A method, device, equipment and medium for dynamically predicting a near-shore microplastic convergence ridge

Abstract

This application discloses a method, apparatus, device, and medium for dynamic prediction of nearshore microplastic aggregation ridges, relating to the field of data processing technology. The method includes: determining a mass point library; reading the mainstream orientation of each mass point in the library; assembling each mass point as a node to obtain a two-layer heterogeneous graph, and writing the shear rate and diffusion weight calculated based on the mainstream orientation into the edge attributes of the upper and lower heterogeneous graphs respectively; splitting the two heterogeneous graphs into a directed convergent adjacency matrix and an undirected diffusion adjacency matrix respectively, and then extracting node representations; selecting ridge seed based on the node representations; calculating the optimal ridge based on the ridge seed; constructing driving vectors for the optimal ridges from multiple historical frames, and generating a predicted ridge sequence with a preset future duration based on the driving vectors and the geometric coordinates of the ridge points in the optimal ridge. This application ensures that each mass point simultaneously possesses continuous dynamic and material properties at any given time, and improves the prediction accuracy of microplastic aggregation based on accurate centroid alignment analysis.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, apparatus, equipment and medium for dynamic prediction of nearshore microplastic convergence ridges. Background Technology

[0002] Existing monitoring systems for microplastics and floating pollutants generally use regional tidal-wave coupled models to provide Eulerian velocity fields. Real-time observations from HF radar, GPS drift rafts, and buoy profiles are then used to perform optimal interpolation or ensemble Kalman assimilation of flow velocity and water level. Virtual particles are then released in a 3D grid, and forward integration along the model step size is performed to obtain the Lagrange distribution of concentration over time. Particle concentration is often mapped to the particle cell after inverting optical substitutions such as suspended sediment and chlorophyll using satellite wide-swath imagery or airborne multispectral imaging, thus fusing the dynamic and material fields within a unified spatiotemporal coordinate system. To characterize nearshore converging fronts, existing methods often calculate instantaneous velocity divergence, strain rate, or the Oille-Lyuvillen index at Eulerian points, combining this with multi-scale Canny or Sobel gradient operations to extract convergence bands, and then using morphological connectivity analysis to obtain slender ridges. Some studies use improved Floyd-Warshall or topographic profile search algorithms to simultaneously constrain ridge extension in both tangential and normal directions to ensure geometric continuity. Meanwhile, the large-scale background field simulated by coarse-resolution drift and the local details of high-resolution remote sensing are stitched together by bilinear or spline interpolation, thus balancing computational efficiency and near-shore spatial accuracy.

[0003] While existing methods can assimilate model flow fields and multi-source observations to a unified coordinate system, velocity interpolation and Lagrange particle integration in the Eulerian grid often belong to different time steps. This can easily lead to phase misalignment during high-frequency tidal oscillations, thus weakening the resolution of nearshore narrow-band fronts. Convergence band extraction steps mostly involve directly thresholding divergence or gradients and relying on morphological connectivity to maintain ridge continuity. This purely geometric approach struggles to maintain a sharp front over long periods under alternating tidal peaks and wind surges, often resulting in excessively wide ridge drift or smoothing by background diffusion. Graph-based deep models typically set adjacency relationships as static, undirected weights. Their message passing does not distinguish between mainstream traction and lateral diffusion; shear intensity and concentration increments are homogenized within a single convolution, resulting in the inability to simultaneously represent the directionality of the convergence axis and the randomness of material diffusion. Cyclic networks used for short-term extrapolation focus on the statistical correlation of time series, lack explicit characterization of astronomical tidal harmonics and sudden storm events, and uncertainty assessment relies on a large number of Monte Carlo perturbations. Although field sampling can reduce errors, the information feedback cycle is long and cannot converge to meter-level spatial accuracy in a timely manner. Summary of the Invention

[0004] The main objective of this application is to propose a method, apparatus, equipment, and medium for dynamic prediction of nearshore microplastic aggregation ridges, so as to improve the prediction accuracy of microplastic aggregation.

[0005] To achieve the above objectives, one aspect of this application proposes a method for dynamic prediction of nearshore microplastic convergence ridges, the method comprising the following steps:

[0006] Identify multiple mass points that are evenly distributed in the tidal reach area;

[0007] Predict the initial trajectory for each of the said particles and interpolate new said particles according to the initial trajectory to obtain a particle library;

[0008] The instantaneous flow velocity of each particle in the particle library at the Eulerian point is read, and the main flow direction of each particle is obtained.

[0009] The mass points are assembled as nodes to obtain two heterogeneous graphs, and the shear rate and diffusion weight calculated according to the mainstream direction are written into the edge attributes of the upper heterogeneous graph and the edge attributes of the lower heterogeneous graph, respectively.

[0010] The two heterogeneous graphs are respectively split into a directed convergent adjacency matrix and an undirected diffusing adjacency matrix;

[0011] The node representation is extracted based on the directed convergent adjacency matrix and the undirected diffusing adjacency matrix;

[0012] Based on the node characterization, a shear stability score is calculated, and nodes with shear stability scores greater than a preset score threshold are selected as ridge seeds.

[0013] The optimal ridge line is calculated based on the ridge line seed;

[0014] A driving vector is constructed for the optimal ridge line in multiple historical frames, and a predicted ridge line sequence with a preset future duration is generated based on the driving vector and the geometric coordinates of the ridge points in the optimal ridge line.

[0015] In some embodiments, predicting the initial trajectory for each of the particles includes the following steps:

[0016] The initial trajectory is obtained by incorporating the three-dimensional flow field of the tidal numerical model onto each of the particles.

[0017] The process of interpolating new mass points based on the initial trajectory to obtain a mass point library includes the following steps:

[0018] Cubic spline interpolation was performed on the model flow velocity and water level on the mass time axis, and the flow velocity from shore-based high-frequency radar, the spectral density of airborne imaging, and the profile density of autonomous underwater vehicles were simultaneously superimposed as the observation mapping.

[0019] By tracking back and forth, using spatial proximity and temporal distance weighting, the observations are mapped to corresponding particle points. After removing anomalies in wind speed, effective wave height, and turbidity, third-order B-spline interpolation is performed along the historical direction of the particle points to obtain a spatiotemporally continuous dynamic and concentration particle library.

[0020] In some embodiments, assembling each of the mass points as nodes to obtain two heterogeneous graphs, and writing the shear rate and diffusion weight calculated according to the mainstream direction into the edge attributes of the upper heterogeneous graph and the edge attributes of the lower heterogeneous graph respectively, includes the following steps:

[0021] The shear strength along the main direction is calculated for each of the mass points using the Lagrange reference grid to extract the shear gradient band as the shear rate;

[0022] Perform temporal consistency checks over two consecutive frames to determine the convergence core node;

[0023] The directed convergence edge is recursively written according to the mainstream direction, and the diffusion weight is calculated and the undirected diffusion edge is selected based on the lateral concentration gradient.

[0024] Assemble the two heterogeneous graphs under the condition that the node indexes are consistent, and write the shear rate and the diffusion weight into the edge attributes of the upper heterogeneous graph and the lower heterogeneous graph respectively.

[0025] In some embodiments, the node representation extracted based on the directed convergent adjacency matrix and the undirected diffusing adjacency matrix includes the following steps:

[0026] Read the node flow velocity modulus, local shear rate and particle concentration, and then splice them into physical features;

[0027] Shear rate-weighted directed convolution is performed on the directed convergent adjacency matrix, and diffusion weight mask transverse convolution is performed on the undirected diffuse adjacency matrix. Then, the gating coefficients are obtained by Sigmoid mapping based on the wind and wave potential energy index and the local relative vorticity.

[0028] The embedding features of the corresponding channels are extracted respectively;

[0029] The embedded features of the two channels are adaptively fused, then added to the residual of the physical features, and then subjected to layer normalization to obtain the node representation.

[0030] In some embodiments, calculating the optimal ridge line based on the ridge line seed includes the following steps:

[0031] Along the main mainstream direction, under the condition that the shear strength is not lower than the preset shear threshold and does not touch the shoreline, a single-scale main ridge is generated iteratively according to the ridge seed. Then, multi-scale ridges are generated in parallel by equal displacement in the normal direction under the condition that the shear strength is not lower than the preset shear threshold and does not touch the shoreline.

[0032] For the single-scale main ridge and each of the multi-scale ridges, the average shear, relative vorticity and wave field energy of each segment are calculated according to a fixed arc length, and the segment attributes are constituted.

[0033] After truncating the ridges according to the set attribute threshold, calculate the comprehensive dynamic score of the remaining ridges and retain the segmented ridges that are greater than the preset dynamic score threshold;

[0034] Based on the segmented ridge line, and through first-order flow direction regression and third-order B-spline global smoothing, the optimal ridge line with geometric continuity is output.

[0035] In some embodiments, constructing the optimal ridge driving vector for multiple historical frames includes the following steps:

[0036] The optimal ridge lines from multiple historical frames are sorted in ascending order by time and resampled to a unified grid, thereby constructing the driving vector that includes tidal phase, shore-directed or coastal current velocity, wind direction and wind speed pair, and wave field energy.

