Non-point source pollution real-time monitoring and intelligent early warning method based on multi-source data fusion

Abstract

The invention discloses a non-point source pollution real-time monitoring and intelligent early warning method based on multi-source data fusion, and the method comprises the steps: laying ground monitoring nodes and water area monitoring nodes, and obtaining the data of the ground monitoring nodes and the data of the water area monitoring nodes; preprocessing the ground monitoring node data and the water area monitoring node data; calculating a CN value, a maximum retention volume S, a runoff volume Q, a dissolved state load and an adsorption state load; and constructing a pollution risk prediction model and training, and obtaining a predicted future pollution risk grade, a predicted COD value and a predicted ammonia nitrogen concentration based on the trained pollution risk prediction model. According to the method, the ground monitoring node data and the water area monitoring node data are integrated, the pollution risk prediction model is trained according to the ground monitoring node data, the water area monitoring node data and the land utilization type, and prediction of the pollution risk level, the COD value and the ammonia nitrogen concentration in the future is achieved.

Description

Technical Field

[0001] This invention belongs to the field of environmental governance. Specifically, it relates to a method for real-time monitoring and intelligent early warning of non-point source pollution based on multi-source data fusion. Background Technology

[0002] Non-point source pollution refers to pollution that occurs when pollutants enter environmental receptors (such as water bodies, soil, and atmosphere) from no clear emission point or from multiple non-fixed and dispersed emission points simultaneously. Situations prone to non-point source pollution include agricultural production activities, urban living activities, and industrial production activities.

[0003] Non-point source pollution is characterized by its dispersed, random, and delayed nature, and traditional monitoring methods have the following problems:

[0004] Limitations of single-point monitoring: It relies on fixed monitoring stations and is difficult to cover large-scale dynamic pollution sources.

[0005] Data isolation: The lack of integration of multi-dimensional data such as meteorological, hydrological, and land use data results in low prediction accuracy.

[0006] Delayed response: Relies on manual sampling and laboratory analysis, making real-time early warning impossible.

[0007] Lack of dynamic modeling: Static thresholds cannot adapt to the spatiotemporal changes in pollutant migration.

[0008] The present invention aims to solve the above problems and proposes a high-precision, real-time, and multi-scenario-supporting method for monitoring and early warning of non-point source pollution. Summary of the Invention

[0009] This invention addresses the aforementioned problems in existing technologies by proposing a method for real-time monitoring and intelligent early warning of non-point source pollution based on multi-source data fusion.

[0010] The above-mentioned objectives of the present invention are achieved through the following technical means:

[0011] A method for real-time monitoring and intelligent early warning of non-point source pollution based on multi-source data fusion includes the following steps:

[0012] Step 1: Deploy ground monitoring nodes and water monitoring nodes, and acquire data from the ground monitoring nodes and water monitoring nodes;

[0013] Step 2: Preprocess the data from ground monitoring nodes and water monitoring nodes;

[0014] Step 3: Calculate the CN value, maximum retention S, runoff Q, dissolved load, and adsorbed load;

[0015] Step 4: Construct and train a pollution risk prediction model. Based on the trained pollution risk prediction model, obtain the predicted pollution risk level for each node in the next 24 / 48 / 72 hours, as well as the predicted COD value and ammonia nitrogen concentration for each water monitoring node in the next 24 / 48 / 72 hours.

[0016] As described above, the ground monitoring nodes are equipped with meteorological monitoring units, soil parameter acquisition units, and runoff monitoring modules. The ground monitoring nodes collect rainfall, wind speed, air temperature, humidity, nitrogen concentration, phosphorus concentration, flow velocity, and flow rate. The water monitoring nodes are equipped with multi-parameter water quality analyzers and hydrodynamic sensors. The water monitoring nodes collect pH value, turbidity value, COD value, ammonia nitrogen concentration value, dissolved oxygen value, water depth pressure value, flow velocity value, and water depth value.

[0017] As described above, the preprocessing includes outlier correction and time alignment.

