Soil pollution risk assessment method and device based on machine learning

Abstract

The invention discloses a machine learning-based soil pollution risk assessment method and a machine learning-based soil pollution risk assessment device, and belongs to the technical field of soil pollution risk assessment. The split biological effect and chemical indexes in traditional evaluation are converted into unified risk driving force measurement. A nonlinear interaction entropy calculation mechanism can adaptively capture an antagonism-collaborative balance relationship between microbial communities and pollutants in a soil environment, so that the degradation potential of functional flora on specific pollutants is reflected, and the dynamic regulation effect of environmental factor fluctuation on bioavailability is quantified. According to the cross-scale feature fusion technology, the evaluation model not only can identify the current pollution space distribution, but also can pre-judge the morphological transformation trend of pollutants in biogeochemical circulation, and provides a quantitative basis with ecological interpretation for risk early warning.

Description

Technical Field

[0001] The present invention belongs to the technical field of soil pollution risk assessment, and particularly relates to a soil pollution risk assessment method and device based on machine learning. Background Art

[0002] Soil pollution risk assessment refers to the process of evaluating the potential risks of soil pollution to the environment and human health through the investigation, detection, and analysis of polluted soil, providing a scientific basis for soil pollution control and risk management. This assessment includes identifying pollution sources, determining pollutant migration and transformation pathways, quantifying pollutant concentrations and exposure levels, and predicting the toxicity and risks of pollutants. By constructing a risk assessment model, the carcinogenic and non-carcinogenic risks of soil pollution are evaluated, and corresponding remediation goals and measures are formulated. Soil pollution risk assessment aims to protect the quality of the soil environment, prevent soil pollution accidents, and ensure the health of the people and ecological security.

[0003] However, the existing technologies mainly rely on static pollutant concentration threshold determination methods, unable to quantify the dynamic impact of microbial activities on pollution transformation, and traditional feature engineering often severs the internal connection between biological indicators and chemical indicators, resulting in the lack of precise regulation ability for pollutant migration and transformation paths in the remediation plan, facing the dual dilemmas of insufficient ecological mechanism analysis and disconnection from engineering practice under complex site conditions. Summary of the Invention

[0004] The purpose of the present invention is to provide a soil pollution risk assessment method and device based on machine learning to solve the above-mentioned problems.

[0005] The technical solution adopted by the present invention is as follows: A soil pollution risk assessment method based on machine learning, which includes the following steps: S1: Obtain surface spectral features through satellite remote sensing technology, combine with a ground sensor network to monitor soil humidity and acidity in real time, and synchronously integrate the historical pollution event database and laboratory physical and chemical analysis reports to construct a cross-scale spatio-temporal dataset; S2: Use an adaptive threshold algorithm to detect abnormal sensor data, and introduce a soil microbial community metabolism model to simulate missing values, and eliminate the resolution differences of multi-source data through spatio-temporal alignment technology; S3: Design an interpretable feature generator to dynamically couple soil microbial activity indicators with environmental factors to generate time-varying features reflecting the pollutant migration and transformation rate; S4: Extract spatial patterns using a deep residual network in pollution hotspots, enable a pre-trained transfer model for data-scarce areas, and use a generative adversarial network to synthesize training samples that conform to local features; S5: Integrate the convolutional neural network and the graph attention mechanism to capture the spatial diffusion characteristics of pollutants and the topological associations of the soil - groundwater system respectively, and output a heat map of the pollution risk probability distribution; S6: Combine geographic information system and augmented reality technology to map the risk assessment results into a three - dimensional hierarchical risk map, and visually display the migration path of pollutants at the soil - plant - atmosphere interface; S7: Access meteorological forecast data and land use change models to dynamically predict the evolution trend of pollution risks within the next three months and generate early warning signals for high - risk areas; S8: Deploy an online learning module to regularly receive on - site sampling verification data, automatically adjust the model hyperparameters by comparing the KL divergence between the predicted and measured results, and then construct a multi - objective optimization tree based on the risk tracing results to provide customized treatment plans for different pollution scenarios.

