A method and system for monitoring soil pollution

Abstract

This invention provides a soil pollution monitoring method and system. It optimizes input data quality through refined preprocessing and adaptive feature filtering of multi-source biological, physicochemical, and spectral time-series data; constructs a time-series-spatial dual-channel coding and cross-attention fusion mechanism to deeply mine pollution dynamics and spatial characteristics; combines fingerprint database matching, transfer learning unmixing, and regression models to achieve high-precision identification of pollutants at the ppb level and generate visualized pollution distribution maps; relies on a pollution propagation knowledge graph to achieve intelligent prediction of pollution trends in multiple scenarios; verifies the predicted trend against the actual pollution evolution trend at fixed time intervals, accurately marks abnormal key monitoring areas based on multi-dimensional deviation thresholds, and iteratively optimizes the inference model in reverse. This invention solves the technical problems of poor data fusion, low identification accuracy, inaccurate trend prediction, and lagging identification of abnormal areas in traditional soil monitoring.

Description

Technical Field

[0001] This invention relates to the field of soil pollution monitoring, and in particular to a soil pollution monitoring method and system. Background Technology

[0002] Soil pollution is characterized by its concealment, delayed effects, and strong spatiotemporal heterogeneity. Accurate monitoring of pollution distribution and dynamic evolution trends is a core prerequisite for efficient soil remediation and risk management. Current soil pollution monitoring technologies suffer from several shortcomings: First, monitoring data sources are limited, failing to effectively integrate biological temporal characteristics, soil physicochemical parameters, and multi-source spectral data. This results in low data utilization, significant noise interference, and difficulty in adapting to complex site pollution characteristics. Second, traditional feature extraction methods are fixed and cannot adaptively select sensitive features for different pollution types. Hyperspectral data suffers from mixed pixel interference, leading to insufficient accuracy in identifying trace pollutants and hindering ppb-level precision detection. Third, conventional monitoring often employs single-channel feature extraction methods, failing to consider both the temporal dynamic changes and spatial distribution characteristics of pollution. This results in blurred pollution plume boundaries and large concentration inversion errors. Fourth, existing pollution trend predictions rely heavily on single mathematical models, lacking support from causal knowledge of pollution propagation. These models have poor generalization capabilities and lack a dynamic verification mechanism between predicted and actual monitoring values. This makes it difficult to identify abnormal pollution areas in a timely manner, and the accuracy of key monitoring area delineation is low, failing to meet the needs of refined soil remediation and management.

[0003] In summary, existing soil monitoring technologies suffer from shortcomings such as poor data fusion, weak feature extraction targeting, low pollution identification accuracy, insufficient reliability of trend prediction, and lack of dynamic early warning capabilities. These shortcomings prevent the realization of high-precision, adaptive dynamic monitoring of soil pollution throughout the entire process, which severely restricts the intelligent and precise implementation of soil remediation. Summary of the Invention

[0004] The purpose of this invention is to provide a soil pollution monitoring method and system that overcomes the shortcomings of existing technologies, such as poor multi-source data fusion, low accuracy in identifying trace pollutants, inaccurate pollution trend prediction, delayed identification of abnormal areas, and lack of adaptive verification and optimization mechanisms. This invention provides precise fusion of multi-source heterogeneous data, high-precision qualitative and quantitative identification of pollutants, visualization of the spatiotemporal distribution of pollution, intelligent prediction and dynamic verification of pollution trends, and adaptive marking of key monitoring areas. This improves the accuracy, real-time performance, and intelligence of soil pollution monitoring, providing reliable data and decision support for precise soil remediation and dynamic risk management.

[0005] To achieve the above objectives, the present invention provides a soil pollution monitoring method, comprising the following steps: Acquire and preprocess biomarker data, physicochemical parameters and spectral data, where the biomarker data and ground spectral data are time series data; Biometric data, physicochemical parameters, and spectral data are processed through a dual-channel encoding mechanism to obtain fused features; Pollutants and their concentrations are identified based on fusion features, and a pollutant distribution map is generated based on the identified pollutants and their concentrations. Based on existing pollutant distribution maps, knowledge graphs are used to predict pollution development trends. Set an interval time, and when the interval time is reached, analyze the changes in the pollutant distribution map during this period to obtain the actual development trend; The predicted development trend is compared with the actual development trend. When the predicted development trend and the actual development trend exceed the set threshold, the location is marked as a key monitoring area.

