A full-process intelligent management system for red tide verification tasks in marine areas

Abstract

This invention relates to the field of marine environmental monitoring technology, specifically disclosing a full-process intelligent management system for red tide verification tasks in marine areas. The system includes a visualization control platform, which is communicatively connected to the following modules: a fragment task integration module, used to integrate fuzzy C-means clustering and spatial topological constraint algorithms to automatically merge fragment tasks at boundaries and assign them to responsible parties. This fragment task integration module includes a fuzzy boundary membership weighting unit and a mandatory assignment unit for the jurisdiction buffer. This invention automatically merges fragment tasks at boundaries and clarifies the responsible parties based on overlap and historical rules by applying fuzzy C-means clustering membership weighting to the edge pixels of remotely sensed water color anomaly areas and applying jurisdiction assignment based on dynamic buffers. This solves the problem of unclaimed areas, ensures rapid implementation of verification tasks, eliminates ambiguity in boundary task assignment, and improves emergency response timeliness.

Description

Technical Field

[0001] This invention relates to the field of marine environmental monitoring technology, and in particular to a full-process intelligent management system for red tide verification tasks in marine areas. Background Technology

[0002] In recent years, red tide disasters have become increasingly frequent and severe, posing a continuous impact on the marine ecological environment, fishery resources, and coastal economic and social development. Currently, red tide monitoring, early warning, and disaster prevention work involves multiple business links, including remote sensing monitoring, on-site verification, trend prediction, and data management, with different functional departments and business categories participating collaboratively. Under the existing working model, red tide verification tasks rely on manual triggering and decentralized management, and there is a lack of unified process support for task generation, allocation, execution, feedback, and review. Task status tracking and accountability are also inadequate.

[0003] In existing red tide verification tasks in marine areas, the blurred boundaries of remotely sensed water color anomaly areas often lead to fragmented tasks or unclaimed areas at the jurisdictional boundaries of automatically segmented verification tasks, causing delays in emergency response and shirking of responsibility. Furthermore, during the 72-hour red tide migration prediction process, the wind and flow field data required for model driving may exhibit spatial discontinuities or local anomalies in actual operations due to observation system errors, unstable data assimilation, or insufficient spatiotemporal resolution. This can further trigger branch drift in the predicted path 24 hours later. The current system cannot provide a single decision point for the subsequent deployment of verification stations. Furthermore, when manually deploying stations, subjective selection of points is often based on areas of abnormal water color from remote sensing, lacking guidance from prior spatial knowledge of areas with historical outbreaks of highly toxic algae. This easily leads to the omission of key areas that, although currently not showing significant water color characteristics, have historically experienced outbreaks of highly toxic algae, resulting in repeated rejections during the verification report review stage and low process efficiency. Therefore, how to construct an intelligent management system that is a closed-loop system covering the entire chain from task generation and path prediction to station deployment, and avoids the accumulation of deviations between levels, is the problem that this invention aims to solve. Summary of the Invention

[0004] To overcome the shortcomings of the prior art, the present invention provides a full-process intelligent management system for red tide verification tasks in marine areas, which can effectively solve the problems involved in the prior art.

[0005] The objective of this invention can be achieved through the following technical solution: This invention provides a full-process intelligent management system for marine red tide verification tasks, including a visual control platform, which is communicatively connected to the following modules: The fragment task integration module is used to address the problem of fragmented tasks and unclaimed areas caused by the blurred boundaries of remote sensing water color anomaly areas in marine red tide verification tasks. It integrates fuzzy C-means clustering and spatial topological constraint algorithms to automatically merge fragmented tasks at the boundary and assign them to the responsible party, eliminate hierarchical deviations in the task generation process, ensure clear verification task boundaries, eliminate fragmented tasks and unclaimed areas at the boundary, and ensure traceability of emergency response responsibilities. The drift ridge generation module is used to address the 72-hour migration prediction branch drift phenomenon caused by grid-level invalid values, outliers, or spatiotemporal continuity abrupt changes in local wind and flow field data. It integrates particle swarm optimization and Kalman filtering algorithms, uses similar environmental field segments in the historical red tide case library as reference particles, performs local dynamic filling and smoothing correction on abnormal grid points, eliminates multi-branch ambiguity, generates a single confidence drift ridge, and outputs a 3-hour interval probability kernel density surface, providing a deterministic decision-making benchmark for station deployment. The multi-objective constraint coding module is used to transform the problem of red tide verification station layout in the sea area into a multi-objective optimization model. It defines constraints including coverage probability contour lines (derived from probability kernel density surface), sensitive area distance and ocean current direction (affecting drift path), and performs Pareto coding on station coordinates and number to construct an initial population. This provides a computable optimization problem expression framework for subsequent genetic iterations, transforming station layout into a computable optimization problem. The Pareto optimization module for station locations uses the NSGA-II algorithm to perform non-dominated sorting and crowding comparison of station locations, iteratively evolves station combination, automatically balances targets including high probability area coverage, high hazard area response and ocean current direction matching, outputs a Pareto front set of stations, corresponding to the theoretically lowest verification rejection risk path, automatically balances multiple conflict targets, and outputs the station location scheme with the theoretically lowest rejection risk. The closed-loop review and feedback module is used to reverse-map the rejection records in the review stage of the verification report to the preceding task generation, path prediction and station optimization stages. By analyzing the reasons for rejection and automatically correcting the fragment merging rules, drift correction parameters and station optimization weights, a full-chain adaptive closed loop is formed, which gradually reduces the accumulation of hierarchical deviations and the process rejection rate.

[0006] Preferably, the fragment task integration module includes a fuzzy boundary membership weighting unit and a jurisdiction buffer mandatory affiliation unit; The fuzzy boundary membership weighting unit is used in the task of verifying red tides in the sea area to calculate the membership degree of the edge pixels of the remote sensing water color anomaly area to the adjacent jurisdiction area using fuzzy C-means clustering, and to weight them according to spatial distance and spectral characteristics, outputting fuzzy membership weights to quantify the degree of boundary fuzziness, providing a continuous membership field for fragment merging, quantifying the degree of boundary fuzziness, and providing a continuous transition judgment basis for fragment merging. The mandatory attribution unit of the jurisdiction buffer zone generates a dynamic buffer zone based on the jurisdiction boundary, combines fuzzy attribution weights to merge fragmented tasks in the intersection area using topological constraints, and assigns tasks with ambiguous attribution to the corresponding responsible party according to the overlap of the buffer zone and the historical task allocation rules. It outputs a verification task package with clear boundaries and definite responsibilities, ensuring that there are no unclaimed areas in the water color anomaly area at the boundary, ensuring that emergency response responsibilities are traceable, and forcibly eliminating attribution ambiguities.

[0007] Preferably, the fuzzy boundary membership weighting unit specifically includes: For remote sensing water color anomaly areas found in red tide verification in the sea area, the spatial coordinates and multi-band spectral reflectance of edge pixels are extracted to construct a pixel-level feature matrix, which serves as the differential input basis for fuzzy C-means clustering, establishes a pixel-level quantization basis, and improves the objectivity of boundary attribution criteria. The number of clusters is set to the total number of adjacent jurisdictions. The membership degree of each edge pixel to each jurisdiction is iteratively calculated. An S-shaped distance decay function is constructed based on the spatial distance from the pixel to the jurisdiction boundary. The membership degree is nonlinearly weighted and corrected. The closer the distance, the weaker the decay and the farther the distance, the stronger the decay. This strengthens the weight of the near boundary and suppresses the ambiguity interference of the far boundary. The weighted membership degree is compared with the preset first threshold and second threshold: pixels above the first threshold are directly assigned to the corresponding jurisdiction, pixels below the second threshold are marked as forced merging seed points, and pixels between the two are output as a continuous membership degree field. Each pixel in this field carries a fuzzy membership weight for each jurisdiction, providing a quantitative basis for determining the boundary transition region for fragment task merging. That is, the first threshold is a high threshold and the second threshold is a low threshold, realizing three-level differential deconstruction and avoiding the omission of membership caused by absolute cutting.