[0018] As described above, the CN value is calculated using the following steps:

[0019] When the ground monitoring node data are cultivated land, forest land, orchard, bare land, and rural roads, the corresponding basic CN values ​​are 78, 55, 70, 85, and 92, respectively.

[0020] When the humidity monitored by the ground monitoring node exceeds 30% of the field capacity, the baseline CN value increases by 5; when the humidity is below 10% of the field capacity, the baseline CN value decreases by 3.

[0021] The baseline CN value increases by an additional 12 within 7 days after fertilization at the ground monitoring node; the baseline CN value decreases by 5 when the ground monitoring node is located in the orchard due to the layer of dead branches and leaves.

[0022] Obtain the adjusted CN value.

[0023] The maximum retention S, as described above, is calculated based on the following formula:

[0024] Maximum retention S = (25400 / CN_adjusted) - 254, where CN_adjusted is the adjusted CN value;

[0025] The runoff volume Q is calculated based on the following steps:

[0026] Calculate the initial runoff volume Q1:

[0027] Q1 = (P - 0.2 × S) 2 / (P+0.8S);

[0028] Where P represents the cumulative rainfall;

[0029] Calculate runoff Q = Q1 × tillage coefficient × crop growth period coefficient;

[0030] The dissolved load is calculated based on the following formula:

[0031] L_d = 0.01 × Q × C_soil × K_d;

[0032] L_d represents the dissolved loading rate;

[0033] C_soil represents the nitrogen and phosphorus concentration in the topsoil.

[0034] K_d is the solution allocation coefficient;

[0035] The adsorbed state loading is calculated based on the following formula.

[0036] L_a=Q×K_erosion×C_soil×(1-K_d);

[0037] L_a represents the adsorbed loading state;

[0038] K_erosion is the soil erosion coefficient.

[0039] As described above, constructing and training a pollution risk prediction model includes the following steps:

[0040] Construct training and validation sets, with training nodes in both sets including ground monitoring nodes and water monitoring nodes;

[0041] A pollution risk prediction model is constructed, which includes a time-series processing channel and a spatial correlation channel.

[0042] The pollution risk prediction model is trained based on the training set.

[0043] As described above, the time-series processing channel includes a 1D-CNN layer, a BiLSTM layer, and a fully connected layer set in sequence. The rainfall, wind speed, air temperature, humidity, nitrogen concentration, phosphorus concentration, and land use type of the ground monitoring nodes in the training set, as well as the pH value, COD value, ammonia nitrogen concentration, turbidity value, and dissolved oxygen value of the water monitoring nodes in the training set, are input into the time-series processing channel. The data are processed sequentially through the 1D-CNN layer, the BiLSTM layer, and the fully connected layer, and the predicted pollution risk level of each node for the next 24 / 48 / 72 hours is output.

[0044] The spatial association channel includes a graph attention network. Based on the ground monitoring nodes and water monitoring nodes in the training set, a node feature matrix and an edge feature matrix are constructed. The node feature matrix and the edge feature matrix are input into the graph attention network to obtain the predicted COD value and the predicted ammonia nitrogen concentration for each water monitoring node in the next 24 / 48 / 72 hours.

[0045] The dynamic weight loss function L for constructing the pollution risk prediction model as described above is based on the following formula:

[0046] L=0.6×L_class+0.3×L_reg+0.1×L_graph;

[0047] Where L_class is the classification loss of the temporal processing channel; L_reg is the regression loss function of the spatial association channel; and L_graph is the graph loss function of the spatial association channel.

[0048] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement steps 2 to 4 of the method described above.

[0049] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements steps 2 to 4 of the above method.

