A marine anomaly monitoring and early warning method and system based on multi-source space-time registration and graph attention fusion

Abstract

The application relates to the technical field of marine anomaly monitoring, in particular to a marine anomaly monitoring and early warning method and system based on multi-source space-time registration and graph attention fusion. The method comprises the following steps: performing data preprocessing based on acquired marine multi-source observation data; performing quality control and space-time registration based on the preprocessed data; performing space-time feature extraction based on the space-time registration result; performing graph attention-based cross-modal correlation fusion based on the extracted space-time features; and performing marine anomaly event identification according to the fused features. The complete processing flow from marine multi-source observation data input, quality control and space-time registration, multi-modal space-time feature extraction, graph attention cross-modal fusion to anomaly event identification and early warning generation is realized, and the accuracy, timeliness and stability of marine anomaly event monitoring and early warning are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of marine anomaly monitoring technology, and in particular to a marine anomaly monitoring and early warning method and system based on multi-source spatiotemporal registration and graph attention fusion. Background Technology

[0002] With the increasing demands for marine resource development, ecological protection, and disaster prevention, marine monitoring is evolving from traditional single-point observation to multi-platform, multi-sensor collaborative sensing. Satellite remote sensing, drones, unmanned vessels, buoys, shore-based monitoring stations, underwater sensors, and sonar equipment can acquire multimodal data such as sea surface images, water color, water temperature, salinity, ocean currents, meteorological data, and underwater acoustic data, providing a data foundation for monitoring and early warning of marine anomalies such as red tides, oil spills, storm surges, and abnormal nuclear power plant cold sources.

[0003] However, existing marine observation data sources are scattered, with significant differences between different platforms in terms of data type, sampling frequency, spatial resolution, coordinate reference, time signature, and noise characteristics. Satellite remote sensing has a wide coverage but limited temporal resolution, buoys and seabed sensors can provide continuous time-series data but their spatial distribution is sparse, and UAV imagery has local high resolution but limited observation range. Because multi-source data are difficult to align under a unified spatiotemporal reference, problems such as information misalignment, missing evidence, and unstable fusion results are prone to occur during marine anomaly monitoring.

[0004] Existing marine anomaly identification methods mostly rely on fixed thresholds, empirical rules, or single models to judge water color anomalies, temperature changes, and pollution spread. These methods are effective in simple scenarios, but are prone to false alarms and missed alarms under complex sea conditions, sensor noise, cloud cover, and regional environmental differences. While some deep learning methods can extract temporal features from remote sensing images or sensors, they typically focus on single modalities or simple feature stitching, making it difficult to express the dynamic relationships between marine grids, observation stations, environmental elements, and anomaly areas.

[0005] Graph neural networks and attention mechanisms offer a new technical approach for the fusion of multi-source ocean observation data. By constructing observation stations, sea area grids, sensor nodes, and anomaly candidate regions into a graph structure, spatial adjacency, temporal propagation, and element coupling relationships between different observation objects can be expressed. Through graph attention mechanisms, weights can be adaptively assigned based on the contribution of different modalities and nodes to anomalous events. However, existing methods still lack effective integration with the multi-source spatiotemporal registration process, making it difficult to form a complete processing flow from heterogeneous data standardization, quality control, spatiotemporal alignment, cross-modal fusion to anomaly early warning generation.

[0006] Therefore, there is an urgent need for a multi-source data processing method for marine anomaly monitoring and early warning, which can uniformly represent, control the quality of and register the multi-source observation data in a spatiotemporal manner, and use graph attention mechanism to model cross-modal correlations, thereby improving the accuracy of marine anomaly identification, the timeliness of early warning, and the stability under complex sea conditions, such as red tides, oil spills, storm surges, and nuclear power plant cold source anomalies. Summary of the Invention

[0007] To address the problems in existing marine anomaly monitoring and early warning processes, such as inconsistent spatiotemporal references for multi-source observation data, insufficient fusion of heterogeneous data, inadequate accuracy in anomaly identification, and weak stability of early warning results under complex sea conditions, this invention provides a marine anomaly monitoring and early warning method and system based on multi-source spatiotemporal registration and graph attention fusion. Through multi-source observation data standardization, quality control and spatiotemporal registration, multimodal spatiotemporal feature extraction, cross-modal graph attention fusion, and anomaly identification and early warning generation mechanisms, the method achieves correlation fusion and anomaly discrimination of marine multi-source observation data under a unified spatiotemporal reference.

[0008] Firstly, the present invention provides a marine anomaly monitoring and early warning method based on multi-source spatiotemporal registration and graph attention fusion, which adopts the following technical solution: A method for monitoring and early warning of marine anomalies based on multi-source spatiotemporal registration and graph attention fusion includes: Acquire multi-source ocean observation data; Data preprocessing is performed based on the acquired multi-source ocean observation data; Quality control and spatiotemporal registration are performed based on the preprocessed data; Spatiotemporal feature extraction is performed based on the spatiotemporal registration results; Cross-modal association fusion based on graph attention is performed based on extracted spatiotemporal features; Identification of marine anomalies based on fused features.

[0009] Secondly, a marine anomaly monitoring and early warning system based on multi-source spatiotemporal registration and graph attention fusion includes: The data acquisition module is configured to acquire multi-source ocean observation data; The preprocessing module is configured to perform data preprocessing based on the acquired multi-source ocean observation data; The spatiotemporal registration module is configured to perform quality control and spatiotemporal registration based on the preprocessed data. The feature extraction module is configured to perform spatiotemporal feature extraction based on the spatiotemporal registration results; The fusion module is configured to perform cross-modal association fusion based on graph attention, based on the extracted spatiotemporal features; The anomaly identification module is configured to identify marine anomalies based on the fused features.

[0010] Thirdly, the present invention provides a computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the aforementioned method for monitoring and early warning of marine anomalies based on multi-source spatiotemporal registration and graph attention fusion.

[0011] Fourthly, the present invention provides a terminal device, including a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to provide a method for monitoring and early warning of marine anomalies based on multi-source spatiotemporal registration and graph attention fusion.

[0012] In summary, the present invention has the following beneficial technical effects: Compared to existing marine anomaly monitoring and early warning methods, which suffer from technical bottlenecks such as inconsistent spatiotemporal benchmarks for multi-source observation data, insufficient fusion of heterogeneous data, inadequate expression of cross-modal correlations, and weak early warning stability under complex sea conditions, this invention addresses the core requirement of "spatiotemporal registration and graph attention fusion early warning of multi-source marine observation data." It proposes a marine anomaly monitoring and early warning method, system, equipment, and storage medium based on multi-source spatiotemporal registration and graph attention fusion. This achieves a complete processing flow from inputting multi-source marine observation data, quality control and spatiotemporal registration, multimodal spatiotemporal feature extraction, cross-modal graph attention fusion, to anomaly event identification and early warning generation, effectively improving the accuracy, timeliness, and stability of marine anomaly monitoring and early warning.

[0013] First, this invention unifies the representation of multi-source observation data from satellite remote sensing, UAVs, and underwater sensors, organizing observation values, time stamps, spatial locations, modal categories, and observation source attributes from different sources into a unified observation sample. This solves the problems of inconsistent data structures, inconsistent coordinate representations, and difficulty in direct data fusion among multiple platforms in traditional marine monitoring systems. Furthermore, this invention enables heterogeneous observation data to have a unified computational representation through observation value encoding, multi-scale time encoding, spatial frequency encoding, gated fusion, and initial confidence calculation, providing a stable data foundation for subsequent quality control and spatiotemporal registration.

[0014] Secondly, this invention employs a multi-source heterogeneous data quality control and spatiotemporal registration mechanism to uniformly address missing observations, temporal asynchrony, inconsistent spatial resolution, and coordinate reference differences in marine observation data. This mechanism, based on observation quality status, anomaly scoring, confidence correction, neighborhood completion, time window synchronization, and marine grid mapping, transforms discrete, asynchronous, and multi-scale observation samples into registered observation results under a unified time window and a unified marine grid. This reduces the impact of observation misalignment, missing local evidence, and low-quality data on anomaly identification results.

[0015] Furthermore, this invention employs a multimodal spatiotemporal feature extraction and graph attention cross-modal association fusion mechanism to perform deep feature modeling on registered marine observation data. On one hand, the system can extract local spatial features, temporal evolution features, and multi-scale enhancement features from different modalities such as visual, hydrological, and meteorological. On the other hand, the system constructs marine grids and observation modalities as graph nodes, and learns the association weights between different marine units, different observation modalities, and different environmental elements through cross-modal association edges and graph attention message passing mechanisms. This enables the system to more fully express the spatial propagation relationships and multi-source evidence synergy relationships in events such as red tide spread, oil spill drift, and storm surge evolution.

[0016] Finally, this invention establishes a graph reasoning process for marine anomaly monitoring through an event recognition and early warning generation mechanism based on an anomaly state graph. The system can directly perform node-level anomaly discrimination, region-level anomaly aggregation, event type identification, risk level assessment, and early warning confidence calculation on the anomaly state graph after graph attention fusion, avoiding the problem of traditional methods only outputting discrete anomaly points or single threshold judgment results. By aggregating anomaly nodes into spatially continuous anomaly candidate regions and combining event type probability, regional anomaly intensity, spatial range, and trend probability to generate structured early warning results, this invention can provide more interpretable and operable early warning information for marine supervision, disaster prevention and control, and emergency response.

