Farmland soil micro-plastic dynamic prediction method based on machine learning

Abstract

The invention discloses a farmland soil micro-plastic dynamic prediction method based on machine learning, and relates to the technical field of farmland soil micro-plastic prediction.The farmland soil micro-plastic dynamic prediction method comprises the steps that geographic boundaries, elevations and water system data of a plurality of farmland parcels in a target area are obtained, a farmland space topological graph is constructed by combining space distances, terrain connectivity and tillage mode similarity, and the farmland micro-plastic dynamic prediction result is obtained; land parcels are used as nodes, space and hydrology association is used as edges, and the connection strength of the edges is dynamically adjusted according to meteorological conditions; the method comprises the following steps: collecting soil physicochemical properties, meteorology, agronomic management and farmer behavior data of each farmland plot, and preprocessing to obtain a node feature set; calculating a micro-plastic input load of each farmland plot based on historical management data, evaluating an output flux in combination with soil degradation and migration capabilities, performing accumulative calculation on long-term deviation of input and output, and generating a dynamic pressure index reflecting system pollution pressure; a farmland space topological graph, a node feature set and a dynamic pressure index are used as input.

Description

Technical Field

[0001] This invention relates to the field of farmland soil microplastic prediction technology, and in particular to a dynamic prediction method for farmland soil microplastics based on machine learning. Background Technology

[0002] Farmland soil microplastic prediction technology refers to a technical system that uses mathematical models, remote sensing monitoring, environmental data analysis, and artificial intelligence to simulate and predict the sources, distribution, migration paths, accumulation trends, and future concentration changes of microplastics in farmland ecosystems. Therefore, how to utilize advanced technologies to improve the intelligence level and safety of farmland soil microplastic prediction has become one of the urgent problems to be solved.

[0003] In the field of microplastic prediction in farmland soil, most existing technologies treat each farmland plot as an independent unit and make predictions based solely on local variables. They do not consider the migration of microplastics between adjacent plots due to factors such as terrain slope, surface runoff, and cross-regional operation of agricultural machinery. This results in an inability to accurately simulate the real-world scenario of upstream pollution and downstream harm, and underestimates the risk of cascading pollution at the regional scale. Furthermore, existing methods often directly use the amount of agricultural film used in the current year as an input variable, without assessing the long-term imbalance between the cumulative effect of historical residues and degradation output. They are unable to distinguish between new pollution and old accumulation, and may misjudge whether the system is approaching the ecological threshold. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a machine learning-based method for dynamic prediction of microplastics in farmland soil. This addresses the problem that most models treat each farmland plot as an independent unit, making predictions based solely on local variables. They fail to consider microplastic migration between adjacent plots due to factors such as terrain slope, surface runoff, and cross-regional agricultural machinery operations. This results in an inability to accurately simulate the real-world scenario of upstream pollution and downstream harm, underestimates the risk of cascading pollution at the regional scale, and, in addition, many existing methods directly use the amount of agricultural film used in the current year as an input variable. They fail to assess the long-term imbalance between the cumulative effect of historical residues and degradation output, cannot distinguish between new pollution and old accumulation, and misjudge whether the system is approaching the ecological threshold.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for dynamic prediction of microplastics in farmland soil based on machine learning, comprising:

[0008] The geographic boundaries, elevations, and water system data of multiple farmland plots within the target area are obtained. A farmland spatial topology map is constructed by combining spatial distance, terrain connectivity, and similarity of farming patterns. Plots are used as nodes, spatial and hydrological relationships are used as edges, and the connection strength of the edges is dynamically adjusted according to meteorological conditions.

[0009] Data on soil physicochemical properties, meteorology, agronomic management, and farmer behavior were collected from various farmland plots and preprocessed to obtain node feature sets.

[0010] The microplastic input load of each farmland plot is calculated based on historical management data. The output flux is assessed by combining soil degradation and migration capacity. The long-term deviation between input and output is accumulated and calculated to generate a dynamic pressure index that reflects the pollution pressure of the system.

[0011] Using farmland spatial topology, node feature set, and dynamic pressure index as input, a spatiotemporal graph neural network model integrating physical process constraints is constructed. Spatial dependencies are extracted through graph convolution, and evolutionary patterns are captured by combining temporal modules to obtain a dynamic prediction model for microplastic concentration.

[0012] Current and future environmental and management parameters are input into the trained model, where the dynamic stress index is used to adjust the model’s response to extreme rainfall events, and the microplastic concentration prediction results for each farmland plot at multiple future time points are output.

[0013] Based on the prediction results, a spatiotemporally continuous pollution risk distribution map is generated to identify high-risk areas of concentration increase and their dominant driving periods.

[0014] Based on the driving characteristics of high-risk areas, a pre-set governance rule library is matched to generate agronomic management optimization suggestions and push them to the terminal.

[0015] As a preferred embodiment of the machine learning-based dynamic prediction method for microplastics in farmland soil described in this invention, the following steps are taken: First, the geographical boundaries, elevations, and water system data of multiple farmland plots within the target area are acquired. Then, a farmland spatial topology map is constructed by combining spatial distance, terrain connectivity, and similarity of farming patterns. Plots are used as nodes, and spatial and hydrological associations are used as edges. The connection strength of the edges is dynamically adjusted according to meteorological conditions.