[0006] In a preferred embodiment, in step S1, satellite remote sensing technology captures the surface reflectance difference with hyperspectral imaging to identify the spatial distribution pattern of potential pollution areas; the ground sensor network deploys electrochemical sensors and ion - selective electrodes to obtain parameters such as soil conductivity and redox potential in real - time, forming vertical profile data with a minute - level monitoring frequency. The historical pollution event database matches with laboratory physical and chemical data through spatio - temporal coding, and finally generates a three - dimensional data cube containing spectral features, physical and chemical parameters, and pollution event labels, providing a multi - dimensional benchmark for subsequent analysis.

[0007] In a preferred embodiment, in step S2, the adaptive threshold algorithm dynamically adjusts the anomaly determination boundary according to soil type. For example, a more relaxed conductivity fluctuation threshold is set for clayey soil; the microbial metabolism model simulates the conversion efficiency of nitrifying bacteria on ammonia nitrogen, and when nitrogen content data is missing from the sensor, the original pollution load is inferred by detecting the nitrite concentration. The spatio - temporal alignment technology uses wavelet transform to unify the kilometer - level resolution of satellite data and the meter - level monitoring accuracy of ground sensors, eliminating the interference of scale differences on the simulation of pollutant diffusion.

[0008] In a preferred embodiment, in step S3, it specifically includes: S3 - 1: Data coupling layer: Integrate soil microbial activity indicators (such as the abundance of specific bacterial groups, the intensity of enzyme activity) and environmental factors (heavy metal ion concentration, temperature gradient, organic matter content) to establish a time - series coupling matrix, retaining the dynamic association between pollutants and ecological responses.

[0009] S3 - 2: Feature generator architecture: Adopt a dual - channel attention mechanism to synchronously analyze the microbial community metabolism rate and environmental factor fluctuations in the time dimension, and generate a candidate feature pool through cross - channel feature intersection.

[0010] S3-3: Nonlinear interaction entropy calculation: The creative formula is as follows: H_t = Σ(P_m(i,t)·log(Q_e(j,t) / (P_m(i,t)^α + Q_e(j,t)^β))) Where: H_t represents the bioremediation potential index at time t. The larger the value, the stronger the conversion ability of microorganisms to pollutants; P_m(i,t) is the normalized abundance of the i-th type of functional microorganism at time t (normalized value in the range of 0-1); Q_e(j,t) is the equivalent concentration of the j-th pollutant at time t (dimensionless value after toxicity weighting); α is the microbial metabolic decay coefficient (calibrated through laboratory culture experiments, typical value 0.6-1.2); β is the pollutant bioavailability regulation factor (calculated dynamically depending on soil pH and organic matter content); S3-4: Feature optimization mechanism: Explain the model through the Shapley value, and select features with a contribution degree to risk assessment greater than 15% to enter the downstream model, ensuring that the generated features have both physical meaning and prediction efficacy.

[0011] In a preferred embodiment, in step S4, in the industrial legacy pollution area, the deep residual network extracts the spatial co-occurrence patterns of heavy metals such as chromium and cadmium through skip connections, and uses dilated convolution to expand the feature receptive field to identify kilometer-scale pollution plumes; for the problem of scarce data in newly developed areas, the pre-trained model uses the migration law of agricultural land pollutants as prior knowledge, and then synthesizes virtual samples that conform to the local soil texture through the generative adversarial network, enabling the model to still capture the vertical infiltration characteristics of organic pollutants such as benzene series under limited data.

[0012] In a preferred embodiment, in step S5, the three-dimensional convolutional neural network extracts the diffusion speed and direction features of pollutants in the soil profile, and enhances the deep soil pollution identification ability through the dilated convolution layer; the graph attention mechanism constructs a topological network of soil pores - groundwater channels, calculates the migration weights of pollutants along the preferred path, and finally tensors and splices the two types of features at the feature layer to output a risk probability heat map with a resolution of 10 m × 10 m, accurately marking the high-risk areas where the leachate pollution boundary overlaps with the underground water source protection area.

