A method and system for releasing risk early warning of non-point source pollution in a water-level fluctuation zone of the Three Gorges Reservoir

Abstract

This invention provides a method and system for early warning of non-point source pollution release risks in the drawdown zone of the Three Gorges Reservoir area, relating to the field of non-point source pollution monitoring and early warning technology. The method includes: acquiring and preprocessing multi-temporal multispectral image data of the drawdown zone to extract vegetation stress characteristic indices and soil spectral characteristics; constructing a synergistic inversion model of vegetation stress indices and soil organic matter content to obtain the spatial distribution of soil organic matter content; establishing a soil organic matter accumulation-release coupled model by combining water level scheduling plans and topographic data to simulate the release flux and its spatiotemporal distribution under different water storage scenarios; constructing a non-point source pollution release risk index and classifying early warning levels, and outputting the risk warning results. This invention achieves synergistic monitoring of vegetation stress status and soil organic matter content in the drawdown zone, improves the spatiotemporal continuity of monitoring, and provides decision support for early warning and control of non-point source pollution release risks under water storage conditions.

Description

Technical Field

[0001] This invention relates to the field of non-point source pollution monitoring and early warning technology, and in particular to a method and system for early warning of non-point source pollution release risks in the drawdown zone of the Three Gorges Reservoir area. Background Technology

[0002] The Three Gorges Reservoir area in the upper reaches of the Yangtze River is an important ecological barrier and water conservation area in my country. After the completion of the Three Gorges Project, the world's largest drawdown zone, with a vertical drop of 30 meters and a total area of ​​over 300 square kilometers, was formed. Due to the reservoir's winter storage and summer discharge operation, the soil in the drawdown zone experiences a periodic flooding-drying process: during the dry period, vegetation recovers and grows, and soil organic matter gradually accumulates; during the storage period, when the water level rises and the soil is submerged, the accumulated organic matter mineralizes and decomposes under flooded conditions, releasing nutrients into the reservoir and becoming a significant endogenous source of water pollution. Therefore, monitoring and early warning of the risk of organic matter release from the soil in the drawdown zone is of great significance.

[0003] Currently, some related technologies have been disclosed, such as patent CN103918373A, which uses multi-level ecological units to physically intercept pollutants. However, this method focuses on spatial retention and cannot quantitatively assess the risk of organic matter release during water level rise, thus it is considered post-event remediation. Ecological restoration experiments in drawdown zones stabilize soil through vegetation restoration, but they cannot directly monitor the dynamic changes in soil organic matter.

[0004] Existing technologies have the following shortcomings: First, existing patents mostly focus on engineering interception or ecological restoration, lacking quantitative early warning capabilities for pollution release risks. Second, sampling-based monitoring methods struggle to achieve large-scale, high-frequency dynamic monitoring. Third, existing technologies have failed to establish a correlation model between vegetation stress, organic matter accumulation, and water level rise release. Fourth, for the Three Gorges Reservoir's winter storage and summer drainage scheduling model, there is a lack of solutions that incorporate the flooding-drainage process into risk assessment. Summary of the Invention

[0005] The main objective of this invention is to provide a method and system for early warning of non-point source pollution release risks in the drawdown zone of the Three Gorges Reservoir area, which solves the problems of low efficiency, limited coverage, poor model adaptability, and poor data coupling in existing monitoring methods.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a method for early warning of non-point source pollution release risks in the drawdown zone of the Three Gorges Reservoir area, comprising the following steps: S1: Acquire multi-temporal multispectral image data of the drawdown zone, preprocess the images to generate standard reflectance images of the study area; based on the standard reflectance images, extract vegetation stress characteristic indices and soil spectral characteristic bands. S2: Construct a synergistic inversion model of vegetation stress index and soil organic matter content. Input vegetation stress characteristic index and soil spectral characteristic band, and output spatial distribution map of soil organic matter content in the drawdown zone and distribution map of vegetation stress level. S3: Establish a coupled model of soil organic matter accumulation and release in the drawdown zone, and simulate the release flux and its spatiotemporal distribution of soil organic matter under different water storage scenarios by combining water level scheduling plans and topographic data. S4: Construct a risk index for non-point source pollution release in the drawdown zone, classify risk warning levels based on the simulation results of release flux, generate a risk warning map, and output a warning report; S5: Generate governance decision support information based on early warning results, and achieve closed-loop optimization of monitoring-early warning-governance through a model dynamic update mechanism.

[0007] In the preferred embodiment, step S1 specifically includes: S101: Acquire multi-temporal multispectral images of the target drawdown zone area by UAV or satellite, wherein the multispectral images include at least blue light band, green light band, red light band, red edge band and near-infrared band; the multi-temporal images include at least three key time nodes: the early, middle and late stages of the drawdown zone drying period. S102: Radiometric calibration and atmospheric correction are performed on the acquired multispectral images to convert the original DN values ​​into surface reflectance; the digital elevation model is used to perform geometric fine correction on the images, and water bodies and buildings are removed by masking to generate a standard reflectance image that only contains the drawdown zone. S103: Based on standard reflectance images, normalized vegetation index NDVI, red-edge normalized vegetation index NDVIred-edge, leaf area index LAI, and vegetation moisture index NDWI were calculated as vegetation stress characteristic indices, and the vegetation stress coefficient VSC was calculated by combining them with historical normal values ​​of the same period. S104: Based on standard reflectance images, red light, near-infrared, and red edge band reflectance were extracted, along with soil modified vegetation index (SAVI), normalized differential moisture index (NDMI), red edge absorption depth, and red edge absorption area as characteristic bands of the soil spectrum.

[0008] In the preferred embodiment, step S2 specifically includes: S201: Construct a training sample set for the co-inversion of vegetation stress index and soil organic matter content; set up typical sampling points in the study area, collect surface soil samples, and determine the soil organic matter content through laboratory chemical analysis; at the same time, extract the vegetation stress characteristic index and soil spectral characteristic bands of the corresponding sampling points to form a training sample set containing input features and output labels. S202: Construct a collaborative inversion model based on random forest or support vector machine; the model input features include vegetation stress characteristic index and soil spectral characteristic bands, and the output layer is soil organic matter content; S203: The trained co-inversion model is applied to the standard reflectance image of the entire study area, and the soil organic matter content is calculated pixel by pixel to generate a spatial distribution map of soil organic matter content in the drawdown zone; at the same time, the drawdown zone is divided into different stress levels according to the numerical range of the vegetation stress coefficient (VSC) to generate a vegetation stress level distribution map. S204: Verify the accuracy of the generated spatial distribution map of soil organic matter content and vegetation stress level distribution map; if the accuracy does not meet the requirements, adjust the model parameters or supplement the training samples, and retrain and invert the model. S205: Construct an active learning sampling optimization mechanism to achieve synergistic improvement in model accuracy and sampling efficiency, including: based on the calculation results of inversion uncertainty, identify high uncertainty regions for encrypted sampling, update the training sample set and retrain the model.

