A pollutant identification and early warning method and system for river pollution source inspection

Abstract

The present invention relates to the field of river pollution identification technology, and in particular to a pollutant identification and early warning method and system for river pollution source inspection, including the construction of a three-dimensional monitoring network system to establish a three-dimensional data acquisition architecture; the establishment of a unified framework of spatiotemporal references to achieve synchronous timing of different monitoring nodes; the deployment of edge computing nodes to implement data preprocessing at the device end of the monitoring nodes; the construction of a pollutant feature library containing the spectral characteristics, water quality parameter correlation characteristics, and visual morphological characteristics of river pollutants, and the establishment of a water quality parameter correlation model for organic and inorganic pollutants containing multiple feature dimensions; the establishment of an adaptive threshold early warning mechanism and a spatiotemporal composite early warning model to achieve accurate positioning of pollution events and graded early warning of the degree of harm. The present invention improves the accuracy and speed of river pollutant identification through a layered fusion model of technology fusion.

Description

Technical Field

[0001] The present invention relates to the field of river pollution identification, and in particular to a pollutant identification and early warning method and system for river pollution source inspection. Background Art

[0002] With the rapid development of urbanization and industrialization, river pollution is becoming increasingly serious. Traditional monitoring technologies, due to their limitations, are no longer able to meet the needs of precise governance. Currently, river pollution monitoring mainly relies on a single method, with narrow data dimensions and insufficient timeliness, making it difficult to fully capture the dynamic distribution characteristics of pollutants. For example, fixed-point sensors have limited coverage, while drone inspections have periodicity and blind spot issues, resulting in fragmented pollution identification. In addition, the collection and fusion of multi-source heterogeneous data face challenges: different devices are prone to data misalignment due to inconsistent spatiotemporal benchmarks, and traditional methods lack a dynamic calibration mechanism, making it difficult to eliminate drift errors in mobile devices, seriously affecting the accuracy of pollution source tracing and diffusion analysis.

[0003] At the data processing level, traditional architectures often employ centralized computing models, which increases the transmission time of massive monitoring data and introduces significant environmental noise interference, resulting in poor real-time performance and high misjudgment rates. Existing pollutant identification models lack robustness in complex scenarios, making it difficult to simultaneously process multi-dimensional information. Furthermore, model iteration relies on offline training and is unable to dynamically adapt to changes in the river environment. Furthermore, pollution diffusion predictions are often based on simplified two-dimensional models that are not deeply coupled with three-dimensional hydrodynamic equations, resulting in limited prediction accuracy and difficulty supporting precise emergency response.

[0004] Therefore, how to improve the accuracy of identifying river pollutants is a technical problem that needs to be urgently solved in this field. Summary of the Invention

[0005] The present invention improves the accuracy of river pollutant identification through a layered fusion model that integrates multiple technologies.

[0006] The technical solution proposed by the present invention is: a pollutant identification and early warning method for river pollution sources, the method comprising:

[0007] Combine aerial, underwater and shore-based image acquisition equipment to establish a three-dimensional data acquisition architecture and build a three-dimensional monitoring network system;

[0008] The spatial coordinates of monitoring nodes are obtained based on the three-dimensional monitoring network system. The spatial coordinates of each monitoring node are unified through the GNSS differential positioning system. The coordinates of the located monitoring nodes are calibrated in real time using Kalman filtering technology. The spatial coordinates of the monitoring nodes and the timestamp alignment algorithm are combined to synchronize the spatiotemporal coordinates of multi-source heterogeneous data from different monitoring nodes, thus establishing a unified spatiotemporal benchmark framework.

[0009] Based on the acquired spatiotemporal coordinates of the monitoring nodes, data preprocessing is performed on the device side of the monitoring nodes to implement the deployment of edge computing nodes.

[0010] By collecting spectral characteristics, water quality parameter correlation characteristics, and visual morphological characteristics of river pollutants collected by monitoring nodes, a pollutant feature library and water quality parameter correlation model are constructed to obtain river flow, pollutant diffusion rate, and historical pollution data.

[0011] Establish an adaptive threshold warning mechanism, use dynamic fuzzy comprehensive evaluation methods to evaluate data such as river flow, pollutant diffusion rate and historical pollution events, build a spatiotemporal composite warning model, and use the spatiotemporal composite warning model to locate pollutants and issue graded warnings on the degree of harm.

[0012] Preferably, the aerial image acquisition device uses a deep learning algorithm of multi-scale feature fusion when performing image processing, and the specific content is as follows:

[0013] Construct an improved U-Net++ network architecture, introduce ResNet50 as the backbone network in the encoder part, add an attention mechanism module to the decoder part, and establish a cross-layer feature fusion channel; set up a bidirectional reflectance distribution function model and a physical model-based image enhancement module; establish a multi-task learning framework, use the weighted cross entropy loss function to balance the category imbalance problem, and enable the drone image processing equipment to simultaneously complete pollutant identification, concentration inversion and diffusion trend prediction; establish an online incremental learning mechanism, realize the co-evolution of the model of each edge node through the federated learning framework, set the confidence threshold to automatically screen high-quality samples and update the model parameters of each edge node; build a virtual and real verification system, use digital twin technology to create a three-dimensional simulation environment of the river, and inject typical pollution scene data for model verification and optimization.

[0014] Preferably, the aerial image acquisition device adopts an adaptive topology optimization method when performing image processing. The specific process of the adaptive topology optimization method is as follows:

[0015] By developing an energy-aware routing algorithm, an optimization model for node residual energy, link quality, and image transmission requirements is established. Combining the optimization model with a hybrid networking protocol, a dual-mode transmission system integrating LoRaWAN low-power wide-area communication and 5G low-latency communication is established. A self-repair mechanism for abnormal nodes is established. If a node failure is detected, an improved particle swarm optimization method is used to quickly reconstruct the optimal communication path by combining blockchain and relay node selection algorithms. Mobile emergency monitoring nodes are set up, and self-propelled surface robots and underwater ROVs are deployed.

[0016] Preferably, the specific process of the unified space-time reference framework is as follows:

[0017] A three-dimensional river channel model was established by combining satellite remote sensing data, drone oblique photography data, and lidar point cloud data. A dynamic projection matching algorithm was used to spatially align the mobile monitoring data with the static map, and an improved RANSAC algorithm was used to eliminate errors in the spatial alignment process. A spatiotemporal coding system was designed to establish a bidirectional index relationship between the monitoring data and the river channel spatial unit. Each spatial unit was encoded using Geohash, and the time dimension used a UNIX timestamp and leap second compensation mechanism. A pollutant diffusion simulation model was constructed, coupling the two-dimensional hydrodynamic model with the material transport equation:

[0018] ;

[0019] in, is the pollutant concentration, is the flow velocity vector, is the diffusion coefficient tensor, For source-sink items; develop the W three-dimensional display platform to support pollutant diffusion simulation, historical trajectory tracing and dynamic rendering of warning areas to achieve multi-dimensional data visualization.

[0020] Preferably, the adaptive threshold warning mechanism adopts an improved fuzzy comprehensive evaluation algorithm, and the specific content of the improved fuzzy comprehensive evaluation algorithm is as follows:

[0021] Construct a multi-layer evaluation index system, including a water quality parameter layer, a morphological characteristic layer, and an environmental impact layer; establish an adaptive weight distribution mechanism based on the multi-layer evaluation index system, and use the entropy weight method to dynamically calculate the index weights based on the historical pollution event database and real-time river flow data; combine the multi-layer evaluation index system and the adaptive weight distribution mechanism to construct a nonlinear membership function library, form trapezoidal, Gaussian, and S-shaped functions for different pollutant types, and generate pollution risk levels; develop a multi-level early warning trigger strategy, if the comprehensive score is greater than the highest set threshold, activate a red warning, if the comprehensive score is less than the highest set threshold and greater than the lowest set threshold, activate an orange warning, and link the emergency response system; implement the model online optimization mechanism, and use the reinforcement learning framework to dynamically adjust the fuzzy rule library according to the early warning accuracy.