[0037] The step of generating a predicted ridgeline sequence with a preset future duration based on the driving vector and the geometric coordinates of the ridge points in the optimal ridgeline includes the following steps:

[0038] The driving vector and the geometric coordinates of the ridge point in the optimal ridge line are combined and input into a recursive network with a tidal harmonic memory gate. The main tidal harmonic reference is explicitly introduced in the update gate and a migration gate is set for sudden changes in wind direction. The ridge point displacement increment is output frame by frame to generate the predicted ridge line sequence with a preset future duration.

[0039] Orthogonal perturbations are applied to future wind fields and tidal amplitudes and extrapolated repeatedly. Spatial confidence radii are calculated based on sample standard deviations, and low-confidence ridges are marked.

[0040] In some embodiments, the method further includes the following steps:

[0041] Scan consecutive ridge segments in the predicted ridge sequence whose confidence radius exceeds a preset radius threshold;

[0042] The system retrieves the location and remaining range of near-shore autonomous surface or underwater platforms, selects the execution platform based on a comprehensive cost function of distance, energy consumption, and sea state, and plans the trajectory along the continuous ridge segment and the cross-sections on both sides.

[0043] The sampling command is issued to drive the execution platform to collect raw flow velocity and concentration profiles along the trajectory;

[0044] The original flow rate and the concentration profile are inserted into the corresponding particles through the Lagrange mapping interface, and the particle library is updated.

[0045] The process involves reading the instantaneous flow velocity of each mass point in the mass point library at Eulerian points to obtain the mainstream direction of each mass point, until the confidence radius of all ridge segments in the predicted ridge sequence is lower than the preset radius threshold, and then the final predicted ridge sequence is output.

[0046] To achieve the above objectives, another aspect of this application proposes a nearshore microplastic convergence ridge dynamic prediction device, the device comprising:

[0047] A particle determination unit is used to determine multiple particles that are uniformly distributed in the tidal reachable area;

[0048] A particle library construction unit is used to predict the initial trajectory of each particle and interpolate new particles according to the initial trajectory to obtain a particle library.

[0049] The mainstream direction reading unit is used to read the instantaneous flow velocity of each particle in the particle library at the Eulerian point, and then obtain the mainstream direction of each particle;

[0050] The heterogeneous graph construction unit is used to assemble each of the mass points as nodes to obtain two-layer heterogeneous graphs, and write the shear rate and diffusion weight calculated according to the mainstream direction into the edge attributes of the upper heterogeneous graph and the edge attributes of the lower heterogeneous graph, respectively.

[0051] The matrix splitting unit is used to split the two layers of the heterogeneous graph into a directed convergent adjacency matrix and an undirected diffusing adjacency matrix, respectively.

[0052] The node representation extraction unit is used to extract node representations based on the directed convergent adjacency matrix and the undirected diffusing adjacency matrix.

[0053] The ridge seed selection unit is used to calculate the shear stability score based on the node characterization and select the nodes whose shear stability scores are greater than a preset score threshold as ridge seeds.

[0054] An optimal ridgeline calculation unit is used to calculate the optimal ridgeline based on the ridgeline seed.

[0055] The ridge prediction unit is used to construct driving vectors for the optimal ridges in multiple historical frames, and generate a predicted ridge sequence with a preset future duration based on the driving vectors and the geometric coordinates of the ridge points in the optimal ridges.

[0056] To achieve the above objectives, another aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0057] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0058] The embodiments of this application include at least the following beneficial effects:

[0059] This application provides a method, apparatus, device, and medium for dynamic prediction of nearshore microplastic convergence ridges. The scheme involves: identifying multiple mass points uniformly distributed in the tidal reachable area; predicting the initial trajectory of each mass point and interpolating new mass points based on the initial trajectories to obtain a mass point library; reading the instantaneous flow velocity of each mass point at Eulerian points in the mass point library to obtain the mainstream direction of each mass point; assembling each mass point as a node to obtain a two-layer heterogeneous graph; and writing the shear rate and diffusion weight calculated based on the mainstream direction into the edge attributes of the upper heterogeneous graph and the lower heterogeneous graph, respectively. The method involves analyzing the edge properties of heterogeneous graphs, splitting the two-layer heterogeneous graph into directed convergent adjacency matrices and undirected diffuse adjacency matrices, extracting node representations from these matrices, calculating shear stability scores based on these node representations, and selecting nodes with shear stability scores greater than a preset threshold as ridge seeds. The optimal ridge is then calculated based on these ridge seeds. Driving vectors are constructed for the optimal ridges across multiple historical frames, and a predicted ridge sequence for a preset future duration is generated based on the driving vectors and the geometric coordinates of the ridge points in the optimal ridge. This approach ensures that each particle possesses continuous dynamic and material properties at any given time, enabling subsequent shear analysis and graph analysis to be conducted in a high-frequency tidal context without phase drift, thus improving alignment performance. Accurate centroid analysis based on alignment can further enhance the prediction accuracy of microplastic aggregation. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 Comparison diagrams between some embodiments of this application and the prior art are provided for illustration purposes;

[0062] Figure 2A flowchart illustrating a method for predicting the dynamics of nearshore microplastic convergence ridges provided in this application embodiment;

[0063] Figure 3 An example flowchart of a nearshore microplastic convergence ridge dynamic prediction method based on a shear-diffusion dual-flow graph neural network provided in this application embodiment;

[0064] Figure 4 A schematic diagram of the process for constructing a microplastic aggregation heterogeneity map is provided for an embodiment of this application;

[0065] Figure 5 A schematic diagram of a dual-channel stream-diffusion GNN encoding process provided for an embodiment of this application;

[0066] Figure 6 This application provides a schematic diagram of a convergence ridge generation and optimization process.

[0067] Figure 7 A schematic diagram of a nearshore microplastic convergence ridge dynamic prediction device provided in this application embodiment;

[0068] Figure 8 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0070] It is understood that the terms “first,” “second,” etc., used in this application may be used herein to describe various concepts, but unless otherwise stated, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of the embodiments of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the words “if,” “when,” or “in response to a determination” as used herein may be interpreted as “when…” or “when…” or “in response to a determination.”

[0071] As used in this application, the terms "at least one", "multiple", "each", "any", etc., "at least one" includes one, two or more, "multiple" includes two or more, "each" refers to each of the corresponding multiples, and "any" refers to any one of the multiples.

[0072] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0073] Before providing a detailed description of the embodiments of this application, some related technologies involved in the embodiments of this application will be described first, as follows:

[0074] Two-stream graph neural networks (BLNNs) are graph neural networks that simultaneously pass messages along two heterogeneous topologies: a convergence (shearing) channel and a diffusion channel. The convergence layer captures the mainstream traction effect, while the diffusion layer simulates lateral spread. The two feature streams are gated and fused to output physically interpretable node embeddings.

[0075] Converging ridges: These are elongated, high-concentration banded structures formed along hydrodynamic convergence axes by microplastics on the ocean surface under tidal and wave action. The ridge line is composed of a series of shear-stable, dynamically continuous, and geometrically interconnected nodes, serving as the core target for localization and extrapolation.

[0076] Tidal harmonic memory gate: Introducing a main half-day / day sinusoidal tidal frequency embedding term into the update gate of a recursive neural unit (such as GRU) enhances the memory ability of tidal periodic drift; and in conjunction with a wind direction migration gate, it can quickly respond to sudden wind and wave events, achieving dual coupling of tide and wind drive.

[0077] Uncertainty-driven sampling: Based on historical data and the confidence radius assessment of extrapolated ridges, incremental profile sampling is automatically assigned to the field platform for low-confidence segments that exceed the threshold; new observation data is assimilated by Lagrange mapping and then iteratively updated to predict and confidence assessment, forming a closed-loop optimization.

[0078] At the data-driven level, convolutional neural networks have been used to segment floating object strips end-to-end from continuous optical images. Graph-based approaches typically treat pattern grid nodes as vertices, establish undirected edges with fixed radii or K-nearest neighbors, and then perform spectral convolution, GAT, or simplified random walks to extract joint flow-concentration features. Most of these methods employ a single-layer homogeneous graph and characterize adjacency relationships with static weights, focusing on reproducing Eulerian gradients in the graph structure, but do not explicitly model the mechanism differences between shear convergence and lateral diffusion.

[0079] In the short-term path extrapolation phase, recurrent neural networks, LSTM, and even variational Seq2Seq structures are used to learn the nonlinear mapping between multi-source time series such as tide level, wind speed, and wave energy and ridge displacement. Uncertainties in future scenarios are typically addressed by applying Monte Carlo perturbations to the atmosphere and boundary conditions, creating multiple drift sets, and then estimating the spatial error ellipse using sample variance. For sea areas with large errors, existing operational platforms often utilize multi-objective planning based on ranging, energy consumption, and sea state to generate profiles or transect survey lines for AUVs or USVs. The value of incremental observations is evaluated using information entropy or the expected reduction in mean square error, thereby achieving a closed-loop prediction-sampling-repreneurial workflow.

