Fusion construction method and system for fishing multi-source data

Abstract

The invention discloses a fishery harvesting multi-source data fusion construction method and system, and relates to the technical field of data fusion, and the method comprises the steps: carrying out the space-time alignment and preprocessing of the heterogeneous data of yield, fishing boats, investigation, environment, social economy and the like, and constructing a graph model comprising multiple types of nodes and multiple relation edges; the two graphs are subjected to weighted fusion based on the fusion weight to obtain a fusion structure graph, so that ordered synthesis of multi-scale information can be realized, and the integrity and robustness of graph representation are improved; on this basis, a node layer and edge layer dual attention mechanism is introduced, adaptive weighted fusion of multi-source information is realized, a high-completeness sea area-time fusion feature data set is generated and output, a fusion structure diagram is input into a fusion network containing a node layer attention module and an edge layer attention module for end-to-end fusion, and a sea area-time fusion feature data set is generated and output. Therefore, more accurate and comprehensive fishery multi-source fusion feature data are obtained, and the accuracy and availability of fishery resource evaluation are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data fusion, and in particular to a fishing multi-source data fusion construction method and system. BACKGROUND

[0002] In the assessment and management of fishery resources, accurate and comprehensive data is the basis for scientific decision-making. However, fishing data is usually derived from multiple heterogeneous channels, such as yield statistics, fishing vessel trajectories, environmental monitoring, scientific surveys, and social and economic factors. These data differ significantly in terms of spatial and temporal scales, resolution, and semantics, leading to information fragmentation, inconsistency, or local bias. If these data are simply associated or spliced, the inherent spatial and temporal, semantic, and scale differences cannot be fully reflected. On the one hand, there are significant inconsistencies in spatial resolution, temporal scale, and sampling frequency among different data sources, and simple association can easily cause information mismatch or statistical bias. On the other hand, the semantic dimensions of data from different sources differ significantly. For example, fishing vessel trajectories reflect fishing behavior, yield data reflect the result state, environmental data reveal driving factors, and survey data reflect local true conditions. If these data are not fused before being associated, it is difficult to establish a clear causal or logical relationship. In addition, multi-source data have noise, missing, and offset problems, and direct association can amplify the interference of low-quality data and weaken the reliability of the overall data set. Therefore, it is urgent to fuse fishing multi-source data to integrate complementary information, eliminate data islands, and construct a comprehensive and highly usable data set. SUMMARY

[0003] The present application provides a fishing multi-source data fusion construction method and system to fuse fishing data from multiple sources and improve the comprehensiveness and usability of fishing data.

[0004] To solve the above technical problems, the present application provides a fishing multi-source data fusion construction method, comprising: Obtaining yield data, fishing vessel trajectory data, environmental data, survey data, and social data of a target area, and performing spatio-temporal alignment on the yield data, fishing vessel trajectory data, environmental data, survey data, and social data based on a preset sea area grid and a preset time window to obtain an initial heterogeneous data set; Dividing the initial heterogeneous data set based on a preset coarse grid resolution and a preset fine grid resolution to obtain a coarse grid multi-source data subset and a fine grid multi-source data subset; For the coarse grid multi-source data subset and the fine grid multi-source data subset, taking a sea area node, a time node, a fishing vessel node, an environmental factor node, and a species node as nodes, calculating the association between each node, and determining the association edge between each node based on the association to construct a first graph structure corresponding to the coarse grid multi-source data subset and a second graph structure corresponding to the fine grid multi-source data subset. calculate a structure similarity of the first graph structure and the second graph structure based on a preset similarity calculation model, and determine a fusion weight based on the structure similarity, and perform weighted fusion on the first graph structure and the second graph structure based on the fusion weight to obtain a fusion structure graph; construct a fusion network based on a node layer attention module and an edge layer attention module, input the fusion structure graph into the fusion network for fusion to obtain fishery catch fusion data of a target area.

[0005] The present application can ensure the comparability and aggregability of multi-source data in the spatial and temporal dimensions by obtaining yield data, fishing vessel trajectory data, environmental data, survey data and social data of a target area and performing spatio-temporal alignment based on a preset sea area grid and a time window, thereby providing high-quality original input for subsequent unified modeling; then, the initial heterogeneous data set is divided based on coarse grid and fine grid resolutions, which can simultaneously consider macro coverage and micro details to meet the needs of different scale analysis and cross-scale fusion; at the same time, a graph structure is constructed with sea area nodes, time nodes, fishing vessel nodes, environmental factor nodes and species nodes, and the node correlation is calculated, which can semantically represent heterogeneous information as graph topology and attributes, so that geographical, temporal, subject and biological information can be interacted in a unified graph domain, thereby obtaining the structure similarity of two scale graph structures based on a preset similarity calculation model and determining the fusion weight, which can provide an objective measurement standard for cross-scale information weighting, avoid single scale bias, and when the two graph structures reflected by the coarse and fine grids are highly similar, it means that the macro trend and micro details are consistent, and the fusion can enhance each other; when the similarity is low, it may indicate that there are local abnormalities or data uncertainties, and the weight of some information can be automatically reduced during fusion, so that the two graphs are weighted and fused based on the fusion weight to obtain a fusion structure graph, which can realize the ordered synthesis of multi-scale information and improve the integrity and robustness of the graph representation; the fusion structure graph is input into a fusion network containing a node layer attention module and an edge layer attention module for end-to-end fusion, which can adaptively allocate attention on the node semantics and edge relationship levels to enhance the expression ability of important information, thereby obtaining more accurate and comprehensive fishery catch multi-source fusion feature data and improving the precision and usability of fishery resource assessment.

[0006] Further, for the coarse grid multi-source data subset and the fine grid multi-source data subset, the correlation between each node is calculated based on the sea area nodes, time nodes, fishing vessel nodes, environmental factor nodes and species nodes, and the associated edges between each node are determined based on the correlation, and the first graph structure corresponding to the coarse grid multi-source data subset and the second graph structure corresponding to the fine grid multi-source data subset are constructed, including: obtain the first grid scale and the second grid scale of the coarse grid multi-source data subset and the fine grid multi-source data subset; generate sea area nodes, time nodes, fishing boat nodes, environmental factor nodes and species nodes based on the first grid scale and the second grid scale respectively, wherein each node comprises a unique identifier and a set of attributes corresponding to the node type thereof; For any two nodes, calculate the correlation degree between the nodes based on a preset semantic correlation rule, and determine the correlation edge between the nodes based on the correlation degree and a preset generation strategy. Based on each node and the correlation edge between the nodes, construct a first graph structure corresponding to the coarse grid multi-source data subset and a second graph structure corresponding to the fine grid multi-source data subset.

[0007] By obtaining the first grid scale and the second grid scale of the coarse grid and the fine grid, the application is beneficial to determining the data resolution and the spatial coverage boundary under different scales, and generating multi-type nodes based on the two grid scales respectively and configuring each node with a unique identifier and a set of attributes can ensure the traceability and attribute completeness of the nodes, facilitating cross-table association and subsequent audit. Calculating the correlation degree between any two nodes based on a preset semantic correlation rule and determining the correlation edge based thereon makes the generation of the edge interpretable and controllable, reduces the risk of introducing noise edges, and further, based on the nodes and the correlation edge, constructing a corresponding first graph structure and a second graph structure can structure the multi-source information within the same scale into graph data.

[0008] Further, the calculation of the correlation degree between the nodes based on the semantic correlation rule and the determination of the correlation edge between the nodes based on the correlation degree and the preset generation strategy comprise: For any two nodes, determine the node type, and match the semantic correlation rule based on the node type; Calculate the correlation degree between the nodes based on the semantic correlation rule, and determine the correlation edge between the nodes based on the correlation degree and the preset generation strategy.

[0009] The application first determines the node type of any two nodes and matches the semantic correlation rule based thereon, which can make the correlation calculation adopt a targeted rule set, improve the semantic relevance and accuracy of the correlation degree calculation, and further, calculate the correlation degree between the nodes based on the matched semantic correlation rule and determine the correlation edge based on the generation strategy, which can ensure that the constructed edge reflects semantic correlation and is constrained by the generation strategy, thereby balancing between accuracy and sparsity and improving the representation quality and calculation efficiency of the graph structure.

[0010] Further, the calculation of the correlation degree between the nodes based on the semantic correlation rule and the determination of the correlation edge between the nodes based on the correlation degree and the preset generation strategy comprise: when the node type is a fishing boat node and a sea area node, calculating a spatial intersection ratio based on a fishing boat trajectory and a sea area boundary, and determining the correlation degree based on the spatial intersection ratio; when the node type is a time node and a sea area node, calculating the correlation degree based on a fishing boat activity count, a catch yield, and a survey sample number within the time node; when the node type is an environmental factor node and a time node, calculating the correlation degree based on an aggregated statistical feature of the environmental data within the time window; when the node type is a species node and any other node, calculating the correlation degree based on a species appearance frequency, a species quantity, and a classification confidence measure.

