioned technology, complex agricultural geological environmental characteristics and meteorological agricultural behaviors are quantitatively transformed into a three-dimensional tensor format that can be directly recognized by neural networks. The three-dimensional root damping tensor provides objective spatial physical constraints for subsequent network operations, and the initial surface pollution concentration tensor provides a standardized input source for the deduction, thus providing the necessary data foundation for improving the accuracy and physical interpretability of the simulated infiltration process.

[0169] Because increasing the network depth of a conventional 3D convolutional neural network model merely represents a stacking of mathematical operation layers and lacks any spatial meaning, when using such models to address non-point source pollution, it is impossible to establish an objective physical correspondence between the feature transmission process within the network and the vertical infiltration displacement of pollutants in a real soil profile, resulting in a lack of interpretability in the intermediate computation process. Therefore, a network inference input module is proposed, as follows:

[0170] The three-dimensional convolutional neural network contains multiple feature extraction layers that map to different soil physical depths. The mapping relationship is established in the following way:

[0171] Obtain the total physical depth from the surface of the target area to the target groundwater level;

[0172] The ratio of the total physical depth to the total number of feature extraction layers in the 3D convolutional neural network is calculated to obtain the physical penetration depth step size equivalent for a single-layer network.

[0173] The product of the interlayer progression number of each feature extraction layer and the physical penetration depth step size is used as the query slice depth corresponding to that layer in the three-dimensional root damping tensor.

[0174] Target groundwater level: refers to the physical depth boundary of the groundwater body below the surface of the target area. It is usually obtained through local hydrogeological survey reports or historical data from groundwater monitoring wells, and serves as the endpoint depth for this pollution infiltration simulation.

[0175] Feature extraction layer: The basic computing unit in a three-dimensional convolutional neural network, in this scheme it is given a physical meaning. It is not only responsible for extracting data features, but also represents a discrete spatial stage in the pollution infiltration process.

[0176] Physical infiltration depth step equivalent: a conversion index that represents the specific physical distance that pollutants infiltrate downward in the real soil physical space for each layer of forward propagation operation in the neural network.

[0177] Interlayer progressive sequence number: A positive integer sequence that sequentially numbers the feature extraction layers from top to bottom according to the data flow direction of the three-dimensional convolutional neural network.

[0178] Query slice depth: The specific vertical spatial coordinates used for data retrieval within the 3D root damping tensor. Based on this value, the system can extract a 2D damping parameter matrix matching the current network layer from slices within the 3D tensor.

[0179] Step 1: Obtain the total physical depth: By reviewing the hydrogeological survey data of the target area, obtain the surface elevation and the elevation of the target groundwater level. Calculate the elevation difference between the two and use this difference as the total physical depth from the surface of the target area to the target groundwater level. The calculation formula is as follows:

[0180] ;

[0181] in: Indicates the total physical depth from the surface of the target area to the target groundwater level; Indicates the average surface elevation of the target area; This represents the elevation of the target groundwater level. By clearly defining the overall physical spatial span of the infiltration simulation, a boundary for the neural network's projection based on the real geological environment is set, ensuring that the simulation does not extend aimlessly downwards.

[0182] Step 2: Calculate the physical penetration depth step size equivalent: Obtain the structural parameters of the pre-constructed 3D convolutional neural network and extract the total number of feature extraction layers (this total number of layers is usually set between 10 and 50, depending on the required inference accuracy). Calculate the ratio between the total physical depth and the total number of feature extraction layers. The calculation formula is as follows: ;

[0183] in: This represents the physical penetration depth step size equivalent for a single-layer network. This represents the total physical depth obtained from previous calculations. This represents the total number of feature extraction layers in a 3D convolutional neural network. By discretizing the continuous physical space of soil, a conversion standard between abstract network layers and specific physical distances is established, giving each layer of the network a clear spatial span meaning.

[0184] Step 3: Establish the mapping relationship for query slice depth: During the layer-by-layer forward inference of the neural network, for any feature extraction layer, its inter-layer progression index is multiplied by the equivalent physical penetration depth step size to generate the actual depth of that layer in physical space, i.e., the query slice depth. The calculation formula is as follows:

[0185] ;

[0186] in: This indicates the query slice depth corresponding to the current feature extraction layer; This represents the interlayer progression number of the current feature extraction layer, with a value of [value missing]. Positive integers in; This represents the physical infiltration depth step size equivalent. By assigning absolute depth coordinates to each network layer, the system is guided to use these coordinates to index the cutoff threshold and attenuation coefficient at the current depth in the three-dimensional root damping tensor, ensuring that each network layer uses the actual soil physical constraints at the current depth during computation.

[0187] Example 4: Continuing from the previous example, assume the environmental parameters of the target area are as follows:

[0188] Obtain the surface elevation of the target area through hydrological monitoring wells. meters, target underground elevation Meters. The total number of feature extraction layers set in the 3D convolutional neural network constructed by the system. .

[0189] Calculate the total physical depth: (rice);

[0190] This indicates that the total estimated depth from the surface to the groundwater is 2.0 meters.

[0191] Calculate the physical infiltration depth step size equivalent: (rice);

[0192] This means that for every layer the network passes down, it is equivalent to the pollutant penetrating 0.2 meters downwards in the soil.

