An agricultural multi-source heterogeneous data processing system based on graph attention fusion

Abstract

This invention discloses an agricultural multi-source heterogeneous data processing system based on graph attention fusion, belonging to the field of agricultural information processing technology. It includes: a data acquisition module for generating a multi-source basic agricultural dataset; a standard registration module for generating a heterogeneous agricultural standard dataset; a topology construction module for constructing a heterogeneous agricultural topology graph; a graph attention fusion module for generating fusion weights; an impedance distance construction module for generating terrain propagation distances; a completion model construction module for constructing an improved CoKriging completion model; a fusion processing module for generating an agricultural fusion processing dataset; and a result output module for generating data completion results, abnormal data records, and agricultural decision input records. This invention improves the accuracy of agricultural heterogeneous data completion and the reliability of anomaly identification.

Description

Technical Field

[0001] This invention relates to the field of agricultural information processing technology, and in particular to an agricultural multi-source heterogeneous data processing system based on graph attention fusion. Background Technology

[0002] With the increasing demand for agricultural information monitoring and smart agricultural management, technologies for cross-source data acquisition, spatiotemporal registration, spatial correlation modeling, and missing data repair in agricultural operations have received widespread attention. Existing agricultural data processing systems mainly rely on sensor tabulation data, remote sensing image interpretation, and conventional spatial interpolation methods for data processing. However, these systems commonly suffer from the following problems in practical applications: Agricultural operational data sources vary greatly, with inconsistent collection times and locations, and difficulties in standardizing field names, units of measurement, and encoding formats, leading to instability in data standardization and spatiotemporal registration results. Agricultural objects are interconnected by plot adjacency, terrain connectivity, and the distribution of collection equipment. Existing processing methods often establish spatial relationships based on planar distances, which fails to adequately represent the impact of slope conduction, field barriers, and terrain access costs on data similarity in high-altitude agricultural regions. Traditional CoKriging algorithms typically use static covariates and Euclidean distance to construct cross-covariance structures. When missing or anomalous collection locations exist in heterogeneous agricultural standard datasets, it is difficult to dynamically adjust completion weights based on node data, relational edge data, and source markers, affecting the reliability of data completion results, anomalous data records, and agricultural decision input records.

[0003] Therefore, how to provide an agricultural multi-source heterogeneous data processing system based on graph attention fusion is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose an agricultural multi-source heterogeneous data processing system based on graph attention fusion. This invention achieves unified organization, missing data completion, and anomaly identification of agricultural multi-source heterogeneous data through agricultural operation data collection, standard registration, heterogeneous topology construction, graph attention fusion, terrain impedance path distance construction, and improved CoKriging completion processing. It has the advantages of high data fusion accuracy, strong terrain adaptability, and reliable support for agricultural decision-making.

[0005] An agricultural multi-source heterogeneous data processing system based on graph attention fusion according to an embodiment of the present invention includes: The data acquisition module is used to collect agricultural operation data within the agricultural area and generate a multi-source basic dataset for agriculture. The standard registration module is used to standardize agricultural multi-source basic datasets and register them according to the time and location of collection to generate agricultural heterogeneous standard datasets. The topology construction module is used to construct agricultural heterogeneous topology graphs based on agricultural object identifiers and spatial relationships, mapping agricultural heterogeneous standard datasets into node data and relation edge data. The graph attention fusion module is used to perform graph attention fusion processing on agricultural heterogeneous topology graphs, and generates fusion weights based on node data and relation edge data. The impedance distance construction module is used to construct the terrain impedance path distance based on the terrain connectivity of agricultural areas, converting spatial relationships into terrain propagation distances. The model completion module is used to write the fusion weights into the cross-covariance structure of the CoKriging algorithm and replace the Euclidean distance with the terrain impedance path distance to build an improved CoKriging completion model. The fusion processing module is used to process missing and anomalous collection locations in the heterogeneous agricultural standard dataset using an improved CoKriging completion model, generating an agricultural fusion processing dataset. The results output module is used to generate data completion results, abnormal data records, and agricultural decision input records based on the agricultural fusion processing dataset.

[0006] Optionally, the terrain impedance path distance is as follows: In the terrain impedance grid diagram, the spatial grid cell where the node data connected by spatial correlation is located is taken as the path endpoint. The terrain propagation path between the path endpoints is calculated along the grid connection edge. The passage impedance value corresponding to the grid connection edge through which the terrain propagation path passes is accumulated. The cumulative passage impedance value corresponding to the terrain propagation path with the smallest cumulative passage impedance value is determined as the terrain impedance path distance.

[0007] Optionally, the data acquisition module specifically comprises: Obtain agricultural region identifiers, agricultural object identifiers, data collection device identifiers, data collection time and location, and spatial location to establish an agricultural data collection index; Based on the agricultural data collection index, the system accesses the field sensing terminal, remote sensing acquisition terminal, terrain data terminal, and agricultural management terminal, reads data content that matches the agricultural area identifier, and generates source-layered collection records. The source-layered collection records are associated with agricultural object identifiers, and the collection device identifier, collection time and location, and spatial location are written into the source-layered collection records to generate agricultural operation data; Agricultural operation data is marked with source tags, equipment tags, and field integrity tags, and the marked agricultural operation data is aggregated into an agricultural multi-source basic dataset.