[0011] When the node type is a fishing boat node and a sea area node, the correlation degree is determined based on the spatial intersection ratio of the trajectory and the sea area boundary, which can accurately reflect the spatial relationship between the actual operation site of the fishing boat and the sea area, and improve the reliability of the spatial attribution judgment. When the node type is a time node and a sea area node, the correlation degree is calculated based on the fishing boat activity count, the catch yield, and the survey sample number within the time window, which can convert the time series activity intensity into the temporal correlation degree within the sea area, facilitating the capture of seasonal or short-term fluctuation signals. When the node type is an environmental factor node and a time node, the correlation degree is calculated based on the aggregated statistical feature of the environmental data within the time window, which can quantify the time series feature of the environmental condition into the correlation with the time window, facilitating the revelation of the environmental-driven catch change. When the node type is a species node and any other node, the correlation degree is calculated based on the species appearance frequency, the quantity, and the classification confidence measure, which can ensure that the correlation degree of the species information in the spatial, temporal, and sample layers has biological significance and data quality guarantee, thereby enhancing the credibility of the species-related inference.

[0012] Further, a structure similarity of the first graph structure and the second graph structure is calculated based on a preset similarity calculation model, a fusion weight is determined based on the structure similarity, the first graph structure and the second graph structure are weighted and fused based on the fusion weight, and a fused structure graph is obtained, including: calculating the structure similarity of the first graph structure and the second graph structure based on a preset similarity calculation model, the structure similarity including a node similarity, a topology similarity, a relationship consistency, and a scale correspondence; correcting the node similarity, the topology similarity, the relationship consistency, and the scale correspondence based on a preset confidence degree, and obtaining a similarity index; mapping the similarity index to a single structure similarity score based on a preset fusion rule, and determining a fusion weight based on the similarity score; weightedly fusing the first graph structure and the second graph structure based on the fusion weight, and obtaining a fused structure graph.

[0013] The application calculates the structural similarity from multiple dimensions of node similarity, topological similarity, relationship consistency and scale correspondence based on a preset similarity calculation model, can comprehensively reflect the consistency of two structure diagrams in semantic, structure and scale mapping, and avoid one-sided judgment caused by a single index; the above similarity indexes are modified according to a preset confidence, the data quality and observation confidence can be included in the similarity evaluation, and the robustness of the similarity measurement is improved; the modified similarity indexes are mapped into a single structure similarity score according to a preset fusion rule, and the fusion weight is determined accordingly, the multi-dimensional information can be summarized as an operable fusion decision basis, which is convenient for actual weighting operation; the two graphs are weighted and fused based on the fusion weight to obtain a fusion structure diagram, which can reasonably integrate the cross-scale information according to its relative confidence and similarity, so that a fusion graph representation considering the consistency of details and the whole is obtained.

[0014] Further, the weighted fusion of the first graph structure and the second graph structure based on the fusion weight to obtain a fusion structure diagram comprises: obtaining the node mapping relationship of the first graph structure and the second graph structure based on the unique identifier of each node, the spatial inclusion relationship or the time window coincidence degree; fusing the first graph structure and the second graph structure based on the node mapping relationship, the preset initial edge weight and the fusion weight to obtain a fusion structure diagram.

[0015] The application obtains the node mapping relationship of two graphs based on the unique identifier of each node, the spatial inclusion relationship or the time window coincidence degree, can realize node alignment in a clear priority manner, retains both explicit identification direct mapping and supports mapping based on space-time relationship, and ensures that cross-graph elements can be correctly paired; fusing two graphs based on the node mapping relationship, the preset initial edge weight and the fusion weight can introduce weighted merging at the edge level, which comes from both original measurement and is constrained by global fusion strategy, so that the weight of the merged edge reflects both original observation and cross-scale consistency, thereby generating a fusion structure diagram which is structured, interpretable and suitable for subsequent graph learning.

[0016] Further, the fusion network is constructed based on the node layer attention module and the edge layer attention module, the fusion structure diagram is input into the fusion network for fusion, and the fishing yield fusion data of the target area is obtained, which comprises: traversing the fusion structure diagram, calculating multi-head attention coefficients for different categories of neighbor nodes of the current node based on the node layer attention module, and weighting and aggregating the nodes in the fusion structure diagram based on the multi-head attention coefficients; calculating the contribution weight of each associated edge of the current node based on the edge layer attention module, and weighting and fusing the associated edges in the fusion structure diagram based on the contribution weight.

[0017] The present application can make the model adaptively allocate the information weight from different neighborhood categories when aggregating nodes by traversing the fusion structure graph and calculating multi-head attention coefficients for different categories of neighbor nodes based on the node layer attention module, thereby highlighting the neighbor information that contributes more to target prediction and suppressing irrelevant noise; further, based on the edge layer attention module, the contribution weight of different associated edges is calculated and the associated edges are weighted and fused accordingly, which can dynamically identify and strengthen the key relationship path at the edge relationship level, improve the capture ability of the graph representation for complex causal and co-occurrence relationships, and thereby enhance the discrimination ability of the fusion network for multi-source information of sea area, time, fishing boat, environment and species and the prediction performance of downstream tasks through the combination of the two.

[0018] In a second aspect, the present application provides a fishing yield multi-source data fusion construction system, comprising: a preprocessing module, a grid division module, a graph construction module, a fusion module and an identification module; The preprocessing module is configured to obtain yield data, fishing boat trajectory data, environmental data, survey data and social data of a target area, and perform spatio-temporal alignment on the yield data, fishing boat trajectory data, environmental data, survey data and social data based on a preset sea area grid and a preset time window to obtain an initial heterogeneous data set. The grid division module is configured to divide the initial heterogeneous data set based on a preset coarse grid resolution and a preset fine grid resolution to obtain a coarse grid multi-source data subset and a fine grid multi-source data subset. The graph construction module is configured to, for the coarse grid multi-source data subset and the fine grid multi-source data subset, take a sea area node, a time node, a fishing boat node, an environmental factor node and a species node as nodes, calculate the association between each node, and determine the associated edges between each node based on the association to construct a first graph structure corresponding to the coarse grid multi-source data subset and a second graph structure corresponding to the fine grid multi-source data subset. The fusion module is configured to calculate the structural similarity of the first graph structure and the second graph structure based on a preset similarity calculation model, determine a fusion weight based on the structural similarity, and perform weighted fusion on the first graph structure and the second graph structure based on the fusion weight to obtain a fusion structure graph. The identification module is configured to construct a fusion network based on a node layer attention module and an edge layer attention module, input the fusion structure graph into the fusion network for fusion, and obtain fishing yield fusion data of the target area.

[0019] Further, the graph construction module is configured to, for the coarse-grid multi-source data subset and the fine-grid multi-source data subset, respectively take a sea area node, a time node, a fishing boat node, an environmental factor node and a species node as a node, calculate the relevance between each node, and determine the association edge between each node based on the relevance, and construct the first graph structure corresponding to the coarse-grid multi-source data subset and the second graph structure corresponding to the fine-grid multi-source data subset, including: obtaining a first grid scale and a second grid scale of the coarse-grid multi-source data subset and the fine-grid multi-source data subset; generating a sea area node, a time node, a fishing boat node, an environmental factor node and a species node based on the first grid scale and the second grid scale respectively, wherein each node includes a unique identifier and an attribute set corresponding to the node type thereof; for any two nodes, calculating the relevance between the nodes based on a preset semantic association rule, and determining the association edge between the nodes based on the relevance and a preset generation strategy; constructing the first graph structure corresponding to the coarse-grid multi-source data subset and the second graph structure corresponding to the fine-grid multi-source data subset based on each node and the association edge between the nodes.

[0020] Further, the graph construction module is configured to, for any two nodes, calculate the relevance between the nodes based on a preset semantic association rule, and determine the association edge between the nodes based on the relevance and a preset generation strategy, including: for any two nodes, determining the node type, and matching the semantic association rule based on the node type; calculating the relevance between the nodes based on the semantic association rule, and determining the association edge between the nodes based on the relevance and a preset generation strategy. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 a flowchart of a fishing catch multi-source data fusion construction method provided by an embodiment of the present application; Figure 2 a structural diagram of a fishing catch multi-source data fusion construction system provided by an embodiment of the present application. DETAILED DESCRIPTION

[0022] The specific embodiments of the present application will be further described in detail below with reference to the accompanying drawings and embodiments. The following embodiments are used to illustrate the present application, but are not used to limit the scope of the present application.

[0023] The terms "first" and "second" and the like in the description and in the claims of the present application and the drawings are used to distinguish between similar objects, not to describe a particular sequential order. Moreover, the terms "include," and "have," and any variations thereof, are intended to cover a non-exclusive inclusion. For example, a process, method, system, product, or device that comprises a list of steps or units is not necessarily limited to the listed steps or units, but can optionally include additional steps or units not expressly listed or can also include steps or units inherent to such process, method, product, or device.