[0016] Obtain vector boundary data for each farmland plot and determine the spatial relative position between each pair of adjacent plots based on geographic coordinates;

[0017] Collect regional digital elevation model data, utilize topographic slope and water accumulation information, analyze surface runoff direction using the D8 algorithm, and determine whether there is a hydrological connectivity path driven by elevation difference between any two plots. If plot A is located in the upstream confluence area of ​​plot B, then there is a one-way hydrological transmission relationship.

[0018] Collect historical agricultural machinery operation trajectories, sowing period and fertilization time records, calculate the time synchronization rate of adjacent plots in key agricultural activities, and determine that there is synergy in farming behavior when the synchronization rate exceeds a preset threshold.

[0019] Treat each farmland plot as a node in the graph structure. If two nodes satisfy any of the following conditions:

[0020] If the straight-line distance in space is less than a set threshold, there is a surface runoff connection, or the synchronization rate of farming behavior is higher than the benchmark value, then an edge is established between the two to form an initial farmland spatial network.

[0021] Each edge is assigned an initial weight, including a geographical proximity factor, a hydrological conduction capacity factor, and a cultivation synergy factor. The geographical proximity factor is inversely proportional to the distance between two nodes, the hydrological conduction capacity factor is determined based on the topographic slope and soil permeability, and the cultivation synergy factor is quantified based on the synchronization rate of agricultural activities.

[0022] Before each prediction task is executed, the rainfall intensity data for the current period is obtained. The rainfall intensity is introduced as an external disturbance factor into the edge weight update mechanism to dynamically adjust the established edges.

[0023] As a preferred embodiment of the machine learning-based dynamic prediction method for microplastics in farmland soil described in this invention, the specific steps for collecting soil physicochemical properties, meteorological data, agronomic management data, and farmer behavior data from various farmland plots, and obtaining node feature sets after preprocessing, are as follows:

[0024] Soil samples were collected from various farmland plots through field sampling, and their pH value, organic matter content, cation exchange capacity, soil texture classification, and moisture content were tested.

[0025] Environmental parameters such as daily average temperature, cumulative precipitation, average wind speed, and sunshine duration are extracted from regional meteorological stations or remote sensing inversion products.

[0026] Collect data on management practices such as the area covered by agricultural mulch film, the total amount of fertilizer applied, the type of irrigation method, and the proportion of straw returned to the field each quarter;

[0027] A structured questionnaire was used to interview farmers to assess their awareness of plastic pollution, their willingness to invest in environmental protection, and their participation in sustainable farming practices. Subjective evaluations were then converted into quantitative scores.

[0028] All collected continuous variables were standardized to eliminate dimensional differences, and categorical variables were converted to numerical values ​​using one-hot coding.

[0029] The processed multi-source data is organized by land parcel to form feature vectors.

[0030] As a preferred embodiment of the machine learning-based dynamic prediction method for microplastics in farmland soil described in this invention, the steps include: calculating the microplastic input load of each farmland plot based on historical management data, assessing the output flux in conjunction with soil degradation and migration capacity, accumulating the long-term deviation between input and output, and generating a dynamic pressure index reflecting the pollution pressure of the system.

[0031] Based on survey data on the years of use, frequency of breakage, amount of agricultural plastics input and residue of plastic film over multiple growing seasons, the amount of microplastics entering the soil system annually for each plot is estimated and recorded as the annual input flux.

[0032] Based on soil microbial activity, ultraviolet radiation intensity, aggregate stability and leaching rate, an empirical model of the natural degradation process of microplastics was established, and the annual average migration output was estimated by combining the surface scouring and subsurface infiltration processes. The sum of the two constitutes the annual output flux.

[0033] Calculate the net input for each year, which is the difference between input flux and output flux. If the difference is positive, it indicates that there is an accumulation of microplastics that have not been absorbed in that year.

[0034] By weighting and accumulating the net input series over the years, and introducing a time decay factor to emphasize the impact of recent loads on the current system state, a dynamic stress index is obtained, expressed as:

[0035] ;

[0036] in, Indicates land parcel In the Annual pollution pressure levels For the first Annual net input This is a time decay factor used to reduce the weight of the impact of early loads on the current state.

[0037] As a preferred embodiment of the machine learning-based dynamic prediction method for microplastics in farmland soil described in this invention, the method involves: constructing a spatiotemporal graph neural network model that integrates physical process constraints using a farmland spatial topology map, node feature set, and dynamic pressure index as inputs; extracting spatial dependencies through graph convolution; and capturing evolutionary patterns using a temporal module to obtain a dynamic prediction model for microplastic concentration. The specific steps are as follows:

[0038] The node feature matrix is ​​input into a multi-layer graph convolutional network. Each layer updates the hidden state of the current node by aggregating the information of neighboring nodes, thereby achieving layer-by-layer abstraction of spatial neighborhood features.

[0039] In graph convolution operations, information propagation is achieved by multiplying the normalized adjacency matrix and the trainable weight matrix, ensuring that neighbors with high connection weights contribute more to the state update of the central node.