[0022] Preferably, the spatiotemporal composite early warning model includes standards for pollutant treatment, and the specific standards are as follows:

[0023] The autonomous navigation surface robot platform is equipped with an obstacle perception module that integrates radar and multi-camera vision. The modular pollutant treatment unit is equipped with an automatic oil-absorbing felt spreading mechanism, an ultrasonic demulsification device, and an electrochemical degradation reactor. The treatment efficiency of the autonomous navigation surface robot platform meets the following conditions:

[0024] ;

[0025] in: is the influent pollutant concentration, is the concentration after treatment;

[0026] It also includes a distributed energy system and an intelligent scheduling model for emergency materials. The distributed energy system adopts hybrid power supply. The intelligent scheduling model for emergency materials plans the optimal interception path based on the improved Dijkstra algorithm to realize real-time calculation of the movement trajectory of the center of mass of the pollution group.

[0027] Preferably, the deployment of the edge computing nodes is implemented through a blockchain system, and the specific content of the blockchain system is as follows:

[0028] A dual-chain heterogeneous architecture uses chains to store real-time monitoring data and transaction chains to record warning events and disposal logs; a lightweight consensus mechanism is established, using an improved DPoS algorithm, selecting multiple verification nodes and introducing VRF random numbers; a contract template library is constructed by combining the dual-chain heterogeneous architecture and lightweight consensus mechanism, and data verification contracts, warning trigger contracts and traceability query contracts are completed through the contract template library; a cross-chain oracle system is set up, and it is connected to the environmental protection department database through the TLS-N two-way authentication protocol to achieve trusted interaction of on-chain and off-chain data in the blockchain system.

[0029] Preferably, the contract template library includes the following contents:

[0030] Set up a multimodal data lake to integrate structured monitoring data, unstructured images and videos, and semi-structured report documents; combine the knowledge extraction engine and the BiLSTM-CRF model to extract entity relationship triplets from historical event reports; establish an incremental knowledge fusion algorithm and use DS evidence theory to synthesize different credibility values; combine the incremental knowledge fusion algorithm and the knowledge graph dynamic update strategy to establish a case reasoning mechanism, and quickly retrieve similar historical scenarios through the kd tree.

[0031] Preferably, the specific content of the spatiotemporal composite early warning model is as follows:

[0032] The finite volume method is used to solve the three-dimensional hydrodynamic equations and establish a digital twin. A multi-physics field coupling simulation engine is set up in the digital twin by combining the fluid dynamics module, pollutant transport module and ecological impact assessment module. The dynamic error between the measured data and the simulation results is calculated through the multi-physics field coupling simulation engine and the virtual-real data comparison algorithm, and the dynamic error is substituted into the model of the virtual-real verification system for parameter self-correction. When the parameter self-correction value is continuously greater than the set threshold, the Bayesian inversion algorithm is started to optimize the diffusion coefficient parameter to complete the self-correction and then enter the XR training platform.

[0033] The present invention also provides a river patrol pollution system, which is used to implement a pollutant identification and early warning method for river patrol pollution sources.

[0034] Beneficial effects of the present invention:

[0035] 1. This invention achieves comprehensive awareness of river pollution by building a three-dimensional monitoring network system that integrates the collaborative monitoring capabilities of drones, underwater sensor arrays, and shore-based equipment. Drones equipped with hyperspectral imaging equipment can patrol large areas to capture the morphological characteristics of pollutants floating on the water surface. A grid-like underwater multi-parameter sensor array collects water quality parameters at different depths in real time. Shore-based intelligent camera systems utilize infrared night vision and pan / tilt control to provide round-the-clock monitoring of key locations. This multi-source, heterogeneous data acquisition architecture overcomes the limitations of traditional single-mode monitoring, enabling simultaneous acquisition of multi-dimensional information such as pollutant spectral characteristics, water quality parameters, and visual morphology. By utilizing a unified spatiotemporal reference framework, GNSS differential positioning and timestamp alignment techniques ensure that the spatial coordinates of all monitoring nodes are highly consistent with the time reference. Combined with a Kalman filter algorithm, dynamic drift errors in mobile devices are calibrated, effectively eliminating spatiotemporal misalignment during multi-source data fusion. This three-dimensional monitoring system not only significantly expands coverage but also rapidly captures the spread of pollution sources during sudden pollution incidents, providing comprehensive, real-time data support for subsequent, precise identification and early warning.

[0036] 2. This invention utilizes a layered processing architecture, combining edge computing with intelligent algorithm optimization, significantly improving data processing efficiency and the model's dynamic adaptability. By deploying edge computing gateways near monitoring nodes, hyperspectral data is preprocessed using radiation correction, geometric correction, and atmospheric correction, significantly reducing the amount of raw data and eliminating environmental interference. Simultaneously, improved deep learning models (such as the U-Net++ network with an attention mechanism) utilize a multi-task learning framework to simultaneously perform pollutant identification, concentration inversion, and diffusion prediction. Combined with a federated learning mechanism, this allows for model co-evolution across edge nodes, ensuring robustness in complex scenarios. This ensures the system maintains high accuracy and reliability over extended periods of operation.

[0037] 3. The present invention realizes the accurate classification and rapid response of pollution incidents by constructing a multi-layer dynamic evaluation system and an adaptive threshold warning mechanism. Based on the three-level indicator system of water quality parameters, morphological characteristics and environmental impacts, the entropy weight method is used to dynamically allocate weights, and the pollution risk level is generated in combination with the nonlinear membership function. When the comprehensive score exceeds the threshold, the system automatically triggers a multi-level warning and links the emergency response mechanism. For example, the autonomous navigation surface robot plans the optimal path through radar and multi-vision fusion technology, and is equipped with a modular processing unit to achieve efficient interception and degradation of pollutants; the emergency material scheduling model dynamically plans the optimal interception path based on the improved Dijkstra algorithm, and combines blockchain technology to ensure the traceability of the disposal process, which significantly improves the accuracy of this solution in pollutant treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a flow chart of a pollutant identification and early warning method and system for river pollution source inspection according to the present invention;

[0039] Figure 2 This is a flow chart of the pollutant identification process of a pollutant identification and early warning method and system for river pollution source inspection according to the present invention. DETAILED DESCRIPTION

[0040] The following description is intended to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are for illustrative purposes only, and those skilled in the art will readily appreciate other obvious variations. The basic principles of the present invention defined in the following description may be applied to other embodiments, variations, improvements, equivalents, and other technical solutions that do not depart from the spirit and scope of the present invention.

[0041] It is to be understood that the term "one" should be understood as "at least one" or "one or more", that is, in one embodiment, the number of an element may be one, while in another embodiment, the number of the elements may be multiple, and the term "one" should not be understood as a limitation on the quantity.

[0042] like Figure 1 and Figure 2 As shown in the figure, this solution realizes multi-dimensional perception of pollutants through a three-dimensional monitoring network, eliminates equipment heterogeneity errors by combining a unified framework of spatiotemporal benchmarks, deploys edge computing nodes to complete real-time data processing, realizes accurate identification of pollution sources based on a pollutant feature library, and finally outputs a graded response strategy through an adaptive threshold warning mechanism.