[0017] In marine anomaly monitoring and early warning applications, the method of this invention was validated in terms of spatiotemporal registration accuracy, cross-modal fusion consistency, anomaly area location accuracy, and early warning timeliness. Results show that the spatiotemporal registration mean square error of this invention is reduced by approximately 18.6% compared to the traditional linear interpolation registration method, the cross-modal fusion consistency score is improved by approximately 12.4%, the intersection-union ratio of anomaly area location is improved by approximately 9.7%, and the average advance warning of anomaly events is increased by approximately 21.3%. Even with an observation missing rate of 20% and enhanced sensor noise, the fluctuation range of early warning confidence is still controlled within 6.5%. Therefore, this invention improves the alignment accuracy and cross-modal correlation expression capability of multi-source marine observation data, while enabling regional, stable, and early warnings for events such as red tides, oil spills, and storm surges. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of a marine anomaly monitoring and early warning method based on multi-source spatiotemporal registration and graph attention fusion according to Embodiment 1 of the present invention; Figure 2 This is a schematic diagram comparing the spatiotemporal registration errors of various methods in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram comparing the cross-modal fusion consistency of various methods in Embodiment 1 of the present invention; Figure 4This is a schematic diagram comparing the intersection and union ratio of abnormal regions in each method of Embodiment 1 of the present invention; Figure 5 This is a schematic diagram comparing the event recognition accuracy of various methods in Embodiment 1 of the present invention; Figure 6 This is a schematic diagram comparing the average advance warning amount of each method in Embodiment 1 of the present invention; Figure 7 This is a schematic diagram comparing the decrease in stability under complex sea conditions for each method in Embodiment 1 of the present invention. Detailed Implementation

[0019] The present invention will be further described in detail below with reference to the accompanying drawings.

[0020] Example 1 Reference Figure 1 This embodiment of a method for monitoring and early warning of marine anomalies based on multi-source spatiotemporal registration and graph attention fusion includes: (1) Input and unified representation module of multi-source marine observation data This module addresses the challenges of diverse observation platforms, inconsistent data structures, and the difficulty in directly fusing raw observations in marine anomaly monitoring and early warning. Its goal is to provide unified access and standardized representation of multi-source marine observation data, including satellite remote sensing, UAVs, and underwater sensors. By establishing a unified descriptive structure encompassing observation values, observation time, spatial location, modal category, and observation source attributes, and further introducing multi-scale time coding, spatial frequency coding, gating fusion, and initial confidence calculation, the module transforms the raw heterogeneous data into a unified observation representation that can be directly processed by subsequent quality control and spatiotemporal registration modules.

[0021] 1) Input of multi-source observation data To enable the unified processing of marine data generated by different observation platforms within the same technical workflow, this module first receives multi-source marine observation data. This multi-source observation data can originate from satellite remote sensing, UAV surveys, unmanned surface vessels, underwater sensors, sonar equipment, or other marine monitoring equipment, without limiting the specific implementation form of the upstream observation platform. By uniformly inputting and representing observation data from different sources, a consistent data entry point can be provided for subsequent sample-level descriptions and unified encoding.

[0022] To uniformly describe ocean observation data from different sources, the input multi-source observation data is represented as a dataset organized by observation mode and observation source: , in, This represents the original set of multi-source ocean observation data. Indicates the first A subset of data from a similar observation source or observation mode. This indicates the number of observation sources or observation modes.

[0023] For the Classified observational data can be further represented as a set consisting of multiple observation samples: , in, Indicates the first The first in the observation sources One observation sample, Indicates the first The number of samples from the observation sources. This sample set is used to receive the original multi-source input and provides a processing object for the unified description of subsequent observation samples.

[0024] 2) Standardized description of observation samples To reduce the differences in field structure, time records, spatial coordinates, and device attribute representations among different data sources, this module further constructs a unified description unit for each observation sample. This description unit simultaneously retains the original observation values, observation time, spatial location, modality category, and observation source attributes, enabling different types of ocean observation data to enter the subsequent encoding process with the same data structure.

[0025] A unified description is constructed for each observation sample as follows: , in, This refers to the raw observations, which include image pixels, sensor readings, acoustic signals, or sequences of environmental parameters. Indicates the observation time; Indicates the observed spatial location; Indicates the observation mode category; This represents the attributes of the observation source, including device number, sampling frequency, spatial resolution, observation depth, coordinate datum, and data units. This unified description unit is used to preserve the key attributes of the observation data.

[0026] To accommodate surface, aerial, and underwater observation data, the spatial location of the observations is described using a three-dimensional spatial vector: , in, Indicates longitude. Indicates latitude, This indicates the observation height or underwater depth. This spatial representation can uniformly describe the three-dimensional spatial information of different observation platforms, providing a coordinate basis for subsequent spatial reprojection and multi-scale grid mapping.

[0027] 3) Joint coding of multi-source observation samples To map different types of observation samples, such as images, time series, acoustic parameters, and environmental parameters, to a computable unified feature space, this module encodes the observation values, time, spatial location, modal category, and observation source attributes separately. This process can eliminate the expression differences between different data formats while preserving the semantics and spatiotemporal attributes of the observations.

[0028] First, modal adaptation coding is performed on the original observations according to the observed modal category: , in, Indicates the first Class 1 The encoding results of the observation values ​​of each observation sample. Indicates the relationship with the first The encoder corresponding to the observation mode, This represents the parameters of the modal encoder. The encoding result is used to carry over the original observations, providing observation content features for the subsequent joint representation of temporal, spatial, and attribute information.

[0029] To express the variation characteristics of ocean observation data at different time scales such as diurnal cycle, tidal cycle, and long-term variation cycle, multi-scale periodic encoding is performed on the observation time: , in, Represents the observation time encoding vector. Indicates the first A time-scale periodic parameter, Indicates the number of time scales.

[0030] To enhance the distinguishability of different locations across different sea areas on a spatial scale, spatial frequency coding is performed on the observed spatial locations: , in, Represents a spatial location encoding vector. Represents the spatial frequency mapping matrix. This represents the aforementioned three-dimensional spatial location vector. The spatial encoding result inherits the observation location. It is used to provide a computable location representation for the subsequent spatiotemporal registration process.

[0031] After obtaining the observation value code, temporal code, and spatial code, they are concatenated with the modality category code and the observation source attribute code to form a joint coding vector for the observation samples: , in, Indicates the first Class 1 The joint encoding vector of each observation sample, Indicates modal category The encoding result, Indicates the attributes of the observation source. The encoding results are as follows. This joint encoding vector integrates observation content, temporal, spatial, modal, and device attribute information, providing input for subsequent gating fusion.

[0032] 4) Generation of gated fusion and unified observation results To avoid the direct impact of differences in data scale, field structure, and reliability among different observation modalities on subsequent processing, this module generates gated weights based on joint encoding vectors and fuses them with residual projection through nonlinear mapping to obtain initial observation features under a unified dimension. This process can adaptively adjust the degree of retention of effective information according to the observation modality and observation source attributes.

[0033] Calculate the gate weights based on the joint encoding vector: , in, Indicates the first Class 1 The gating weight vector of each observation sample, This represents the Sigmoid activation function. Represents the gated mapping matrix, This represents the gating bias vector.

[0034] To ensure that the dimensions of the vectors on both sides of the gating fusion are consistent, the joint encoding vector is first projected: , in, This represents the projection result of the joint encoding vector. Represents the projection matrix. This represents the projection bias vector. The projection result has the same dimension as the gating weights and is used in subsequent residual gating fusion.

[0035] Based on the gating weights and projection results, a gated fused representation of the observed samples is generated: , in, This represents the observed samples after gating and fusion. This represents element-wise multiplication. Represents a non-linear activation function. Represents a nonlinear mapping matrix. Represents the nonlinear mapping bias vector. This represents a vector of all 1s. The fusion representation incorporates the gating weights. and projection results It is used to generate a unified feature vector.

[0036] After obtaining the gated fusion representation, it is subjected to multi-layer mapping and normalization to obtain the initial observation feature vector under a unified dimension: , in, This represents the initial observed feature vector under a unified dimension. Presentation layer normalization operation, , Represents the feature mapping matrix, , Represents the feature map bias vector. This represents a unified feature dimension.

[0037] Considering the differences in missing data, noise levels, timeliness, and spatial positioning reliability among different observation samples, an observation quality vector is constructed: , in, Represents the observation quality vector. Indicates the degree of missing information. Indicates the noise level score. Indicates timeliness score, This indicates the reliability score for spatial positioning.

[0038] Based on the initial observation features, observation quality vector, and observation source attribute encoding, the initial confidence weights of the observation samples are generated: , in, Indicates the first Class 1 The initial confidence weights for each observation sample , , Represents the weight parameters. This represents the confidence bias term. This confidence weight comprehensively reflects observation quality, feature representation, and device attributes.

[0039] Furthermore, the initial observation feature vector, observation time, spatial location, modality category, and initial confidence weight are combined into a unified representation of the observation sample: , in, This represents the observation sample after unified characterization. This sample retains the spatiotemporal attributes of the original observations while possessing unified computational characteristics, providing standardized input for the next module to carry out quality control and spatiotemporal registration.

[0040] Finally, this module outputs a unified, characterized set of multi-source ocean observation data: , in, This represents the unique output of the multi-source ocean observation data input and unified characterization module. This result serves as input for subsequent modules, used to further complete outlier handling, missing data completion, time synchronization, and spatial alignment.

[0041] (2) Multi-source heterogeneous data quality control and spatiotemporal registration module This module builds upon the unified observation dataset output from the previous module, addressing issues such as noise interference, missing observations, outliers, temporal asynchrony, inconsistent spatial coordinate references, and resolution differences in multi-source marine observation data. Its goal is to transform observation samples from different sources, frequencies, and spatial scales to a unified spatiotemporal reference. This module generates a multi-source marine observation dataset with completed quality control and spatiotemporal registration through quality status modeling, reliability correction, missing data completion, time window synchronization, spatial coordinate transformation, and marine grid mapping.

[0042] 1) Modeling of observation quality status In order for the standardized observation samples output by the previous module to enter the quality control process, this module first performs... For each observation sample, a quality status model is performed based on its features, confidence level, and effective observation dimensions. This process does not repeatedly parse the original observation data; instead, it determines the usability of the sample under the current observation conditions based on the unified feature vector and initial confidence weights formed in the previous module.

