Intelligent early warning method for soil and groundwater pollution based on multi-modal data fusion

Abstract

This invention discloses an intelligent early warning method for soil and groundwater pollution based on multimodal data fusion, comprising: acquiring multi-source environmental monitoring data of the area under study to obtain the multimodal fusion characteristics of the area under study; constructing a pollutant prediction model through parameter inversion and SEAWAT numerical simulation; determining the risk index of the area under study; and then using at least one of the risk index and the distribution of sensitive receptors, pollutant diffusion rate, and predicted concentration exceedance multiple in the multimodal fusion characteristics as the judgment criteria to determine the early warning level and issue an early warning. This invention provides a complete data fusion-migration simulation-risk assessment-dynamic early warning scheme, solving the problems of data dispersion, response lag, and low early warning accuracy in the prior art, and improving the intelligent level of soil and groundwater pollution risk management.

Description

Technical Field

[0001] This invention relates to the field of environmental information and pollution early warning technology, specifically to an intelligent early warning method for soil and groundwater pollution based on multimodal data fusion. Background Technology

[0002] With the acceleration of industrialization, soil and groundwater pollution incidents around high-risk facilities such as chemical industrial parks and tank farms occur frequently, seriously threatening the ecological environment and human health. Furthermore, soil and groundwater pollution is characterized by its concealment, delayed effects, and difficulty in remediation, posing a significant challenge to current environmental risk management. An effective pollution early warning system can provide crucial decision support for pollution prevention and emergency response; however, traditional monitoring and early warning methods often rely on single data sources, such as periodic sampling and testing or single sensor monitoring, making it difficult to comprehensively and in real-time capture the dynamic evolution of pollution, leading to delayed early warning responses and missed opportunities for optimal pollution control.

[0003] In existing technologies, pollution early warning methods mainly include numerical simulation based on physical mechanisms and data-driven methods based on statistics. Physical models (such as groundwater flow and solute transport models) can characterize the migration and transformation patterns of pollutants, but they require accurate hydrogeological parameters and boundary conditions, and are computationally time-consuming, making it difficult to meet the needs of real-time early warning. Data-driven methods can quickly predict pollution trends, but they often ignore physical laws, have insufficient model generalization ability, and have high requirements for data quality and sample size. In addition, existing technologies usually process heterogeneous data such as meteorological, hydrological, remote sensing, and online monitoring in isolation, lacking effective data fusion mechanisms and failing to fully utilize the spatiotemporal correlation between multimodal information, resulting in low early warning accuracy and widespread false alarms and missed alarms.

[0004] Therefore, there is an urgent need to develop an intelligent method that can efficiently integrate multi-source heterogeneous data, couple physical mechanisms with artificial intelligence, and achieve rapid and accurate early warning. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides an intelligent early warning method for soil and groundwater pollution based on multimodal data fusion.

[0006] A smart early warning method for soil and groundwater pollution based on multimodal data fusion includes the following steps: Multi-source environmental monitoring data of the area under study is obtained. At least two neural network models are used to extract the spatiotemporal features of the multi-source environmental monitoring data. Feature alignment and fusion are performed through an attention fusion mechanism to obtain the multimodal fusion features of the area under study. By using parameter inversion methods and numerical simulation methods for pollution migration such as SEAWAT, GMS, Feflow, or TOUGHREACT, a dataset of the spatiotemporal distribution of pollutant concentrations corresponding to the multimodal fusion features is obtained; based on the multimodal fusion features and the dataset of the spatiotemporal distribution of pollutant concentrations, a pollutant prediction model is constructed; the input of the pollutant prediction model is the multimodal fusion features, and the output is the predicted spatiotemporal distribution of pollutant concentrations. Based on the comparison between the predicted spatiotemporal distribution of pollutant concentration and the pollution concentration threshold, the risk index of the area under study is determined; then, the risk index is combined with at least one of the following in the multimodal fusion features: sensitive receptor distribution, pollutant diffusion rate, and predicted concentration exceedance multiple, to determine the warning level and issue a warning.

[0007] Explanation: The above method combines multi-source environmental monitoring data with various neural networks, and utilizes the attention mechanism to achieve efficient alignment and fusion of cross-modal spatiotemporal features, significantly improving the comprehensiveness and robustness of feature representation. Furthermore, it integrates physical-driven (SEAWAT numerical simulation and parameter inversion) and data-driven methods to construct a pollutant prediction model, enhancing the model's generalization ability and interpretability. By combining multimodal fusion features with receptor distribution, diffusion rate, etc., it realizes an intelligent decision-making closed loop from pollutant concentration prediction to multi-dimensional risk assessment and graded early warning, thereby greatly improving the accuracy, timeliness, and systematic nature of environmental risk early warning.