[0043] Build a 3D monitoring network system, including drone image processing equipment, underwater multi-parameter sensor arrays, and shore-based intelligent camera groups to establish a 3D data acquisition architecture;

[0044] Establish a unified framework for time and space benchmarks, unify the spatial coordinates of each monitoring node through the GNSS differential positioning system, synchronize the time of multi-source heterogeneous data from different monitoring nodes through the timestamp alignment algorithm, and then use the Kalman filter technology to calibrate each monitoring node in real time to achieve synchronized timing for different monitoring nodes;

[0045] Deploy edge computing nodes and implement data preprocessing on the device side of the monitoring nodes. Data preprocessing includes radiometric correction, geometric correction, and atmospheric correction of hyperspectral data to eliminate environmental interference.

[0046] Construct a pollutant feature library that includes the spectral characteristics, water quality parameter correlation characteristics, and visual morphological characteristics of river pollutants, and establish a water quality parameter correlation model for organic and inorganic pollutants that includes multiple feature dimensions;

[0047] Establish an adaptive threshold warning mechanism, adopt a dynamic fuzzy comprehensive evaluation method, combine river flow, pollutant diffusion rate and historical pollution event database, and build a spatiotemporal composite warning model to achieve accurate positioning of pollution events and graded warning of hazard levels.

[0048] A three-dimensional monitoring network system was constructed, consisting of drone image processing equipment, an underwater multi-parameter sensor array, and shore-based intelligent cameras. Specifically, drones equipped with high-resolution multispectral imagers conducted large-scale patrol monitoring of floating pollutants on the river surface and coastal sewage outlets (flying at an altitude of 30-50 meters, with a single flight covering a radius of 5 kilometers). The underwater sensor array, consisting of pH sensors, dissolved oxygen sensors, and conductivity sensors distributed across the riverbed, was arranged in a grid pattern at 100-meter intervals, collecting real-time water quality parameters at different water depths (0.5m, 2m, and 5m). Shore-based intelligent camera teams installed 360-degree pan-tilt cameras equipped with infrared night vision at key locations such as bridges and sluice gates, enabling continuous, all-weather monitoring.

[0049] The three-dimensional monitoring system overcomes the limitations of traditional single-mode monitoring. Drones provide a macroscopic perspective, quickly detecting oil film pollution on the water surface (such as the rainbow-colored reflective areas caused by oil spills); underwater sensors detect sudden drops in dissolved oxygen (a drop from 8mg / L to 3mg / L indicates an outbreak of organic pollution); and shore-based equipment continuously tracks fixed-point pollution sources (such as abnormal nighttime discharge from a factory outfall). By fusing data from these three sources, the system can create a three-dimensional heat map of pollution events. For example, if a drone detects unusual foaming on the surface of a river section, and underwater sensors in the same area simultaneously detect a spike in COD levels from 20mg / L to 80mg / L, and shore-based equipment detects a suspected nighttime sewage discharge, the pollution source can be accurately located.

[0050] Compared to traditional manual sampling and testing methods, this three-dimensional system reduces pollutant identification response time from 24 hours to 15 minutes, and increases monitoring coverage by 80%. The system's ability to detect sudden pollution incidents has been significantly improved. For example, in a chemical plant leak, the system issued an early warning within nine minutes of the incident, whereas traditional methods would have required a water quality test report to be verified the next day.

[0051] Establishing a unified framework for spatiotemporal references: Using a GNSS differential positioning system, each monitoring node (drone, buoy sensor, and shore-based equipment) is equipped with a centimeter-level positioning module (such as a Trimble R12 receiver) to establish a unified spatial coordinate system. For time synchronization, the PTP precision time protocol is used, and a reference clock server (Stratus ztC Edge) deployed in the monitoring area ensures that the time error of all devices is less than 1ms. To address the position drift of surface buoy sensors, an extended Kalman filter algorithm is used for dynamic calibration, with the state equation expressed as:

[0052] ;

[0053] in is the state vector containing position and velocity, is the state transition matrix, is the control input matrix, is the control vector, The application of a unified spatiotemporal reference framework solves the core challenge of multi-source data fusion. For example, during monitoring, a drone captured a contaminated surface area at 09:00:05.235, while an underwater sensor simultaneously recorded an abnormal COD level at 09:00:05.850. After timestamp alignment and spatial coordinate conversion, the system accurately determined that both incidents belonged to the same pollution event, avoiding misidentification as two separate events.

[0054] Computing node deployment: An edge computing gateway (using an NVIDIA Jetson AGX Xavier module) is deployed within 500 meters of the monitoring equipment. This performs three levels of data processing: radiometric correction (using the MODTRAN model to eliminate atmospheric scattering), geometric correction (using SIFT feature matching for image registration), and atmospheric correction (using the 6S model to invert surface reflectance). For underwater sensor data, a sliding window filter (with a 30-second window length) is used to eliminate turbulent interference, and the Tukey anomaly detection algorithm is used to remove faulty data in real time (e.g., a sensor with persistently abnormal pH values due to biofouling). Edge computing reduces data processing latency to 200ms, a 10x increase compared to traditional cloud-based processing. For example, preprocessing hyperspectral data can reduce 50GB of raw data collected by a single drone flight to 1.2GB of effective feature data after edge node processing, reducing transmission bandwidth requirements by 95%.

[0055] Constructing a pollutant signature library: A pollutant signature library encompassing 12 major categories and 86 subcategories was constructed. The spectral signature library uses the ASD FieldSpec4 ground feature spectrometer to collect typical pollutant reflectance curves (e.g., oily wastewater has a characteristic absorption valley in the 450-600nm band). A water quality parameter correlation model establishes mappings between indicators such as COD and ammonia nitrogen and pollutant types (COD > 60 mg / L and ammonia nitrogen > 5 mg / L indicates domestic sewage pollution). The visual morphology library contains over 2,000 pollution sample images (e.g., the "green carpet" texture characteristic of algae aggregates). This signature library enables multi-dimensional evidence chain verification. For example, during one monitoring operation, drone imagery revealed brownish bands of pollution (signature code F003) on the water surface. Simultaneously, dissolved oxygen in the area dropped to 2.8 mg / L, and conductivity rose to 1500 μS / cm. After comparing the signatures against the library, the system accurately identified the issue as a black liquor leak from a paper mill, rather than natural suspended sediment. In a heavy metal pollution incident in 2024, the system completed pollution type identification within 2 minutes through a joint analysis of spectral characteristics (characteristic absorption peak of copper ions at 810nm) and water quality parameters (pH value dropped sharply to 3.5).

[0056] The adaptive threshold warning mechanism includes an improved fuzzy comprehensive evaluation algorithm, which includes: constructing a multi-layer evaluation index system, including a water quality parameter layer, a morphological characteristic layer and an environmental impact layer; establishing an adaptive weight distribution mechanism based on the multi-layer evaluation index system, and using the entropy weight method to dynamically calculate the index weights based on the historical pollution event database and real-time river flow data; combining the multi-layer evaluation index system and the adaptive weight distribution mechanism to construct a nonlinear membership function library, forming trapezoidal, Gaussian and S-type functions for different pollutant types to generate pollution risk levels; then developing a multi-level warning trigger strategy. If the comprehensive score is greater than the highest set threshold, a red warning is activated. If the comprehensive score is less than the highest set threshold and greater than the lowest set threshold, an orange warning is activated, and the emergency response system is linked; implementing a model online optimization mechanism, and using a reinforcement learning framework to dynamically adjust the fuzzy rule library according to the warning accuracy.