[0024] Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the application. The appearances of the phrase in various places in the specification are not necessarily all referring to the same embodiment, nor are they necessarily mutually exclusive or alternative embodiments. It is expressly understood that the embodiments described herein are combinable with each other.

[0025] Embodiment 1 Reference Figure 1 , Figure 1 A flowchart diagram of a fishing multi-source data fusion construction method provided by an embodiment of the application. The embodiment of the application provides a fishing multi-source data fusion construction method, which comprises steps 101 to 105, and specifically as follows: Step 101: Obtain yield data, fishing vessel trajectory data, environmental data, survey data and social data of a target area, and perform spatio-temporal alignment on the yield data, fishing vessel trajectory data, environmental data, survey data and social data based on a preset sea area grid and a preset time window to obtain an initial heterogeneous data set; In this embodiment, multi-source original data such as yield data, fishing vessel trajectory data, environmental observation data, survey sample data and social and economic data of a target area are obtained, and the multi-source original data such as yield data, fishing vessel trajectory data, environmental observation data, survey sample data and social and economic data are spatio-temporally aligned by a preset sea area grid and a preset time window to obtain an initial heterogeneous data set for subsequent construction of a graph structure.

[0026] In this embodiment, the yield data is the weighing and reporting record at a fishing port or on a ship, including fishing time, fishing location, species name and weight, etc.; the fishing vessel trajectory data is a ship positioning sequence, including a time point and a latitude-longitude sequence; the environmental data is from an observation station or a remote sensing product, including observation time, observation location and environmental factor values; the survey data is a scientific or sampling survey record, including sample time, sample point location, species category and quantity; and the social and economic data is aggregated by region or period and includes fishing yield price, management measures and other information.

[0027] In this embodiment, the unified time reference is converted to the same time zone representation, and each record is mapped to the corresponding time period according to the preset time window, which can be set to several hours or day windows; at the same time, based on the preset spatial grid system, a coarse grid and a fine grid are established respectively, and the point data with latitude and longitude are mapped to the corresponding grid unit according to the point in the polygon, the number of trajectory points, the number of entries and the residence time of the trajectory data are counted in each grid according to the trajectory point, and the spatial intersection ratio of the trajectory and the grid is calculated.

[0028] In this embodiment, the spatial range and time range of the target area are first determined, the target area can be a specific sea area, fishing ground or management unit, and the time range can be several months or several fishing seasons. Access to multi-source channels including yield data, fishing vessel trajectory data, environmental data, survey data and social data, etc.

[0029] Among them, the yield data comes from the landing declaration system of the fishing port or the shipborne monitoring terminal, records the fishing boat number, fishing time, fishing location, species category and yield information; the fishing vessel trajectory data comes from the Beidou or automatic identification system equipment, contains time, latitude and longitude, speed, heading and other information, and is used to reflect the operation trajectory and activity area of the fishing vessel; the environmental data comes from the ocean observation station, buoy or remote sensing monitoring system, including water temperature, salinity, dissolved oxygen, chlorophyll concentration and other environmental factors; the survey data comes from scientific research or sampling survey tasks, records the sampling point location, time, species composition and quantity; the social data comes from the fishery management department or statistical agency, including fishery input-output, price index, policy measures and social and economic conditions, etc.

[0030] In this embodiment, the format, coordinate system and time reference of various types of data are unified, all data time fields are converted to a unified time standard, and spatial discretization is performed based on a preset sea area grid. The sea area grid divides the target sea area into a plurality of regular or irregular grid units, each grid unit has a unique identifier, and is used to carry multi-source data in the region. The time window is used for time segmentation and aggregation of data at different times, for example, a sliding window with hour, day or week as a unit can be set.

[0031] In this embodiment, in the alignment process, all records containing location information are mapped to the corresponding sea area grid according to their latitude and longitude coordinates, and the time information is aligned to the corresponding time window, so as to realize the unification of space and time. For missing or abnormal data, the system is based on statistical rules or time series interpolation method to complete or correct, for example, through adjacent period mean, adjacent grid average or data source reliability to correct, to ensure the continuity and rationality of the data.

[0032] In this embodiment, within each "sea area grid and time window" unit, the yield, fishing vessel activity, environmental factors, survey results and socio-economic indicators are aggregated respectively to form a unified data record containing multiple types of fields. Each record is accompanied by data source identification, time label and quality mark to ensure data traceability and reliability.

[0033] Step 102: based on the preset coarse grid resolution and fine grid resolution, the initial heterogeneous data set is divided respectively to obtain a coarse grid multi-source data subset and a fine grid multi-source data subset; In this embodiment, the initial heterogeneous data set that has been spatio-temporally aligned is divided into a coarse grid multi-source data subset and a fine grid multi-source data subset according to different spatial resolutions to support multi-scale feature extraction and graph structure construction. First, based on the range of the target sea area, management scale and data density, two spatial resolutions are preset: one is coarse grid resolution, which is used to describe the fishing distribution characteristics on a macro scale; the other is fine grid resolution, which is used to depict the fine differences and dynamic changes in the local area. Coarse grid is generally used to reflect large-scale fishing activities, environmental characteristics and socio-economic factors, while fine grid is used to accurately analyze the yield changes, fishing vessel distribution and ecological response of local fishing grounds.

[0034] In this embodiment, the latitude and longitude range of the target sea area can be divided into regular squares or hexagonal units according to the set interval. The unit spacing of the coarse grid can be set to several kilometers to several tens of kilometers, and the unit spacing of the fine grid can be set to several hundred meters to several kilometers. The specific resolution can be adjusted according to the monitoring accuracy and computing power. Each grid unit is assigned a unique spatial identifier for subsequent indexing and data fusion.

[0035] In this embodiment, all data records in the initial heterogeneous data set, including yield data, fishing vessel trajectory data, environmental data, survey data and social data, are classified into corresponding coarse and fine grids according to their geographic coordinates. For trajectory data, the system determines the grid unit based on the latitude and longitude coordinates of the trajectory points and records the number, duration or activity frequency of the trajectory points in the grid. For regional environmental observation data, the observation station location or remote sensing pixel center location is projected into the corresponding grid to calculate the environmental mean or statistical characteristics within the grid range. For survey data and yield data, the sampling points or fishing locations are assigned to the corresponding grids. Social and economic data are allocated to matching grid units according to the spatial overlap relationship of administrative regions or fishing management units.

[0036] In the embodiment, a coarse grid multi-source data subset is generated at a coarse grid level for representing the spatio-temporal variation trend of the whole sea area, and a fine grid multi-source data subset is generated at a fine grid level for analyzing the high-resolution characteristics of a local sea area. Each subset contains aggregated attributes of multi-source data, such as total yield, fishing vessel activity intensity, environmental index average, survey sample quantity, and social and economic indicators, and is accompanied by data source, time window, and quality flag information.

[0037] In step 103, for the coarse grid multi-source data subset and the fine grid multi-source data subset, the correlation between nodes is calculated respectively with sea area nodes, time nodes, fishing vessel nodes, environmental factor nodes, and species nodes as nodes, and the correlation edges between nodes are determined based on the correlation, and the first graph structure corresponding to the coarse grid multi-source data subset and the second graph structure corresponding to the fine grid multi-source data subset are constructed. In the embodiment, for the coarse grid multi-source data subset and the fine grid multi-source data subset, the correlation between nodes is calculated respectively with sea area nodes, time nodes, fishing vessel nodes, environmental factor nodes, and species nodes as nodes, and the correlation edges between nodes are determined based on the correlation, and the first graph structure corresponding to the coarse grid multi-source data subset and the second graph structure corresponding to the fine grid multi-source data subset are constructed, including: The first grid scale and the second grid scale of the coarse grid multi-source data subset and the fine grid multi-source data subset are obtained. The sea area nodes, the time nodes, the fishing vessel nodes, the environmental factor nodes, and the species nodes are generated based on the first grid scale and the second grid scale respectively, wherein each node includes a unique identifier and a set of attributes corresponding to the node type. For any two nodes, the correlation degree between the nodes is calculated based on a preset semantic correlation rule, and the correlation edge between the nodes is determined based on the correlation degree and a preset generation strategy. The first graph structure corresponding to the coarse grid multi-source data subset and the second graph structure corresponding to the fine grid multi-source data subset are constructed based on the nodes and the correlation edges between the nodes.

[0038] In the embodiment, based on the spatial division result of the initial data set, the first grid scale of the coarse grid multi-source data subset and the second grid scale of the fine grid multi-source data subset are determined. The first grid scale usually corresponds to a larger spatial range and a longer time span, and is used to represent the macroscopic fishing activity pattern; the second grid scale corresponds to finer spatial and temporal resolution, and is used to depict local operation characteristics and short-term dynamic changes.