[0008] Preferably, the mandatory attribution unit of the jurisdiction buffer specifically includes: Dynamic jurisdictional buffer zones are generated by buffering outwards at a preset distance along the administrative boundaries of each jurisdiction, and further buffered inwards to generate a core jurisdictional inner circle. The annular area between the two is defined as the attribution negotiation zone. The intersection area between different buffer zones is identified as a potential attribution ambiguity area for fragmented tasks, ensuring that fragmented tasks at the intersection are fully included in the judgment scope and eliminating blind spots. Within the ambiguous region of belonging, a continuous membership field is superimposed, and a spatial topological adjacency graph is constructed for multiple fragmented tasks in the intersection region. Connectivity component analysis is used to merge the tasks and form a task package to be assigned, thereby realizing the automated merging of fragmented tasks and avoiding omissions from manual splicing. For each task package to be assigned, the overlap ratio between it and each managed buffer is calculated, and a mandatory unique assignment decision function is introduced: this function uses the buffer overlap as the main factor and the default responsibility party weight at the boundary in the historical task allocation rule base as the secondary factor, and uses a weighted voting mechanism to force assignment to a unique primary responsible party, while pushing it to one or more secondary responsible parties. The primary responsible party must accept the task, and the secondary responsible parties can choose the task, ensuring that there are no unclaimed areas at the boundary, ensuring that each task package has a responsible party, and preventing the shirking of responsibility.

[0009] Preferably, the drift ridge generation module includes a correction unit and a single-confidence drift ridge generation unit; The correction unit is used to detect and mark grid-level invalid values, outliers, or spatiotemporal continuity abrupt changes in wind field and flow field data in the 72-hour migration prediction of red tide verification in marine areas. It retrieves similar environmental field fragments from the historical red tide case library as reference particles, uses particle swarm optimization algorithm to search for the optimal local correction parameters for the abnormal grid points, and uses Kalman filtering to smoothly replace the outliers. It outputs a continuous and consistent environmental driving field after anomaly correction to ensure prediction stability. The single-confidence drift ridge generation unit, based on the modified continuous environmental driving field, uses a particle tracking and probability density fusion method to eliminate the multi-branch drift ambiguity of the original model, and outputs a unique confidence drift ridge (optimal migration path) and a probability kernel density surface at 3-hour intervals, providing unambiguous and quantifiable red tide migration prediction basis for subsequent verification station deployment.

[0010] Preferably, the correction unit specifically includes: Real-time monitoring of wind and flow field grid data; identification of local outliers (including invalid values, outliers, or abrupt changes in continuity) in the time or space dimensions based on the spatiotemporal continuity scoring function; and morphological dilation processing of the outlier grid region to determine the boundary of the outlier influence range. An environmental field feature hash index is constructed from the historical red tide case library. Several historical environmental fields most similar to the current abnormal area environmental features are retrieved as reference particles. A dynamic time warping algorithm is introduced to align environmental field sequences of different time phases. A particle swarm optimization algorithm is used to search for the optimal local correction parameters for the abnormal grid points. The inertial weight in the particle velocity update formula is designed as an adaptive function that is negatively correlated with the severity of the abnormality. The optimized local correction parameters are input into the Kalman filter predictor to dynamically correct and smoothly replace the abnormal grid point values ​​at the current time (normal grid points remain unchanged), and output a continuous and consistent environmental driving field sequence after anomaly correction.

[0011] Preferably, the single-confidence drift ridge generation unit specifically includes: Based on the continuous and consistent environment driving field of the output, a multi-scale virtual particle swarm is initialized using the Lagrange particle tracking framework, and a fourth-order Runge-Kutta method is introduced into the coupled integral of the wind field and the flow field to improve the trajectory calculation accuracy, which significantly improves the calculation accuracy of particle trajectories in complex flow fields. A time-varying kernel density estimation model is constructed for the spatial positions of all particles within a 72-hour prediction period. An adaptive bandwidth selection mechanism is introduced to perform differentiated smoothing for different drift stages. By identifying the global principal mode ridge in the density field, ambiguous branch drift paths are automatically removed to ensure the extraction of the main drift trend. The graph cut algorithm is used to optimize the spatial continuity of the principal mode ridge line. The single curve after topological repair is extracted as the unique confidence drift ridge line. The probability kernel density surface is generated in the form of the predicted confidence interval envelope every 3 hours to ensure that the probability field satisfies the mass conservation constraint in the forward integration. The single optimal path and the probability density grid that satisfies the conservation constraint are output.

[0012] Preferably, the multi-objective constraint encoding module specifically includes: From the probability kernel density surface at 3-hour intervals, the contour line tracking algorithm is used to extract several coverage probability contour lines corresponding to the preset probability threshold. The area enclosed by the contour lines is discretized into candidate station deployment grid units, and an initial coverage probability weight is assigned to each unit to ensure that the candidate stations are concentrated in the red tide high-incidence area and improve the deployment targeting. A multi-objective optimization function is constructed, which includes coverage probability, sensitive area distance, and ocean current direction consistency. The coverage probability objective is obtained by integrating the kernel density value of the station location in the probability kernel density surface. The sensitive area distance objective is obtained by calculating the shortest spatial distance between the candidate station and the adjacent sensitive area (including bathing beach, aquaculture area, nuclear power plant water intake, etc.) and then performing inverse normalization. The closer the distance, the higher the priority. The ocean current direction consistency objective is obtained by calculating the cosine value of the angle between the direction of the line connecting the stations and the current mainstream current direction and performing inverse normalization. This achieves a multi-dimensional overall balance between coverage effect, ecological risk, and drift dynamics. The coordinates and number of stations in the grid cells of the candidate stations are jointly Pareto encoded. Each encoded bit corresponds to the selected state of a grid cell, generating an initial population containing multiple combinations of station numbers and distributions. At the same time, at least one pair of control stations is forcibly included in the optimization constraints, with one station located in the high-value area of ​​the probability kernel density surface and the other located in the low-value area or outside the boundary of the anomaly area, to ensure the comparative monitoring capability between red tide areas and non-red tide areas.

[0013] Preferably, the station Pareto optimization module specifically includes: Using the constructed initial population as the starting point for optimization, the fitness vector of each individual's position combination under the multi-objective optimization function is calculated. An improved fast non-dominated sorting algorithm is used to divide the Pareto front into multiple levels, and a Pareto front thickness coefficient is introduced to identify the boundary solutions and internal solutions in the non-dominated solution set, ensuring that extreme position schemes are not eliminated and improving the diversity of the solution set. Within each non-dominated layer, the crowding distance of each station individual is calculated. This distance takes into account both the spatial distribution density of the station and the neighborhood density in the objective function space. Individuals with high crowding are preferentially retained to maintain population diversity. Offspring populations are generated by simulating binary crossover and polynomial mutation operations. An elite retention strategy is used to merge the parent and offspring populations to prevent premature convergence of the algorithm and ensure balanced coverage of the station space. The process iterates and evolves until the rate of change of the hypervolume index of the Pareto front is lower than the convergence threshold within a consecutive preset number of algebras. The final Pareto front station set is then output. Each station combination in this set is a non-dominated solution, corresponding to the station layout scheme with the lowest theoretical risk of rejection. This scheme can be directly called by the on-site verification task, and the station layout scheme with the lowest rejection risk is automatically generated from the optimal solution set.