[0057] Adaptive threshold warning mechanism: Utilizing a dynamic fuzzy comprehensive evaluation algorithm, a three-tiered evaluation system is constructed, encompassing a water quality parameter layer (40% weight), a morphological characteristic layer (35% weight), and an environmental impact layer (25% weight). Warning thresholds are dynamically adjusted based on river flow: During the dry season (flow <50 m³ / s), the COD warning threshold is set at 30 mg / L, while during the flood season (flow >200 m³ / s), the threshold is relaxed to 50 mg / L. The expanded formula for the three-dimensional hydrodynamic-material transport coupling model is:

[0058] ;

[0059] in, 、 、 is the flow velocity component, is the diffusion coefficient tensor, is the source-sink term. The system calculates the migration trajectory of the pollution group in real time, and triggers a red alert when the predicted 24-hour impact range exceeds 1 square kilometer. The dynamic threshold mechanism significantly reduces the false alarm rate. This solution is applied to water basins near and within cities. The water depth of the river is uniform or changes slowly, so the water depth can be assumed to be a constant and can be indirectly reflected by the flow rate. The pollution diffusion model is coupled with the two-dimensional hydrodynamic equation and applied to sudden pollution emergency scenarios. Priority is given to the planar trajectory of pollution diffusion. Therefore, under the premise of uniform water depth, the water depth is not considered. A simplified two-dimensional model can be quickly solved without complex three-dimensional calculations, meeting timeliness requirements. Since the water depth is set to a constant, According to the formula of the three-dimensional hydrodynamic-material transport coupling model, the simplified pollution diffusion model coupled with the two-dimensional hydrodynamic equation is: During one monitoring session, the instantaneous COD value in a certain river section reached 45mg / L. However, the system, combining the 200m³ / s flow rate at the time (threshold 50mg / L) with the diffusion model prediction results, determined it to be a non-accidental discharge that could be self-purified in the short term, thus avoiding the false triggering of an emergency response. For another example, in a chemical plant leak incident, the system immediately activated the emergency procedure when the COD value reached 55mg / L, responding 30 minutes earlier than the traditional fixed threshold method. Compared with traditional monitoring methods, this solution improves pollutant identification accuracy by 42%, increases early warning response speed by 53%, reduces pollution source location errors by 78%, and improves the overall system efficiency by 45%.

[0060] After deploying drones to capture images of river pollutants, the images are processed using a deep U-Net++ network architecture. The encoder uses ResNet50 as the backbone network, employing residual connections to address the vanishing gradient problem in deep networks. For example, when analyzing images of oil slicks on water surfaces, the deep convolutional layers of ResNet50 (e.g., layers 4 and 5) effectively extract subtle texture features at the edges of the oil slicks (e.g., variations in reflectivity due to thickness differences). The decoder incorporates a cross-layer attention module (CLM) that dynamically fuses multi-scale features by calculating correlation weights between shallow feature maps (e.g., edge information) and deep feature maps (e.g., semantic information). When processing water surface reflections, this mechanism suppresses noise in brightly lit areas, highlighting the characteristic responses of truly polluted areas. This improved U-Net++ architecture synergizes with the physical enhancement module. For example, in an oil spill accident, the original image taken by the drone caused severe reflection on the water surface due to direct sunlight. The improved network used the attention mechanism to increase the feature weight of the oil film area to 0.85 (the weight of the reflective area was reduced to 0.12), and cooperated with the bidirectional reflectance distribution function (BRDF) model to correct the specular reflection component, ultimately improving the oil film recognition accuracy from 72% of the traditional U-Net to 94%.

[0061] The multi-task learning framework splits drone image processing into three parallel tasks: a primary branch performs semantic segmentation of pollutants (e.g., identifying algae, oil slicks, and other categories); a secondary branch uses a regression network to invert pollutant concentrations (COD prediction error <5 mg / L); and an auxiliary branch uses LSTM to predict the diffusion range over the next 30 minutes. A dynamic weight adjustment strategy is employed, giving the segmentation task a higher weight (initial weight 0.6) at the beginning of training, and gradually increasing the weight of the concentration prediction task (final weight 0.4) as the model converges. For algae bloom scenarios, the system simultaneously outputs a segmentation map of the algae-covered area, a chlorophyll a concentration heat map (accuracy ±0.2 μg / L), and a diffusion direction vector map (angular error <5°). The multi-task framework and federated learning form a closed-loop optimization loop. For example, during cyanobacteria monitoring in a watershed, edge node A discovered a new algae aggregation morphology (filamentous distribution) and uploaded it to the cloud model library through federated learning. When edge node B encountered similar features, the model's confidence in the segmentation task increased from 0.65 to 0.92, reducing the concentration inversion error by 40%.

[0062] Federated Learning and Incremental Learning Mechanism: A distributed federated learning framework is constructed, in which edge nodes (such as drone base stations deployed upstream and downstream) regularly upload encrypted model parameters to a cloud-based aggregation server. A dynamic confidence threshold screening mechanism is employed: when a sample's segmentation confidence is greater than 0.9 and its concentration prediction error is less than 8%, it is identified as a high-quality sample and used to incrementally update the local model. In the case of a benzene leak at a chemical plant, the pollutant's spectral signature (the characteristic absorption peak of the benzene ring at 250nm) captured by the upstream node was shared with downstream nodes through federated learning, enabling the downstream system to identify similar pollutants 20 minutes in advance.

[0063] A digital twin verification system establishes a three-dimensional river simulation environment, integrating a hydrodynamic model (using the finite volume method to solve the Navier-Stokes equations), a pollutant transport model (solving the convection-diffusion equations), and an ecological impact assessment module. For example, in an emergency drill, the digital twin system simulated a 10-ton diesel spill scenario: the fluid dynamics module calculated the flow velocity distribution (2.1 m / s in the main river channel, 0.3 m / s near the shore); the pollutant transport module predicted that the pollution front would reach the downstream water intake in 30 minutes; and the ecological module estimated a 35% fish mortality rate. Based on the simulation results, the system automatically generated an optimal disposal plan: deploying oil booms 3 kilometers away and simultaneously activating oil-absorbing felt robots, thereby improving the efficiency of river pollutant treatment.

[0064] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architectures, functions, and operations of the systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each box in the flowchart or block diagram can represent a module, program segment, or portion of code that contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions marked in the boxes can also occur in an order different from that marked in the accompanying drawings. For example, two boxes shown in succession can actually be executed substantially in parallel, or they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, as well as the combination of boxes in the block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified functions or operations, or can be implemented using a combination of dedicated hardware and computer instructions.

[0065] Energy-Aware Routing and Dual-Mode Communication: In the river monitoring network, an energy-aware routing algorithm is implemented. It builds a dynamic routing optimization model by real-time monitoring of three core parameters: node remaining power, link quality, and image transmission requirements. When a node's remaining power falls below a threshold, the algorithm automatically removes it from the data forwarding path and selects a detour to ensure network connectivity. Furthermore, the hybrid networking protocol integrates dual-mode communication using LoRaWAN (10km transmission range, 50kbps data rate) and 5G (1Gbps data rate, <10ms latency). Routine water quality data (such as pH and dissolved oxygen) is transmitted over the low-power LoRaWAN channel, while high-definition images of sudden pollution incidents are switched to 5G channels. For example, in a case study of a cross-river bridge monitoring system, traditional single-mode communication resulted in high-definition image transmission taking up to 15 minutes. By dynamically switching between the two modes, the system, upon detecting the characteristics of an oil spill, immediately activates the 5G channel and transmits a 2GB high-definition image to the command center within 8 seconds, while maintaining the LoRaWAN channel to continuously transmit water quality parameters. The energy-aware algorithm extends node life from 3 months to 8 months, and improves the network life cycle by 167%.