[0040] The graph convolution output is fed into the temporal attention mechanism module, which adaptively focuses on the past state that has the greatest influence on the current prediction by calculating the importance weights of features at different time steps.

[0041] By introducing prior knowledge of physical processes, the prediction results are constrained to conform to the basic migration-reaction laws of microplastics in soil media during model training.

[0042] Define a composite loss function, which is a weighted sum of a data fitting term and a physical consistency term. The data fitting term measures the deviation between the model output and the measured concentration, while the physical consistency term measures the residual between the time rate of change of the predicted concentration field and the theoretical evolution equation.

[0043] By minimizing the composite loss function to optimize the model parameters, a dynamic prediction model for microplastic concentration is obtained.

[0044] The update process of the hidden state of nodes in the graph convolutional layer is expressed as follows:

[0045] ;

[0046] in, Represents a node The neighborhood group, For dynamic edge weights, For the first The trainable parameter matrix of the layer, It is a non-linear activation function;

[0047] The calculation process for the physical consistency loss term is expressed as follows:

[0048] ;

[0049] in, For the model in the first Predicted concentration at each spatial sampling point The diffusion coefficient is... Let be the water flow velocity vector. For degradation rate, For source and sink items;

[0050] The process of constructing the composite loss function is expressed as follows:

[0051] ;

[0052] in, For data fitting loss, The weighting coefficients for the physical constraint terms are used to balance the contributions of data-driven and mechanism-guided approaches.

[0053] As a preferred embodiment of the machine learning-based dynamic prediction method for microplastics in farmland soil described in this invention, the steps include: inputting current and future environmental and management parameters into the trained model, wherein the dynamic pressure index is used to adjust the model's response intensity to extreme rainfall events, and outputting the predicted microplastic concentration results for each farmland plot at multiple future time points; the specific steps are as follows:

[0054] The latest observed node features and the updated farmland spatial topology map are input into the trained dynamic prediction model for microplastic concentration.

[0055] The dynamic pressure index is introduced as a regulation signal into the model's response mechanism to construct a nonlinear gated function;

[0056] Under heavy rainfall scenarios, the output of the gating function is used to amplify the edge weights with hydrological connectivity, thereby increasing the microplastic migration flux between high-pressure plots.

[0057] The model synchronously updates the state of all nodes at each time step to simulate the continuous diffusion process of microplastics in a spatial network.

[0058] Output predicted values ​​of microplastic concentration at multiple future time points to form a dynamic evolution sequence with spatiotemporal resolution;

[0059] The dynamic adjustment process of edge weights is expressed as follows:

[0060] ;

[0061] in, The initial edge weights, For dynamic stress index The gain factor of regulation, This is a sensitivity parameter used to adjust the steepness of the gain response curve.

[0062] As a preferred embodiment of the machine learning-based dynamic prediction method for microplastics in farmland soil described in this invention, the specific steps for generating a spatiotemporally continuous pollution risk distribution map based on the prediction results and identifying high-risk areas of increasing concentrations and their dominant driving periods are as follows:

[0063] The predicted concentration values ​​of all plots are mapped back to their geographic coordinates, and a continuous concentration surface is generated using spatial interpolation methods to form a microplastic pollution distribution map at different times.

[0064] Based on the preset risk level classification criteria, areas with concentrations exceeding the set threshold are marked as high-risk areas;

[0065] For each high-risk area, calculate its concentration change rate within the prediction period and screen out areas with a rapidly deteriorating trend;

[0066] By combining attribution analysis techniques, the contribution of each input feature to the predicted concentration is quantified over time.

[0067] By analyzing the peak occurrence time of the contribution sequence, the main driving factors leading to the concentration increase and their initial action time window can be determined, thus identifying the dominant driving period.

[0068] As a preferred embodiment of the machine learning-based dynamic prediction method for microplastics in farmland soil described in this invention, the specific steps of generating agronomic management optimization suggestions and pushing them to the terminal based on the driving characteristics of high-risk areas and matching them with a preset governance rule library are as follows:

[0069] Establish a governance strategy rule base, which stores the logical mapping relationship between different pollution driving types and corresponding intervention measures;

[0070] When the main driver of a region is identified as the accumulation of residual plastic film, strategies such as promoting biodegradable plastic film or strengthening residual film recycling should be matched accordingly.

[0071] When the primary driver is irrigation erosion leading to intensified migration, strategies such as matching and optimizing irrigation methods or constructing ecological interception zones should be adopted.

[0072] When farmers’ non-standard management practices are a key factor, strategies such as organizing environmental protection training or providing green subsidies should be adopted.

[0073] The matched governance measures are transformed into structured recommendation texts, including the measure name, implementation suggestions, expected effects, and operational guidelines.

[0074] The recommendations will be pushed to the linked mobile terminal, village-level management platform, or agricultural technology service system via wireless communication network.

[0075] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the machine learning-based dynamic prediction method for microplastics in farmland soil as described in the first aspect of the present invention.

[0076] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the machine learning-based dynamic prediction method for microplastics in farmland soil as described in the first aspect of the present invention.