[0008] Furthermore, the multi-source environmental monitoring data includes meteorological data, hydrological data, remote sensing images, pollutant survey data, online monitoring data, high-risk facility leakage monitoring data, multiple fingerprint profiles, and isotope tracer data.

[0009] Note: The above method, by incorporating heterogeneous data of multiple dimensions and scales, performs comprehensive data fusion, which not only significantly improves the completeness and accuracy of pollution feature characterization, but also provides a solid and multidimensional evidence foundation for subsequent physical model inversion and intelligent early warning, and significantly enhances the system's ability to analyze and predict complex pollution scenarios.

[0010] Furthermore, the neural network model includes graph neural networks, deep residual networks, and temporal neural networks.

[0011] The method employs at least two neural network models to extract spatiotemporal features from multi-source environmental monitoring data, and uses an attention fusion mechanism to align and fuse these features, resulting in multimodal fused features of the area under study; including: Graph neural networks are used to extract topological features of pollution diffusion paths based on hydrological connectivity. Utilizing deep residual networks to extract multi-scale spatial structure features from remote sensing images; Temporal neural networks are used to capture the temporal evolution of pollutant concentrations, and the spatiotemporal correlation weights between meteorological data and pollutant concentrations are calculated through a cross-modal attention mechanism. Finally, feature alignment and fusion are performed through an attention fusion mechanism to obtain the multimodal fused features of the region under study.

[0012] Explanation: The above method integrates graph neural networks, deep residual networks, and temporal neural networks, and introduces a cross-modal attention mechanism to achieve high-precision collaborative extraction and fusion of multi-dimensional features of environmental systems. Graph neural networks accurately depict the topological relationships of pollution diffusion along hydrological pathways, deep residual networks extract multi-scale spatial structure information from remote sensing images, and temporal neural networks effectively capture the dynamic evolution of pollutant concentrations. The cross-modal attention mechanism can dynamically quantify the spatiotemporal correlation weights of meteorological and other factors on the pollution process, thereby achieving deep alignment and fusion of multi-source heterogeneous data at the physical level. This architecture not only fully extracts and unifies spatiotemporally heterogeneous features, but also significantly improves the completeness and interpretability of feature representation by coupling data-driven approaches with physical mechanisms (diffusion pathways, spatial structures, and temporal evolution), laying a reliable and efficient data foundation for subsequent accurate pollutant concentration prediction, pollution source tracing analysis, and dynamic risk assessment.

[0013] Furthermore, the step of obtaining a dataset of the spatiotemporal distribution of pollutant concentrations corresponding to the multimodal fusion features through parameter inversion methods and pollution migration numerical simulation methods such as SEAWAT, GMS, Feflow, or TOUGHREACT includes: Based on the hydrogeological parameters and pollution source parameters in the multimodal fusion features, the parameter inversion method is used to estimate the hydrogeological parameter field and the pollution source release history. Based on the hydrogeological parameter field obtained by inversion and the release history of pollution sources, a SEAWAT variable density fluid-mass coupling numerical model is constructed to simulate the dynamic interaction process of coastal groundwater pollution and seawater intrusion, and output the spatiotemporal distribution of pollutant concentration. By using the Monte Carlo sampling method, hydrogeological parameters and pollution source parameters are randomly sampled to generate multiple sets of parameter combinations as model inputs. The SEAWAT variable density fluid-mass coupling numerical model is run in batches and the corresponding pollutant concentration output results are collected to establish a dataset of the spatiotemporal distribution of pollutant concentrations corresponding to the multimodal fusion features.

[0014] Note: The above method generates a high-quality, physically consistent dataset of spatiotemporal distributions of pollutant concentrations covering a wide range of scenarios. Batch simulations generate a complete dataset encompassing multiple possible scenarios. This not only provides ample and diverse training samples for subsequent data-driven prediction models to enhance their generalization ability, but also achieves complementary advantages and deep integration between data-driven methods and physical mechanism models, laying a solid and reliable data foundation for accurate prediction and risk assessment of pollutant transport in complex hydrogeological environments.