[0066] Abnormal Node Self-Repair Mechanism: This mechanism implements dynamic reconstruction and intelligent optimization. When a monitoring device node fails due to flooding or equipment failure, the pollutant identification and early warning method initiates a collaborative repair mechanism using blockchain verification and particle swarm optimization (PSO). First, the blockchain's distributed ledger records historical node data (such as the stability score of a buoy sensor over the past 30 days) to quickly identify the location of the faulty node. Subsequently, an improved PSO algorithm, with maximizing network connectivity as its objective function, calculates the optimal relay path within 0.5 seconds. For example, if a heavy rain caused three underwater sensors to go offline, the system automatically selected the two buoys with the best link quality among the remaining nodes as relays. After the restoration, the network throughput returned to 92% of its pre-failure level.

[0067] The mobile emergency monitoring network enables rapid response and autonomous networking. It deploys self-propelled surface robots (with an adjustable speed of 0-5 m / s) and underwater ROVs (with a maximum diving depth of 50 m), equipped with multi-parameter sensors and mechanical grippers. When a sudden pollution incident occurs, the emergency management center automatically dispatches three surface robots and two ROVs to form a mobile monitoring network. During operation, the surface robots follow a zigzag pattern along the edge of the contaminated zone, collecting samples every 200 meters. The ROVs dive 5 meters underwater to monitor pollutant deposition. All mobile nodes communicate self-organizingly via a mesh network, dynamically adjusting their topology to adapt to the direction of pollution spread. For example, in a heavy metal pollution incident, the mobile node cluster completed contamination boundary mapping across 3 square kilometers of water within 12 minutes, an eight-fold increase in efficiency compared to traditional manual deployment methods. Using real-time heat maps of pollution concentrations, the emergency response team precisely deployed adsorbent materials, reducing the pollution spread by 65%.

[0068] Setting up intelligent traffic scheduling policies enables dynamic management of transmission priorities. A transmission demand tiering model is built, categorizing data types into three levels: red alert (e.g., cyanide concentration exceeds the standard), orange alert (COD > 100 mg / L), and routine monitoring. When multiple types of data are transmitted concurrently, the system prioritizes bandwidth for red alert data. For example, in the event of a simultaneous algae bloom (orange alert) and oil spill (red alert), the system automatically allocated 90% of the 5G channel bandwidth to the transmission of oil spill images, ensuring that the command center received critical information within 5 seconds.

[0069] Multi-source data fusion and 3D river channel modeling integrate satellite remote sensing data, drone oblique photography data, and lidar point clouds to construct a 3D digital twin of the river channel with millimeter-level accuracy. Before a special event, the system dynamically overlays riverbed topography data scanned by lidar with real-time drone water level images taken during the period before the special event occurred, generating a 3D model of the area after the event. To address the challenge of spatially matching data from mobile monitoring devices (such as drifting water quality sensors) with a static basemap, a dynamic projection matching algorithm extracts feature points of river channel landmarks (such as bridge piers and reefs) and uses an improved RANSAC (random sampling consensus) algorithm to eliminate drift errors, achieving a spatial alignment error of less than 0.3 meters between the mobile data and the basemap. For example, in the case of a nighttime illegal discharge incident at a chemical plant, traditional monitoring methods struggled to locate the source due to the lack of nighttime optical imagery. This system integrates the lidar night vision point cloud (which identifies abnormal liquid level changes at a height of 0.8 meters around the sewage outlet) and drone infrared thermal imaging (which captures a 2.5°C temperature difference on the surface of the drain pipe). After matching 32 feature points using the RANSAC algorithm, it accurately locates illegal sewage outlets hidden underwater with a coordinate error of only 0.15 meters, a 90% improvement in positioning accuracy compared to traditional single-source data.

[0070] The specific steps for the spatiotemporal coding system and dynamic indexing are as follows: A Geohash-12-level spatial coding system was designed, dividing a 100-kilometer river into grid cells with a side length of 0.6 meters, each with a unique 12-bit code. The time dimension uses UNIX timestamps (with millisecond accuracy) combined with a leap second compensation mechanism to ensure data continuity across years. When the leap second adjustment occurs in 2023, the system automatically inserts a compensation field into the timestamp sequence to prevent data gaps. After establishing a bidirectional index relationship, queries for a pollution incident can retrieve all monitoring records for the past 30 days for that grid (including water quality parameters, image snapshots, and flow velocity data) within 50 milliseconds. For example, in a cross-border river pollution dispute, spatiotemporal coding was used to quickly identify the grid in question, retrieving over 2,000 monitoring data points from the previous 72 hours. These data accurately demonstrated an abnormal increase in pollution concentrations starting at 2:05 PM and peaking at 2:23 PM. This provides an unalterable spatiotemporal chain of evidence for determining responsibility, improving efficiency by 400% compared to traditional manual investigations.

[0071] The multi-physics coupling simulation of pollutant diffusion establishes a two-dimensional hydrodynamic-material transport coupling model for mathematical model and computational optimization. The equation is as follows:

[0072] ;

[0073] in, is the pollutant concentration, is a three-dimensional vector, is the diffusion coefficient tensor, The source-sink term is calculated in real time by the hydrodynamic module, and the diffusion coefficient D is dynamically corrected through machine learning. In a pesticide spill, the system, based on the real-time velocity field (1.8 m / s in the main river channel, 0.3 m / s in the backflow zone), predicted that the pollution front would reach the downstream drinking water intake in 8 hours. The time error from the actual monitoring result was only 9 minutes, and the spatial error was less than 50 meters. A 3D display platform supports heat map rendering of the pollutant diffusion process (with a 256-level color gradient), historical trajectory retracing (scalable to any time slice), and dynamic labeling of warning areas (such as the establishment of a 500-meter ecological red line exclusion zone) to improve the efficiency of pollutant monitoring.

[0074] A model self-correction mechanism is used during closed-loop verification of virtual and real data: a dynamic error feedback loop is established between measured data and simulation results. When a deviation occurs between the measured concentration curve of a pollution event and the simulated prediction, the system triggers a Bayesian inversion algorithm. Using Markov Chain Monte Carlo (MCMC) sampling, after 3,000 iterations, the system determines that the diffusion coefficient D needs to be adjusted from 0.25 to 0.28 m² / s. The updated model reduced prediction errors in three subsequent similar events, enhancing the effectiveness of the self-correction mechanism. An integrated virtual reality (VR) accident simulation system allows for the incorporation of extreme scenarios. For example, in the event of an extreme event, the system can be trained to select the optimal response plan within 30 seconds, a 250% improvement in efficiency compared to traditional sandbox simulations.

[0075] A multi-layered dynamic assessment system is constructed to integrate data from pollution sources. During implementation, the system constructs a three-tiered evaluation index system consisting of a water quality parameter layer (dissolved oxygen, COD, heavy metal concentrations, etc.), a morphological characteristic layer (pollutant diffusion area, color and texture), and an environmental impact layer (distance to water intake, coverage of sensitive ecological zones). For example, in the case of a benzene leak at a chemical plant, the system integrates multi-dimensional data in real time: The water quality layer shows a sudden increase in benzene concentration from 0.1 mg / L to 8.2 mg / L (164 times the standard); the morphological layer shows a drone identifying an 800m x 50m rainbow-colored oil slick; and the environmental layer shows a fish spawning area 3.2 kilometers downstream of the contaminated zone. The index weights are dynamically calculated using an entropy weighting method. During flood season (flow velocities > 2 m / s) the morphological layer's weight increases to 45% (due to rapid diffusion). During dry season (flow velocities < 0.5 m / s), the water quality layer's weighting increases to 60% (due to the high concentration of pollutants). When sensitive ecological areas are detected, the weight of the environmental impact layer is automatically increased from 20% to 35%.