[0043] For any uniform observation sample output by the previous module Construct its effective observation mask vector: , in, Indicates the first Class 1 The effective observation mask vector for each observation sample Represents the first eigenvector in the unified eigenvector. Each dimension is effective. This indicates that there are missing or invalid observations in this dimension. This represents a unified feature dimension.

[0044] Based on a unified feature vector, initial confidence weights, and an effective observation mask, a quality assessment input vector is constructed: , in, This represents the input vector for observation quality assessment. Represents the initial observed feature vector. Indicates the initial credibility weight. Represents the effective observation mask vector. This indicates the proportion of effective observations.

[0045] To identify observation samples affected by sensor noise, cloud cover, equipment drift, or anomalous readings, anomaly scores are calculated based on the quality assessment input vector: , in, Indicates the first Class 1 Anomaly score for each observed sample, Represents the quality state mapping matrix. This represents the quality state mapping bias vector. Represents the abnormal rating weight vector. This indicates anomaly scoring bias. Represents a non-linear activation function. This represents the Sigmoid function.

[0046] After the above processing, an anomaly score is obtained for each observed sample. This will serve as the basis for subsequent credibility corrections and missing information completion.

[0047] 2) Credibility Correction and Missing Information Completion To avoid interference from low-quality samples in the spatiotemporal registration process, the initial confidence level given in the previous module is corrected based on the anomaly score, and missing features are completed by combining observations from adjacent time periods, adjacent spaces, and similar modalities. This process can reduce the impact of anomalies on subsequent spatiotemporal alignment while preserving valid observation information.

[0048] The reliability of the observed samples is corrected based on anomaly scores and the proportion of valid observations: , in, This represents the credibility weight after quality control. Indicates the initial credibility weight. Indicates abnormal rating. This indicates the proportion of effective observations.

[0049] For observation samples with missing dimensions, a neighborhood sample set is first constructed based on temporal distance, spatial distance, and modal consistency: , in, Indicates sample The spatiotemporal neighborhood sample set, express Neighborhood observation samples, Represents the time neighborhood threshold. This represents the spatial neighborhood threshold.

[0050] Based on the neighborhood sample set, calculate the completion weight of the neighborhood samples to the current sample: , in, Representing neighborhood samples For the current sample The supplementary weights, Indicates the time-distance scale parameter. This represents a spatial distance scale parameter. Represents the modal category consistency function. This represents the confidence weight after quality control of neighboring samples.

[0051] Based on the completion weights and the effective observation mask, missing features are filled in on the unified feature vector: , in, This represents the observed feature vector after missing data completion. This represents element-wise multiplication. Represents a vector consisting entirely of 1s. This represents the initial observed feature vector of the current sample. This represents the initial observation feature vector of the neighborhood samples.

[0052] After the above processing, the corrected confidence level is obtained. and completion features This is used for subsequent time window synchronization.

[0053] 3) Time window synchronization To eliminate differences in sampling frequency and observation timestamps among different observation platforms, this module maps the completed missing observation samples to a unified time window sequence. This process enables comparison and fusion of satellite remote sensing, UAV surveys, buoy continuous monitoring, and underwater sensor observations on a unified time scale.

[0054] Let the unified time window sequence be represented as: in, Represents a unified time window sequence. Indicates the first A time window, This indicates the center moment of the time window. Indicates the radius of the time window. Indicates the number of time windows.

[0055] Based on the distance between the observation time and the center of the time window, calculate the synchronization weight of the sample to the time window: , in, Indicates sample For the first Synchronization weights for each time window Indicates the sample observation time. and They represent the first The and the first The center moment of a time window This represents the time synchronization scale parameter.

[0056] Based on the time synchronization weights and the completed observation features, time synchronization features are generated: , in, This represents the observational characteristics after time synchronization is complete. This represents the observed feature vector after missing data completion. Indicates the first Time index encoding for each time window This indicates a vector concatenation operation.

[0057] After the above processing, asynchronous observation samples are mapped to synchronous features under a unified time window. It is used for subsequent spatial coordinate transformation and mesh registration.

[0058] 4) Spatial coordinate transformation and mesh registration To address the inconsistencies in coordinate references, spatial resolutions, and coverage scales among different observation platforms, this module maps time-synchronized observation samples to a unified marine spatial grid. This process unifies point observations, regional observations, and image pixel observations into a computable marine grid representation through coordinate transformation matrices and spatial kernel weights.

[0059] Let the unified sea area grid set be: , in, Represents a unified sea area grid set. Indicates the first Each marine grid unit Indicates the number of grid cells.

[0060] For the sample The spatial location is transformed based on the coordinate reference parameters corresponding to its observation source attributes: , in, This indicates the sample location after conversion to a unified spatial reference. Indicates the first The spatial coordinate transformation matrix of the observation source. Represents a spatial translation vector. Indicates the original spatial location of the sample observation.

[0061] Based on the distance between the transformed spatial location and the center of the marine grid, the spatial registration weights from the sample to the grid cell are calculated: , in, Indicates sample To the Spatial registration weights for each sea area grid cell and Representing grid cells respectively and The center coordinates, Indicates the first The spatial uncertainty covariance matrix of the observation source.

[0062] After the above processing, the observation samples were transformed to a unified spatial reference and spatial registration weights for the marine grid were obtained. .

[0063] 5) Registration result generation To generate a data representation usable by the subsequent multimodal spatiotemporal feature extraction module, this module jointly calculates quality control weights, time synchronization weights, and spatial registration weights, and aggregates multi-source observation features on a unified time window and a unified sea area grid. This process can transform discrete, asynchronous, and heterogeneous ocean observation samples into regularized spatiotemporal registration results.

[0064] Based on quality control weights, time synchronization weights, and spatial registration weights, the joint registration weights of samples with respect to a specified time window and grid cell are calculated: , in, Indicates sample For the first The first time window and the first Joint registration weights for each sea area grid cell, This represents the credibility weight after quality control. Indicates the time synchronization weight. Indicates spatial registration weights. This represents the smallest positive number that prevents the denominator from being zero.

[0065] Based on the joint registration weights, a registration observation representation on a unified spatiotemporal unit is generated: , in, Indicates the first The first time window, the first Registration observation features on each sea area grid cell Presentation layer normalization operation, This shows the observation characteristics after time synchronization is completed.

[0066] Furthermore, the registration observation features across all time windows and sea area grid cells are organized and output: , in, This represents the set of multi-source ocean observation data after quality control and spatiotemporal registration have been completed. Indicates a unified time window Marine grid unit The corresponding registration observation results. The output of this module will serve as the input for subsequent modules.

[0067] (3) Ocean multimodal spatiotemporal feature extraction module This module builds upon the spatiotemporal registration results output by the previous module, addressing the challenges in marine anomaly monitoring such as inconsistent feature representations across different modalities, unclear local anomaly changes, and difficulties in uniformly modeling temporal evolution features. Its goal is to extract multimodal spatiotemporal features from marine observation data within a unified time window and a unified marine grid. This module transforms the registered observation data into a unified set of multimodal spatiotemporal features through modal decoupling mapping, local spatial feature extraction, temporal dependency modeling, and multi-scale feature enhancement.

[0068] 1) Registration observation feature input In order for the observation results after quality control and spatiotemporal registration to enter the feature extraction process, this module first receives the output of the previous module. The input has already undergone time synchronization and spatial grid mapping, so this module will no longer repeat the issues of original data source, missing data completion, and coordinate registration. Instead, it will focus on deep feature representation of the observation features in a unified spatiotemporal unit.

[0069] Any registration observation result output by the previous module is represented as: , in, Indicates the first The first time window and the first Registration observation results corresponding to each marine grid cell Indicates the characteristics of the registration observations. Indicates a unified time window. This represents a unified marine grid unit.

[0070] To organize continuous observation states under a unified spatiotemporal reference, the registration observation characteristics of the same sea area grid within a continuous time window are represented as a time series matrix: , in, Indicates the first The time-series observation matrix corresponding to each sea area grid cell Indicates the first Registration observation characteristics under each time window Indicates the number of time windows.

[0071] After the above processing, the registration observation results are organized into a time series representation oriented towards the sea area grid, providing input for subsequent modal decoupling and temporal feature extraction.

[0072] 2) Modal decoupling and fundamental feature mapping To separate information from different observation modes, such as visual, hydrological, meteorological, and acoustic, from the registered observation features, this module performs modal decoupling mapping on the registered observation features. This process calculates the contribution intensity of different modes in the unified features using a modal gating matrix and generates the corresponding modal basic features.

[0073] Let the modal set be represented as: , in, Represents the set of ocean observation modes. Indicates the first Observational modes, Indicates the number of modes.

[0074] Calculate modal gating weights based on registration observation features: , in, Indicates the first The first time window, the first In the first sea area grid cell Gating weights for observation modes, Indicates the first Modal mapping matrix, This represents the corresponding bias vector. This represents the modal scoring weight vector.

[0075] Based on modal gating weights, modal-based mapping is performed on the registered observation features: , in, Indicates the first Observation mode in the first The first time window, the first Basic features under each marine grid cell Indicates the first The feature projection matrix of the modality, This represents the feature projection bias vector. Presentation layer normalization operation.

[0076] After the above processing, the basic features \(\mathbf{f}_{l,g}^{s}\) corresponding to different observation modes are obtained, which are used for subsequent local spatial feature extraction and temporal dependency modeling.

[0077] 3) Local spatial feature extraction To capture the local diffusion, spatial texture, and regional continuous variation characteristics of marine anomalies between adjacent sea area grids, this module performs local spatial convolution extraction on modal features based on the adjacency relationship of sea area grids. This process can aggregate spatially relevant information from adjacent grids, enhancing the expressive power of spatial patterns such as red tide diffusion, oil spill drift, and anomaly regions in water temperature.

[0078] Let the first The spatial neighborhood set of a sea area grid cell is represented as follows: , in, Indicates the first The spatial neighborhood set of each marine grid cell Indicates the first Each marine grid unit and Representing grid cells respectively and The center coordinates, Represents the radius of the spatial neighborhood.