[0009] In the preferred scheme, the random forest-based collaborative inversion model constructed in step S202 specifically includes: S2021: Perform Bootstrap sampling on the training sample set to generate K training subsets; S2022: For each training subset, construct a decision regression tree; when splitting at each node, randomly select m features from all input features, and select the optimal feature for splitting; S2023: Repeat steps S2021-S2022 to generate K decision trees, forming a random forest; for a new input sample, take the average of all decision tree predictions as the final prediction result. S2024: Calculate the prediction error and feature importance of the model using out-of-bag data, and sort and select key features based on feature importance.

[0010] In the preferred scheme, the active learning sampling optimization mechanism in step S205 specifically includes: S2051: Based on the preliminary inversion results, the inversion uncertainty of each pixel is calculated. The sources of uncertainty include spatial sparsity of spectral features, terrain complexity, and vegetation heterogeneity. The formula for calculating the uncertainty index is: ; in, It is a comprehensive uncertainty index. The Mahalanobis distance between the pixel's spectral features and the nearest training sample. This represents the local topographic variation coefficient. To account for the local NDVI variance, all indicators were normalized to the range of 0-1. , , These represent the maximum value of Mahalanobis distance, the maximum value of local topographic variation coefficient, and the maximum value of local NDVI variance, respectively. S2052: Generate a sampling uncertainty distribution map, identify high uncertainty areas, plan denser sampling routes, and add sampling points; S2053: Merge the newly added sampling point data with the original training samples, retrain the model, and improve the inversion accuracy; S2054: Establish a historical sampling database, and use a reinforcement learning strategy to dynamically adjust the weight coefficients of the uncertainty index to optimize subsequent sampling decisions.

[0011] In the preferred embodiment, step S3 specifically includes: S301: Obtain digital elevation model data and reservoir scheduling plan data for the study area; calculate the inundation range and water depth distribution of the drawdown zone under different water level conditions based on DEM data and water level elevation data; S302: Construct a model for soil organic matter accumulation in the drawdown zone to describe the temporal variation of soil organic matter content during the drying period; the model expression is: ; in, For the first dry period Soil organic matter content of the day This represents the initial soil organic matter content during the drying period. This is the organic matter accumulation rate coefficient; S303: Construct a soil organic matter release model in the drawdown zone to describe the release pattern of soil organic matter during flooding; the model expression is: ; in, For organic matter release flux, This refers to the soil organic matter content at the start of flooding. Water temperature This refers to the depth of floodwater. This refers to the duration of the flooding. S304: Substitute the spatial distribution map of soil organic matter content, inundation range and water depth distribution into the release model to simulate the release flux and its spatiotemporal distribution of organic matter under different water storage scenarios, and generate a spatial distribution map of release flux. S305: Construct a digital twin of non-point source pollution in the drawdown zone to achieve closed-loop dynamic evolution, including grid-based modeling based on the ST-GCN architecture, introducing the PINN physical information neural network mechanism, multi-scenario inference, and online learning and updating.

[0012] In the preferred embodiment, step S305, constructing a digital twin of non-point source pollution in the drawdown zone, specifically includes: S3051: Based on the inverted spatial distribution map of soil organic matter content, the accumulation-release model, and real-time access to water level and meteorological data, a digital twin is constructed using a spatiotemporal graph convolutional neural network architecture. The drawdown zone is divided into grid cells, with each cell serving as a graph node. Node features include SOM content, VSC value, slope, distance from the river channel, and historical release flux. Edge weights are calculated based on hydrological connectivity. S3052: Introducing a physical information neural network mechanism, the partial differential equation of organic matter accumulation-release is added as a physical constraint term to the loss function; S3053: Built-in multiple preset scenarios, supports managers to customize parameters for simulation, and outputs release flux change curves and risk level evolution diagrams; S3054: After acquiring new monitoring data each time, the deviation between the predicted value and the measured value is automatically calculated, and the model parameters are updated using an online learning algorithm.

[0013] In the preferred embodiment, step S4 specifically includes: S401: Construct a risk index for non-point source pollution release in the drawdown zone, comprehensively considering multiple factors such as soil organic matter content, vegetation stress level, topographic slope, and distance from the river or reservoir bay; the risk index calculation formula is: ; in, This represents the normalized soil organic matter content. This represents the vegetation stress coefficient. This is the normalized terrain slope. This is the normalized distance from the river channel or reservoir bay; , , , The weight coefficients for each factor are determined using the analytic hierarchy process (AHP) or the entropy weight method. S402: Based on the risk index The numerical range is used to divide the drawdown zone into multiple risk levels, generating a distribution map of the risk levels of non-point source pollution release in the drawdown zone. S403: Generate an early warning report based on the risk level distribution map and the release flux simulation results; the early warning report shall include at least: the spatial distribution of high-risk areas and extremely high-risk areas, area statistics, the estimated total release of organic matter, the impact assessment on downstream water quality, and treatment recommendations for different risk levels; S404: Push the generated early warning report to the relevant management department through the GIS platform or mobile application.

[0014] In the preferred embodiment, step S5 specifically includes: S501: Based on the risk level distribution and risk source identification results, a multi-criteria decision analysis method is used to generate a governance priority ranking; S502: Construct a multi-objective optimization model for governance solutions to achieve Pareto optimal decisions in terms of ecological benefits, economic benefits, and governance cycle; S503: Establish a monitoring-feedback-update mechanism, repeat steps S1-S3 during the dry season of the following year, to achieve online updates and accuracy improvement of the model.

[0015] Secondly, the present invention provides a multispectral inversion-based area source pollution release risk early warning system for implementing the method, comprising: The multispectral image acquisition and preprocessing module is used to control the UAV equipped with a multispectral camera to acquire multi-temporal images of the study area, and to perform radiometric calibration, atmospheric correction, geometric correction and drawdown mask extraction, and output standard reflectance images. The vegetation stress index calculation module is used to calculate the vegetation index based on standard reflectance imagery, and construct the vegetation stress coefficient (VSC) by comparing it with historical data from the same period, and generate a vegetation stress level distribution map. The soil organic matter inversion module is used to extract spectral features of sampling points, construct a collaborative inversion model based on random forest, invert soil organic matter content pixel by pixel, and generate a spatial distribution map; this module has a built-in active learning sampling optimization mechanism. The accumulation-release simulation module is used to fit the organic matter accumulation rate based on multi-time back-evolution results, and to construct an accumulation-release coupled model by combining DEM data and water level scheduling plan to simulate the organic matter release flux and its spatiotemporal distribution under different water storage scenarios. The risk warning module is used to construct a risk release index, classify risk warning levels, generate risk level distribution maps and warning reports, and push them to management departments through the WebGIS platform and mobile terminals. The system management module is used for user permission management, data interfaces, log auditing, and system monitoring.