[0076] The multi-level dynamic triggering of early warnings utilizes a hierarchical response strategy, establishing a three-tiered early warning mechanism linked to emergency response plans. In this application, a red alert is issued when the score is >0.85: the emergency response system is automatically activated, surface robots are dispatched to deploy oil booms, and the water intake is simultaneously closed. When the score is 0.6 < ≤ 0.85, an orange alert is issued: an alert message is sent to the environmental protection department, and monitoring frequency is increased to every five minutes. When the score is 0.4 < ≤ 0.6, a yellow alert is issued: data is recorded and local grid personnel are notified for investigation. Using this implementation, during a heavy metal pollution incident, the system detected a comprehensive score of 0.72 (orange) at 2:05 PM and 0.89 (red) at 2:17 PM, triggering an emergency deployment of a fleet of robots, improving response time by 40 minutes compared to manual decision-making. By establishing a closed-loop feedback loop for early warning effectiveness, when the actual pollution losses after an early warning are less than 20% of the predicted value, the system automatically strengthens the relevant rule weights, increasing the accuracy of red alerts from 78% to 96%.

[0077] Autonomous Navigation Surface Robot Platform: Through autonomous control, the platform enables intelligent perception and dynamic obstacle avoidance. The steps are as follows: The surface robot is equipped with an obstacle perception system that integrates radar and multi-camera vision. It uses millimeter-wave radar (with a detection range of 300 meters) and a high-frame-rate camera (60fps) to construct a real-time 3D semantic map of the river. When it detects floating objects, bridge piers, or other vessels, it dynamically plans a path using an improved A* algorithm. During execution, the robot visually identifies a fishing vessel operating at a set distance ahead and initiates a serpentine maneuver, reducing its speed to maintain a safe distance while maintaining pollution adsorption operations and controlling the error. The robot platform integrates a robotic arm, adsorption device, and sensor array, completing a closed-loop "identification-processing-verification" process. For example, in a chemical raw material barrel leak incident, the robot used its robotic arm to grasp the damaged container (with a grasping accuracy of ±2cm) and simultaneously activated an ultrasonic demulsifier (at a frequency of 28kHz) to decompose the surfactant. The real-time monitoring of the treatment efficiency reached 89%, meeting the 85% target. The specific calculation method for the relevant technical indicators in this application is as follows: ;

[0078] in: is the influent pollutant concentration, is the concentration after treatment.

[0079] The intelligent emergency material dispatch model improves emergency dispatch efficiency through dynamic path planning and resource optimization. During execution, a multi-constraint optimization model based on an improved Dijkstra algorithm is constructed, taking into account water velocity, pollutant diffusion rate, and equipment transportation time. The algorithm incorporates real-time river topology data (water depth and flow velocity distribution). When predicting the trajectory of the center of mass of the pollution cluster, it uses the second-order Runge-Kutta method to solve the advection-diffusion equation and perform path planning to reduce response time. For example, in a transboundary pollution incident, the system mapped out three interception paths: the optimal path with an estimated evacuation distance and interception time of 8.2 kilometers and an estimated interception time of 42 minutes; the optimal path with an estimated evacuation distance and interception time of 9.1 kilometers and 48 minutes for Path A (avoiding the construction section of the river); and the alternative path B with an estimated evacuation distance and interception time of 7.8 kilometers and 39 minutes (requiring coordination with navigation control).

[0080] Ultimately, the optimal path was selected to successfully intercept 85% of the polluted clusters, achieving a 400% improvement in efficiency compared to traditional manual dispatching. Compared to traditional emergency response systems, this solution achieved a 30%-800% improvement in core metrics such as timeliness, depth of treatment, and sustained operational capability. This model increased the pollution control rate in the waters surrounding a polluted area from 55% to 92%, reducing actual losses and thus improving the efficiency and accuracy of pollutant identification and early warning methods.

[0081] The blockchain system utilizes a dual-chain heterogeneous architecture, building two independently operating blockchains: a monitoring chain and a transaction chain. This allows for data classification and storage, as well as permission isolation. The monitoring chain utilizes a highly efficient key-value database to store high-frequency data such as real-time water quality parameters and equipment status. Each block can hold 5,000 records, with a block generation interval of three seconds. The transaction chain utilizes a relational database structure to record transactional data, such as warning events and disposal logs, and supports complex queries. While processing 2,000 sensor data points per second, the monitoring chain also comprehensively records 78 operational logs from warning triggering, emergency response, to disposal acceptance, forming an immutable chain of evidence. For example, when environmental law enforcement officers need to trace a pollution incident, they can quickly locate the relevant warning ID through the transaction chain and directly link all raw sensor data from that period within the monitoring chain, achieving a 10-fold increase in query efficiency compared to a traditional single-chain architecture. This dual-chain architecture enables data write throughput of 12,000 transactions per second, a 400% increase compared to traditional environmental monitoring systems. The blockchain also includes a VRF-enhanced consensus mechanism for random trusted node selection. This mechanism utilizes a modified DPoS consensus algorithm and introduces a Verifiable Random Function (VRF). Each validator generates a public-private key pair and uses the VRF algorithm to output a random certificate. The system then selects 21 validators based on the hash value of the certificate. During node selection, nodes A and B generate random numbers 0x7a3f... and 0x9e1b..., respectively. These numbers are then sorted and selected as validators. A dynamic rotation mechanism is also implemented to reselect validators every 15 minutes. The system automatically marks three nodes in the contaminated zone as "high risk," reducing their selection weight to 30% of their normal value to ensure consensus network stability. Field tests have shown that this mechanism reduces block confirmation time from 5 seconds in traditional PBFT to 1.2 seconds, improving consensus efficiency by 316%.

[0082] A smart contract template library was established, including a standardized contract development framework, to construct three core contract templates: a data verification contract, which automatically verifies the validity of sensor data. In actual execution, a device will be flagged as abnormal if it reports COD > 1000 mg / L five times in a row; an early warning trigger contract, which presets threshold conditions for 32 pollution scenarios and automatically generates an orange alert when dissolved oxygen < 2 mg / L and ammonia nitrogen > 5 mg / L; and a traceability query contract, which supports multi-dimensional joint queries. A cross-chain oracle system was established in conjunction with the smart contract template library. A trusted data exchange channel was established within the blockchain system, connecting to the environmental protection department's database via the TLS 1.3 two-way authentication protocol, establishing a cross-chain bridge between the monitoring chain and the government affairs chain. When pollutant discharge permit data is needed, the oracle node automatically encrypts the query request and returns structured data after gateway verification, improving the accuracy of the smart contract template library's execution.

[0083] Simultaneously, a spatial index structure based on a kd-tree is constructed through rapid retrieval of multidimensional features, mapping pollution feature vectors into a multidimensional semantic space. If a new pollution incident occurs, the system matches historical cases through the following steps: extracting key features of the incident, performing a nearest neighbor search within the kd-tree, and returning the top five similar cases (with a similarity >85%) and their resolution plans. Once the plans are returned, the contract template is dynamically adapted. Based on the case matching results, the system automatically populates contract elements, such as the pollution cleanup contract, compensation calculation contract, and enforcement basis contract, further improving the accuracy of the smart contract template library.