[0014] Preferably, the closed-loop review feedback module specifically includes: During the review of verification reports, rejection records are captured, and rejection opinions are analyzed through natural language processing to extract the categories of rejection reasons, including incorrect boundary attribution, deviation of predicted path, insufficient station coverage, or omission of high-hazard areas. This enables automated and accurate identification of rejection reasons, reducing the time-consuming manual judgment. The reasons for rejection are mapped in reverse to the fragment task integration module, the drift ridge generation module, and the multi-objective constraint encoding module, which trigger adaptive correction mechanisms respectively: dynamically adjust the distance attenuation coefficient in the membership weighting function for the fragment merging rules, update the process noise covariance matrix of the Kalman filter for the drift correction parameters, and redistribute the weight vector of the three objective functions for the station optimization weights using the gradient descent method, so as to achieve fixed-point correction of the root cause of the problem and avoid blind adjustment of global parameters; Record the rejection rate and adjustment parameter combination after each adjustment, build a rejection-correction mapping library, and iteratively optimize the correction rules through reinforcement learning strategy to form a closed-loop feedback archive, which is used to reduce the accumulation of hierarchical deviations and process rejection rate one by one. This allows the system to learn the optimal parameters autonomously during continuous operation, and the rejection rate gradually converges and decreases.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. This intelligent management system for the entire process of red tide verification tasks in marine areas automatically merges fragmented tasks at the boundary by performing fuzzy C-means clustering and weighted membership of edge pixels in remotely sensed water color anomaly areas and assigning jurisdiction based on dynamic buffers. It also clarifies the responsible party based on overlap and historical rules, solves the problem of unclaimed areas, ensures the rapid implementation of verification tasks, eliminates ambiguity in the attribution of boundary tasks, and improves the timeliness of emergency response.

[0016] 2. This intelligent management system for the entire process of red tide verification in marine areas integrates particle swarm optimization and Kalman filtering algorithms to correct grid-level invalid values, outliers, or prediction branch drift caused by spatiotemporal continuity abrupt changes in local wind and flow field data. It generates a single confidence drift ridge and a probability kernel density surface at 3-hour intervals, eliminates multipath ambiguity, provides deterministic migration prediction paths, supports precise station deployment, and enhances the reliability of red tide migration prediction.