[0077] The beneficial effects of this invention are as follows: By constructing a spatiotemporal graph neural network model that integrates spatial topological relationships and physical process constraints, a dynamic pressure index is introduced to quantify the cumulative load of soil microplastics, and the model's response intensity under extreme weather conditions is adjusted using this index. This enables high-precision dynamic prediction of the spatiotemporal evolution of microplastic concentration in farmland soil. Combined with attribution analysis and rule base matching mechanisms, actionable agronomic management optimization suggestions are generated, forming a closed loop of perception-prediction-decision. This improves the intelligence and precision of agricultural non-point source pollution control and provides technical support for risk warning and targeted prevention and control of microplastic pollution in farmland. Attached Figure Description

[0078] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0079] Figure 1 This is a flowchart of the machine learning-based dynamic prediction method for microplastics in farmland soil in Example 1.

[0080] Figure 2 This is a flowchart of the dominant driving period in Example 1. Detailed Implementation

[0081] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0082] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0083] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0084] Example, refer to Figure 1 and Figure 2 This embodiment of the invention provides a machine learning-based method for dynamically predicting microplastics in farmland soil, comprising the following steps:

[0085] S1. Obtain the geographical boundaries, elevations, and water system data of multiple farmland plots within the target area. Combine spatial distance, terrain connectivity, and similarity of farming patterns to construct a farmland spatial topology map. Use plots as nodes and spatial and hydrological associations as edges, and dynamically adjust the connection strength of the edges according to meteorological conditions.

[0086] Furthermore, the vector boundary data of each farmland plot is obtained, and the spatial relative position between each two adjacent plots is determined based on geographic coordinates;

[0087] Collect regional digital elevation model data, utilize topographic slope and water accumulation information, analyze surface runoff direction using the D8 algorithm, and determine whether there is a hydrological connectivity path driven by elevation difference between any two plots. If plot A is located in the upstream confluence area of ​​plot B, then there is a one-way hydrological transmission relationship.

[0088] Collect historical agricultural machinery operation trajectories, sowing period and fertilization time records, calculate the time synchronization rate of adjacent plots in key agricultural activities, and determine that there is synergy in farming behavior when the synchronization rate exceeds a preset threshold.

[0089] Treat each farmland plot as a node in the graph structure. If two nodes satisfy any of the following conditions:

[0090] If the straight-line distance in space is less than a set threshold, there is a surface runoff connection, or the synchronization rate of farming behavior is higher than the benchmark value, then an edge is established between the two to form an initial farmland spatial network.

[0091] Each edge is assigned an initial weight, including a geographical proximity factor, a hydrological conduction capacity factor, and a cultivation synergy factor. The geographical proximity factor is inversely proportional to the distance between two nodes, the hydrological conduction capacity factor is determined based on the topographic slope and soil permeability, and the cultivation synergy factor is quantified based on the synchronization rate of agricultural activities.

[0092] Before each prediction task is executed, the rainfall intensity data for the current period is obtained, and the rainfall intensity is introduced as an external disturbance factor into the edge weight update mechanism to dynamically adjust the established edges.

[0093] It should be noted that by constructing a farmland spatial topology map by integrating the triple relationships of geospatial, hydrological connectivity, and synergy of farming behavior, the limitations of the traditional static adjacency assumption are overcome, and a refined modeling of potential microplastic migration paths is achieved. The dynamic edge weight mechanism introduces real-time rainfall intensity as an external perturbation factor, which enables the network connectivity strength to be adaptively adjusted according to meteorological conditions, thereby improving the model's ability to characterize the microplastic migration process driven by surface runoff under extreme weather conditions and enhancing the real-world responsiveness and spatiotemporal sensitivity of the prediction system.

[0094] S2. Collect soil physicochemical properties, meteorological data, agronomic management data, and farmer behavior data for each farmland plot. After preprocessing, obtain the node feature set.

[0095] Furthermore, soil samples were obtained from various farmland plots through on-site sampling to test their pH value, organic matter content, cation exchange capacity, soil texture classification, and moisture content.

[0096] Environmental parameters such as daily average temperature, cumulative precipitation, average wind speed, and sunshine duration are extracted from regional meteorological stations or remote sensing inversion products.

[0097] Collect data on management practices such as the area covered by agricultural mulch film, the total amount of fertilizer applied, the type of irrigation method, and the proportion of straw returned to the field each quarter;

[0098] A structured questionnaire was used to interview farmers to assess their awareness of plastic pollution, their willingness to invest in environmental protection, and their participation in sustainable farming practices. Subjective evaluations were then converted into quantitative scores.

[0099] All collected continuous variables were standardized to eliminate dimensional differences, and categorical variables were converted to numerical values ​​using one-hot coding.

[0100] The processed multi-source data is organized by land parcel to form feature vectors;

[0101] It should be noted that by constructing a multi-source heterogeneous data fusion framework, the system integrates multi-level characteristics such as soil physicochemical properties, meteorological environment, agronomic management, and farmer behavior, comprehensively reflecting the key driving factors affecting microplastic accumulation and migration. Through standardization of continuous variables and one-hot encoding of categorical variables, the system effectively eliminates the problems of dimensional differences and inconsistent numerical types, ensuring the comparability and compatibility of input features in machine learning models, and providing a high-quality, structurally unified data foundation for subsequent high-dimensional nonlinear relationship mining.