[0016] This invention provides a method and system for early warning of non-point source pollution release risks in the drawdown zone of the Three Gorges Reservoir area. It acquires vegetation stress characteristics and soil spectral characteristics in the drawdown zone through multi-temporal multispectral imagery, constructs a soil organic matter synergistic inversion model, and establishes a soil organic matter accumulation-release coupled model by combining water level scheduling plans and topographic data, thereby generating a non-point source pollution release risk index and early warning results. Compared with existing technologies, this invention achieves synergistic monitoring of vegetation stress status and dynamic changes in soil organic matter in the drawdown zone, and can provide early warnings of pollution release risks under different water storage scenarios, thus improving the continuity, timeliness, and applicability of monitoring. Attached Figure Description

[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is an overall flowchart of the method of the present invention; Figure 2 This is a schematic diagram illustrating the calculation of the vegetation stress coefficient in this invention; Figure 3 This is a structural diagram of the random forest inversion model in this invention; Figure 4 This is a diagram showing the simulation results of the soil organic matter accumulation-release process and release flux under different water storage scenarios during the winter storage and summer drainage cycle in this invention. Figure 5 This is the final risk level distribution map of non-point source pollution release in the drawdown zone generated by this invention. Detailed Implementation

[0018] Example 1 like Figure 1-5 As shown, a method for early warning of non-point source pollution release risks in the drawdown zone of the Three Gorges Reservoir area includes the following steps: S1: Acquire multi-temporal multispectral image data of the drawdown zone, preprocess the images to generate standard reflectance images of the study area; based on the standard reflectance images, extract vegetation stress characteristic indices and soil spectral characteristic bands.

[0019] S2: Construct a synergistic inversion model of vegetation stress index and soil organic matter content. Input vegetation stress characteristic index and soil spectral characteristic bands, and output a spatial distribution map of soil organic matter content in the drawdown zone and a distribution map of vegetation stress level.

[0020] S3: Establish a coupled model of soil organic matter accumulation and release in the drawdown zone, and combine water level scheduling plan and topographic data to simulate the release flux of soil organic matter and its spatiotemporal distribution under different water storage scenarios.

[0021] S4: Construct a risk index for non-point source pollution release in the drawdown zone, classify risk warning levels based on the simulation results of release flux, generate a risk warning map, and output a warning report.

[0022] S5: Generate governance decision support information based on early warning results, and achieve closed-loop optimization of monitoring-early warning-governance through a model dynamic update mechanism.

[0023] This embodiment preprocesses multi-temporal multispectral image data of the drawdown zone, extracts vegetation stress characteristic indices and soil spectral characteristic bands, constructs a synergistic inversion model of vegetation stress index and soil organic matter content, and a soil organic matter accumulation-release coupling model of the drawdown zone, constructs a non-point source pollution release risk index of the drawdown zone, classifies risk warning levels, and generates warning results of risk information. It solves the technical problems of existing technologies being unable to coordinate the monitoring of dynamic changes in vegetation stress and soil organic matter, and being unable to quantitatively predict the risk of pollutant release after water storage. Through a dynamic update mechanism, it realizes closed-loop optimization of monitoring, early warning and treatment, realizes early warning of non-point source pollution in the drawdown zone, improves inversion accuracy and work efficiency, improves coverage and model adaptability, and improves the accuracy of risk identification.

[0024] This embodiment selects the drawdown zone near the confluence of a tributary in the Three Gorges Reservoir area as the research object. The drawdown zone in this area is about 200-500 meters wide and about 5 kilometers long, and has typical winter storage and summer drainage processes and obvious vegetation gradient distribution characteristics.

[0025] In the preferred embodiment, step S1 specifically includes: S101: Acquire multi-temporal multispectral image data of the drawdown zone, preprocess the images, and generate standard reflectance images of the study area.

[0026] S1011: At three key time points—June 15, 2023 (early dry season), August 10, 2023 (mid-dry season), and October 5, 2023 (late dry season)—multispectral images of the study area were acquired using a DJI M300 RTK drone equipped with a MicaSense RedEdge-MX multispectral camera, flying at an altitude of 200 meters with 80% forward overlap and 70% lateral overlap. The multispectral camera includes five bands: blue (475nm center, 20nm bandwidth), green (560nm center, 20nm bandwidth), red (668nm center, 10nm bandwidth), red-edge (717nm center, 10nm bandwidth), and near-infrared (842nm center, 40nm bandwidth). The ground resolution of the images reached 0.15 meters per pixel. S1012: Acquire 1:2000 digital elevation model (DEM) data of the study area with a resolution of 2 meters; collect the Three Gorges Reservoir's 2023-2024 scheduling plan, including monthly water level and elevation data (145m-175m, fluctuation rate 0.5-1.0 m / day).

[0027] Step S102: Import the acquired multispectral image into Pix4Dmapper software for radiometric calibration and geometric correction. First, perform radiometric calibration using the reflectance panel after whiteboard calibration, converting the original DN values ​​into radiance values, and then converting them into surface reflectance using the reflectance calibration file provided by MicaSense. Orthorectify the image using the DEM of the study area, controlling the correction accuracy within 0.5 pixels to eliminate geometric distortion caused by topographic relief. The water area was extracted by manual interpretation combined with the NDWI index (threshold set to 0.2), and a drawdown zone mask was generated. Only the area within the elevation range of 145m-175m was retained as the study area to obtain the standard reflectance image. After radiometric calibration and geometric correction, the standard reflectance image that meets the requirements of subsequent inversion analysis was obtained.

[0028] Radiometric calibration and atmospheric correction were performed on the acquired multispectral images to convert the original DN values ​​into surface reflectance. A digital elevation model was used to perform geometric fine correction on the images, and water bodies and buildings were removed through masking to generate a standard reflectance image that only includes the drawdown zone.

[0029] S103: Based on standard reflectance images, normalized vegetation index NDVI, red-edge normalized vegetation index NDVIred-edge, leaf area index LAI, and vegetation moisture index NDWI were calculated as vegetation stress characteristic indices, and the vegetation stress coefficient VSC was calculated by combining them with historical normal values ​​of the same period. The formula for calculating NDVI is: ; in, For near-infrared reflectivity, Reflectivity in the red light band This refers to the reflectivity of the red-edge band.

[0030] Obtain historical data for the same period over the past five years for this region. Data, calculation of historical data for the same period mean In this embodiment, the historical data for the study area in June... The mean is 0.62.

[0031] The VSC calculation formula is: ; in, This represents the current vegetation index. This represents the average normal vegetation index for the same period in history.

[0032] Vegetation stress levels are classified based on VSC values: VSC < 0.1 indicates no stress, 0.1 ≤ VSC < 0.3 indicates mild stress, 0.3 ≤ VSC < 0.5 indicates moderate stress, and VSC ≥ 0.5 indicates severe stress. A vegetation stress level distribution map is generated, as shown below. Figure 2 As shown in (c), the red area in the figure represents the area of ​​severe vegetation stress (VSC>0.5), which is mainly concentrated within 50-100 meters from the river channel, covering an area of ​​about 42 hectares. This area is affected by periodic flooding, and the vegetation recovery is poor.

[0033] S104: Extract sensitive spectral features related to soil organic matter content from standard reflectance images, including the original reflectance of red, near-infrared, and red-edge bands, as well as the Soil Adjusted Vegetation Index (SAVI) and Normalized Difference Moisture Index (NDMI) constructed through band operations. At the same time, construct red-edge related feature parameters, including red-edge band reflectance, red-edge normalized index, and red-edge feature quantities constructed based on the differences between adjacent bands, to form a multidimensional soil spectral feature dataset.

[0034] Step S2 constructs a collaborative inversion model to generate spatial distribution maps of soil organic matter and vegetation stress.