[0013] In a preferred embodiment, in step S6, the Geographic Information System (GIS) overlays the two-dimensional risk heat map with the Digital Elevation Model (DEM) to generate an animation simulating the diffusion of the pollution plume along the elevation gradient. The augmented reality device, through spatial registration technology, superimposes and displays the enrichment concentration of pollutants in crop roots during on-site inspections, and uses a particle tracking algorithm to dynamically render the path of volatile organic compounds escaping from soil pores to the atmosphere, enabling decision-makers to intuitively grasp the multi-media pollution coupling effect.

[0014] In a preferred embodiment, in step S7, rainfall intensity and wind direction data from meteorological forecasts are accessed. The vertical migration increment of pollutants caused by rainwater infiltration is calculated through Darcy's law. The land use change model predicts the location and production scale of a proposed factory within half a year, and combines with the pollutant emission inventory to predict the threat level of the new pollution source to the groundwater vulnerability area. Finally, a risk evolution curve with a timestamp is generated, marking in advance, two weeks ahead, the warning of excessive carcinogen concentration in the residential area downwind of the industrial park.

[0015] In a preferred embodiment, in step S8, after obtaining the latest soil sampling data from the environmental protection department every month, the online learning module compares the predicted risk level with the laboratory test results. When the deviation between the predicted concentration and the measured value of heavy metal arsenic exceeds 20%, the characteristic weighting coefficient is automatically adjusted by calculating the KL divergence between the two, and at the same time, the retraining process of the generative adversarial network is triggered, making the synthetic data more conform to the characteristics of soil moisture content changes in the current season, and ensuring the prediction stability of the model in the wet-dry alternating environment.

[0016] In summary, due to the adoption of the above technical solutions, the beneficial effects of the present invention are as follows: 1. In the present invention, the microbial metabolic process is real-time associated and modeled with the migration and transformation of pollutants, so that the biological effects and chemical indicators that are separated in traditional assessments are transformed into a unified risk driving force metric. The non-linear interaction entropy calculation mechanism can adaptively capture the antagonistic-synergistic balance relationship between the microbial community and pollutants in the soil environment, reflecting both the degradation potential of functional bacteria for specific pollutants and quantifying the dynamic regulation effect of environmental factor fluctuations on bioavailability. This cross-scale feature fusion technology enables the assessment model to not only identify the current pollution spatial distribution, but also predict the morphological transformation trend of pollutants in the biogeochemical cycle, providing a quantitative basis with ecological interpretability for risk warning.

[0017] 2. In the present invention, the time-varying remediation potential index output by the dynamic interactive feature generator can directly guide the selection of bioremediation technologies and the planning of implementation nodes. For example, by quantifying the metabolic efficiency differences of different microbial communities in the dry-wet alternating environment, the dosing ratio and timing of microbial agents can be optimized. Meanwhile, the closed-loop system formed by the migration-enhanced learning architecture and the self-optimizing feedback mechanism can continuously adapt to the non-linear changes of soil environmental parameters, ensuring that the evaluation model maintains stable and reliable prediction performance in the long time series. This technical path that deeply couples microscopic biological processes with macroscopic risk expression provides full-chain decision-making support for soil pollution control from risk tracing to solution optimization, effectively solving the problem of the disconnection between ecological mechanisms and engineering practices in traditional methods, thereby improving the scientificity and accuracy of overall decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a schematic diagram of the process principle of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0020] Embodiment: Referring to Figure 1 , a method for assessing soil pollution risk based on machine learning, comprising the following steps: S1: Obtain surface spectral features through satellite remote sensing technology, combine with a ground sensor network to monitor soil humidity and acidity in real time, and synchronously integrate the historical pollution event database and laboratory physical and chemical analysis reports to construct a cross-scale spatio-temporal dataset; S2: Use an adaptive threshold algorithm to detect abnormal sensor data, and introduce a soil microbial community metabolism model to simulate missing values, and eliminate the resolution differences of multi-source data through spatio-temporal alignment technology; S3: Design an interpretable feature generator to dynamically couple soil microbial activity indicators with environmental factors to generate time-varying features reflecting the migration and transformation rate of pollutants; S4: Use a deep residual network to extract spatial patterns in pollution hotspots, enable a pre-trained transfer model for data-scarce regions, and use a generative adversarial network to synthesize training samples that conform to local features; S5: Integrate a convolutional neural network and a graph attention mechanism to capture the spatial diffusion characteristics of pollutants and the topological association of the soil-groundwater system respectively, and output a heat map of the pollution risk probability distribution; S6: Combine the geographic information system with augmented reality technology to map the risk assessment results into a three-dimensional hierarchical risk map, visually showing the migration path of pollutants at the soil-plant-atmosphere interface; S7: Access meteorological forecast data and land use change models to dynamically predict the evolution trend of pollution risks within the next three months and generate early warning signals for high-risk areas; S8: Deploy an online learning module to regularly receive on-site sampling verification data, automatically adjust the model hyperparameters by comparing the KL divergence between the predicted and measured results, and then construct a multi-objective optimization tree based on the risk tracing results to provide customized treatment plans for different pollution scenarios.

