Water pollution risk early warning and tracing method based on multi-source data fusion

Abstract

A water pollution risk early warning and tracing method based on multi-source data fusion belongs to the technical field of water pollution monitoring and early warning, and comprises the following steps: step 1, constructing a multi-source heterogeneous data acquisition network and realizing real-time data transmission; 2, performing multi-source data fusion and feature enhancement processing based on space-time reference; 3, constructing a water pollution risk dynamic early warning system based on a WOA-LSSVM model; fourthly, reverse positioning of the pollution source is completed on the basis of a CNN-GRU-SE Attention model; and 5, carrying out development and emergency response on a multi-dimensional visual decision support system, and positioning a pollution source. Multi-source information such as water quality sensor data, unmanned aerial vehicle image data and geographic information data is fused, and an intelligent monitoring network is constructed; through multi-source data fusion and an intelligent algorithm, water pollution risk early warning accuracy and traceability efficiency are effectively improved, and the method has the advantages of high response speed, wide monitoring range, accurate positioning and the like, can be widely applied to the fields of urban water supply, drainage basin management and the like, and meets the requirements of water environment safety guarantee.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of water pollution monitoring and early warning, and relates to a water pollution risk early warning and real-time monitoring system based on multi-source data fusion and intelligent algorithm driving and application thereof. BACKGROUND

[0002] With the rapid development of industrialization and urbanization, water pollution has gradually become a major challenge to the ecological environment and human health. Traditional water quality monitoring technology usually relies on manual sampling and laboratory analysis, which can obtain high-precision data, but due to its poor timeliness, limited spatial coverage and high cost, it is difficult to meet the real-time dynamic monitoring needs of large-scale water areas. Therefore, it is urgent to develop a new water pollution monitoring technology to cope with the limitations of the current water quality monitoring system.

[0003] In recent years, remote sensing technology as a new water quality monitoring method has been widely used in water pollution detection. Water spectral data such as chlorophyll, turbidity and suspended matter concentration obtained by satellite or unmanned aerial vehicle inspection can provide certain help for water quality monitoring. However, the spatial resolution of remote sensing technology is still limited, and its response to sudden water pollution incidents is lagging, which cannot achieve accurate positioning of the pollution source. In addition, the existing water quality monitoring system relies on a single data source and lacks deep fusion of multi-source data, resulting in the shortcomings of the existing early warning model in accuracy, timeliness and pollution source tracing ability.

[0004] The existing water pollution risk early warning and tracing research faces multiple technical bottlenecks. First, the fusion of multi-source data is insufficient. The existing sensor data, remote sensing images and geographic information have not been effectively integrated, making it difficult to comprehensively depict the pollution dynamics and its spatial distribution characteristics

JIANG F, WU Q X, LI Q, et al. Integrated quantitative tracing for Karst groundwater contamination: A case study of landfill in Zunyi, Guizhou Province, China [J]. ENVIRONMENTAL POLLUTION, 2025, 368

MOHAN K, LAKSHMANAN VR. A critical review of the recent trends in source tracing of microplastics in the environment [J]. ENVIRONMENTAL RESEARCH, 2023, 239

[0005] Therefore, in order to effectively solve the above problems, it is urgent to develop a water pollution risk early warning and tracing real-time monitoring system based on multi-source data fusion and intelligent algorithm driving. The system should be able to integrate various water quality monitoring data (including sensor data, remote sensing image data and geographic information data), and improve the data fusion and analysis ability through intelligent algorithms. At the same time, with the help of GIS technology, the system can accurately depict the spatio-temporal distribution characteristics of water pollution, realize dynamic early warning of water pollution risk, and improve the accuracy and efficiency of pollution source positioning through advanced data-driven tracing technology, thereby providing reliable technical support for environmental governance and water pollution emergency response. SUMMARY

[0006] This invention aims to address the shortcomings of existing water pollution risk early warning and source tracing systems in terms of multi-source data fusion, precise pollution source location, and emergency response speed. Traditional water pollution early warning systems often rely on a single data source, resulting in low pollution prediction accuracy, inaccurate pollution source location, and slow emergency response. By combining sensor data, remote sensing imagery, and laboratory testing data, this invention provides a water pollution risk early warning and source tracing system based on multi-source data fusion. This system not only improves the accuracy of pollution risk prediction and pollution source location but also accelerates environmental emergency response, thereby effectively supporting water pollution prevention and environmental governance.

[0007] To achieve the above technical objectives, the technical solution of the present invention is as follows:

[0008] A method for water pollution risk early warning and source tracing based on multi-source data fusion includes the following steps:

[0009] Step 1: Construct a multi-source heterogeneous data acquisition network and achieve real-time data transmission;

[0010] Within the target water area, a wireless sensor network is strategically deployed based on factors such as the water's geographical environment and water flow characteristics. Water quality parameter data is collected in real-time through this wireless sensor network. Let the set of water quality parameters be Q = {q1, q2, ..., qn}, where qi represents the i-th parameter, i = 1, 2, ..., n, such as pH and dissolved oxygen content. Simultaneously, satellite remote sensing imagery of the water body is acquired using satellite remote sensing technology. Unmanned aerial vehicles (UAVs) are used for water body inspection, obtaining UAV inspection data containing information such as the water body's spectral characteristics. Let the set of data acquired by satellite remote sensing and UAV inspection be S. Furthermore, water samples are collected and tested in a laboratory, obtaining a set of laboratory test data L. The water quality parameter data Q collected by the wireless sensor network, the satellite remote sensing and UAV inspection data S, and the laboratory test data L are integrated to obtain a multi-source data set (Q, S, L) of water quality parameters, satellite remote sensing and UAV inspection data, and laboratory test data. This multi-source data set is then input into the system, providing a comprehensive data foundation for subsequent processing.

[0011] Step 2: Multi-source data fusion and feature enhancement processing based on spatiotemporal benchmarks;

[0012] Based on the ArcGIS platform, spatial registration is performed on the multi-source data (Q, S, L) acquired in the first step. Spatial registration unifies data from different sources and coordinate systems to the same geographic spatial coordinate system, ensuring data consistency in geographic space. After spatial registration, data cleaning algorithms are used to process the multi-source data, removing noisy and outlier data. Let the cleaned datasets be Qclean, Sclean, and Lclean. Then, semantic segmentation algorithms are used to process the cleaned data, extracting effective features to form a dataset F with spatiotemporal fusion characteristics. Data set F is a water quality feature dataset unified under the same spatiotemporal benchmark, integrating real-time monitoring parameters from Qclean, remote sensing inspection features from Sclean, and laboratory test results from Lclean. The effective features include time-series feature data. This refers to the timestamp information of data collection and the changing trends of water quality parameters over time. Spatial characteristic data. This refers to the latitude and longitude coordinates of each monitoring point after spatial registration on the ArcGIS platform. This data can be directly used as standardized input data for subsequent water quality analysis, integrating information from both temporal and spatial dimensions to provide high-quality data support for later analysis.

[0013] Step 3: Construct a dynamic early warning system for water pollution risks based on the WOA-LSSVM model;

[0014] Step 3.1: First, a Least Squares Support Vector Machine (LSSVM) model is constructed to learn the mapping relationship between the water element dataset and migration rates. The LSSVM model is represented as follows:

[0015] (1)

[0016] in, This indicates that for the input The predicted migration amount; Represents the weight vector; This represents a kernel function mapping, used to map input data to a high-dimensional feature space; This indicates the bias term.

[0017] To achieve high-precision prediction, the LSSVM model minimizes the prediction error through a loss function, which is expressed as:

[0018] (2)

[0019] in, Indicates the actual migration amount; γ represents the bias term; γ represents the penalty coefficient, used to control the degree of penalty applied to the slack variables. Indicates the number of samples; Indicates the concentration of the target element in the water body; Represents the concentration gradient; Indicates the diffusion coefficient; This represents the second partial derivative of concentration; Represents the regularization parameter. This indicates the speed of water flow.

[0020] The optimization objective of the LSSVM model is to minimize the model parameter complexity and prediction error, where the error is expressed as:

[0021] (3)

[0022] The optimization objective aims to minimize the model parameters. and This allows the predicted migration values ​​of elements in water to approximate the actual migration values ​​as closely as possible, thereby revealing the nonlinear laws governing the migration behavior of elements in water.

[0023] Step 3.2: To improve the prediction accuracy and reduce the computational complexity of the LSSVM model, the Whale Optimization Algorithm (WOA) is introduced to optimize the LSSVM model parameters. WOA optimizes parameters by simulating whale predation behavior; its core principle is to approximate the global optimum by updating the individual whale positions. The specific update rule is as follows: when a whale randomly selects a companion for searching, i.e., p < 0.5 and | When |≥1, the position update formula is:

[0024] (4)

[0025] in, Indicates algorithm control parameters; Indicates the individual ID; Indicates the spatial dimension number; Indicates the first The individual in the first Position in 3D space; This indicates that the location of a companion is randomly selected; This indicates the relationship between the current individual and randomly selected peer individuals at the [number]th ... The distance in the dimension, where p represents the position.

[0026] When the whale moves closer to the globally optimal position (i.e., p < 0.5 and |A| < 1), the position update formula is:

[0027] (5)

[0028] in, This indicates that the best individual in the current group is at the 1st position. Position in 3D space; This indicates the difference between the current individual and the optimal individual at the 1st rank. Distance on the dimension.