[0102] S3. Calculate the microplastic input load of each farmland plot based on historical management data, assess the output flux in combination with soil degradation and migration capacity, accumulate the long-term deviation between input and output, and generate a dynamic pressure index that reflects the pollution pressure of the system.

[0103] Furthermore, based on survey data on the years of use, frequency of breakage, amount of agricultural plastic input and residue of plastic film over multiple growing seasons, the annual amount of microplastics entering the soil system for each plot is estimated and recorded as the annual input flux.

[0104] Based on soil microbial activity, ultraviolet radiation intensity, aggregate stability and leaching rate, an empirical model of the natural degradation process of microplastics was established, and the annual average migration output was estimated by combining the surface scouring and subsurface infiltration processes. The sum of the two constitutes the annual output flux.

[0105] Calculate the net input for each year, which is the difference between input flux and output flux. If the difference is positive, it indicates that there is an accumulation of microplastics that have not been absorbed in that year.

[0106] By weighting and accumulating the net input series over the years, and introducing a time decay factor to emphasize the impact of recent loads on the current system state, a dynamic stress index is obtained, expressed as:

[0107] ;

[0108] in, Indicates land parcel In the Annual pollution pressure levels For the first Annual net input This is a time decay factor used to reduce the weight of the impact of early loads on the current state;

[0109] It should be noted that the dynamic pressure index not only quantifies the long-term cumulative effect of net microplastic input, but also assigns higher weight to recent loads through a time decay factor, more accurately reflecting the current pollution carrying capacity of the soil system. The index simplifies the complex input-output imbalance process into a calculable comprehensive indicator, which retains the inertial characteristics of historical accumulation and enhances the sensitivity to recent management changes, providing key state adjustment signals for prediction models.

[0110] S4. Using farmland spatial topology map, node feature set and dynamic pressure index as input, construct a spatiotemporal graph neural network model that integrates physical process constraints. Extract spatial dependence through graph convolution and capture evolution law by combining temporal module to obtain a dynamic prediction model of microplastic concentration.

[0111] Furthermore, the node feature matrix is ​​input into a multi-layer graph convolutional network. Each layer updates the hidden state of the current node by aggregating information from neighboring nodes, thereby achieving a layer-by-layer abstraction of spatial neighborhood features.

[0112] In graph convolution operations, information propagation is achieved by multiplying the normalized adjacency matrix and the trainable weight matrix, ensuring that neighbors with high connection weights contribute more to the state update of the central node.

[0113] The graph convolution output is fed into the temporal attention mechanism module, which adaptively focuses on the past state that has the greatest influence on the current prediction by calculating the importance weights of features at different time steps.

[0114] By introducing prior knowledge of physical processes, the prediction results are constrained to conform to the basic migration-reaction laws of microplastics in soil media during model training.

[0115] Define a composite loss function, which is a weighted sum of a data fitting term and a physical consistency term. The data fitting term measures the deviation between the model output and the measured concentration, while the physical consistency term measures the residual between the time rate of change of the predicted concentration field and the theoretical evolution equation.

[0116] By minimizing the composite loss function to optimize the model parameters, a dynamic prediction model for microplastic concentration is obtained.

[0117] The update process of the hidden state of nodes in the graph convolutional layer is expressed as follows:

[0118] ;

[0119] in, Represents a node The neighborhood group, For dynamic edge weights, For the first The trainable parameter matrix of the layer, It is a non-linear activation function;

[0120] The calculation process for the physical consistency loss term is expressed as follows:

[0121] ;

[0122] in, For the model in the first Predicted concentration at each spatial sampling point The diffusion coefficient is... Let be the water flow velocity vector. For degradation rate, For source and sink items;

[0123] The process of constructing the composite loss function is expressed as follows:

[0124] ;

[0125] in, For data fitting loss, These are the weighting coefficients for the physical constraint terms, used to balance the contributions of data-driven and mechanism-guided approaches;

[0126] It should be noted that by integrating a spatiotemporal graph neural network with physical constraints, the prior mechanism of microplastic migration-degradation is embedded while data-driven learning is carried out, which effectively prevents the model prediction results from violating the basic laws of environmental processes. The design of the composite loss function enables the model to minimize the residual of the concentration field evolution dynamic equation while fitting the observed data, thereby improving the extrapolation reliability and physical interpretability of the model in sparse data regions and solving the problem of weak generalization ability of pure black box models.

[0127] S5. Input the current and future environmental and management parameters into the trained model, where the dynamic stress index is used to adjust the model's response intensity to extreme rainfall events, and output the microplastic concentration prediction results for each farmland plot at multiple future time points.

[0128] Furthermore, the latest observed node features and the updated farmland spatial topology map are input into the trained dynamic prediction model for microplastic concentration.

[0129] The dynamic pressure index is introduced as a regulation signal into the model's response mechanism to construct a nonlinear gated function;

[0130] Under heavy rainfall scenarios, the output of the gating function is used to amplify the edge weights with hydrological connectivity, thereby increasing the microplastic migration flux between high-pressure plots.

[0131] The model synchronously updates the state of all nodes at each time step to simulate the continuous diffusion process of microplastics in a spatial network.