[0080] To address the aforementioned shortcomings, this application aims to construct a physical-data coupled representation that can simultaneously preserve the sharpness of the shear front and the lateral diffusion morphology, and to complete the real-time alignment and fusion of multi-source observations in a high-frequency tidal-wave background, thereby generating short-term drift prediction results of microplastic convergence ridges with clear physical stratification and quantitative uncertainty, and further providing spatiotemporal guidance for adaptive encrypted sampling. Figure 1 The diagram shows some embodiments of this application compared with the prior art.

[0081] This application provides a method, apparatus, device, and medium for dynamic prediction of nearshore microplastic convergence ridges, relating to the field of data processing technology. The method, apparatus, device, and medium provided in this application can be applied to a terminal, a server, or software running on a terminal or server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, or vehicle terminal, but is not limited thereto; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network; the software can be an application implementing the method, apparatus, device, and medium for dynamic prediction of nearshore microplastic convergence ridges, but is not limited to the above forms.

[0082] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0083] Reference Figure 2 This application provides a method for dynamic prediction of nearshore microplastic convergence ridges. This method may include, but is not limited to, steps S200 to S280, as detailed below:

[0084] S200: Determines multiple mass points that are uniformly distributed in the tidal reach area;

[0085] S210: Predict the initial trajectory for each of the said particles and interpolate new said particles according to the initial trajectory to obtain a particle library;

[0086] S220: Read the instantaneous flow velocity of each particle in the particle library at the Eulerian point, and then obtain the main flow direction of each particle;

[0087] S230: Assemble each of the mass points as nodes to obtain two heterogeneous graphs, and write the shear rate and diffusion weight calculated according to the mainstream direction into the edge attributes of the upper heterogeneous graph and the edge attributes of the lower heterogeneous graph respectively.

[0088] S240: Decompose the two heterogeneous graphs into a directed convergent adjacency matrix and an undirected diffusing adjacency matrix, respectively;

[0089] S250: Node representations extracted based on the directed convergent adjacency matrix and the undirected diffusing adjacency matrix;

[0090] S260: Calculate the shear stability score based on the node characterization and select the nodes whose shear stability scores are greater than a preset score threshold as ridge seeds;

[0091] S270: The optimal ridge line is calculated based on the ridge line seed;

[0092] S280: Construct driving vectors for the optimal ridges in multiple historical frames, and generate a predicted ridge sequence with a preset future duration based on the driving vectors and the geometric coordinates of the ridge points in the optimal ridges.

[0093] Optionally, predicting the initial trajectory for each of the said particles includes the following steps:

[0094] The initial trajectory is obtained by incorporating the three-dimensional flow field of the tidal numerical model onto each of the particles.

[0095] The process of interpolating new mass points based on the initial trajectory to obtain a mass point library includes the following steps:

[0096] Cubic spline interpolation was performed on the model flow velocity and water level on the mass time axis, and the flow velocity from shore-based high-frequency radar, the spectral density of airborne imaging, and the profile density of autonomous underwater vehicles were simultaneously superimposed as the observation mapping.

[0097] By tracking back and forth, using spatial proximity and temporal distance weighting, the observations are mapped to corresponding particle points. After removing anomalies in wind speed, effective wave height, and turbidity, third-order B-spline interpolation is performed along the historical direction of the particle points to obtain a spatiotemporally continuous dynamic and concentration particle library.

[0098] Optionally, the step of assembling each of the mass points as nodes to obtain two-layer heterogeneous graphs, and writing the shear rate and diffusion weight calculated according to the mainstream direction into the edge attributes of the upper-layer heterogeneous graph and the edge attributes of the lower-layer heterogeneous graph respectively, includes the following steps:

[0099] The shear strength along the main direction is calculated for each of the mass points using the Lagrange reference grid to extract the shear gradient band as the shear rate;

[0100] Perform temporal consistency checks over two consecutive frames to determine the convergence core node;

[0101] The directed convergence edge is recursively written according to the mainstream direction, and the diffusion weight is calculated and the undirected diffusion edge is selected based on the lateral concentration gradient.

[0102] Assemble the two heterogeneous graphs under the condition that the node indexes are consistent, and write the shear rate and the diffusion weight into the edge attributes of the upper heterogeneous graph and the lower heterogeneous graph respectively.

[0103] Optionally, the node representation extracted based on the directed convergent adjacency matrix and the undirected diffusing adjacency matrix includes the following steps:

[0104] Read the node flow velocity modulus, local shear rate and particle concentration, and then splice them into physical features;

[0105] Shear rate-weighted directed convolution is performed on the directed convergent adjacency matrix, and diffusion weight mask transverse convolution is performed on the undirected diffuse adjacency matrix. Then, the gating coefficients are obtained by Sigmoid mapping based on the wind and wave potential energy index and the local relative vorticity.

[0106] The embedding features of the corresponding channels are extracted respectively;

[0107] The embedded features of the two channels are adaptively fused, then added to the residual of the physical features, and then subjected to layer normalization to obtain the node representation.

[0108] Optionally, calculating the optimal ridge line based on the ridge line seed includes the following steps:

[0109] Along the main mainstream direction, under the condition that the shear strength is not lower than the preset shear threshold and does not touch the shoreline, a single-scale main ridge is generated iteratively according to the ridge seed. Then, multi-scale ridges are generated in parallel by equal displacement in the normal direction under the condition that the shear strength is not lower than the preset shear threshold and does not touch the shoreline.

[0110] For the single-scale main ridge and each of the multi-scale ridges, the average shear, relative vorticity and wave field energy of each segment are calculated according to a fixed arc length, and the segment attributes are constituted.

[0111] After truncating the ridges according to the set attribute threshold, calculate the comprehensive dynamic score of the remaining ridges and retain the segmented ridges that are greater than the preset dynamic score threshold;

[0112] Based on the segmented ridge line, and through first-order flow direction regression and third-order B-spline global smoothing, the optimal ridge line with geometric continuity is output.

[0113] Optionally, constructing the optimal ridge driving vector for multiple historical frames includes the following steps:

[0114] The optimal ridge lines from multiple historical frames are sorted in ascending order by time and resampled to a unified grid, thereby constructing the driving vector that includes tidal phase, shore-directed or coastal current velocity, wind direction and wind speed pair, and wave field energy.

[0115] The step of generating a predicted ridgeline sequence with a preset future duration based on the driving vector and the geometric coordinates of the ridge points in the optimal ridgeline includes the following steps:

[0116] The driving vector and the geometric coordinates of the ridge point in the optimal ridge line are combined and input into a recursive network with a tidal harmonic memory gate. The main tidal harmonic reference is explicitly introduced in the update gate and a migration gate is set for sudden changes in wind direction. The ridge point displacement increment is output frame by frame to generate the predicted ridge line sequence with a preset future duration.

[0117] Orthogonal perturbations are applied to future wind fields and tidal amplitudes and extrapolated repeatedly. Spatial confidence radii are calculated based on sample standard deviations, and low-confidence ridges are marked.

[0118] Optionally, the method further includes the following steps:

[0119] Scan consecutive ridge segments in the predicted ridge sequence whose confidence radius exceeds a preset radius threshold;

[0120] The system retrieves the location and remaining range of near-shore autonomous surface or underwater platforms, selects the execution platform based on a comprehensive cost function of distance, energy consumption, and sea state, and plans the trajectory along the continuous ridge segment and the cross-sections on both sides.

[0121] The sampling command is issued to drive the execution platform to collect raw flow velocity and concentration profiles along the trajectory;

[0122] The original flow rate and the concentration profile are inserted into the corresponding particles through the Lagrange mapping interface, and the particle library is updated.

[0123] The process involves reading the instantaneous flow velocity of each mass point in the mass point library at Eulerian points to obtain the mainstream direction of each mass point, until the confidence radius of all ridge segments in the predicted ridge sequence is lower than the preset radius threshold, and then the final predicted ridge sequence is output.

[0124] The following sections will provide a detailed description and explanation of some optional embodiments of this application, using specific application examples.

[0125] The application scenarios of this embodiment may include:

[0126] Ocean Multi-Source Data Fusion and Lagrange Tracking Platform: This platform integrates numerical tidal models, HF radar, airborne telemetry, and AUV observation data, providing visualized Lagrange mass placement, three-dimensional flow field interpolation, and trajectory backtracking functions to help users quickly generate high-precision dynamic-material spatiotemporal fields.

[0127] Dual-flow graph neural network prediction engine: Designed specifically for heterogeneous graphs (convergence layer + diffusion layer), it embeds directional shearing convolution, lateral diffusion convolution and wind-wave-vortex gating modules, enabling efficient training and online inference on GPUs, and achieving real-time localization and short-term drift prediction of microplastic converging ridges.

[0128] Uncertainty-driven field sampling scheduling system: Integrates parameters such as available energy, remaining flight time, and sea state and waves of the platform, automatically plans sampling tracks and profiles for AUV / ASV / USV based on confidence radius, supports task issuance, real-time monitoring and data feedback, and forms an uncertainty closed-loop iterative sampling system.