[0035] In the preferred embodiment, step S2 specifically includes: S201: Construct a training sample set for the co-inversion of vegetation stress index and soil organic matter content; set up typical sampling points in the study area, collect surface soil samples, and determine the soil organic matter content through laboratory chemical analysis; at the same time, extract the vegetation stress characteristic index and soil spectral characteristic bands of the corresponding sampling points to form a training sample set containing input features and output labels.

[0036] In August 2023, field ground sampling was conducted simultaneously. Fifty typical sampling points were set up in the study area using stratified random sampling, comprehensively covering areas with different vegetation stress levels, terrain slopes, and distances from river channels. At each sampling point, a five-point sampling method was used to collect soil samples from the top 0-20 cm layer. After air-drying, grinding, and sieving through a 0.25 mm sieve, the soil organic matter content was determined using the potassium dichromate titration method (external heating method), with a detection accuracy of ±0.1%. At the same time, the vegetation stress characteristic index and soil spectral characteristics corresponding to each sampling point were extracted to form a 25-dimensional input feature vector. Using the measured soil organic matter content as the output label, a training sample set was constructed, which was divided into a training set (40 samples) and a validation set (10 samples) in an 8:2 ratio.

[0037] S202: Construct a collaborative inversion model based on random forest or support vector machine; the model input features include vegetation stress characteristic index and soil spectral characteristic bands, and the output layer is soil organic matter content.

[0038] The constructed model was trained using the scikit-learn library (version 1.2.0) in Python 3.9 to build a random forest regression model. The main parameters were set as follows: number of decision trees n_estimators = 200, maximum depth max_depth = 12, minimum number of samples per node split min_samples_split = 5, minimum number of samples per leaf node min_samples_leaf = 2, maximum number of features max_features = 'sqrt', and random seed random_state = 42. The model was trained using the training set, and 10-fold cross-validation was used to optimize the model parameters.

[0039] S203: The trained co-inversion model is applied to the standard reflectance image of the entire study area, and the soil organic matter content is calculated pixel by pixel to generate a spatial distribution map of soil organic matter content in the drawdown zone; at the same time, the drawdown zone is divided into different stress levels according to the numerical range of the vegetation stress coefficient (VSC) to generate a vegetation stress level distribution map.

[0040] The trained model was applied to standard reflectance imagery of the study area to calculate soil organic matter content pixel by pixel, generating a spatial distribution map of soil organic matter content with a spatial resolution of 0.15 meters. The validation set accuracy test results were: coefficient of determination R² = 0.86, root mean square error RMSE = 0.28%, mean absolute error MAE = 0.21%, indicating that the model inversion accuracy meets the requirements for operational applications. In this embodiment, the soil organic matter content in the study area ranged from 1.2% to 3.8%, with a mean of 2.3%. High-value areas (>3.0%) were mainly distributed near tributary confluences and in low-lying areas, covering an area of ​​approximately 35 hectares, with a spatial overlap of 78% with high-value areas of vegetation stress.

[0041] After training, the importance of each input feature is calculated, and the results are as follows: Figure 3 As shown on the left, NDVI and red-edge NDVI contribute the most to the model, with importance of 32% and 28% respectively, while red reflectance has an importance of 18% and near-infrared reflectance has an importance of 22%.

[0042] S204: Verify the accuracy of the generated spatial distribution map of soil organic matter content and vegetation stress level distribution map; if the accuracy does not meet the requirements, adjust the model parameters or supplement the training samples, and retrain and invert the model.

[0043] S205: Construct an active learning sampling optimization mechanism to achieve synergistic improvement in model accuracy and sampling efficiency, including: based on the calculation results of inversion uncertainty, identify high uncertainty regions for encrypted sampling, update the training sample set and retrain the model.

[0044] In the preferred scheme, the random forest-based collaborative inversion model constructed in step S202 specifically includes: S2021: Perform Bootstrap sampling on the training sample set to generate K training subsets.

[0045] S2022: For each training subset, construct a decision regression tree; when splitting at each node, randomly select m features from all input features, and select the optimal feature for splitting.

[0046] S2023: Repeat steps S2021-S2022 to generate K decision trees, forming a random forest; for a new input sample, take the average of all the predictions from the decision trees as the final prediction result.

[0047] S2024: Calculate the prediction error and feature importance of the model using out-of-bag data, and sort and select key features based on feature importance.

[0048] In the preferred scheme, the active learning sampling optimization mechanism in step S205 specifically includes: S2051: Based on the preliminary inversion results, the inversion uncertainty of each pixel is calculated. The sources of uncertainty include spatial sparsity of spectral features, terrain complexity, and vegetation heterogeneity. The formula for calculating the uncertainty index is: ; in, It is a comprehensive uncertainty index. The Mahalanobis distance between the pixel's spectral features and the nearest training sample. This represents the local topographic variation coefficient. To account for the local NDVI variance, all indicators were normalized to the range of 0-1. , , These represent the maximum value of Mahalanobis distance, the maximum value of local topographic variation coefficient, and the maximum value of local NDVI variance, respectively.

[0049] S2052: Generate a sampling uncertainty distribution map, identify high uncertainty areas, plan denser sampling routes, and add sampling points.

[0050] S2053: Merge the newly added sampling data with the original training samples, retrain the model, and improve the inversion accuracy.

[0051] S2054: Establish a historical sampling database, and use a reinforcement learning strategy to dynamically adjust the weight coefficients of the uncertainty index to optimize subsequent sampling decisions.

[0052] Through the above steps, based on the preliminary inversion results, the comprehensive uncertainty index of each pixel (fusing spectral feature Mahalanobis distance, topographic variation coefficient, and NDVI local variance) was calculated, and 10 high uncertainty areas (single area > 0.5 hectares) were identified. 3-5 additional encrypted sampling points were added to each area. After merging the 40 new sampling points with the original 50 samples, the model was retrained, which improved the model accuracy, R²=0.91, reduced RMSE, and improved sampling efficiency.

[0053] Step S3 establishes a cumulative-release coupling model to simulate the release flux under different water storage scenarios.

[0054] In the preferred embodiment, step S3 specifically includes: S301: Obtain digital elevation model data and reservoir scheduling plan data for the study area; based on DEM data and water level elevation data, calculate the inundation range and water depth distribution of the drawdown zone under different water level conditions.

[0055] Based on the DEM data of the study area and the 2023-2024 water level scheduling plan of the Three Gorges Reservoir, this embodiment calculates the inundation range, water depth distribution, and inundation start time and duration of each pixel under different water level elevations, providing basic topographic and hydrological data for subsequent simulations.

[0056] S302: Construct a model for soil organic matter accumulation in the drawdown zone to describe the temporal variation of soil organic matter content during the drying period. The model expression is as follows:

[0057] in, For the first dry period Soil organic matter content of the day This represents the initial soil organic matter content during the drying period. This is the organic matter accumulation rate coefficient.

[0058] In this embodiment, based on the soil organic matter content retrieved from three periods of multispectral inversion (2.1% on June 15, 2.4% on August 10, and 2.7% on October 5), the cumulative rate coefficient of organic matter during the drying period was calculated using the linear fitting method. The average cumulative rate in the study area of ​​this embodiment was 0.12% / month, and the high value area at the tributary confluence could reach 0.25% / month. The spatial variation coefficient of the cumulative rate was CV=0.35.