[0008] Optionally, the standard registration module specifically comprises: Agricultural operation data is read from a multi-source agricultural basic dataset, and source markers, equipment markers, field integrity markers, agricultural object identifiers, collection time and location, and spatial location are extracted to establish a standard registration index. Based on the standard registration index, agricultural operation data is processed to unify field names, convert data units, convert data types, and unify encoding formats, generating standard field records. Perform time window alignment on standard field records according to the collection time location, and write standard field records with consistent agricultural object identification and collection time location falling into the same time window into the time registration record; Perform coordinate datum transformation and spatial grid positioning on the time registration records according to their spatial location, and write the time registration records with consistent agricultural object identification and unified spatial location into the spatial registration records; Based on the field integrity markers, the spatial registration records are marked for missing locations, and the spatial registration records with missing locations marked are summarized into an agricultural heterogeneous standard dataset.

[0009] Optionally, the topology building module specifically comprises: Read the agricultural object identifiers, collection time and location, spatial location and missing location labeling results from the agricultural heterogeneous standard dataset, and build a topology construction index; Based on the topology, the data content corresponding to the agricultural object identifier is converted into node data, and the collection time location, spatial location, and missing location marking results are written into the node attribute field. The spatial adjacency relationship between node data is calculated based on spatial location, and the spatial adjacency relationship is bound to the agricultural object identifier to generate spatial association relationship; Based on spatial relationships, relation edge data is constructed, and relation edge data is connected with node data to generate an agricultural heterogeneous topology graph; The heterogeneous agricultural standard dataset is mapped to the heterogeneous agricultural topology graph according to the agricultural object identifier, so that the heterogeneous agricultural standard dataset forms node data and relation edge data.

[0010] Optionally, the graph attention fusion module specifically comprises: Read node data and relation edge data from the heterogeneous topology graph of agriculture, extract node attribute fields and edge attribute fields, and build a graph attention calculation index; Based on the graph attention calculation index, feature vector transformation is performed on the node data to generate node feature vectors, and relation vector transformation is performed on the relation edge data to generate relation edge feature vectors. Input the node feature vector and the relation edge feature vector into the graph attention network, calculate the node association strength between node data, and generate node attention weights based on the node association strength. The contribution of relation edge data to the connection relationship between node data is calculated based on the relation edge feature vector, and relation edge attention weights are generated based on the contribution. Based on the node attention weight and the relation edge attention weight, the contribution of the source labels in the heterogeneous standard agricultural dataset is aggregated to generate the data source contribution weight; The node attention weight, relation edge attention weight, and data source contribution weight are aggregated into a fusion weight.

[0011] Optionally, the impedance distance construction module specifically comprises: Agricultural region identifiers, spatial locations, and spatial relationships are read from agricultural heterogeneous standard datasets and agricultural heterogeneous topology maps. Based on the agricultural region identifiers, the topographic connectivity relationships of agricultural regions are extracted, and an impedance distance index is built. The agricultural area is divided into spatial grid cells by constructing an index based on impedance distance, and the spatial location is mapped to the spatial grid cells to generate grid location records; Based on the terrain connectivity, configure the passage impedance value for adjacent spatial grid cells, and write the passage impedance value into the grid connection edge between adjacent spatial grid cells to generate a terrain impedance grid map; Based on the grid location records, locate the node data connected by spatial relationships in the topographic impedance grid map, and determine the spatial grid cell where the node data is located as the path endpoint; The terrain propagation path with the minimum cumulative impedance value between the endpoints of the calculated path is determined as the terrain impedance path distance. The terrain impedance path distance is then bound to the spatial correlation to generate the terrain propagation distance.

[0012] Optionally, the completion model construction module specifically comprises: The model is constructed by extracting agricultural object identifiers, collection time and location, spatial location, missing location calibration results, node data, relation edge data, node attention weights, relation edge attention weights, data source contribution weights, and terrain impedance path distances from agricultural heterogeneous standard datasets, agricultural heterogeneous topology maps, fusion weights, and terrain propagation distances, and building a complete model index. Based on the completion model, an index is constructed to read the missing position marking results. The data fields corresponding to the missing position marking results are identified as records to be completed, and the data fields that have consistency with agricultural object identification and spatial association with the records to be completed are identified as collaborative field records. Based on node attention weight, relation edge attention weight, and data source contribution weight, the variable contribution relationship between the record to be completed and the collaborative field record is weighted to generate a collaborative variable contribution record; The contribution records of the collaborative variables are written into the cross covariance structure of the CoKriging algorithm, so that the cross covariance structure represents the collaborative change relationship between the records of the field to be completed and the collaborative field records according to the fusion weight. Replace the Euclidean distance in the CoKriging algorithm with the terrain impedance path distance, and generate a semi-variogram input record based on the terrain impedance path distance; An improved CoKriging completion model is constructed based on the cross-covariance structure and the semi-variogram input records.

[0013] Optionally, the fusion processing module specifically comprises: Agricultural object identifiers, collection time and location, spatial location, missing location calibration results, abnormal collection locations and data fields are read from the agricultural heterogeneous standard dataset and the improved CoKriging completion model to build a fusion processing index; Based on the fusion processing index, the data fields corresponding to the missing location marking results are identified as the missing fields to be processed. An improved CoKriging completion model is used to perform completion calculations on the missing fields to be processed, generating complete field records. Based on the fusion processing index, the data fields corresponding to the abnormal collection locations are identified as abnormal fields to be verified. An improved CoKriging completion model is used to estimate and calculate the abnormal fields to be verified, generating abnormal field estimation records. The records of abnormal fields to be verified are compared for consistency, and abnormal field estimation records are generated based on the consistency comparison results. Completed field records and anomaly identification records are written into the agricultural heterogeneous standard dataset, and then linked and organized according to agricultural object identifiers, collection time and location, and spatial location to generate an agricultural fusion processing dataset.