[0006] Preferably, the preprocessing of biometric data, physicochemical parameters, and hyperspectral data specifically includes: The obtained biometric data, physicochemical parameters and spectral data are spatiotemporally registered. For hyperspectral data, a dynamic feature selection mechanism is set up to screen the full-band hyperspectral data and full physicochemical parameters, and to screen the features most sensitive to the current pollution type online to obtain the optimal feature subset. The hyperspectral data in the optimal feature subset are fine-tuned and unmixed using standard spectral data to obtain the adjusted preferred hyperspectral features.

[0007] Preferably, biometric data, physicochemical parameters, and spectral data are processed through a dual-channel encoding mechanism to obtain fused features, specifically as follows: Physicochemical parameters are encoded over time to extract long-term dependencies, capture the rising / falling trends and abrupt changes in pollution concentration, and obtain dynamic feature vectors of pollution. Spectral data is spatially encoded to extract multi-scale spatial features, accurately segment the pollution plume boundary, and obtain a pollution spatial feature map. The pollution dynamic feature vector and pollution spatial feature map are fused through cross-attention, specifically including... Pollution dynamic feature vector guides pollution spatial feature map: When the concentration in a certain area suddenly increases, the weight of the pollution spatial feature map of that area is automatically enhanced; Pollution spatial feature map guides pollution dynamic feature vector: when a new pollution plume is discovered, the historical time series data of the area is automatically retrieved.

[0008] Preferably, pollutants and their concentrations are identified based on fusion features, and a pollutant distribution map is generated based on the identified pollutants and their concentrations, specifically including: The fused features are input into the classification subnetwork, and the results of hyperspectral unmixing through transfer learning are combined with the pollution component fingerprint database. Through similarity matching and neural network classification, the pollutant type in the current area is determined, and the pollutant category label is assigned at the plot / pixel level. A feature-pollutant concentration regression model was constructed, incorporating complete time-series data and endmember abundance information after GAN (Generative Adversarial Network) completion. The trained regression model was used to calculate the concentration values ​​of heavy metals, organic pollutants, etc. for each grid cell. The model was then corrected by combining physicochemical parameters to achieve the inversion of trace concentrations at the ppb (parts per billion) level. The pollutant concentration values ​​for each spatial cell were obtained. By mapping pollutant types and concentration values ​​to a unified spatial grid, continuous and fully covered two-dimensional soil pollution raster data is obtained. A semantic segmentation algorithm is used to extract the precise boundary of the pollution plume. Based on the pollutant concentration threshold, the pollution areas are divided into light, moderate and heavy pollution areas to obtain the continuous pollution spatial distribution area.

[0009] Using a geographic coordinate system as the base map, spatially interpolated concentration raster is overlaid, and different colors are rendered according to the concentration gradient. Pollution type, concentration range, and pollution boundary are marked to create a vector / raster format spatial distribution map of pollutants, including location, range, concentration gradient, and pollution level.

[0010] Preferably, the pollution development trend is predicted using a knowledge graph based on existing pollutant distribution maps, specifically as follows: Soil grids, pollutant information, and environmental factors were extracted from pollutant distribution maps to construct a pollution propagation knowledge graph with soil plots as nodes and pollution migration as the correlation. By integrating soil physicochemical, hydrological, and topographic data to assign node attributes and edge weights, and combining causal reasoning to improve the graph logic; Based on knowledge graphs, the diffusion paths and concentration evolution patterns of pollutants can be deduced to predict pollution development trends under multiple scenarios; Finally, by comparing the predicted results with the actual monitoring data at fixed time intervals, areas with deviations exceeding the threshold are designated as key monitoring areas, thus completing the dynamic early warning of pollution.

[0011] Preferably, an interval time is set, and when the interval time is reached, the changes in the pollutant distribution map during this period are analyzed to obtain the actual development trend, specifically: Obtain the measured pollutant distribution map within the current interval; Extract soil grids, pollutant information, and environmental factors from the current interval of the pollutant distribution map; Based on pollutant information and environmental factors, the comprehensive concentration change trend, spatial expansion trend, migration path, and grade evolution of the corresponding soil grid are calculated to obtain the actual pollution development trend.

[0012] Preferably, the predicted development trend is compared with the actual development trend. When the predicted development trend and the actual development trend exceed a set threshold, the location is marked as a key monitoring area, specifically including: Under the same spatiotemporal soil grid, the pollution development trend predicted by the knowledge graph and the actual pollution development trend are calculated in multiple dimensions, including pollutant concentration error, spatial morphology error and migration direction error. A comprehensive deviation threshold is pre-set. When the regional deviation exceeds the threshold, the pollution evolution of the region is determined to be abnormal, and it is marked as a key monitoring area. The monitoring intensity is increased and feedback is used to optimize the knowledge graph reasoning rules.