[0193] Calculate the query slice depth for each layer:

[0194] With the first layer of the network ( ) and the 5th floor ( For example:

[0195] Query slice depth of level 1: (rice);

[0196] Query slice depth at level 5: (rice);

[0197] Network layer to physical depth mapping table:

[0198] The mapping relationship between network layers and physical depth is shown in the table below (partial list of layers):

[0199]

[0200] In the aforementioned technology, the abstract structure of a three-dimensional convolutional neural network is endowed with specific agricultural hydrological and physical meaning. The layer-by-layer forward propagation process of the network is transformed into the spatial evolution process of pollutants infiltrating the soil layer by layer with water. This method enables the neural network to strictly follow the vertical distribution sequence of real geological space while performing high-dimensional feature extraction, providing an accurate depth coordinate system and basic extrapolation data for subsequent layered physical interception and retention determination.

[0201] Existing deep learning networks commonly employ purely mathematical activation functions such as ReLU or Sigmoid for nonlinear mapping during inter-layer feature transfer. These mathematical functions fail to represent the physical barriers encountered by pollutants in the soil; that is, they cannot identify the objective physical limitations that pollutant concentrations must reach to overcome local soil capillary forces and root resistance to downward infiltration. This leads to a discrepancy between the pollutant distribution patterns output by the network and the actual hydrodynamic processes. Therefore, an inter-layer feature correction and transfer module is proposed, as follows:

[0202] Input tensor: In the forward inference process of a three-dimensional convolutional neural network, the current feature extraction layer receives a multi-dimensional data matrix from the previous layer, which represents the concentration distribution of pollutants at the previous soil depth.

[0203] Spatial convolution operation: a mathematical operation that uses a local sliding window (convolution kernel) to perform weighted summation on the input tensor, used to extract the concentration gradient features between spatially adjacent nodes and simulate the physical diffusion process of pollutants in the three-dimensional space of soil.

[0204] Preliminary infiltration tensor: The intermediate result matrix obtained after the spatial convolution operation of the input tensor of the current layer, representing the theoretical infiltration concentration distribution before local root interception and pore physical attenuation.

[0205] Corrected characteristic value: After filtering by physical mechanisms (interception threshold determination and attenuation coefficient conversion), the actual pollution characteristic value that can penetrate the current soil depth and continue to infiltrate downwards is finally determined.

[0206] Sub-layer transfer tensor: The output matrix, which is recombined from the corrected eigenvalues ​​of all nodes in the current feature extraction layer, serves as the direct input source for the next deep feature extraction layer.

[0207] Step 1: Perform spatial convolution to generate the initial percolation tensor:

[0208] Because the infiltration of pollutants into soil is not a vertical straight line, but rather involves complex lateral and vertical spatial diffusion, traditional one-dimensional numerical calculation models struggle to efficiently capture the interactions between adjacent nodes in this local space. This results in the derived pollution plume edges not conforming to natural diffusion patterns. Therefore, this paper proposes obtaining the preliminary infiltration tensor of the current layer, specifically including:

[0209] Obtain the feature extraction tensor kernel containing the three-dimensional diffusion resistance weights of water in the soil;

[0210] The feature extraction tensor kernel is extracted by sliding across the input tensor of the current layer with a preset step size, and a local matrix dot product summation operation is performed at each sliding pause position;

[0211] The summation results of all sliding pause positions are reorganized according to their spatial positions to generate the preliminary infiltration tensor of the current layer.

[0212] Feature extraction tensor kernel: In a three-dimensional convolutional neural network, this refers to a multidimensional weight parameter matrix used to extract features from input data in a spatial dimension. In this technical solution, its internal numerical distribution is given specific physical meaning to simulate the hydrodynamic diffusion characteristics in a natural environment.

[0213] Three-dimensional diffusion resistance weights: specific parameter values ​​inside the feature extraction tensor kernel, used to characterize the proportion of spatial transmission resistance and concentration loss caused by the soil matrix when pollutants are physically diffused in the soil in all directions and vertically with water.

[0214] Sliding pause position: During the convolution operation, the center coordinate point at which the feature extraction tensor kernel stops and spatially aligns with the underlying local data matrix each time it moves on the data plane of the input tensor with a fixed stride.

[0215] Local matrix dot product summation: This is a mathematical calculation process that multiplies each resistance weight value in the feature extraction tensor kernel with the initial or previous layer pollution concentration value at the overlapping position in the input tensor one by one, and then sums all the multiplications into a single value.

[0216] Step 1: Obtain the feature extraction tensor kernel containing 3D diffusion resistance weights: The system calls a pre-initialized and trained feature extraction tensor kernel from the current feature extraction layer of the 3D convolutional neural network. The two-dimensional spatial size of this tensor kernel is usually preset to an odd-numbered matrix (the value range is mostly...). or The three-dimensional diffusion resistance weight distribution inside the structure shows the highest value at the center and decreases towards the edge, in order to simulate the physical law that pollutants penetrate the center most strongly and diffuse gradually weaken towards the edges.

[0217] Step 2: Perform local matrix dot product summation: Overlay the feature extraction tensor kernel onto the starting position of the input tensor of the current layer (or the initial surface pollution concentration tensor if it is the first layer). Slide the tensor across the input tensor with a preset step size (usually 1 to maintain consistent spatial resolution). At each sliding pause, perform a local matrix dot product summation. The calculation formula is as follows:

[0218] ;

[0219] in: This indicates the center coordinates of the sliding pause position in the current layer. The preliminary infiltration characteristic values ​​were calculated at the location. This represents the spatial side length of the feature extraction tensor kernel (e.g., 3). This represents the value of the input tensor node in the current layer. This represents the preset step size by which the feature extraction tensor kernel slides on the input tensor. and This represents the local horizontal and vertical coordinate indices within the feature extraction tensor kernel. The feature extraction tensor kernel is located at the local index. The three-dimensional diffusion resistance weight at the location. This represents the constant of the calculated bias term for the current layer (simulating the baseline environmental concentration of the current environment, usually initialized to 0). The convolution calculation formula is used to comprehensively evaluate the joint contribution of the pollution concentration of the target node and its surrounding adjacent areas to the vertical infiltration of the node, thereby transforming isolated single-point concentration data into a spatially coherent theoretical diffusion concentration, consistent with the physical characteristics of water diffusing in porous media towards the direction of least resistance.