[0021] In step S1, satellite remote sensing technology captures the surface reflectance difference with hyperspectral imaging to identify the spatial distribution pattern of potential pollution areas; the ground sensor network deploys electrochemical sensors and ion-selective electrodes to obtain parameters such as soil conductivity and redox potential in real time, forming vertical profile data with a monitoring frequency of minutes. The historical pollution event database matches with laboratory physical and chemical data through spatio-temporal coding, and finally generates a three-dimensional data cube containing spectral features, physical and chemical parameters, and pollution event labels, providing a multi-dimensional benchmark for subsequent analysis.

[0022] In step S2, the adaptive threshold algorithm dynamically adjusts the anomaly determination boundary according to the soil type. For example, a looser conductivity fluctuation threshold is set for clayey soils; the microbial metabolism model simulates the conversion efficiency of nitrifying bacteria on ammonia nitrogen. When the nitrogen content data is missing from the sensor, the original pollution load is inferred by detecting the nitrite concentration. The spatio-temporal alignment technology uses wavelet transform to unify the kilometer-level resolution of satellite data and the meter-level monitoring accuracy of ground sensors, eliminating the interference of scale differences on the simulation of pollutant diffusion.

[0023] In step S3, it specifically includes: S3-1: Data coupling layer: Integrate soil microbial activity indicators (such as the abundance of specific bacterial communities and the intensity of enzyme activity) and environmental factors (heavy metal ion concentration, temperature gradient, organic matter content) to establish a time series coupling matrix, retaining the dynamic association between pollutants and ecological responses.

[0024] S3-2: Feature generator architecture: Adopt a dual-channel attention mechanism to synchronously analyze the microbial community metabolism rate and environmental factor fluctuations in the time dimension, and generate a candidate feature pool through cross-channel feature crossing.

[0025] S3-3: Nonlinear interaction entropy calculation: The creative formula is as follows: H_t = Σ(P_m(i,t)·log(Q_e(j,t) / (P_m(i,t)^α + Q_e(j,t)^β))) Where: $H_t$ represents the bioremediation potential index at time $t$, and a larger value indicates a stronger ability of microorganisms to transform pollutants; $P_m(i,t)$ is the normalized abundance of the $i$-th type of functional microorganism at time $t$ (a normalized value in the range of 0 - 1); $Q_e(j,t)$ is the equivalent concentration of the $j$-th pollutant at time $t$ (a dimensionless value weighted by toxicity); $\alpha$ is the microbial metabolic decay coefficient (calibrated through laboratory culture experiments, with typical values ranging from 0.6 to 1.2); $\beta$ is the pollutant bioavailability adjustment factor (calculated dynamically depending on soil pH and organic matter content); S3 - 4: Feature optimization mechanism: Through the Shapley value interpretation model, features with a contribution degree greater than 15% to risk assessment are selected to enter the downstream model, ensuring that the generated features have both physical meaning and prediction efficacy.