[0014] Optionally, the result output module specifically comprises: Read agricultural object identifiers, collection time and location, spatial location, missing location calibration results, abnormal collection locations, complete field records, abnormal calibration records, and data fields from the agricultural fusion processing dataset, and build a result output index; Based on the output index, extract the records of the completed fields, bind the records of the completed fields with the missing location marking results, and arrange them according to the agricultural object identifier, collection time and location, and spatial location to generate data completion results; Based on the output index, extract the anomaly calibration records, bind the anomaly calibration records to the anomaly collection locations, and read the consistency comparison results from the anomaly calibration records to generate anomaly data records; Based on the results, the output index is used to associate and organize the data completion results, abnormal data records, and data fields to generate agricultural decision input records; The data completion results, abnormal data records, and agricultural decision input records are output according to the agricultural object identifier, collection time and location, and spatial location.

[0015] The beneficial effects of this invention are: This invention, through a data acquisition module, a standard registration module, and a topology construction module, unifies scattered agricultural operational data within agricultural areas into a unified agricultural multi-source basic dataset and an agricultural heterogeneous standard dataset. This is further mapped into node data and relational edge data, enabling data from different sources, collection times, and spatial locations to form stable correspondences within the same heterogeneous agricultural topology graph. Compared to processing methods relying solely on tabular records or single sensor data, this invention improves the organizational consistency of agricultural operational data, reduces data fragmentation caused by inconsistent field formats, spatial mismatches, and misaligned collection times, and provides a structured and traceable data foundation for agricultural data completion, anomaly identification, and agricultural decision-making.

[0016] This invention generates fusion weights through a graph attention fusion module, converts spatial relationships into terrain propagation distances through an impedance distance construction module, and then writes the fusion weights into the cross-covariance structure of the CoKriging algorithm through a completion model construction module. Simultaneously, it replaces Euclidean distance with terrain impedance path distance to construct an improved CoKriging completion model. This allows the processing of missing and anomalous data locations to simultaneously consider the importance of node data, the connectivity contribution of relational edge data, the data contribution of source markers, and the terrain connectivity of agricultural areas. This improves the accuracy and scenario adaptability of agricultural fusion processing datasets, making the output data completion results, anomalous data records, and agricultural decision input records more reliable. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a structural diagram of an agricultural multi-source heterogeneous data processing system based on graph attention fusion proposed in this invention; Figure 2 This is a flowchart illustrating the construction process of an improved CoKriging completion model for an agricultural multi-source heterogeneous data processing system based on graph attention fusion, as proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figures 1-2 A graph attention-based multi-source heterogeneous data processing system for agriculture, comprising: The data acquisition module is used to collect agricultural operation data within the agricultural area and generate a multi-source basic dataset for agriculture. The standard registration module is used to standardize agricultural multi-source basic datasets and register them according to the time and location of collection to generate agricultural heterogeneous standard datasets. The topology construction module is used to construct agricultural heterogeneous topology graphs based on agricultural object identifiers and spatial relationships, mapping agricultural heterogeneous standard datasets into node data and relation edge data. The graph attention fusion module is used to perform graph attention fusion processing on agricultural heterogeneous topology graphs, and generates fusion weights based on node data and relation edge data. The impedance distance construction module is used to construct the terrain impedance path distance based on the terrain connectivity of agricultural areas, converting spatial relationships into terrain propagation distances. The model completion module is used to write the fusion weights into the cross-covariance structure of the CoKriging algorithm and replace the Euclidean distance with the terrain impedance path distance to build an improved CoKriging completion model. The fusion processing module is used to process missing and anomalous collection locations in the heterogeneous agricultural standard dataset using an improved CoKriging completion model, generating an agricultural fusion processing dataset. The results output module is used to generate data completion results, abnormal data records, and agricultural decision input records based on the agricultural fusion processing dataset.

[0020] In this embodiment, the terrain impedance path distance is as follows: In the terrain impedance grid diagram, the spatial grid cell where the node data connected by spatial correlation is located is taken as the path endpoint. The terrain propagation path between the path endpoints is calculated along the grid connection edge. The passage impedance value corresponding to the grid connection edge through which the terrain propagation path passes is accumulated. The cumulative passage impedance value corresponding to the terrain propagation path with the smallest cumulative passage impedance value is determined as the terrain impedance path distance.

[0021] In this embodiment, the data acquisition module specifically comprises: Obtain agricultural region identifier, agricultural object identifier, data collection device identifier, data collection time and location, and spatial location. Write the agricultural region identifier into the data collection range field, the agricultural object identifier into the object attribution field, the data collection device identifier into the source device field, and the data collection time and location into the location field. Then, establish an agricultural data collection index according to the binding relationship between the fields. Based on the agricultural data collection index, access the field sensing terminal, remote sensing acquisition terminal, terrain data terminal and agricultural management terminal, read the data content that matches the agricultural area identifier, write the data content into the source level field according to the access terminal type, and use the acquisition device identifier as the verification basis for the source level field to generate source hierarchical acquisition records; The data collected from different sources is associated with agricultural object identifiers, so that the data content under the same agricultural object identifier forms an object record. Then, the data collection device identifier, data collection time and location, and spatial location are written into the object record, so that the object record has the source device, data collection time sequence and spatial location at the same time, thus generating agricultural operation data. Agricultural operation data is marked with source tags, equipment tags, and field integrity tags. Source tags are used to distinguish the access terminal type of data content, equipment tags are used to record the identification of the collection equipment, and field integrity tags are used to record the writing status of agricultural area identification, agricultural object identification, collection time and location, and spatial location. The marked agricultural operation data are summarized into an agricultural multi-source basic dataset.