[0017] 3. This intelligent management system for the entire process of red tide verification tasks in marine areas transforms the station deployment problem into a multi-objective optimization model, automatically balancing multiple conflicting objectives such as coverage probability, distance to sensitive areas and consistency with ocean current direction. It uses the NSGA-II algorithm to output a Pareto front set of stations and forcibly includes control station pairs to avoid the problem of missing low-probability, high-hazard areas by manual deployment, thereby reducing the rejection probability of verification reports from the source. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the workflow of a full-process intelligent management system for marine red tide verification tasks according to the present invention. Figure 2 This is a schematic diagram of the system architecture of a full-process intelligent management system for marine red tide verification tasks according to the present invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0020] Example 1, please refer to Figure 1 , Figure 2 This invention provides a technical solution: a full-process intelligent management system for red tide verification tasks in marine areas, including a visual control platform, which has the following communication connections: The fragment task integration module is designed to address the issue of fragmented tasks and unclaimed areas caused by blurred boundaries of remotely sensed water color anomaly areas in marine red tide verification tasks. It integrates fuzzy C-means clustering and spatial topological constraint algorithms to automatically merge fragmented tasks at the boundary and assign them to the responsible party. This eliminates hierarchical bias in the task generation process, ensures clear verification task boundaries and unique responsibility, eliminates fragmented tasks and unclaimed areas at the boundary, and ensures traceability of emergency response responsibility. The fragment task integration module includes a fuzzy boundary membership weighted unit and a mandatory assignment unit for the jurisdiction buffer zone. The fuzzy boundary membership weighting unit is used in marine red tide verification tasks to calculate the membership degree of edge pixels in remotely sensed water color anomaly areas to adjacent jurisdictional areas using fuzzy C-means clustering. Weighting is applied based on spatial distance and spectral characteristics, outputting fuzzy membership weights to quantify the degree of boundary fuzziness. This provides a continuous membership field for fragment merging, quantifies the degree of boundary fuzziness, and provides a continuous transitional judgment basis for fragment merging. For remotely sensed water color anomaly areas in marine red tide verification, the spatial coordinates and multi-band spectral reflectance of edge pixels are extracted to construct a pixel-level feature matrix, serving as the differential input basis for fuzzy C-means clustering. This establishes a pixel-level quantization foundation, improves the objectivity of boundary membership judgment, and sets the number of clusters to the total number of adjacent jurisdictional areas, iteratively calculating the membership degree of each edge pixel. For each jurisdiction, the membership degree is determined, and an S-shaped distance decay function is constructed based on the spatial distance from the pixel to the jurisdiction boundary. The membership degree is then nonlinearly weighted and corrected, where the decay is weaker when the distance is closer and stronger when the distance is farther, thus strengthening the weight of near-boundary assignment and suppressing the fuzzy interference of far-boundary assignment. The weighted membership degree is then compared with a preset first threshold and a second threshold: pixels above the first threshold are directly assigned to the corresponding jurisdiction, pixels below the second threshold are marked as forced merging seed points, and pixels between the two are output as a continuous membership degree field. Each pixel in this field carries a fuzzy assignment weight for each jurisdiction, providing a quantitative basis for determining the boundary transition area for fragment task merging. That is, the first threshold is a high threshold and the second threshold is a low threshold, realizing three-level differential deconstruction and avoiding the omission of assignment caused by absolute cutting. It should be noted that edge detection is performed on the identified water color anomaly areas in the remote sensing image. The spatial coordinates of each pixel on the boundary of the anomaly area and its multi-band spectral reflectance data are extracted. The spatial coordinates adopt the WGS-84 geodetic coordinate system. The spectral reflectance is extracted from the visible to near-infrared range of the GOCI sensor or domestic satellites (such as HY-1C / D). The spatial coordinates of each pixel and the reflectance of each band are concatenated to form a multi-dimensional feature vector, which constitutes a pixel-level feature matrix. The number of clusters in the fuzzy C-means clustering is set to be equal to the total number of all jurisdictional areas adjacent to the water color anomaly area, usually 2 to 4. During the clustering process, the fuzziness index is set to 2.0, and the maximum number of iterations is set to 200. When the change in the objective function between two consecutive iterations is less than 1×10, the clustering is considered successful. -5Convergence is determined upon iteration. After the iteration is complete, the initial membership degree of each pixel to each jurisdiction is output, and the sum of the membership degrees of the same pixel to all jurisdictions is 1. Based on the initial membership degree, a spatial distance correction factor is further introduced. For each edge pixel, the spatial distance between it and the administrative boundary of each jurisdiction is calculated in meters. An S-shaped distance decay function is constructed, where the decay coefficient is 0.0015 and the decay center distance is set to 500 meters. When the pixel distance to the administrative boundary is less than 500 meters, the decay function value approaches 1, and the membership degree remains at a high level; when the distance is greater than 500 meters, the decay function value approaches 1. At 0 meters, the decay function value gradually approaches 0, significantly reducing the membership degree. The initial membership degree is multiplied by this decay function value to obtain a weighted, corrected membership degree. A first threshold of 0.75 and a second threshold of 0.25 are set. Pixels with a weighted membership degree of 0.75 or higher are directly assigned to the corresponding jurisdiction and no longer participate in subsequent merging negotiations. Pixels with a weighted membership degree of 0.25 or lower are marked as forced merging seed points, serving as the basic unit for task merging between different jurisdictions. For pixels with a membership degree between 0.25... Pixels with a membership degree between 0.75 and 0.75 are output as a continuous membership field, in which each pixel carries a fuzzy membership weight value for each jurisdiction. For pixels with membership degrees in the middle range, a data structure for the continuous membership field is constructed. This field is stored in raster form with the same raster resolution as the original remote sensing image. Each raster cell stores a membership vector, the dimension of which is equal to the total number of adjacent jurisdictions. Each component represents the weighted membership value of the pixel to the corresponding jurisdiction. The continuous membership field covers the transition zone within 3 kilometers on both sides of the boundary of the water color anomaly area. This range covers most of the areas where debris tasks are generated at the boundary. After the continuous membership field is output, it is transmitted to the mandatory membership unit of the jurisdiction buffer and overlaid with the dynamic jurisdiction buffer for analysis. During the overlay analysis, the fuzzy membership weight in the continuous membership field directly affects the membership determination result of each debris task package within the membership negotiation zone. The jurisdiction with the higher weight gets a larger share of votes in the membership decision of the pixel. The generation of the continuous membership field is completed independently in each water color anomaly detection task, without relying on historical data, ensuring the real-time and objectivity of the boundary membership determination. The jurisdictional buffer system mandates attribution units, generating dynamic buffers based on jurisdictional boundaries. It combines fuzzy attribution weights to perform topological constraint merging of fragmented tasks in the intersection area. For tasks with ambiguous attribution, it assigns them to the corresponding responsible party based on buffer overlap and historical task allocation rules, outputting a clear-boundary, responsibility-defined verification task package. This ensures there are no unclaimed areas in the water color anomaly zone at the boundary, guaranteeing traceability of emergency response responsibility and forcibly eliminating attribution ambiguity. Dynamic jurisdictional buffers are generated outwards at a preset distance along the administrative boundaries of each jurisdiction, and further buffered inwards to generate a core jurisdictional inner circle. The annular area between these two is defined as the attribution negotiation zone. The intersection area between different buffers is identified as a potential attribution ambiguity area for fragmented tasks, ensuring that fragmented tasks at the boundary are fully included in the judgment scope. In addition to blind spots, within the ambiguous attribution area, a continuous membership field is superimposed. A spatial topological adjacency graph is constructed for multiple fragmented tasks in the intersection area. Connectivity component analysis is used to merge tasks and form task packages to be assigned, realizing the automated merging of fragmented tasks and avoiding omissions from manual splicing. For each task package to be assigned, the overlap ratio between it and each managed buffer is calculated, and a mandatory unique attribution decision function is introduced: this function uses the buffer overlap as the main factor and the default responsibility party weight at the boundary in the historical task allocation rule base as the auxiliary factor. A weighted voting mechanism is used to force assignment to a unique primary responsibility party, while pushing it to one or more secondary responsibility parties. The primary responsibility party must accept the task, and the secondary responsibility parties can choose the task, ensuring that there are no unclaimed areas at the boundary and guaranteeing that each task package has a responsible party, thus preventing the shirking of responsibility. It should be noted that when performing the mandatory attribution operation for the jurisdiction buffer zone, a dynamic jurisdiction buffer zone is first generated by buffering outwards by 3000 meters along the administrative boundaries of each jurisdiction area to cover the potential impact range of the water color anomaly zone at the boundary; simultaneously, a core jurisdiction inner circle is generated by buffering inwards by 500 meters. The annular area between the dynamic jurisdiction buffer zone and the core jurisdiction inner circle is defined as the attribution negotiation zone, with a width of 2500 meters. Subsequently, the spatial intersection areas between different jurisdiction buffer zones are identified and marked as potential attribution ambiguities for the fragment task. The generation of attribution ambiguities is based on the spatial overlay analysis of the administrative boundary and the spatial location of the real-time remote sensing water color anomaly zone, and is only considered when two or more jurisdictions... The subsequent attribution determination process is triggered only when there is overlap in the buffer zones and edge pixels of water color anomalies exist within the overlapping area. The parameters of this buffer zone are determined statistically based on the actual spatial distribution at historical red tide boundary areas, ensuring coverage of over 95% of the historical debris task generation areas. Within the attribution ambiguity zone, a continuous membership field generated by fuzzy boundary membership weighted units is first loaded. This field has a raster resolution of 10 meters, consistent with the resolution after resampling of the GOCI remote sensing image. Then, a spatial topological adjacency graph is constructed for all debris task polygons within the attribution ambiguity zone. Nodes represent individual debris task polygons, and edges represent spatial adjacency or overlap relationships between two polygons, using connected component analysis. The quantitative analysis algorithm traverses the topology graph, merging all fragmented tasks belonging to the same connected component into a single task package to be assigned. The boundary of the merged task package is the outer contour of the union of the original fragmented task polygons. The attribute table records the number of original fragmented tasks and their source jurisdiction identifiers. Connected component analysis does not rely on any external threshold parameters, ensuring that all fragmented tasks within the same connected component are completely merged. For each task package to be assigned, the ratio of its overlap area with each jurisdictional buffer is calculated, and this ratio is normalized into an overlap weight. The mandatory unique assignment decision function adopts a weighted voting mechanism, with the main factor being the buffer overlap weight, accounting for 70% of the voting weight, and the secondary factor... The default responsibility party weight for this boundary is determined by the historical task allocation rule base, accounting for 30%. The historical task allocation rule base stores the responsibility attribution records of all completed verification tasks for each boundary in the past three years. The default responsibility party weight is generated by frequency statistics. A weighted voting mechanism calculates the total score for each candidate jurisdiction and determines the jurisdiction with the highest total score as the sole primary responsibility party. If the highest scores are tied, the one with the higher historical default responsibility party weight is selected and pushed to one or more secondary responsibility parties. The primary responsibility party must accept the task, while the secondary