[0026] In step S4, in the industrial legacy pollution area, the deep residual network extracts the spatial co - occurrence patterns of heavy metals such as chromium and cadmium through skip connections, and uses dilated convolutions to expand the feature receptive field to identify kilometer - scale pollution plumes; for the problem of scarce data in newly developed areas, the pre - trained model takes the pollutant migration law of agricultural land as prior knowledge, and then synthesizes virtual samples that conform to the local soil texture through a generative adversarial network, enabling the model to still capture the vertical infiltration characteristics of organic pollutants such as benzene series under limited data.

[0027] In step S5, the 3D convolutional neural network extracts the diffusion speed and direction features of pollutants in the soil profile, and enhances the ability to identify deep - layer soil pollution through the dilated convolution layer; the graph attention mechanism constructs a topological network of soil pores - groundwater channels, calculates the migration weights of pollutants along the preferential paths, and finally tensors and concatenates the two types of features at the feature layer to output a risk probability heat map with a resolution of 10 m × 10 m, accurately marking the high - risk areas where the leachate pollution boundary overlaps with the underground water source protection area.

[0028] In step S6, the geographic information system overlays the 2D risk heat map with the digital elevation model to generate an animation of the diffusion simulation of the pollution plume along the altitude gradient; the augmented reality device, through spatial registration technology, overlays and displays the enrichment concentration of pollutants in crop roots during on - site inspections, and uses the particle tracking algorithm to dynamically render the path of volatile organic compounds escaping from soil pores to the atmosphere, enabling decision - makers to intuitively grasp the multi - medium pollution coupling effect.

[0029] In step S7, rainfall intensity and wind direction data in the meteorological forecast data are accessed, and the vertical migration increment of pollutants caused by rainwater infiltration is calculated through Darcy's law. The land use change model predicts the location and production scale of the planned factory within half a year, and combines the pollutant emission inventory to predict the threat level of the new pollution source to the groundwater vulnerability area. Finally, a risk evolution curve with a timestamp is generated, and a carcinogen concentration exceeding warning for the residential area downwind of the industrial park is marked two weeks in advance.

[0030] In step S8, after obtaining the latest soil sampling data from the environmental protection department every month, the online learning module compares the predicted risk level with the laboratory test results. When the deviation between the predicted concentration and the measured value of heavy metal arsenic exceeds 20%, the characteristic weighting coefficient is automatically adjusted by calculating the KL divergence between the two, and at the same time, the retraining process of the generative adversarial network is triggered to make the synthetic data more conform to the soil moisture content change characteristics of the current season, ensuring the prediction stability of the model in the wet-dry alternating environment.

[0031] A soil pollution risk assessment device using machine learning, which operates the soil pollution risk assessment method of the above-mentioned embodiment.

[0032] In the present invention, the microbial metabolic process is real-time associated and modeled with the pollutant migration and transformation, so that the biological effects and chemical indicators that are separated in traditional assessments are transformed into a unified risk driving force metric. The non-linear interaction entropy calculation mechanism can adaptively capture the antagonistic-synergistic balance relationship between the microbial community and pollutants in the soil environment, which not only reflects the degradation potential of functional bacteria groups for specific pollutants, but also quantifies the dynamic regulation effect of environmental factor fluctuations on bioavailability. This cross-scale feature fusion technology enables the assessment model to not only identify the current pollution spatial distribution, but also predict the morphological transformation trend of pollutants in the biogeochemical cycle, providing a quantitative basis with ecological interpretability for risk warning.