[0015] Furthermore, the construction of the pollutant prediction model includes: A model architecture for establishing a pollutant prediction model is provided; the model architecture of the pollutant prediction model includes a convolutional neural network layer, a long short-term memory network layer, and a generative adversarial network layer. The pollutant prediction model is trained by using the dataset of the spatiotemporal distribution of pollutant concentrations corresponding to the multimodal fusion features. The hydrogeological parameters and pollution source parameters in the multimodal fusion features are used as input features, and the spatiotemporal distribution of pollutant concentrations output by the SEAWAT numerical model is used as the prediction label.

[0016] Note: The innovative advantage of the above model architecture lies in its deep integration of the spatial feature extraction capability of convolutional neural networks, the temporal dynamic capture capability of long short-term memory networks, and the high-fidelity generation capability of generative adversarial networks, thereby constructing an intelligent prediction engine that can accurately simulate the complex spatiotemporal evolution of pollutants. By directly using the dataset generated by the physical mechanism model (SEAWAT) for supervised learning, the model not only inherits the core scientific mechanism of the physical model to ensure the rationality and interpretability of the prediction results, but also achieves an exponential improvement in inference speed. It can transform the traditional numerical simulation process that takes several hours or even days into a near real-time rapid prediction, which provides key technical support for high-frequency assessment and rapid emergency early warning of dynamic environmental risks.

[0017] Further, determining the risk index of the area under study based on the comparison between the predicted spatiotemporal distribution of pollutant concentrations and the pollution concentration threshold includes: If the predicted pollutant concentration is less than the standard limit, the risk index is 0. If the standard limit is less than or equal to the predicted pollutant concentration but less than 5 times the standard limit, then the risk index is 1. If the predicted pollutant concentration is ≥ 5 times the standard limit, the risk index is 2.

[0018] Furthermore, the step of determining the warning level and issuing a warning by using at least one of the following criteria from the risk index and multimodal fusion features: sensitive receptor distribution, pollutant diffusion rate, and predicted concentration exceedance multiple; includes: The risk index is obtained by weighting and summing at least one of the following features from the multimodal fusion characteristics: sensitive receptor distribution, pollutant diffusion rate, and predicted concentration exceedance multiple. The comprehensive risk index is compared with the warning threshold to determine the warning level and issue a warning.

[0019] Note: The above method upgrades the early warning mechanism from a single concentration threshold judgment to a multi-dimensional dynamic risk assessment system. Its core advantage lies in its ability to not only obtain highly reliable concentration predictions through a physical mechanism-driven model, but also to innovatively quantify risk into an index and weight it with key dynamic elements of the environmental system (sensitive receptor distribution, diffusion rate, etc.). This allows the early warning level determination to comprehensively reflect the objective intensity of pollution (concentration exceeding the standard multiple), potential exposure risk (receptor sensitivity), and the urgency of pollution (diffusion rate), thus achieving a leap from "predicting pollution concentration" to "assessing actual hazards and emergency priorities." The resulting early warning signals are more targeted, scientific, and provide action guidance, offering direct and reliable decision-making basis for refined environmental management and emergency response.

[0020] The present invention also provides an intelligent early warning system for soil and groundwater pollution based on multimodal data fusion, which is used to implement the above method, including: a data acquisition module, a multimodal fusion module, an intelligent early warning model construction and execution module, and an early warning output and feedback optimization module.

[0021] The data acquisition module is used to acquire multi-source environmental monitoring data of the area under study. A multimodal fusion module is used to generate multimodal fusion features based on at least two neural network models and an attention fusion mechanism. The intelligent early warning model construction and execution module includes a mechanism-data fusion simulation unit and a multi-factor dynamic early warning decision unit, which are used to process the multimodal fusion features through an architecture that couples data assimilation-driven physical simulation with multi-factor dynamic risk assessment to generate early warning levels; The mechanism-data fusion simulation unit is used to obtain a dataset of the spatiotemporal distribution of pollutant concentrations corresponding to the multimodal fusion features by using parameter inversion methods and pollution migration numerical simulation methods such as SEAWAT, GMS, Feflow, or TOUGHREACT, based on the multimodal fusion features; and to construct a pollutant prediction model based on the multimodal fusion features and the dataset of the spatiotemporal distribution of pollutant concentrations; the input of the pollutant prediction model is the multimodal fusion features, and the output is the predicted spatiotemporal distribution of pollutant concentrations. The multi-factor dynamic early warning decision unit is used to determine the risk index of the area under study based on the comparison between the predicted spatiotemporal distribution of pollutant concentration and the pollution concentration threshold; then, the risk index is used as a basis for determining the early warning level and issuing an early warning by using at least one of the following characteristics from the multimodal fusion features: sensitive receptor distribution, pollutant diffusion rate, and predicted concentration exceedance multiple.