[0039] In this embodiment, after determining the grid scale, the system generates five types of nodes based on each grid cell and the time window, including sea area nodes, time nodes, fishing boat nodes, environmental factor nodes, and species nodes. Among them, the sea area node is used to represent the spatial location characteristics, including grid number, latitude and longitude range, sea area area, etc. attributes; the time node is used to represent the time characteristics, including time period number, start and end time, season information, etc. attributes; the fishing boat node is used to represent the production subject characteristics, including the fishing boat number, operation type, power, tonnage, sea time, track range, etc. attributes; the environmental factor node is used to represent the external ecological environment conditions, including sea surface temperature, salinity, dissolved oxygen, wind speed, flow rate, etc. attributes; the species node is used to represent the catch composition, including species name, body length, body weight, catch quantity, habitat depth, etc. attributes. Each node is provided with a unique identifier for indexing and association in the graph structure.

[0040] In this embodiment, for any two nodes, the system calculates the correlation degree between the nodes based on the preset semantic association rule. The semantic association rule can be defined according to the similarity of node type and attribute characteristics, spatial proximity, temporal overlap, behavior co-occurrence, or ecological coupling relationship.

[0041] In this embodiment, for any two nodes, the correlation degree between the nodes is calculated based on the preset semantic association rule, and the associated edges between the nodes are determined based on the correlation degree and the preset generation strategy, including: For any two nodes, determine the node type, and match the semantic association rule based on the node type; Calculate the correlation degree between the nodes based on the semantic association rule, and determine the associated edges between the nodes based on the correlation degree and the preset generation strategy.

[0042] In this embodiment, in order to accurately depict the semantic relationship between different types of nodes, the system performs pairwise association analysis on each type of node included in the coarse grid and fine grid multi-source data subsets to determine the semantic correlation degree between the nodes, and generates associated edges accordingly, thereby forming a graph structure with clear semantic expression ability. Specifically, the system first traverses all node pairs, and for any two nodes to be analyzed, first determines the node type. The node type includes sea area node, time node, fishing boat node, environmental factor node and species node, and different types of nodes have different semantic relationships. For example, the sea area node and the fishing boat node represent the spatial operation relationship, the fishing boat node and the time node represent the operation time sequence relationship, and the environmental factor node and the species node represent the ecological response relationship.

[0043] In this embodiment, after determining the node type, the system matches the corresponding semantic association rule based on the node type. The semantic association rule is pre-configured according to the combination of node types, and different rules are adopted for different types of node pairs.

[0044] In this embodiment, the association degree between nodes is calculated based on the semantic association rule, and the associated edges between nodes are determined based on the association degree and a preset generation strategy, including: When the node type is a fishing boat node and a sea area node, the spatial intersection ratio is calculated based on the fishing boat trajectory and the sea area boundary, and the association degree is determined based on the spatial intersection ratio; When the node type is a time node and a sea area node, the association degree is calculated based on the fishing boat activity count, the catch yield, and the survey sample number in the time node; When the node type is an environmental factor node and a time node, the association degree is calculated based on the aggregated statistical features of the environmental data within the time window; When the node type is a species node and any other node, the association degree is calculated based on the species appearance frequency, the species quantity, and the classification confidence measure.

[0045] In this embodiment, in order to construct a graph structure that can reflect the multi-dimensional relationships of the fishery ecosystem, the system formulates multiple sets of semantic association rules for the semantic feature differences between different types of nodes to calculate the association degree between nodes and generate the associated edges between nodes, thereby establishing a unified data expression structure in the spatial, temporal, environmental, and species dimensions. First, obtain each type of node in the multi-source data subsets divided by coarse and fine grids, and extract the attribute information corresponding to each node, such as the fishing boat trajectory point column, the sea area boundary range, the time window definition, the environmental element sequence, and the species identification record. Subsequently, for any two nodes, match the corresponding semantic association rule according to the node type and perform association degree calculation.

[0046] In the present embodiment, when the node type is a fishing boat node and a sea area node, the calculation is based on the spatial relationship between the fishing boat trajectory data and the sea area grid boundary. Specifically, first, the trajectory point set of the fishing boat within a specified time window is extracted, and the trajectory point is spatially superimposed and analyzed with the sea area boundary to calculate the residence ratio or passing ratio of the trajectory in the sea area, and the spatial intersection ratio is obtained. The higher the ratio, the more frequent the fishing boat's operation activities in the sea area. Accordingly, the correlation degree between the two nodes is determined, which is used to reflect the operation intensity relationship between the fishing boat and the sea area. Since the spatial intersection ratio of the fishing boat trajectory and the sea area directly quantifies the "share" of the fishing boat's activities in the sea area, including passing, wandering and staying, etc., it can better reflect the possibility and contribution of the fishing boat to the fishery catch in the sea area, and the intersection ratio can be measured in the form of "the proportion of the time spent in the sea area to the total time" or "the proportion of the length of the trajectory in the sea area to the total length", which equivalently introduces time and distance factors into the correlation degree calculation, thereby reflecting the operation persistence rather than one-time passing. Thus, it helps to distinguish between substantive operation and transit behavior. That is, the fishing boat may pass through a sea area without operation (high-speed passing) or stay in the sea area at low speed (actual operation), and simply relying on whether a point appears can lead to misjudgment. By calculating the intersection and combining information such as speed and heading (for example, only including low-speed / suspected operation period in the intersection), it can effectively suppress the transit interference and improve the semantic accuracy of the correlation degree.

[0047] In the present embodiment, when the node type is a time node and a sea area node, the fishing activity level in the sea area within the time period is analyzed from the time dimension. Specifically, the number of fishing boat activities, catch yield, and survey sample size in the sea area within the corresponding time window are counted, and the above indicators are standardized or weighted summed to obtain a comprehensive activity intensity value. The activity intensity value is used as the correlation degree between the two nodes, which is used to represent the fishing activity level of the sea area in the time period. First, the number of fishing boat activities, catch yield, and survey sample size, which are three types of indicators, reflect the fishing activity situation of the space-time unit from three complementary dimensions of behavior intensity, output result, and field verification, respectively. The number of fishing boat activities directly represents the operation frequency and effort of the fishing boat in the sea area within the time period, which is the first-hand signal for measuring fishing pressure and potential fishing opportunities; the catch yield is the final output measure, which can reflect the actual resource utilization and output level of the sea area in the time period; and the number of survey samples provides gold standard information for calibration and verification as high-confidence observations from field sampling, which helps to correct the bias of reported or remotely estimated data. Second, the combination of the three can compensate for the shortcomings of each other: simply counting the number of operations may misjudge high-speed transit as operation, simply reporting the yield may be affected by underreporting or delay, and simply using survey data cannot represent the whole due to sparse coverage; the combination of the three can reduce noise and bias, and improve the robustness of the correlation degree.

[0048] In this embodiment, when the node type is an environmental factor node and a time node, the correlation degree is calculated based on the change characteristics of the environmental monitoring data within a specified time window. Specifically, the environmental data such as sea surface temperature, salinity, chlorophyll concentration, dissolved oxygen, etc. are aggregated and statistically analyzed within the time window to obtain mean, variance or change trend, etc. characteristic indexes, reflecting the stability and dynamics of the environmental factors in this time period. According to the significance and fluctuation amplitude of these statistical characteristics, the correlation degree of the environmental factor and the time node is determined, thus reflecting the correlation between the environmental conditions and the time change. When the node type is an environmental factor node and a time node, the change characteristics of the environmental monitoring data within a specified time window are used to calculate the correlation degree, mainly based on the dynamic coupling characteristics and time sensitivity of the fishery ecosystem. The distribution, migration and catchability of fishery resources are closely related to the marine environmental conditions, and these environmental factors (such as sea surface temperature, salinity, dissolved oxygen, chlorophyll concentration, flow velocity, etc.) often show obvious time sequence change characteristics. By extracting the change trend, fluctuation amplitude and statistical characteristics of the environmental data within the time window, the stability and dynamics of the environmental state in the sea area within this time period can be effectively described, thus reflecting the potential driving effect of the environmental factors on the fishing activities and the fishing results.

[0049] In this embodiment, when the node type is a species node and any other type of node, the system calculates the correlation degree based on the species distribution and activity characteristics information. Specifically, the occurrence frequency, quantity and classification confidence of the species within the corresponding range of the relevant nodes (such as sea area, time, environment or fishing boat) are statistically analyzed. The occurrence frequency is used to reflect the distribution breadth of the species under certain conditions, the quantity is used to reflect the population size, and the classification confidence is used to evaluate the reliability of species identification. By integrating these indicators, the correlation degree between the species node and other nodes is calculated to reflect the response relationship between the species distribution and the operation activities, time change and environmental conditions.