[0013] A soil pollution monitoring system, comprising: The data processing module is used to acquire and preprocess biometric data, physicochemical parameters, and spectral data, where the biometric data and ground spectral data are both time-series data. The dual-channel module is used to obtain fused features by using a dual-channel encoding mechanism to process biometric data, physicochemical parameters, and spectral data. The distribution map generation module is used to identify pollutants and their concentrations based on fusion features, and generate a pollutant distribution map based on the identified pollutants and their concentrations. The prediction module is used to predict pollution development trends based on existing pollutant distribution maps using knowledge graphs. The comparison module is used to set the interval time. When the interval time is reached, the changes in the pollutant distribution map during this period are analyzed to obtain the actual development trend. The predicted development trend is compared with the actual development trend. When the predicted development trend and the actual development trend exceed the set threshold, this location is marked as a key monitoring area.

[0014] Therefore, the present invention employs the above-mentioned soil pollution monitoring method and system, and the technical effects are as follows: This invention achieves multi-source fusion of biological time-series data, physicochemical parameters, and hyperspectral time-series data. Through dynamic feature selection, spectral fine-tuning and mixing, and GAN data completion technology, it solves the problems of traditional data redundancy, noise interference, mixed pixel distortion, and data missing, and significantly improves the accuracy and effectiveness of input data.

[0015] This invention employs a dual-channel encoding + cross-attention bidirectional fusion mechanism to separately mine the temporal dynamic change patterns and spatial distribution characteristics of pollution, achieving deep complementarity between temporal and spatial features. This overcomes the shortcomings of single-channel feature extraction, such as one-sided information extraction, blurred pollution boundary delineation, and inaccurate trend capture.

[0016] This invention combines fingerprint database matching, transfer learning unmixing, physicochemical parameter correction and regression model to achieve trace identification and quantitative inversion of pollutants at the ppb level. At the same time, it accurately classifies pollution levels and renders standardized pollution distribution maps, realizing visualization and refined characterization of soil pollution.

[0017] This invention constructs a pollution propagation simulation system based on knowledge graphs and causal reasoning, which closely matches the real propagation mechanism of soil pollution, predicts pollution development trends in multiple scenarios, and has a prediction logic and reliability far superior to traditional mathematical models.

[0018] This invention establishes a closed-loop dynamic monitoring mechanism of "prediction-measurement-comparison-verification-optimization". Through multi-dimensional deviation threshold judgment, it accurately marks the key monitoring areas of anomalies. At the same time, it iteratively optimizes the knowledge graph in reverse to realize the adaptive upgrading of the monitoring model, continuously improve the long-term monitoring accuracy, and adapt to complex and ever-changing soil pollution scenarios. Attached Figure Description

[0019] Figure 1 This is a spatial distribution map of pollutants; Figure 2 This is a pollution evolution prediction diagram; Figure 3 Mark key monitoring areas. Detailed Implementation

[0020] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0022] Example 1 A method for monitoring soil pollution includes the following steps: S1. Acquisition and Refined Preprocessing of Multi-Source Heterogeneous Data Continuous collection of multi-source monitoring data from target contaminated sites, including time-series soil biological characteristic data, soil physicochemical parameters (pH, redox potential, soil porosity, organic matter content, moisture content, etc.), and time-series ground hyperspectral data. First, the three types of data are uniformly registered in time and space, unifying the geographic coordinate system and timestamps to eliminate differences in time and space scales caused by satellite, ground sensor, and UAV monitoring.

[0023] For full-band hyperspectral data and all physicochemical parameters, a dynamic feature selection mechanism is initiated. Based on the main pollution types of the current site (heavy metals, organic pollutants, and compound pollution), highly sensitive spectral bands and key physicochemical parameters are screened online, redundant and invalid features are eliminated, and an optimal feature subset with reduced dimensions and strong targeting is constructed. A standard spectral database is retrieved, and transfer learning is used to fine-tune the hyperspectral data in the optimal feature subset, decomposing mixed pixels into pure substance endmember abundance, breaking through the mixed pixel limitation, and obtaining the preferred hyperspectral features after noise reduction and optimization. At the same time, a generative adversarial network (GAN) is used to fill in missing data, strictly following the constraints of soil physical mechanisms to ensure data integrity and authenticity.