[0022] Note: The above system, through modular integrated design, integrates complex processes such as multi-source data perception, cross-modal feature fusion, mechanism-data collaborative simulation, and multi-factor dynamic decision-making into an efficient, interpretable, and automated intelligent early warning closed loop. It can realize a complete technical chain of data-driven feature extraction, physical mechanism-assured reliability, intelligent model-accelerated inference, and multi-factor comprehensive decision-making, which significantly improves the timeliness, accuracy, and decision support capabilities of pollution early warning.

[0023] The beneficial effects of this invention are: This invention combines multi-source environmental monitoring data with various neural networks, utilizing an attention mechanism to achieve efficient alignment and fusion of cross-modal spatiotemporal features, significantly improving the comprehensiveness and robustness of feature representation. Furthermore, it integrates physics-driven (SEAWAT numerical simulation and parameter inversion) and data-driven methods to construct a pollutant prediction model, enhancing the model's generalization ability and interpretability. By combining multimodal fusion features with receptor distribution, diffusion rate, etc., it achieves an intelligent decision-making closed loop from pollutant concentration prediction to multi-dimensional risk assessment and graded early warning, thereby significantly improving the accuracy, timeliness, and systematic nature of environmental risk early warning. Attached Figure Description

[0024] Figure 1 This is a schematic diagram illustrating the method principle of an embodiment of the present invention; Figure 2 This is the early warning response diagram in an embodiment of the present invention; Figure 3 This is a schematic diagram of the method framework in an embodiment of the present invention. Detailed Implementation

[0025] To further illustrate the methods and effects of this invention, the technical solution of this invention will be clearly and completely described below in conjunction with experiments.

[0026] Example 1: A smart early warning method for soil and groundwater pollution based on multimodal data fusion, such as Figure 1 As shown, it includes the following steps: S101. Obtain multi-source environmental monitoring data of the area to be studied, use at least two neural network models to extract the spatiotemporal features of the multi-source environmental monitoring data respectively, and perform feature alignment and fusion through an attention fusion mechanism to obtain the multimodal fusion features of the area to be studied. The multi-source environmental monitoring data includes meteorological data, hydrological data, remote sensing images, pollutant survey data, online monitoring data, high-risk facility leakage monitoring data, multiple fingerprint profiles, and isotope tracing data. As a preferred option, such as Figure 3 As shown, this embodiment of the invention uses a coastal industrial cluster as the area under study and establishes a multi-source heterogeneous spatiotemporal database for storing and retrieving the multi-source environmental monitoring data. Specifically, the database establishment includes the following steps: Data standardization and alignment: Using a unified coordinate system and time standard, spatial data (high-resolution remote sensing images, GIS maps), temporal data (sensor data, meteorological data), and feature data (fingerprint maps, isotopes) are spatially registered and temporally aligned to ensure consistency and usability. Data cleaning and imputation: Through an ETL process, machine learning algorithms (such as KNN) are used to imput missing values ​​and automatically remove sensor outliers. Finally, a multi-source heterogeneous spatiotemporal database covering "soil, groundwater, and surface environment" is formed, including data layers such as hydrology and meteorology, remote sensing images, pollution surveys, online monitoring, facility leakage, and multiple fingerprint maps. Those skilled in the art should understand that the multi-source heterogeneous spatiotemporal database can be a physically existing database instance, a logically unified data set, or a data cache in memory, as long as it enables unified access to multi-source data.

[0027] Obtain the multimodal fusion features of the region to be studied; including: 1) Extract topological features of pollution diffusion paths based on hydrological connectivity using graph neural networks; 2) Extracting multi-scale spatial structure features from remote sensing images using deep residual networks; 3) Utilize temporal neural networks to capture the temporal evolution of pollutant concentrations, and calculate the spatiotemporal correlation weights between meteorological data and pollutant concentrations through a cross-modal attention mechanism; 4) Finally, feature alignment and fusion are performed through an attention fusion mechanism to obtain the multimodal fusion features of the region under study.