[0050] Since the species node is the core biological element in the fishing data, its spatial distribution, seasonal activity pattern, and group behavior directly determine the spatiotemporal distribution characteristics of the fishing yield. Different species have different ecological adaptation intervals and sensitive thresholds for environmental factors. If the ecological distribution characteristics of the species are not considered, and only environmental or human activity information is relied on for correlation calculation, the actual variation pattern of the fishery resources cannot be accurately described. By introducing the information of the species distribution density, migration path, habitat depth, and activity period, the coupling strength between the species node and other nodes (such as sea area, time, environmental factor, or fishing boat) can be more accurately reflected, so that the fused graph structure has ecological consistency. Second, the association between the species and the fishing boat or sea area reflects the spatial response of the fishing behavior. For example, if the occurrence frequency of a certain species in a specific sea area or time period significantly increases, it will guide the fishing boat trajectory to concentrate in that area, forming a high fishing density area. Such a high-frequency interaction mode can only be obtained by analyzing the co-occurrence relationship between the species distribution characteristics and the fishing boat activity characteristics. By calculating the species occurrence frequency, the proportion in the fishing sample, the fishing time distribution, and the classification confidence, the association degree between the species and the fishing boat node or sea area node can be quantified, reflecting the spatial selectivity and target of the fishing activity. Third, the association between the species and the time node reflects the seasonal nature of the species and the matching of the fishery dynamics. Most economic fish species have obvious spawning seasons, migration seasons, or overwintering periods, and the species distribution and fishing yield fluctuate periodically with time. By calculating the occurrence probability and quantity variation trend of the species in the time window, the time sequence correspondence between the species dynamics and the time node can be revealed, thereby improving the accuracy of the multi-source data fusion in the temporal dimension.

[0051] After calculating the node association degree, the system determines the association edges between the nodes according to the preset generation strategy. The generation strategy can include threshold screening, similarity sorting, adjacency number limitation, etc., for controlling the number and connection density of the edges, ensuring that the graph structure can express the association relationship while maintaining computational efficiency.

[0052] Finally, based on the above node set and association edge set, the first graph structure corresponding to the coarse grid multi-source data subset and the second graph structure corresponding to the fine grid multi-source data subset are constructed. The first graph structure can reflect the macroscopic relationship between different sea areas, time periods, and fishing boat activities, and is suitable for regional fishery situation analysis; the second graph structure can reflect the fine-grained relationship of local sea area, specific time period, and specific fishing boat operation, and is suitable for micro production characteristics and environmental interaction analysis. Through this multi-level graph structure construction method, the unified expression of multi-source heterogeneous data in space, time, object, and environment is realized, providing a structured foundation for subsequent multi-layer graph neural network fusion.

[0053] Step 104: calculating the structural similarity of the first graph structure and the second graph structure based on a preset similarity calculation model, determining a fusion weight based on the structural similarity, and performing weighted fusion on the first graph structure and the second graph structure based on the fusion weight to obtain a fused structure graph; In this embodiment, the structural similarity of the first graph structure and the second graph structure is calculated based on a preset similarity calculation model, the fusion weight is determined based on the structural similarity, and the first graph structure and the second graph structure are weighted fused based on the fusion weight to obtain a fused structure graph, which includes: The structural similarity of the first graph structure and the second graph structure is calculated based on a preset similarity calculation model, and the structural similarity includes node similarity, topological similarity, relationship consistency and scale correspondence. The node similarity, topological similarity, relationship consistency and scale correspondence are corrected based on a preset confidence level to obtain a similarity index. The similarity index is mapped to a single structural similarity score based on a preset fusion rule, and the fusion weight is determined based on the structural similarity score. The first graph structure and the second graph structure are weighted fused based on the fusion weight to obtain a fused structure graph.

[0054] In this embodiment, in order to realize effective fusion of fishery multi-source data graph structures under different scales, a complementary feature of comprehensive coarse grid and fine grid data is used to form a fused structure graph with unified topological expression. First, a first graph structure and a second graph structure are constructed from a coarse grid multi-source data subset and a fine grid multi-source data subset, respectively. Each graph structure includes a plurality of type nodes and weighted relationship edges between them. Since the two are derived from data sets of different resolutions, there are differences in node quantity, edge density and topological connection mode, etc. Therefore, a structural similarity calculation model is needed for unified analysis.

[0055] In the embodiment, the similarity between the first graph structure and the second graph structure is measured from four aspects based on a preset similarity calculation model, including node similarity, topological similarity, relationship consistency and scale correspondence. The node similarity is used to measure the similarity of nodes of the same type in the two graphs in attribute characteristics. For example, the corresponding sea area nodes can calculate the similarity through the overlap ratio of the geographical coverage range, the time nodes can measure the similarity through the time window coincidence degree, and the environment factor nodes can measure the similarity through the difference of the environment variable statistical characteristics. The topological similarity is used to reflect the consistency of the two graphs in the node connection structure. The adjacency relationship of the corresponding nodes in the two graphs is compared, and the similarity of the topological statistical characteristics such as node degree distribution, clustering coefficient and path connectivity is calculated to reflect the corresponding relationship at the structure level. The relationship consistency is used to measure the semantic consistency of the associated edges between the nodes of the same type in different graphs. For example, if the distribution trend of the fishing boat and sea area connection edge weight under the coarse grid and the fine grid is similar, it indicates that the relationship consistency of the two is high. The scale correspondence is used to measure the corresponding relationship of the node aggregation or decomposition under the two grid scales, that is, to judge whether a node in the coarse grid is composed of several nodes in the fine grid, or whether the nodes in the fine grid belong to the same coarse grid unit.

[0056] In the embodiment, considering the accuracy difference of different types of data sources and observation noise, the confidence of each type of similarity result calculated is corrected. The confidence is determined according to the reliability, time integrity and spatial resolution of data acquisition. For example, the confidence of the environment data obtained based on the automatic sensor is high, and the confidence of the manual survey data is low. The confidence can be used to weight and correct each similarity index, so that the high-confidence data contributes more to the final similarity index, thereby obtaining a set of corrected comprehensive similarity indexes.

[0057] In the embodiment, after obtaining the corrected similarity indexes, the system maps the node similarity, the topological similarity, the relationship consistency and the scale correspondence according to a preset fusion rule to form a single structure similarity score. The score reflects the structural closeness of the two graph structures as a whole. According to the score result, the fusion weight is determined. When the structure similarity is high, the detail features of the fine grid graph have a high weight in the fusion; when the structure similarity is low, the weight of the coarse grid graph is increased to ensure the overall stability and physical consistency of the fused structure.

[0058] In the embodiment, the weighted fusion of the first graph structure and the second graph structure based on the fusion weight to obtain a fused structure graph includes: obtaining the node mapping relationship of the first graph structure and the second graph structure based on the unique identifier of each node, the spatial inclusion relationship or the time window coincidence degree; Fusing the first graph structure and the second graph structure based on the node mapping relationship, the preset initial edge weight and the fusion weight to obtain a fused structure graph.

[0059] In the embodiment, first, unique identifiers of the nodes are obtained. The unique identifiers can be node numbers automatically generated based on a data source system, entity numbers defined in business logic or spatial feature encodings, and are used to identify the same object in different graph structures. For example, a fishing boat node in the first graph structure can be identified by a fishing boat identification code, and a yield node in the second graph structure can be identified by a fishing ground location number.

[0060] Subsequently, the node mapping relationship between the first graph structure and the second graph structure is determined based on a spatial inclusion relationship of the nodes or a time window coincidence degree. The spatial inclusion relationship is used to describe the corresponding relationship of the nodes in the geographical or regional layer, such as the spatial membership relationship between the fishing ground node and the fishing boat node. The time window coincidence degree is used to describe the overlapping degree of the nodes in the time dimension, such as the coincidence of the fishing yield statistical time period and the ship operation time period. Based on these spatial and time attributes, the matching relationship of the nodes in different graph structures can be determined, and a node mapping table can be generated for subsequent fusion calculation.

[0061] In the embodiment, after the node mapping relationship is obtained, the first graph structure and the second graph structure are fused based on the preset initial edge weight and the fusion weight. The initial edge weight is used to reflect the strength of the edges in the original graph structure, such as the fishing frequency between the fishing boat and the fishing ground, the correlation strength between the environmental node and the yield node, and the like. The fusion weight is determined by a multi-source data fusion strategy and is used to balance the importance of different data sources. For example, when the environmental data has a greater impact on the yield, the fusion weight of the environmental graph structure can be increased. When the social and economic factors have a weak influence, the fusion weight of the social and economic graph can be correspondingly reduced.

[0062] In the embodiment, in the fusion process, the associated edges of the corresponding nodes in the two graph structures are merged according to the node mapping relationship, the edge weight values are weighted based on the fusion weight, and a new fused edge set is formed. If a node exists only in one of the graph structures, the original edge connection form is retained, and the edge weight is attenuated or completed according to a preset strategy, so as to ensure the connectivity and data consistency of the fused structure.

[0063] In the embodiment, when the first graph structure and the second graph structure are fused according to the node mapping relationship, the system first performs corresponding matching on the nodes mapped to the same object and extracts information of adjacent edges. For the case that corresponding nodes exist in both graph structures, the system identifies all associated edges between the two nodes and performs weighted processing according to a preset fusion weight. The fusion weight can be dynamically set according to the confidence of the data source, the data sampling accuracy or the time freshness. For example, when the update period of the second graph structure is shorter and the data is more real-time, the edge weight of the second graph structure can be given a higher weight during fusion to reflect the latest dynamic characteristics.