[0024] S2, Dual-channel cross-attention feature fusion Construct a dual-channel feature extraction mechanism that combines temporal and spatial coding: Temporal coding channel: The preprocessed soil physicochemical parameters, biological time-series data, and complete spectral time-series data are input into the temporal coding module. The Transformer encoder is used to mine the long-term temporal dependencies of the data, accurately capture the phased rise and fall patterns and abrupt change nodes of pollution concentration, and output a pollution dynamic feature vector with unified dimensions.

[0025] Spatial coding channel: The optimized hyperspectral spatial feature data is input into the multi-scale spatial coding module, and the spatial distribution features of pollution at different scales are extracted through the convolutional network to accurately segment the edge boundary of the pollution plume and output a refined spatial feature map of pollution.

[0026] A cross-attention fusion mechanism is adopted to achieve bidirectional feature complementarity: First, temporal features guide spatial features. When a sudden increase in the temporal data of pollution concentration in a local area is detected, the spatial feature weight of that area is automatically increased to strengthen the spatial representation of the abnormal pollution area. Second, spatial features guide temporal features. When the spatial coding module identifies a new pollution plume area, it automatically retrieves the historical temporal data of that area to explore the pollution accumulation and evolution pattern, and finally fuses to obtain multi-source fusion features that take into account both temporal dynamism and spatial accuracy.

[0027] S3. High-precision identification and spatial distribution map generation of pollutants like Figure 1 As shown, the fused features are input into the trained classification subnetwork, a pre-constructed pollution component fingerprint database is retrieved, and endmember features obtained by transfer learning hyperspectral unmixing are combined. Through feature similarity matching and neural network classification algorithms, the pollutant type of each grid unit is automatically identified, and plot-level and pixel-level pollution category labels are output.

[0028] A fusion feature-pollutant concentration quantitative regression model is constructed, which takes the complete time-series data and spectral endmember abundance information of GAN completion as the core input, and combines the physicochemical parameters such as soil pH, organic matter, and porosity for error correction, so as to realize the inversion of ppb-level trace concentrations of heavy metals and organic pollutants and accurately output the real-time concentration values ​​of pollutants in each spatial grid cell.

[0029] All grid cells are mapped to a unified standard spatial grid for pollutant types and concentrations. Spatial interpolation algorithms are used to fill in the blank areas of discrete points, generating two-dimensional soil pollution raster data with continuous coverage across the entire area. Semantic segmentation algorithms are used to refine the pollution plume boundaries, classifying pollution levels into light, moderate, and severe according to national soil pollution control standards, removing noise interference, and smoothing boundary contours. Based on a geographic coordinate system base map, color-coded rendering is performed according to concentration gradient differences, labeling pollution types, concentration ranges, pollution boundaries, and pollution levels to generate standardized vector and raster format spatial distribution maps of pollutants.

[0030] S4. Pollution Development Trend Prediction Based on Knowledge Graph like Figure 2 As shown, the generated spatial distribution map of pollutants is analyzed in a structured manner. The soil in the whole region is divided into grids, and each soil grid is used as a knowledge graph entity node. The pollutant type, concentration, pollution level, and soil physicochemical properties of each node are extracted. At the same time, environmental factors such as topography, groundwater runoff, rainfall, and temperature are included as auxiliary entity nodes.

[0031] The relationships between nodes are defined, with pollutant migration to groundwater, surface diffusion, vertical infiltration, soil adsorption, microbial degradation, and environmental disturbance as the associated edges. The weights of each associated edge are assigned using hydrological and hydraulic data and soil migration parameters to characterize the intensity and probability of pollution transmission. A Bayesian causal reasoning model is integrated to clarify the causal logic of pollution source-transmission pathway-environmental factors-pollution evolution, thus improving the reasoning rules of the pollution transmission knowledge graph.

[0032] Based on the completed pollution propagation knowledge graph, multi-scenario pollution evolution simulations are conducted to simulate the diffusion paths, concentration changes, and expansion or contraction patterns of pollutants under different hydrological and meteorological conditions. The system predicts the overall pollution development trend within a set time period and outputs a predictive pollution evolution distribution map.

[0033] S5. Fixed-interval time-series monitoring and extraction of actual pollution trends A fixed monitoring time interval can be preset, which can be adaptively configured to 7 days, 15 days or 30 days according to the site pollution risk level. The system automatically keeps track of the time and automatically triggers the monitoring and analysis process after the interval expires.