[0028] In this embodiment of the invention, spatial / geographic data is processed using 3D convolutional neural networks / 3D U-Net, graph convolutional networks, and VisionTransformer / ResNet single-modal data feature extraction algorithms. LSTM / GRU / Temporal CNN algorithms are used to process concentration time series from single monitoring wells, capturing the natural decay and periodic fluctuations of pollutants (such as the impact of seasonal water level changes). The Transformer Encoder algorithm processes long sequences of multiple monitoring points and multiple pollutant indicators, capturing global dependencies through a self-attention mechanism to identify pollution events and complex migration patterns. Domain BERT algorithms can be further pre-trained on a large number of geological and environmental science documents and reports to understand survey reports and lithological description texts, extracting key entities (such as pollutants and strata) and semantic relationships. Simultaneously, strategies such as deep fusion based on spatiotemporal graph networks, encoder-decoder multi-task learning, multimodal pre-training + domain fine-tuning, and physically-enhanced neural networks are employed to achieve multimodal data fusion, enabling pollution source identification and migration prediction, site pollution risk assessment, historical report data mining and rapid screening of new sites, and high-precision simulation in data-scarce scenarios.

[0029] In the specific implementation process, taking a soil and groundwater pollution early warning project in a coastal industrial park as an example, the data acquisition and preprocessing involved collecting multi-source data for the study area over one year, including: a) daily time series of chlorinated hydrocarbon concentrations from 10 monitoring wells; b) monthly multispectral remote sensing images from the Gaofen-2 satellite; c) daily rainfall and temperature data from regional meteorological stations; and d) a hydrogeological structure model constructed based on geological exploration data, clarifying the distribution of aquifers and impermeable layers. Graph Neural Network (GNN) feature extraction: using monitoring wells and known potential pollution sources (such as chemical plant tank areas) as nodes, and hydraulic connections inferred from the hydrogeological model and groundwater flow direction as edges, a spatial relationship graph was constructed. The GNN was used to extract the features of each node in the topological structure, characterizing its position and connectivity in the pollution diffusion network. Deep Residual Network (e.g., ResNet) feature extraction: using a pre-trained ResNet-50 model to extract spatial features at different scales from the preprocessed remote sensing images. For example, deep features can identify large industrial facilities and land use changes, while shallow features can capture subtle texture information such as soil moisture and vegetation anomalies. Temporal neural networks (such as LSTM) and cross-modal attention: The concentration sequence from monitoring wells and the corresponding meteorological sequence are input into an LSTM-attention coupled module. The LSTM captures the intrinsic evolution of concentration (such as decay and fluctuation), while a cross-modal attention layer dynamically calculates the contribution weights of meteorological factors such as rainfall at different time steps to concentration changes, outputting a temporal feature vector that integrates key meteorological driving information. Attention fusion mechanism alignment and fusion: The topological feature vector output by the GNN, the multi-scale spatial feature vector output by the ResNet, and the temporal feature vector output by the LSTM-attention layer are concatenated and input into a fusion module consisting of a multilayer perceptron (MLP) and a self-attention layer. This module uses a self-attention mechanism to automatically learn and assign importance weights to each modality feature for the final pollution prediction task, achieving semantic and scale alignment of features. Finally, it outputs a unified, high-dimensional "multimodal fusion feature" vector, which comprehensively encodes the regional pollution diffusion structure, spatial environmental state, and temporal dynamics influenced by meteorology.

[0030] S102. Obtain the dataset of the spatiotemporal distribution of pollutant concentration corresponding to the multimodal fusion feature by using the parameter inversion method and the SEAWAT numerical simulation method; construct a pollutant prediction model based on the multimodal fusion feature and the dataset of the spatiotemporal distribution of pollutant concentration; the input of the pollutant prediction model is the multimodal fusion feature, and the output is the predicted spatiotemporal distribution of pollutant concentration. 1) Based on the hydrogeological parameters and pollution source parameters in the multimodal fusion features, the parameter inversion method is used to estimate the hydrogeological parameter field and the pollution source release history; 2) Based on the hydrogeological parameter field obtained by inversion and the release history of pollution sources, a SEAWAT variable density fluid-mass coupling numerical model is constructed to simulate the dynamic interaction process of coastal groundwater pollution and seawater intrusion, and output the spatiotemporal distribution of pollutant concentration. 3) Using the Monte Carlo sampling method, hydrogeological parameters and pollution source parameters are randomly sampled to generate multiple sets of parameter combinations as model inputs. The SEAWAT variable density fluid-mass coupling numerical model is run in batches and the corresponding pollutant concentration output results are collected to establish a dataset of the spatiotemporal distribution of pollutant concentrations corresponding to the multimodal fusion features.