[0129] Reference Figure 3This embodiment provides a method for dynamic prediction of nearshore microplastic convergence ridges based on a shear-diffusion dual-flow graph neural network, including the following steps S1~S6:

[0130] S1. Under the constraints of the shoreline digital elevation model and the intertidal zone boundary, mass points are uniformly distributed in the tidal reachable area. The three-dimensional flow field of the tidal numerical model is called to perform forward integration on each mass point to obtain the initial trajectory. Cubic spline interpolation is performed on the model flow velocity and water level on the mass point time axis. The shore-based high-frequency radar flow velocity, airborne imaging spectral concentration and autonomous underwater vehicle profile concentration are simultaneously superimposed. The observations are mapped to the corresponding mass points through forward and backward tracking, spatial proximity and temporal distance weighting. After removing anomalies in wind speed, effective wave height and turbidity, third-order B-spline interpolation is performed along the historical direction of the mass points to obtain a spatiotemporally continuous dynamic-concentration mass point library.

[0131] S2. Read the instantaneous flow velocity of the mass point library on the Eulerian grid points, calculate the shear intensity along the mainstream direction for each grid point to extract the shear gradient band (shear rate), and then perform a temporal consistency check in two consecutive frames to determine the convergence core node; further, recursively write directed convergence edges according to the mainstream direction, calculate the diffusion weights according to the lateral concentration gradient and filter undirected diffusion edges, assemble two heterogeneous graphs under the condition of consistent node indices, and write the shear rate and diffusion weights into the edge attributes of the upper and lower layers respectively. For example, Figure 4 A schematic diagram of the process for constructing a microplastic aggregation heterogeneous diagram.

[0132] S3. The heterogeneous graph is split into a directed convergent adjacency matrix and an undirected diffuse adjacency matrix. The node velocity modulus, local shear rate, and particle concentration are concatenated to form physical features. A shear rate-weighted directed convolution is performed on the convergent layer, and a diffusion-weighted transverse convolution is performed on the diffuse layer. Then, gating coefficients are obtained by Sigmoid mapping based on the wind and wave potential energy index and local relative vorticity. The two-channel embeddings are adaptively fused and added to the input embedding residuals, along with layer normalization, to obtain the final node representation. For example, Figure 5 This is a schematic diagram of a dual-channel stream-diffusion GNN encoding process.

[0133] S4. Based on node characterization, calculate the shear stability score and select high-scoring nodes as ridge seed nodes. Iteratively generate a single-scale main ridge along the mainstream direction, provided the shear intensity is not lower than the threshold and the ridge does not touch the shoreline. Then, generate multi-scale ridges in parallel by arithmetic shift along the normal direction and applying the same iteration rule. For each ridge, calculate the segment-average shear, relative vorticity, and wave field energy in segments with a fixed arc length. After truncation according to a set threshold, calculate the comprehensive dynamic score of the remaining ridges and retain high-scoring ridges. Finally, output the geometrically continuous optimal ridge through first-order flow direction regression and third-order B-spline overall smoothing. For example, Figure 6 This is a schematic diagram of a process for generating and optimizing converged ridges.

[0134] S5. The optimal ridge lines from multiple historical frames are sorted in ascending order by time and resampled to a unified grid. A driving vector containing tidal phase, shore / coastal current velocity, wind direction-speed pair, and wave field energy is constructed. This vector is then merged with the geometric coordinates of the ridge points and input into a recursive network with a tidal harmonic memory gate. The main tidal harmonic reference is explicitly introduced into the update gate, and a migration gate is set for sudden changes in wind direction. The ridge point displacement increment is output frame by frame to generate a predicted ridge line sequence for a preset future duration. Orthogonal perturbations are applied to the future wind field and tidal amplitude and extrapolated repeatedly. The spatial confidence radius is calculated based on the sample standard deviation, and low-confidence ridge segments are marked.

[0135] S6. Scan the continuous ridge segments in the predicted ridge sequence whose confidence radius exceeds the threshold, retrieve the location and remaining endurance of near-shore autonomous surface or underwater platforms, select the execution platform based on the comprehensive cost function of distance, energy consumption and sea state, and plan the trajectory along the ridge and the cross profiles on both sides. After issuing the sampling command, receive the raw velocity and concentration profiles in real time. Insert the back-transmitted observations into the corresponding mass points through the Lagrange mapping interface of step S1 and update the mass point library. Then, re-execute steps S3 to S5 to form a short closed loop until the confidence radius of all ridge segments decreases to the target threshold and the final prediction result is output.

[0136] More specifically, the above steps may include the following steps:

[0137] Step S1 includes the following steps:

[0138] S11. Establish a Lagrange reference frame based on the shoreline DEM and the intertidal zone boundary, at the initial time... Evenly distributed within the tidal range A point mass Then, the three-dimensional flow field of the tidal numerical model is invoked, and the integral is applied to each mass point up to the future time. This yields the set of initial trajectories for each particle. .

[0139] S12, Numerical power flow synchronization, on the particle time axis Above, cubic spline interpolation is performed on the flow velocity and water level of the tidal model to ensure that for each particle... All of these can yield Lagrange numerical dynamic pairs that are completely consistent with the measured time series. The time synchronization error is limited to the single-step integration error. Within this range, ensure that there is no timescale drift when the numerical background field and the measured flow velocity are subsequently fused.

[0140] S13, Lagrange mapping of current velocity from shore-based radar, calling up the real-time Euler current velocity from shore-based HF radar. Perform forward and backward tracking on each grid point and place it in... displacement path before and after Compare and select the index of the mass point with the smallest spatial distance. Write the corresponding mass entry to achieve fine shear correction in the nearshore shallow water area; if the mass lacks a corresponding radar grid point, the numerical model velocity is retained.

[0141] S14, Airborne telemetry concentration Lagrange interpolation, for airborne imaging spectrum at time... Obtained cell center coordinates and its concentration Perform countercurrent drift compensation: based on the inversion of pixels in the tidal field displacement vector within Calculate the position after drift and compare it with the mass capture radius. Comparison; if the spatial proximity condition is met, a weighted average concentration is used to write to the particles. The relationship between the concentration interpolation weight and the squared exponential decay of the spatial deviation is given by the following formula:

[0142] ;

[0143] in, Point mass At any moment The telemetry fusion concentration; Point mass Spatial coordinates; : No. Original coordinates of the center of each airborne pixel; The pixel is in Current flow reverse drift displacement; The observed concentration value of this pixel; The number of pixels that meet the capture radius requirement; Capture radius parameters to control the spatial influence range.

[0144] S15. Arrange AUV profile observations in ascending chronological order. Project onto the nearest particle; to avoid early observations having an excessive impact on the current scene, a time-difference-based approach is used. The concentration is weighted by multiplying the exponentially decaying temporal weight by the spatial Gaussian weight. Through this decaying mechanism, the contribution of observations from distant time periods and spatial locations to the current concentration is automatically reduced.

[0145] S16. Wind and wave-turbidity anomaly removal and interpolation: Continuously read wind speeds from on-site weather stations and buoys at the mass level. Effective wave height and turbidity If it appears or or The corresponding dynamic and concentration records of the particles are then marked as anomalies and removed. For the spatial voids created by the removal, third-order B-splines are applied for local interpolation along the particle history trajectory to ensure the dynamic grid library is maintained. With concentration grid library Without any gaps in time and space, the integrated processing of ocean dynamics and concentration fields is completed.

[0146] Step S2 includes the following steps:

[0147] S21. Shear gradient thin band extraction at each dynamic grid point. First, read the main flow direction and construct the neighborhood along the flow direction. To address the objective characteristic of the sharp, abrupt change in the aggregation ridges of microplastics, the shear strength is calculated using the following formula. Then Percentile thresholds set according to the scenario The grid points are sorted, retaining the set of points with the steepest gradients and a continuous, banded distribution. This set reduces the search space, laying the geometric foundation for subsequently locking the convergence axis.

[0148] ;

[0149] in, Grid Shear strength; Grid The instantaneous velocity vector; and A unit vector along the main current direction; and The displacement vector; :and Set of adjacent grid points in the same direction; : Percentage of quantiles extracted from the thin band.

[0150] S22. Velocity abrupt changes within the shear band are often affected by short-period wind and wave noise, easily leading to transient spurious peaks. To improve the physical reliability of core nodes, a consistency check is performed on the same shear band grid point in two consecutive frames of the dynamic field. Only if the grid point satisfies the condition in both frames is a consistency check performed. When the node is in its convergence core, it is marked as such. The time-consistency screening simultaneously considers particle flux conservation and convergence front stability to filter out transient misjudgments.

[0151] S23, Downstream directed convergent edges, for each convergent core node Read its main direction unit vector Iteratively search for first-order neighbors along this direction. When the flow velocity at the neighboring point is... The included angle satisfies At that time, in the node pair Write directed convergent edges between them The process proceeds recursively until it terminates at a shoreline obstruction or a low-speed buffer zone. The process outputs a chain-like structure that strictly follows the hydrodynamic direction, clearly defining the main channel through which microplastic particles are drawn within a narrow band.