[0064] In the embodiment, when the edge weight is weighted, the edges between the same node pairs in the two graph structures are merged to generate new fusion edges. If the node pair has edge connection in both graph structures, a new fusion edge weight is calculated based on the corresponding edge weight and the fusion weight, so as to ensure that the fusion result not only retains the topological characteristics of the original graph, but also takes into account the weight distribution of different data sources. For example, for the connection relationship between the same fishing ground node and the fishing boat node in the fishing boat operation behavior graph and the environmental influence graph, the system will comprehensively consider the fishing frequency and the environmental suitability to generate a fusion edge reflecting the comprehensive strength of fishing activities.

[0065] In the embodiment, for the nodes existing only in a single graph structure, the system retains the original edge connection form, and simultaneously attenuates or completes the edge weight based on a preset strategy. Specifically, when a node exists only in the first graph structure, the system attenuates the edge weight by a set proportion to reflect the degree of information loss in the multi-source fusion scenario; when a node exists only in the second graph structure, a virtual edge is generated through a completion mechanism, and the completion method can be determined based on spatial proximity, time series trend or similar node characteristics, so as to ensure the connectivity and overall consistency of the fusion structure graph. For example, in the fusion process of fishing yield data and weather data, if a fishing ground node exists in the weather graph but is missing in the yield graph, the system can complete the edge connection for the node according to the spatial distance of adjacent fishing grounds and the similarity of characteristics such as wind speed and sea temperature, and assign a lower initial edge weight to represent the inferred relationship rather than the measured relationship.

[0066] In the embodiment, the final output fusion structure graph retains the main topological relationship and weight information in the first graph structure and the second graph structure, and can reflect the comprehensive correlation characteristics between multi-source data, thereby providing a unified data basis for subsequent graph neural network modeling, feature extraction and prediction analysis.

[0067] In the embodiment, the node mapping relationship of the two maps is obtained based on the unique identifier of each node, the spatial inclusion relationship or the time window coincidence degree, the node alignment is realized in a priority clear manner, the direct mapping of the explicit identification is reserved and the mapping based on the space-time relationship is supported, the cross-map elements can be correctly paired, the two maps are fused based on the node mapping relationship, the preset initial edge weight and the fusion weight, the weighted merging of the edge level is introduced from the original measurement and the global fusion strategy, the weight value of the merged edge reflects the original observation and considers the cross-scale consistency, and thus the fusion structure map is generated, which is structured, interpretable and applicable to subsequent graph learning.

[0068] Step 105: Constructing a fusion network based on the node layer attention module and the edge layer attention module, inputting the fusion structure map into the fusion network for fusion, and obtaining the fishing yield fusion data of the target region.

[0069] In the embodiment, the fusion network is constructed based on the node layer attention module and the edge layer attention module, the fusion structure map is input into the fusion network for fusion, and the fishing yield fusion data of the target region is obtained, including: Traversing the fusion structure map, calculating a multi-head attention coefficient for different categories of neighbor nodes of a current node based on the node layer attention module, and performing weighted aggregation on the nodes in the fusion structure map based on the multi-head attention coefficient; Calculating a contribution weight of each associated edge of a current node based on the edge layer attention module, and performing weighted fusion on the associated edges in the fusion structure map based on the contribution weight.

[0070] In the embodiment, first, the fusion structure map is taken as input data, the fusion structure map is generated after the multi-source fishery data is subjected to space-time alignment, associated rule calculation and weighted fusion, and contains multiple types of nodes such as fishing boat nodes, sea area nodes, environment nodes, time nodes and species nodes, and corresponding multiple associated edges.

[0071] In the embodiment, in order to fully mine the potential relationship between nodes and edges in different data sources, the system constructs a fusion network, which is composed of a node layer attention module and an edge layer attention module, for modeling the weight distribution of nodes and edges in the graph structure.

[0072] In this embodiment, during the operation of the fusion network, first, the fusion structure diagram is traversed, and for each target node (current node), the node layer attention module identifies the neighbor node set of different categories (such as fishing boats, sea areas, time or environmental factors, etc.) according to its category. For each category of neighbor nodes, the module calculates the attention coefficients of multiple attention heads, which are used to measure the influence degree of different neighbor nodes on the information update of the current node. The multi-head attention mechanism can learn multi-dimensional dependency from different feature subspaces, thereby improving the comprehensiveness of the node feature expression. Then, based on the weighted results of each attention head, the current node feature is aggregated and updated to realize cross-node category information fusion. For example, in the aggregation process of the fishing boat node feature, the node layer attention module will comprehensively consider the spatial adjacency relationship with the sea area node, the operation cycle relationship with the time node, and the climate adaptability association with the environment node, so as to obtain the updated feature reflecting the comprehensive production conditions.

[0073] In this embodiment, after the node aggregation is completed, the edges in the fusion structure diagram are further processed by the edge layer attention module. The edge layer attention module analyzes the node types connected by the edges, the semantic attributes (such as “operated on”, “affected on”, “output on”, etc.) of the edges, and the historical weights of the edges, and calculates the contribution weights of each associated edge. The contribution weight reflects the relative importance of different edges in information transmission and relationship expression. For example, among the multiple edges between fishing boats and sea areas, the edges of the frequently operated tracks will obtain higher contribution weights, while the edges of occasional associations will be automatically reduced in weight. Based on these contribution weights, the associated edges in the fusion structure diagram are weighted and fused to form new edge feature representations, thereby realizing information integration at the relationship level at the edge level.

[0074] In this embodiment, the catch fusion data specifically includes: spatial and time index and metadata: unique unit identification, spatial boundary, time window and data source summary, ensuring that each record is locatable and traceable. Multi-source aggregated observation values: yield and species composition, fishing boat activity statistics, environmental factor statistics, survey sample statistics, and social and economic indicators. Graph structure and relationship information: node set, edge set, edge weight, edge type, and multi-channel adjacency representation. Fusion decision information: scale fusion weight, structure similarity score, edge / node level fusion weight, and sparsification strategy description. Confidence and uncertainty evaluation: field-level quality annotation, confidence interval, uncertainty measure, and source weight, facilitating risk measurement and decision evaluation. Attention and explanatory output: node layer and edge layer attention coefficient summary, used to explain why the model gives higher weight to certain information.

[0075] In this embodiment, the fusion record can be directly used for downstream tasks such as catch estimation, resource distribution modeling, and temporary control decision-making, and also supports manual review.

[0076] In this embodiment, through the joint calculation of the node layer and edge layer attention modules, the fusion network can simultaneously capture the heterogeneous relationships between nodes and the semantic differences between edges, realizing the deep fusion of multi-source fishing data. Finally, the system outputs the catch fusion data of the target area, which comprehensively reflects the coupling characteristics between multi-dimensional factors (fishing boat activities, sea area environment, time dynamics, species distribution, etc.), providing reliable data support for catch prediction, resource assessment and fishery management.

[0077] In this embodiment, through the hierarchical fusion strategy of the double attention mechanism, global feature aggregation from the node layer to the edge layer is realized, enabling the fusion network to adaptively allocate information weights in complex multi-source data graph structures, thereby significantly improving the accuracy and interpretability of the fusion results of the catch data.

[0078] In this embodiment, by traversing the fusion structure graph and calculating multi-head attention coefficients for different categories of neighbor nodes based on the node layer attention module, the model can adaptively allocate information weights from different neighborhood categories when aggregating nodes, thereby highlighting neighbor information that contributes more to target prediction and suppressing irrelevant noise. Further, based on the edge layer attention module, the contribution weights of different associated edges are calculated and the associated edges are weighted and fused accordingly, which can dynamically identify and strengthen key relationship paths at the edge relationship level, improving the ability of the graph representation to capture complex causal and co-occurrence relationships, thereby enhancing the discriminability of the fusion network for multi-source information of sea area, time, fishing boat, environment and species and the prediction performance of downstream tasks through the combination of the two.

[0079] As a specific example of an embodiment of the present application, the catch data of a certain archipelago and adjacent sea area is described. The yield data, fishing boat trajectory data, environmental data, survey data and social data were obtained from August 1, 2024 00:00 to August 7, 2024 24:00 (7 days in total), including: fishing boat trajectory data: Global Fishing Watch fishing activity data; environmental data: sea surface temperature data from Copernicus Marine Service and chlorophyll concentration data from MODIS satellite; survey data: recent scientific trawl and fixed-point survey samples in the sea area; yield data: fishing port landing declaration or on-board reporting data; social data: provincial or city fishery market weekly report; Download or access each type of raw data within the preset time period, and keep the original file snapshot and download or access time stamp in the original storage area. Record metadata for each data set: source name, collection method, time coverage, spatial reference system, sampling frequency and quality description, and generate data access records for traceability.