[0029] When simulating spiral ascent predation behavior, i.e., p > 0.5, the position update formula is:

[0030] (6)

[0031] in, Indicates the individual ID; This represents the distance between an individual whale and the globally optimal position; , Here is the spiral parameter. The formula for calculating the distance between an individual whale and its random companions or the global optimum is:

[0032] (7)

[0033] (8)

[0034] in, Indicates the algorithm control parameters. This indicates the relationship between the current individual and randomly selected peer individuals at the [number]th ... Distance in dimensional space; This indicates the difference between the current individual and the optimal individual at the 1st rank. Distance on the dimension.

[0035] To adapt to the migration and diffusion characteristics of elements in water, control parameters are introduced into the WOA calculation. These parameters are optimized in conjunction with the estimated diffusion coefficient and flow velocity, as shown in the following formula:

[0036] (9)

[0037] (10)

[0038] (11)

[0039] in, , These are the initial values ​​for the diffusion coefficient and the flow velocity, respectively. , , For adjustment coefficients; This represents the current iteration number. This represents the estimated diffusion coefficient. This represents the estimated diffusion coefficient. Indicates control parameters; Indicates control parameters The initial value.

[0040] At the same time, the search strategy is dynamically adjusted through parameters:

[0041] (12)

[0042] (13)

[0043] in, This represents the current iteration number; This represents the maximum number of iterations. This indicates that the parameters are adapted to migration and diffusion characteristics, enhancing the search capability in complex regions; This indicates the convergence speed of parameter optimization, gradually approaching the optimal solution.

[0044] When the maximum number of iterations is reached After that, the algorithm terminates, and the solution corresponding to the global optimal position is the optimal parameter combination of the LSSVM model. At this point, the model parameters are minimized, and the computation time is significantly reduced.

[0045] Step 3.3: Utilize the dataset with spatiotemporal fusion characteristics obtained in step 2. The optimized WOA-LSSVM model was trained.

[0046] Data set Time series data in It can be represented as .in Represents the total number of samples in the time series. For time steps The input data, including dimensions such as temperature, pH, permanganate concentration, total ammonia concentration, and conductivity, all came from a dataset. Feature extraction results; The target value at the corresponding time step represents the amount of element migration in the water. During training, the dataset... The model fitting direction is dynamically adjusted based on the spatiotemporal characteristics, and the model parameters are continuously optimized until the prediction error meets the preset threshold.

[0047] After the model is trained, the pollutant concentration is fused and predicted by combining the mechanistic model, as shown in the following formula:

[0048] (14)

[0049] in, Preliminary concentration distribution calculated for the mechanistic model; The prediction results are from the WOA-LSSVM model. These are weighting coefficients; This indicates the final concentration distribution result.

[0050] Leveraging cloud computing capabilities, the data set collected in real time and processed through steps 2 and 3 is analyzed. The analysis is performed, and the water pollution risk early warning results are output. The early warning level is determined based on the predicted pollutant concentration sequence for future time steps, as shown in formula (15):

[0051] (15)

[0052] in, Indicates the current time step; Indicates a future time step; Represents a sequence of predicted values; Different levels of warning thresholds ( The threshold This represents the upper limit of the Class III standard for surface water, signifying the critical value for slight pollution. =10 mg / L, which is the upper limit of the Class IV standard for surface water, representing the critical value for moderate pollution; This represents the upper limit of the Class V standard for surface water, signifying the critical value for severe pollution.

[0053] The warning range is determined based on the predicted pollutant concentration at a spatial point, as shown in formula (16):

[0054] (16)

[0055] in, Target waters; The spatial early warning threshold ranges from 60% to 80%. Represents the x-coordinate of a point in space; Represents the ordinate of a point in space; This represents the predicted value at a spatial point. Through the above steps, the dynamic early warning system for water pollution risks is constructed, enabling real-time prediction, level determination, and extent delineation of water pollution.

[0056] Step 4: Complete the reverse localization of pollution sources based on the CNN-GRU-SE Attention model;

[0057] Step 4.1, based on the dataset with spatiotemporal fusion characteristics obtained in step 2. First, we construct the CNN-GRU-SE Attention model. The model construction process is as follows. This model is a combined architecture of temporal feature extraction and spatial feature extraction. It achieves multi-dimensional feature fusion through a three-layer structure. The construction of each module and the formulas are as follows:

[0058] The first layer is the CNN module, which is a spatial feature extraction module used to capture local correlation features between spatial points. The construction formula is as follows:

[0059] (17)

[0060] (18)

[0061] in, The convolution kernel represents the core parameters of the model, used to extract spatial features; Indicates the bias term; ReLU represents the convolution operation; ReLU represents the activation function. This represents the max pooling operation; Indicates input data; This represents the output features of the convolutional layer; This represents the features after pooling.

[0062] The second layer is the GRU module, which is a time feature extraction module. It captures long-term dependencies in time series through a gating mechanism. The construction formula is as follows:

[0063] (19)

[0064] (20)

[0065] (twenty one)

[0066] (twenty two)

[0067] in, , , These are weight matrices representing the reset gate, update gate, and hidden state, respectively, which are used to construct the core parameters of the model. express The input feature vector at time step; , , These represent the corresponding bias terms; The circle represents the sigmoid activation function; ⊙ represents element-wise multiplication. Indicates that the door is being reset; Indicates an update to the door; express The hidden state at any given moment; Indicates the candidate hidden state; for The final hidden state of the moment.

[0068] The third layer is the SE-Attention module, which is a key feature enhancement module. It highlights the contribution of important features through attention weights, and the construction formula is as follows:

[0069] (twenty three)

[0070] (twenty four)

[0071] in, This is the attention weight matrix; For bias terms; This is a global average pooling operation; The sigmoid activation function is used; ⊙ represents element-wise multiplication. Attention weights; Indicates the first [module name] extracted by the preceding module. 1 eigenvector; These are the fusion features after attention weighting.

[0072] The three modules described above together constitute the CNN-GRU-SE Attention model, which is used to extract key information in the spatiotemporal and feature dimensions from multi-source data. After the model is built using the above steps, the dataset formed in the second step is then... Data set as model input Represented as ,in This is a spatiotemporal fusion dataset; In terms of time dimension, a total One time step; For spatial dimensions, a total One spatial point; For feature dimensions, a total of Water quality-related characteristics, such as dissolved oxygen and pH.

[0073] Step 4.2: Apply the constructed CNN-GRU-SE Attention model to the input dataset. Processing: The spatial features are obtained after pooling by the CNN module. After processing by the GRU module, the hidden state with fused temporal features is obtained. After processing by the SE-Attention module, the attention-weighted fused features are obtained. This ultimately forms a spatiotemporal feature data set. ={ , ,…, },in for The fusion features at any given time include spatial distribution, temporal variation, and key feature weight information.

[0074] Then, the time series feature data obtained in the second step are used. and spatial feature data The constructed CNN-GRU-SE Attention model is trained. Training samples are drawn from the dataset. The feature extraction results, input is The spatiotemporal feature sequence in { The output is the corresponding historical pollution source location label. During training, the dataset is used... The spatiotemporal characteristics are dynamically adjusted to adjust the model parameters, i.e., the convolution kernel. Size, GRU weight matrix / / Attention weight matrix The goal is to minimize the mean squared error, and the model is continuously optimized until the prediction accuracy meets a preset threshold of 0.05. After the dataset model is trained, a reverse diffusion model is used to simulate the diffusion path of pollution in water bodies, enabling reverse tracing of pollution sources. The reverse diffusion model is constructed based on the following formula.

[0075] The fundamental equation for reverse diffusion:

[0076] (25)

[0077] in, Indicates pollutant concentration; Indicates time; Indicates the diffusion coefficient; Indicates the speed of water flow; Represents the Laplace operator; Represents the gradient operator; This indicates the source term, i.e., the intensity of pollutant release.

[0078] To meet computational requirements, the basic equations are discretized, resulting in the discretized reverse diffusion equations:

[0079] (26)

[0080] in, express Time and Space Point The concentration of pollutants; Indicates the concentration at time t; Indicates the time step; express The water flow velocity at that location. This equation allows for the inverse calculation of historical concentrations from the current concentration, enabling the reverse tracing of diffusion paths.

[0081] Step 4.3, combining hydrological information of the target water area Information on the distribution of sewage outlets The spatiotemporal feature set based on the output of the CNN-GRU-SEAttention model Based on the results of the reverse diffusion model, multi-source information fusion localization is achieved through four parts of calculation:

[0082] The first part, the source term estimation formula based on spatiotemporal characteristics, is as follows:

[0083] (27)

[0084] in, Represents attention weights over the time dimension; express time The spatiotemporal fusion characteristic value of a location reflects the degree of correlation between that location and the pollution source; Represents the total sequence of spatiotemporal features; It represents a specific moment in a spatiotemporal feature sequence.

[0085] The second part, the formula for calculating hydrological information for hydrological impact correction, is as follows:

[0086] (28)

[0087] in, This indicates the average flow velocity of the target water area; This indicates the degree of influence of the adjustment coefficient, i.e., the deviation from the control flow rate; express The Euclidean distance between the actual flow velocity and the average flow velocity is used. The closer the flow velocity is to the average flow velocity, the higher the weight, thus reducing the interference of abnormal water flow on the positioning. It indicates hydrological information.