[0079] Modal spatial aggregation weights are constructed based on spatial neighborhood sets. , in, Indicates the first Modal mesh elements For grid cells Spatial aggregation weights, This represents a spatial distance scale parameter. Indicates the first Spatial correlation matrix of modalities and They represent the neighboring grid cells respectively. and Modal fundamental characteristics.

[0080] Generate local spatial augmentation features based on spatial aggregation weights: , in, Indicates the first Class modality in the first The first time window, the first Local spatial enhancement features under each marine grid cell Indicates the first The neighborhood feature transformation matrix of the modality.

[0081] After the above processing, local spatial enhancement features containing neighborhood spatial correlation information are obtained. This is used for subsequent time-dependent feature extraction.

[0082] 4) Temporal Dependency Feature Extraction To capture the evolutionary trends of marine anomalies within continuous time windows, this module constructs modal temporal inputs based on local spatial enhancement features and extracts temporal dependencies through gated recurrent units and a temporal attention mechanism. This process can express dynamic features such as slow changes in environmental elements, sudden enhancement of anomalies, and continuous diffusion of anomaly regions.

[0083] For the first The first sea area grid, the first Temporal input for constructing local spatial augmentation features of a modality: , in, Indicates the first The first sea area grid, the first Modal spatial augmentation time series matrix, Indicates the first Local spatial enhancement features within a time window.

[0084] Gated timing recursion based on spatially enhanced timing matrix: , in, Indicates the first Class modality in the first The first time window, the first Temporal hidden states under each sea area grid cell This indicates the hidden state of the time sequence in the previous time window. Indicates the first The gated loop unit corresponding to the modality, This indicates its parameters.

[0085] To highlight key time windows that contribute significantly to anomalous changes, temporal attention weights are calculated based on temporal hidden states: , in, Indicates the first Class modality in the first In the first sea area grid Time attention weights for each time window Represents the time attention mapping matrix, This represents the temporal attention bias vector. This represents the time attention score vector.

[0086] Modal spatiotemporal features are generated based on temporal attention weights: , in, Indicates the first The first sea area grid, the first Spatiotemporal characteristics of modalities Indicates time attention weights. This indicates the hidden state of the time sequence within the corresponding time window.

[0087] After the above processing, modal spatiotemporal features that combine local spatial structure and temporal evolution trend are obtained. .

[0088] 5) Multi-scale feature enhancement and output generation To further enhance the expressive power of ocean anomaly features at different scales, this module performs multi-scale mapping and residual normalization on the spatiotemporal features of each modality. This process can simultaneously preserve information at different scales, such as local anomalies, regional diffusion, and overall environmental changes, providing a node feature foundation for subsequent graph attention cross-modal association fusion.

[0089] Constructing multi-scale enhanced representations based on modal spatiotemporal features: , in, Indicates the first The first sea area grid, the first Modal-like multi-scale enhancement features Indicates the scale quantity. Indicates the first Class modality in the first Enhanced weights at each scale Indicates the first Feature mapping matrix at each scale This represents the corresponding bias vector.

[0090] Multi-scale enhancement features of different modalities are organized into multi-modal spatiotemporal feature units under the same sea area grid: , in, Indicates the first The multimodal spatiotemporal feature units corresponding to each sea area grid Indicates the first Multi-scale enhancement features of modalities.

[0091] Finally, the multimodal spatiotemporal feature units corresponding to all sea area grids are organized into the unique output of this module: , in, Represents the set of multimodal spatiotemporal features of the ocean. This indicates the number of grid cells in the sea area. The output of this module will serve as the input for subsequent modules.

[0092] (4) Cross-modal association fusion module based on graph attention This module inherits the multimodal spatiotemporal feature set of the ocean output from the previous module. Addressing the difficulty in explicitly expressing the relationships between different sea area grids, observation modalities, and environmental elements, it aims to construct a graph structure from the multimodal spatiotemporal features and learn cross-modal and cross-regional association weights through a graph attention mechanism. This module focuses on resolving the challenge of single-modal features or simple splicing methods failing to characterize the spatial diffusion, element coupling, and multi-source evidence synergy of ocean anomalies, ultimately generating a graph-attention fused representation of the ocean anomaly state. To avoid confusion with the output of this module, the number of sea area grids will be denoted as [missing information]. and order .

[0093] 1) Construction of multimodal graph nodes To convert the multimodal spatiotemporal features output from the previous module into node representations suitable for graph association modeling, this module constructs graph nodes using marine grids and observation modes as basic organizational units. Each node corresponds to a class of modal features within a specific marine grid, enabling the subsequent graph attention mechanism to simultaneously characterize spatial proximity relationships and cross-modal association relationships.

[0094] For any sea area grid feature unit output by the previous module It contains multi-scale enhancement features of different modalities: , in, Indicates the first The multimodal spatiotemporal feature units corresponding to each sea area grid Indicates the first In the first sea area grid Modal-like multi-scale enhancement features Indicates the number of modes.

[0095] In order to preserve the spatial location information of the marine grid, for the first Location coding is performed on the center coordinates of each sea area grid: , in, Indicates the first Location encoding vectors of each sea area grid This indicates the center coordinates of the grid for that sea area. Represents the spatial frequency mapping matrix. This indicates a vector concatenation operation.

[0096] Based on modal features and grid position encoding, an initial representation of the graph nodes is constructed: , in, Indicates the first The first sea area grid, the first Initial features of graph nodes corresponding to class modalities Indicates the first Modal node feature mapping matrix, Represents the position encoding mapping matrix, This represents the node mapping bias vector. Presentation layer normalization operation.

[0097] Furthermore, all sea area grids and the nodes corresponding to all modes are organized into a graph node set: , in, Represents the set of nodes in a multimodal graph. Indicates the first In the first sea area grid Graph nodes corresponding to class modalities This represents the initial features of the node. After the above processing, a multimodal node set for graph attention fusion is obtained. .

[0098] 2) Construction of cross-modal association edges To express the relationships between different marine grids and different observation modes, this module establishes cross-modal association edges between graph nodes. These association edges simultaneously consider spatial proximity, modal correlation, and feature similarity, enabling the graph structure to reflect the inherent connections between marine anomalies in terms of spatial diffusion, element coupling, and multi-source evidence synergy.

[0099] For any two graph nodes and Construct its associated feature vector: , in, Represents a node With nodes The associated feature vectors between them and These represent the initial features of the two nodes, This represents element-wise multiplication. This represents the spatial relationship encoding between two sea area grids. This represents the encoding of the modal relationship between two types of observation modes.

[0100] Spatial relationship encoding uses both distance attenuation and direction difference for representation: , in, Represents the sea area grid and Encoding of spatial relationships between them and These represent the coordinates of the centers of the two grids, This indicates the smallest positive number whose denominator is zero. This represents the spatial distance scale parameter.

[0101] Calculate the edge connection strength between nodes based on the associated feature vectors: , in, Represents a node With nodes The strength of the edge connection between them Represents the edge feature mapping matrix. This represents the edge feature mapping bias vector. Represents the edge strength score vector. This indicates the edge strength scoring bias term. This represents the Sigmoid function.

[0102] Filter cross-modal associated edges based on edge connection strength: , in, Represents the set of cross-modal associated edges. Represents a node To the node Associated edges, This represents the edge connection strength threshold. After the above processing, a set of cross-modal associated edges is obtained, encompassing spatial relationships, modal relationships, and feature similarity relationships.

[0103] 3) Graph attention information fusion To adaptively allocate fusion weights based on the contribution of different observation nodes to the anomalous state, this module performs graph attention message passing on the constructed cross-modal graph structure. This process takes node features and edge connection strength as input, calculates the attention weights of neighboring nodes to the target node, and aggregates cross-modal and cross-regional association information.

[0104] For the target node Define its graph neighborhood node set as: , in, Represents the target node The set of graph neighborhood nodes, This represents a neighboring node that has an edge connected to the target node. This represents the set of cross-modal associated edges.

[0105] To calculate the contribution of neighboring nodes to the target node, a graph attention score is constructed: , in, Representing neighboring nodes For the target node Graph attention weights, Represents the query mapping matrix. Represents the key mapping matrix, Represents the attention score vector. Indicates the edge connection strength. This represents a linear activation function with leakage correction.

[0106] Based on graph attention weights, neighborhood messages are aggregated to obtain node update representations: , in, Represents a node Updated features after graph attention fusion This represents the initial characteristics of the target node. Indicates the first The value mapping matrix corresponding to the neighborhood mode. This represents the characteristics of neighboring nodes.

[0107] To enhance the expressive power of cross-modal information propagation, multi-head graph attention is further employed for fusion: , in, This represents the node features after attention fusion in a multi-head graph. Indicates the first The node update features are obtained from the attention head. Indicates the number of attention heads. This represents the multi-head fusion mapping matrix. This represents the multi-head fusion bias vector. After the above processing, the node features enhanced with cross-modal and cross-regional association are obtained. .

[0108] 4) Abnormal state diagram generation To transform the node features obtained after graph attention fusion into a graph representation for anomaly identification, this module performs modal-level aggregation on nodes of different modes within the same sea area grid, and further forms a graph structure representation that includes node, edge, and anomaly state features. This result is used to support the aforementioned graph attention fusion process and serves as input for the subsequent anomaly event identification and early warning generation module.

[0109] For the same sea area grid Modality aggregation weights are calculated based on the fusion features of nodes from different modalities: , in, Indicates the first In the first sea area grid Modal aggregation weights of modal nodes Represents the modality aggregation mapping matrix. This represents the modal aggregation bias vector. This represents the modal aggregation score vector.

[0110] Generating a grid-level anomaly state representation of the sea area based on modal aggregation weights: , in, Indicates the first Characterization of anomalous states in a sea area grid Represents modal aggregation weights. The modal node features are represented after graph attention fusion.

[0111] Further construct the set of abnormal state nodes after graph attention fusion: , in, Represents the set of nodes in an abnormal state. Indicates the first Each sea area grid corresponds to an abnormal state node. Represents a marine grid cell. This represents the abnormal state of the grid.