[0080] All timestamps are unified to the same time zone, e.g. to UTC or to East 8, and the original time is saved as an ISO style time string. At the same time, spatial alignment is performed, and all longitude and latitude coordinates are in WGS-84. For remote sensing or grid data, the center coordinates of the pixel and the coverage polygon are recorded.

[0081] Pre-set grid and time window: coarse grid resolution: 0.25 degree x 0.25 degree; fine grid resolution: 0.05 degree x 0.05 degree; time window: every 6 hours as a window (also can be changed to daily scale for comparison).

[0082] Point-in-grid determination is performed on point data with longitude and latitude (yield, survey, observation station), and the points are mapped into coarse and fine grids; For fishing boat trajectory data, the trajectory points are sliced and mapped to the grid, and the number of trajectory points, the number of entries, and the length of stay in each grid are counted; the spatial intersection ratio of the trajectory and the grid is calculated to represent the activity intensity of the ship in the grid.

[0083] Further, in each sea area and time window unit, aggregate quantities and statistical quantities are calculated and saved according to the source, such as: total yield, species composition, number of entering ships, number of independent ships, cumulative operation time, sea surface temperature average, chlorophyll average, number of survey samples, and average weight of samples, etc.

[0084] If there are aggregate outliers, outlier labeling is performed on the aggregation results (based on historical distribution or empirical threshold), time interpolation or neighborhood mean completion is used for short-term missing, and long-term missing is kept as missing mark and record quality mark.

[0085] Get the data obtained by mapping or aggregating coarse and fine grids, respectively generate coarse grid multi-source data subset and fine grid multi-source data subset, each subset is indexed by "sea area unit and time window", records aggregation attributes and data quality fields, and generates a unique identifier for each subset entry (e.g. "grid type_grid number_time window"), and saves the source list and processing serial number for traceability.

[0086] Then, the following node types are generated at each scale, and the node contains a unique identifier and a set of attributes, such as: Sea area node: attributes include: grid number, boundary, multi-day historical yield average, bottom material label, etc. Time node: attributes include: time window identifier, belonging day or tidal phase, season label; Fishing boat node: attributes include: ship number (or de-sensitized number), ship type, tonnage, trajectory summary (coverage grid set); Environmental factor node: attributes include: window average, variance, number of observations, quality mark of a certain environmental variable; Species node: attributes include: species code, window occurrence times, average weight, classification confidence.

[0087] Further, the correlation strength (original correlation degree) is calculated for any node pair according to the matching rules by type, for example: The fishing boat and the sea area are associated with the intersection proportion of the trajectory and the sea area boundary and the stay time length to obtain the correlation degree; The time and the sea area are associated with the ship activity count, the yield aggregation and the investigation sample number of the sea area in the time window to obtain the correlation degree; The environment and the time are associated with the statistical characteristics (mean, variance, mutation) of the environmental variables in the window to obtain the correlation degree; The species and other nodes are associated with the occurrence frequency, the number and the classification confidence to obtain the correlation degree.

[0088] The original correlation degree calculated is applied to generate a set of edges to form an edge set, and each edge records the original edge weight, the source explanation and the time window identifier. The generation strategy includes threshold screening or retaining the top K highest correlations for each node.

[0089] The first graph structure of the coarse grid and the second graph structure of the fine grid are generated respectively, and the above node and edge generation is completed, and two graph structure data are output. The graph structure includes a node set, an edge set, an attribute matrix and an adjacency representation.

[0090] Four indicators of structural similarity (specific values are obtained by rules or statistical methods, and then modified by confidence): node similarity: comparing the similarity of attributes of the same type of nodes (such as geographical coverage overlap, time window overlap, etc.); Topological similarity: comparing the closeness of the structural characteristics of the two graphs in the adjacency mode, degree distribution, etc.; Relationship consistency: comparing the weight distribution and semantic consistency of corresponding edge types; Scale correspondence: measuring the coverage or correspondence of the aggregation of fine grid nodes to coarse grid.

[0091] First, the above indicators are weighted and modified according to the data source quality (such as observation times, sensor accuracy, artificial investigation confidence), to obtain robust similarity indicators, then the modified similarity indicators are mapped to a single structure similarity score according to the preset rules, and further mapped to the fusion weight of the first graph structure and the second graph structure, for example, if the score is high, the weight of the second graph structure is larger; if the score is low, the weight of the first graph structure is increased to ensure stability.

[0092] In this embodiment, the weight of the first graph structure is 0.58 and the weight of the second graph structure is 0.42.

[0093] In the process of fusion, first, the node mapping relationship between the two graphs is established based on unique identification, spatial inclusion or time window coincidence; Then the mapped nodes are combined according to the fusion weight, and the source and confidence information are retained; Next, the corresponding edges are weighted and superimposed according to the edge level fusion weight; For the edges existing only in one graph, its weight is attenuated according to the preset strategy or a virtual edge is generated based on the neighborhood similarity to maintain connectivity, and the decision is recorded in the merging log, finally the fusion structure graph is output: including node feature matrix, multi-channel adjacency representation, edge feature and merging metadata (source list, initial edge weight, mapping confidence, merging log).

[0094] Further, the fusion network uses the fusion structure graph as input, and includes a node layer attention module (calculating multi-head attention for different categories of neighbors and aggregating) and an edge layer attention module (calculating contribution weight for different edges and adjusting information transmission weight).

[0095] The fusion network uses supervised samples for supervised training of part of the nodes, and uses unsupervised signals to train unlabeled nodes. For example, using high-confidence production of scientific research survey points or species observation as labels for supervised training; applying graph autoencoder reconstruction loss or contrastive learning loss to unlabeled nodes to provide full graph learning signals.

[0096] After the fusion network is trained, the fusion features of each sea area and time unit are output (used for downstream fishing catch estimation, resource analysis or visualization), and the attention weight is recorded for explanatory analysis.

[0097] In this embodiment, by acquiring production data, fishing vessel trajectory data, environmental data, survey data, and social data of the target area and performing spatiotemporal alignment based on a preset sea area grid and time window, the comparability and aggregability of multi-source data in spatial and temporal dimensions can be ensured, thus providing high-quality raw input for subsequent unified modeling. Furthermore, the initial heterogeneous dataset is divided based on coarse and fine grid resolutions, simultaneously considering both macroscopic coverage and microscopic details, meeting the needs of analysis at different scales and cross-scale fusion. Simultaneously, by constructing a graph structure using sea area nodes, time nodes, fishing vessel nodes, environmental factor nodes, and species nodes and calculating node correlations, heterogeneous information can be semantically transformed into graph topology and attributes, enabling geographical, temporal, subject, and biological information to interact within a unified graph domain. This allows for the obtaining of the structural similarity of the two-scale graph structure based on a preset similarity calculation model. Determining fusion weights provides an objective standard for weighting cross-scale information, avoiding bias from a single scale. Furthermore, if the graph structures reflected by coarse and fine grids are highly similar, it indicates consistency between macroscopic trends and microscopic details, allowing for mutual reinforcement during fusion. Conversely, low similarity may indicate local anomalies or data uncertainty, automatically reducing the weight of certain information during fusion. This weighted fusion of two graphs based on the fusion weights to obtain a fused structure graph enables the orderly synthesis of multi-scale information, improving the integrity and robustness of graph representation. Inputting the fused structure graph into a fusion network containing node-level and edge-level attention modules for end-to-end fusion allows for adaptive allocation of attention at both the node semantic and edge relation levels, enhancing the expressive power of important information. This results in more accurate and comprehensive multi-source fusion feature data for fisheries, improving the accuracy and usability of fisheries resource assessment.

[0098] Please refer to Figure 2 , Figure 2 This is a schematic diagram of a multi-source fish catch data fusion construction system provided in an embodiment of the present invention, including: a preprocessing module 201, a grid division module 202, a graph construction module 203, a fusion module 204, and an identification module 205.

[0099] The preprocessing module 201 is used to acquire production data, fishing vessel trajectory data, environmental data, survey data and social data of the target area, and to perform spatiotemporal alignment of the production data, fishing vessel trajectory data, environmental data, survey data and social data based on a preset sea area grid and a preset time window to obtain an initial heterogeneous dataset. The grid partitioning module 202 is used to partition the initial heterogeneous dataset based on preset coarse grid resolution and fine grid resolution respectively, to obtain coarse grid multi-source data subsets and fine grid multi-source data subsets; The graph construction module 203 is configured to, for the coarse grid multi-source data subset and the fine grid multi-source data subset, respectively take a sea area node, a time node, a fishing boat node, an environmental factor node and a species node as a node, calculate the relevance between each node, determine the association edge between each node based on the relevance, and construct a first graph structure corresponding to the coarse grid multi-source data subset and a second graph structure corresponding to the fine grid multi-source data subset; The fusion module 204 is configured to calculate the structural similarity of the first graph structure and the second graph structure based on a preset similarity calculation model, determine a fusion weight based on the structural similarity, perform weighted fusion on the first graph structure and the second graph structure based on the fusion weight, and obtain a fusion structure graph. The recognition module 205 is configured to construct a fusion network based on a node layer attention module and an edge layer attention module, input the fusion structure graph into the fusion network for fusion, and obtain fishing catch fusion data of a target area.