[0088] The third part, prioritization of sewage outlets, i.e., the formula for screening potential pollution sources, is as follows:

[0089] (29)

[0090] in, for Distance to the nearest sewage outlet; This is the sensitivity coefficient, which controls the impact of distance on the degree of association. The closer the distance, the higher the weight. Information on the distribution of sewage outlets.

[0091] Part Four, the formula for integrating the positioning results is as follows:

[0092] (30)

[0093] in, This indicates the inversion of the reverse diffusion model. time The concentration at a given point reflects the historical diffusion path; λ represents the time decay factor, 0 < λ < 1, which weakens the interference of long-term concentrations. Indicates hydrological weight; Indicates the prior weight of the sewage outlet; This indicates the location result.

[0094] The final output is a source location map of the pollution source, which is normalized to determine the possible location of the pollution source.

[0095] (31)

[0096] in, This represents the normalized location probability, with values ​​ranging from 0 to 1, and a total of 1. The spatial coordinates with the highest probability of location are used, which are the most likely locations of the pollution source, to obtain the pollution source location results; The index variable represents all spatial locations, corresponding to the coordinates of each grid cell within the water area. Represents all spatial coordinates In the middle, looking for The coordinates corresponding to the maximum value .

[0097] Step 5: Development of a multi-dimensional visualization decision support system and emergency response;

[0098] Step 5.1: Based on the secondary development interface of the open-source geographic information platform QGIS and the ECharts visualization library, a multi-dimensional visualization decision support platform was developed, named the Water Environment Intelligent Management and Control Visualization Platform. This platform integrates three core visualization modules to achieve intuitive display and real-time push of pollution-related data. Its underlying map rendering engine reuses the geospatial data processing kernel of QGIS, and the chart rendering function is developed based on the open-source components of ECharts.

[0099] The spatiotemporal dynamic visualization module is built based on QGIS's map projection and layer management functions, reusing its WGS84 coordinate system adaptation and spatial element rendering interface. It integrates the water pollution risk warning results obtained in the third step (including the warning level W and warning range A) with the pollution source location results obtained in the fourth step (including the location probability). and core position This is mapped to a geospatial coordinate system. The heatmap is then displayed by calling ECharts' heatmap component. The spatial distribution of the location is such that darker colors indicate a higher probability of localization, based on a reverse diffusion model. The inversion results use dynamic vector arrows to mark the pollutant diffusion path and use polygons of different colors to divide the warning range A, namely, green for the safe zone, yellow for the attention zone, orange for the warning zone, and red for the danger zone.

[0100] The data visualization module directly reuses ECharts' line chart, bar chart, and radar chart components. It interfaces with the system's data interface through a custom data adapter, displaying key water quality parameters in real time, such as dissolved oxygen and pH values, as well as other characteristic indicators from the dataset. Radar charts present pollution source location indicators, including source terms. Hydrological weight Sewage outlet inspection The contribution level helps managers quickly grasp core data.

[0101] The response status visualization module is developed based on the Vue framework and the Element UI component library. It reuses ElementUI's flowchart component and status label component to display the emergency response progress, namely, early warning triggering → level determination → measure execution → result feedback. The status labels update the status of each stage in real time, including started, in progress, and completed. The statistical charts are implemented based on the ECharts component to ensure that the handling process is traceable.

[0102] The above modules work in conjunction with the preceding steps system through a cloud data interface to synchronize processed water quality data, early warning results, and location information in real time. These data are then pushed to environmental management personnel in the form of charts, maps, etc., providing intuitive data support for decision-making.

[0103] Step 5.2 integrates a policy recommendation generation module into the intelligent water environment management and visualization platform. This module is built on the DeepSeek open-source policy knowledge base framework, reusing its core semantic matching algorithm code and supplementing it with a corpus of documents and laws and regulations from the Ministry of Environmental Protection to form a dedicated knowledge base. It extracts key features of current pollution events, such as pollutant type, impact range, and pollution source type, from early warning results. and positioning results The module uses the DeepSeek semantic matching algorithm to match relevant clauses from the knowledge base, such as clauses on sewage outlet management and emergency response in the "Water Pollution Prevention and Control Law of the People's Republic of China," and generates targeted policy recommendations based on event characteristics. Simultaneously, the module reuses the HTTPS protocol and the OAuth2.0 open-source authentication component to implement API encryption, ensuring the security of data calls and transmissions for the clauses referenced in the policy recommendations and preventing the leakage of sensitive information.

[0104] When the warning result output in the third step meets the triggering conditions, the system automatically initiates an emergency response. The triggering condition formula is as follows:

[0105] When Trigger=1, and (32)

[0106] Where Trigger represents the trigger condition; For the future The predicted pollutant concentration sequence at each time step; T1 is the first-level warning threshold. This represents the normalized localization probability. The location warning threshold is preset to 0.6, meaning that when the location probability is ≥60%, it is judged as a potential pollution source. When Trigger=1, it is judged as a pollution event and enters the level classification process.

[0107] Based on the DeepSeek event classification model, its pre-trained model weights are reused and fine-tuned based on historical pollution event data. The characteristics of pollution events, including pollutant concentration, diffusion rate, and impact range, are analyzed, and event levels are output. The level classification formula is as follows:

[0108] (33)

[0109] in, This represents the pollutant diffusion rate retrieved based on the reverse diffusion model; This indicates the actual area affected by the pollution incident; core water areas and ecologically sensitive areas are preset geographical areas, reusing the spatial data storage format of QGIS.

[0110] Step 5.3 Pollution Level Classification. Level 1, light pollution, is triggered when a single pollution indicator approaches the warning threshold, such as dissolved oxygen at 3.2-3.5 mg / L. It is predicted that local water areas may be affected in the next 6-12 hours. Countermeasures include using the intelligent water environment management and visualization platform to call the sensor control interface, automatically instructing the sensor data collection interval to be shortened to 2 minutes (from 10 minutes), increasing the frequency of drone inspections to once per hour (from once every 4 hours), initiating sewage outlet flow monitoring, integrating data with the real-time updated version of the dataset, and combining satellite remote sensing imagery with drone inspection data to identify suspected pollution sources; and sending alerts to management personnel via SMS and email open-source interfaces through the platform, simultaneously displaying the warning area and the location of potential pollution sources.

[0111] Level 2 moderate pollution is triggered by the exceeding of certain indicators. These indicators refer to two key water quality parameters exceeding the corresponding warning threshold, such as either permanganate concentration or total ammonia concentration exceeding the standard, or one significantly exceeding the standard and the other approaching the threshold, with the exceedance ranging from 10% to 30%, without causing substantial damage to the aquatic ecosystem. For example, dissolved oxygen ≤ 3.2 mg / L. It is predicted that pollution may spread to the core water area within the next 3-6 hours. Response measures include water source control: remotely closing secondary water intakes and activating backup water supply via the water environment intelligent management and visualization platform's intake control API, restricting non-essential water use; deploying 50-100 meter floating pollution barriers, with locations defined based on warning range A, and adding adsorbent material at a rate of 10 kg per hour, with the placement points determined based on location results; mobilizing emergency response teams, dispatching a 3-5 person on-site working group to obtain real-time location results and spread paths through the platform, and conducting on-site verification of pollution sources.

[0112] Level 3 severe pollution is triggered by multiple indicators exceeding standards simultaneously. These multiple indicators refer to three or more core water quality parameters, including temperature, pH, conductivity, permanganate concentration, and total ammonia concentration, exceeding the warning threshold, or at least one indicator exceeding the standard by more than 50%, posing a serious threat to the aquatic ecosystem, drinking water safety, or the surrounding environment; for example, dissolved oxygen ≤3.0 mg / L corresponds to a maximum concentration of 4.0 mg / L. The forecast predicts threats to drinking water safety or ecologically sensitive areas within the next 1-3 hours. Response measures include ensuring water supply: closing the main water intake, activating a 50m³ / h emergency purification system, and monitoring the system's operational status and purified water quality via a water environment intelligent management and visualization platform. Pollution prevention is based on location data. The target pipeline was located, and a smart valve was activated to block the suspected sewage discharge. Sealant was delivered by drone, covering 10 square meters each time, with delivery points dynamically planned by the platform. At least five responsible personnel, including those from environmental protection, water resources, and emergency response departments, were notified, and a cross-departmental video conference was initiated using the platform's open-source video conferencing component. Simultaneously, warning information was disseminated through community broadcasts and government apps, advising the public to reduce contact with the contaminated water body.

[0113] The system constructs a closed loop of "response-feedback-optimization". During emergency response, all operation records, including the initiation time of measures, the amount of resources deployed, and the results of on-site verification, are synchronously archived to the water environment intelligent management and visualization platform through the government emergency platform interface. This interface reuses the HTTPS protocol and OAuth2.0 authentication component. After the response is completed, the actual pollution elimination effect, which includes the time for water quality to recover to the safe threshold and the pollutant removal rate, is fed back into the DeepSeek training set to optimize the grading weights of the event classification model. Based on historical response data, the platform calls the ECharts statistical component to generate a response efficiency analysis report. This report provides a basis for subsequent emergency resource allocation and measure optimization.

[0114] Through the above process, the intelligent water environment management and visualization platform reuses existing platforms and open-source code such as QGIS, ECharts, Vue, and Element UI, and customizes core functions to meet the needs of pollution early warning and location. This enables full-process support from pollution early warning and location to emergency response, assists managers in making efficient decisions, and improves the response speed and handling effect of water pollution incidents.