[0059] S303: Construct a soil organic matter release model in the drawdown zone to describe the release pattern of soil organic matter during flooding; referencing the experimental research results on soil organic matter release in the drawdown zone of the Three Gorges Reservoir area (indoor simulation experiment, water temperature 5-25℃, flooding time 0-180 days), construct an organic matter release model: ; in, Organic matter release flux (kg / m²) The percentage of soil organic matter at the start of flooding. The water temperature is (°C). In this example, the average winter water temperature is taken as 12°C. The duration of flooding (in days). Flood depth (m). Model validation. The average relative error is 12%.

[0060] S304: Substitute the spatial distribution map of soil organic matter content, inundation range and water depth distribution into the release model to simulate the release flux of organic matter and its spatiotemporal distribution under different water storage scenarios, and generate a spatial distribution map of release flux.

[0061] This embodiment sets up three typical water storage scenarios based on the Three Gorges Reservoir scheduling plan, and simulates the spatiotemporal distribution of soil organic matter release flux under different scenarios. The scenario settings and simulation results are as follows: 1) Scenario A (Normal water storage): Water storage begins on November 1st, reaches 175m on December 20th, with a storage rate of 0.5m / day, an average release flux of 1.8kg / m², and a total release of 156.8 tons.

[0062] 2) Scenario B (Pre-storage): Storage begins on October 15 and reaches 175m on December 5, with a storage rate of 0.7m / day, an average release flux of 2.1kg / m², and a total release of 219.5 tons, which is 40% higher than the normal scenario.

[0063] 3) Scenario C (Delayed water storage): Water storage begins on November 15th and reaches 175m by January 5th of the following year, with a storage rate of 0.4m / day, an average release flux of 1.2kg / m², and a total release of 109.8 tons, which is 30% less than the normal scenario.

[0064] Based on the DEM data of the study area and the three water level control schemes mentioned above, the flooding start time, flooding depth, and flooding duration for each pixel under different scenarios were calculated. Using the soil organic matter content distribution map generated in step S203 as input, the release model was substituted to calculate the organic matter release flux under each scenario. The results are as follows: Figure 4 As shown in (b)-(d).

[0065] Simulation results show that in Scenario A (normal water storage), the average release flux is 1.8 kg / m², and the total release is 156.8 tons; in Scenario B (early water storage), the average release flux is 2.1 kg / m², and the total release is 219.5 tons, an increase of 40% compared to the normal scenario; in Scenario C (delayed water storage), the average release flux is 1.2 kg / m², and the total release is 109.8 tons, a decrease of 30% compared to the normal scenario. High-risk areas (release flux greater than 2.5 kg / m²) are mainly distributed in the 145 m–160 m elevation range, covering an area of ​​approximately 35 hectares.

[0066] S305: A digital twin of non-point source pollution in the drawdown zone is constructed based on a spatiotemporal graph convolutional neural network (ST-GCN) architecture to achieve closed-loop dynamic evolution. The study area is divided into 100m×100m grid cells, and node features include SOM content, VSC value, slope, distance from the river channel, and historical release flux. Edge weights are calculated based on hydrological connectivity. A physical information neural network (PINN) mechanism is introduced, incorporating the organic matter accumulation-release partial differential equation as a physical constraint term into the loss function. This digital twin has 12 built-in preset scenarios, supports custom parameter extrapolation, and can complete simulation within 5 minutes, outputting release flux change curves and risk level evolution diagrams. The model parameters can be updated based on new monitoring data to improve the applicability of subsequent extrapolation results.

[0067] In the preferred embodiment, step S305, constructing a digital twin of non-point source pollution in the drawdown zone, specifically includes: S3051: Based on the inverted spatial distribution map of soil organic matter content, the accumulation-release model, and real-time access to water level and meteorological data, a digital twin is constructed using a spatiotemporal graph convolutional neural network architecture. The drawdown zone is divided into grid cells, with each cell serving as a graph node. Node features include SOM content, VSC value, slope, distance from the river channel, and historical release flux. Edge weights are calculated based on hydrological connectivity.

[0068] S3052: Introducing a physical information neural network mechanism, the partial differential equation of organic matter accumulation-release is added as a physical constraint term to the loss function.

[0069] S3053: Built-in multiple preset scenarios, supports administrators to customize parameters for simulation, and outputs release flux change curves and risk level evolution diagrams.

[0070] S3054: After acquiring new monitoring data each time, the deviation between the predicted value and the measured value is automatically calculated, and the model parameters are updated using an online learning algorithm.

[0071] Furthermore, a knowledge graph of non-point source pollution in the drawdown zone was constructed, and a graph neural network (GNN) source tracing algorithm was used to analyze the pollution sources in high-risk areas. In this embodiment, the pollution sources in the extremely high-risk area consist of: upstream agricultural non-point sources 42%, local accumulation in the drawdown zone 38%, urban point sources 15%, and atmospheric deposition 5%. The key pollution transmission path (contribution rate 23%) was identified as the north bank tributary → confluence → extremely high-risk area.

[0072] A1: Construct a knowledge graph of non-point source pollution in the drawdown zone. Entity types include: pollution source type (upstream farmland, urban point sources, local sources in the drawdown zone, atmospheric deposition), transport pathway (surface runoff, interflow, groundwater exchange), receptor unit (grid unit), and influencing factors (slope, vegetation, soil type). Relationship types include: contribution (contribution of pollution sources to grid units), connection (hydrological connection), and influence (regulation of processes by influencing factors).

[0073] A2: Based on the simulated spatial distribution of released flux, a graph neural network (GNN) source tracing algorithm is used to trace the contribution ratio of pollution sources in each high-risk grid cell. The core formula of the algorithm is: ; in, upstream source region For downstream grid Contribution rate, For hydrological connectivity weights, The hydrological connectivity weight in the denominator (from upstream) to downstream network ), Distance along the water flow path The distance along the water flow path in the denominator (from upstream) to downstream network ), For transmission attenuation factor, The transmission attenuation factor in the denominator comes from the upstream to downstream network. and For empirical parameters (in this embodiment, we take...) , ).

[0074] A3: The source tracing results show that the pollution sources in the extremely high-risk area (28.7 hectares) of the study area are as follows: upstream agricultural non-point source pollution accounts for 42% (mainly from farmland areas upstream of 3 tributaries), local cumulative pollution from the drawdown zone accounts for 38% (mainly from soil organic matter loss caused by vegetation degradation), urban point source pollution accounts for 15% (from the tailwater discharge outlet of 1 sewage treatment plant), and atmospheric deposition accounts for 5%.

[0075] A4: Generates a Sankey diagram for risk tracing, visually demonstrating the pollution transmission process from source to pathway to sink. The system automatically identifies key transmission pathways (such as the north bank tributary → sink inlet → extremely high-risk area pathway, which contributes 23%) and provides targeted governance recommendations: It is recommended to add an ecological interception zone before the north bank tributary flows into the reservoir, thereby reducing pollution input to the high-risk area.

[0076] Step S4 involves constructing a risk index, classifying warning levels, and generating warning reports.