[0100] In the embodiments of the present application, a terminal device is also provided, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the computer program is executed by the processor, the fishing catch multi-source data fusion construction method described above is implemented.

[0101] In the embodiments of the present application, a computer readable storage medium is also provided, which includes a stored computer program. When the computer program is running, the computer readable storage medium controls the device where the computer readable storage medium is located to execute the fishing catch multi-source data fusion construction method described above.

[0102] The above specific embodiments further specifically describe the purpose, technical solutions and advantages of the present application. It should be understood that the above description is only for specific embodiments of the present application and is not intended to limit the protection scope of the present application. It is particularly pointed out that any modification, equivalent replacement, improvement, etc. made by those skilled in the art within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for constructing a fishing multi-source data fusion, characterized in that, The method comprises the following steps: acquiring yield data, fishing vessel trajectory data, environmental data, survey data and social data of a target area, and performing spatio-temporal alignment on the yield data, fishing vessel trajectory data, environmental data, survey data and social data based on a preset sea area grid and a preset time window to obtain an initial heterogeneous data set; dividing the initial heterogeneous data set based on a preset coarse grid resolution and a preset fine grid resolution to obtain a coarse grid multi-source data subset and a fine grid multi-source data subset; for the coarse grid multi-source data subset and the fine grid multi-source data subset, taking a sea area node, a time node, a fishing vessel node, an environmental factor node and a species node as nodes, calculating the relevance between each node, and determining the association edge between each node based on the relevance, to construct a first graph structure corresponding to the coarse grid multi-source data subset and a second graph structure corresponding to the fine grid multi-source data subset; calculating the structural similarity of the first graph structure and the second graph structure based on a preset similarity calculation model, determining a fusion weight based on the structural similarity, and performing weighted fusion on the first graph structure and the second graph structure based on the fusion weight to obtain a fusion structure graph; constructing a fusion network based on a node layer attention module and an edge layer attention module, inputting the fusion structure graph into the fusion network for fusion to obtain fishing catch fusion data of the target area.

2. The method of claim 1, wherein, The method comprises the following steps: acquiring the first grid scale and the second grid scale of the coarse grid multi-source data subset and the fine grid multi-source data subset; generating a sea area node, a time node, a fishing vessel node, an environmental factor node and a species node based on the first grid scale and the second grid scale respectively, wherein each node comprises a unique identifier and an attribute set corresponding to its node type; for any two nodes, calculating the association degree between the nodes based on a preset semantic association rule, and determining the association edge between the nodes based on the association degree and a preset generation strategy; constructing the first graph structure corresponding to the coarse grid multi-source data subset and the second graph structure corresponding to the fine grid multi-source data subset based on each node and the association edge between the nodes.

3. The method of claim 2, wherein, The method comprises the following steps: for any two nodes, determining the node type, and matching the semantic association rule based on the node type; calculating the association degree between the nodes based on the semantic association rule, and determining the association edge between the nodes based on the association degree and a preset generation strategy.

4. The method of claim 3, wherein, The association degree between nodes is calculated based on the semantic association rule, and an associated edge between nodes is determined based on the association degree and a preset generation strategy, including: When the node type is a fishing boat node and a sea area node, a spatial intersection ratio is calculated based on a fishing boat trajectory and a sea area boundary, and the association degree is determined based on the spatial intersection ratio; When the node type is a time node and a sea area node, the association degree is calculated based on a fishing boat activity count, a catch yield, and a survey sample number in the time node; When the node type is an environmental factor node and a time node, the association degree is calculated based on an aggregated statistical feature of the environmental data in the time window; When the node type is a species node and any other node, the association degree is calculated based on a species appearance frequency, a species quantity, and a classification confidence measurement.

5. The method of claim 4, wherein, The structural similarity of the first graph structure and the second graph structure is calculated based on a preset similarity calculation model, a fusion weight is determined based on the structural similarity, the first graph structure and the second graph structure are weighted and fused based on the fusion weight, and a fusion structure graph is obtained, including: The structural similarity of the first graph structure and the second graph structure is calculated based on a preset similarity calculation model, the structural similarity including node similarity, topological similarity, relationship consistency, and scale correspondence; The node similarity, the topological similarity, the relationship consistency, and the scale correspondence are corrected based on a preset confidence, and a similarity index is obtained; The similarity index is mapped to a single structural similarity score based on a preset fusion rule, and the fusion weight is determined based on the structural similarity score; The first graph structure and the second graph structure are weighted and fused based on the fusion weight, and a fusion structure graph is obtained.

6. The method of claim 5, wherein, The first graph structure and the second graph structure are weighted and fused based on the fusion weight, and a fusion structure graph is obtained, including: A node mapping relationship of the first graph structure and the second graph structure is obtained based on unique identifiers of each node, spatial inclusion relationships, or time window overlap degrees; The first graph structure and the second graph structure are fused based on the node mapping relationship, a preset initial edge weight, and the fusion weight, and a fusion structure graph is obtained.

7. A method for constructing a fishing multisource data fusion according to any one of claims 1 to 6, characterized in that, The fusion network is constructed based on the node layer attention module and the edge layer attention module, the fusion structure graph is input into the fusion network for fusion, and the fishing catch fusion data of the target area is obtained, including: The fusion structure graph is traversed, multi-head attention coefficients of different category neighbor nodes of a current node are calculated based on the node layer attention module, and the nodes in the fusion structure graph are weighted and aggregated based on the multi-head attention coefficients; The contribution weights of each associated edge are calculated based on the edge layer attention module, and the associated edges in the fusion structure graph are weighted and fused based on the contribution weights.

8. A fishing multi-source data fusion construction system, characterized in that, It includes: a preprocessing module, a grid division module, a graph construction module, a fusion module, and an identification module; The preprocessing module is configured to acquire yield data, fishing vessel trajectory data, environmental data, survey data, and social data of a target area, and perform spatio-temporal alignment on the yield data, the fishing vessel trajectory data, the environmental data, the survey data, and the social data based on a preset sea area grid and a preset time window to obtain an initial heterogeneous data set; The grid division module is configured to divide the initial heterogeneous data set based on a preset coarse grid resolution and a preset fine grid resolution to obtain a coarse grid multi-source data subset and a fine grid multi-source data subset; The graph construction module is configured to, for the coarse grid multi-source data subset and the fine grid multi-source data subset, take a sea area node, a time node, a fishing vessel node, an environmental factor node, and a species node as nodes, calculate the relevance between each node, determine the association edges between each node based on the relevance, and construct a first graph structure corresponding to the coarse grid multi-source data subset and a second graph structure corresponding to the fine grid multi-source data subset; The fusion module is configured to calculate the structural similarity of the first graph structure and the second graph structure based on a preset similarity calculation model, determine a fusion weight based on the structural similarity, perform weighted fusion on the first graph structure and the second graph structure based on the fusion weight, and obtain a fusion structure graph; The identification module is configured to construct a fusion network based on a node layer attention module and an edge layer attention module, input the fusion structure graph into the fusion network for fusion, and obtain fishing catch fusion data of the target area.

9. The fishing multisource data fusion construction system of claim 8, wherein, The graph construction module is configured to, for the coarse grid multi-source data subset and the fine grid multi-source data subset, take a sea area node, a time node, a fishing vessel node, an environmental factor node, and a species node as nodes, calculate the relevance between each node, determine the association edges between each node based on the relevance, and construct a first graph structure corresponding to the coarse grid multi-source data subset and a second graph structure corresponding to the fine grid multi-source data subset, including: obtaining a first grid scale and a second grid scale of the coarse grid multi-source data subset and the fine grid multi-source data subset; generating a sea area node, a time node, a fishing vessel node, an environmental factor node, and a species node based on the first grid scale and the second grid scale, respectively, wherein each node includes a unique identifier and an attribute set corresponding to the node type thereof; for any two nodes, calculating the relevance between the nodes based on a preset semantic association rule, and determining the association edges between the nodes based on the relevance and a preset generation strategy; constructing the first graph structure corresponding to the coarse grid multi-source data subset and the second graph structure corresponding to the fine grid multi-source data subset based on each node and the association edges between the nodes.

10. The fishing multisource data fusion construction system of claim 9, wherein, The graph construction module is configured to, for any two nodes, calculate the relevance between the nodes based on a preset semantic association rule, and determine the association edges between the nodes based on the relevance and a preset generation strategy, including: for any two nodes, determining the node type, and matching a semantic association rule based on the node type; Calculate the correlation degree between nodes based on the semantic correlation rules, and determine the correlation edges between nodes based on the correlation degree and a preset generation strategy.

Images