[0115] Design principles and innovation analysis of this invention:

[0116] This invention addresses the challenge of data-driven models failing to adequately consider the migration, diffusion, spatiotemporal variability, and multimodal synergy of water bodies during water pollution monitoring and treatment. Specifically, in water pollution risk assessment and source tracing, the diffusion of pollutants in water bodies is influenced by factors such as flow velocity, temperature, and topography. Furthermore, the diverse properties and types of pollutants lead to significant spatiotemporal variability in their behavior within water bodies. While data-driven models can provide water quality monitoring information through sensor and remote sensing data, they often fail to accurately capture the migration, diffusion, and complex interactions between multimodal pollution sources. Therefore, this invention integrates the migration and diffusion mechanisms of water pollution with data-driven models, thereby better addressing the difficulties in predicting water pollution.

[0117] First, when considering the migration and diffusion mechanisms of pollutants in water bodies, this invention employs a pollutant diffusion model based on hydrodynamics. In this model, the flow characteristics of water bodies, such as flow velocity and direction, are closely related to the physicochemical properties of pollutants, such as solubility and diffusion coefficient. By combining information such as topography and climate, a migration and diffusion equation for pollutants is established, accurately simulating the trajectory and concentration changes of pollutants in water bodies. Second, this invention comprehensively considers the spatiotemporal variability of pollutant concentrations through spatiotemporal analysis methods. In practical applications, the degree of water pollution changes over time, and the concentration of pollutants varies significantly across different regions. Therefore, this invention introduces spatiotemporal modeling technology, achieving accurate early warning of water quality change trends through time-series analysis and spatial distribution feature extraction of pollutant concentration data in water bodies. Furthermore, this invention also solves the problem of the inability to effectively integrate different sensor and remote sensing data in data-driven methods through multimodal collaborative analysis. Water pollution monitoring data comes from a wide range of sources, including ground sensor data, remote sensing image data, and meteorological data, which have different timeliness and spatial scales. To this end, the present invention employs an algorithm based on multimodal data fusion to perform consistency matching of these data in terms of time, space and scale, thereby improving the accuracy and real-time performance of data analysis.

[0118] Combining the three mechanisms mentioned above, this invention integrates the migration and diffusion mechanisms of water pollution with data-driven technology, enabling more accurate prediction of water pollution occurrence and precise source tracing. The diffusion and concentration changes of pollutants in water bodies, as well as the spatiotemporal distribution of pollution sources, can all be effectively analyzed and predicted through this model, thus providing a scientific basis for water pollution control.

[0119] The beneficial effects of this invention are as follows:

[0120] (1) This invention integrates hydrodynamic mechanism models, real-time monitoring data, and multi-source heterogeneous information to construct a dynamic analysis framework that combines physical interpretability with the advantages of data-driven approaches. It retains the migration and diffusion simulation capabilities of traditional water quality models while overcoming the limitations of single data sources in terms of spatiotemporal resolution and pollutant interaction analysis. It can be widely applied in fields such as industrial wastewater discharge supervision, watershed pollution dynamic early warning, urban pipeline leakage tracing, and emergency response to sudden pollution events, providing full-chain technical support for intelligent water environment governance.

[0121] (2) Regarding industrial wastewater pollution monitoring and source tracing, traditional monitoring methods struggle to quickly locate pollutants and trace responsible parties, especially given the hidden discharge from multiple pollution sources and complex pollutant migration paths in chemical industrial parks. This invention utilizes a ground-based water quality sensor network, UAV spectral remote sensing, and meteorological and hydrological monitoring equipment to collect real-time data on the concentrations of various pollutants such as COD, heavy metals, and volatile organic compounds in water bodies. This data is then combined with watershed topography, pipeline layout, and enterprise discharge permit information to construct a multi-source spatiotemporal database. Based on a hydrodynamic diffusion model and deep learning algorithms, the system can dynamically simulate the migration trajectory of pollutants in rivers and utilize multimodal technology to trace pollution sources. For example, in the 2021 pollution incident in City H, the system, by fusing hyperspectral remote sensing data and multimodal data, located the leak point within two hours, achieving a source tracing accuracy of 95%, a 67% improvement in efficiency compared to traditional manual investigation. Attached Figure Description

[0122] Figure 1 This is a structural diagram of a water pollution risk early warning and source tracing system that integrates multi-source data.

[0123] Figure 2 This is a schematic diagram of the data transmission process acquired by the sensor network in Embodiment 1 of the present invention.

[0124] Figure 3 This is a schematic diagram of the water spectral feature extraction process of the remote sensing data module in Embodiment 2 of the present invention.

[0125] Figure 4 This is a comparison chart of the pollution diffusion prediction results of the WOA-LSSVM prediction model in Embodiment 4 of the present invention.

[0126] Figure 5 This is the pollution source tracing and location map in Embodiment 5 of the present invention.

[0127] Figure 6 This is a distribution map of sensor nodes (grid layout + water depth labeling).

[0128] Figure 7 This is a fitness curve.

[0129] Figure 8 This is a diagram illustrating the attention mechanism transition process. Detailed Implementation

[0130] The present invention will be further described below with reference to specific implementation examples.

[0131] This embodiment provides a water pollution risk early warning and source tracing method based on multi-source data fusion. This embodiment consists of seven parts, constructed in the order of data acquisition, feature extraction, spectral registration, early warning, source tracing, and platform construction. Specifically, as follows... Figure 1 As shown. Includes the following steps:

[0132] The first step is to construct a multi-source heterogeneous data acquisition network and achieve real-time data transmission, i.e., to perform sensor network data acquisition and transmission, such as... Figure 2 As shown:

[0133] Within the target water area, due to the complex geographical environment, including shallows and deep channels, the water flow exhibits a characteristic of alternating slow-moving and fast-moving zones. Based on these characteristics, an equilateral triangular grid layout was adopted to deploy the wireless sensor network, with the grid side length precisely set at 200m, covering a total area of ​​1.5 km². The sensor nodes were precisely placed 0.5m below the water surface, a depth that effectively avoids interference from floating debris, enabling the acquisition of more accurate and stable water quality data.

[0134] This wireless sensor network collects a wealth of water quality parameter data in real time, denoted as Q = {q1, q2, ..., qn}. Here, q1 represents pH, allowing real-time monitoring of changes in water pH, ranging from approximately 6.5 to 8.5; q2 is dissolved oxygen content, crucial for assessing the water body's self-purification capacity, typically between 5 and 10 mg / L; q3 is chemical oxygen demand, reflecting the degree of organic pollution in the water, generally between 10 and 30 mg / L; and q4 is ammonia nitrogen content, directly reflecting the water body's nitrogen-containing organic pollution level, usually between 0.1 and 1 mg / L.

[0135] Meanwhile, satellite remote sensing technology, utilizing high-resolution Sentinel-2 satellite imagery to acquire water body data, enables macroscopic monitoring of information such as water area and turbidity. Drones are used for water body inspections; a DJI M300 drone equipped with a RedEdge-MX multispectral camera acquires drone inspection data containing detailed spectral characteristics of the water body. Let S be the data set acquired by satellite remote sensing and drone inspections. In addition, water samples are periodically collected and tested in the laboratory to obtain a high-precision laboratory test data set L containing data on heavy metal content, microbial indicators, etc. Subsequently, the water quality parameter data Q collected by the wireless sensor network, the satellite remote sensing and drone inspection data S, and the laboratory test data L are organically integrated and input into the system, providing a comprehensive and diverse data foundation for subsequent processing.

[0136] In terms of specific technical details, the sensor nodes employ a combination of solar power and lithium batteries. The lithium batteries have a capacity of 10,000mAh, ensuring sufficient power for each node even under severe conditions of insufficient sunlight for several consecutive days. The system's battery life can reach over 30 days, meeting the needs of long-term environmental monitoring. The solar panels charge efficiently during the day, minimizing dependence on external energy sources and enabling longer-term stable operation. Each node incorporates an advanced filtering algorithm using a moving average window with a window size of 5 sampling points, effectively smoothing data fluctuations and reducing the interference of instantaneous outliers. Simultaneously, the system rigorously removes outliers from the collected data using a scientifically set threshold of ±3σ, ensuring accurate and reliable data transmission. The sensor node distribution map (corresponding to Figure 6), extracted from the transmitted data, clearly shows the distribution of nodes in the water area, detailing the location of each node and its water depth, which helps to deeply understand the close relationship between network coverage and water stratification.

[0137] The sensor parameters are shown in Table 1.

[0138] Table 1 Sensor Parameter Description

[0139]

[0140] Through the above technical solutions, sensor networks not only have high monitoring accuracy, but can also maintain stable operation for a long time in complex aquatic environments, providing solid data support for subsequent data analysis and decision-making.

[0141] The second step: Multi-source data fusion and feature enhancement processing based on spatiotemporal references, i.e., spectral feature extraction from remote sensing images, such as... Figure 3 As shown;

[0142] Based on the ArcGIS platform, spatial registration was performed on the multi-source data (Q, S, L) acquired in the first step. Spatial registration aims to unify data from different sources and coordinate systems into the same geospatial coordinate system. For example, the WGS84 coordinate system of satellite imagery, the custom coordinate system of UAV imagery, and the local coordinate system of sensor data were all converted to the unified CGCS2000 coordinate system to ensure data consistency in geospatial space and lay the foundation for subsequent processing.