[0050] The present invention has the following advantages over the prior art:

[0051] Ground-based monitoring nodes are deployed via mobile platforms in pollution-sensitive areas such as farmland boundaries and urban stormwater runoff areas. These mobile platforms are autonomously navigable and remotely visualized. Miniature monitoring vessels deployed from the mobile platforms serve as waterway monitoring nodes, integrating data from both ground and waterway nodes. Based on this data and land use type, pollution risk prediction models are trained to predict future pollution risk levels, COD values, and ammonia nitrogen concentrations, thus completing a closed-loop management mechanism of monitoring, early warning, and response. This approach improves early warning accuracy by over 30% compared to traditional methods and reduces response time to within one hour, making it suitable for large-scale non-point source pollution control in river basins and urban clusters. Attached Figure Description

[0052] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0053] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0054] Example 1:

[0055] A method for real-time monitoring and intelligent early warning of non-point source pollution based on multi-source data fusion includes the following steps:

[0056] Step 1: Deploy a multi-source sensor network on the mobile monitoring platform to build a three-dimensional monitoring system for area source pollution that integrates space, ground, and ground.

[0057] Ground-based monitoring nodes are deployed using a combination of fixed and mobile methods, focusing on pollution-sensitive areas such as farmland boundaries and urban stormwater runoff areas. They integrate meteorological monitoring units (including high-precision rain gauges, wind speed sensors, and temperature sensors), soil parameter acquisition units (including capacitive humidity sensors and spectroscopic nitrogen and phosphorus detectors), and runoff monitoring modules (including Doppler current meters and water level / flow meters) to achieve dynamic monitoring of the entire precipitation-soil-runoff process. Water area monitoring nodes are covered by a network of intelligent miniature monitoring vessels deployed via mobile platforms. An adaptive gridded deployment strategy is used (focusing on key control areas such as sewage outlets and agricultural drainage outlets). These nodes are equipped with multi-parameter water quality analyzers (including pH, turbidity, UV-based COD, fluorescence-based ammonia nitrogen, and optical dissolved oxygen monitoring) and hydrodynamic sensors (including acoustic Doppler current profilers and pressure-based depth sensors) to construct a high spatiotemporal resolution water quality-hydrological synchronous monitoring network. By using LoRa / 5G hybrid networking technology, real-time correlation and transmission of data from ground monitoring nodes (ground runoff pollution source data) and water monitoring nodes (water quality response data) are achieved, providing comprehensive three-dimensional perception data support for tracing and early warning of non-point source pollution.

[0058] Ground-based monitoring nodes are deployed via mobile platforms in pollution-sensitive areas such as farmland boundaries and urban stormwater runoff catchments. These nodes collect data and are equipped with meteorological monitoring units (collecting rainfall, wind speed, and temperature), soil parameter collection units (collecting humidity, nitrogen concentration, and phosphorus concentration), and runoff monitoring modules (measuring flow velocity and flow rate), forming a ground-based monitoring network. The mobile platforms can be fixed in place or dynamically inspected, and the real-time data collected from the ground-based monitoring nodes is transmitted back wirelessly.

[0059] The water monitoring nodes consist of miniature monitoring vessels deployed from a mobile platform, distributed in a grid pattern across water bodies (such as rivers and drainage outlets), with denser deployment in key areas (sewage outlets and farmland drainage ditches). These monitoring vessels collect data from the water monitoring nodes. They are equipped with multi-parameter water quality analyzers (collecting pH, turbidity, COD, ammonia nitrogen concentration, dissolved oxygen, and water depth / pressure) and hydrodynamic sensors (collecting flow velocity and water depth), enabling real-time tracking and source tracing of pollutants. The ground-based nodes and water-based nodes work collaboratively to form an integrated air-ground non-point source pollution monitoring system.

[0060] Step 2: Transmit the data from the ground monitoring nodes and the water monitoring nodes to the mobile platform for data preprocessing, including outlier correction and time alignment.

[0061] Outlier correction:

[0062] Physical threshold filtering: Remove values ​​exceeding the limit (e.g., pH>14 or <0) based on the sensor's measurement range.

[0063] Statistical outlier detection: Data collected by sensors in chronological order constitutes time-series parameter data. For time-series parameter data of the same type (e.g., COD), the median (Xmed) and absolute median difference (MAD) are calculated within a sliding window (window size of 1 hour). Outliers are defined as those exceeding Xmed ± 3 × 1.4826 × MAD (1.4826 is the Gaussian distribution correction coefficient). Outliers are replaced by linear interpolation or the mean of adjacent time points.