[0022] In this embodiment, the standard registration module specifically includes: Agricultural operation data is read from the multi-source basic dataset of agriculture, and source markers, equipment markers, field integrity markers, agricultural object identifiers, collection time and location and spatial location are extracted. The source markers are written into the source identification field, the equipment markers are written into the equipment identification field, the field integrity markers are written into the integrity identification field, and the agricultural object identifiers, collection time and location and spatial location are written into the registration and positioning field to establish a standard registration index. Based on the standard registration index, agricultural operation data undergoes unified processing of field names, data unit conversion, data type conversion, and encoding format. Unified field name processing is used to group fields from different sources into the same field naming rule. Data unit conversion processing is used to convert the values ​​of the same indicator to the same unit of measurement. Data type conversion processing is used to convert text, numerical, and time data to the matching data type. Unified encoding format processing is used to write agricultural operation data into the same encoding format and generate standard field records. The standard field records are aligned with the time window according to the collection time location. First, the time window is determined based on the collection time location. Then, the standard field records with the same agricultural object identification and whose collection time location falls into the same time window are merged so that the data content corresponding to the same agricultural object identification forms a data record with consistent time and is written into the time registration record. Perform coordinate datum transformation and spatial grid positioning on the time registration records according to their spatial location. Coordinate datum transformation is used to transform the spatial location to the same coordinate datum, and spatial grid positioning is used to map the spatial location that has completed coordinate datum transformation to the spatial grid corresponding to the agricultural area. Write the time registration records with consistent agricultural object identification and unified spatial location into the spatial registration records. Based on the field integrity marker, the spatial registration records are used to perform missing location calibration. First, the data fields that have not been written in the spatial registration records are identified. Then, the data fields that have not been written are bound with the agricultural object identifier, the collection time and location and the spatial location to form the missing location calibration result. The spatial registration records with completed missing location calibration are summarized into an agricultural heterogeneous standard dataset.

[0023] In this embodiment, the topology construction module specifically includes: Read agricultural object identifiers, collection time and location, spatial location and missing location calibration results from the agricultural heterogeneous standard dataset, write agricultural object identifiers into the object index field, collection time and location into the time series index field, spatial location into the location index field, and missing location calibration results into the missing index field, and build a topology index according to the field binding relationship; Based on the topology, an index is constructed to extract the data content corresponding to the agricultural object identifier. The data content associated with the same agricultural object identifier is merged into node records. The node records are converted into node data. Then, the collection time location, spatial location, and missing location marking results are written into the node attribute fields of the node data, so that the node data has object ownership, time location, spatial location, and missing status. Based on the spatial location in the node attribute field, calculate the coordinate distance and spatial grid proximity status between node data. Establish node correspondence for node data whose coordinate distance meets the adjacency judgment condition and whose spatial grid proximity status meets the connection judgment condition. Then bind the node correspondence with the agricultural object identifier to generate spatial association relationship. Based on spatial relationships, the connection objects and directions between node data are determined. The connection objects, directions, and spatial relationships are written into the edge attribute fields to form relation edge data. Then, the relation edge data is used as the connection carrier between node data to generate an agricultural heterogeneous topology graph. The agricultural object identifiers in the agricultural heterogeneous standard dataset are matched with the node data in the agricultural heterogeneous topology graph for consistency, and the spatial relationships in the agricultural heterogeneous standard dataset are matched with the relation edge data in the agricultural heterogeneous topology graph for consistency, so that the agricultural heterogeneous standard dataset can form node data and relation edge data.

[0024] In this embodiment, the graph attention fusion module specifically comprises: Read node data and relation edge data from the agricultural heterogeneous topology graph, locate node data according to agricultural object identifier, locate relation edge data according to spatial association, extract node attribute fields from node data, extract edge attribute fields from relation edge data, write node attribute fields and edge attribute fields into the same calculation record, and establish graph attention calculation index. Based on the graph attention calculation index, feature vector transformation is performed on the node data. The sampling time location, spatial location, and missing location marking results in the node attribute field are numerically processed and arranged according to a uniform vector length to generate node feature vectors. Based on the graph attention calculation index, relation vector transformation is performed on the relation edge data. The connection object, connection direction, and spatial association relationship in the edge attribute field are numerically processed to generate relation edge feature vectors. The node feature vectors and relation edge feature vectors are input into the graph attention network. First, the node feature vectors are linearly mapped, and then the relation edge feature vectors are used as connection constraints to participate in the attention scoring, so that the node data with connection relationships form an association score. The node association strength between node data is generated based on the association score, and the node attention weight is generated based on the node association strength. Based on the feature vector of the relation edge, the connection object, connection direction and spatial relationship corresponding to the relation edge data are read. The matching degree of the connection object, the correspondence of the connection direction and the tightness of the spatial relationship are comprehensively calculated to obtain the contribution of the relation edge data to the connection relationship between the node data. The relation edge attention weight is generated based on the contribution degree. Based on the node attention weight and the relation edge attention weight, the contribution of the source labels in the agricultural heterogeneous standard dataset is aggregated. First, the source labels are mapped to the node data. Then, the node attention weight and relation edge attention weight corresponding to the node data are assigned to the same source label. The data contribution degree corresponding to the source label is calculated to generate the data source contribution weight. The node attention weight, relation edge attention weight, and data source contribution weight are associated and summarized according to the agricultural object identifier, collection time and location, and spatial location. The fusion weight simultaneously represents the importance of node data, the degree of connection contribution of relation edge data, and the degree of data contribution of source labeling, thus generating the fusion weight.