[0132] Output predicted values ​​of microplastic concentration at multiple future time points to form a dynamic evolution sequence with spatiotemporal resolution;

[0133] The dynamic adjustment process of edge weights is expressed as follows:

[0134] ;

[0135] in, The initial edge weights, For dynamic stress index The gain factor of regulation, This is a sensitivity parameter used to adjust the steepness of the gain response curve;

[0136] It should be noted that by using the dynamic pressure index as a gating gain factor to regulate the side weights, a stronger migration response of high-pollution-risk sites in extreme rainfall events was achieved, reflecting the ecological principle that the more fragile the system, the more sensitive it is to external disturbances. This mechanism endows the model with a nonlinear amplification effect, enabling it to capture the risk transition behavior of rapid microplastic diffusion under critical conditions, thereby improving the early warning capability for sudden pollution transmission processes.

[0137] S6. Generate a spatiotemporally continuous pollution risk distribution map based on the prediction results, and identify high-risk areas of concentration increase and their dominant driving periods;

[0138] Furthermore, the predicted concentration values ​​of all plots are mapped back to their geographic coordinates, and a continuous concentration surface is generated using spatial interpolation methods to form a microplastic pollution distribution map at different times.

[0139] Based on the preset risk level classification criteria, areas with concentrations exceeding the set threshold are marked as high-risk areas;

[0140] For each high-risk area, calculate its concentration change rate within the prediction period and screen out areas with a rapidly deteriorating trend;

[0141] By combining attribution analysis techniques, the contribution of each input feature to the predicted concentration is quantified over time.

[0142] By analyzing the peak occurrence time of the contribution sequence, the main driving factors leading to the increase in concentration and their initial time window can be determined, and the dominant driving period can be identified.

[0143] It should be noted that by combining spatial interpolation and attribution analysis, the prediction results of discrete plots are scaled up to a continuous spatiotemporal risk distribution map, which intuitively presents the evolution trend of pollution hotspots. The identification of the dominant driving period not only locates high-risk areas, but also reveals the window of action of key anthropogenic or natural driving factors behind them, providing a scientific basis for formulating precise time-series intervention strategies and achieving a deeper understanding from knowing what to know why.

[0144] S7. Match the preset governance rule base with the driving characteristics of high-risk areas, generate agronomic management optimization suggestions and push them to the terminal;

[0145] Furthermore, a governance strategy rule base should be established, which stores the logical mapping relationship between different pollution driving types and corresponding intervention measures;

[0146] When the main driver of a region is identified as the accumulation of residual plastic film, strategies such as promoting biodegradable plastic film or strengthening residual film recycling should be matched accordingly.

[0147] When the primary driver is irrigation erosion leading to intensified migration, strategies such as matching and optimizing irrigation methods or constructing ecological interception zones should be adopted.

[0148] When farmers’ non-standard management practices are a key factor, strategies such as organizing environmental protection training or providing green subsidies should be adopted.

[0149] The matched governance measures are transformed into structured recommendation texts, including the measure name, implementation suggestions, expected effects, and operational guidelines.

[0150] The recommendations will be pushed to the linked mobile terminal, village-level management platform or agricultural technology service system via wireless communication network;

[0151] It should be noted that the governance rule base established in this step realizes the intelligent transformation from predictive output to management decision-making, forming a closed-loop smart agriculture governance chain. By transforming the results of technical analysis into structured suggestions that farmers and managers can understand and implement, and by using a digital platform to achieve precise delivery, it effectively bridges the gap between scientific research and field practice, and improves the implementation and operability of agricultural non-point source pollution prevention and control.

[0152] This embodiment also provides a computer device applicable to the dynamic prediction method of microplastics in farmland soil based on machine learning, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the dynamic prediction method of microplastics in farmland soil based on machine learning as proposed in the above embodiment.

[0153] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0154] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the machine learning-based dynamic prediction method for microplastics in farmland soil as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0155] In summary, this invention constructs a spatiotemporal graph neural network model that integrates spatial topological relationships and physical process constraints. It introduces a dynamic pressure index to quantify the cumulative load of soil microplastics and utilizes this index to adjust the model's response intensity under extreme weather conditions. This enables high-precision dynamic prediction of the spatiotemporal evolution of microplastic concentration in farmland soil. Combined with attribution analysis and rule base matching mechanisms, it generates actionable agronomic management optimization suggestions, forming a closed loop of perception-prediction-decision. This improves the intelligence and precision of agricultural non-point source pollution control and provides technical support for risk warning and targeted prevention and control of microplastic pollution in farmland.