[0220] Step 3: Reorganize and generate the initial percolation tensor of the current layer: After the feature extraction tensor kernel has traversed all valid sliding pause positions on the input tensor, output all positions. The values ​​are based on the corresponding planar spatial coordinates. Topological sorting and splicing are performed to generate a complete preliminary infiltration tensor for the current layer. This tensor serves as the input basis for subsequent "root and pore retention threshold determination".

[0221] Through deep coupling of data-driven approaches and physical mechanisms, each feature value update rigorously corresponds to the actual situation of pollutants encountering interception and loss at a specific depth, thereby significantly improving the spatial resolution and physical reliability of the dynamic simulation of the infiltration process.

[0222] Example 5: Continuing from the previous example, the planar coordinates of the grid nodes in the target area are... In the preceding steps, the initial surface pollution concentration tensor values ​​for this coordinate and its neighboring nodes have all been calculated to be 60 (assuming uniform fertilization and rainfall in the area). At this point, the system enters the first layer of the network (i.e.,...). Coordinates need to be calculated. The initial infiltration tensor value of the current layer. The spatial dimensions of the feature extraction tensor kernel are set. (Right now Matrix), sliding step size Bias term Since the simulation measures water diffusion resistance, the center of infiltration has the highest weight, gradually decreasing towards the periphery, resulting in the extracted... Three-dimensional diffusion resistance weight The distribution is as follows:

[0223] Center point weight: 0.516;

[0224] The weights of the four adjacent points (top, bottom, left, and right) are 0.08 each.

[0225] Weights of the four diagonal edge points: 0.02 for each;

[0226] Kernel alignment and data extraction: The center sliding pause position of the convolution kernel is aligned with the coordinates of the input tensor. Since the input concentration of the surrounding nodes is all 60, the system extracts the corresponding... Local input matrix.

[0227] Perform local matrix dot product summation: According to the formula, multiply the values ​​of the local input matrix with the diffusion resistance weights one by one and then add them together.

[0228] .

[0229] Step 2: Threshold Extraction and Determination Based on the Query Slice Depth of the Current Layer Extract the corresponding two-dimensional slices from the pre-generated three-dimensional root damping tensor, and obtain the cutoff threshold of each coordinate node at that depth. Each node in the initial infiltration tensor of the current layer is checked to determine whether its eigenvalue is greater than the cutoff threshold.

[0230] Step 3: Physical attenuation correction or zeroing of eigenvalues: Perform flow splitting based on the determination result.

[0231] If (i.e.) Extract the physical attenuation coefficient corresponding to this coordinate. .

[0232] The eigenvalues ​​of this node are corrected based on the physical attenuation coefficient to obtain the corrected eigenvalues, as follows:

[0233] Existing prediction models typically treat soil physical damping parameters as static constants that do not change over time. However, under real continuous rainfall or high-intensity sewage discharge conditions, local soil pores are gradually filled by water and pollutants, causing subsequent pollutants to penetrate directly without attenuation. Static models cannot identify this dynamic saturation breakthrough phenomenon, leading to calculation biases in long-term simulations. Therefore, this paper proposes: dynamically acquiring a saturation penalty factor, specifically including:

[0234] Read the preset cumulative status register to obtain the amount of soil pollution retained by the current node within the historical time step;

[0235] Extract the cutoff threshold and physical attenuation coefficient corresponding to the current node from the three-dimensional root damping tensor, calculate the ratio of the cutoff threshold to the physical attenuation coefficient, and generate the dynamic saturation sensitivity constant of the current node.

[0236] The saturation index of the current node is obtained by multiplying the residual soil pollution at the node with the dynamic saturation sensitivity constant.

[0237] After taking a negative value for the saturation index, a natural exponent mapping operation is performed to generate a dynamic saturation penalty factor whose value is limited to the range of zero and one, and the value in the cumulative state register is updated simultaneously using the modified eigenvalue.

[0238] Cumulative Status Register: An internal storage matrix allocated by the system to store the cumulative status of data over time series. In this scheme, it is specifically used to record the total amount of pollutants intercepted and retained in that local area by each spatial node over all past time steps.

[0239] Nodal soil pollution retention: The characteristic concentration of pollutants that has been accumulated and stored at a specific soil spatial node before the current time step due to physical interception and adsorption.

[0240] Dynamic saturation sensitivity constant: A conversion parameter characterizing the sensitivity of a specific spatial node to pollutant accumulation. This value is determined by the proportional relationship between local physical retention capacity and physical penetration attenuation characteristics.

[0241] Saturation index: A dimensionless intermediate state variable used to comprehensively measure whether the accumulated pollution stock of the current node is approaching the saturation extreme value that the physical properties of the node can withstand.

[0242] Dynamic saturation penalty factor: A dynamic multiplier whose value is strictly limited to between zero and one. It is used to proportionally reduce the subsequent physical retention and attenuation capacity of soil as it approaches saturation.

[0243] Step 1: Read the residual soil pollution at each node: When processing the infiltration simulation at the current time step, the system reads the preset cumulative status register inside the system. The depth corresponding to the current feature extraction layer is extracted from this register. Below, coordinates are The accumulated amount of soil pollution at each node within a historical time step. This amount is preset to zero for all nodes during the initial system operation (i.e., the first time step).