[0077] S401: Taking into account soil organic matter content, vegetation stress level, topographic slope, and distance from the river channel, a release risk index is constructed ( ): ; in, To normalize soil organic matter content, , These are measured or retrieved values ​​of soil organic matter content. It is the minimum value of soil organic matter content in the study area. This is the maximum value of soil organic matter content in the study area; This represents the vegetation stress coefficient. To normalize the slope, (45° is the maximum slope in the study area) This is the measured slope value; To normalize the distance from the river channel, , Distance from the river channel The distance is 500m. The weights of each factor were determined using the analytic hierarchy process (AHP), with soil organic matter content having the highest weight (0.45), reflecting that organic matter accumulation is the core driver of risk release.

[0078] S402: Based on the RI value, the drawdown zone is divided into four risk levels: low risk (RI < 0.25), medium risk (0.25 ≤ RI < 0.5), high risk (0.5 ≤ RI < 0.75), and extremely high risk (RI ≥ 0.75). A risk level distribution map is generated, as follows. Figure 5 As shown in (a).

[0079] S403: Area statistics for each risk level: Low-risk area 45.2 hectares (21.7%), Medium-risk area 78.5 hectares (37.6%), High-risk area 56.3 hectares (27.0%), Very high-risk area 28.7 hectares (13.7%), Total drawdown zone area 208.7 hectares. Generate an area statistics bar chart, as follows. Figure 5 As shown in (b).

[0080] S404: Generate an early warning report based on the risk level distribution and flux simulation results. Set warning thresholds: a red warning is triggered when the area of ​​an extremely high-risk zone > 20 hectares or the estimated total release > 150 tons; an orange warning is triggered when the area of ​​a high-risk zone > 50 hectares or the estimated total release > 120 tons; and a yellow warning is triggered when the area of ​​a high-risk zone > 30 hectares. In this example, the extremely high-risk zone area is 28.7 hectares (exceeding the threshold by 43.5%), and the estimated total release is 156.8 tons (exceeding the threshold by 4.5%), triggering a red warning.

[0081] The generated early warning report will be pushed to management departments such as the Municipal Water Resources Bureau, the Ecological Environment Bureau, and the Three Gorges Reservoir Dispatch Center via the WebGIS platform and mobile application, with a push delay of less than 30 seconds. The report includes a risk distribution map (GeoTIFF format, spatial resolution 0.15 meters), risk area vector boundaries (Shapefile format), statistical data tables (CSV format), and a text report (PDF format). It is recommended to activate the emergency response mechanism and prioritize the treatment of extremely high-risk areas.

[0082] Table 1 Comparison of Simulated Flux Release Results under Various Water Storage Scenarios

[0083] Step S5 generates governance decision support information and establishes a dynamic model update mechanism.

[0084] In the preferred embodiment, step S5 specifically includes: S501: Based on the risk level distribution and risk source identification results, a multi-criteria decision analysis method is used to generate a governance priority ranking. The governance priority scoring function comprehensively considers risk level, risk source type, governance difficulty, and ecological sensitivity. ; in, As a risk index, Risk source type scores (organic matter accumulation area 0.8, vegetation degradation area 0.7, steep slope area 0.6). The difficulty coefficient for governance is (0-1). The ecological sensitivity coefficient is (0-1).

[0085] Based on the risk level distribution and risk source identification results, a multi-criteria decision analysis method was adopted to generate a governance priority ranking by comprehensively considering the risk index, risk source type, governance difficulty, and ecological sensitivity. The study area was divided into 23 governance units, and the governance priority of each unit was clarified.

[0086] Furthermore, a governance measure suggestion library is generated, matching corresponding governance technologies to different risk source types. These suggestions are overlaid with a risk distribution map to generate a governance decision support map, marking the location, area, recommended technologies, estimated costs, and governance cycle of key governance areas. A governance task list is generated through a GIS platform, supporting management departments in decomposing and assigning governance tasks.

[0087] S502: Construct a multi-objective optimization model for governance solutions to achieve Pareto optimal decisions in terms of ecological benefits, economic benefits, and governance cycle. The multi-objective optimization model is expressed as follows: ; in, For ecological benefits (total SOM reduction, tons). The cost of governance (in ten thousand yuan). The governance cycle is in months. Decision variables. For the combination of technologies selected for each governance unit.

[0088] This embodiment constructs a parameterized model library containing 15 technologies for managing drawdown zones (phytoremediation: Bermuda grass, vetiver, mulberry; engineering interception: ecological ditches, ecological bag slope protection; soil improvement: microbial agents, organic fertilizers, etc.). With the optimization objectives of maximizing total soil moisture loss (SOM) reduction, minimizing management costs, and minimizing management cycle, the NSGA-III multi-objective evolutionary algorithm is used to solve for the Pareto optimal solution set, generating three typical management schemes: 1) Cost-first approach: Total cost 2.85 million yuan, soil organic matter reduction of 86 tons, remediation period of 1.2 years; 2) Balanced type: Total cost 4.12 million yuan, soil organic matter reduction 124 tons, treatment cycle 1.5 years; 3) Results-oriented: Total cost of RMB 5.68 million, soil organic matter reduction of 156 tons, significantly reducing the risk of pollution release, and treatment cycle of 2.3 years.

[0089] Simultaneously, a governance scheme comparison dashboard is generated, which quantifies and scores the solutions from six dimensions: cost, benefits, cycle, technological maturity, and ecological security, providing management departments with flexible decision-making basis. The optimal solution is broken down into a specific governance task list, which clarifies the task location, governance technology, workload, budget, responsible person, and completion deadline, and supports integration with the government project management system.

[0090] S503: Establish a monitoring-feedback-update mechanism, repeat steps S1-S3 during the dry season of the following year, to achieve online updates and accuracy improvement of the model.

[0091] A dynamic model update and monitoring feedback mechanism was established. Steps S1-S3 of this invention were repeated during the drying season in 2024 to obtain a new round of soil organic matter inversion results and field sampling data. The results were compared with the data from 2023 to evaluate the treatment effect. At least 20 newly collected samples were added to the training sample set to retrain the random forest inversion model. At the same time, the organic matter accumulation-release model parameters were calibrated according to the actual water level scheduling and monitoring data to achieve online model updates and continuous improvement in accuracy. It is expected that the model inversion accuracy can be improved to R²>0.90, forming a closed-loop optimization system of monitoring, early warning, treatment, feedback and update. Table 3 shows the comparison of the effects of this invention and traditional methods.

[0092] Table 3. Comparison of the effects of the present invention and traditional methods

[0093] This embodiment demonstrates how the method of this invention can be applied to the drawdown zone of a tributary in the Three Gorges Reservoir area. It enables coordinated monitoring of vegetation stress and soil organic matter content in the drawdown zone, and provides early warning of pollution release risks based on water storage conditions. Specifically: 1. Under the conditions of this embodiment, the cycle for generating a monitoring and early warning report is 7 days, which shortens the monitoring cycle compared to traditional sampling methods.

[0094] 2. Under the conditions of this embodiment, the coefficient of determination R² of the soil organic matter inversion results reached 0.86, indicating that the method can meet the needs of operational application in the study area.