[0152] S24. Lateral diffusion weight calculation: Switch to the concentration grid library for each node. Its lateral neighbor Concentration gradient detection is performed. Considering the lateral spread effect caused by wind and wave disturbance, the diffusion weight is obtained by coupling and quantizing the concentration difference and lateral distance between nodes using the following formula. The weights not only reflect the increasing concentration trend but also incorporate spatial decay, providing a normalized assessment of the diffusion contribution at different particle size scales. ;

[0153] in, Node pairs diffusion weights; :node particle concentration; : The lateral displacement vector between nodes; Spatial attenuation scale; Numerical stability constant; and The set of horizontally adjacent grid points; : Diffusion connection threshold.

[0154] S25. Diffusion edge filtering and write weighting, based on the results generated in the previous step. Implement full-court positional normalization and select Establish undirected diffusion edges between nodes ,Will Write directly into the edge attributes. This ensures that only diffusion paths that significantly contribute to lateral spread are retained, reducing the interference of low-intensity noise on subsequent GNN propagation.

[0155] S26. Assemble the two-layer heterogeneous topology, ensuring complete consistency of node indices, and merge all convergent core nodes and their directed convergent edges into the upper-layer graph. Merge all diffusion nodes and their undirected diffusion edges into the lower-level graph. Node sharing ensures that the information from both layers can be aligned in the same spatial reference frame, while avoiding redundant storage, resulting in a compact two-layer heterogeneous graph with clear physical meaning. .

[0156] S27, on the upper edge Write shear rate As a weight, in the lower layer edge Write diffusion weights Then, a node-level cross-layer mapping table is generated. The table provides a synchronization index for the dual-flow graph neural network, enabling a one-to-one correspondence between shear flow and diffusion flow at each node in the subsequent gated fusion stage. This completes the construction of a shear-dominated heterogeneous graph, providing a physically interpretable graph structure prior for accurate short-term prediction of the dynamic drift trajectory of microplastic convergence ridges.

[0157] Step S3 includes the following steps:

[0158] S31. Topology entry splitting, targeting heterogeneous graphs. The upper-level directed convergent edges are analyzed as adjacency matrices. Simultaneously, the lower-level undirected diffusion edges are analyzed as adjacency matrices. By splitting the topology entry point, we can avoid the convergence chain and the diffusion network from mixing in the initial stage, and provide clear edge direction constraints for the subsequent design of flow-related convolutional kernels. This ensures that the convergence information is not diluted by diffusion smoothing during high-frequency tidal oscillations, preserving the geometric sharpness of the narrowband front.

[0159] S32. Node physical feature injection, targeting any node within the sea area. Read instantaneous flow rate model Local shear rate and particle concentration Concatenate them to form the original feature vector. Perform a linear mapping and add sinusoidal positional encoding to convert it into a dimensionless array. The initial embedding .

[0160] This ensures that the convergence-diffusion dual channels possess both dynamic gradient and material gradient semantics when entering the network, which is suitable for the tidal contraction + lateral diffusion coupling scenario.

[0161] S33, Convergence Flow Message Passing, along Perform a directed convolution, applying a shearing rate to the upstream node information. Weighted aggregation, and for those below a threshold Weak gradient edges undergo exponential decay. This maintains a sharp front during tidal surges and prevents weak gradient noise from widening the convergence axis. Output embedding The focus is on describing the narrow-band aggregation trend of particles being pulled by the mainstream, which is crucial for locating the main ridge of microplastics.

[0162] S34, Diffusion Flow Message Passing, Independent Along When performing a horizontally diffusing convolution, the aggregation function uses diffusing weights. A soft mask is used, and a random walk regularization is superimposed within the kernel function to simulate the random diffusion caused by microscale agitation from wind and waves. The effect of concentration increments on the lateral neighborhood is amplified, making the network more sensitive to the spread within water masses, and the output embedding is improved. It mainly depicts the lateral diffusion pattern of matter.

[0163] S35. Wind and wave potential energy – vortex gating calculation, for nodes. First, it is determined by the effective wave height of the buoy. Calculate the wind and wave potential energy index Then, the relative vorticity is obtained by calculating the curl of the local velocity field. The gating factor is obtained by linearly combining the two and applying a Sigmoid mapping. This is used to measure the wave drive shear intensity at the current moment. Wave height rise or vorticity enhancement will increase... This makes the model more confident in the shear channel; conversely, when the sea surface is calm and the eddy current is reduced, the model is more likely to trust the shear channel. The model automatically favors diffusion channels as the number of channels decreases.

[0164] ;

[0165] in, Gating coefficient; : Sigmoid mapping function; Trainable weights and biases; : Wave potential energy index; Relative vorticity modulus; Converging flow embedding; : Diffusion flow embedding; : Fusion and embedding.

[0166] S36. Based on S35, when the wind and wave potential energy is high and the vortex is significant. near When the convergent current weights are amplified, the model focuses on maintaining the sharpness of the front; when the sea state is calm, near The diffusion flow weights are increased, and the model focuses on lateral spread details. An adaptive strategy is used to fuse embeddings. It contains two types of mechanism features, reducing the ridge positioning error caused by single-flow bias.

[0167] S37. To avoid numerical scale drift caused by dual-stream fusion, With initial embedding Add the residuals and perform layer specification to obtain the final node representation. The characterization maintains strong physical interpretability and gradient stability, and can seamlessly characterize inputs to multi-scale ridge growth processes, thereby enabling accurate short-term prediction of nearshore microplastic convergence ridges in high-frequency tidal-wave coupled environments.

[0168] Step S4 includes the following steps:

[0169] S41, Fusion Embedding The weights of both shear and diffusion mechanisms have been integrated, but the actual aggregation sites of microplastics still require further screening. This step first calculates the shear stability score for all nodes. . Instantaneous shear weights at nodes The primary factor is used, and a Gaussian attenuation is applied to the embedding difference between adjacent frames to measure the continuity and intensity dual criteria. A high score indicates that the convergence is both sharp and sustained, consistent with the physical characteristics of a narrow converging ridge source. After traversing the entire field, only the... The nodes serve as seeds, ensuring that the starting points for subsequent ridge growth have a stable physical foundation and are not affected by short-lived storm peaks:

[0170] ;

[0171] in, :node The current shearing weights come directly from the convergence channel of the two-stream GNN; , : The fused embedding vector of a node in two consecutive frames, used to characterize temporal stability; : The attenuation scale that controls the range of cross-frame difference influence; Seed threshold, set to the entire field. Percentile.

[0172] S42. Single-scale main ridge extension: Using reliable seeds, extend them into a continuous ridge line. This step is performed at each seed node. Reading local mainstream direction Then proceed according to a fixed step size Proceed in this direction. To avoid the ridge line accidentally entering shelf shoals or shear attenuation zones, the iterative process must simultaneously ensure that the shear strength is not less than [a certain value]. The two conditions, namely "not touching the shoreline," are logically and geometrically constrained and embodied in the following formula. For each legal advance, the coordinates of the new node are recorded. It updates the local flow direction until any condition fails, at which point it stops. The resulting main ridge naturally aligns with the hydrodynamic transport conveyor belt, covering the maximum length of the microplastic main polymerization axis.

[0173] Iteration conditions:

[0174] ;

[0175] in, Iteration number The coordinates of the ridge node; The single-step advance distance along the flow direction is adaptively set based on the grid resolution. :node The current main direction unit vector; : Nodal shear strength, used to determine whether the front is still sharp; : Shear attenuation threshold; below this value, the front is considered to have disappeared; : Shoreline discrimination function, value The time indicates that it hit the shore.

[0176] S43. Multi-scale parallel ridge growth: Microplastics often accumulate not only along a single axis but also form parallel ridges of varying widths on either side. To capture this multi-scale spatial hierarchy, this step uses the main ridge as the geometric center and grows along the normal direction... Upper arithmetic increasing distance The starting point is copied, and then extended according to the same rules as S42. The copy offset relationship and scale parameters are reflected in the following formula. Through multi-scale scanning from meters to hundreds of meters, fine textures (which facilitate the estimation of the upper limit of polymerization concentration) and coarse-scale channels (which determine the overall drift velocity) can be identified simultaneously, providing a richer pool of candidates for subsequent kinetic optimization.

[0177] ;

[0178] in, Parallel Ridge In step The node coordinates; : The normal unit vector orthogonal to the main current direction; Minimum normal offset, usually equal to the single-cell spacing; The scale increment step size is set based on experience with the scene's diffusion width. Scale index It refers to the main ridge itself.

[0179] S44. Ridge segment dynamic attribute annotation: To select the most credible main ridge from numerous candidate ridges, dynamic scoring needs to be performed on each local segment of each ridge. This step assigns candidate ridges to fixed arc lengths. Segmentation and cutting Relative vorticity With wave energy Calculate the average of each segment to form a segment attribute vector. See the formula below for details. The vector maps the three dynamic factors to the same scale, quantitatively revealing whether shear and wave energy are sufficient to overcome vortex tearing, and is the core basis for subsequent optimization decisions.

[0180] ;

[0181] in, Arc length The local shear strength at a point reflects the strength of the convergent front; Arc length The relative vorticity at a given location measures the tendency of local rotational tearing. Arc length The energy of the wave field indicates the potential energy supply that the wind and waves use to maintain the front. : Arc length coordinates of paragraph endpoints; The segmented arc length interval is usually 3–5 grids.