[0244] Step 2: Calculate the dynamic saturation sensitivity constant: The system extracts the dynamic saturation sensitivity constant from the three-dimensional root damping tensor and the coordinates of the current node. The corresponding cutoff threshold and physical attenuation coefficient are then calculated. Subsequently, the ratio of the cutoff threshold to the physical attenuation coefficient is calculated to generate the dynamic saturation sensitivity constant for the current node. The calculation formula is as follows:

[0245] ;

[0246] in: This represents the dynamic saturation sensitivity constant of the current node; This indicates the extracted threshold value; This represents the extracted physical attenuation coefficient. The interception threshold represents the maximum local static resistance, while the physical attenuation coefficient represents the local dynamic flow smoothness. The ratio of the two objectively reflects the degree of congestion at this node when facing continuous infiltration, thus transforming static physical spatial parameters into a sensitivity indicator of the cumulative response over time.

[0247] Step 3: Calculate the saturation index: Multiply the obtained soil pollution retention volume at each node by the dynamic saturation sensitivity constant to obtain the saturation index of the current node. The calculation formula is as follows:

[0248] ;

[0249] in: This represents the saturation index of the current node; This indicates the amount of soil pollution retained at the read node. By coupling the absolute amount of pollution accumulation with local sensitivity characteristics, a unified quantitative measurement standard for the state of "saturation" is established for different soil physical environments.

[0250] Step 4: Generate a dynamic saturation penalty factor and update the register: Negate the calculated saturation exponent and use it as the exponent of the natural constant. Perform a natural exponent mapping operation to generate the dynamic saturation penalty factor. The calculation formula is as follows:

[0251] ;

[0252] in: This represents the dynamic saturation penalty factor; This represents the base of the natural logarithm. By using the exponential decay formula to simulate the non-linear decline in system buffer capacity with increasing load in nature, the generated penalty factor is strictly limited to the range of zero and one. When the stock is small, the penalty factor approaches one (no penalty); when the stock is extremely large, the penalty factor approaches zero (failure due to interception).

[0253] After completing the feature transfer calculation for the current time step, the system synchronously extracts the corrected feature value generated by the current node, adds it to the accumulated state register, and completes the numerical update. The update formula is as follows:

[0254] ;

[0255] in, This indicates the amount of soil pollution retained in the new generation node after the update is written to the register; This represents the corrected eigenvalue generated at this node at the current time step.

[0256] By endowing the inference system with "time memory" capabilities, the network can adaptively adjust its current attenuation capacity in real time based on the accumulated state of previous time steps when performing inter-layer feature correction. This accurately recreates the nonlinear dynamic process of soil from absorption and buffering to saturation breakdown, improving the model's predictive reliability in extreme conditions.

[0257] Example 6: Continuing from the previous example, the coordinates of the grid nodes in the target area are... .

[0258] The fixed physical parameters of the node coordinates are known to be:

[0259] Retention threshold ;

[0260] Physical attenuation coefficient ;

[0261] Compare the initial feature values ​​with the cutoff threshold: The result is "yes", indicating that the pollution can penetrate the root system and pores at this depth.

[0262] Suppose that when the system reaches the current time step, the accumulated status register is read, and it is found that the node has accumulated a certain amount of pollutants during the historical time steps, and the soil pollution retention volume of the node is obtained. .

[0263] Calculate the dynamic saturation sensitivity constant: using the ratio formula.

[0264] ;

[0265] Calculate the saturation index: Multiply the historical stock by the sensitivity constant:

[0266] ;

[0267] Generate dynamic saturation penalty factor: Perform natural exponential mapping operation:

[0268] ;

[0269] This result indicates that, due to existing historical accumulation, the physical decay capability of this node can only achieve a fraction of its initial state. .

[0270] The corrected feature value at the current time step is used for cumulative updates. This updated value will be saved and used for the saturation calculation of the node in the next time step.

[0271] Traditional neural networks or static hydrological models typically use a fixed attenuation coefficient when calculating infiltration attenuation. However, in real-world environments, as rainfall and infiltration continue, soil pore water saturation rises, and the attenuation rate of pollutants along their path dynamically changes. Static attenuation calculations cannot reflect this dynamic evolution, leading to deviations from actual deep pollution infiltration rates in long-term simulations or under continuous rainfall conditions. Therefore, this paper proposes correcting the eigenvalues ​​of this node based on the physical attenuation coefficient, resulting in corrected eigenvalues, specifically including:

[0272] The difference between the feature values ​​that are greater than the cutoff threshold and the cutoff threshold is calculated to obtain the preliminary remaining feature values.

[0273] The saturation penalty factor is dynamically obtained, and the physical attenuation coefficient at the corresponding coordinate position of the node is fused with the saturation penalty factor to obtain the real-time attenuation coefficient.

[0274] The preliminary remaining characteristic values ​​are corrected based on the real-time attenuation coefficient to obtain the corrected characteristic values ​​that characterize the actual amount of pollutant infiltration through the current soil pores.

[0275] Preliminary residual characteristic value: refers to the effective pollution characteristic value remaining after deducting the critical value corresponding to the interception resistance after the initial infiltration characteristic value of the current node exceeds the local interception threshold. It represents the initial effective pollution amount that has the ability to continue to infiltrate downward.

[0276] Real-time attenuation coefficient: refers to the pollution infiltration attenuation ratio constant obtained by combining the static physical attenuation coefficient with the dynamic saturation penalty factor that characterizes the cumulative saturation state of the soil, which can reflect the current time step and the actual saturation degree of the soil.