[0143] After spatial registration, advanced data cleaning algorithms are used to meticulously process the multi-source data. For the water quality parameter data Q collected by the wireless sensor network, noise and abnormal data caused by sensor failure, communication interference, etc., are removed to obtain the cleaned dataset Qclean; for the satellite remote sensing and UAV inspection data S, cloud noise, bad pixels, and abnormal data caused by unstable UAV flight attitude are removed from the images to generate Sclean; for the laboratory test data L, outliers caused by experimental operation errors and instrument deviations are checked to form Lclean.

[0144] Subsequently, a semantic segmentation algorithm was used to perform deep processing on the cleaned data to extract effective features. These effective features include time-series feature data. This refers to the timestamp information of data collection and the changing trends of water quality parameters over time. Spatial characteristic data. This refers to the latitude and longitude coordinates of each monitoring point after spatial registration on the ArcGIS platform. This embodiment utilizes the Sentinel-2 satellite with a spatial resolution of 10 meters, covering bands including the B3 green band (center wavelength approximately 560nm), the B8 near-infrared band (center wavelength approximately 842nm), and B8a, B11, and B12, along with a RedEdge-MX multispectral camera mounted on a DJI M300 drone, with a resolution of 1.2 cm / pixel and a band range of 475-840 nm, to simultaneously acquire images of the target water area. The satellite imagery undergoes precise radiometric correction using the Sen2Cor tool, while the drone imagery undergoes geometrical correction by carefully deploying 10 ground control points per square kilometer to ensure an RMSE ≤ 0.3 pixels, significantly improving image accuracy. Based on satellite and drone imagery data, the normalized differential water index is calculated using the green band reflectance... Near-infrared band reflectivity The calculation shows that the ratio of the difference between the two to the sum of the two is NDWI = (B3 - B8) / (B3 + B8).

[0145] Based on the laboratory-measured endpoint value of chlorophyll concentration (2.5 μg / L), an ARIMA model was successfully established within the technical range of 2.5-15.6 μg / L. The coefficient of determination R² of the ARIMA equation reached 0.92, indicating that the model can accurately predict chlorophyll concentration. UAV data, with its high resolution advantage, effectively compensated for the blurring problem in detailed areas of satellite imagery, especially in identifying edge pollution zones, exhibiting higher clarity and accuracy, ultimately forming a dataset F with spatiotemporal fusion characteristics. This dataset integrates Qclean's real-time monitoring parameters, Sclean's remote sensing inspection features, and Lclean's laboratory detection results, incorporating the time dimension. and spatial dimensions This information provides high-quality data support for subsequent analysis.

[0146] Perform GIS spatiotemporal data registration, specifically:

[0147] In the ArcGIS platform, this embodiment focuses on spatial registration between the sensor data in the WGS84 coordinate system and the satellite imagery in the UTMZone 50N coordinate system. Bilinear interpolation is used for projection transformation, strictly controlling the registration error to ≤0.5 pixels. Within the 0.3-0.8 pixel range provided by the technical solution, high-precision spatial matching is ensured. Accurate time synchronization is achieved using hourly sampled laboratory test data and sensor data at 5-minute intervals through linear interpolation. The overlay analysis of the registered data clearly shows that the impact of spatiotemporal deviation on pollution feature extraction is significantly reduced.

[0148] The dataset used in this embodiment comes from the National Surface Water Quality Automatic Monitoring Real-time Data Release System. This system collects key water quality parameters such as total nitrogen in real time through sensors at carefully deployed monitoring points across the country, spanning from 2021 to 2024. Geospatial information is based on the "2019 Map of China" (Approval No. GS (2019) 1822), and remote sensing technology is used to accurately extract watershed boundaries, river system distribution, and monitoring point coordinates. Water quality sensors automatically upload monitoring values ​​to the central database every four hours, which are precisely matched with the geographic coordinates interpreted from remote sensing using unique site codes. Simultaneously, a spatiotemporal attribute table is constructed by combining spatial coordinates, ultimately forming a structured dataset containing the three elements of time, location, and concentration, providing standardized input for spatiotemporal analysis.

[0149] Subsequently, the collected data was imported into the ArcGIS Pro system, using total nitrogen concentration as an example to visualize the total nitrogen concentration in water bodies of Chinese provinces. First, attribute data containing geographic location information and corresponding total nitrogen concentrations were imported into the system. Georegistration or spatial linking tools were used to closely link the attribute table with the base map, generating a point layer with spatial attributes. Then, based on different ranges of total nitrogen concentration, five intervals were scientifically preset in the symbol system, using gradient color schemes to map different pollution levels, intuitively and clearly presenting the regional pollution distribution characteristics. Finally, combined with base map elements, further in-depth analysis of the spatial correlation of pollution can be conducted. Furthermore, this visualization result establishes a basic framework for pollution risk prediction and precise source tracing, providing strong technical support for the next step of water pollution risk prediction and source tracing result display, and effectively contributing to the construction of an integrated management framework of "monitoring-early warning-source tracing-treatment".

[0150] Step 3: Construct a dynamic early warning system for water pollution risks based on the WOA-LSSVM model; that is, WOA-LSSVM model training and early warning, such as...Figure 4 As shown, specifically:

[0151] Step 3.1: First, a least squares support vector machine (LSSVM) model is constructed to learn the mapping relationship between the water element dataset and the migration amount. The LSSVM model is shown in Equation (1). To achieve high-precision prediction, the LSSVM model minimizes the prediction error through a loss function, which is shown in Equation (2).

[0152] The optimization objective of the LSSVM model is to minimize the model parameter complexity and prediction error, as shown in equation (3); this optimization objective aims to minimize the model parameter complexity. and This allows the predicted migration values ​​of elements in water to approximate the actual migration values ​​as closely as possible, thereby revealing the nonlinear laws governing the migration behavior of elements in water.

[0153] Step 3.2: To improve the prediction accuracy and reduce the computational complexity of the LSSVM model, the Whale Optimization Algorithm (WOA) is introduced to optimize the LSSVM model parameters. WOA optimizes parameters by simulating whale predation behavior; its core principle is to approximate the global optimum by updating the individual whale positions. The specific update rule is as follows: when a whale randomly selects a companion for searching, i.e., p < 0.5 and | When |≥1, the position update formula is as shown in formula (4).

[0154] When the whale moves closer to the global optimal position, i.e., p<0.5 and |A|<1, the position update formula is as shown in formula (5).

[0155] When simulating spiral ascent predation behavior, i.e., p>0.5, the position update formula is as shown in formula (6).

[0156] The distance between an individual whale and its random companions or the global optimal position is calculated as shown in formulas (7) and (8).

[0157] To adapt to the migration and diffusion characteristics of water elements, control parameters are introduced into the WOA calculation, which are optimized in conjunction with the estimated diffusion coefficient and the estimated flow velocity, as shown in formulas (9), (10), and (11).

[0158] Meanwhile, by dynamically adjusting the search strategy using parameters as shown in formulas (12) and (13), the algorithm terminates when the maximum number of iterations M is reached. The solution corresponding to the global optimal position is the optimal parameter combination of the LSSVM model. At this time, the model parameters are minimized and the computation time is significantly reduced.

[0159] Step 3.3: Utilize the dataset with spatiotemporal fusion characteristics obtained in step 2. In this process, the optimized WOA-LSSVM model is trained.

[0160] Data set Time series data in Represented as .in Represents the total number of samples in the time series. For time steps The input data, including dimensions such as temperature, pH, permanganate concentration, total ammonia concentration, and conductivity, all came from a dataset. Feature extraction results; The target value at the corresponding time step represents the amount of element migration in the water. During training, the dataset... The model fitting direction is dynamically adjusted based on the spatiotemporal characteristics, and the model parameters are continuously optimized until the prediction error meets the preset threshold.

[0161] After the model training is completed, the pollutant concentration is fused and predicted by combining the mechanistic model, as shown in formula (14).

[0162] Leveraging cloud computing capabilities, the spatiotemporal fusion characteristics of the data set collected in real time and processed in steps 2 and 3 are analyzed. The analysis is performed, and the water pollution risk warning results are output. The warning level is determined based on the predicted pollutant concentration sequence for future time steps, as shown in formula (15).

[0163] The warning range is determined based on the predicted pollutant concentration at a spatial point, as shown in formula (16).

[0164] Through the above steps, a dynamic early warning system for water pollution risks is constructed, enabling real-time prediction, level determination, and extent delineation of water pollution. The fitness curve illustrates the model's convergence; the error value reaches its minimum after 6-7 calculations, indicating convergence. Figure 7 .

[0165] The Whale Optimization Algorithm was used to optimize the hyperparameters of the LSSVM model, with 100 iterations. The optimized model had 128 neurons in the hidden layer and a learning rate of 0.001. Water quality monitoring data from 2021-2024, including key water quality indicators such as permanganate ions, dissolved oxygen, water temperature, pH, and conductivity, were input at 4-hour intervals to predict the pollution concentration for the next 4 hours. The average MAE after 5-fold cross-validation was 0.71887 mg / L, MBE was 0.054245 mg / L, and RMSE was 1.4239 mg / L. Compared with the unoptimized LSSVM model and the hydrodynamic model, the prediction accuracy improved by 54.0%. The average error of WOA-LSSVM was reduced by 10% compared to LSSVM, and by 4.38% compared to the hydrodynamic model. See Table 2 for the evaluation index table.