[0033] In the present invention, the time-varying remediation potential index output by the dynamic interactive feature generator can directly guide the selection of bioremediation technologies and the planning of implementation nodes. For example, by quantifying the metabolic efficiency differences of different bacteria groups in the wet-dry alternating environment, the dosing ratio and timing of microbial agents can be optimized. At the same time, the closed-loop system formed by the migration-enhanced learning architecture and the self-optimizing feedback mechanism can continuously adapt to the non-linear changes of soil environmental parameters, ensuring the stable and reliable prediction performance of the assessment model in the long time series. This technical path that deeply couples microscopic biological processes with macroscopic risk expression provides full-chain decision support for soil pollution control from risk tracing to solution optimization, effectively solving the problem of the disconnection between ecological mechanisms and engineering practices in traditional methods, thereby improving the scientificity and accuracy of overall decision-making.

[0034] It should be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, article, or device comprising the said element.

[0035] The above description enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the broadest scope consistent with the principles and novel features disclosed herein.

Claims

1. A soil pollution risk assessment method based on machine learning, characterized by: The steps include: S1: Obtain surface spectral characteristics through satellite remote sensing technology, combine with ground sensor network to monitor soil moisture and pH in real time, and simultaneously integrate historical pollution event database and laboratory physical and chemical analysis reports to construct a cross-scale spatiotemporal data set; S2: Adaptive threshold algorithm is used to detect abnormal sensor data, and soil microbial community metabolism model is introduced to simulate missing values, and the resolution difference of multi-source data is eliminated through spatiotemporal alignment technology; S3: Design an interpretable feature generator to dynamically couple soil microbial activity indicators with environmental factors to generate time-varying features reflecting the migration and transformation rates of pollutants; S4: Use deep residual networks to extract spatial patterns in pollution hotspots, enable pre-trained transfer models for data-scarce areas, and use adversarial generative networks to synthesize training samples that meet local characteristics; S5: Integrate convolutional neural network and graph attention mechanism to capture the spatial diffusion characteristics of pollutants and the topological correlation of soil-groundwater system, and output the heat map of pollution risk probability distribution; S6: Combine geographic information system and augmented reality technology to map risk assessment results into a three-dimensional layered risk map to intuitively display the migration path of pollutants at the soil-plant-atmosphere interface; S7: Access to meteorological forecast data and land use change models to dynamically predict the evolution trend of pollution risks in the next three months and generate early warning signals for high-risk areas; S8: Deploy an online learning module, regularly receive field sampling verification data, automatically adjust the model hyperparameters by comparing the KL divergence between the predicted and measured results, and then build a multi-objective optimization tree based on the risk tracing results to provide customized governance solutions for different pollution scenarios.

2. The soil pollution risk assessment method based on machine learning as claimed in claim 1, characterized in that: In step S1, satellite remote sensing technology uses hyperspectral imaging to capture differences in surface reflectivity and identify the spatial distribution patterns of potential pollution areas; the ground sensor network deploys electrochemical sensors and ion selective electrodes to obtain soil conductivity and redox potential parameters in real time, forming vertical profile data with a minute-level monitoring frequency; the historical pollution event database is matched with laboratory physical and chemical data through spatiotemporal coding, and finally generates a three-dimensional data cube containing spectral characteristics, physical and chemical parameters, and pollution event labels, providing a multi-dimensional benchmark for subsequent analysis.

3. The soil pollution risk assessment method based on machine learning as claimed in claim 1, characterized in that: In step S2, the adaptive threshold algorithm dynamically adjusts the abnormal judgment boundary according to the soil type, and sets a looser conductivity fluctuation threshold for clay soil; the microbial metabolism model simulates the conversion efficiency of nitrifying bacteria to ammonia nitrogen, and when the sensor lacks nitrogen content data, the original pollution load is inferred from the detected nitrite concentration; the spatiotemporal alignment technology uses wavelet transform to unify the kilometer-level resolution of satellite data and the meter-level monitoring accuracy of ground sensors, eliminating the interference of scale differences on pollutant diffusion simulation.