[0277] Step 1: Calculate the initial residual eigenvalues:

[0278] For nodes in the initial percolation tensor of the current feature extraction layer whose eigenvalues ​​are greater than the cutoff threshold, the difference between their eigenvalues ​​and the cutoff threshold at the corresponding extracted coordinates is calculated to obtain the initial remaining eigenvalues. The calculation formula is as follows:

[0279] ;

[0280] in: Indicates the spatial coordinates of the current layer And the depth slice is The initial residual eigenvalues ​​at the location. This represents the node eigenvalues ​​of the initial infiltration tensor of the current layer at the corresponding coordinates. This represents the interception threshold corresponding to the current node depth extracted from the three-dimensional root damping tensor. By subtracting the critical amount of pollution that can be intercepted by soil pores and root adsorption, the residual pollution characteristic base that truly possesses the ability to flow and infiltrate after overcoming static physical resistance is accurately separated.

[0281] Step 2: Calculate the real-time attenuation coefficient: The system obtains the dynamic saturation penalty factor calculated in the previous step. Simultaneously, the physical attenuation coefficient corresponding to the current node is extracted from the three-dimensional root damping tensor. The real-time attenuation coefficient is obtained by multiplying and fusing the physical attenuation coefficient and the dynamic saturation penalty factor. The calculation formula is as follows:

[0282] ;

[0283] in: This represents the real-time decay coefficient of the current node at the current time step. This represents the inherent physical attenuation coefficient of the current node. This represents the dynamic saturation penalty factor generated in the preceding steps, with a value range limited to 0 to 1. By dynamically adjusting the static physical attenuation coefficient based on the historical cumulative saturation state of the soil, the attenuation coefficient is adjusted accordingly as the soil saturation level increases, accurately characterizing the dynamically changing hydrodynamic transport efficiency.

[0284] Step 3: Calculate the corrected characteristic value: Based on the real-time attenuation coefficient, perform a product correction calculation on the preliminary remaining characteristic value to obtain the corrected characteristic value representing the actual amount of contaminated infiltration through the current soil pores. The calculation formula is as follows:

[0285] ;

[0286] in: This represents the corrected feature value generated after the current layer has been modified. This represents the initial residual characteristic value obtained from the first step of the calculation. This represents the real-time attenuation coefficient obtained in the second step of the calculation. By quantitatively combining the amount of pollution after exceeding the interception limit with the dynamic attenuation rate, the actual amount of pollution infiltrating to the next soil depth after deducting adsorption interception and dynamic infiltration loss is obtained, ensuring that the calculation results strictly follow the laws of mass transport and conservation.

[0287] The model achieves a deep integration of static physical damping parameters and dynamic spatiotemporal evolution. It can adaptively adjust the attenuation ratio according to the real-time saturation level of the soil, improving the accuracy and physical consistency of simulating agricultural non-point source pollution infiltration under complex conditions (such as continuous heavy rainfall).

[0288] Example 7: Continuing from the previous example, for the grid node coordinates The relevant parameters are consistent with the calculation results above as follows:

[0289] Preliminary infiltration characteristic values ​​of the current layer: ;

[0290] Retention threshold: ;

[0291] Static physical attenuation coefficient: ;

[0292] The dynamic saturation penalty factor calculated in the previous "saturation penalty factor" step is as follows:

[0293] .

[0294] Calculate the preliminary residual characteristic value: Subtract the preliminary infiltration characteristic value from the retention threshold:

[0295] ;

[0296] Calculate the real-time attenuation coefficient: Multiply the physical attenuation coefficient by the dynamic saturation penalty factor:

[0297] ;

[0298] Calculate the corrected eigenvalue: Multiply the initial residual eigenvalue by the real-time attenuation coefficient:

[0299] .

[0300] If not (i.e.) If the concentration is less than a certain value, it indicates that the pollution concentration is insufficient to overcome the physical barriers at the current depth and is either completely adsorbed by the soil or intercepted by the roots. The calculation formula is as follows:

[0301] The forced zeroing is to simulate physical blocking phenomena and filter out noise transmission that does not conform to physical laws.

[0302] Step 4: Reorganize into a secondary transit tensor: After the feature values ​​of all planar spatial coordinate nodes in the current layer have been updated according to the above rules, they are spliced ​​and combined in the original spatial topology array order to generate a secondary transit tensor, which is then passed as an input tensor to the next feature extraction layer.

[0303] In the aforementioned technology, the microscopic physical mechanism of pollutant infiltration in soil—"overcoming resistance-attenuation-retention and stagnation"—is directly encoded into the feature activation and transmission rules of the bottom layer of a neural network. This effectively filters out noise transmission in the network that does not conform to physical laws, enabling each calculation of the feature extraction layer to truly reflect the mass loss and blockage state of pollutants at a specific soil depth, thereby improving the physical accuracy of dynamic simulation.

[0304] Because deep learning models typically output high-dimensional, dimensionless abstract feature tensors after forward propagation, this type of data lacks direct physical meaning and geospatial attributes. Agricultural and environmental protection workers cannot directly understand the specific pollutant concentrations and actual geographical locations it represents, making the algorithm's inference results unsuitable for direct application in actual groundwater environmental monitoring and pollution control projects. Therefore, an inverse mapping result generation module is proposed, as follows:

[0305] Generate a deep groundwater pollution concentration distribution map, specifically including:

[0306] Read the pre-stored conversion ratio parameter matrix from dimensionless characteristic values ​​to actual chemical dimensions;

[0307] The elemental feature values ​​in the target feature map output at the bottom layer are multiplied by the conversion ratio parameter matrix to convert them into actual chemical oxygen demand and nitrogen and phosphorus concentration values.

[0308] A deep groundwater pollution concentration distribution map is generated by combining real chemical oxygen demand (COD) and nitrogen and phosphorus concentration values ​​with the three-dimensional geographic coordinate system of the target area.