[0064] Time alignment:

[0065] The timing data of parameters with different sampling frequencies are uniformly interpolated to the same sampling frequency.

[0066] Step 3: Estimate runoff using the SCS-CN model and calculate diffusion paths using pollutant migration equations.

[0067] 1. Data Acquisition and Input

[0068] This system collects real-time data on key parameters from ground monitoring nodes deployed in farmland areas. The meteorological monitoring unit provides minute-level rainfall (P) monitoring, while the soil parameter acquisition unit measures soil moisture and nitrogen and phosphorus content (including nitrate nitrogen and available phosphorus) in the 0-30cm soil layer. This data is combined with land use type data from each ground monitoring node in the GIS system (classes 1-5 correspond to cultivated land, forest land, orchard, bare land, and rural roads, respectively). This ground monitoring node data is transmitted in real-time to the central processing platform via a wireless network.

[0069] 2. Calculation of adjusted CN values ​​for ground monitoring nodes

[0070] Based on the traditional SCS-CN model, this invention develops a dynamic correction algorithm specifically for agricultural non-point source pollution to obtain adjusted CN values ​​(runoff curve values). The adjusted CN values ​​are determined based on the following methods: the baseline CN value is determined according to land use type: arable land 78, forest land 55, orchard 70, bare land 85, and rural roads 92. Dynamic correction considers two key factors: first, soil moisture; when the moisture level monitored by the ground monitoring node exceeds 30% of field capacity, the baseline CN value increases by 5; when it is below 10% of field capacity, the baseline CN value decreases by 3. Second, agricultural activities; within 7 days after fertilization at the ground monitoring node, the baseline CN value increases by an additional 12; when the ground monitoring node is located in an orchard, due to the interception effect of the litter layer, the baseline CN value decreases by 5. The adjusted baseline CN value is the adjusted CN value.

[0071] 3. Calculation of runoff and pollutant load

[0072] Runoff (Q) was calculated using the modified SCS-CN formula, where maximum retention (S) was determined by the adjusted CN value. Nitrogen and phosphorus loss was obtained by multiplying runoff by the nitrogen and phosphorus concentrations in the topsoil and the erosion coefficient. The erosion coefficient was based on the USLE model, taking into account factors such as soil type, slope, and vegetation cover, and its value ranged from 0 to 1.

[0073] 3.1 Runoff Calculation

[0074] Based on the dynamic SCS-CN model, the runoff calculation process is as follows:

[0075] (a) Determining basic parameters:

[0076] Maximum retention S = (25400 / CN_adjusted) - 254, where CN_adjusted is the adjusted CN value.

[0077] Runoff calculation is initiated when the cumulative rainfall P > 0.2 × S.

[0078] (b). Preliminary formula for calculating runoff:

[0079] Q1 = (P - 0.2 × S) 2 / (P+0.8S)

[0080] in:

[0081] Q1: Preliminary runoff volume (mm);

[0082] P: Cumulative rainfall (mm);

[0083] S: Maximum retention (mm).

[0084] (c) Agricultural characteristic correction:

[0085] Tillage measure coefficient (1.2 for tilled plots, 0.8 for no-till plots, and 1 for other plots);

[0086] Crop growth period coefficient (0.9 for seedling stage, 0.6 for vigorous growth stage, 0.8 for maturity stage, and 1 for other conditions);

[0087] Calculate the runoff Q = Q1 × tillage coefficient × crop growth period coefficient.

[0088] 3.2 Pollutant Load Calculation

[0089] An improved load function method is adopted:

[0090] (a) Dissolved loading:

[0091] L_d = 0.01 × Q × C_soil × K_d;

[0092] L_d: Dissolved loading (kg / ha);

[0093] C_soil: Nitrogen and phosphorus concentration in the topsoil (mg / kg);

[0094] K_d: Partition coefficient of solubility. When C_soil is taken as the nitrogen concentration, the corresponding partition coefficient of solubility is 0.35. When C_soil is taken as the phosphorus concentration, the corresponding partition coefficient of solubility is 0.15.