4. The soil pollution risk assessment method based on machine learning as claimed in claim 1, characterized in that: The step S3 specifically includes: S3-1: Data coupling layer: Integrate soil microbial activity indicators and environmental factors to establish a time series coupling matrix to retain the dynamic association between pollutants and ecological responses; S3-2: Feature generator architecture: A dual-channel attention mechanism is used to simultaneously analyze the metabolic rate of the microbial community and the fluctuation of environmental factors in the time dimension, and a candidate feature pool is generated by cross-channel feature cross-section; S3-3: Nonlinear interaction entropy calculation: The creative formula is as follows: H_t=Σ(P_m(i,t)·log(Q_e(j,t) / (P_m(i,t)^α+Q_e(j,t)^β))) in: H_t represents the bioremediation potential index at time t, and the larger the value, the stronger the microorganism's ability to transform pollutants; P_m(i,t) is the standardized abundance of the i-th type of functional microorganism at time t; Q_e(j,t) is the equivalent concentration of the jth pollutant at time t; α is the microbial metabolic decay coefficient; β is the pollutant bioavailability regulator; S3-4: Feature optimization mechanism: Through the Shapley value interpretation model, the features that contribute more than 15% to risk assessment are screened and enter the downstream model to ensure that the generated features have both physical meaning and predictive effectiveness.

5. The soil pollution risk assessment method based on machine learning as claimed in claim 1, characterized in that: In step S4, in the industrial legacy pollution area, the deep residual network extracts the spatial co-occurrence pattern of heavy metals such as chromium and cadmium through jump connections, and uses dilated convolution to expand the feature receptive field to identify kilometer-level pollution plumes.

6. The soil pollution risk assessment method based on machine learning as claimed in claim 1, characterized in that: In step S5, the three-dimensional convolutional neural network extracts the diffusion speed and direction characteristics of pollutants in the soil profile, and enhances the deep soil pollution identification ability through the void convolution layer; the graph attention mechanism constructs the topological network of soil pore-groundwater channels, calculates the migration weights of pollutants along the priority path, and finally performs tensor splicing of the two types of features in the feature layer to output a risk probability heat map with a resolution of 10 meters × 10 meters, marking the high-risk areas where the leachate pollution boundary overlaps with the groundwater source protection area.

7. The soil pollution risk assessment method based on machine learning as claimed in claim 1, characterized in that: In step S6, the geographic information system superimposes the two-dimensional risk heat map with the digital elevation model to generate a diffusion simulation animation of the pollution plume on the altitude gradient; the augmented reality device uses spatial registration technology to superimpose and display the enrichment concentration of pollutants in the crop root system during field inspections, and uses a particle tracking algorithm to dynamically render the path of volatile organic compounds escaping from soil pores to the atmosphere, so that decision makers can intuitively grasp the multi-media pollution coupling effect.

8. The soil pollution risk assessment method based on machine learning as claimed in claim 1, characterized in that: In step S7, the rainfall intensity and wind direction data in the meteorological forecast data are accessed, and the vertical migration increment of pollutants caused by rainwater infiltration is calculated using Darcy's law; the land use change model predicts the location and production scale of the factory to be built within half a year, and the threat level of new pollution sources to groundwater vulnerable areas is predicted in combination with the pollutant emission inventory, and finally a risk evolution curve with a timestamp is generated.

9. The soil pollution risk assessment method based on machine learning as claimed in claim 1, characterized in that: In step S8, after obtaining the latest soil sampling data from the environmental protection department every month, the online learning module compares the predicted risk level with the laboratory test results. When the predicted concentration of heavy metal arsenic deviates from the measured value by more than 20%, the feature weighting coefficient is automatically adjusted by calculating the KL divergence between the two, and at the same time, the retraining process of the adversarial generative network is triggered to make the synthetic data more in line with the soil moisture change characteristics of the current season, thereby ensuring the prediction stability of the model in a dry-wet alternating environment.

10. A soil pollution risk assessment device based on machine learning, characterized in that: The device runs the soil pollution risk assessment method based on machine learning as described in any one of claims 1 to 9.

Images