[0034] Collect on-site measured data at the end of the current time interval, and regenerate the spatial distribution map of pollutants for the current period through the aforementioned steps S1-S3. Extract the pollutant type, real-time concentration, pollution range, boundary location, and corresponding environmental factor data for each soil grid in the distribution map. Calculate the pollutant concentration change and concentration change rate within each grid interval, and statistically analyze the evolution of pollution levels. Compare the pollution distribution maps of the two periods to extract the spatial expansion and contraction areas of pollution and the actual migration paths of pollutants. Integrate the temporal changes in concentration, spatial morphological changes, migration patterns, and level evolution characteristics to obtain the actual pollution development trend within the monitoring interval.

[0035] S6. Trend Deviation Detection and Adaptive Marking of Key Monitoring Areas like Figure 3 As shown, a unified spatiotemporal grid scale, coordinate system, and evaluation index system for predicting and actual trends are established, and multi-dimensional deviation quantification calculations are carried out: including relative and absolute errors of pollutant concentration, spatial morphological errors such as pollution plume boundary overlap and pollution area deviation, as well as angular deviations between predicted and actual pollutant migration paths.

[0036] Based on site control requirements and industry standards, a comprehensive deviation threshold is pre-set, and each grid area in the entire region is verified and judged. When the multi-dimensional comprehensive deviation of a certain area exceeds the preset threshold, it is determined that the pollution evolution pattern of that area is abnormal and the model prediction adaptability has decreased, and the area is permanently marked as a key monitoring area.

[0037] For key monitoring areas, the system automatically increases the frequency of monitoring and sampling, densifies the density of sensor deployment and the frequency of spectral monitoring, so as to achieve refined management and control of key areas. At the same time, the measured evolution data of the area is input into the pollution propagation knowledge graph to iteratively update node attributes and edge weights, optimize causal reasoning rules and diffusion inference logic, realize adaptive iterative upgrade of the model, and continuously improve the accuracy of subsequent pollution trend prediction.

[0038] Therefore, the present invention adopts the above-mentioned soil pollution monitoring method and system, which solves the technical problems of poor data fusion, low identification accuracy, inaccurate trend prediction, and lagging identification of abnormal areas in traditional soil monitoring.

[0039] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. 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 still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for monitoring soil pollution, characterized in that, Includes the following steps: Acquire and preprocess biomarker data, physicochemical parameters and spectral data, where the biomarker data and ground spectral data are time series data; Biometric data, physicochemical parameters, and spectral data are processed through a dual-channel encoding mechanism to obtain fused features; Pollutants and their concentrations are identified based on fusion features, and a pollutant distribution map is generated based on the identified pollutants and their concentrations. Based on existing pollutant distribution maps, knowledge graphs are used to predict pollution development trends. Set an interval time, and when the interval time is reached, analyze the changes in the pollutant distribution map during this period to obtain the actual development trend; The predicted development trend is compared with the actual development trend. When the predicted development trend and the actual development trend exceed the set threshold, the location is marked as a key monitoring area.

2. The soil pollution monitoring method according to claim 1, characterized in that, Preprocessing of biometric data, physicochemical parameters, and hyperspectral data specifically includes: The obtained biometric data, physicochemical parameters and spectral data are spatiotemporally registered. For hyperspectral data, a dynamic feature selection mechanism is set up to screen the full-band hyperspectral data and full physicochemical parameters, and to screen the features most sensitive to the current pollution type online to obtain the optimal feature subset. The hyperspectral data in the optimal feature subset are fine-tuned and unmixed using standard spectral data to obtain the adjusted preferred hyperspectral features.

3. The soil pollution monitoring method according to claim 1, characterized in that, Biometric data, physicochemical parameters, and spectral data are processed through a dual-channel encoding mechanism to obtain fused features, specifically: Physicochemical parameters are encoded over time to extract long-term dependencies, capture the rising / falling trends and abrupt changes in pollution concentration, and obtain dynamic feature vectors of pollution. Spectral data is spatially encoded to extract multi-scale spatial features, accurately segment the pollution plume boundary, and obtain a pollution spatial feature map. The pollution dynamic feature vector and pollution spatial feature map are fused through cross-attention, specifically including... Pollution dynamic feature vector guides pollution spatial feature map: When the concentration in a certain area suddenly increases, the weight of the pollution spatial feature map of that area is automatically enhanced; Pollution spatial feature map guides pollution dynamic feature vector: when a new pollution plume is discovered, the historical time series data of the area is automatically retrieved.