responsibility parties can choose to accept the task, ensuring that there are no unclaimed areas at the boundary. After the assignment is completed, the historical task allocation rule base is automatically updated, and the assignment result is recorded. The drift ridge generation module is used to address the 72-hour migration prediction branch drift phenomenon caused by grid-level invalid values, outliers, or spatiotemporal continuity abrupt changes in local wind and flow field data. It integrates particle swarm optimization and Kalman filtering algorithms, uses similar environmental field segments in the historical red tide case library as reference particles, performs local dynamic filling and smoothing correction on abnormal grid points, eliminates multi-branch ambiguity, generates a single-confidence drift ridge, and outputs a 3-hour interval probability kernel density surface, providing a deterministic decision-making benchmark for station deployment. The drift ridge generation module includes a correction unit and a single-confidence drift ridge generation unit. The correction unit is used in the 72-hour migration prediction of red tide verification in marine areas to detect and mark grid-level invalid values, outliers, or spatiotemporal continuity abrupt changes in wind and flow field data. It retrieves similar environmental field fragments from a historical red tide case database as reference particles, uses a particle swarm optimization algorithm to search for optimal local correction parameters for anomalous grid points, and performs smooth replacement of outliers using Kalman filtering. The output is a continuous and consistent environmental driving field after anomaly correction. It monitors wind and flow field grid data in real time, identifies local outliers (including invalid values, outliers, or continuity abrupt changes) in the time or spatial dimensions based on a spatiotemporal continuity scoring function, and performs morphological dilation processing on the anomalous grid point regions to ensure... Define the boundary of the anomaly's impact range, construct an environmental field feature hash index from the historical red tide case library, retrieve several historical environmental fields most similar to the current anomaly area's environmental features as reference particles, introduce a dynamic time warping algorithm to align environmental field sequences of different time phases, use a particle swarm optimization algorithm to search for the optimal local correction parameters for the anomaly grid points, and design the inertia weight in the particle velocity update formula as an adaptive function negatively correlated with the severity of the anomaly. Input the optimized local correction parameters into a Kalman filter predictor to dynamically correct and smoothly replace the anomaly grid point values ​​at the current moment (normal grid points remain unchanged), and output a continuous and consistent environmental driving field sequence after anomaly correction. It should be noted that real-time access to wind and flow field grid data is used, and a spatiotemporal continuity scoring function is employed to perform continuity checks in both the temporal and spatial dimensions. In the temporal dimension, a quadratic difference sequence is constructed using data from the same grid point across three adjacent time intervals (1-hour intervals). A time abrupt change anomaly is identified when the difference exceeds three times the root mean square error (RMSE) of the historical data for that grid point. In the spatial dimension, the mean deviation between each grid point and its eight neighboring grid points is calculated. A spatial abrupt change anomaly is identified when the deviation exceeds a preset 30% threshold. Invalid or missing values ​​in the data are directly identified as anomalous grid points. After anomaly labeling is completed... A circular structuring element with a radius of 3 grid points is used to perform morphological expansion on the anomalous region. This process is repeated twice to connect the scattered, isolated anomalous grid points into a continuous region. This region is then expanded outwards by 5 grid points as the boundary of the anomalous influence range, ensuring that the corrected region fully covers the potential space for anomalous propagation. An environmental field feature hash index is constructed. The index key is generated by local sensitive hashing of four features: regional average wind speed, prevailing wind direction, average flow velocity, and vorticity variance. The hash table has a length of 1024. Hash values ​​matching the current anomalous region's environmental features (i.e., the vector formed by the above four features) are retrieved from the historical red tide case database. For cases where the Hamming distance is less than 3, the 20 most similar historical environmental field segments are selected as reference particles. A dynamic time warping algorithm is introduced to align environmental field sequences from different time phases. The warping window width is set to 6 hours. A particle swarm optimization algorithm is used to search for the optimal local correction parameters for anomalous grid points. 50 particles are initialized, and the search space consists of three dimensions for filling the parameters: interpolation radius, time smoothing coefficient, and spatial gradient weight. The inertia weight in the particle velocity update formula is designed as an adaptive function negatively correlated with the severity of the anomaly, where the severity of the anomaly is defined as the ratio of the total number of anomalous grid points to the monitored... The ratio of the total number of grid points within the region, the inertial weight decreases linearly from 0.9 to 0.4 in actual iterations, and the higher the severity of the anomaly, the smaller the inertial weight, in order to enhance the local search capability; the optimal filling parameters output by the particle swarm optimization algorithm are input into the Kalman filter predictor. The state vector of the Kalman filter contains the wind field U / V components and the flow field X / Y direction velocities of each grid point at the current moment, a total of four-dimensional state variables. The process noise covariance matrix is ​​initially set as a diagonal matrix, and the values ​​of the diagonal elements refer to the time smoothing coefficient in the filling parameters; the observation noise covariance is set according to the nominal accuracy of the sensor, and the wind speed observation noise is set to 0.With a flow rate of 5 m / s and a flow direction observation noise of 5°, the filter's prediction-update cycle is executed in 1-hour increments. The prediction step employs a linear extrapolation model, while the update step fuses the observed data with the predicted values. For grid points marked as missing, the filter skips the observation update step and fills in the missing data solely through the prediction step. For grid points marked as abrupt anomalies, the filter adaptively weights the data based on the Mahalanobis distance between the predicted and observed values. When the Mahalanobis distance exceeds 3, the observation weights are significantly reduced. The filtered environmental driving field sequence simultaneously satisfies both spatial continuity and temporal smoothness requirements and serves as the input data for the downstream drift prediction model. A single-confidence drift ridge generation unit, based on a modified continuous environmental driving field, employs a particle tracking and probability density fusion method to eliminate the multi-branch drift ambiguity of the original model. It outputs a unique confidence drift ridge (optimal migration path) and probability kernel density surfaces at 3-hour intervals, providing unambiguous and quantifiable red tide migration prediction data for subsequent verification station deployment. Based on the output continuous and consistent environmental driving field, a Lagrange particle tracking framework is used to initialize a multi-scale virtual particle swarm. A fourth-order Runge-Kutta method is introduced into the coupled integral of the wind and flow fields to improve trajectory calculation accuracy, significantly enhancing the calculation accuracy of particle trajectories in complex flow fields. This results in a significant improvement in the accuracy of particle trajectory calculation in 7... A time-varying kernel density estimation model is constructed based on the spatial positions of all particles within a 2-hour prediction period. An adaptive bandwidth selection mechanism is introduced to perform differentiated smoothing for different drift stages. By identifying the global principal mode ridge in the density field, ambiguous paths of branch drift are automatically removed to ensure the extraction of the main drift trend. A graph cut algorithm is used to optimize the spatial continuity of the principal mode ridge. A single curve after topological repair is extracted as the unique confidence drift ridge. A probability kernel density surface is generated every 3 hours in the form of the prediction confidence interval envelope to ensure that the probability field satisfies the mass conservation constraint in the forward integration. The single optimal path and the probability density grid that satisfies the conservation constraint are output. It should be noted that when performing the 72-hour red tide migration prediction, based on the continuous and consistent environmental driving field after missing field correction, a multi-scale virtual particle swarm was initialized using a Lagrange particle tracking framework. The initial particle positions covered the outline of the water color anomaly area and its internal region, and the spatial density was set to no less than 10 particles per square kilometer to ensure statistical significance. In the coupled integration process of the wind field and the flow field, the fourth-order Runge-Kutta method was used for time-step integration, with the time step fixed at 600 seconds, controlling the truncation error to the fifth power of the step size, thus improving the calculation accuracy of particle trajectories in complex flow fields. During the integration process, each particle carried its position and time scale. The system identifies the initial jurisdiction and outputs a sequence of spatial coordinates hourly. The independent trajectories of all particles constitute the original drift path set for a 72-hour prediction period. After integrating the particle trajectories, a time-varying kernel density estimation model is constructed for the spatial positions of all particles within the 72-hour prediction period. A Gaussian kernel is selected as the kernel function. Density estimation is performed on the positions of all particles at each hour. An adaptive bandwidth selection mechanism is introduced, with an initial bandwidth of 5 kilometers. In the early stage of prediction (0-24 hours), the bandwidth is linearly reduced to 3 kilometers to maintain path compactness. In the later stage of prediction (48-72 hours), the bandwidth is expanded to 8 kilometers to accommodate diffusion uncertainty. To achieve differentiated smoothing for different drift stages, the system identifies global principal mode ridges in the density field—continuous spatial connections between density peak points—and automatically removes ambiguous paths caused by initial field uncertainties. The retained ridge segments must satisfy the condition that their density value is not less than 40% of the maximum density value at that moment, ensuring the extracted results represent the main drift trend. After extracting the principal mode ridges, a graph cut algorithm is further used to optimize their spatial continuity. The candidate point set on the ridge is constructed as a graph structure, where the weight of the edges between nodes is negatively correlated with the spatial distance and density gradient difference between the two points. By solving for the minimum cut set of the graph, isolated points or short branches with low connectivity are removed, resulting in… The single continuous curve after topology repair is the unique confidence drift ridge. This ridge represents the optimal path for red tide migration. A ridge point is output every 3 hours. Based on this ridge, a probability kernel density surface is generated in the form of a prediction confidence interval envelope. A two-dimensional Gaussian distribution is constructed with the ridge point as the center and the given standard deviation (default is 5 km) according to the local historical drift error statistics. The width of the positive and negative two standard deviations is intercepted along the direction perpendicular to the ridge as the confidence envelope. The particle positions within each 3-hour window are accumulated by forward integration so that the probability field at each time moment satisfies the normalization conservation constraint in terms of quality, forming a probability density grid that can be directly called for station deployment. The multi-objective constraint coding module is used to transform the problem of red tide verification station layout in the sea area into a multi-objective optimization model. It defines constraints including coverage probability contour lines (derived from probability kernel density surface), sensitive area distance and ocean current direction (affecting drift path), and performs Pareto coding on station coordinates and number to construct an initial population. This provides a computable optimization problem expression framework for subsequent genetic iterations, transforming station layout into a computable optimization problem. The Pareto optimization module for station locations uses the NSGA-II algorithm to perform non-dominated sorting and crowding comparison of station locations, iteratively evolves station combination, automatically balances targets including high probability area coverage, high hazard area response and ocean current direction matching, outputs a Pareto front set of stations, corresponding to the theoretically lowest verification rejection risk path, automatically balances multiple conflict targets, and outputs the station location scheme with the theoretically lowest rejection risk. The closed-loop review and feedback module is used to reverse-map the rejection records in the review stage of the verification report to the preceding task generation, path prediction and station optimization stages. By analyzing the reasons for rejection and automatically correcting the fragment merging rules, drift correction parameters and station optimization weights, a full-chain adaptive closed loop is formed, which gradually reduces the accumulation of hierarchical deviations and the process rejection rate.