[0309] Target Feature Map: After passing through all feature extraction layers of the three-dimensional convolutional neural network and undergoing physical damping correction, the multidimensional data tensor output at the bottom layer of the network (i.e., the depth corresponding to the target groundwater level) represents the final pollution distribution state.

[0310] Dimensionless feature values: Pure mathematical feature values ​​generated internally by a neural network that represent only relative magnitude and have no specific physical or chemical units.

[0311] Real chemical units: Standard physicochemical units used in actual environmental engineering to measure the concentration of substances; in this system, they refer to milligrams per liter.

[0312] Transformation scaling parameter matrix: A set of fixed parameters pre-calibrated based on historical groundwater quality sampling data of the target area and the output feature values ​​of the model at corresponding depths, obtained through multiple regression analysis. This matrix is ​​used to linearly map the multi-channel abstract features of the network output to specific concentrations of various chemical pollutants.

[0313] Deep groundwater pollution concentration distribution map: A two-dimensional or three-dimensional visualization layer that combines Geographic Information System (GIS) coordinates with actual chemical concentration values ​​to intuitively show the pollution status of groundwater at a target depth.

[0314] Step 1: Extract the target feature map from the lowest layer output: The system internally sets up loop control logic to sequentially traverse and execute the aforementioned inter-layer correction and propagation steps. When the traversal reaches the last layer of the 3D convolutional neural network (let the total number of layers be...), ... After completing the sub-layer transfer tensor reorganization of this layer, the downward derivation stops. The tensor output from this last layer is directly extracted as the target feature map. The extraction formula is as follows:

[0315] ;

[0316] in: Represents the target feature map in planar grid coordinates And the feature channels are The dimensionless characteristic value at the location. Represents the lowest layer of the network (i.e., the first layer) The final transfer tensor is calculated from the layer. By determining the endpoint boundary of the infiltration simulation, the pollution characteristics reaching the preset groundwater level depth are accurately extracted, while intermediate process data that do not reach that depth are ignored.

[0317] Step 2: Inverse mapping to true chemical dimensions: The system reads the pre-stored conversion ratio parameter matrix from the local database. The elemental eigenvalues ​​in the target feature map are multiplied by this parameter matrix using the channel dimension, and then summed to obtain the true chemical oxygen demand (COD) and nitrogen and phosphorus concentrations. The calculation formula is as follows:

[0318] ;

[0319] in: Represents grid coordinates The corresponding number Similar to real chemical pollutants (such as Represents chemical oxygen demand. The value represents the concentration of total nitrogen, expressed in milligrams per liter. This indicates the total number of feature channels contained in the target feature map. This represents the parameter used in the conversion scaling matrix to convert the first... Data from the first feature channel is mapped to the first... Weighting coefficients for pollutants are typically set between 0.1 and 50.0 based on the background concentration of the pollutants. By utilizing established historical regression relationships, multi-dimensional abstract characteristics are decoded into concentration indicators with clear environmental scientific significance.

[0320] Step 3: Render and generate a deep groundwater pollution concentration distribution map: Use affine transformation to convert the grid coordinates... This is mapped to the actual latitude and longitude coordinates in the three-dimensional geographic coordinate system of the target area. The calculation formula is as follows:

[0321] ; ;

[0322] in: and These represent the actual geographical longitude and latitude after mapping, respectively. and These represent the reference longitude and latitude of the starting point (origin) of the grid in the target area, respectively. This indicates the preset grid resolution (e.g., a single grid represents 5 meters in actual physical space). By performing spatial registration, the concentration values ​​are given a true spatial location. Then, the system uses these coordinates as a reference and renders the data according to the preset concentration color mapping rules (e.g., low concentration is blue, high concentration is red), outputting a deep groundwater pollution concentration distribution map.

[0323] Example 8: Continuing from the previous example, the planar coordinates of the grid nodes in the target area are... Assuming that after multi-layer feature transfer and attenuation correction at all previous depths, the system simulation reaches the lowest level corresponding to the groundwater level (e.g., the 10th level, i.e., ...). ).

[0324] In the extracted target feature map In the middle, coordinates This contains two feature channels ( Their dimensionless characteristic values ​​are as follows:

[0325] Feature values ​​corresponding to channel 1: ;

[0326] Feature values ​​corresponding to channel 2: ;

[0327] The system sets two categories of pollutants to be converted: Chemical Oxygen Demand (COD). ) and total nitrogen ( ).

[0328] Read the pre-stored conversion scaling parameter matrix as follows:

[0329] (Channel weights used to calculate chemical oxygen demand)

[0330] (Channel weights used to calculate total nitrogen)

[0331] Longitude of the reference origin of the target area ,latitude Grid resolution Spend.

[0332] Calculate the actual chemical oxygen demand concentration (COD) ):

[0333] (mg / L);

[0334] Calculate the true total nitrogen concentration ( ):

[0335] (mg / L);

[0336] Calculate the actual geographic coordinates:

[0337] ;

[0338] ;

[0339] Rendering attributes for generating distribution maps:

[0340] After inverse mapping and coordinate registration, the coordinates The generated deep groundwater pollution rendering attributes are shown in the table below. The system will then use this table to represent the coordinates on the map. Render the corresponding concentration color patch at that location.

[0341] In the aforementioned technologies, by quantifying and reconstructing the high-dimensional output of complex neural networks into a standard and intuitive groundwater pollution concentration distribution map, the model output results are given precise stoichiometric meaning and geospatial location, providing intuitive and quantitative engineering data support for tracing the source of agricultural non-point source pollution, risk assessment, and protection of deep groundwater resources.