[0031] A model architecture for establishing a pollutant prediction model is provided; the model architecture of the pollutant prediction model includes a convolutional neural network layer, a long short-term memory network layer, and a generative adversarial network layer. The pollutant prediction model is supervised learning using the dataset of the spatiotemporal distribution of pollutant concentrations corresponding to the multimodal fusion features. The hydrogeological parameters and pollution source parameters in the multimodal fusion features are used as input features, and the spatiotemporal distribution of pollutant concentrations output by the SEAWAT numerical model is used as the prediction label to train the model. The model training and validation are the same as the methods in the prior art, and will not be described in detail here.

[0032] For example, this embodiment of the invention takes chlorinated hydrocarbon pollution in a coastal industrial area as an example. First, it uses multimodal fusion features (such as spatial structure and temporal patterns parsed from remote sensing and monitoring data) to assist in parameter inversion, and estimates key hydrogeological parameters (such as permeability field) and historical release fluxes of pollution sources through Bayesian inversion methods. Then, based on the inversion results, a high-resolution SEAWAT variable density fluid-mass coupling numerical model is constructed to simulate the migration process of pollutants under the coupling effect of groundwater and seawater intrusion, generating high-fidelity concentration spatiotemporal distribution data. Next, the Monte Carlo method is used to probabilistically sample parameters such as permeability, dispersion, and source strength, driving SEAWAT to run in batches, thereby constructing a concentration field dataset covering different hydrogeological conditions and pollution scenarios. Finally, using this dataset as training samples, a pollutant prediction model that integrates CNN (extracts spatial distribution features), LSTM (captures temporal evolution patterns), and GAN (improves the authenticity and resolution of prediction results) is constructed and trained. Its input is multimodal fusion features (such as geological parameters and source strength features), and the output is a fast and accurate prediction result of the spatiotemporal distribution of pollutant concentration.

[0033] S103. Based on the comparison results between the predicted spatiotemporal distribution of pollutant concentration and the pollutant concentration threshold, determine the risk index of the area to be studied; then, use at least one of the following as the basis for judgment: the distribution of sensitive receptors, the pollutant diffusion rate, and the predicted concentration exceedance multiple in the multimodal fusion features, determine the warning level, and issue a warning.

[0034] The step of determining the risk index of the area under study based on the comparison between the predicted spatiotemporal distribution of pollutant concentrations and the pollution concentration threshold includes: If the predicted pollutant concentration is less than the standard limit, the risk index is 0. If the standard limit is less than or equal to the predicted pollutant concentration but less than 5 times the standard limit, then the risk index is 1. If the predicted pollutant concentration is ≥ 5 times the standard limit, the risk index is 2.

[0035] The method of determining the warning level and issuing a warning by using at least one of the following criteria from the risk index and multimodal fusion features: sensitive receptor distribution, pollutant diffusion rate, and predicted concentration exceedance multiple: includes: The risk index is obtained by weighting and summing at least one of the following features from the multimodal fusion characteristics: sensitive receptor distribution, pollutant diffusion rate, and predicted concentration exceedance multiple. The comprehensive risk index is compared with the warning threshold to determine the warning level and issue a warning.

[0036] For example, taking the early warning of chlorinated hydrocarbon pollution in groundwater in coastal industrial areas as an example, the early warning decision-making process is as follows: First, the pollutant concentration field (e.g., 1,2-dichloroethane concentration) output by the prediction model for the next 72 hours is compared grid by grid with the limit value in the "Groundwater Quality Standard" (e.g., 30 μg / L). According to the rules (e.g., <30 is 0, 30-150 is 1, ≥150 is 2), a basic risk index spatial distribution map is generated. Then, a comprehensive risk index is constructed: the distribution of sensitive receptors is extracted (e.g., residential areas within 1 km downstream are assigned a weight of 0.4, and schools are assigned a weight of 0.6), and their spatial superposition risk value is calculated; combined with the pollutant diffusion rate (calculated based on groundwater flow velocity and pollutant attenuation coefficient, a faster rate results in a higher risk bonus, assigned a weight of 0.3) and the predicted concentration exceeding the standard multiple (e.g., the predicted concentration in a certain area is 180 μg / L, exceeding the standard by 6 times, assigned a weight of 0.3). The risk index is calculated using a weighted formula (e.g., Comprehensive Risk Index = Basic Risk Index × 0.4 + Receptor Risk Value × 0.3 + Diffusion Rate Coefficient × 0.2 + Exceedance Multiple Coefficient × 0.1). Finally, the warning level is automatically determined based on the comprehensive risk index thresholds (e.g., 0-0.5 is Blue Concern Level, 0.5-1.2 is Yellow Alert Level, 1.2-2.0 is Orange Alert Level, and >2.0 is Red Emergency Level). The platform then disseminates tiered warning information and protective recommendations to the relevant environmental management personnel and affected communities.