[0025] In this embodiment, the impedance distance construction module is specifically as follows: Agricultural region identifiers, spatial locations, and spatial relationships are read from agricultural heterogeneous standard datasets and agricultural heterogeneous topology maps. Agricultural region identifiers are used as the matching criteria for agricultural region ranges, spatial locations are used as the location criteria for node data placement, and spatial relationships are used as the objects for calculating distances between node data. Then, the terrain connectivity relationships corresponding to the agricultural region are retrieved based on the agricultural region identifiers, so that agricultural region identifiers, spatial locations, spatial relationships, and terrain connectivity relationships form the same index record, and an impedance distance index is established. The agricultural area is divided into spatial grid cells by constructing an index based on impedance distance, so that the spatial grid cells cover the agricultural area range corresponding to the agricultural area identifier. Then, the spatial position is matched to the corresponding spatial grid cell one by one, the correspondence between the spatial position and the spatial grid cell is recorded, and the grid position record is generated. Based on the terrain connectivity, determine whether terrain connectivity can be formed between adjacent spatial grid cells. Establish grid connection edges for adjacent spatial grid cells that can form terrain connectivity. Then, configure the passage impedance value for the grid connection edge according to the passage difficulty corresponding to the terrain connectivity. Make the passage impedance value represent the terrain propagation cost between adjacent spatial grid cells and generate a terrain impedance grid map. Based on the grid location records, locate the node data of spatial relationship connection in the topographic impedance grid map, match the spatial location corresponding to the node data to the spatial grid cell, and then determine the spatial grid cell where the node data at both ends of the spatial relationship is located as the path endpoint, so that the path endpoint and the spatial relationship remain in correspondence; The terrain propagation path is calculated by connecting the endpoints of the path along the grid connection edges. The passage impedance values ​​corresponding to the grid connection edges traversed by the terrain propagation path are accumulated to obtain the cumulative passage impedance value. Then, the terrain propagation path with the smallest cumulative passage impedance value is selected from the terrain propagation paths between the endpoints. The cumulative passage impedance value corresponding to the selected terrain propagation path is determined as the terrain impedance path distance. The terrain impedance path distance is then bound to the spatial correlation to generate the terrain propagation distance.

[0026] In this embodiment, the model completion module specifically includes: The model extracts agricultural object identifiers, collection time and location, spatial location, missing location calibration results, node data, relation edge data, node attention weights, relation edge attention weights, data source contribution weights, and terrain impedance path distances from agricultural heterogeneous standard datasets, agricultural heterogeneous topology maps, fusion weights, and terrain propagation distances. It uses agricultural object identifiers, collection time and location, and spatial location as field location criteria, missing location calibration results as the basis for object identification, and node data, relation edge data, node attention weights, relation edge attention weights, data source contribution weights, and terrain impedance path distances as the basis for model construction, and establishes an index for model construction. Based on the completion model, an index is constructed to read the missing location calibration results, locate the data fields corresponding to the missing location calibration results, and determine the located data fields as the records to be completed. Then, according to the consistency of agricultural object identifiers, data fields with object associations to the records to be completed are retrieved from the agricultural heterogeneous standard dataset, and data fields with connection relationships to the records to be completed are retrieved from the agricultural heterogeneous topology graph according to spatial association relationships. The retrieved data fields are determined as collaborative field records. Based on node attention weight, relation edge attention weight, and data source contribution weight, the variable contribution relationship between the record to be completed and the collaborative record is weighted. The node attention weight is used to characterize the influence strength of node data on the record to be completed, the relation edge attention weight is used to characterize the influence strength of spatial association on field connection relationship, and the data source contribution weight is used to characterize the influence strength of source tag on field credibility, thus generating collaborative variable contribution records. The collaborative variable contribution record is written into the cross covariance structure of the CoKriging algorithm, so that the cross covariance structure reads the variable contribution relationship between the field record to be completed and the collaborative field record, and adjusts the collaborative change relationship between the field record to be completed and the collaborative field record according to the fusion weight, so that the cross covariance structure simultaneously corresponds to field association, node association and source contribution. The terrain impedance path distance replaces the Euclidean distance in the CoKriging algorithm, and a semi-mutation function input record is generated based on the terrain impedance path distance. This allows the semi-mutation function input record to express the spatial correlation between the field record to be completed and the collaborative field record according to the terrain propagation distance. An improved CoKriging completion model is constructed based on the cross-covariance structure and the semi-variogram input records. The cross-covariance structure is used as a covariance variable constraint, and the semi-variogram input records are used as a spatial distance constraint. This enables the improved CoKriging completion model to solve the completion model according to the fusion weights and the path distance of the terrain impedance.

[0027] This invention transforms the traditional CoKriging's static covariance relationship, which relies on static covariates, into a dynamic relationship jointly regulated by node association, edge connection contribution, and source credibility, by incorporating node attention weights, relation edge attention weights, and data source contribution weights into the cross-covariance structure of the CoKriging algorithm. Simultaneously, it generates a semi-variogram input record by replacing Euclidean distance with terrain impedance path distance, allowing spatial correlation to be expressed according to terrain propagation distance. The improved CoKriging completion model can simultaneously utilize fusion weights and terrain impedance path distance to solve the completion model, thereby reducing interference from invalid data sources on records to be completed, improving the accuracy of missing field completion in complex agricultural areas, enhancing the model's adaptability to terrain connectivity and spatial correlation, and improving the consistency between the results of abnormal data collection and the actual agricultural situation.