[0112] Finally, the set of abnormal state nodes, the set of cross-modal associated edges, and the node-level fused features are combined to form the unique output of this module: , in, This represents the ocean anomaly state graph obtained based on graph attention fusion. Represents the set of nodes in an abnormal state. Represents the set of cross-modal associated edges. This represents the multimodal node features after graph attention fusion. The output of this module is... And serve as input for subsequent modules.

[0113] (5) Marine Anomaly Identification and Early Warning Generation Module This module builds upon the marine anomaly state graph representation output from the previous module. Addressing the challenges of marine anomalies exhibiting regional diffusion in space, multi-node associations in graph structure, and diverse categories such as red tides, oil spills, storm surges, and nuclear power plant cold source anomalies, this module aims to achieve node-level anomaly discrimination, regional anomaly aggregation, event type identification, and risk warning generation based on anomaly state nodes, cross-modal association edges, and graph attention fusion features. Instead of reprocessing the original observation data, this module directly infers from the graph representation generated in the previous module, ultimately outputting a set of marine anomaly event identification and warning results.

[0114] 1) Abnormal state graph input and node-level abnormality detection In order for the graph attention fusion result generated by the fourth module to enter the early warning reasoning process, this module first receives the abnormal state graph representation. The graph already includes marine grid-level anomalous state nodes, cross-modal correlation edges, and fused multimodal node features. Therefore, this module focuses on estimating the anomalous probability of the graph nodes.

[0115] The abnormal state diagram output by the previous module is represented as follows: , in, The diagram represents anomalies in the ocean. Represents the set of nodes in an abnormal state. Represents the set of cross-modal associated edges. Indicates the first Sea area grid, the first Graph attention fusion of node features corresponding to class modalities Indicates the number of grid cells in the sea area. Indicates the number of modes.

[0116] For any node in the set of abnormal state nodes, it is represented as: , in, Indicates the first Each sea area grid corresponds to an abnormal state node. Represents a marine grid cell. This represents the abnormal state of the grid.

[0117] To determine whether anomalies exist in each sea area grid, the node-level anomaly probability is calculated based on the anomaly state characterization: , in, Indicates the first Node-level anomaly probability of a sea area grid Represents the anomaly detection mapping matrix. This represents the anomaly detection mapping bias vector. Represents the anomaly probability score vector. This represents the anomaly probability scoring bias term. This represents the Sigmoid function.

[0118] Filter the set of candidate nodes for anomalies based on node-level anomaly probability: , in, Represents the set of candidate nodes for anomalies. This represents the threshold for identifying node anomalies. After the above processing, a set of candidate anomaly nodes is obtained. This is used for subsequent regional-level anomaly aggregation.

[0119] 2) Aggregation of abnormal candidate regions To avoid outputting warning results only as discrete grid points, this module further performs region-level aggregation based on the graph connectivity and spatial adjacency relationships between anomaly candidate nodes. This process can group spatially continuous nodes with closely related graph structures and high anomaly probabilities into anomaly candidate regions.

[0120] For any two abnormal candidate nodes and Build regional connectivity strength: , in, Indicates abnormal candidate nodes and The strength of the regional connection between them Indicates an indicator function, and Representing the sea area grid and The center coordinates, Indicates the spatial connectivity scale parameter. This represents the set of cross-modal associated edges from the fourth module. The average inter-mesh connectivity strength obtained through aggregation. , and This represents the weighting coefficient.

[0121] The average inter-mesh connectivity strength is expressed as: , in, Represents the sea area grid and Average plot connection strength between them Indicates the first The first grid Modal nodes and the first The first grid Associative edges between modal nodes Indicates the connection strength of the corresponding edges. This represents the smallest positive number that prevents the denominator from being zero.

[0122] Constructing a set of anomaly candidate regions based on region connectivity strength: , in, Represents the set of candidate regions for anomalies. Indicates the first One abnormal candidate region, Indicates by The resulting region connectivity strength matrix Indicates the region connectivity threshold. Indicates the number of abnormal candidate regions. This represents a connected component partitioning operation based on connection strength. After the above processing, a set of anomaly candidate regions is obtained, which is used for subsequent event type identification and risk level assessment.

[0123] 3) Regional-level event representation and type identification To classify anomaly candidate regions into different event categories, this module performs weighted aggregation of node state representations, anomaly probabilities, and graph connectivity strengths within each candidate region to form a regional event representation. This representation is used to distinguish between different types of anomalous events, such as red tides, oil spills, storm surges, and nuclear power plant cold source anomalies.

[0124] For the Anomalous candidate regions Calculate the aggregate weights of nodes within the region: , in, Indicates anomaly candidate region internal nodes Regional aggregation weights, Represents the region aggregation mapping matrix. This represents the region aggregation bias vector. Represents the region aggregated score vector. This represents the probability of node-level anomalies.

[0125] Generate regional event representations based on the aggregated weights of nodes within the region: , in, Indicates the first Regional-level event representation of anomaly candidate regions. This represents the aggregate weight of nodes within the region. This represents the abnormal state of a node. Presentation layer normalization operation.

[0126] Let the set of marine anomaly event types be represented as: , in, Represents a set of types of marine anomalous events. Indicates the first Class of exception event types, Indicates the number of abnormal event types.

[0127] Calculating the probability of anomalous event types based on regional-level event representation: , in, Indicates the first The anomaly candidate region belongs to the first... The probability of a class of abnormal events. Represents an event type mapping matrix. This represents the event type mapping bias vector. Indicates the first Classification weight vector of class events, Indicates the first Category bias for class events.

[0128] Determine the first Event types for each anomaly candidate region: , in, Indicates the first Predicted event types for each anomaly candidate region This represents the probability of the corresponding event type. Through the above processing, the regional-level event characterization and predicted event type for each anomaly candidate region are obtained, which are used for subsequent risk level and early warning confidence calculation.

[0129] 4) Risk level assessment and early warning result generation To transform the results of anomaly identification into early warning information that can be used for marine surveillance and emergency response, this module further calculates the risk level, trend probability, and early warning confidence of the anomaly area, and generates structured early warning results. This process comprehensively considers anomaly probability, area size, event type probability, and graph connectivity stability.

[0130] First, calculate the regional anomaly intensity of the anomaly candidate region: , in, Indicates the first The intensity of regional anomalies in candidate anomaly regions This represents the aggregate weight of nodes within the region. This represents the probability of node-level anomalies.

[0131] Calculate the spatial extent of the anomaly candidate region: , in, Indicates the first The spatial range of the candidate anomaly regions Represents marine grid cells The corresponding area.

[0132] Calculate the probability of risk level based on regional event characterization, regional anomaly intensity, and spatial extent: , in, Indicates the first The anomaly candidate region belongs to the first... The probability of each risk level. Indicates the number of risk levels. This represents the risk level mapping matrix. This represents the risk level mapping bias vector. Indicates the first A scoring vector for each risk level. This indicates the corresponding bias term.

[0133] Determine the first Risk level of each abnormal candidate region: , in, Indicates the first Predicted risk level of each abnormal candidate region This indicates the probability of the corresponding risk level.

[0134] To characterize the likelihood of an anomalous event continuing to develop within subsequent time windows, the trend probability is calculated: , in, Indicates the first The trend probability of each abnormal candidate region This represents the trend prediction mapping matrix. This represents the bias vector for trend prediction mapping. Represents the trend probability score vector. This represents the bias term in the trend probability score.

[0135] The confidence level of the early warning is calculated by combining the probability of the event type, the probability of the risk level, the intensity of the regional anomaly, and the probability of the trend. , in, Indicates the first The warning confidence level of each abnormal candidate region This indicates the weighting of the confidence level for early warning.

[0136] Based on event type, risk level, spatial range, trend probability, and early warning confidence level, the first... Results of alerts for individual abnormal events: , in, Indicates the first An abnormal event warning result, Indicates the type of event to be predicted. Indicates the predicted risk level. Indicates an abnormal candidate region. Indicates the range of abnormal space. Indicates the intensity of regional anomalies. Indicates the probability of a trend. This indicates the confidence level of the warning.

[0137] Finally, the early warning results of abnormal events that meet the early warning confidence requirements will be organized and output: , in, This represents the set of results for identifying and issuing early warnings of marine anomalies. This indicates the warning confidence threshold. The output of this module... This is used to represent the final generated monitoring and early warning results for marine anomalies.

[0138] Experimental verification To verify the effectiveness and reliability of the proposed marine anomaly monitoring and early warning method based on multi-source spatiotemporal registration and graph attention fusion in terms of multi-source observation data alignment, cross-modal feature fusion, anomaly region identification, event type discrimination, and early warning stability under complex sea conditions, a marine multimodal anomaly monitoring and early warning dataset was constructed. This dataset originates from various marine observation platforms, including satellite remote sensing, UAV patrols, buoy monitoring, shore-based monitoring stations, underwater sensors, sonar equipment, and temperature, salinity, and depth (TSD) observation equipment. It includes sea surface remote sensing images, UAV nearshore images, water color parameters, water temperature and salinity, ocean current velocity, meteorological data, underwater acoustic signals, and manually labeled anomaly event records, covering typical samples under different sea area types, observation scales, sea conditions, and anomaly event types.

[0139] ① Remote sensing and UAV visual samples mainly include image data such as sea surface color changes, suspected oil spill patches, red tide spots, floating objects on the sea surface, and nearshore anomaly areas, used to verify the ability to extract spatial features of visual modalities and locate anomaly areas; ② Buoy, shore-based monitoring station, and temperature, salinity, and depth observation samples mainly include time-series data such as water temperature, salinity, chlorophyll concentration, turbidity, dissolved oxygen, current velocity and direction, and meteorological parameters, used to verify the ability to control the quality of multi-source time-series data, time synchronization, and trend identification; ③ Underwater sensor and sonar samples mainly include data such as underwater acoustic echoes, cold source water intake anomalies, underwater target disturbances, and local environmental abrupt changes, used to verify the cross-modal correlation and fusion ability between underwater modalities and sea surface, meteorological, and hydrological modalities. All anomaly samples were manually labeled to form labeled data including anomaly type, occurrence area, time window, risk level, and event boundary.