[0182] S45. Based on the results obtained in the previous step Scan segment by segment from the origin of the ridge downstream: if a certain segment Exceeding the stability threshold This indicates that the vortex's tearing effect is strong, and it should be immediately cut off; if the wave energy... Below the maintenance threshold Similarly, it is truncated. This prevents the ridge line from mistakenly extending into the structurally vulnerable area and maintains the physical stability of the predicted axis.

[0183] S46. For the ridge line that remains after truncation, harmonize the vorticity by applying exponential decay to the shear strength and wave energy of all segments, and calculate the comprehensive score. According to the whole field percentile threshold Only the high-pitched ridges are retained. Because... By superimposing strong shear and high wave energy, and applying an exponential penalty to high vorticity, it is possible to effectively distinguish between true ridges and false ridges, thereby excluding error sources from subsequent extrapolation.

[0184] S47, Ridge Geometric Smoothing and Output:

[0185] Finally, perform two-step geometry optimization on the preserved ridge lines:

[0186] Local hydrodynamic regression: The position of the fine-tuning point is fitted with a first-order linear fit along the flow direction to remove the jagged edges caused by integer grids;

[0187] Third-order B-spline smoothing: Reconstructs the entire ridge curve, suppressing short-wavelength artifacts while maintaining the overall orientation. The optimized output... It simultaneously meets the triple requirements of geometric continuity, sharp shear, and dynamic stability, and is the initial ridge line with the smallest error and highest reliability when performing short-term drift extrapolation.

[0188] Step S5 includes the following steps:

[0189] S51. Historical ridge line sequence organization, in the data window. Internal collection of high-confidence converged ridge centerlines for each frame Sort by observation time in ascending order and resample to a uniform The mesh size is minimized to ensure tidal continuity between adjacent frames. The alignment operation provides the recursive model with a complete and uninterrupted ridge evolution path, avoiding memory gate synchronization issues caused by unequal time intervals.

[0190] S52. Driving vector construction, for any ridge point in the sequence. Adding the tidal phase at the same time shore-to-shore current velocity Coastal current velocity Wind direction and wind speed and wave field energy Concatenate them into a fixed-dimensional driving vector And concatenate it with geometric coordinates to form an input vector. The driving information propagates along with the nodes, enabling the recursive unit to adaptively switch the dynamic weights between tidal surges and wind jumps.

[0191] S53, Tidal Harmonic Memory Gate Recursive Propagation, Feeding into the recursive unit with tidal harmonic memory gate: The basic gating structure follows the GRU, but the main tidal harmonic reference is explicitly introduced in the update gate. and Simultaneously, additional migration gates are set for sudden wind direction events. The gating update relationship is as follows:

[0192] ;

[0193] in, : Input vector at the current step; , The current and previous hidden states; Update and reset doors; : Tidal harmonic embedding matrix; : Astronomical semi-diurnal and solar tidal frequency; : The difference in wind direction between the current frame and the previous frame; : Migration door, sensitive to sudden changes in wind direction; : Wind mutation migration vector; , Sigmoid and hyperbolic tangent activation; Hadamard multiplication by element.

[0194] Through tidal harmonic terms, the recursive unit can cyclically memorize along the tidal phase during the steady segment; when During the sudden rise, Fast injection of transfer vectors in open and hidden states Capture sudden drift acceleration.

[0195] S54. Frame-by-frame shift extrapolation, starting from the latest observed frame. Start, according to the preset step size Iteratively run the recursive unit and output the ridge point displacement increment. Then, the coordinates are superimposed onto the coordinates of the previous frame to generate the future. ( The predicted ridge sequence is obtained. The displacement increments are derived directly from the hidden state map, ensuring that tidal periodicity and wind-triggered nonlinear drifts are synchronized to the geometry.

[0196] S55, to quantify the uncertainty of extrapolation, regarding future wind fields With tidal range Apply Group of orthogonal perturbations, repeated S54 to obtain Group ridge line samples. Calculate the sample standard deviation for the coordinates of the same point. And mapped to spatial confidence radius :

[0197] ;

[0198] in, : No. Coordinates of the ridge point of the sample after the second perturbation; : The coordinates of the mean of the ridge point sample; Set size; :correspond Coefficient at the % confidence level.

[0199] S56. If there are nodes on a certain ridge line that satisfy... If so, the segment is marked as low trust and the trigger threshold is recorded. The corresponding main disturbance factor (wind or tide) is written into the metadata so that the mobile platform can quickly locate it during incremental sampling.

[0200] S57. Integrate the predicted ridge sequence with node confidence radius labels to generate a time-series ridge package. The forecast is written to the forecast database by time index and published in real time to the navigation, marine monitoring and assimilation modules through the message bus, so as to realize the dynamic push and sharing of drift paths and potential risk areas.

[0201] Step S6 includes the following steps:

[0202] S61. Scan the ridge packet with confidence radius output from the previous time period point by point, and calculate the average uncertainty of each continuous over-threshold segment according to the following formula. And extract its center coordinates With corresponding time window :

[0203] ;

[0204] in, : No. A continuous line segment exceeding the threshold at time 1 The average confidence radius; The number of ridge points contained in a line segment; : The set of ridge point indices for this line segment; : Ridge point The single-point confidence radius.

[0205] when When this happens, the line segment is marked as low trust and added to the encryption candidate pool to ensure that subsequent tasks focus only on real error sources.

[0206] S62. For each low-confidence line segment, retrieve the location sets of near-shore AUV, ASV, and USV platforms in real time. and remaining battery life ; Calculation platform arrival time from the center of the line segment ,like and Then the platform Considered as an executable unit, it is preferred. The smallest value is used to lock the execution carrier. The filtering process combines real-time tidal flow windows to avoid the high time consumption of platform reverse flow.

[0207] S63. Adopt a comprehensive cost function based on distance, energy consumption, and sea state. Planned flight path:

[0208] ;

[0209] in, :platform Execution segment The overall cost of the task; The current distance from the platform to the center of the line segment on the water surface; Platform cruising speed; Energy required to complete the planned profile; The platform's remaining available energy; : Effective wave height of the sea state along the line segment; Platform security risk ceiling; Adjustable weights.

[0210] By minimizing It automatically obtains the survey lines, sampling depths and frequencies along the ridge and the cross sections on both sides, forming an incremental observation list to ensure maximum information gain sampling of uncertain areas under limited energy consumption conditions.

[0211] S64. The observation list, along with wave height and navigation limit thresholds, is sent to the selected platform control terminal; the platform automatically collects concentration and current velocity profiles according to the planned trajectory and transmits the raw data packets back via high-frequency narrowband communication. During the mission execution, the sea state monitoring module monitors the data in real time. and Perform threshold comparison; if an over-limit is triggered, immediately adjust the test line or stop the test to ensure platform safety.

[0212] S65. After receiving the returned data, call the Lagrange coordinate mapping interface of step S1 to insert the newly acquired concentration profile into the corresponding particle node and replace the original low-confidence observation; synchronously update the concentration grid library and driving vector set to provide the latest status for rapid assimilation.

[0213] S66, Trigger short loop mode: Re-execute S3 dual-stream coding, S4 ridge generation, and S5 timing extrapolation, and output a new confidence radius assessment. If it still exists... If the line segment meets the criteria, return to S62 to re-select the platform, forming a closed-loop iteration; if all line segments meet the criteria, then enter the termination logic.

[0214] S67, When the confidence radius of the entire vertebral segment is... The system automatically sends a termination command and records the remaining flight time of each platform. and remaining energy The data is archived to the resource scheduling database, providing prior information for the next round of encryption. This completes a small-batch field encryption operation driven by uncertainty, enabling continuous reduction of the uncertainty in the location of microplastic convergence ridges with minimal fieldwork costs.

[0215] This embodiment first introduces a full Lagrange coordinate system in the data preprocessing layer, and uses this as the core to complete the strict spatiotemporal alignment of multi-source observations. This embodiment uses the shoreline DEM and intertidal boundary as constraints, uniformly distributing mass points in the tidal reach area, and then performing forward integration using a tidal three-dimensional numerical model. High-order interpolation of model current velocity and water level is achieved using cubic splines on the mass point's own time axis. In this unified reference frame, the Euler current velocity from the shore-based HF radar is mapped to the mass points through forward and backward tracking. Then, using airborne imaging spectral pixels and autonomous underwater vehicle profile observations, concentration information is written according to spatial proximity and temporal distance exponential weights. If wind speed, effective wave height, or turbidity triggers anomaly thresholds, the corresponding mass point is recorded as missing, and third-order B-spline local interpolation is performed along its historical trajectory. The above process ensures that each mass point simultaneously possesses continuous dynamic and material properties at any given time, enabling subsequent shearing analysis and graph convolution to be carried out in a high-frequency tidal background without phase drift. This precise alignment capability is unattainable by the traditional Euler interpolation-particle integration separation framework.