[0028] In this embodiment, the fusion processing module specifically comprises: Agricultural object identifiers, collection time and location, spatial location, missing location calibration results, abnormal collection locations, and data fields are read from the agricultural heterogeneous standard dataset and the improved CoKriging completion model. The agricultural object identifier is used as the basis for object attribution, the collection time and location is used as the basis for temporal positioning, the spatial location is used as the basis for location positioning, the missing location calibration results are used as the basis for missing field identification, the abnormal collection locations are used as the basis for abnormal field identification, and the data fields are used as the processing objects to establish a fusion processing index. Based on the fusion processing index, the data fields corresponding to the missing location marking results are determined as the missing fields to be processed. The agricultural object identifier, collection time location, and spatial location corresponding to the missing fields to be processed are read. Then, the improved CoKriging completion model is called to perform completion calculations on the missing fields to be processed. The field values, agricultural object identifier, collection time location, and spatial location obtained from the completion calculations are written into the same record to generate a completion field record. Based on the fusion processing index, the data field corresponding to the abnormal collection location is determined as the abnormal field to be verified. The agricultural object identifier, collection time and location and spatial location corresponding to the abnormal field to be verified are read. Then, the improved CoKriging completion model is called to estimate the abnormal field to be verified. The estimated field value, agricultural object identifier, collection time and location and spatial location are written into the same record to generate an abnormal field estimation record. The abnormal fields to be verified and the estimated abnormal field records are matched according to the agricultural object identifier, collection time and location, and spatial location. The field value consistency comparison is performed on the matched abnormal fields to be verified and the estimated abnormal field records, and the comparison differences are written into the consistency comparison results. Based on the consistency comparison results, the data fields corresponding to the abnormal collection locations are marked, and abnormal marking records are generated. The completion field records are written into the data fields corresponding to the missing location calibration results in the agricultural heterogeneous standard dataset, and the anomaly calibration records are written into the data fields corresponding to the abnormal collection locations in the agricultural heterogeneous standard dataset. Then, the written agricultural heterogeneous standard dataset is associated and organized according to the agricultural object identifier, collection time and location, and spatial location to generate the agricultural fusion processing dataset.

[0029] In this embodiment, the result output module is specifically as follows: The agricultural object identifier, collection time and location, spatial location, missing location calibration result, abnormal collection location, completion field record, abnormal calibration record, and data field are read from the agricultural fusion processing dataset. The agricultural object identifier is used as the result attribution basis, the collection time and location is used as the result time sequence basis, the spatial location is used as the result location basis, the missing location calibration result and abnormal collection location are used as the result classification basis, and the completion field record, abnormal calibration record, and data field are used as the result organization objects to establish a result output index. Based on the output index, extract the records of the completed fields, bind the field values ​​in the completed field records to the missing location marking results, and arrange the completed field records according to the agricultural object identifier, collection time location, and spatial location, so that the completed field records have object ownership, temporal location, and spatial location, and generate data completion results; Based on the output index, extract the anomaly labeling records, bind the anomaly labeling records to the anomaly collection locations, read the consistency comparison results from the anomaly labeling records, use the consistency comparison results as the basis for anomaly judgment of the data fields corresponding to the anomaly collection locations, and generate anomaly data records. Based on the output index, the data completion results, abnormal data records, and data fields are associated with agricultural object identifiers, time sequence associations, and spatial location associations, so that the data completion results, abnormal data records, and data fields form a data input relationship oriented towards agricultural decision-making, and agricultural decision input records are generated. The data completion results, abnormal data records, and agricultural decision input records are output according to the agricultural object identifier, collection time and location, and spatial location, so that the output content maintains consistency in object attribution, temporal arrangement, and spatial location.

[0030] Example 1: To verify the feasibility of this invention in practice, it was applied to a smart monitoring scenario in a high-altitude terraced agricultural area. This agricultural area suffers from significant slope undulations, irregular field boundaries, winding irrigation paths, and dispersed deployment of on-site sensors. Agricultural operation data comes from on-site sensing terminals, remote sensing acquisition terminals, terrain data terminals, and agricultural management terminals. Due to unstable communication signals in some slope areas, soil moisture, temperature and humidity, and crop growth fields were missing during the acquisition process. Simultaneously, some acquisition devices were affected by rain, direct sunlight, and power fluctuations, resulting in abnormal acquisition locations in the agricultural operation data. Traditional spatial interpolation methods based on Euclidean distance tend to classify fields that are close in line but separated by field ridges as strongly correlated, causing the missing data completion results to deviate from the actual agricultural conditions.

[0031] During implementation, the data acquisition module first connects to the field sensing end, remote sensing acquisition end, terrain data end, and agricultural management end within the agricultural area to obtain agricultural area identifiers, agricultural object identifiers, acquisition device identifiers, acquisition time and location, and spatial location, forming a multi-source basic dataset for agriculture. The standard registration module unifies field names, converts data units, transforms data types, and unifies encoding formats for the multi-source basic dataset. It then aligns time windows according to acquisition time and location, and performs coordinate benchmark transformation and spatial grid positioning according to spatial location, forming a heterogeneous standard dataset for agriculture. The topology construction module converts the data content corresponding to agricultural object identifiers into node data and calculates spatial adjacency relationships based on spatial location, generating spatial association relationships and relational edge data to form a heterogeneous agricultural topology graph. The graph attention fusion module generates fusion weights based on node data and relational edge data. The impedance distance construction module constructs terrain impedance path distances based on terrain connectivity relationships, enabling irrigation connectivity, slope propagation, and field blockage to be included in spatial distance calculations. The completion model construction module writes the fusion weights into the cross-covariance structure of the CoKriging algorithm and replaces Euclidean distance with terrain impedance path distance, forming an improved CoKriging completion model. The fusion processing module uses an improved CoKriging completion model to process missing and abnormal data collection locations. Finally, the result output module generates data completion results, abnormal data records, and agricultural decision input records.