[0140] The multi-source observation data in the dataset first underwent unified characterization and spatiotemporal registration processing, and was divided into three task levels according to the anomaly monitoring tasks: ① Basic spatiotemporal registration tasks, including time synchronization, coordinate transformation, grid mapping, and missing data completion of data from different observation platforms; ② Cross-modal fusion tasks, including correlation modeling and consistency fusion between remote sensing images, hydrological time series, meteorological data, and underwater acoustic signals; ③ Anomaly monitoring and early warning tasks, including regional location, type identification, risk level assessment, and early warning generation for events such as red tides, oil spills, storm surges, nuclear power plant cold source anomalies, and nearshore water quality anomalies. A total of 1,800 sets of marine multimodal observation samples were constructed, of which 1,200 sets were used for model training and parameter adjustment, and 600 sets were used for validation and testing. The test samples covered different sea areas, observation missing ratios, sensor noise levels, cloud and fog obstruction degrees, and anomaly event complexity.

[0141] To comprehensively evaluate system performance, four comparison methods were selected: ① Linear-Interp is a traditional linear interpolation registration method that maps multi-source observation data to a unified spatiotemporal grid through temporal and spatial interpolation, but does not perform deep cross-modal fusion; ② CNN-LSTM is a combination of visual convolutional networks and temporal recurrent networks that extracts remote sensing image features and sensor temporal features respectively and then performs simple concatenation; ③ Transformer-Fusion is a Transformer-based multimodal feature fusion method that fuses different modal features through a self-attention mechanism, but does not explicitly model the graph structure relationship between marine grids; ④ GNN-Fusion is a graph neural network-based marine observation data fusion method that can utilize the adjacency relationship of marine grids for feature propagation, but lacks multi-source spatiotemporal registration confidence constraints and cross-modal graph attention edge weight modeling; ⑤ The method of this invention achieves regional-level monitoring and early warning of marine anomalies through unified representation of multi-source observation data, quality control and spatiotemporal registration, multimodal spatiotemporal feature extraction, cross-modal graph attention fusion, and anomaly state graph inference mechanism.

[0142] All methods were evaluated under the same multi-source observation samples, the same anomaly event annotations, and a unified training and testing partition. To verify the system's stability in a real marine environment, three levels of complexity were set: ① Low-complexity samples, with relatively complete observation data, stable sea states, and clear anomaly event boundaries, mainly used to verify basic registration and routine anomaly identification capabilities; ② Medium-complexity samples, with some missing observations, mild cloud and fog obstruction, sensor noise, and multi-source time asynchrony issues, mainly used to verify quality control and cross-modal fusion capabilities; ③ High-complexity samples, with asynchronous observations from multiple platforms, ocean current disturbances, cloud and fog obstruction, acoustic noise, anomaly area diffusion, and multiple event superposition, mainly used to verify graph attention fusion and early warning stability under complex sea states.

[0143] The following evaluation metrics were used: ① Spatiotemporal Registration Error (STE) to measure the average spatiotemporal deviation after mapping multi-source observation data to a unified time window and sea area grid; ② Cross-modal Fusion Consistency (MFC) to measure the degree of consistent expression of different modal features within the same anomaly region; ③ Anomaly Region Location Intersection over Union (A-IoU) to evaluate the degree of overlap between the system's output anomaly region and the manually labeled anomaly region; ④ Event Recognition Accuracy (E-ACC) to measure the accuracy of identifying event types such as red tides, oil spills, storm surges, and nuclear power plant cold source anomalies; ⑤ Average Warning Lead Time (WLT) to evaluate the system's ability to provide early warnings compared to manual confirmation time or baseline warning time; ⑥ Stability Decline in Complex Sea State (S-DR) to measure the percentage decrease in system performance under high-complexity samples compared to low-complexity samples.

[0144] Experimental results show that the method of this invention exhibits good overall performance in marine anomaly monitoring and early warning tasks. Linear-Interp completes basic registration based on temporal and spatial interpolation, with a simple calculation process and certain stability in low-complexity samples, achieving a spatiotemporal registration error of 0.181. However, due to the lack of deep feature representation and cross-modal correlation modeling, the cross-modal fusion consistency is only 0.704, and the intersection-union ratio of anomaly region localization is 0.638. CNN-LSTM can extract image spatial features and sensor temporal features respectively, improving the event recognition accuracy to 84.7%. However, it uses a simple stitching method to fuse multi-source features, and under conditions of missing observations and multi-source asynchronous operation, the spatiotemporal registration error is still 0.196, with an average early warning lead time of 25.4 minutes.

[0145] Transformer-Fusion enhances information interaction between different modal features through a self-attention mechanism, achieving a cross-modal fusion consistency of 0.812 and an event recognition accuracy of 86.9%. However, due to the lack of explicit modeling of spatial adjacency and anomaly propagation relationships between sea area grids, the intersection-union ratio (IUU) for anomaly region localization is 0.721. GNN-Fusion utilizes the adjacency relationships of sea area grids for graph structure feature propagation, improving the IUU for anomaly region localization to 0.759 and reducing stability in complex sea conditions by 9.2%. However, it does not fully integrate multi-source registration confidence and cross-modal edge weight differences, resulting in a cross-modal fusion consistency of 0.798 and an early warning lead time of 30.8 minutes. In comparison, the spatiotemporal registration error of the method of this invention is reduced to 0.149, the cross-modal fusion consistency reaches 0.856, the intersection-union ratio of anomaly region localization reaches 0.795, the event recognition accuracy reaches 91.8%, the average early warning lead time reaches 37.6 minutes, and the stability decline in complex sea states is controlled at 6.3%. It outperforms the comparative methods such as Linear-Interp, CNN-LSTM, Transformer-Fusion, and GNN-Fusion in all core indicators.

[0146] In summary, the experimental results verify the effectiveness of the method of this invention in spatiotemporal registration, cross-modal fusion, anomaly region localization, event type identification, and early warning generation of multi-source marine observation data. This method not only reduces spatiotemporal mismatch errors in multi-source heterogeneous observation data but also improves the regional-level localization capability, cross-modal fusion consistency, and early warning capability of marine anomalies, providing an efficient, stable, and scalable technical solution for monitoring and early warning of marine anomalies such as red tides, oil spills, and storm surges.

[0147] Table 1 Comparison of data from different methods under six major indicators In performance evaluation, cross-modal fusion consistency (MFC), intersection-in-union ratio (A-IoU) of anomaly area location, event recognition accuracy (E-ACC), and mean advance time (WLT) are positive indicators. MFC measures the fusion consistency of different observation modalities within the same anomaly area. A-IoU evaluates the spatial overlap between the system's output anomaly area and the manually labeled anomaly area. E-ACC measures the accuracy of identifying event types such as red tides, oil spills, and storm surges. WLT measures the system's ability to provide early warnings compared to manual confirmation time or baseline warning time. Spatiotemporal registration error (STE) and stability degradation in complex sea states (S-DR) are negative indicators. STE measures the average deviation after mapping multi-source observation data to a unified time window and sea area grid. S-DR measures the percentage decrease in system performance under highly complex sea states.

[0148] The experimental results are shown in Table 1. Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 , Figure 7 As shown in the figure, in marine anomaly monitoring and early warning tasks, traditional linear interpolation registration methods can complete basic temporal and spatial interpolation, but they are difficult to handle nonlinear misalignment, observation gaps, and cross-modal semantic differences in multi-source observation data. CNN-LSTM methods can extract remote sensing image features and sensor temporal features respectively, but their fusion method mainly relies on feature splicing, making it difficult to express the correlation weights between different modalities. Transformer-Fusion methods can enhance the information interaction between features of different modalities, but lack marine grid structure constraints, and are prone to location discontinuity in anomaly diffusion and regional linkage scenarios. GNN-Fusion methods can utilize the adjacency relationship of marine grids for spatial feature propagation, but do not fully combine multi-source registration confidence and cross-modal edge weight differences. In contrast, the method of this invention introduces unified representation of multi-source observation data, quality control and spatiotemporal registration, multimodal spatiotemporal feature extraction, cross-modal graph attention fusion, and anomaly state graph reasoning mechanism, realizing a complete processing flow from multi-source data alignment, cross-modal correlation modeling to regional anomaly early warning, and has better fusion expression ability and early warning stability in complex marine observation environments.

[0149] The results show that the spatiotemporal registration error (STE) of the method of this invention is reduced to 0.149, which is about 17.7% lower than that of Linear-Interp and about 11.3% lower than that of GNN-Fusion; the cross-modal fusion consistency (MFC) reaches 0.856, which is 0.044 higher than that of Transformer-Fusion; the intersection-union ratio (A-IoU) of anomaly region localization reaches 0.795, which is 0.036 higher than that of GNN-Fusion; the event recognition accuracy (E-ACC) reaches 91.8%, which is 4.2 percentage points higher than that of GNN-Fusion; the mean advance warning time (WLT) reaches 37.6 minutes, which is 6.8 minutes higher than that of GNN-Fusion; and the stability reduction rate (S-DR) in complex sea states is controlled at 6.3%, which is significantly lower than that of the comparative methods such as CNN-LSTM, Transformer-Fusion and GNN-Fusion. The results demonstrate that the method of the present invention has significant advantages in spatiotemporal registration of multi-source marine observation data, cross-modal fusion, anomaly area location, event type identification, and early warning, and verifies its engineering feasibility and practical application potential in marine anomaly monitoring and early warning tasks such as red tides, oil spills, and storm surges.

[0150] Example 2 This embodiment provides a marine anomaly monitoring and early warning system based on multi-source spatiotemporal registration and graph attention fusion.