[0021] Example 2, as Figure 1 , Figure 2 As shown, based on Embodiment 1, this invention provides a technical solution: the multi-objective constraint coding module specifically includes: extracting several coverage probability contour lines corresponding to a preset probability threshold from the probability kernel density surface at 3-hour intervals using a contour tracking algorithm; discretizing the area enclosed by the contour lines into candidate station layout grid units; and assigning an initial coverage probability weight to each unit to ensure that candidate stations are concentrated in areas prone to red tides, thereby improving the targeting of the layout; and constructing a multi-objective optimization function that includes coverage probability, sensitive area distance, and consistency with ocean current direction. The coverage probability objective is obtained by integrating the kernel density value of the station location in the probability kernel density surface, and the sensitive area distance objective is obtained by calculating the distance between the candidate station and adjacent sensitive areas. The shortest spatial distance (including bathing areas, aquaculture areas, nuclear power plant intakes, etc.) is obtained after inverse normalization. The closer the distance, the higher the priority. The ocean current direction consistency target is obtained by inverse normalization of the cosine value of the angle between the line connecting the stations and the current mainstream current direction. The station coordinates and number of stations in the grid cells of the candidate stations are jointly Pareto encoded. Each encoded bit corresponds to the selected state of a grid cell. An initial population containing multiple combinations of station numbers and distributions is generated. At the same time, at least one pair of control stations is forcibly included in the optimization constraints. One of them is located in the high value area of ​​the probability kernel density surface, and the other is located in the low value area or outside the boundary of the anomaly area to ensure the comparative monitoring capability between red tide areas and non-red tide areas. It should be noted that during the multi-objective constrained coding process, candidate areas for station deployment are extracted from the probability kernel density surface of the completed drift prediction. This probability kernel density surface is output at 3-hour intervals, with each time segment corresponding to a spatial raster covering the red tide migration path and confidence envelope. The raster values ​​represent the relative probability of red tide organisms appearing at that location. A contour tracking algorithm is used, with a probability threshold set to 0.6. The coverage probability contour lines corresponding to this threshold are extracted, and the area enclosed by the contour lines is the high-probability coverage area. This area is discretized into square grid cells with a spatial resolution of 10 meters. Each grid cell is considered a candidate station deployment point. The kernel density value at the center point of each grid cell is read as the initial coverage probability weight for that cell, and they are sorted from high to low. After completing the discretization and weight allocation of the candidate station grid cells, an objective function containing three optimization objectives is constructed. The first objective is to maximize the coverage probability, calculated as the integral sum of the kernel density values ​​of the selected station grid cells. The second objective is to minimize the sensitive area distance, calculated by comparing the candidate station with adjacent sensitive areas (including bathing areas, etc.). The shortest spatial distance (such as to aquaculture areas and nuclear power plant water intakes) is used, and inverse normalization is applied based on this distance to give higher priority to stations closer to sensitive areas. The third objective is to ensure consistency of ocean current direction. The current mainstream current direction is read, and the cosine value of the angle between the direction of the line connecting adjacent stations and the mainstream current direction is calculated. After inverse normalization, the smaller the value, the stronger the consistency. A set of grid cells is deployed for candidate stations, and joint Pareto coding is performed using binary encoding. Each coded bit corresponds to the selection status of a grid cell, with 1 indicating selection. The grid cell is used to deploy stations, with 0 indicating no stations. The number of stations is not preset to a fixed value, but is used as an optimization variable in the encoding. That is, the number of 1s in the encoding string is the number of stations corresponding to that individual. The initial population adopts a strategy of combining random generation and heuristic initialization to generate 200 individuals. 100 individuals are generated completely randomly, and the remaining 100 individuals are set to 1 in descending order of coverage probability weight. The number of stations is taken from 5 to 15, with several examples. This initial population covers different combinations of station numbers and spatial distributions. The Pareto optimization module specifically includes: using the constructed initial population as the starting point for optimization, calculating the fitness vector of each individual's position combination under the multi-objective optimization function; employing an improved fast non-dominated sorting algorithm to divide into multiple levels of Pareto fronts; and introducing a Pareto front thickness coefficient to identify boundary and internal solutions in the non-dominated solution set, ensuring that extreme position schemes are not eliminated and improving the diversity of the solution set. Within each non-dominated layer, calculating the crowding distance of each individual at each position, which simultaneously considers the spatial distribution density of positions and the neighborhood density in the objective function space, prioritizing the retention of individuals with high crowding. To maintain population diversity, offspring populations are generated by simulating binary crossover and polynomial mutation operations. An elite retention strategy is used to merge parent and offspring populations to prevent premature convergence of the algorithm and ensure balanced coverage of station space. Iterative evolution is repeated until the rate of change of the hypervolume index of the Pareto front is lower than the convergence threshold within a preset number of generations. The final Pareto front station set is output. Each station combination in this set is a non-dominated solution, corresponding to the station layout scheme with the lowest theoretical rejection risk, which can be directly called by the on-site verification task. The station layout scheme with the lowest rejection risk is automatically generated from the optimal solution set. It should be noted that the initial population constructed using the multi-objective constraint coding module serves as the starting point for optimization. This population contains 200 individuals, each represented by a binary encoded string indicating the selected state of candidate station grid cells. First, for all selected station cells within each individual, the fitness vectors under the three objectives of maximizing coverage probability, minimizing sensitive area distance, and consistency of ocean current direction are calculated. Subsequently, an improved fast non-dominated sorting algorithm is employed, dividing the front into multiple levels based on the Pareto dominance relationship between individuals. A Pareto front thickness coefficient is introduced during the sorting process, which is determined by statistically analyzing the number of neighboring solutions in the target space for each non-dominated solution. The number of enclosing layers is used to distinguish between boundary solutions and internal solutions, preserving boundary solutions with extreme objective values ​​to avoid optimization convergence to local regions. After sorting, each individual is assigned a non-dominated layer number. For individual stations within the same non-dominated layer, their crowding distance is further calculated. This distance consists of two parts: first, in the spatial geographic dimension, the average Euclidean distance between any two stations in the station combination is calculated, reflecting the spatial density of station distribution; second, in the three-dimensional objective function space, the sum of the differences between the individual and its immediate neighbors in each objective dimension is calculated, reflecting the diversity contribution of the solution. The two distances are then weighted and merged. After merging, individuals with higher crowding are prioritized for retention to maintain balanced coverage of the population in high-probability, high-risk, and flow-consistent areas. Then, a simulated binary crossover operation is performed on the retained parent population with a crossover probability of 0.9. Polynomial mutation is applied to the newborn individuals, with the mutation probability being the reciprocal of the decision variable dimension, generating a population of 200 offspring of equal size. An elite retention strategy is used to merge the parent and offspring generations, sorting them by non-dominant level and crowding. The top 200 individuals are selected as the new generation, ensuring no excellent solutions are lost. This iterative evolutionary process is repeated, and the hypervolume index of the current Pareto front is calculated at the end of each generation. The indicator measures convergence and diversity by the volume enclosed by all individuals in the front and the reference point in the three-dimensional target space. When the rate of change of the hypervolume indicator is less than 0.1% for 20 consecutive generations, the optimization is considered to have converged, the iteration is terminated, and the final Pareto front station set is output. This set contains multiple non-dominated station combinations. Each combination cannot be surpassed by other solutions at the same time in terms of coverage probability, hazard level, and flow direction consistency. It corresponds to the station layout scheme with the lowest theoretical verification rejection risk. The Pareto front is output in list form for direct use by on-site verification tasks. Users can choose any set of stations to deploy according to their actual operational capabilities. The closed-loop review and feedback module specifically includes: capturing rejection records during the review of verification reports, parsing rejection opinions through natural language processing, extracting rejection reason categories, including boundary attribution errors, predicted path deviations, insufficient station coverage, or omissions of high-hazard areas, achieving automated and accurate identification of rejection reasons, reducing the time-consuming manual judgment, and mapping the rejection reasons back to the fragment task integration module, drift ridge generation module, and multi-objective constraint encoding module, triggering adaptive correction mechanisms respectively: dynamically adjusting the distance attenuation coefficient in the membership weighting function for fragment merging rules, updating the process noise covariance matrix of the Kalman filter for drift correction parameters, and redistributing the weight vector of the three objective functions for station optimization weights using gradient descent, achieving point-to-point correction of the root cause of the problem, avoiding blind adjustment of global parameters, recording the rejection rate and adjustment parameter combination after each adjustment, constructing a rejection-correction mapping library, iteratively optimizing correction rules through reinforcement learning strategies, forming a closed-loop feedback archive, used to successively reduce the accumulation of hierarchical deviations and the process rejection rate, enabling the system to autonomously learn the optimal parameters during continuous operation, and the rejection rate to converge and decrease successively; It should be noted that during the review of the verification report, the system automatically captures the rejection records submitted by the reviewers, extracts the rejection text, and uses a natural language processing method based on a combination of domain dictionaries and rules to perform structured parsing of the rejection opinions. During the parsing process, the rejection opinions are first segmented into Chinese words and tagged with parts of speech. Key feature words are matched against the red tide verification business terminology database, and the reasons for rejection are categorized into four types: incorrect boundary attribution, deviation from predicted path, insufficient station coverage, or omission of high-hazard areas. Errors in boundary attribution are identified by the presence of words such as "jurisdictional boundary" and "unclear attribution" in the rejection opinions; deviation from predicted path is identified by expressions such as "drift direction" and "actual location does not match prediction"; insufficient station coverage is identified by words such as "omission" and "insufficient density"; and omission of high-hazard areas is identified by matching the algal species name with words such as "high-risk" and "toxin". After parsing, the system outputs the rejection reason classification labels and their confidence levels (if the confidence level is below 0.7, it is transferred to manual review), and the classification results are displayed. The correction trigger signal is transmitted to the corresponding module. Based on the rejection reason classification results, a reverse mapping operation is performed to map boundary assignment errors to the fragment task integration module, predicted path deviations to the drift ridge generation module, and insufficient station coverage or omission of high-hazard areas to the multi-objective constraint coding module. For boundary assignment errors, the attenuation coefficient of the S-shaped distance attenuation function in the fuzzy boundary membership weighting unit is dynamically adjusted. The original coefficient of 0.0015 is increased in increments of 0.0002 (enhancing boundary sharpness) or decreased (enhancing transition smoothness). After each adjustment, the continuous membership field is recalculated. For predicted path deviations, the process noise covariance diagonal matrix elements of the Kalman filter in the correction unit are updated. The original time smoothing coefficient value is adjusted with a 10% amplitude gradient to adjust the sensitivity of the filter to new observation data. For insufficient station coverage or omission of high-hazard areas, the gradient descent method is used to redistribute the weight vector of the three objective functions in the multi-objective constraint coding module, and the gradient step size of the coverage probability weight is set to 0.05. The hazard level weight and flow consistency weight are adjusted in the opposite direction accordingly, and the sum of the three weights is always kept at 1. After each rejection correction operation, the reason category of this rejection, the parameter combination before adjustment, the parameter combination after adjustment, and the change in the rejection rate of subsequent verification reports after this adjustment (using the most recent 10 reports as the statistical window) are recorded to form a rejection-correction mapping record. All records constitute a rejection-correction mapping library. On this basis, the Q-learning reinforcement learning strategy is used to iteratively optimize the correction rules: the current system parameter combination (including attenuation system) is adjusted accordingly. The state space is defined by the number of parameters, the diagonal elements of the Kalman filter process noise covariance matrix, and the three-objective weight vector. The action space is defined by the three types of parameter adjustment actions. The negative value of the rejection rate is used as the reward function. After each complete rejection-correction-re-report closed loop, the Q-value table is updated. The optimal parameter adjustment strategy for different rejection reason categories is gradually learned. As the sample size of the mapping library accumulates to more than 50, stability correction rules are output to gradually reduce the cumulative hierarchical deviation of subsequent tasks. The process rejection rate shows a convergent downward trend, forming a traceable and reusable closed-loop feedback archive.