[0032] During a 30-day continuous operation verification, 128 monitoring points within the high-altitude terraced agricultural area were selected as verification objects. Agricultural operation data was generated for each monitoring point according to a unified collection time and location. During the verification period, the agricultural multi-source basic dataset accumulated over 46,000 agricultural operation data entries, of which 3,126 were manually verified as missing fields, and 784 were found to be abnormally collected fields. This invention was compared with the ordinary CoKriging processing method. Ordinary CoKriging uses Euclidean distance and static covariates for completion, while this invention uses fused weights and terrain impedance path distance to construct an improved CoKriging completion model. The verification results are shown in Table 1. Table 1 Comparison of Treatment Effects

[0033] As shown in Table 1, this invention outperforms the conventional CoKriging method in terms of missing field completion, abnormal data collection location identification, and processing efficiency. Because the fusion weights reflect the contribution of node data, relational edge data, and source markers to the fields being processed, the completion results do not solely rely on fixed covariates. Furthermore, since the terrain impedance path distance characterizes the actual propagation cost of spatial relationships in high-altitude terraced agricultural areas, the completion calculation avoids misjudgments caused by field boundary barriers and strengthens effective connections on irrigation connectivity paths. After the agricultural fusion processing dataset is output, the data completion results and abnormal data records are more stable, and the agricultural decision input records can more accurately reflect the true state of agricultural objects in terms of collection time and spatial location, thereby improving the supporting value of agricultural data processing results for irrigation scheduling, equipment inspection, and crop management.

[0034] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An agricultural multi-source heterogeneous data processing system based on graph attention fusion, characterized in that, include: The data acquisition module is used to collect agricultural operation data within the agricultural area and generate a multi-source basic dataset for agriculture. The standard registration module is used to standardize agricultural multi-source basic datasets and register them according to the time and location of collection to generate agricultural heterogeneous standard datasets. The topology construction module is used to construct agricultural heterogeneous topology graphs based on agricultural object identifiers and spatial relationships, mapping agricultural heterogeneous standard datasets into node data and relation edge data. The graph attention fusion module is used to perform graph attention fusion processing on agricultural heterogeneous topology graphs, and generates fusion weights based on node data and relation edge data. The impedance distance construction module is used to construct the terrain impedance path distance based on the terrain connectivity of agricultural areas, converting spatial relationships into terrain propagation distances. The model completion module is used to write the fusion weights into the cross-covariance structure of the CoKriging algorithm and replace the Euclidean distance with the terrain impedance path distance to build an improved CoKriging completion model. The fusion processing module is used to process missing and anomalous collection locations in the heterogeneous agricultural standard dataset using an improved CoKriging completion model, generating an agricultural fusion processing dataset. The results output module is used to generate data completion results, abnormal data records, and agricultural decision input records based on the agricultural fusion processing dataset.

2. The agricultural multi-source heterogeneous data processing system based on graph attention fusion according to claim 1, characterized in that, The terrain impedance path distance is as follows: In the terrain impedance grid diagram, the spatial grid cell where the node data connected by spatial correlation is located is taken as the path endpoint. The terrain propagation path between the path endpoints is calculated along the grid connection edge. The passage impedance value corresponding to the grid connection edge through which the terrain propagation path passes is accumulated. The cumulative passage impedance value corresponding to the terrain propagation path with the smallest cumulative passage impedance value is determined as the terrain impedance path distance.

3. The agricultural multi-source heterogeneous data processing system based on graph attention fusion according to claim 1, characterized in that, The data acquisition module is specifically: Obtain agricultural region identifiers, agricultural object identifiers, data collection device identifiers, data collection time and location, and spatial location to establish an agricultural data collection index; Based on the agricultural data collection index, the system accesses the field sensing terminal, remote sensing acquisition terminal, terrain data terminal, and agricultural management terminal, reads data content that matches the agricultural area identifier, and generates source-layered collection records. The source-layered collection records are associated with agricultural object identifiers, and the collection device identifier, collection time and location, and spatial location are written into the source-layered collection records to generate agricultural operation data; Agricultural operation data is marked with source tags, equipment tags, and field integrity tags, and the marked agricultural operation data is aggregated into an agricultural multi-source basic dataset.

4. The agricultural multi-source heterogeneous data processing system based on graph attention fusion according to claim 1, characterized in that, The standard registration module is specifically as follows: Agricultural operation data is read from a multi-source agricultural basic dataset, and source markers, equipment markers, field integrity markers, agricultural object identifiers, collection time and location, and spatial location are extracted to establish a standard registration index. Based on the standard registration index, agricultural operation data is processed to unify field names, convert data units, convert data types, and unify encoding formats, generating standard field records. Perform time window alignment on standard field records according to the collection time location, and write standard field records with consistent agricultural object identification and collection time location falling into the same time window into the time registration record; Perform coordinate datum transformation and spatial grid positioning on the time registration records according to their spatial location, and write the time registration records with consistent agricultural object identification and unified spatial location into the spatial registration records; Based on the field integrity markers, the spatial registration records are marked for missing locations, and the spatial registration records with missing locations marked are summarized into an agricultural heterogeneous standard dataset.

5. The agricultural multi-source heterogeneous data processing system based on graph attention fusion according to claim 1, characterized in that, The topology construction module is specifically as follows: Read the agricultural object identifiers, collection time and location, spatial location and missing location labeling results from the agricultural heterogeneous standard dataset, and build a topology construction index; Based on the topology, the data content corresponding to the agricultural object identifier is converted into node data, and the collection time location, spatial location, and missing location marking results are written into the node attribute field. The spatial adjacency relationship between node data is calculated based on spatial location, and the spatial adjacency relationship is bound to the agricultural object identifier to generate spatial association relationship; Based on spatial relationships, relation edge data is constructed, and relation edge data is connected with node data to generate an agricultural heterogeneous topology graph; The heterogeneous agricultural standard dataset is mapped to the heterogeneous agricultural topology graph according to the agricultural object identifier, so that the heterogeneous agricultural standard dataset forms node data and relation edge data.