[0166] Evaluation Indicator Table

[0167]

[0168] Step 4: Perform reverse localization of pollution sources based on the CNN-GRU-SE Attention model, such as... Figure 5 As shown, specifically:

[0169] Step 4.1, based on the dataset with spatiotemporal fusion characteristics obtained in step 2. First, we construct the CNN-GRU-SE Attention model. The model construction process is as follows. This model is a combined architecture of temporal feature extraction and spatial feature extraction. It achieves multi-dimensional feature fusion through a three-layer structure. The construction of each module and the formulas are as follows:

[0170] The first layer is the CNN module, which is the spatial feature extraction module, used to capture the local correlation features between spatial points. The construction process is shown in formulas (17) and (18).

[0171] The second layer is the GRU module, which is the time feature extraction module. It captures the long-term dependencies of time series through a gating mechanism. The construction process is shown in formulas (19) to (22).

[0172] The third layer is the SE-Attention module, which is the key feature enhancement module. It highlights the contribution of important features through attention weights. The construction process is shown in formulas (23) and (24).

[0173] The three modules described above together constitute the CNN-GRU-SE Attention model, which is used to extract key information in the spatiotemporal and feature dimensions from multi-source data. After the model is built using the above steps, the dataset formed in the second step is then... Data set as model input Represented as ,in This is a spatiotemporal fusion dataset; In terms of time dimension, a total One time step; For spatial dimensions, a total One spatial point; For feature dimensions, a total of Water quality-related characteristics, such as dissolved oxygen and pH.

[0174] Step 4.2: Apply the constructed CNN-GRU-SE Attention model to the input dataset. Processing: The spatial features are obtained after pooling by the CNN module. After processing by the GRU module, the hidden state with fused temporal features is obtained. After processing by the SE-Attention module, the attention-weighted fused features are obtained. This ultimately forms a spatiotemporal feature data set. ={ , ,…, },in for The fusion features at any given time include spatial distribution, temporal variation, and key feature weight information.

[0175] Then, the time series feature data obtained in the second step are used. and spatial feature data The constructed CNN-GRU-SE Attention model is trained. Training samples are drawn from the dataset. The feature extraction results, input is The spatiotemporal feature sequence in { The output is the corresponding historical pollution source location label. During training, the dataset is used... The spatiotemporal characteristics are dynamically adjusted to adjust the model parameters, i.e., the convolution kernel. Size, GRU weight matrix / / Attention weight matrix The goal is to minimize the mean squared error, and the model is continuously optimized until the prediction accuracy meets a preset threshold. Utilizing... After the dataset model is trained, a reverse diffusion model is used to simulate the diffusion path of pollution in water bodies, enabling reverse tracing of pollution sources. The reverse diffusion model is constructed based on the following equations:

[0176] The basic equation for reverse diffusion is shown in equation (25). To meet computational requirements, the basic equation is discretized to obtain the discretized reverse diffusion equation, as shown in equation (26).

[0177] Step 4.3, combining hydrological information of the target water area Information on the distribution of sewage outlets The spatiotemporal feature set based on the output of the CNN-GRU-SEAttention model Based on the results of the reverse diffusion model, multi-source information fusion localization is achieved through four parts of calculation:

[0178] The first part is the source term estimation formula based on spatiotemporal characteristics, as shown in formula (27).

[0179] Part Two: The formula for calculating hydrological information for hydrological impact correction is shown in Formula (28).

[0180] Part Three, prior to discharge outlets, i.e., the formula for screening potential pollution sources is as shown in Formula (29).

[0181] Part Four: The formula for integrating the positioning results is shown in Formula (30).

[0182] Finally, the source location map of the pollution source is output. The possible location of the pollution source is determined by normalization, as shown in formula (31).

[0183] Spatial features were extracted from 5 channels of 256×256 pixel remote sensing imagery using a CNN. A GRU was used to process 24 steps of long-term data within a 6-hour time window. The SE Attention mechanism was employed to enhance the turbidity feature weight contribution to 73%. Combining a hydrological flow velocity of 0.3 m / s with the coordinates of the discharge outlet, a reverse diffusion model generated a pollution location map with a confidence level ≥87%. The location results indicated that the pollution source was located at... Figure 5 The three pollution sources include one cross-basin pollution zone, B. The source location results are consistent with the 2021 central environmental protection inspection results for H City. The attention mechanism shift process corresponds to... Figure 8 This section provides a detailed explanation of the attention mechanism transformation process of CNN-GRU-SE Attention and the parameter values ​​for CNN.

[0184] Step 5: Development of a multi-dimensional visualization decision support system and emergency response, i.e., emergency response policy triggering, specifically:

[0185] The developed multi-dimensional visualization decision support system—the Intelligent Water Environment Management and Control Visualization Platform—is built upon the secondary development interface of the open-source geographic information platform QGIS and the ECharts visualization library. The platform reuses QGIS's geospatial data processing kernel as its map rendering engine, while the chart rendering function is developed and extended based on open-source components of ECharts. It integrates three core visualization modules, enabling real-time push of water pollution risk warning results and pollution source location results to environmental management personnel in the form of dynamic charts and GIS maps, achieving intuitive display and real-time push of pollution-related data.

[0186] Meanwhile, the platform has a built-in policy recommendation generation module. This module relies on the DeepSeek policy knowledge base, which covers documents from the Ministry of Environmental Protection. It can automatically match relevant legal provisions for early warning events as the basis for generating policy recommendations. Combined with the API encryption module, it can ensure information security during data transmission and interaction, and provide full-process support for pollution emergency decision-making, including data visualization, policy basis matching, and data support.

[0187] The spatiotemporal dynamic visualization module is built based on QGIS's map projection and layer management functions, reusing its WGS84 coordinate system adaptation and spatial element rendering interface. It integrates the water pollution risk warning results obtained in the third step (including the warning level W and warning range A) with the pollution source location results obtained in the fourth step (including the location probability). and core position This is mapped to a geospatial coordinate system. The heatmap is then displayed by calling ECharts' heatmap component. The spatial distribution of the location is such that darker colors indicate a higher probability of localization, based on a reverse diffusion model. The inversion results use dynamic vector arrows to mark the pollutant diffusion path and use polygons of different colors to divide the warning area A: green for the safe zone, yellow for the attention zone, orange for the warning zone, and red for the danger zone. The vector arrow animation effect is developed based on ECharts' path animation API.

[0188] The data visualization module directly reuses ECharts' line chart, bar chart, and radar chart components. It connects to the system's data interface through a custom data adapter, displaying key water quality parameters such as dissolved oxygen and pH values ​​from the dataset in real time. The time-series changes of characteristic indicators are presented in radar charts to show pollution source location-related indicators, including source terms. Hydrological weight Sewage outlet inspection The contribution level helps managers quickly grasp core data.

[0189] The response status visualization module is developed based on the Vue framework and the Element UI component library. It reuses ElementUI's flowchart component and status label component to display the emergency response progress, namely, early warning triggering → level determination → measure execution → result feedback. The status labels update the status of each stage in real time, including started, in progress, and completed. The statistical charts are implemented based on the ECharts component to ensure that the handling process is traceable.

[0190] The above modules work in conjunction with the preceding steps system through a cloud data interface to synchronize processed water quality data, early warning results, and location information in real time. These data are then pushed to environmental management personnel in the form of charts, maps, etc., providing intuitive data support for decision-making.

[0191] A policy recommendation generation module is integrated into the intelligent water environment management and visualization platform. This module is built on the DeepSeek open-source policy knowledge base framework, reusing its core semantic matching algorithm code and supplementing it with a corpus of documents and laws and regulations from the Ministry of Environmental Protection to form a dedicated knowledge base. Key features of current pollution events are extracted, including pollutant type, scope of impact, and pollution source type, derived from early warning results. and positioning results The module uses the DeepSeek semantic matching algorithm to match relevant clauses from the knowledge base, such as the provisions on sewage outlet management and emergency response in the "Water Pollution Prevention and Control Law of the People's Republic of China," and generates targeted suggestions based on event characteristics. Simultaneously, the module reuses the HTTPS protocol and the OAuth2.0 open-source authentication component to implement API encryption, ensuring the security of data calls and transmission of the clauses cited in the policy suggestions and preventing the leakage of sensitive information.

[0192] At the data interface level, the system obtains real-time JSON data of industrial discharge permits within the watershed via the DeepSeek API, extracts fields such as enterprise discharge type, pollutant code, and violation records from the past three years, and combines this with a historical pollution event database to call the DeepSeek multimodal model to generate dynamic threshold adjustment suggestions. When the lead concentration exceeds the standard by >0.01mg / L, the model outputs an emergency plan of closing the water intake → starting the purification equipment → notifying the environmental protection department, with a confidence level of 0.88, based on relevant provisions of the "Water Pollution Prevention and Control Law of the People's Republic of China". Regarding equipment control, the DeepSeek IoT module sends commands to emergency equipment via the MQTT protocol, monitoring equipment response latency in real time. The total time from command issuance to operation is ≤30 minutes; if the timeout occurs, log checks are automatically triggered. Predicted data is presented as a spatiotemporal distribution map of pollutant concentrations, and solutions are generated by combining the pollution source location results with the warning range and level correlation.

[0193] The above embodiments are merely illustrative of the implementation methods of the present invention, but should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.