[0084] A multimodal data lake is established to logically integrate data across all dimensions. During execution, a cross-media data storage architecture is constructed to uniformly store structured monitoring data, unstructured images and videos, and semi-structured report documents in a distributed object storage system. An adaptive metadata tagging system is used to add semantic tags such as time stamps, device IDs, and pollution types to each type of data. When a pollution incident occurs, the system automatically aggregates structured data, such as continuous monitoring records from water quality sensors around the leak point; unstructured data, such as drone videos and thermal images; and semi-structured data, such as corporate environmental emergency plans and emergency response logs. The data lake is divided into hot data SSD storage and cold data Blu-ray archiving through intelligent tiered storage technology, reducing data query response time from minutes in traditional systems to less than 200ms, and storage costs are reduced by 65%. In the process of summarizing data in the data lake, a deep semantic parsing engine is activated through knowledge extraction and entity association. At launch, a three-layer parsing system was constructed using the BiLSTM-CRF model. This includes an entity recognition layer, which extracts key elements from historical event reports; a relationship extraction layer, which establishes associations such as "pollutant → impact → ecologically sensitive area"; and an event reconstruction layer, which combines discrete entities to form complete event chains. For example, in a pollution incident, the system automatically extracts the key pattern from historical reports, such as "pollution diffusion speed increases by 40% when flow velocity exceeds 2 m / s during the rainy season," and generates association rules for "pollution diffusion rate - river flow velocity," providing knowledge support for subsequent smart contract generation.

[0085] Incremental knowledge fusion and graph updates utilize a dynamic learning algorithm, implementing a closed-loop learning mechanism of "data triggering - model fine-tuning - graph iteration." When the volume of new pollution incident data reaches a threshold, the system automatically initiates incremental training and performs subgraph sampling of the knowledge graph (retaining core nodes and three-hop neighbor relationships). It then updates embedding vectors using a contrastive learning algorithm to maintain the stability of the existing knowledge structure, and adjusts node connection weights using a topology optimization algorithm. For example, in a new pollutant incident, the system completed the following within 72 hours after receiving the first 50 test reports: adding a new "pollutant" entity node, establishing an edge relationship from "PFOA → Toxic Effects → Fish Embryonic Deformities," updating the "Water Treatment Process" node attributes, and adding a warning label stating "Activated Carbon Adsorption Efficiency ≤ 30%," effectively improving the efficiency of early warning implementation.

[0086] A multi-physics coupled digital twin is constructed, using the finite volume method to solve the three-dimensional hydrodynamic equations. This digital twin is then built, and a multi-physics coupled simulation engine is set up within the digital twin, combining the fluid dynamics module, the pollutant transport module, and the ecological impact assessment module. The dynamic error between the measured data and the simulation results is calculated using the multi-physics coupled simulation engine and a virtual-real data comparison algorithm. This dynamic error is then substituted into the model of the virtual-real verification system for parameter self-correction. When the parameter self-correction value remains above the set threshold, a Bayesian inversion algorithm is used to optimize the diffusion coefficient parameters. After self-correction, the model is fed into the XR training platform, integrating virtual reality accident simulation with augmented reality simulation to achieve precise location of pollution incidents and graded warnings of hazard levels. This three-dimensional hydrodynamic simulation engine, based on the finite volume method (FVM), discretizes the river channel into several grid cells, and characterizes water flow using the Navier-Stokes equations. When using the coupled pollutant transport module, the Lagrangian particle tracking method is used to simulate the diffusion process. The advection term corresponds to driving particle displacement based on real-time velocity field data, the diffusion term corresponds to the introduction of a turbulent pulsation model (k-ε model) to enhance accuracy, and the reaction term corresponds to the integration of a chemical kinetics database. This improves the computational speed of the digital twin and simulates the diffusion of contaminated belts to the downstream of the polluted river, thereby improving the consistency of the simulation results with the actual monitoring trajectory.

[0087] In the aforementioned simulation of the spread of the contaminated belt to the downstream of the polluted river, dynamic error correction and Bayesian inversion were employed, combined with a closed-loop feedback mechanism for virtual and real data and a virtual and real data comparison algorithm, to compare the simulation results with the actual values at the monitoring points every second. When the error between the diffusion rate in the monitoring area and the simulation result consistently exceeds 15%, the correction process was initiated. The correction process is as follows: data cleaning to eliminate sensor fault data; sensitivity analysis to determine the contribution of the diffusion coefficient to the error; Bayesian inversion to iteratively optimize the diffusion coefficient based on Markov Chain Monte Carlo (MCMC) sampling; and model hot update to adjust the optimized diffusion coefficient. For example, in one oil spill incident, this mechanism completed three rounds of parameter optimization within two hours of the incident, reducing the 72-hour diffusion prediction error from an initial 22% to 4.7%.

[0088] The XR fusion training platform includes a multimodal emergency simulation system, which serves as a hybrid of virtual and real-world XR training. This system includes virtual reality accident reconstruction, which generates interactive 3D scenes based on historical pollution data; augmented reality on-site overlay, which uses smart glasses to overlay simulated pollution zone outlines onto the actual river; and mixed reality collaborative drills, where remote command personnel use holographic projections to annotate key response points in real time. This setup allows pollution drills to be conducted through the XR platform, improving the efficiency of subsequent pollutant identification and treatment in actual rivers.

[0089] The process described above with reference to the flowchart can be implemented as a computer software program. The embodiments disclosed in the present invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network via a communication portion, and / or installed from a removable medium. When the computer program is executed by a central processing unit (CPU), the above-mentioned functions defined in the method of the present application are performed. It should be noted that the computer-readable medium mentioned above in the present application can be a computer-readable signal medium or a computer-readable storage medium or any combination of the above. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared segment or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media can include, but are not limited to: an electrical connection with one or more wire segments, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this application, a computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. Furthermore, in this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, which carries computer-readable program code. This propagated data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium that can transmit, propagate, or transfer a program for use by or in conjunction with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, electrical, optical, RF, or any suitable combination thereof.

[0090] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architectures, functions, and operations of the systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each box in the flowchart or block diagram can represent a module, program segment, or portion of code that contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions marked in the boxes can also occur in an order different from that marked in the accompanying drawings. For example, two boxes shown in succession can actually be executed substantially in parallel, or they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, as well as the combination of boxes in the block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified functions or operations, or can be implemented using a combination of dedicated hardware and computer instructions.

[0091] Those skilled in the art should understand that the embodiments of the present invention described above and shown in the accompanying drawings are only examples and do not limit the present invention. The objectives of the present invention have been fully and effectively achieved. The functional and structural principles of the present invention have been demonstrated and explained in the embodiments. Without departing from the principles, the implementation methods of the present invention may be subject to any deformation or modification.