[0342] A deep learning-based method for dynamic simulation and prediction of agricultural non-point source pollution includes the following steps:

[0343] The original surface pixel matrix and soil basic sampling sequence of the target area are obtained and preprocessed to generate a three-dimensional root damping tensor; the rainfall numerical sequence and fertilizer application numerical sequence are obtained and preprocessed to obtain the initial surface pollution concentration tensor.

[0344] The initial surface pollution concentration tensor is input into a three-dimensional convolutional neural network for infiltration simulation. The three-dimensional convolutional neural network contains multiple feature extraction layers that map different soil physical depths.

[0345] In the inter-layer feature propagation process of a 3D convolutional neural network, the following correction propagation step is performed for each feature extraction layer:

[0346] Perform spatial convolution on the input tensor of the current layer to obtain the initial percolation tensor of the current layer;

[0347] Extract the cut-off threshold corresponding to the current layer depth from the three-dimensional root damping tensor, and determine whether the feature value of each node in the initial infiltration tensor of the current layer is greater than the cut-off threshold;

[0348] If so, extract the physical attenuation coefficient corresponding to the current layer depth from the three-dimensional root damping tensor, and correct the eigenvalue of the node based on the physical attenuation coefficient to obtain the corrected eigenvalue.

[0349] If not, then set the eigenvalue of that node to zero;

[0350] After updating the feature values ​​of all nodes, the data is reorganized into a sub-layer transfer tensor and input into the next feature extraction layer.

[0351] After iterating through all feature extraction layers in a loop, the target feature map of the lowest output layer is extracted and inversely mapped to generate a deep groundwater pollution concentration distribution map.

[0352] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A deep learning-based dynamic simulation and prediction system for agricultural non-point source pollution, characterized in that, include: The data acquisition and preprocessing module is used to acquire the original surface pixel matrix and soil basic sampling sequence of the target area, and to preprocess them to generate a three-dimensional root damping tensor. The numerical sequences of rainfall and fertilizer application were obtained and preprocessed to obtain the initial surface pollution concentration tensor. The network inference input module is used to input the initial surface pollution concentration tensor into a three-dimensional convolutional neural network for infiltration inference inference. The three-dimensional convolutional neural network contains multiple feature extraction layers that map different soil physical depths. The inter-layer feature correction and propagation module is used to perform the following correction and propagation steps for each feature extraction layer during the inter-layer feature propagation process in a 3D convolutional neural network: Perform spatial convolution on the input tensor of the current layer to obtain the initial percolation tensor of the current layer; Extract the cut-off threshold corresponding to the current layer depth from the three-dimensional root damping tensor, and determine whether the feature value of each node in the initial infiltration tensor of the current layer is greater than the cut-off threshold; If so, extract the physical attenuation coefficient corresponding to the current layer depth from the three-dimensional root damping tensor, and correct the eigenvalue of the node based on the physical attenuation coefficient to obtain the corrected eigenvalue. If not, then set the eigenvalue of that node to zero; After updating the feature values ​​of all nodes, the data is reorganized into a sub-layer transfer tensor and input into the next feature extraction layer. The inverse mapping result generation module is used to iteratively execute the correction and propagation step, traverse all feature extraction layers, extract the target feature map output from the bottom layer, and generate a deep groundwater pollution concentration distribution map through inverse mapping.

2. The deep learning-based dynamic simulation and prediction system for agricultural non-point source pollution according to claim 1, characterized in that: The eigenvalues ​​of this node are corrected based on the physical attenuation coefficient to obtain the corrected eigenvalues, which specifically include: The difference between the feature values ​​that are greater than the cutoff threshold and the cutoff threshold is calculated to obtain the preliminary remaining feature values. The saturation penalty factor is dynamically obtained, and the physical attenuation coefficient at the corresponding coordinate position of the node is fused with the saturation penalty factor to obtain the real-time attenuation coefficient. The preliminary remaining characteristic values ​​are corrected based on the real-time attenuation coefficient to obtain the corrected characteristic values ​​that characterize the actual amount of pollutant infiltration through the current soil pores.

3. The deep learning-based dynamic simulation and prediction system for agricultural non-point source pollution according to claim 2, characterized in that: Dynamically obtaining the saturation penalty factor, specifically including: Read the preset cumulative status register to obtain the amount of soil pollution retained by the current node within the historical time step; Extract the cutoff threshold and physical attenuation coefficient corresponding to the current node from the three-dimensional root damping tensor, calculate the ratio of the cutoff threshold to the physical attenuation coefficient, and generate the dynamic saturation sensitivity constant of the current node. The saturation index of the current node is obtained by multiplying the residual soil pollution at the node with the dynamic saturation sensitivity constant. After taking a negative value for the saturation index, a natural exponent mapping operation is performed to generate a dynamic saturation penalty factor whose value is limited to the range of zero and one, and the value in the cumulative state register is updated simultaneously using the modified eigenvalue.

4. The deep learning-based dynamic simulation and prediction system for agricultural non-point source pollution according to claim 1, characterized in that: The generation of the three-dimensional root damping tensor specifically includes: Convert the original surface pixel matrix into a grayscale matrix; By using a preset edge detection operator to perform local gradient calculation on the gray value matrix, the set of gradient coordinates representing the crop canopy boundary is extracted; The pixel nodes within the gradient coordinate set are assigned the first identifier value, and the pixel nodes outside the boundary are assigned the second identifier value to construct a two-dimensional canopy contour tensor. Obtain the vertical downward growth step length of the root system corresponding to the crop variety in the target area. Based on the two-dimensional canopy contour tensor, perform multi-layer tensor replication and layer-by-layer element value reduction calculation along the depth direction according to the vertical downward growth step length. The tensors after multi-level decreasing calculation are stacked in three dimensions to generate a three-dimensional root system occupancy tensor whose elemental values ​​represent the root growth volume within the three-dimensional spatial coordinate points. The soil basic sampling sequence is mapped into a grid according to spatial coordinates to generate a basic porosity matrix; The three-dimensional root system occupancy tensor is fused with the basic porosity matrix to generate the corresponding interception threshold and physical attenuation coefficient, thus constructing the three-dimensional root system damping tensor.