Claims

1. A method for water pollution risk early warning and source tracing based on multi-source data fusion, characterized in that, The water pollution risk early warning and source tracing method includes the following steps: Step 1: Construct a multi-source heterogeneous data acquisition network and achieve real-time data transmission; specifically: The water quality parameter data Q collected by the wireless sensor network, the data set S obtained by satellite remote sensing and UAV inspection, and the water body detection data L are integrated to obtain a multi-source data set (Q, S, L); Step 2: Perform multi-source data fusion and feature enhancement processing based on a spatiotemporal benchmark; specifically: After spatial registration and cleaning of the multi-source dataset (Q, S, L) obtained in the first step, effective features are extracted to form a dataset F with spatiotemporal fusion characteristics. The dataset F is a water quality feature dataset unified under the same spatiotemporal benchmark, which serves as standardized input data for subsequent water quality analysis. Step 3: Construct a dynamic early warning system for water pollution risks based on the WOA-LSSVM model; specifically: Step 3.1: Construct the Least Squares Support Vector Machine (LSSVM) model, which minimizes the prediction error through a loss function and determines the optimization objective of the LSSVM model. Step 3.2: The Whale Optimization Algorithm (WOA) is introduced to optimize the parameters of the LSSVM model, resulting in the WOA-LSSVM model. Step 3.3: Use the dataset F obtained in step 2 to train the WOA-LSSVM model; Through the above steps, the construction of a dynamic early warning system for water pollution risks is completed, enabling real-time prediction, level determination, and scope delineation of water pollution. Step 4: Perform reverse localization of pollution sources based on the CNN-GRU-SE Attention model; specifically: Step 4.1: Based on the dataset F with spatiotemporal fusion characteristics obtained in the second step, construct the CNN-GRU-SEAttention model, which is a combination architecture of temporal feature extraction and spatial feature extraction, and achieves multi-dimensional feature fusion through a three-layer structure; Step 4.2: Process the input dataset F using the CNN-GRU-SE Attention model to obtain the spatiotemporal feature data set T = {T1, T2, ..., T}. i }, where T i The fused features at time i are used; and the CNN-GRU-SE Attention model is trained. Step 4.3: Calculate and realize the multi-source information fusion positioning, and finally output the source location map of pollution source. The possible location of pollution source is determined by normalization processing. Step 5: Development of a multi-dimensional visualization decision support system and emergency response.

2. The water pollution risk early warning and source tracing method based on multi-source data fusion according to claim 1, characterized in that, The specific steps for obtaining the multi-source data set in the first step are as follows: Water quality parameter data are collected in real time through a wireless sensor network. Let the set of water quality parameters be Q = {q1, q2, ..., qn}, where qi represents the i-th parameter, i = 1, 2, ..., n. At the same time, satellite remote sensing image data of the water body is acquired through satellite remote sensing technology. UAVs are used to inspect the water area and acquire UAV inspection data containing spectral feature information of the water body. Let the set of data acquired by satellite remote sensing and UAV inspection be S. Water samples were collected and tested to obtain a dataset L.

3. The water pollution risk early warning and source tracing method based on multi-source data fusion according to claim 1, characterized in that, The second step is specifically as follows: Step 2.1: Based on the ArcGIS platform, perform spatial registration on the multi-source dataset (Q, S, L) obtained in the first step. After completing the spatial registration, use data cleaning algorithms to clean the multi-source data and remove noisy and abnormal data. Let the cleaned datasets be Qclean, Sclean, and Lclean. Step 2.2: Use a semantic segmentation algorithm to extract features from the cleaned multi-source data, extract effective features, and form a dataset F with spatiotemporal fusion characteristics; the effective features include time series feature data D, namely the timestamp information of data collection and the changing trend of water quality parameters over time; Spatial feature data E refers to the latitude and longitude coordinates of each monitoring point after spatial registration based on the ArcGIS platform.

4. The water pollution risk early warning and source tracing method based on multi-source data fusion according to claim 1, characterized in that, The third step is specifically as follows: Step 3.1, the LSSVM model is used to learn the mapping relationship between the water element dataset and the migration amount; the LSSVM model is represented as: f(x i )=w T φ(x s )+b (1) Where, f(x) s ) represents the input x s The predicted migration amount; w represents the weight vector; φ(x) s ) represents the kernel function mapping, used to map input data to a high-dimensional feature space; b represents the bias term; To achieve high-precision prediction, the LSSVM model minimizes the prediction error through a loss function, which is expressed as: Among them, t i γ represents the actual migration amount; b represents the bias term; γ represents the penalty coefficient, used to control the degree of penalty applied to slack variables; N represents the sample size; C s Indicates the concentration of the target element in the water body; D represents the concentration gradient; D represents the diffusion coefficient. λ represents the second partial derivative of concentration; λ represents the regularization parameter; v represents the water flow velocity. The optimization objective of the LSSVM model is to minimize the model parameter complexity and prediction error, where the error is expressed as: The optimization objective aims to minimize the model parameters w and b to make the predicted migration of elements in water as close as possible to the actual migration values, thereby revealing the nonlinear laws governing the migration behavior of elements in water. Step 3.2: The Whale Optimization Algorithm (WOA) optimizes parameters by simulating whale foraging behavior. It approximates the global optimum by updating the individual whale positions. The update rule is as follows: When a whale randomly selects a companion for searching (i.e., p < 0.5 and |A| ≥ 1), the position update formula is: Where A represents the algorithm control parameters; s represents the individual ID; j represents the spatial dimension ID; Positions(s,j) represents the position of the s-th individual in the j-th dimension; X rand (j) indicates that the position of a companion is randomly selected; p represents the distance between the current individual and a randomly selected companion individual in the j-th dimension; The formula for calculating the distance between an individual whale and its random companions or the global optimal position is: D leader =|C×Leader_pos(j)-Positions(s,j)|(8) Where C represents the algorithm control parameter, D represents the distance between the current individual and a randomly selected peer individual in the j-th dimension; Leader This represents the distance between the current individual and the best individual in the j-th dimension; To adapt to the migration and diffusion characteristics of elements in water, control parameters are introduced into the WOA calculation. These parameters are optimized in conjunction with the estimated diffusion coefficient and flow velocity, as shown in the following formula: D estimated =D initial +k2·iteration(9) v estimated =v initial +k3·iteration(10) e=initial-k1·interaction+k2·D estimated -k3·v estimated (11) Among them, D initial v initial These are the initial values ​​for the diffusion coefficient and flow velocity, respectively; k1, k2, and k3 are adjustment coefficients; interation is the current iteration number; D estimated This represents the estimated diffusion coefficient; v estimated This represents the estimated diffusion coefficient; e represents the control parameter; initial represents the initial value of the control parameter e; At the same time, the search strategy is dynamically adjusted through parameters: Where iteration is the current iteration number; M is the maximum iteration number; a1 represents the parameter adaptation migration and diffusion characteristics, enhancing the search capability in complex regions; a2 represents the parameter optimization convergence speed, gradually approaching the optimal solution; When the maximum number of iterations M is reached, the algorithm terminates. The solution corresponding to the global optimal position is the optimal parameter combination of the LSSVM model, thus obtaining the WOA-LSSVM model. Step 3.3: Use the dataset F with spatiotemporal fusion characteristics obtained in the second step to train the optimized WOA-LSSVM model; The time series data D in dataset F is represented as follows: Where N represents the total number of samples in the time series, x i The input data at time step i includes dimensions such as temperature, pH, permanganate concentration, total ammonia concentration, and conductivity, all derived from feature extraction results of dataset F; t i The target value at the corresponding time step represents the amount of element migration in the water. During training, the model fitting direction is dynamically adjusted based on the spatiotemporal characteristics of the dataset F, and the model parameters are continuously optimized until the prediction error meets the preset threshold. After training, the pollutant concentration is fused and predicted by combining the mechanistic model, as shown in the following formula: C final =β·C phys +(1-β)·f(x s (14) Among them, C phys Preliminary concentration distribution calculated for the mechanistic model; f(x) s ) represents the prediction result of the WOA-LSSVM model; β is the weighting coefficient; C final This indicates the final concentration distribution result; The spatiotemporal fusion characteristics of the data set F, which is collected in real time and processed in steps 2 and 3, are analyzed to output water pollution risk early warning results. The warning range is determined based on the predicted pollutant concentration at a specific spatial point.

5. A water pollution risk early warning and source tracing method based on multi-source data fusion according to claim 4, characterized in that, In step 3.2: When the whale moves closer to the globally optimal position (i.e., p < 0.5 and |A| < 1), the position update formula is: |A|<1,Positions(s,j)=Leader_pos(j)-A×D Leader (5) Where Leader_pos(j) represents the position of the best individual in the current group in the j-th dimension; D Leader This represents the distance between the current individual and the best individual in the j-th dimension; When simulating spiral ascent predation behavior, i.e., p > 0.5, the position update formula is: Positions(k,j)=distanceLeader×exp(b×l)×co s(l×2π)+Leader_pos(j)(6) Where k represents the individual ID; distanceLeader represents the distance between the individual whale and the global optimal position; b and l are the spiral parameters.