4. The soil pollution monitoring method according to claim 1, characterized in that, Based on the fusion of features, pollutants and their concentrations are identified, and a pollutant distribution map is generated according to the identified pollutants and their concentrations. Specifically, this includes: The fused features are input into the classification subnetwork, and the results of hyperspectral unmixing through transfer learning are combined with the pollution component fingerprint database. Through similarity matching and neural network classification, the pollutant type in the current area is determined, and the pollutant category label is assigned at the plot / pixel level. A feature-pollutant concentration regression model was constructed, incorporating complete time-series data and endmember abundance information after GAN completion. The trained regression model was used to calculate the concentration values ​​of heavy metals and organic pollutants for each grid cell. The model was then corrected using physicochemical parameters to achieve ppb-level trace concentration inversion, thus obtaining the pollutant concentration values ​​for each spatial cell. By mapping pollutant types and concentration values ​​to a unified spatial grid, continuous and fully covered two-dimensional soil pollution raster data is obtained. A semantic segmentation algorithm is used to extract the precise boundary of the pollution plume, and the pollution areas are divided into light, moderate and heavy pollution areas based on the pollutant concentration threshold to obtain the continuous pollution spatial distribution area. Using a geographic coordinate system as the base map, spatially interpolated concentration raster is overlaid, and different colors are rendered according to the concentration gradient. Pollution type, concentration range, and pollution boundary are marked to create a vector / raster format spatial distribution map of pollutants, including location, range, concentration gradient, and pollution level.

5. A soil pollution monitoring method according to claim 1, characterized in that, Based on existing pollutant distribution maps, knowledge graphs are used to predict pollution development trends, specifically: Soil grids, pollutant information, and environmental factors were extracted from pollutant distribution maps to construct a pollution propagation knowledge graph with soil plots as nodes and pollution migration as the correlation. By integrating soil physicochemical, hydrological, and topographic data to assign node attributes and edge weights, and combining causal reasoning to improve the graph logic; Based on knowledge graphs, the diffusion paths and concentration evolution patterns of pollutants can be deduced to predict pollution development trends under multiple scenarios; Finally, by comparing the predicted results with the actual monitoring data at fixed time intervals, areas with deviations exceeding the threshold are designated as key monitoring areas, thus completing the dynamic early warning of pollution.

6. The soil pollution monitoring method according to claim 1, characterized in that, Set an interval time, and when the interval time is reached, analyze the changes in the pollutant distribution map during this period to obtain the actual development trend, specifically: Obtain the measured pollutant distribution map within the current interval; Extract soil grids, pollutant information, and environmental factors from the current interval of the pollutant distribution map; Based on pollutant information and environmental factors, the comprehensive concentration change trend, spatial expansion trend, migration path, and grade evolution of the corresponding soil grid are calculated to obtain the actual pollution development trend.

7. The soil pollution monitoring method according to claim 1, characterized in that, The predicted development trend is compared with the actual development trend. When the predicted and actual development trends exceed a set threshold, the location is marked as a key monitoring area, specifically including: Under the same spatiotemporal soil grid, the pollution development trend predicted by the knowledge graph and the actual pollution development trend are calculated in multiple dimensions, including pollutant concentration error, spatial morphology error and migration direction error. A comprehensive deviation threshold is pre-set. When the regional deviation exceeds the threshold, the pollution evolution of the region is determined to be abnormal, and it is marked as a key monitoring area. The monitoring intensity is increased and feedback is used to optimize the knowledge graph reasoning rules.

8. A soil pollution monitoring system, characterized in that, include: The data processing module is used to acquire and preprocess biometric data, physicochemical parameters, and spectral data, where the biometric data and ground spectral data are both time-series data. The dual-channel module is used to obtain fused features by using a dual-channel encoding mechanism to process biometric data, physicochemical parameters, and spectral data. The distribution map generation module is used to identify pollutants and their concentrations based on fusion features, and generate a pollutant distribution map based on the identified pollutants and their concentrations. The prediction module is used to predict pollution development trends based on existing pollutant distribution maps using knowledge graphs. The comparison module is used to set the interval time. When the interval time is reached, the changes in the pollutant distribution map during this period are analyzed to obtain the actual development trend. The predicted development trend is compared with the actual development trend. When the predicted development trend and the actual development trend exceed the set threshold, this location is marked as a key monitoring area.

Images