[0156] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for dynamic prediction of microplastics in farmland soil based on machine learning, characterized in that: The method comprises the following steps: Obtain the geographical boundaries, elevations, and water system data of multiple farmland plots in the target area, and construct a farmland spatial topology graph by combining spatial distance, terrain connectivity, and similarity of tillage patterns, taking plots as nodes and spatial and hydrological relationships as edges, and dynamically adjusting the connection strength of the edges according to meteorological conditions; Collect soil physical and chemical properties, meteorological data, agronomic management data, and farmer behavior data of each farmland plot, and obtain a node feature set after preprocessing; Calculate the microplastic input load of each farmland plot based on historical management data, evaluate the output flux by combining soil degradation and migration ability, and calculate the long-term deviation between input and output to generate a dynamic pressure index reflecting the pollution pressure of the system; Input the farmland spatial topology graph, node feature set, and dynamic pressure index to construct a spatio-temporal graph neural network model that integrates physical process constraints, extract spatial dependencies through graph convolution, and capture evolution rules through a time sequence module to obtain a dynamic prediction model of microplastic concentration; Input the current and future environmental and management parameters into the trained model, where the dynamic pressure index is used to adjust the response strength of the model to extreme rainfall events, and output the prediction results of microplastic concentration at multiple time points in the future for each farmland plot; Generate a spatio-temporal continuous pollution risk distribution map based on the prediction results, identify high-risk areas with rising concentrations and their dominant driving periods; Match the driving characteristics of the high-risk areas with a pre-set rule library to generate agronomic management optimization suggestions and push them to the terminal.

2. The machine learning based dynamic prediction method of microplastics in farmland soil according to claim 1, characterized in that: The method comprises the following steps: Obtain the vector boundary data of each farmland plot, and determine the spatial relative position between each two adjacent plots based on geographic coordinates; Collect regional digital elevation model data, use terrain slope and water flow accumulation information, and analyze the surface runoff direction through the D8 algorithm to determine whether there is a hydrological connection path driven by elevation difference between any two plots. If plot A is located in the upstream catchment area of plot B, there is a one-way hydrological conduction relationship; Collect historical agricultural operation trajectories, sowing period, and fertilizer application time records, and calculate the time synchronization rate of adjacent plots during key agricultural activities. When the synchronization rate exceeds the pre-set threshold, it is determined that there is tillage behavior coordination; Treat each farmland plot as a node in the graph structure. If two nodes meet any of the following conditions: The spatial straight-line distance is less than a set threshold, there is a surface runoff connection relationship, or the tillage behavior synchronization rate is higher than a reference value, then a edge is established between the two nodes to form an initial farmland spatial network; Assign an initial weight to each edge, including a geographical proximity factor, a hydrological conduction ability factor, and a tillage coordination degree factor. The geographical proximity factor is inversely proportional to the distance between the two nodes, the hydrological conduction ability factor is determined according to the terrain slope and soil permeability, and the tillage coordination degree factor is quantified based on the synchronization rate of agricultural activities. Before each execution of the prediction task, the rainfall intensity data of the current period is obtained, and the rainfall intensity is introduced as an external disturbance factor into the edge weight updating mechanism to dynamically adjust the established edges.

3. The machine learning based dynamic prediction method of microplastics in farmland soil according to claim 2, characterized in that: The soil physical and chemical properties, meteorology, agronomic management, and farmer behavior data of each farmland plot are collected, and after preprocessing, a node feature set is obtained. The specific steps are as follows: Through field sampling, the soil samples of each farmland plot are obtained, and the pH value, organic matter content, cation exchange capacity, soil texture classification, and water content are detected. The daily average temperature, cumulative precipitation, average wind speed, and sunshine duration environmental parameters are extracted from regional meteorological stations or remote sensing inversion products. The management behavior data of each season, such as plastic mulch coverage area, total amount of chemical fertilizer application, irrigation method type, and straw return ratio, are statistically analyzed. A structured questionnaire is used to interview farmers to assess their cognitive level of plastic pollution, environmental protection investment willingness, and sustainable farming practice participation, and to convert subjective evaluation into quantitative scores. All collected continuous variables are standardized to eliminate dimensional differences, and categorical variables are numerically converted using one-hot encoding. The processed multi-source data is organized by plot to form a feature vector.

4. The machine learning based dynamic prediction method of microplastics in farmland soil according to claim 3, characterized in that: The historical management data is used to calculate the microplastic input load of each farmland plot, and the soil degradation and migration ability is used to evaluate the output flux. The long-term deviation between input and output is calculated to generate a dynamic pressure index reflecting the system pollution pressure. The specific steps are as follows: Based on the mulch usage life, damage frequency, agricultural plastic input, and residual amount survey data of multiple growth seasons in the past, the mass of microplastics entering the soil system of each plot per year is estimated, denoted as the annual input flux. Based on soil microbial activity, ultraviolet radiation intensity, aggregate stability, and leaching rate, an empirical model of the natural degradation process of microplastics is established, and the annual average migration output is estimated based on surface erosion and underground infiltration processes. The sum of the two constitutes the annual output flux. The net input of each year, which is the difference between the input flux and the output flux, is calculated. If the difference is positive, it indicates that there is an accumulation of microplastics that has not been consumed in the current year. The sequence of net input over the years is weighted and accumulated, and a time decay factor is introduced to emphasize the impact of recent loads on the current system state to obtain the dynamic pressure index, which is expressed as: ; wherein, represents a plot In the first year of pollution pressure level, is the net input for the year, is a time decay factor to reduce the weight of the influence of early loads on the current state.