5. The deep learning-based dynamic simulation and prediction system for agricultural non-point source pollution according to claim 4, characterized in that: The generation of the basic porosity matrix includes: Analyze the basic soil sampling sequence and extract the geospatial coordinate data and corresponding measured porosity values ​​of each physical sampling point; The system acquires the three-dimensional geographic boundary information of the target area and the preset grid division resolution, and generates a three-dimensional spatial base grid that matches the target area. The three-dimensional spatial base grid contains multiple discrete grid nodes. Traverse each grid node in the three-dimensional spatial base grid and calculate the spatial Euclidean distance between the current grid node and each physical sampling point; Physical sampling points whose spatial Euclidean distance is less than a preset neighborhood threshold are selected as valid reference points. The reciprocal of the spatial Euclidean distance is used as the distance attenuation weight. The measured porosity values ​​of all valid reference points are weighted and summed to obtain the inferred porosity value of the current grid node. The inferred porosity values ​​of all grid nodes are spliced ​​and assigned values ​​according to the spatial topology array order to generate a basic porosity matrix containing continuous spatial porosity distribution data.

6. The deep learning-based dynamic simulation and prediction system for agricultural non-point source pollution according to claim 1, characterized in that: The initial surface pollution concentration tensor is obtained, specifically including: Both the rainfall and fertilizer application data sequences were resampled to the same target timestamp. The resampled fertilizer application numerical sequence is multiplied with the preset dissolution conversion constant, and the result is multiplied and weighted by the resampled rainfall numerical sequence to generate the initial single-point pollution characteristic value. The initial single-point pollution characteristic values ​​are extended and assigned according to the two-dimensional geospatial grid of the target area to generate the initial surface pollution concentration tensor.

7. The deep learning-based dynamic simulation and prediction system for agricultural non-point source pollution according to claim 1, characterized in that: The three-dimensional convolutional neural network contains multiple feature extraction layers that map to different soil physical depths, and their mapping relationship is established in the following way: Obtain the total physical depth from the surface of the target area to the target groundwater level; The ratio of the total physical depth to the total number of feature extraction layers in the 3D convolutional neural network is calculated to obtain the physical penetration depth step size equivalent for a single-layer network. The product of the interlayer progression number of each feature extraction layer and the physical penetration depth step size is used as the query slice depth corresponding to that layer in the three-dimensional root damping tensor.

8. The deep learning-based dynamic simulation and prediction system for agricultural non-point source pollution according to claim 1, characterized in that: The initial infiltration tensor of the current layer is obtained, specifically including: Obtain the feature extraction tensor kernel containing the three-dimensional diffusion resistance weights of water in the soil; The feature extraction tensor kernel is extracted by sliding across the input tensor of the current layer with a preset step size, and a local matrix dot product summation operation is performed at each sliding pause position; The summation results of all sliding pause positions are reorganized according to their spatial positions to generate the preliminary infiltration tensor of the current layer.

9. The deep learning-based dynamic simulation and prediction system for agricultural non-point source pollution according to claim 1, characterized in that: Generate a deep groundwater pollution concentration distribution map, specifically including: Read the pre-stored conversion ratio parameter matrix from dimensionless characteristic values ​​to actual chemical dimensions; The elemental feature values ​​in the target feature map output at the bottom layer are multiplied by the conversion ratio parameter matrix to convert them into actual chemical oxygen demand and nitrogen and phosphorus concentration values. A deep groundwater pollution concentration distribution map is generated by combining real chemical oxygen demand (COD) and nitrogen and phosphorus concentration values ​​with the three-dimensional geographic coordinate system of the target area.

10. A method for dynamic simulation and prediction of agricultural non-point source pollution based on deep learning, characterized in that, Includes the following steps: The original surface pixel matrix and soil basic sampling sequence of the target area are obtained and preprocessed to generate a three-dimensional root damping tensor. The numerical sequences of rainfall and fertilizer application were obtained and preprocessed to obtain the initial surface pollution concentration tensor. The initial surface pollution concentration tensor is input into a three-dimensional convolutional neural network for infiltration simulation. The three-dimensional convolutional neural network contains multiple feature extraction layers that map different soil physical depths. In the inter-layer feature propagation process of a 3D convolutional neural network, the following correction propagation step is performed for each feature extraction layer: Perform spatial convolution on the input tensor of the current layer to obtain the initial percolation tensor of the current layer; Extract the cut-off threshold corresponding to the current layer depth from the three-dimensional root damping tensor, and determine whether the feature value of each node in the initial infiltration tensor of the current layer is greater than the cut-off threshold; If so, extract the physical attenuation coefficient corresponding to the current layer depth from the three-dimensional root damping tensor, and correct the eigenvalue of the node based on the physical attenuation coefficient to obtain the corrected eigenvalue. If not, then set the eigenvalue of that node to zero; After updating the feature values ​​of all nodes, the data is reorganized into a sub-layer transfer tensor and input into the next feature extraction layer. After iterating through all feature extraction layers in a loop, the target feature map of the lowest output layer is extracted and inversely mapped to generate a deep groundwater pollution concentration distribution map.

Images