[0095] 3. Under different water storage scenarios, the present invention can output the corresponding release flux and risk level distribution results. Among them, the total release in the delayed water storage scenario is 30% lower than that in the normal water storage scenario, indicating that the method can provide a reference for water level scheduling and risk prevention and control.

[0096] 4. This invention can generate risk distribution maps, statistical results, and early warning reports, which helps to improve the spatiotemporal continuity of risk identification and the pertinence of management decisions.

[0097] 5. By supplementing samples and updating model parameters in subsequent monitoring cycles, the method of this invention has the ability to be continuously optimized and expanded in application.

[0098] Example 2 To further illustrate with reference to Example 1, a source pollution release risk early warning system based on multispectral inversion for implementing the method in Example 1 includes: The multispectral image acquisition and preprocessing module is used to control the UAV equipped with a multispectral camera to acquire multi-temporal images of the study area, and to perform radiometric calibration, atmospheric correction, geometric correction and drawdown mask extraction, and output standard reflectance images.

[0099] The vegetation stress index calculation module is used to calculate the vegetation index based on standard reflectance images, and construct the vegetation stress coefficient (VSC) by comparing it with historical data from the same period, and generate a vegetation stress level distribution map.

[0100] The soil organic matter inversion module is used to extract spectral features of sampling points, construct a collaborative inversion model based on random forest, invert soil organic matter content pixel by pixel, and generate a spatial distribution map; this module has a built-in active learning sampling optimization mechanism.

[0101] The accumulation-release simulation module is used to fit the organic matter accumulation rate based on multi-time back-evolution results. Combined with DEM data and water level scheduling plan, it constructs an accumulation-release coupled model to simulate the organic matter release flux and its spatiotemporal distribution under different water storage scenarios.

[0102] The risk warning module is used to construct a risk release index, classify risk warning levels, generate risk level distribution maps and warning reports, and push them to management departments through the WebGIS platform and mobile terminals.

[0103] The system management module is used for user permission management, data interfaces, log auditing, and system monitoring.

[0104] This system adopts a cloud-edge-device collaborative architecture: On-device (drone): Equipped with a multispectral camera and edge computing unit (NVIDIA Jetson Xavier NX), it enables image acquisition and preliminary preprocessing, and supports resume transmission after network outage.

[0105] Side-side (local server): Deploy M2, M3, and M4 core computing modules, configure GPU servers (NVIDIA TeslaT4×2, 64GB memory, 20TB storage), and support batch processing and model training.

[0106] Cloud-based (Government Cloud): Deploy M5, M6, and M7 management modules, configure a web server (Tomcat), a database server (PostgreSQL+PostGIS), and a GIS server (GeoServer), supporting concurrent access by multiple users and mobile push notifications.

[0107] After the system was deployed in this embodiment, the measured performance indicators are as follows: Table 2 Measured Performance Indicators of the System

[0108] This embodiment provides a method for early warning of non-point source pollution release risks in the drawdown zone of the Three Gorges Reservoir area. The working process, working details and technical effects are described in Embodiment 1, and will not be repeated here.

[0109] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A method for early warning of non-point source pollution release risks in the drawdown zone of the Three Gorges Reservoir area, characterized in that, Includes the following steps: S1: Acquire multi-temporal multispectral image data of the drawdown zone, preprocess the images to generate standard reflectance images of the study area; based on the standard reflectance images, extract vegetation stress characteristic indices and soil spectral characteristic bands. S2: Construct a synergistic inversion model of vegetation stress index and soil organic matter content. Input vegetation stress characteristic index and soil spectral characteristic band, and output spatial distribution map of soil organic matter content in the drawdown zone and distribution map of vegetation stress level. S3: Establish a coupled model of soil organic matter accumulation and release in the drawdown zone, and simulate the release flux and its spatiotemporal distribution of soil organic matter under different water storage scenarios by combining water level scheduling plans and topographic data. S4: Construct a risk index for non-point source pollution release in the drawdown zone, classify risk warning levels based on the simulation results of release flux, generate a risk warning map, and output a warning report; S5: Generate governance decision support information based on early warning results, and achieve closed-loop optimization of monitoring-early warning-governance through a model dynamic update mechanism.

2. The method for early warning of non-point source pollution release risk in the drawdown zone of the Three Gorges Reservoir area according to claim 1, characterized in that, Step S1 specifically includes: S101: Acquire multi-temporal multispectral images of the target drawdown zone area by UAV or satellite, wherein the multispectral images include at least blue light band, green light band, red light band, red edge band and near-infrared band; the multi-temporal images include at least three key time nodes: the early, middle and late stages of the drawdown zone drying period. S102: Radiometric calibration and atmospheric correction are performed on the acquired multispectral images to convert the original DN values ​​into surface reflectance; the digital elevation model is used to perform geometric fine correction on the images, and water bodies and buildings are removed by masking to generate a standard reflectance image that only contains the drawdown zone. S103: Based on standard reflectance images, normalized vegetation index NDVI, red-edge normalized vegetation index NDVIred-edge, leaf area index LAI, and vegetation moisture index NDWI were calculated as vegetation stress characteristic indices, and the vegetation stress coefficient VSC was calculated by combining them with historical normal values ​​of the same period. S104: Based on standard reflectance images, red light, near-infrared, and red edge band reflectance were extracted, along with soil modified vegetation index (SAVI), normalized differential moisture index (NDMI), red edge absorption depth, and red edge absorption area as characteristic bands of the soil spectrum.

3. The method for early warning of non-point source pollution release risk in the drawdown zone of the Three Gorges Reservoir area according to claim 1, characterized in that, Step S2 specifically includes: S201: Construct a training sample set for the co-inversion of vegetation stress index and soil organic matter content; set up typical sampling points in the study area, collect surface soil samples, and determine the soil organic matter content through laboratory chemical analysis; at the same time, extract the vegetation stress characteristic index and soil spectral characteristic bands of the corresponding sampling points to form a training sample set containing input features and output labels. S202: Construct a collaborative inversion model based on random forest or support vector machine; the model input features include vegetation stress characteristic index and soil spectral characteristic bands, and the output layer is soil organic matter content; S203: The trained co-inversion model is applied to the standard reflectance image of the entire study area, and the soil organic matter content is calculated pixel by pixel to generate a spatial distribution map of soil organic matter content in the drawdown zone; at the same time, the drawdown zone is divided into different stress levels according to the numerical range of the vegetation stress coefficient (VSC) to generate a vegetation stress level distribution map. S204: Verify the accuracy of the generated spatial distribution map of soil organic matter content and vegetation stress level distribution map; if the accuracy does not meet the requirements, adjust the model parameters or supplement the training samples, and retrain and invert the model. S205: Construct an active learning sampling optimization mechanism to achieve synergistic improvement in model accuracy and sampling efficiency, including: based on the calculation results of inversion uncertainty, identify high uncertainty regions for encrypted sampling, update the training sample set and retrain the model.