Claims

1. A pollutant identification and early warning method for river pollution sources, characterized in that: The method comprises: Combine aerial, underwater and shore-based image acquisition equipment to establish a three-dimensional data acquisition architecture and build a three-dimensional monitoring network system; The spatial coordinates of monitoring nodes are obtained based on the three-dimensional monitoring network system. The spatial coordinates of each monitoring node are unified through the GNSS differential positioning system. The coordinates of the located monitoring nodes are calibrated in real time using Kalman filtering technology. The spatial coordinates of the monitoring nodes and the timestamp alignment algorithm are combined to synchronize the spatiotemporal coordinates of multi-source heterogeneous data from different monitoring nodes, thus establishing a unified spatiotemporal benchmark framework. Based on the acquired spatiotemporal coordinates of the monitoring nodes, data preprocessing is performed on the device side of the monitoring nodes to implement the deployment of edge computing nodes. By collecting spectral characteristics, water quality parameter correlation characteristics, and visual morphological characteristics of river pollutants collected by monitoring nodes, a pollutant feature library and water quality parameter correlation model are constructed to obtain river flow, pollutant diffusion rate, and historical pollution data. Establish an adaptive threshold warning mechanism, use a dynamic fuzzy comprehensive evaluation method to evaluate river flow, pollutant diffusion rate, and historical pollution events, and construct a spatiotemporal composite warning model to locate pollutants and provide graded warnings on the severity of their damage. The adaptive threshold warning mechanism adopts an improved fuzzy comprehensive evaluation algorithm. The specific content of the improved fuzzy comprehensive evaluation algorithm is as follows: Construct a multi-layer evaluation index system, including a water quality parameter layer, a morphological characteristic layer, and an environmental impact layer; establish an adaptive weight allocation mechanism based on the multi-layer evaluation index system, and use the entropy weight method to dynamically calculate the index weights based on the historical pollution event database and real-time river flow data; combine the multi-layer evaluation index system and the adaptive weight allocation mechanism to construct a nonlinear membership function library, forming trapezoidal, Gaussian, and S-shaped functions for different pollutant types to generate pollution risk levels; develop a multi-level early warning trigger strategy, if the comprehensive score is greater than the highest set threshold, activate a red warning; if the comprehensive score is less than the highest set threshold but greater than the lowest set threshold, activate an orange warning and link the emergency response system; implement a model online optimization mechanism, and use the reinforcement learning framework to dynamically adjust the fuzzy rule library according to the early warning accuracy; The specific content of the spatiotemporal composite early warning model is as follows: The finite volume method is used to solve the three-dimensional hydrodynamic equations and establish a digital twin. A multi-physics field coupling simulation engine is set up in the digital twin by combining the fluid dynamics module, pollutant transport module and ecological impact assessment module. The dynamic error between the measured data and the simulation results is calculated through the multi-physics field coupling simulation engine and the virtual-real data comparison algorithm, and the dynamic error is substituted into the model of the virtual-real verification system for parameter self-correction. When the parameter self-correction value is continuously greater than the set threshold, the Bayesian inversion algorithm is started to optimize the diffusion coefficient parameter to complete the self-correction and then enter the XR training platform.

2. The pollutant identification and early warning method for river pollution sources according to claim 1 is characterized in that: The aerial image acquisition device uses a multi-scale feature fusion deep learning algorithm for image processing, the specific content of which is as follows: Build an improved U-Net++ network architecture, introduce ResNet50 as the backbone network in the encoder part, add an attention mechanism module in the decoder part, and establish a cross-layer feature fusion channel; Set up a bidirectional reflectance distribution function model and an image enhancement module based on a physical model; establish a multi-task learning framework, use a weighted cross-entropy loss function to balance the category imbalance problem, and enable the drone image processing equipment to simultaneously complete pollutant identification, concentration inversion and diffusion trend prediction; establish an online incremental learning mechanism, realize the co-evolution of the model of each edge node through the federated learning framework, set the confidence threshold to automatically screen high-quality samples and update the model parameters of each edge node; build a virtual and real verification system, use digital twin technology to create a three-dimensional simulation environment for the river, and inject typical pollution scene data for model verification and optimization.

3. The pollutant identification and early warning method for river pollution sources according to claim 2 is characterized in that: The aerial image acquisition device uses an adaptive topology optimization method when performing image processing. The specific process of the adaptive topology optimization method is as follows: By developing an energy-aware routing algorithm, an optimization model for node residual energy, link quality, and image transmission requirements is established. Combining the optimization model with a hybrid networking protocol, a dual-mode transmission system integrating LoRaWAN low-power wide-area communication and 5G low-latency communication is established. A self-repair mechanism for abnormal nodes is established. If a node failure is detected, an improved particle swarm optimization method is used to quickly reconstruct the optimal communication path by combining blockchain and relay node selection algorithms. Mobile emergency monitoring nodes are set up, and self-propelled surface robots and underwater ROVs are deployed.

4. The pollutant identification and early warning method for river pollution sources according to claim 3 is characterized in that: The specific process of the unified spatiotemporal reference framework is as follows: A three-dimensional river channel model was established by combining satellite remote sensing data, drone oblique photography data, and lidar point cloud data. A dynamic projection matching algorithm was used to spatially align the mobile monitoring data with the static map, and an improved RANSAC algorithm was used to eliminate errors in the spatial alignment process. A spatiotemporal coding system was designed to establish a bidirectional index relationship between the monitoring data and the river channel spatial unit. Each spatial unit was encoded using Geohash, and the time dimension used a UNIX timestamp and leap second compensation mechanism. A pollutant diffusion simulation model was constructed, coupling the two-dimensional hydrodynamic model with the material transport equation: ； in, is the pollutant concentration, is the flow velocity vector, is the diffusion coefficient tensor, For source-sink items; develop the W three-dimensional display platform to support pollutant diffusion simulation, historical trajectory tracing and dynamic rendering of warning areas to achieve multi-dimensional data visualization.

5. The pollutant identification and early warning method for river pollution sources according to claim 4 is characterized in that: The spatiotemporal composite early warning model includes standards for pollutant treatment, the specific standards are as follows: The autonomous navigation surface robot platform is equipped with an obstacle perception module that integrates radar and multi-camera vision. The modular pollutant treatment unit is equipped with an automatic oil-absorbing felt spreading mechanism, an ultrasonic demulsification device, and an electrochemical degradation reactor. The treatment efficiency of the autonomous navigation surface robot platform meets the following conditions: ； in: is the influent pollutant concentration, is the concentration after treatment; It also includes a distributed energy system and an intelligent scheduling model for emergency materials. The distributed energy system adopts hybrid power supply. The intelligent scheduling model for emergency materials plans the optimal interception path based on the improved Dijkstra algorithm to realize real-time calculation of the movement trajectory of the center of mass of the pollution group.

6. The pollutant identification and early warning method for river pollution sources according to claim 5 is characterized in that: The deployment of the edge computing nodes is achieved through the blockchain system. The specific content of the blockchain system is as follows: A dual-chain heterogeneous architecture uses chains to store real-time monitoring data and transaction chains to record warning events and disposal logs; a lightweight consensus mechanism is established, using an improved DPoS algorithm, selecting multiple verification nodes, and introducing VRF random numbers; Combining the dual-chain heterogeneous architecture and lightweight consensus mechanism, a contract template library is built, and data verification contracts, early warning trigger contracts, and traceability query contracts are completed through the contract template library; a cross-chain oracle system is set up, and the environmental protection department database is connected through the TLS-N two-way authentication protocol to realize trusted interaction of on-chain and off-chain data in the blockchain system.

7. The pollutant identification and early warning method for river pollution sources according to claim 6 is characterized in that: The contract template library includes the following contents: Set up a multimodal data lake to integrate structured monitoring data, unstructured images and videos, and semi-structured report documents; Combine the knowledge extraction engine and the BiLSTM-CRF model to extract entity relationship triplets from historical event reports; establish an incremental knowledge fusion algorithm and use DS evidence theory to synthesize different credibility values; combine the incremental knowledge fusion algorithm and the knowledge graph dynamic update strategy to establish a case reasoning mechanism, and quickly retrieve similar historical scenarios through the kd tree.

8. A system for patrolling river pollution, characterized in that: The system is used to execute a pollutant identification and early warning method for river patrol pollution sources according to any one of claims 1-7.

Images