[0095] (b) Adsorbed loading:

[0096] L_a=Q×K_erosion×C_soil×(1-K_d);

[0097] L_a: Adsorbed state loading;

[0098] K_erosion: Soil erosion coefficient (which can be calculated using the USLE model).

[0099] (c) Pollutant migration simulation

[0100] A one-dimensional slope flow model was established to simulate nitrogen and phosphorus migration processes. The model considered four main processes: diffusion, convection, infiltration loss, and biodegradation. The diffusion coefficient was determined based on the pollutant type (nitrate nitrogen 1.2 × 10⁻⁶). -4 m 2 / s, phosphorus 0.8×10 -4 m 2 The runoff velocity ( / s) was calculated using the Manning formula and corrected for slope based on DEM data.

[0101] Step 4: Construct a pollution risk prediction model with multi-node spatiotemporal characteristics

[0102] 1. Data Input Standards

[0103] The pollution risk prediction model receives data from both ground-based monitoring nodes and water monitoring nodes.

[0104] The ground monitoring node data includes: three meteorological data points (rainfall, wind speed, and temperature) collected every 5 minutes, and three soil data points (humidity, nitrogen concentration, and phosphorus concentration).

[0105] The water monitoring node data includes five water quality indicators collected simultaneously: pH value, COD value, ammonia nitrogen concentration, turbidity value, and dissolved oxygen value.

[0106] All ground monitoring node data and water monitoring node data are uploaded to the cloud platform in real time through the LoRa gateway, and missing values ​​are filled in using the spatiotemporal KNN algorithm.

[0107] The training nodes in the training and validation sets include ground monitoring nodes and water monitoring nodes.

[0108] 2. Feature Engineering Processing

[0109] Spatial features are standardized and encoded:

[0110] Land use types of ground monitoring nodes are converted into 5-dimensional one-hot vectors (arable land)

[0111] [1,0,0,0,0], woodland [0,1,0,0,0], orchard [0,0,1,0,0], bare land [0,0,0,1,0], rural roads [0,0,0,0,1], etc.), and the terrain features (elevation, slope, aspect) are normalized using Min-Max.

[0112] 3. Construct a pollution risk prediction model

[0113] The pollution risk prediction model adopts a dual-channel hybrid architecture:

[0114] 1) Timing processing channel:

[0115] The temporal processing path includes sequentially configured 1D-CNN layers, BiLSTM layers, and fully connected layers.

[0116] A 1D-CNN layer (kernel_size=3) extracts local features;

[0117] The BiLSTM layer (hidden_size=64) captures long-term dependencies;

[0118] The fully connected layer outputs the predicted pollution risk level.

[0119] The rainfall, wind speed, air temperature, humidity, nitrogen concentration, phosphorus concentration, and land use type of the ground monitoring nodes in the training set, as well as the pH value, COD value, ammonia nitrogen concentration, turbidity value, and dissolved oxygen value of the water monitoring nodes in the training set, are input into the time-series processing channel. The data are processed sequentially through a 1D-CNN layer, a BiLSTM layer, and a fully connected layer to output the predicted pollution risk level of each node for the next 24 / 48 / 72 hours.

[0120] 2) Spatial association channel:

[0121] Graph attention network (3-head attention) is used to model the pollution diffusion relationship between nodes.

[0122] Based on the ground monitoring nodes and water monitoring nodes in the training set, node feature matrices and edge feature matrices are constructed. The edge weights in the edge feature matrices are calculated based on hydrological connectivity. The node feature matrices and edge feature matrices are input into a graph attention network to obtain the predicted COD values ​​and predicted ammonia nitrogen concentrations for each water monitoring node in the next 24 / 48 / 72 hours.