4. The method for early warning of non-point source pollution release risk in the drawdown zone of the Three Gorges Reservoir area according to claim 3, characterized in that, The random forest-based collaborative inversion model constructed in step S202 specifically includes: S2021: Perform Bootstrap sampling on the training sample set to generate K training subsets; S2022: For each training subset, construct a decision regression tree; when splitting at each node, randomly select m features from all input features, and select the optimal feature for splitting; S2023: Repeat steps S2021-S2022 to generate K decision trees, forming a random forest; for a new input sample, take the average of all decision tree predictions as the final prediction result. S2024: Calculate the prediction error and feature importance of the model using out-of-bag data, and sort and select key features based on feature importance.

5. The method for early warning of non-point source pollution release risk in the drawdown zone of the Three Gorges Reservoir area according to claim 3, characterized in that, The active learning sampling optimization mechanism in step S205 specifically includes: S2051: Based on the preliminary inversion results, the inversion uncertainty of each pixel is calculated. The sources of uncertainty include spatial sparsity of spectral features, terrain complexity, and vegetation heterogeneity. The formula for calculating the uncertainty index is: ； in, It is a comprehensive uncertainty index. The Mahalanobis distance between the pixel's spectral features and the nearest training sample. This represents the local topographic variation coefficient. To account for the local NDVI variance, all indicators were normalized to the range of 0-1. , , These represent the maximum value of Mahalanobis distance, the maximum value of local topographic variation coefficient, and the maximum value of local NDVI variance, respectively. S2052: Generate a sampling uncertainty distribution map, identify high uncertainty areas, plan denser sampling routes, and add sampling points; S2053: Merge the newly added sampling point data with the original training samples, retrain the model, and improve the inversion accuracy; S2054: Establish a historical sampling database, and use a reinforcement learning strategy to dynamically adjust the weight coefficients of the uncertainty index to optimize subsequent sampling decisions.

6. The method for early warning of non-point source pollution release risk in the drawdown zone of the Three Gorges Reservoir area according to claim 1, characterized in that, Step S3 specifically includes: S301: Obtain digital elevation model data and reservoir scheduling plan data for the study area; calculate the inundation range and water depth distribution of the drawdown zone under different water level conditions based on DEM data and water level elevation data; S302: Construct a model for soil organic matter accumulation in the drawdown zone to describe the temporal variation of soil organic matter content during the drying period; the model expression is: ； in, For the first dry period Soil organic matter content of the day This represents the initial soil organic matter content during the drying period. This is the organic matter accumulation rate coefficient; S303: Construct a soil organic matter release model in the drawdown zone to describe the release pattern of soil organic matter during flooding; the model expression is: ； in, For organic matter release flux, This refers to the soil organic matter content at the start of flooding. Water temperature This refers to the depth of floodwater. This refers to the duration of the flooding. S304: Substitute the spatial distribution map of soil organic matter content, inundation range and water depth distribution into the release model to simulate the release flux and its spatiotemporal distribution of organic matter under different water storage scenarios, and generate a spatial distribution map of release flux. S305: Construct a digital twin of non-point source pollution in the drawdown zone to achieve closed-loop dynamic evolution, including grid-based modeling based on the ST-GCN architecture, introducing the PINN physical information neural network mechanism, multi-scenario inference, and online learning and updating.

7. The method for early warning of non-point source pollution release risk in the drawdown zone of the Three Gorges Reservoir area according to claim 6, characterized in that, Step S305, which involves constructing a digital twin of area source pollution in the drawdown zone, specifically includes: S3051: Based on the inverted spatial distribution map of soil organic matter content, the accumulation-release model, and real-time access to water level and meteorological data, a digital twin is constructed using a spatiotemporal graph convolutional neural network architecture. The drawdown zone is divided into grid cells, with each cell serving as a graph node. Node features include SOM content, VSC value, slope, distance from the river channel, and historical release flux. Edge weights are calculated based on hydrological connectivity. S3052: Introducing a physical information neural network mechanism, the partial differential equation of organic matter accumulation-release is added as a physical constraint term to the loss function; S3053: Built-in multiple preset scenarios, supports managers to customize parameters for simulation, and outputs release flux change curves and risk level evolution diagrams; S3054: After acquiring new monitoring data each time, the deviation between the predicted value and the measured value is automatically calculated, and the model parameters are updated using an online learning algorithm.

8. The method for early warning of non-point source pollution release risk in the drawdown zone of the Three Gorges Reservoir area according to claim 1, characterized in that, Step S4 specifically includes: S401: Construct a risk index for non-point source pollution release in the drawdown zone, comprehensively considering multiple factors such as soil organic matter content, vegetation stress level, topographic slope, and distance from the river or reservoir bay; the risk index calculation formula is: ； in, This represents the normalized soil organic matter content. This represents the vegetation stress coefficient. This is the normalized terrain slope. This is the normalized distance from the river channel or reservoir bay; , , , The weight coefficients for each factor are determined using the analytic hierarchy process (AHP) or the entropy weight method. S402: Based on the risk index The numerical range is used to divide the drawdown zone into multiple risk levels, generating a distribution map of the risk levels of non-point source pollution release in the drawdown zone. S403: Generate an early warning report based on the risk level distribution map and the release flux simulation results; the early warning report shall include at least: the spatial distribution of high-risk areas and extremely high-risk areas, area statistics, the estimated total release of organic matter, the impact assessment on downstream water quality, and treatment recommendations for different risk levels; S404: Push the generated early warning report to the relevant management department through the GIS platform or mobile application.

9. The method for early warning of non-point source pollution release risk in the drawdown zone of the Three Gorges Reservoir area according to claim 1, characterized in that, Step S5 specifically includes: S501: Based on the risk level distribution and risk source identification results, a multi-criteria decision analysis method is used to generate a governance priority ranking; S502: Construct a multi-objective optimization model for governance solutions to achieve Pareto optimal decisions in terms of ecological benefits, economic benefits, and governance cycle; S503: Establish a monitoring-feedback-update mechanism, repeat steps S1-S3 during the dry season of the following year, to achieve online updates and accuracy improvement of the model.

10. A multispectral inversion-based area source pollution release risk early warning system for implementing the method of any one of claims 1-9, characterized in that, include: The multispectral image acquisition and preprocessing module is used to control the UAV equipped with a multispectral camera to acquire multi-temporal images of the study area, and to perform radiometric calibration, atmospheric correction, geometric correction and drawdown mask extraction, and output standard reflectance images. The vegetation stress index calculation module is used to calculate the vegetation index based on standard reflectance imagery, and construct the vegetation stress coefficient (VSC) by comparing it with historical data from the same period, and generate a vegetation stress level distribution map. The soil organic matter inversion module is used to extract spectral features of sampling points, construct a collaborative inversion model based on random forest, invert soil organic matter content pixel by pixel, and generate a spatial distribution map; this module has a built-in active learning sampling optimization mechanism. The accumulation-release simulation module is used to fit the organic matter accumulation rate based on multi-time back-evolution results, and to construct an accumulation-release coupled model by combining DEM data and water level scheduling plan to simulate the organic matter release flux and its spatiotemporal distribution under different water storage scenarios. The risk warning module is used to construct a risk release index, classify risk warning levels, generate risk level distribution maps and warning reports, and push them to management departments through the WebGIS platform and mobile terminals. The system management module is used for user permission management, data interfaces, log auditing, and system monitoring.

Images