[0151] A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the method for monitoring and early warning of marine anomalies based on multi-source spatiotemporal registration and graph attention fusion.

[0152] A terminal device includes a processor and a computer-readable storage medium, the processor being used to implement various instructions; the computer-readable storage medium being used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to provide a method for monitoring and early warning of marine anomalies based on multi-source spatiotemporal registration and graph attention fusion.

[0153] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for monitoring and early warning of marine anomalies based on multi-source spatiotemporal registration and graph attention fusion, characterized in that, include: Acquire multi-source ocean observation data; Data preprocessing is performed based on the acquired multi-source ocean observation data; Quality control and spatiotemporal registration are performed based on the preprocessed data; Spatiotemporal feature extraction is performed based on the spatiotemporal registration results; Cross-modal association fusion based on graph attention is performed based on extracted spatiotemporal features; Identification of marine anomalies based on fused features; The quality control and spatiotemporal registration based on preprocessed data includes: modeling the quality status of each observation sample and constructing a quality assessment input vector; calculating anomaly scores based on the quality assessment input vector; correcting the credibility of observation samples based on anomaly scores and the proportion of valid observations; for observation samples with missing dimensions, constructing a neighborhood sample set based on temporal distance, spatial distance, and modal consistency, and calculating the completion weights of neighborhood samples for the current sample based on the neighborhood sample set; completing missing features in a unified feature vector based on the completion weights and a valid observation mask; and mapping the completed observation samples to a unified time window sequence. Based on the distance between the observation time and the center of the time window, the synchronization weight of the sample to the time window is calculated, and based on the time synchronization weight and the completed observation features, the time synchronization features are generated. The time-synchronized observation samples are mapped to a unified marine spatial grid. For the spatial location of the sample, spatial coordinate transformation is performed based on the coordinate reference parameters corresponding to its observation source attributes. At the same time, the spatial registration weight from the sample to the grid cell is calculated. Finally, the joint registration weight of the sample to the specified time window and grid cell is calculated. And generate a registered observation representation on a unified spatiotemporal unit; The cross-modal association fusion based on graph attention, which is based on the extracted spatiotemporal features, includes position encoding of the center coordinates of the sea area grid and constructing an initial representation of the graph nodes based on the modal features and the grid position encoding. All sea area grids and mode-corresponding nodes are organized into a graph node set to obtain a multimodal node set for graph attention fusion. Cross-modal association edges are established between graph nodes. For any two graph nodes, an association feature vector is constructed. Finally, cross-modal association edges are selected based on the edge connection strength. Graph attention message passing is performed on the constructed cross-modal graph structure. A graph attention score is constructed, and neighborhood messages are aggregated based on the graph attention weights to obtain the node update representation. Multi-head image attention is used for fusion; For the same sea area grid, the modal aggregation weight is calculated based on the fusion features of nodes of different modalities; Based on modal aggregation weights, a grid-level representation of marine anomalies is generated, and then a set of anomaly nodes after graph attention fusion is constructed.

2. The marine anomaly monitoring and early warning method based on multi-source spatiotemporal registration and graph attention fusion according to claim 1, characterized in that, The data preprocessing based on the acquired multi-source ocean observation data includes representing the input multi-source observation data as a dataset organized by observation mode and observation source: in, This represents the original set of multi-source ocean observation data. Indicates the first A subset of data from a similar observation source or observation mode. Indicates the number of observed modes; for the th Classified observational data can be represented as a set consisting of multiple observation samples: in, Indicates the first The first in the observation sources One observation sample, Indicates the first The number of samples in the observation sources; then a uniform description is constructed for each observation sample: in, Represents the original observation values; Indicates the observation time; Indicates the observed spatial location; Indicates the observation mode category; The observation source attributes are then represented; subsequently, modal adaptation encoding is performed on the original observations based on the observation modality category. in, Indicates the first Class 1 The encoding results of the observation values ​​of each observation sample. Indicates the relationship with the first The encoder corresponding to the observation mode, The parameters of the modal encoder are represented. To express the variation characteristics of marine observation data at different time scales, the observation time is periodically encoded at multiple scales. At the same time, to enhance the distinguishability of different sea area locations at spatial scales, the observation spatial location is spatially frequency encoded. After obtaining the observation value code, time code, and spatial code, they are concatenated with the modal category code and the observation source attribute code to form a joint encoding vector of the observation samples. Finally, gating weights are generated based on the joint encoding vector, and the initial observation features under a unified dimension are obtained by fusing them through nonlinear mapping and residual projection.

3. The marine anomaly monitoring and early warning method based on multi-source spatiotemporal registration and graph attention fusion according to claim 2, characterized in that, The quality control and spatiotemporal registration based on the preprocessed data includes first... For each observation sample, a quality state model is performed based on its features, confidence level, and effective observation dimensions. Construct an effective observation mask vector: in, Indicates the first Class 1 The effective observation mask vector for each observation sample Represents the first eigenvector in the unified eigenvector. Each dimension is effective. This indicates that there are missing or invalid observations in this dimension. We represent a unified feature dimension; then, based on the unified feature vector, initial confidence weights, and effective observation mask, we construct a quality assessment input vector: in, This represents the input vector for observation quality assessment. Represents the initial observed feature vector. Indicates the initial credibility weight. Represents the effective observation mask vector. This represents the proportion of valid observations; to identify observation samples affected by sensor noise, cloud cover, equipment drift, or abnormal readings, anomaly scores are calculated based on the quality assessment input vector. in, Indicates the first Class 1 Anomaly score for each observed sample, Represents the quality state mapping matrix. This represents the quality state mapping bias vector. Represents the abnormal rating weight vector. This indicates anomaly scoring bias. Represents a non-linear activation function. This represents the Sigmoid function; to avoid interference from low-quality samples in the spatiotemporal registration process, the reliability of the observed samples is corrected based on anomaly scoring and the proportion of valid observations. in, This indicates the credibility weight after quality control. Indicates the initial credibility weight. Indicates abnormal rating. This represents the proportion of valid observations; for observation samples with missing dimensions, a neighborhood sample set is constructed based on temporal distance, spatial distance, and modal consistency. in, Indicates sample The spatiotemporal neighborhood sample set, express Neighborhood observation samples, Represents the time neighborhood threshold. The threshold of the spatial neighborhood is represented; and the completion weight of the neighborhood samples to the current sample is calculated based on the neighborhood sample set; finally, the missing feature vector is completed according to the completion weight and the effective observation mask.

4. The marine anomaly monitoring and early warning method based on multi-source spatiotemporal registration and graph attention fusion according to claim 3, characterized in that, The quality control and spatiotemporal registration based on the preprocessed data also includes mapping the completed missing data observation samples to a unified time window sequence to eliminate differences in sampling frequency and observation timestamps between different observation platforms. Let the unified time window sequence be represented as: in, Represents a unified time window sequence. Indicates the first A time window, This indicates the center moment of the time window. Indicates the radius of the time window. This indicates the number of time windows; based on the distance between the observation time and the center of the time window, the synchronization weight from the sample to the time window is calculated: ,in, Indicates sample For the Synchronization weights for each time window Indicates the sample observation time. and They represent the first The and the first The center moment of a time window Represents the time synchronization scale parameters; and generates time synchronization features based on the time synchronization weights and the completed observation features: in, This represents the observational characteristics after time synchronization is complete. This represents the observed feature vector after missing data completion. Indicates the first Time index encoding for each time window This represents a vector concatenation operation. To address the inconsistencies in coordinate references, spatial resolutions, and coverage scales among different observation platforms, time-synchronized observation samples are mapped to a unified marine spatial grid. Let the unified marine grid set be: in, Represents a unified sea area grid set. Indicates the first Each marine grid unit Indicates the number of grid cells; for a sample The spatial location is transformed based on the coordinate reference parameters corresponding to its observation source attributes: in, This indicates the sample location after conversion to a unified spatial reference. Indicates the first The spatial coordinate transformation matrix of the observation source. Represents a spatial translation vector. The original spatial location of the sample is represented; the spatial registration weight of the sample to the grid cell is calculated based on the distance between the transformed spatial location and the center of the sea area grid; finally, the joint registration weight of the sample to the specified time window and grid cell is calculated based on the quality control weight, time synchronization weight and spatial registration weight; and based on the joint registration weight, the registered observation representation on the unified spatiotemporal cell is generated.

5. The marine anomaly monitoring and early warning method based on multi-source spatiotemporal registration and graph attention fusion according to claim 4, characterized in that, The spatiotemporal feature extraction based on the spatiotemporal registration results includes, to ensure that the observation results after quality control and spatiotemporal registration can enter the feature extraction process, firstly representing any registered observation result output by the previous module as follows: in, Indicates the first The first time window and the first Registration observation results corresponding to each marine grid cell Indicates the characteristics of the registration observations. Indicates a unified time window. This represents a unified marine grid cell; to organize continuous observation states under a unified spatiotemporal reference, the registration observation characteristics of the same marine grid within a continuous time window are represented as a time series matrix: in, Indicates the first The time-series observation matrix corresponding to each sea area grid cell Indicates the first Registration observation characteristics under each time window This represents the number of time windows. To separate information from different observation modes from the registered observation features, modal decoupling mapping is performed on the registered observation features. The contribution intensity of different modes in the unified features is calculated through the modal gating matrix, and the corresponding modal basic features are generated. Let the modality set be represented as: in, Represents the set of ocean observation modes. Indicates the first Observational modes, Indicate the number of modes; calculate the mode gating weights based on registration observation features: in, Indicates the first The first time window, the first In the first sea area grid cell Gating weights for observation modalities Indicates the first Modal mapping matrix, This represents the corresponding bias vector. This represents the modal scoring weight vector; finally, based on the modal gating weights, modal basis mapping is performed on the registered observation features: in, Indicates the first Observation mode in the first The first time window, the first Basic characteristics under each marine grid cell Indicates the first The feature projection matrix of the modality, This represents the feature projection bias vector. Presentation layer normalization operation.