[0022] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A fully intelligent management system for marine red tide verification tasks, comprising a visual control platform, characterized in that, The visualization control platform has the following communication connections: The fragment task integration module is used to integrate fuzzy C-means clustering and spatial topological constraint algorithms, automatically merge fragment tasks at the boundary and assign them to the responsible party, and eliminate hierarchical bias in the task generation process. The drift ridge generation module is used to address the 72-hour migration prediction branch drift phenomenon caused by grid-level invalid values, outliers, or spatiotemporal continuity abrupt changes in local wind and flow field data. It integrates particle swarm optimization and Kalman filtering algorithms, uses similar environmental field fragments in the historical red tide case library as reference particles, performs local dynamic filling and smoothing correction on abnormal grid points, eliminates multi-branch ambiguity, generates a single confidence drift ridge, and outputs a 3-hour interval probability kernel density surface. The multi-objective constraint coding module is used to transform the problem of red tide verification station layout in the sea area into a multi-objective optimization model, define constraints, perform Pareto coding on station coordinates and quantity, and construct the initial population. The Pareto optimization module for station locations uses the NSGA-II algorithm to perform non-dominated sorting and crowding comparison of station layouts, iteratively evolves station combinations, automatically balances targets including high-probability zone coverage, high-hazard zone response, and ocean current direction matching, and outputs a Pareto front set of stations, corresponding to the theoretically lowest verification rejection risk path. The closed-loop review feedback module is used to reverse-map the rejection records of the verification report review stage to the previous stage, forming a full-chain adaptive closed loop, and gradually reducing the accumulation of hierarchical deviations and the process rejection rate.