[0037] In some embodiments, the determination is made directly based on a multi-factor early warning rule matrix that combines the pollutant exceedance multiple (Pi) and the concentration change rate multiple (Δx); as shown in Table 1 below: Table 1 Early Warning Matrix Determination

[0038] In Table 1, Pi represents the multiple of the pollutant relative to the Class IV standard value; Δx represents the ratio of the current data to the previous monitoring data.

[0039] The advantages of the embodiments of the present invention are as follows: ① Significantly enhanced early warning capabilities: Early warning accuracy has increased from approximately 70% in traditional models to over 80%, reducing false alarms and missed alarms. ② More rapid emergency response: Emergency response activation time has been shortened by 50%, gaining crucial time for pollution control. ③ Support for scientific decision-making and management: Visualized risk maps and trend forecasts are provided to guide the optimization of monitoring sites, evaluation of remediation plans, and determination of remediation priorities. ④ Achieving a paradigm shift in management: Environmental risk management is being transformed from a "passive monitoring and post-event handling" model to a new intelligent paradigm of "proactive early warning and pre-event prevention," ensuring sustainable regional development.

[0040] In summary, this embodiment of the study constructed a complete intelligent early warning model for soil and groundwater pollution through multimodal data fusion and artificial intelligence technology. This model possesses multiple functions, including data integration, feature fusion, migration simulation, risk assessment, and dynamic early warning, which can significantly improve the intelligence level of environmental risk management and provide a solid environmental security guarantee for regional sustainable development.

Claims

1. A method for intelligent early warning of soil and groundwater pollution based on multimodal data fusion, characterized in that, Includes the following steps: Multi-source environmental monitoring data of the area under study is obtained. At least two neural network models are used to extract the spatiotemporal features of the multi-source environmental monitoring data. Feature alignment and fusion are performed through an attention fusion mechanism to obtain the multimodal fusion features of the area under study. A dataset of the spatiotemporal distribution of pollutant concentrations corresponding to the multimodal fusion features is obtained by using parameter inversion and pollution migration numerical simulation methods. Based on the multimodal fusion features and the dataset of the spatiotemporal distribution of pollutant concentrations, a pollutant prediction model is constructed. The input of the pollutant prediction model is the multimodal fusion features, and the output is the predicted spatiotemporal distribution of pollutant concentrations. Based on the comparison between the predicted spatiotemporal distribution of pollutant concentration and the pollution concentration threshold, the risk index of the area under study is determined; then, the risk index is combined with at least one of the following in the multimodal fusion features: sensitive receptor distribution, pollutant diffusion rate, predicted concentration, and exceedance multiple, to determine the warning level and issue a warning.

2. The intelligent early warning method for soil and groundwater pollution based on multimodal data fusion as described in claim 1, characterized in that, The multi-source environmental monitoring data includes meteorological data, hydrological data, remote sensing images, pollutant survey data, online monitoring data, high-risk facility leakage monitoring data, multiple fingerprint profiles, and isotope tracing data.

3. The intelligent early warning method for soil and groundwater pollution based on multimodal data fusion as described in claim 1, characterized in that, The neural network model is at least one of graph neural network, deep residual network and temporal neural network, and the pollution migration numerical simulation method is at least one of SEAWAT, GMS, Feflow or TOUGHREACT.

4. The intelligent early warning method for soil and groundwater pollution based on multimodal data fusion as described in claim 3, characterized in that, The method employs at least two neural network models to extract spatiotemporal features from multi-source environmental monitoring data, and uses an attention fusion mechanism to align and fuse these features, resulting in multimodal fused features of the area under study; including: Graph neural networks are used to extract topological features of pollution diffusion paths based on hydrological connectivity. Utilizing deep residual networks to extract multi-scale spatial structure features from remote sensing images; Temporal neural networks are used to capture the temporal evolution of pollutant concentrations, and the spatiotemporal correlation weights between meteorological data and pollutant concentrations are calculated through a cross-modal attention mechanism. Finally, feature alignment and fusion are performed through an attention fusion mechanism to obtain the multimodal fused features of the region under study.