5. The machine learning based dynamic prediction method of microplastics in farmland soil according to claim 4, characterized in that: The spatial topology graph of farmland, node feature set, and dynamic pressure index are used as inputs to construct a spatio-temporal graph neural network model that integrates physical process constraints. Through graph convolution, spatial dependence is extracted, and temporal modules are used to capture evolution rules to obtain a dynamic prediction model of microplastic concentration. The specific steps are as follows: The node feature matrix is input into a multi-layer graph convolution network. Each layer updates the hidden state of the current node by aggregating the information of neighboring nodes, achieving layer-by-layer abstraction of spatial neighborhood features. In the graph convolution operation, the product of the normalized adjacency matrix and the trainable weight matrix is used to realize information propagation, ensuring that neighbors with high connection weights have a greater contribution to the center node state update. The graph convolution output is sent to the temporal attention mechanism module, which calculates the importance weight of features at different time steps to adaptively focus on the past states that have the most influence on the current prediction. The physical process prior knowledge is introduced to constrain the prediction results to comply with the basic migration-reaction rules of microplastics in soil media during the model training process; A composite loss function is defined, which is composed of a data fitting term and a physical consistency term, wherein the data fitting term measures the deviation between the model output and the measured concentration, and the physical consistency term measures the residual between the time rate of change of the predicted concentration field and the theoretical evolution equation; The model parameters are optimized by minimizing the composite loss function to obtain the dynamic prediction model of microplastic concentration; The updating process of the node hidden state in the graph convolution layer is expressed as: ; wherein, represents a set of neighbors of a node , is a dynamic edge weight, is a trainable parameter matrix of the layer, is a non-linear activation function; The calculation process of the physical consistency loss term is expressed as: ; wherein, is the predicted concentration of the model at the th spatial sampling point, is the diffusion coefficient, is the water flow velocity vector, is the degradation rate, is the source / sink term; The construction process of the composite loss function is expressed as: ； wherein, is a data fitting loss, is a weighting coefficient for the physical constraint term, balancing the contributions of data-driven and mechanism-guided.

6. The machine learning based dynamic prediction method of microplastics in farmland soil according to claim 5, characterized in that: The current and future environment and management parameters are input into the trained model, wherein the dynamic stress index is used to adjust the response intensity of the model to extreme rainfall events, and the predicted results of the microplastic concentration at multiple time points in the future are output for each farmland plot, and the specific steps are as follows: The latest observed node features and the updated farmland spatial topology graph are input into the trained microplastic concentration dynamic prediction model; The dynamic stress index is introduced as an adjustment signal into the response mechanism of the model to construct a nonlinear gating function; In the scenario of heavy rainfall, the output of the gating function is used to amplify the edge weight with hydrological connectivity, so that the microplastic migration flux between high stress plots is improved; The model updates the state of all nodes at each time step to simulate the continuous diffusion process of microplastics in the spatial network; The predicted values of microplastic concentration at multiple time points in the future are output to form a dynamic evolution sequence with spatial and temporal resolution; The dynamic adjustment process of the edge weight is expressed as: ; wherein, is an initial edge weight, is a gain factor modulated by a dynamic pressure index is a gain factor modulated by a dynamic pressure index is a sensitivity parameter used to adjust the steepness of the gain response curve.

7. The machine learning based dynamic prediction method of microplastics in farmland soil according to claim 6, characterized in that: Based on the predicted results, a spatiotemporal continuous pollution risk distribution map is generated to identify high-risk areas with rising concentrations and their dominant driving time periods, and the specific steps are as follows: The predicted concentration values of all plots are mapped back to their geographic coordinate positions, and a continuous concentration surface is generated using a spatial interpolation method to form a microplastic pollution distribution map at different times; According to the preset risk level division standard, the areas with concentrations higher than the set threshold are marked as high-risk areas; For each high-risk area, the concentration change rate within the prediction period is calculated to screen out rapidly deteriorating areas with an increasing trend; Combined with the attribution analysis technology, the contribution of each input feature to the predicted concentration is quantified over time; By analyzing the peak value occurrence time of the contribution sequence, the main driving factors causing the concentration rise and their initial action time window are determined to identify the dominant driving time period.

8. The machine learning based dynamic prediction method of microplastics in farmland soil according to claim 7, characterized in that: According to the driving characteristics of the high-risk areas, a preset management rule library is matched to generate agronomic management optimization suggestions and push them to the terminal, and the specific steps are as follows: A management strategy rule library is established, which stores the logical mapping relationship between different pollution driving types and corresponding intervention measures; When the main driving of a certain area is the accumulation of mulch residues, the promotion of degradable mulch or the strengthening of residue recycling strategy is matched; When the main driving is the intensification of migration caused by irrigation scouring, the optimization of irrigation mode or the construction of ecological interception belt strategy is matched; When the non-standard management behavior of farmers is the key factor, the organization of environmental protection training or the provision of green subsidies strategy is matched; Convert the matched governance measures into structured suggestion text, including measure name, implementation suggestion, expected effect description, and operation guide; Push the suggestions to the bound mobile terminal, village-level management platform, or agricultural technology service system through the wireless communication network. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is characterized in that: The processor executes the computer program to realize the steps of the machine learning-based farmland soil microplastic dynamic prediction method of any one of claims 1-8.

10. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to realize the steps of the machine learning-based farmland soil microplastic dynamic prediction method of any one of claims 1-8.

Images