[0123] 4. Multi-task output layer

[0124] The model outputs two types of results simultaneously:

[0125] 1) Classification and forecast: Pollution risk level (level 1-3) for the next 24 / 48 / 72 hours

[0126] 2) Regression Prediction: Specific concentration values ​​of major pollutants (COD, ammonia nitrogen) for the next 24 / 48 / 72 hours.

[0127] A dynamic weighted loss function L is used to balance the two types of tasks:

[0128] L=0.6×L_class+0.3×L_reg+0.1×L_graph;

[0129] Where L_class is the classification loss of the temporal processing channel; L_reg is the regression loss function of the spatial association channel; and L_graph is the graph loss function of the spatial association channel.

[0130] 5. Training Optimization Scheme

[0131] Data augmentation: Training samples are generated using a sliding window method. Ground monitoring nodes and water monitoring nodes in the training set are selected from all ground monitoring nodes and water monitoring nodes using a sliding window method.

[0132] Optimizer: AdamW (lr = 0.001, weight_decay = 0.01);

[0133] Regularization: Dropout = 0.3 + LabelSmoothing = 0.1;

[0134] Early stopping mechanism: Training is terminated if the dynamic weight loss function L on the validation set does not decrease for 10 consecutive rounds.

[0135] Based on the trained pollution risk prediction model, the pollution risk level, COD value, and ammonia nitrogen concentration for each node are predicted for the next 24 / 48 / 72 hours.

[0136] 6. Deployment of Predictive Services for Pollution Risk Prediction Models

[0137] Execute automatically every 5 minutes:

[0138] 1) Check data quality (missing rate <15%)

[0139] 2) Trigger pollution risk prediction model inference (GPU accelerated)

[0140] 3) Based on the predicted pollution risk level, predicted COD value, and predicted ammonia nitrogen concentration for each node in the next 24 / 48 / 72 hours, a heat map of the risk of formation and early warning information are generated.

[0141] 4) Push to the monitoring terminal via WebSocket.

[0142] Those skilled in the art will understand that steps 2, 3, and 4 in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the flow of the embodiments of the above methods.

[0143] Example 2:

[0144] In this embodiment, a computer device is also provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement steps 2, 3, and 4 in embodiment 1 above.

[0145] Example 3:

[0146] In this embodiment, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements steps 2, 3, and 4 in embodiment 1 above.

[0147] Example 4:

[0148] In this embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements steps 2, 3, and 4 in embodiment 1 above.

[0149] It should be noted that the specific embodiments described in this invention are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains can make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A method for real-time monitoring and intelligent early warning of non-point source pollution based on multi-source data fusion, characterized in that, Includes the following steps: Step 1: Deploy ground monitoring nodes and water monitoring nodes, and acquire data from the ground monitoring nodes and water monitoring nodes; Step 2: Preprocess the data from ground monitoring nodes and water monitoring nodes; Step 3: Calculate the CN value, maximum retention S, runoff Q, dissolved load, and adsorbed load; Step 4: Construct and train a pollution risk prediction model. Based on the trained pollution risk prediction model, obtain the predicted pollution risk level for each node in the next 24 / 48 / 72 hours, as well as the predicted COD value and ammonia nitrogen concentration for each water monitoring node in the next 24 / 48 / 72 hours.

2. The method for real-time monitoring and intelligent early warning of non-point source pollution based on multi-source data fusion according to claim 1, characterized in that, The ground monitoring nodes are equipped with meteorological monitoring units, soil parameter acquisition units, and runoff monitoring modules. The ground monitoring nodes collect rainfall, wind speed, air temperature, humidity, nitrogen concentration, phosphorus concentration, flow velocity, and flow rate. The water monitoring nodes are equipped with multi-parameter water quality analyzers and hydrodynamic sensors. The water monitoring nodes collect pH value, turbidity value, COD value, ammonia nitrogen concentration value, dissolved oxygen value, water depth pressure value, flow velocity value, and water depth value.