6. The agricultural multi-source heterogeneous data processing system based on graph attention fusion according to claim 1, characterized in that, The graph attention fusion module is specifically as follows: Read node data and relation edge data from the heterogeneous topology graph of agriculture, extract node attribute fields and edge attribute fields, and build a graph attention calculation index; Based on the graph attention calculation index, feature vector transformation is performed on the node data to generate node feature vectors, and relation vector transformation is performed on the relation edge data to generate relation edge feature vectors. Input the node feature vector and the relation edge feature vector into the graph attention network, calculate the node association strength between node data, and generate node attention weights based on the node association strength. The contribution of relation edge data to the connection relationship between node data is calculated based on the relation edge feature vector, and relation edge attention weights are generated based on the contribution. Based on the node attention weight and the relation edge attention weight, the contribution of the source labels in the heterogeneous standard agricultural dataset is aggregated to generate the data source contribution weight; The node attention weight, relation edge attention weight, and data source contribution weight are aggregated into a fusion weight.

7. The agricultural multi-source heterogeneous data processing system based on graph attention fusion according to claim 1, characterized in that, The impedance distance construction module is specifically as follows: Agricultural region identifiers, spatial locations, and spatial relationships are read from agricultural heterogeneous standard datasets and agricultural heterogeneous topology maps. Based on the agricultural region identifiers, the topographic connectivity relationships of agricultural regions are extracted, and an impedance distance index is built. The agricultural area is divided into spatial grid cells by constructing an index based on impedance distance, and the spatial location is mapped to the spatial grid cells to generate grid location records; Based on the terrain connectivity, configure the passage impedance value for adjacent spatial grid cells, and write the passage impedance value into the grid connection edge between adjacent spatial grid cells to generate a terrain impedance grid map; Based on the grid location records, locate the node data connected by spatial relationships in the topographic impedance grid map, and determine the spatial grid cell where the node data is located as the path endpoint; The terrain propagation path with the minimum cumulative impedance value between the endpoints of the calculated path is determined as the terrain impedance path distance. The terrain impedance path distance is then bound to the spatial correlation to generate the terrain propagation distance.

8. The agricultural multi-source heterogeneous data processing system based on graph attention fusion according to claim 1, characterized in that, The completion model construction module is specifically as follows: The model is constructed by extracting agricultural object identifiers, collection time and location, spatial location, missing location calibration results, node data, relation edge data, node attention weights, relation edge attention weights, data source contribution weights, and terrain impedance path distances from agricultural heterogeneous standard datasets, agricultural heterogeneous topology maps, fusion weights, and terrain propagation distances, and building a complete model index. Based on the completion model, an index is constructed to read the missing position marking results. The data fields corresponding to the missing position marking results are identified as records to be completed, and the data fields that have consistency with agricultural object identification and spatial association with the records to be completed are identified as collaborative field records. Based on node attention weight, relation edge attention weight, and data source contribution weight, the variable contribution relationship between the record to be completed and the collaborative field record is weighted to generate a collaborative variable contribution record; The contribution records of the collaborative variables are written into the cross covariance structure of the CoKriging algorithm, so that the cross covariance structure represents the collaborative change relationship between the records of the field to be completed and the collaborative field records according to the fusion weight. Replace the Euclidean distance in the CoKriging algorithm with the terrain impedance path distance, and generate a semi-variogram input record based on the terrain impedance path distance; An improved CoKriging completion model is constructed based on the cross-covariance structure and the semi-variogram input records.

9. The agricultural multi-source heterogeneous data processing system based on graph attention fusion according to claim 1, characterized in that, The fusion processing module specifically comprises: The agricultural object identifier, collection time and location, spatial location, missing location calibration results, abnormal collection locations and data fields are read from the agricultural heterogeneous standard dataset and the improved CoKriging completion model to build a fusion processing index; Based on the fusion processing index, the data fields corresponding to the missing location marking results are identified as the missing fields to be processed. An improved CoKriging completion model is used to perform completion calculations on the missing fields to be processed, generating complete field records. Based on the fusion processing index, the data fields corresponding to the abnormal collection locations are identified as abnormal fields to be verified. An improved CoKriging completion model is used to estimate and calculate the abnormal fields to be verified, generating abnormal field estimation records. The records of abnormal fields to be verified are compared for consistency, and abnormal field estimation records are generated based on the consistency comparison results. Completed field records and anomaly identification records are written into the agricultural heterogeneous standard dataset, and then linked and organized according to agricultural object identifiers, collection time and location, and spatial location to generate an agricultural fusion processing dataset.

10. The agricultural multi-source heterogeneous data processing system based on graph attention fusion according to claim 1, characterized in that, The result output module is specifically as follows: Read agricultural object identifiers, collection time and location, spatial location, missing location calibration results, abnormal collection locations, complete field records, abnormal calibration records, and data fields from the agricultural fusion processing dataset, and build a result output index; Based on the output index, extract the records of the completed fields, bind the records of the completed fields with the missing location marking results, and arrange them according to the agricultural object identifier, collection time and location, and spatial location to generate data completion results; Based on the output index, extract the anomaly calibration records, bind the anomaly calibration records to the anomaly collection locations, and read the consistency comparison results from the anomaly calibration records to generate anomaly data records; Based on the results, the output index is used to associate and organize the data completion results, abnormal data records, and data fields to generate agricultural decision input records; The data completion results, abnormal data records, and agricultural decision input records are output according to the agricultural object identifier, collection time and location, and spatial location.

Images