6. The water pollution risk early warning and source tracing method based on multi-source data fusion according to claim 4, characterized in that, In step 3.3: The water pollution risk early warning result is determined based on the predicted pollutant concentration sequence for future time steps, as shown in formula (15): where \(i\) represents the current time step; \(k\) represents the future time step; represents the predicted value sequence; \(T_1\), \(T_2\), \(T_3\) are early warning thresholds at different levels (\(T_1 < T_2 < T_3\)), where the threshold \(T_1 = 6\ mg / L\), which is the upper limit of the Class III surface water standard and represents the critical value of mild pollution; \(T_2 = 10\ mg / L\), which is the upper limit of the Class IV surface water standard and represents the critical value of moderate pollution; \(T_3 = 15\ mg / L\), which is the upper limit of the Class V surface water standard and represents the critical value of severe pollution; The warning range is shown in formula (16): Where D represents the target water area; T alert The spatial early warning threshold ranges from 60% to 80%; x represents the abscissa of the spatial point; y represents the ordinate of the spatial point. This represents the predicted value at a spatial point.

7. A water pollution risk early warning and source tracing method based on multi-source data fusion according to claim 4, characterized in that, The fourth step is specifically as follows: Step 4.1, the three-layer structure and construction process of the CNN-GRU-SE Attention model are as follows: The first layer is the CNN module, which is a spatial feature extraction module used to capture local correlation features between spatial points; The second layer is the GRU module, which is a time feature extraction module that captures long-term dependencies in time series through a gating mechanism. The third layer is the SE-Attention module, which is a key feature enhancement module that highlights the contribution of important features through attention weights. The three modules described above together constitute the CNN-GRU-SE Attention model, used to extract key information in the spatiotemporal and feature dimensions from multi-source data. After the model is built using the above steps, the data set F formed in the second step is used as the model input. The data set F is represented as F = {f} i,j,k |i=1,…,I;j=1,…,J;k=1,…,K}, where F is the spatiotemporal fusion dataset; i is the time dimension, with I time steps; j is the spatial dimension, with J spatial points; and k is the feature dimension, with K water quality-related features. Step 4.2: Process the input dataset F using the CNN-GRU-SE Attention model: After processing by the CNN module, obtain the pooled spatial features P. i After processing by the GRU module, the hidden state h with fused temporal features is obtained. i After processing by the SE-Attention module, the attention-weighted fused feature T is obtained. i This ultimately forms a spatiotemporal feature data set T = {T1, T2, ..., T}. i }, where T i The fusion feature at time i includes spatial distribution, temporal variation, and key feature weight information; Then, using the time series feature data D and spatial feature data E from the dataset F with spatiotemporal fusion characteristics obtained in the second step, the CNN-GRU-SE Attention model is trained; the training samples come from the feature extraction results of the dataset F, the input is the spatiotemporal feature sequence {f(ijk)} in F, and the output is the corresponding historical pollution source location label; after training, the reverse diffusion model is combined to simulate the diffusion path of pollution in the water body to realize the reverse tracking of pollution sources; Step 4.3: Combining the hydrological information H of the target water area and the sewage outlet distribution information P, based on the spatiotemporal feature set T output by the CNN-GRU-SE Attention model and the results of the backdiffusion model, multi-source information fusion localization is achieved through four parts of calculation: Part 1: Source term estimation based on spatiotemporal characteristics; Part Two: Calculation of Hydrological Information for Hydrological Impact Correction; The third part is the preliminary inspection of sewage outlets, namely the screening of potential pollution sources; Part Four: Integration of Location Results; The final output is a source location map of the pollution source, which is normalized to determine the possible location of the pollution source. in, This represents the normalized location probability, taking values ​​from 0 to 1, with a total sum of 1; (x * ,y * (i,j) represents the spatial coordinates with the highest probability of location, i.e., the most likely location of the pollution source, thus obtaining the pollution source location result; (i,j) represents the index variable of all spatial locations, corresponding to the coordinates of each grid within the water area; This means finding, in all spatial coordinates (x, y), the... The coordinates (x, y) corresponding to the maximum value.

8. A water pollution risk early warning and source tracing method based on multi-source data fusion according to claim 7, characterized in that, In the fourth step mentioned above: In step 4.2, during the training of the CNN-GRU-SE Attention model, the model parameters are dynamically adjusted based on the spatiotemporal features of the dataset F, with the goal of minimizing the mean squared error, and the model is continuously optimized until the prediction accuracy meets the preset threshold. The preset threshold is 0.05; In step 4.2, the reverse diffusion model is constructed based on the following formula: The fundamental equation for reverse diffusion: Where C represents pollutant concentration; t represents time; D represents diffusion coefficient; and u represents water flow velocity. Represents the Laplace operator; represents the gradient operator; S represents the source term, i.e., the pollutant release intensity; To meet computational requirements, the basic equations are discretized, resulting in the discretized reverse diffusion equations: Among them, C t-1 (x,y) represents the pollutant concentration at spatial point (x,y) at time t-1; C t (x,y) represents the concentration at time t; Δt represents the time step; u(x,y) represents the water flow velocity at (x,y); this equation allows us to deduce historical concentrations from the current concentration, thus enabling reverse tracing of the diffusion path. The four parts of step 4.3 are as follows: The first part, the source term estimation formula based on spatiotemporal characteristics, is as follows: Among them, w t T represents the attention weights over the time dimension; t (x,y) represents the spatiotemporal fusion feature value at (x,y) time t, reflecting the correlation between this location and the pollution source; T represents the total spatiotemporal feature sequence; t represents a certain time in the spatiotemporal feature sequence; The second part, the formula for calculating hydrological information for hydrological impact correction, is as follows: in, The average flow velocity of the target water area is represented by α; α represents the adjustment coefficient, i.e., the degree of influence of the control flow velocity deviation. The value represents the Euclidean distance between the actual flow velocity and the average flow velocity at (x,y). The closer the flow velocity is to the average flow velocity, the higher the weight, thus reducing the interference of abnormal water flow on the positioning. h represents hydrological information. The third part, prioritization of sewage outlets, i.e., the formula for screening potential pollution sources, is as follows: Where d(x,y) is the distance from (x,y) to the nearest sewage outlet; β is the sensitivity coefficient, which controls the influence of distance on correlation; the closer the distance, the higher the weight; and P is the sewage outlet distribution information. Part Four, the formula for integrating the positioning results is as follows: in, λ represents the concentration at time t (x,y) retrieved by the reverse diffusion model, reflecting the historical diffusion path; λ represents the time decay factor, 0<λ<1; H(x,y) represents the hydrological weight; P(x,y) represents the prior weight of the discharge outlet; R(x,y) represents the location result.

9. A water pollution risk early warning and source tracing method based on multi-source data fusion according to claim 7, characterized in that, The fifth step is specifically as follows: Step 5.1: Design a water environment intelligent management and visualization platform, integrating a spatiotemporal dynamic visualization module, a data indicator visualization module, and a response status visualization module to realize intuitive display and real-time push of pollution-related data; Step 5.2: Integrate the policy recommendation generation module into the intelligent water environment management and visualization platform to extract key features of the current pollution event; when the early warning results output in step 3 meet the triggering conditions, automatically initiate the emergency response; analyze the characteristics of the pollution event and output the event level; Step 5.3: Classify the pollution level.

10. A water pollution risk early warning and source tracing method based on multi-source data fusion according to claim 9, characterized in that, In the fifth step mentioned above: In step 5.1, the three modules are specifically as follows: The spatiotemporal dynamic visualization module is used to visualize the water pollution risk early warning results obtained in the third step, including the early warning level W and the early warning range A, and the pollution source location results obtained in the fourth step, including the location probability. and core location (x) * ,y * ), mapped to a geospatial coordinate system; presented Spatial distribution, based on the reverse diffusion model The inversion results are marked with dynamic vector arrows to indicate the pollutant diffusion path, and the warning range A is divided with polygons of different colors; The data indicator visualization module displays key water quality parameters in real time and presents indicators related to pollution source location, including the contribution of source term S(x,y), hydrological weight H(x,y), and prior P(x,y) of the discharge outlet. The aforementioned response status visualization module displays the emergency response progress, namely, early warning triggering → level determination → measure implementation → result feedback. It updates the status of each stage in real time through status tags and ensures the traceability of the handling process through statistical charts. In step 5.2: The triggering condition is: Where Trigger represents the trigger condition; This is a predicted sequence of pollutant concentrations for the next k time steps; T1 is the first-level warning threshold. T represents the normalized localization probability. alert The location warning threshold is preset to 0.6, meaning that when the location probability is ≥60%, it is judged as a potential pollution source. When Trigger=1, it is judged as a pollution event and enters the level classification process. The formula for classifying event levels is as follows: Where z represents the pollutant diffusion rate inverted based on the reverse diffusion model; B represents the actual impact area of ​​the pollution event; core water areas and ecologically sensitive areas are preset geographical areas, and the spatial data storage format of QGIS is reused; Step 5.3, the pollution level classification, is as follows: Level 1 light pollution is triggered when a single pollution indicator approaches the warning threshold. Level 2 moderate pollution is triggered by the exceeding of certain indicators; these indicators refer to two key water quality parameters exceeding the corresponding warning threshold, such as any two of the permanganate concentration and total ammonia concentration exceeding the standard, or one of them significantly exceeding the standard and the other approaching the threshold, with the exceedance range being 10%-30%, and not yet causing substantial damage to the aquatic ecological function; Level 3 severe pollution is triggered by multiple indicators exceeding the standard simultaneously. These multiple indicators refer to three or more core water quality parameters, including temperature, pH, conductivity, permanganate concentration, and total ammonia concentration, exceeding the warning threshold, or at least one indicator exceeding the standard by more than 50%, which has posed a serious threat to the aquatic ecosystem, drinking water safety, or the surrounding environment.

Images