[0216] In the feature modeling and representation learning layer, this embodiment constructs a two-layer heterogeneous graph consisting of an upper-layer directed convergent chain and a lower-layer undirected diffusion network. The convergence layer recursively generates chain edges based on the mainstream gradient and writes the shear rate into the edge weights; the diffusion layer quantizes the edge weights through lateral concentration difference and spatial decay, retaining the main channels of lateral spread. The node index remains consistent in both layers, enabling shear and diffusion information to be convolved in parallel and then fused at the node level. Inside the graph neural network, shear rate-weighted directed convolution is used in the convergence channel to enhance the traction feature of particles along the mainstream; in the diffusion channel, diffusion weight soft masking is used and superimposed with random walk regularization to characterize the randomness of lateral diffusion induced by wind and waves. In the fusion stage, dynamic gating coefficients are generated by mapping the wind and wave potential energy index and local relative vorticity through Sigmoid, adaptively adjusting the contributions of the two channels, and finally, after residual connection with the input features, the output is normalized. The aforementioned layered graphing, dual-channel convolution, and physical quantity-driven gating mechanism avoid the problem of homogenization of shearing and diffusion in existing static undirected graphs, while ensuring that the sharpness of convergence peaks is not diluted by lateral smoothing during high-frequency tides.

[0217] In the temporal extrapolation and on-site closed-loop layer, this embodiment employs a harmonic memory gate (GRU) with explicitly embedded astronomical semi-diurnal tide and diurnal tidal frequency to recursively propagate historical ridge sequence and driving vectors, outputting ridge displacement increments in real time. When encountering sudden wind direction changes, an additional migration gate rapidly injects wind jump information into the hidden state, thus simultaneously describing periodic and transient dynamics. Uncertainties in future scenarios are addressed by applying orthogonal perturbations to the wind field and tidal amplitude to obtain a ridge set; the sample standard deviation is mapped to the node confidence radius and used to locate low-confidence ridge segments requiring encryption. Subsequent task scheduling uses distance, energy consumption, and sea state as a comprehensive cost function to plan trajectories along ridges and normal cross-sections for AUVs, ASVs, or USVs. Newly acquired profile data is written back to the mass library in real-time via a Lagrange mapping interface, triggering short-loop re-encoding and re-prediction, forming an adaptive closed loop updated in minutes. This multi-level collaborative process significantly shortens error convergence time under the same observational resources and ensures that the prediction results remain consistent with the latest tidal-wave conditions.

[0218] Compared to existing schemes that primarily rely on Eulerian grid assimilation combined with single-layer static graph convolution, the full Lagrangian alignment mechanism proposed in this embodiment enables the model flow field and multi-source observations to obtain a unified time reference at every mass point, fundamentally eliminating the phase mismatch problem during the high-frequency tidal oscillation phase. Therefore, it can maintain resolution for convergence frontal sharpness at the sub-grid scale. The bilayer heterogeneous graph constructed based on this unified reference frame exists in parallel with a shear-driven directed convergence chain and a transverse concentration difference-driven undirected diffusion network. Through node index sharing and physical quantity weighting, shear traction and diffusion propagation can be separated and modeled within the network before being gated and fused, avoiding frontal blunting caused by the homogenization process of traditional single-layer undirected graphs.

[0219] In the time-series extrapolation phase, the memory-gated GRU embedded with astronomical tidal harmonics, combined with a wind direction shift gate, can simultaneously maintain periodic tidal current memory and a sensitive response to sudden wind and wave events, providing a more stable drift vector output than conventional LSTM or pure data-driven methods. The field closed loop guided by the confidence radius further maps the predicted uncertainty directly to the platform sampling trajectory planning conditions, achieving minute-level data backfeeding and model re-prediction. This adaptive encryption strategy targeting error sources overcomes the time-sensitivity limitation of traditional batch processing modes that require waiting for a full tidal cycle to update. Overall, this embodiment exhibits differentiated advantages in structural aspects such as spatiotemporal alignment accuracy, shear-diffusion mechanism decoupling, and real-time closed-loop updates, enabling continuous output of physically consistent convergent ridge drift results with quantified uncertainty under complex nearshore sea conditions.

[0220] One possible specific embodiment:

[0221] From July 2023 to June 2024, the Marine Ecological Environment Monitoring Center of L Province conducted a special mission on the dynamic tracking and emergency response of microplastic convergence ridges in the nearshore waters of F Bay. In the past two years, this sea area has experienced multiple incidents of floating microplastics landing on fishing ports due to the migration of discharge outlets from upstream plastic processing parks and monsoon shifts. Traditional Eulerian grid forecasts based on offshore buoys and single-station HF radar often have a deviation of hundreds of meters in frontal position, failing to meet the precise requirements of emergency salvage windows. This mission was supported throughout the entire process by the technology platform of this invention, from multi-source data access to adaptive encrypted sampling closed-loop operation, aiming to reduce the positioning error of the convergence ridge to within 50 meters.

[0222] During the data acquisition phase, the technical team deployed a dual-station WERA HF radar array spanning 30 km, outputting surface current velocities at a 15-minute resolution. Simultaneously, three AUV profile lines were deployed, transmitting current velocities and turbidity at depths of 0–20 m daily. During the early morning low-cloud period, a fixed-wing UAV carrying a 400–1000 nm hyperspectral camera completed 18 flights along the coastal zone, acquiring areal images at a resolution of 1.2 m. The numerical background used the CMA 2022 tidal-wave coupled model with a grid resolution of 150 m and a time step of 300 s. All data, after coordinate unification and time-scale correction, were input into the Lagrange preprocessing module of this invention, ultimately releasing approximately 450,000 particles in the tidal-accessible area, forming a particle library with continuous dynamic-concentration attributes for 48 hours.

[0223] The shear-diffusion two-layer heterogeneous graph is constructed online with an update frequency of 5 minutes. The upper-layer directed convergent chain grows recursively through the mainstream gradient, with an average chain length of 1.8 km. The lower-layer undirected diffusion network selects edges based on the lateral concentration difference, with the edge density controlled at 6.2%, ensuring that the GPU memory usage is less than 32 GB. The dual-flow graph neural network was pre-trained for 18 rounds on historical observations from 2021 to 2023 before being put into inference. Each inference cycle takes 42 seconds and can continuously produce convergent ridge extrapolation sequences with 1-hour intervals and a maximum duration of 48 hours, with a 95% confidence radius attached to each node.

[0224] In the first month after deployment, the system identified 57 high-confidence main ridges, of which 9 were marked as low-confidence sections. Based on the confidence radius, the platform automatically planned 14 ridge-lined and intersecting profile tracks for the two USVs. The USVs actually traveled 137 km, transmitting 112 profiles back. The data was mapped back to the mass library in real time via Starlink before returning to port, facilitating 5 short-cycle re-predictions. Ultimately, the average distance error between the measured and predicted distances of the convergence ridge centerline was 34 m.

[0225] To verify the performance, the monitoring center simultaneously ran two control schemes: one was the "static Eulerian grid + single-layer GCN" method, and the other was the LCS analytical method based on the flow field divergence threshold. A comparative analysis of the three methods during the 12 spring tides from April to June 2024 is shown in Table 1.

[0226] Table 1

[0227]

[0228] As can be seen from the convergence ridge positioning evaluation results of the method in this embodiment and the control scheme in Table 1, this embodiment is significantly better than the two control schemes in terms of positioning accuracy and ridge detection rate. At the same time, it maintains the real-time level in terms of inference time, providing timely and reliable decision-making basis for emergency deployment during rapid tidal reversal.

[0229] Reference Figure 7 This application also provides a nearshore microplastic converging ridge dynamic prediction device, which can realize the above-mentioned nearshore microplastic converging ridge dynamic prediction method. The device includes:

[0230] A particle determination unit is used to determine multiple particles that are uniformly distributed in the tidal reachable area;

[0231] A particle library construction unit is used to predict the initial trajectory of each particle and interpolate new particles according to the initial trajectory to obtain a particle library.

[0232] The mainstream direction reading unit is used to read the instantaneous flow velocity of each particle in the particle library at the Eulerian point, and then obtain the mainstream direction of each particle;

[0233] The heterogeneous graph construction unit is used to assemble each of the mass points as nodes to obtain two-layer heterogeneous graphs, and write the shear rate and diffusion weight calculated according to the mainstream direction into the edge attributes of the upper heterogeneous graph and the edge attributes of the lower heterogeneous graph, respectively.

[0234] The matrix splitting unit is used to split the two layers of the heterogeneous graph into a directed convergent adjacency matrix and an undirected diffusing adjacency matrix, respectively.

[0235] The node representation extraction unit is used to extract node representations based on the directed convergent adjacency matrix and the undirected diffusing adjacency matrix.

[0236] The ridge seed selection unit is used to calculate the shear stability score based on the node characterization and select the nodes whose shear stability scores are greater than a preset score threshold as ridge seeds.

[0237] An optimal ridgeline calculation unit is used to calculate the optimal ridgeline based on the ridgeline seed.

[0238] The ridge prediction unit is used to construct driving vectors for the optimal ridges in multiple historical frames, and generate a predicted ridge sequence with a preset future duration based on the driving vectors and the geometric coordinates of the ridge points in the optimal ridges.

[0239] It is understood that the content of the above method embodiments is applicable to the present device embodiments. The specific functions implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0240] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method of this application. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0241] It is understood that the content of the above method embodiments is applicable to the device embodiments. The specific functions implemented by the device embodiments are the same as those of the methods of this application, and the beneficial effects achieved are the same as those achieved by the methods of this application.