3. The method for real-time monitoring and intelligent early warning of non-point source pollution based on multi-source data fusion according to claim 1, characterized in that, The preprocessing includes outlier correction and time alignment.

4. The method for real-time monitoring and intelligent early warning of non-point source pollution based on multi-source data fusion according to claim 1, characterized in that, The CN value The calculation includes the following steps: When the ground monitoring node data are cultivated land, forest land, orchard, bare land, and rural roads, the corresponding basic CN values ​​are 78, 55, 70, 85, and 92, respectively. When the humidity monitored by the ground monitoring node exceeds 30% of the field capacity, the baseline CN value increases by 5; when the humidity is below 10% of the field capacity, the baseline CN value decreases by 3. The baseline CN value increases by an additional 12 within 7 days after fertilization at the ground monitoring node; the baseline CN value decreases by 5 when the ground monitoring node is located in the orchard due to the layer of dead branches and leaves. Obtain the adjusted CN value.

5. The method for real-time monitoring and intelligent early warning of non-point source pollution based on multi-source data fusion according to claim 4, characterized in that, The maximum retention S is calculated based on the following formula: Maximum retention S = (25400 / CN_adjusted) - 254, where CN_adjusted is the adjusted CN value; The runoff volume Q is calculated based on the following steps: Calculate the initial runoff volume Q1: Q1=(P-0.2×S) 2 / (P+0.8S); Where P represents the cumulative rainfall; Calculate runoff Q = Q1 × tillage coefficient × crop growth period coefficient; The dissolved load is calculated based on the following formula: L_d = 0.01 × Q × C_soil × K_d; L_d represents the dissolved loading rate; C_soil represents the nitrogen and phosphorus concentration in the topsoil. K_d is the solution allocation coefficient; The adsorbed state loading is calculated based on the following formula. L_a=Q×K_erosion×C_soil×(1-K_d); L_a represents the adsorbed loading state; K_erosion is the soil erosion coefficient.

6. The method for real-time monitoring and intelligent early warning of non-point source pollution based on multi-source data fusion according to claim 2, characterized in that, The construction and training of the pollution risk prediction model includes the following steps: Construct training and validation sets, with training nodes in both sets including ground monitoring nodes and water monitoring nodes; A pollution risk prediction model is constructed, which includes a time-series processing channel and a spatial correlation channel. The pollution risk prediction model is trained based on the training set.

7. The method for real-time monitoring and intelligent early warning of non-point source pollution based on multi-source data fusion according to claim 6, characterized in that, The time-series processing channel includes a 1D-CNN layer, a BiLSTM layer, and a fully connected layer arranged in sequence. The rainfall, wind speed, air temperature, humidity, nitrogen concentration, phosphorus concentration, and land use type of the ground monitoring nodes in the training set, as well as the pH value, COD value, ammonia nitrogen concentration, turbidity value, and dissolved oxygen value of the water monitoring nodes in the training set, are input into the time-series processing channel and processed sequentially through the 1D-CNN layer, BiLSTM layer, and fully connected layer to output the predicted pollution risk level of each node for the next 24 / 48 / 72 hours. The spatial association channel includes a graph attention network. Based on the ground monitoring nodes and water monitoring nodes in the training set, a node feature matrix and an edge feature matrix are constructed. The node feature matrix and the edge feature matrix are input into the graph attention network to obtain the predicted COD value and the predicted ammonia nitrogen concentration for each water monitoring node in the next 24 / 48 / 72 hours.

8. The method for real-time monitoring and intelligent early warning of non-point source pollution based on multi-source data fusion according to claim 7, characterized in that, The dynamic weight loss function L for constructing the pollution risk prediction model is based on the following formula: L=0.6×L_class+0.3×L_reg+0.1×L_graph; Where L_class is the classification loss of the temporal processing channel; L_reg is the regression loss function of the spatial association channel; and L_graph is the graph loss function of the spatial association channel.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements steps 2 to 4 of the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements steps 2 to 4 of the method according to any one of claims 1 to 8.

Images