5. The intelligent early warning method for soil and groundwater pollution based on multimodal data fusion as described in claim 1, characterized in that, The process involves obtaining a dataset of the spatiotemporal distribution of pollutant concentrations corresponding to the multimodal fusion features through parameter inversion and pollution migration numerical simulation methods; including: Based on the hydrogeological parameters and pollution source parameters in the multimodal fusion features, the parameter inversion method is used to estimate the hydrogeological parameter field and the pollution source release history. Based on the hydrogeological parameter field obtained by inversion and the release history of pollution sources, a SEAWAT variable density fluid-mass coupling numerical model is constructed to simulate the dynamic interaction process of coastal groundwater pollution and seawater intrusion, and output the spatiotemporal distribution of pollutant concentration. By using the Monte Carlo sampling method, hydrogeological parameters and pollution source parameters are randomly sampled to generate multiple sets of parameter combinations as model inputs. The SEAWAT variable density fluid-mass coupling numerical model is run in batches and the corresponding pollutant concentration output results are collected to establish a dataset of the spatiotemporal distribution of pollutant concentrations corresponding to the multimodal fusion features.

6. The intelligent early warning method for soil and groundwater pollution based on multimodal data fusion as described in claim 4, characterized in that, The construction of the pollutant prediction model includes: A model architecture for establishing a pollutant prediction model is provided; the model architecture of the pollutant prediction model includes a convolutional neural network layer, a long short-term memory network layer, and a generative adversarial network layer. The pollutant prediction model is trained using the dataset of the spatiotemporal distribution of pollutant concentrations corresponding to the multimodal fusion features. The hydrogeological parameters and pollution source parameters in the multimodal fusion features are used as input features, and the spatiotemporal distribution of pollutant concentrations output by the SEAWAT numerical model is used as the prediction label.

7. The intelligent early warning method for soil and groundwater pollution based on multimodal data fusion as described in claim 1, characterized in that, The step of determining the risk index of the area under study based on the comparison between the predicted spatiotemporal distribution of pollutant concentrations and the pollution concentration threshold includes: If the predicted pollutant concentration is less than the standard limit, the risk index is 0. If the standard limit is less than or equal to the predicted pollutant concentration but less than 5 times the standard limit, then the risk index is 1. If the predicted pollutant concentration is ≥ 5 times the standard limit, the risk index is 2.

8. The intelligent early warning method for soil and groundwater pollution based on multimodal data fusion as described in claim 1, characterized in that, The method of determining the warning level and issuing a warning by using at least one of the following criteria from the risk index and multimodal fusion features: sensitive receptor distribution, pollutant diffusion rate, and predicted concentration exceedance multiple: includes: The risk index is obtained by weighting and summing at least one of the following features from the multimodal fusion characteristics: sensitive receptor distribution, pollutant diffusion rate, and predicted concentration exceedance multiple. The comprehensive risk index is compared with the warning threshold to determine the warning level and issue a warning.

9. A smart early warning system for soil and groundwater pollution based on multimodal data fusion, used to implement the method described in any one of claims 1 to 8, characterized in that, include: The system includes a data acquisition module, a multimodal fusion module, an intelligent early warning model construction and execution module, and an early warning output and feedback optimization module. The data acquisition module is used to acquire multi-source environmental monitoring data of the area under study. A multimodal fusion module is used to generate multimodal fusion features based on at least two neural network models and an attention fusion mechanism. The intelligent early warning model construction and execution module includes a mechanism-data fusion simulation unit and a multi-factor dynamic early warning decision unit, which are used to process the multimodal fusion features through an architecture that couples data assimilation-driven physical simulation with multi-factor dynamic risk assessment to generate early warning levels; The mechanism-data fusion simulation unit is used to obtain a dataset of the spatiotemporal distribution of pollutant concentrations corresponding to the multimodal fusion features through parameter inversion and SEAWAT numerical simulation methods, based on the multimodal fusion features; and to construct a pollutant prediction model based on the multimodal fusion features and the dataset of the spatiotemporal distribution of pollutant concentrations; the input of the pollutant prediction model is the multimodal fusion features, and the output is the predicted spatiotemporal distribution of pollutant concentrations. The multi-factor dynamic early warning decision unit is used to determine the risk index of the area under study based on the comparison between the predicted spatiotemporal distribution of pollutant concentration and the pollution concentration threshold; then, the risk index is used as a basis for determining the early warning level and issuing an early warning by using at least one of the following characteristics from the multimodal fusion features: sensitive receptor distribution, pollutant diffusion rate, and predicted concentration exceedance multiple.

Images