[0242] Please see Figure 8 , Figure 8 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes:

[0243] The processor 801 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0244] The memory 802 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 802 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 802 and is called and executed by the processor 801.

[0245] The 803 input / output interface is used to implement information input and output.

[0246] The communication interface 804 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0247] Bus 805 transmits information between various components of the device (e.g., processor 801, memory 802, input / output interface 803, and communication interface 804);

[0248] The processor 801, memory 802, input / output interface 803, and communication interface 804 are connected to each other within the device via bus 805.

[0249] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method of this application.

[0250] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0251] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0252] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0253] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0254] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0255] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0256] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0257] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0258] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0259] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0260] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0261] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0262] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A method for dynamic prediction of nearshore microplastic convergence ridges, characterized in that, The method includes the following steps: Identify multiple mass points that are evenly distributed in the tidal reach area; Predict the initial trajectory for each of the said particles and interpolate new said particles according to the initial trajectory to obtain a particle library; The instantaneous flow velocity of each particle in the particle library at the Eulerian point is read, and the main flow direction of each particle is obtained. The mass points are assembled as nodes to obtain two heterogeneous graphs, and the shear rate and diffusion weight calculated according to the mainstream direction are written into the edge attributes of the upper heterogeneous graph and the edge attributes of the lower heterogeneous graph, respectively. The two heterogeneous graphs are respectively split into a directed convergent adjacency matrix and an undirected diffusing adjacency matrix; The node representation is extracted based on the directed convergent adjacency matrix and the undirected diffusing adjacency matrix; Based on the node characterization, a shear stability score is calculated, and nodes with shear stability scores greater than a preset score threshold are selected as ridge seeds. The optimal ridge line is calculated based on the ridge line seed; A driving vector is constructed for the optimal ridge line in multiple historical frames, and a predicted ridge line sequence with a preset future duration is generated based on the driving vector and the geometric coordinates of the ridge points in the optimal ridge line.

2. The method for dynamic prediction of nearshore microplastic convergence ridges according to claim 1, characterized in that, The prediction of the initial trajectory for each of the aforementioned particles includes the following steps: The initial trajectory is obtained by incorporating the three-dimensional flow field of the tidal numerical model onto each of the particles. The process of interpolating new mass points based on the initial trajectory to obtain a mass point library includes the following steps: Cubic spline interpolation was performed on the model flow velocity and water level on the mass time axis, and the flow velocity from shore-based high-frequency radar, the spectral density of airborne imaging, and the profile density of autonomous underwater vehicles were simultaneously superimposed as the observation mapping. By tracking back and forth, using spatial proximity and temporal distance weighting, the observations are mapped to corresponding particle points. After removing anomalies in wind speed, effective wave height, and turbidity, third-order B-spline interpolation is performed along the historical direction of the particle points to obtain a spatiotemporally continuous dynamic and concentration particle library.

3. The method for dynamic prediction of nearshore microplastic convergence ridges according to claim 1, characterized in that, The process of assembling each of the aforementioned particles as nodes to obtain two layers of heterogeneous graphs, and writing the shear rate and diffusion weight calculated based on the mainstream direction into the edge attributes of the upper and lower heterogeneous graphs respectively, includes the following steps: The shear strength along the main direction is calculated for each of the mass points using the Lagrange reference grid to extract the shear gradient band as the shear rate; Perform temporal consistency checks over two consecutive frames to determine the convergence core node; The directed convergence edge is recursively written according to the mainstream direction, and the diffusion weight is calculated and the undirected diffusion edge is selected based on the lateral concentration gradient. Assemble the two heterogeneous graphs under the condition that the node indexes are consistent, and write the shear rate and the diffusion weight into the edge attributes of the upper heterogeneous graph and the lower heterogeneous graph respectively.

4. The method for dynamic prediction of nearshore microplastic convergence ridges according to claim 1, characterized in that, The node representation extracted based on the directed convergent adjacency matrix and the undirected diffusing adjacency matrix includes the following steps: Read the node flow velocity modulus, local shear rate and particle concentration, and then splice them into physical features; Shear rate-weighted directed convolution is performed on the directed convergent adjacency matrix, and diffusion weight mask transverse convolution is performed on the undirected diffuse adjacency matrix. Then, the gating coefficients are obtained by Sigmoid mapping based on the wind and wave potential energy index and the local relative vorticity. The embedding features of the corresponding channels are extracted respectively; The embedded features of the two channels are adaptively fused, then added to the residual of the physical features, and then subjected to layer normalization to obtain the node representation.

5. The method for dynamic prediction of nearshore microplastic convergence ridges according to claim 1, characterized in that, The process of calculating the optimal ridge line based on the ridge line seed includes the following steps: Along the main mainstream direction, under the condition that the shear strength is not lower than the preset shear threshold and does not touch the shoreline, a single-scale main ridge is generated iteratively according to the ridge seed. Then, multi-scale ridges are generated in parallel by equal displacement in the normal direction under the condition that the shear strength is not lower than the preset shear threshold and does not touch the shoreline. For the single-scale main ridge and each of the multi-scale ridges, the average shear, relative vorticity and wave field energy of each segment are calculated according to a fixed arc length, and the segment attributes are constituted. After truncating the ridges according to the set attribute threshold, calculate the comprehensive dynamic score of the remaining ridges and retain the segmented ridges that are greater than the preset dynamic score threshold; Based on the segmented ridge line, and through first-order flow direction regression and third-order B-spline global smoothing, the optimal ridge line with geometric continuity is output.

6. The method for dynamic prediction of nearshore microplastic convergence ridges according to claim 1, characterized in that, The construction of the optimal ridge driving vector for multiple historical frames includes the following steps: The optimal ridge lines from multiple historical frames are sorted in ascending order by time and resampled to a unified grid, thereby constructing the driving vector that includes tidal phase, shore-directed or coastal current velocity, wind direction and wind speed pair, and wave field energy. The step of generating a predicted ridgeline sequence with a preset future duration based on the driving vector and the geometric coordinates of the ridge points in the optimal ridgeline includes the following steps: The driving vector and the geometric coordinates of the ridge point in the optimal ridge line are combined and input into a recursive network with a tidal harmonic memory gate. The main tidal harmonic reference is explicitly introduced in the update gate and a migration gate is set for sudden changes in wind direction. The ridge point displacement increment is output frame by frame to generate the predicted ridge line sequence with a preset future duration. Orthogonal perturbations are applied to future wind fields and tidal amplitudes and extrapolated repeatedly. Spatial confidence radii are calculated based on sample standard deviations, and low-confidence ridges are marked.

7. A method for dynamic prediction of nearshore microplastic convergence ridges according to any one of claims 1 to 6, characterized in that, The method further includes the following steps: Scan consecutive ridge segments in the predicted ridge sequence whose confidence radius exceeds a preset radius threshold; The system retrieves the location and remaining range of near-shore autonomous surface or underwater platforms, selects the execution platform based on a comprehensive cost function of distance, energy consumption, and sea state, and plans the trajectory along the continuous ridge segment and the cross-sections on both sides. The sampling command is issued to drive the execution platform to collect raw flow velocity and concentration profiles along the trajectory; The original flow rate and the concentration profile are inserted into the corresponding particles through the Lagrange mapping interface, and the particle library is updated. The process involves reading the instantaneous flow velocity of each mass point in the mass point library at Eulerian points to obtain the mainstream direction of each mass point, until the confidence radius of all ridge segments in the predicted ridge sequence is lower than the preset radius threshold, and then the final predicted ridge sequence is output.

8. A dynamic prediction device for nearshore microplastic convergence ridges, characterized in that, The device includes: A particle determination unit is used to determine multiple particles that are uniformly distributed in the tidal reachable area; A particle library construction unit is used to predict the initial trajectory of each particle and interpolate new particles according to the initial trajectory to obtain a particle library. The mainstream direction reading unit is used to read the instantaneous flow velocity of each particle in the particle library at the Eulerian point, and then obtain the mainstream direction of each particle; The heterogeneous graph construction unit is used to assemble each of the mass points as nodes to obtain two-layer heterogeneous graphs, and write the shear rate and diffusion weight calculated according to the mainstream direction into the edge attributes of the upper heterogeneous graph and the edge attributes of the lower heterogeneous graph, respectively. The matrix splitting unit is used to split the two layers of the heterogeneous graph into a directed convergent adjacency matrix and an undirected diffusing adjacency matrix, respectively. The node representation extraction unit is used to extract node representations based on the directed convergent adjacency matrix and the undirected diffusing adjacency matrix. The ridge seed selection unit is used to calculate the shear stability score based on the node characterization and select the nodes whose shear stability scores are greater than a preset score threshold as ridge seeds. An optimal ridgeline calculation unit is used to calculate the optimal ridgeline based on the ridgeline seed. The ridge prediction unit is used to construct driving vectors for the optimal ridges in multiple historical frames, and generate a predicted ridge sequence with a preset future duration based on the driving vectors and the geometric coordinates of the ridge points in the optimal ridges.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

Images