6. The marine anomaly monitoring and early warning method based on multi-source spatiotemporal registration and graph attention fusion according to claim 5, characterized in that, The spatiotemporal feature extraction based on the spatiotemporal registration results also includes, in order to capture the local diffusion, spatial texture, and regional continuous change features of marine anomalies between adjacent sea area grids, performing local spatial convolution extraction on the modal basic features based on the adjacency relationship of the sea area grids. Let the first... The spatial neighborhood set of a sea area grid cell is represented as follows: in, Indicates the first The spatial neighborhood set of each marine grid cell Indicates the first Each marine grid unit and Representing grid cells respectively and The center coordinates, Represents the spatial neighborhood radius; and constructs modal space aggregation weights based on the spatial neighborhood set: in, Indicates the first Modal mesh elements For grid cells Spatial aggregation weights, This represents a spatial distance scale parameter. Indicates the first Spatial correlation matrix of modalities and They represent the neighboring grid cells respectively. and Modal base features; then, local spatial enhancement features are generated based on spatial aggregation weights: in, Indicates the first Class modality in the first The first time window, the first Local spatial enhancement features under each marine grid cell Indicates the first The neighborhood feature transformation matrix of the modality; furthermore, to capture the evolution trend of marine anomalies within a continuous time window, modal temporal input is constructed based on local spatial enhancement features, and temporal dependencies are extracted through gated recurrent units and a temporal attention mechanism; for the th The first sea area grid, the first Temporal input for constructing local spatial augmentation features of a modality: in, Indicates the first The first sea area grid, the first Modal spatial augmentation time series matrix, Indicates the first Local spatial augmentation features under a time window, and gated temporal recursion based on the spatial augmentation temporal matrix: in, Indicates the first Class modality in the first The first time window, the first Temporal hidden states under each sea area grid cell This indicates the hidden state of the time sequence in the previous time window. Indicates the first The gated loop unit corresponding to the modality, The parameters are represented; to highlight the key time windows that contribute significantly to the anomaly changes, the temporal attention weights are calculated based on the temporal hidden states; and modal spatiotemporal features are generated based on the temporal attention weights; finally, to enhance the expressive power of ocean anomaly features at different scales, multi-scale mapping and residual normalization are performed on each modal spatiotemporal feature, and then the multi-scale enhanced features of different modalities are organized into multi-modal spatiotemporal feature units under the same sea area grid, so as to obtain the unique output of the multi-modal spatiotemporal feature units corresponding to all sea area grids.

7. The marine anomaly monitoring and early warning method based on multi-source spatiotemporal registration and graph attention fusion according to claim 6, characterized in that, The cross-modal association fusion based on graph attention, performed on the extracted spatiotemporal features, includes converting multimodal spatiotemporal features into node representations for graph association modeling. For any sea area grid feature unit output by the previous module... It contains multi-scale enhancement features of different modalities: in, Indicates the first The multimodal spatiotemporal feature units corresponding to each sea area grid Indicates the first In the first sea area grid Modal-like multi-scale enhancement features Indicates the number of modes; to preserve the spatial location information of the sea area grid, the th mode is... Location coding is performed on the center coordinates of each sea area grid: in, Indicates the first Location encoding vectors of each sea area grid This indicates the center coordinates of the grid for that sea area. Represents the spatial frequency mapping matrix. The vector concatenation operation, based on modal features and grid position encoding, constructs the initial representation of the graph nodes: in, Indicates the first The first sea area grid, the first Initial features of graph nodes corresponding to class modalities Indicates the first Modal node feature mapping matrix, Represents the position encoding mapping matrix, This represents the node mapping bias vector. Representation layer normalization operation; organize all sea area grids and nodes corresponding to all modes into a graph node set: in, Represents the set of nodes in a multimodal graph. Indicates the first In the first sea area grid Graph nodes corresponding to class modalities This represents the initial features of the node; after processing, a multimodal node set for graph attention fusion is obtained. To express the relationships between different sea area grids and different observation modes, cross-modal association edges are established between graph nodes. For any two graph nodes... and The associated feature vector is constructed, in which spatial relationship encoding is represented by both distance decay and direction difference, and the edge connection strength between nodes is calculated based on the associated feature vector. Finally, cross-modal associated edges are selected based on the edge connection strength.

8. The marine anomaly monitoring and early warning method based on multi-source spatiotemporal registration and graph attention fusion according to claim 7, characterized in that, The cross-modal association fusion based on extracted spatiotemporal features and graph attention also includes adaptively allocating fusion weights according to the contribution of different observation nodes to the abnormal state, performing graph attention message passing on the constructed cross-modal graph structure, and for the target node... Define its graph neighborhood node set as: in, Represents the target node The set of graph neighborhood nodes, This represents a neighboring node that has an edge connected to the target node. This represents the set of cross-modal associated edges; to calculate the contribution of neighboring nodes to the target node, a graph attention score is constructed: in, Representing neighboring nodes For the target node Graph attention weights, Represents the query mapping matrix. Represents the key mapping matrix, Represents the attention score vector. Indicates the edge connection strength. This represents a linear activation function with leakage correction, and the node update representation is obtained by aggregating neighborhood messages based on graph attention weights: in, Represents a node Updated features after graph attention fusion Indicates the initial characteristics of the target node. Indicates the first The value mapping matrix corresponding to the neighborhood mode. This represents the features of neighboring nodes; to improve the expressive power of cross-modal information propagation, multi-head graph attention is further used for fusion: in, This represents the node features after attention fusion in a multi-head graph. Indicates the first The node update features are obtained from the attention head. Indicates the number of attention heads. This represents the multi-head fusion mapping matrix. This represents the multi-head fusion bias vector; to transform the node features after graph attention fusion into a graph representation for anomaly detection, for the same sea area grid... Modality aggregation weights are calculated based on the fusion features of nodes from different modalities: in, Indicates the first In the first sea area grid Modal aggregation weights of modal nodes Represents the modality aggregation mapping matrix. This represents the modal aggregation bias vector. Represent the modal aggregation score vector; and generate a grid-level anomaly state representation of the sea area based on the modal aggregation weights: in, Indicates the first Characterization of anomalous states in a sea area grid Represents modal aggregation weights. The modal node features after graph attention fusion are represented, and the set of abnormal state nodes after graph attention fusion is further constructed: in, Represents a set of nodes in an abnormal state. Indicates the first Each sea area grid corresponds to an abnormal state node. Represents a marine grid cell. This represents the abnormal state characterization of the grid; finally, the set of abnormal state nodes, the set of cross-modal associated edges, and the node-level fusion features are organized together as the unique output of this module.

9. A marine anomaly monitoring and early warning method based on multi-source spatiotemporal registration and graph attention fusion as described in claim 8, characterized in that, The step of identifying marine anomalies based on the fused features includes representing the anomaly state diagram output by the previous module as follows: in, The diagram represents anomalies in the ocean. Represents a set of nodes in an abnormal state. Represents the set of cross-modal associated edges. Indicates the first Sea area grid, the first Graph attention fusion of node features corresponding to class modalities Indicates the number of grid cells in the sea area. This represents the number of modalities; for any node in the set of abnormal state nodes, it is represented as: in, Indicates the first Each sea area grid corresponds to an abnormal state node. Represents a marine grid cell. This represents the anomalous state characterization of the grid; to determine whether an anomalous state exists in each sea area grid, the node-level anomaly probability is calculated based on the anomalous state characterization: in, Indicates the first Node-level anomaly probability of a sea area grid Represents the anomaly detection mapping matrix. This represents the anomaly detection mapping bias vector. This represents the anomaly probability score vector. This represents the anomaly probability scoring bias term. This represents the Sigmoid function and filters the set of candidate nodes for anomalous nodes based on node-level anomaly probabilities: in, Represents the set of candidate nodes for anomalies. This represents the threshold for node anomaly detection; to avoid outputting warning results only as discrete grid points, for any two candidate anomaly nodes... and Build regional connectivity strength: in, Indicates abnormal candidate nodes and The strength of the regional connection between them Indicates an indicator function, and Representing the sea area grid and The center coordinates, Indicates the spatial connectivity scale parameter. This represents the set of cross-modal associated edges from the fourth module. The average inter-mesh connectivity strength obtained through aggregation. , and This represents the weighting coefficient; where the average inter-mesh connectivity strength is expressed as: in, Represents the sea area grid and Average plot connection strength between them Indicates the first The first grid Modal nodes and the first The first grid Associative edges between modal nodes Indicates the connection strength of the corresponding edges. This prevents extremely small positive numbers with a denominator of zero, and constructs a set of anomalous candidate regions based on region connectivity strength: in, Represents the set of candidate regions for anomalies. Indicates the first One abnormal candidate region, Indicates by The resulting region connectivity strength matrix Indicates the region connectivity threshold. Indicates the number of abnormal candidate regions. This represents a connected component partitioning operation based on connection strength. To determine the event category of anomaly candidate regions, the node state representation, anomaly probability, and graph connection strength within each anomaly candidate region are weighted and aggregated to form a region-level event representation. For the ... Anomalous candidate regions Calculate the node aggregation weights within the region; generate a regional event representation based on the node aggregation weights within the region; and calculate the probability of abnormal event types based on the regional event representations.

10. A marine anomaly monitoring and early warning system based on multi-source spatiotemporal registration and graph attention fusion, executing the marine anomaly monitoring and early warning method based on multi-source spatiotemporal registration and graph attention fusion as described in claim 1, characterized in that, include: The data acquisition module is configured to acquire multi-source ocean observation data; The preprocessing module is configured to perform data preprocessing based on the acquired multi-source ocean observation data; The spatiotemporal registration module is configured to perform quality control and spatiotemporal registration based on the preprocessed data. The feature extraction module is configured to perform spatiotemporal feature extraction based on the spatiotemporal registration results; The fusion module is configured to perform cross-modal association fusion based on graph attention, based on the extracted spatiotemporal features; The anomaly identification module is configured to identify marine anomalies based on the fused features.

Images