2. The intelligent management system for the entire process of red tide verification tasks in marine areas according to claim 1, characterized in that: The fragment task integration module includes a fuzzy boundary membership weighting unit and a mandatory attribution unit for the jurisdiction buffer. The fuzzy boundary membership weighting unit is used in the task of verifying red tides in the sea area to calculate the membership degree of the edge pixels of the remote sensing water color anomaly area to the adjacent jurisdiction area using fuzzy C-means clustering, and to output the fuzzy membership weight based on spatial distance and spectral characteristics to quantify the degree of boundary fuzziness. The mandatory attribution unit of the jurisdiction buffer generates a dynamic buffer based on the jurisdiction boundary, combines fuzzy attribution weights to merge fragmented tasks in the intersection area with topological constraints, and assigns tasks with ambiguous attribution to the corresponding responsible party according to the buffer overlap and historical task allocation rules, and outputs a verification task package with clear boundaries and determined responsibilities.

3. The intelligent management system for the entire process of red tide verification tasks in marine areas according to claim 2, characterized in that: The fuzzy boundary membership weighting unit specifically includes: For remote sensing water color anomaly areas found in red tide verification in the sea area, the spatial coordinates and multi-band spectral reflectance of edge pixels are extracted to construct a pixel-level feature matrix; The number of clusters is set to the total number of adjacent jurisdictions. The membership degree of each edge pixel to each jurisdiction is iteratively calculated. An S-shaped distance decay function is constructed based on the spatial distance from the pixel to the jurisdiction boundary to perform nonlinear weighted correction on the membership degree. The weighted membership degree is compared with the preset first threshold and second threshold: pixels above the first threshold are directly assigned to the corresponding jurisdiction, pixels below the second threshold are marked as forced merging seed points, and pixels in between are output as a continuous membership degree field, in which each pixel carries a fuzzy assignment weight to each jurisdiction.

4. The intelligent management system for the entire process of red tide verification tasks in marine areas according to claim 2, characterized in that: The mandatory attribution unit for the jurisdiction buffer specifically includes: Dynamic jurisdictional buffer zones are generated by buffering outwards at a preset distance along the administrative boundaries of each jurisdiction, and further buffered inwards to generate a core jurisdictional inner circle. The annular area between the two is defined as the attribution negotiation zone, and the intersection area between different buffer zones is identified as a potential attribution ambiguity area for fragmented tasks. Within the ambiguous region of attribution, a continuous membership field is superimposed, and a spatial topological adjacency graph is constructed for multiple fragmented tasks in the intersection region. Connectivity component analysis is used to merge the tasks and form a task package to be assigned. For each task package to be assigned, calculate its overlap ratio with each managed buffer zone, and introduce a mandatory unique assignment decision function: using the buffer overlap as the main factor and the default responsibility party weight at the boundary in the historical task allocation rule base as the secondary factor, a weighted voting mechanism is used to assign it to a unique primary responsibility party, while simultaneously pushing it to one or more secondary responsibility parties. The primary responsibility party must accept the task, while the secondary responsibility parties can choose the task, ensuring that there are no unclaimed areas at the boundary.

5. The intelligent management system for the entire process of red tide verification tasks in marine areas according to claim 2, characterized in that: The drift ridge generation module includes a correction unit and a single-confidence drift ridge generation unit; The correction unit is used to detect and mark grid-level invalid values, outliers, or spatiotemporal continuity abrupt changes in wind field and flow field data in the 72-hour migration prediction of red tide verification in the sea area. It retrieves similar environmental field fragments from the historical red tide case library as reference particles, uses particle swarm optimization algorithm to search for the optimal local correction parameters for the abnormal grid points, and uses Kalman filtering to smoothly replace the outliers, outputting a continuous and consistent environmental driving field after anomaly correction. The single-confidence drift ridge generation unit, based on the modified continuous environmental driving field, uses a particle tracking and probability density fusion method to eliminate multi-branch drift ambiguity and output a unique confidence drift ridge and a probability kernel density surface with a 3-hour interval.

6. The intelligent management system for the entire process of red tide verification tasks in marine areas according to claim 5, characterized in that: The correction unit specifically includes: Real-time monitoring of wind and flow field grid data; identification of local outliers in the time or space dimension based on the spatiotemporal continuity scoring function; and morphological dilation processing of the outlier grid region to determine the boundary of the outlier influence range. An environmental field feature hash index is constructed from the historical red tide case library. Several historical environmental fields most similar to the current abnormal area environmental features are retrieved as reference particles. A dynamic time warping algorithm is introduced to align environmental field sequences of different time phases. A particle swarm optimization algorithm is used to search for the optimal local correction parameters for the abnormal grid points. The inertial weight in the particle velocity update formula is designed as an adaptive function that is negatively correlated with the severity of the abnormality. The optimized local correction parameters are input into the Kalman filter predictor to dynamically correct and smooth the abnormal grid point values ​​at the current time, and output a continuous and consistent environmental driving field sequence after anomaly correction.

7. The intelligent management system for the entire process of red tide verification tasks in marine areas according to claim 5, characterized in that: The single-confidence drift ridge generation unit specifically includes: Based on the continuous and consistent environment driving field of the output, a multi-scale virtual particle swarm is initialized using the Lagrange particle tracking framework, and a fourth-order Runge-Kutta method is introduced into the coupled integral of the wind field and the flow field to improve the accuracy of trajectory calculation. A time-varying kernel density estimation model is constructed for the spatial positions of all particles within a 72-hour prediction period. An adaptive bandwidth selection mechanism is introduced to perform differential smoothing for different drift stages. Ambiguous branch drift paths are automatically pruned by identifying the global principal mode ridge in the density field. The spatial continuity of the principal mode ridge is optimized by using a graph cut algorithm. The single curve after topological repair is extracted as the unique confidence drift ridge, and a probability kernel density surface is generated every 3 hours in the form of a predicted confidence interval envelope.

8. The intelligent management system for the entire process of red tide verification tasks in marine areas according to claim 5, characterized in that: The multi-objective constraint encoding module specifically includes: From the probability kernel density surface at 3-hour intervals, the contour line tracking algorithm is used to extract several coverage probability contour lines corresponding to the preset probability threshold. The area enclosed by the contour lines is discretized into candidate station grid cells, and an initial coverage probability weight is assigned to each cell. A multi-objective optimization function is constructed, which includes coverage probability, sensitive area distance, and ocean current direction consistency. The coverage probability objective is obtained by integrating the kernel density value of the station location in the probability kernel density surface. The sensitive area distance objective is obtained by calculating the shortest spatial distance between the candidate station and the adjacent sensitive area and then performing inverse normalization. The ocean current direction consistency objective is obtained by calculating the cosine of the angle between the direction of the line connecting the stations and the current mainstream current direction and performing inverse normalization. The station coordinates and number of stations in the grid cells of the candidate stations are jointly Pareto encoded. Each encoded bit corresponds to the selected state of a grid cell, generating an initial population containing multiple combinations of station numbers and distributions. At the same time, at least one pair of control stations is forcibly included in the optimization constraints, one of which is located in the high value region of the probability kernel density surface, and the other is located in the low value region or outside the boundary of the anomaly region.

9. The intelligent management system for the entire process of red tide verification tasks in marine areas according to claim 8, characterized in that: The Pareto optimization module for station locations specifically includes: Using the constructed initial population as the starting point for optimization, the fitness vector of each individual under the multi-objective optimization function is calculated for the station combination. An improved fast non-dominated sorting algorithm is used to divide the multi-level Pareto front, and a Pareto front thickness coefficient is introduced to identify the boundary solutions and internal solutions in the non-dominated solution set. Within each non-dominated layer, the crowding distance of each station individual is calculated. This distance takes into account both the spatial distribution density of the station and the neighborhood density in the objective function space. Individuals with high crowding are preferentially retained to maintain population diversity. Offspring populations are generated by simulating binary crossover and polynomial mutation operations, and an elite retention strategy is adopted to merge parent and offspring populations. Repeatedly iterate and evolve until the rate of change of the hypervolume index of the Pareto front is lower than the convergence threshold within a continuous preset number of algebras, and output the final Pareto front station set. Each station combination in this set is a non-dominated solution, corresponding to the station layout scheme with the lowest theoretical risk of rejection.

10. The intelligent management system for the entire process of red tide verification tasks in marine areas according to claim 9, characterized in that: The closed-loop review feedback module specifically includes: During the review of the verification report, rejection records were captured, and the rejection opinions were analyzed through natural language processing to extract the categories of rejection reasons, including incorrect boundary attribution, deviation of predicted path, insufficient station coverage, or omission of high-hazard areas. The reasons for rejection are reverse-mapped to the fragment task integration module, the drift ridge generation module, and the multi-objective constraint encoding module, triggering adaptive correction mechanisms respectively. Record the rejection rate and adjustment parameter combination after each adjustment, build a rejection-correction mapping library, and iteratively optimize the correction rules through reinforcement learning strategy to form a closed-loop feedback archive, which is used to reduce the accumulation of hierarchical deviations and the process rejection rate step by step.

Images