A multi-source heterogeneous geographic data processing method based on a deep fusion algorithm

Abstract

The application discloses a kind of based on deep fusion algorithm's multi-source heterogeneous geographic data processing method, comprising: collecting multi-source heterogeneous geographic data, and write into uniform index library;Perform coordinate reference normalization, time series complement and integrity-correctness joint verification, obtain standardized geographic data block;Carry out semantic-space joint matching, generate candidate cross-modal entity pair;Construct cross-modal attention network, extract fusion feature and calculate posterior fusion confidence;Select the highest confidence fusion result in each candidate cross-modal entity pair as target geographic entity, and generate unique fusion identification;The target geographic entity holding unique fusion identification is written into fusion database, and obtains uniform geographic entity dataset.The application solves the problems of low data fusion quality, insufficient fusion process credibility and cross-modal feature matching difficulty in the prior art, improves the fusion precision, efficiency and credibility.

Description

Technical Field

[0001] This invention relates to the technical field of geographic information processing and data fusion, and in particular to a method for processing multi-source heterogeneous geographic data based on a deep fusion algorithm. Background Technology

[0002] With the rapid development of spatial information technology, the means of collecting geographic data and information sources have become increasingly diversified, expanding from single vector maps to multi-source heterogeneous information such as remote sensing raster imagery, lidar point clouds, POIs, UAV low-altitude data, and social crowdsourced data. The fusion and unified processing of these multi-source heterogeneous geographic data has become one of the important topics in current geographic information science research and application. The methods currently widely used mainly include fusion techniques based on traditional data matching and spatial overlay, as well as intelligent fusion techniques represented by deep learning. However, due to the large modal differences, diverse semantic description methods, and significant differences in spatial resolution and time scale among multi-source heterogeneous geographic data, traditional fusion methods are mostly centered on single modal data and cannot achieve fine-grained fusion of different modal data in terms of semantics and dimensions. Deep learning methods, on the other hand, usually do not fully consider the cross-modal interaction mechanism between data, which limits the accuracy and confidence of the fusion results.

[0003] CN115269947A discloses a data processing method for multi-source heterogeneous geographic information, which is applied to the measurement of earth resistivity in geological exploration. It realizes multi-channel data acquisition by switching between repeaters and smart electrodes. However, the method focuses on the physical structure design of sensors and measuring equipment, and does not consider the semantic-spatial correlation matching and fusion between cross-modal geographic data. It cannot meet the high requirements of complex geographic information applications for fusion accuracy and semantic consistency.

[0004] CN116719898A discloses a method for generating geographic entities based on multi-source heterogeneous data. Although it considers the standardization and fusion processing of multi-source data, it only performs fusion through a preset entity generation model and does not adopt an effective cross-modal attention mechanism, which limits the accuracy of the fused data. Especially for data sources with significant feature differences, its unified fusion effect is poor. Furthermore, the calculation of geographic entity confidence in the fusion process is not clear enough, and the credibility of data fusion cannot be effectively quantified.

[0005] In summary, existing multi-source heterogeneous geographic data fusion technologies generally suffer from insufficient accuracy in semantic matching and feature fusion of cross-modal data. The present invention aims to solve the problem of how to efficiently and accurately achieve deep fusion processing of multi-source heterogeneous geographic data. Summary of the Invention

[0006] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section, as well as in the abstract and title of the present application, to avoid obscuring the purpose of this section, the abstract and title of the invention. Such simplifications or omissions shall not be used to limit the scope of the present invention.

[0007] In view of the aforementioned existing problems, the present invention is proposed.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution: collecting multi-source heterogeneous geographic data and writing it into a unified index library according to the collection time-spatial metadata;

[0009] The written multi-source heterogeneous geographic data is subjected to coordinate benchmark normalization, time series completion, and joint integrity-correctness verification to obtain standardized geographic data blocks.

[0010] Based on the unique identifiers of geographic entities and spatial topological constraints, semantic-spatial joint matching is performed on the standardized geographic data blocks to generate candidate cross-modal entity pairs.

[0011] A cross-modal attention network is constructed, which includes a raster convolution branch, a vector convolution branch, and a text encoding branch. The candidate cross-modal entity pairs are input into the cross-modal attention network to extract fusion features and calculate the posterior fusion confidence.

[0012] In each candidate cross-modal entity pair, the fusion result with the highest confidence is selected as the target geographic entity, and a unique fusion identifier is generated for the target geographic entity.

[0013] The target geographic entity holding a unique fusion identifier is written into the fusion database, and a unified geographic entity dataset is output.

[0014] As a preferred embodiment of the multi-source heterogeneous geographic data processing method based on deep fusion algorithm described in this invention, the multi-source heterogeneous geographic data includes at least vector feature data, remote sensing raster image data, lidar point cloud data, and POI data.

[0015] As a preferred embodiment of the multi-source heterogeneous geographic data processing method based on deep fusion algorithm described in this invention, the multi-source heterogeneous geographic data is written into a unified index library according to the collection time-spatial metadata, including:

[0016] The collected multi-source heterogeneous geographic data is stored in an index database using a spatial quadtree index and GeoHash grid encoding method, based on metadata information from a unified timestamp and spatial reference system.

[0017] The spatial quadtree index is used to manage the vector feature data and the lidar point cloud data;

[0018] The GeoHash grid encoding is used to manage the remote sensing raster image data and the POI data.

[0019] As a preferred embodiment of the multi-source heterogeneous geographic data processing method based on deep fusion algorithm described in this invention, the standardized geographic data block is obtained, comprising:

[0020] A dynamic seven-parameter spatial transformation model is used to normalize data from different coordinate datums;

[0021] For raster cells with temporal missing data, bidirectional spatiotemporal interpolation is used to fill in the missing data, and adaptive interpolation based on word vector cosine similarity is used for missing vector attributes.

[0022] A joint verification rule base consisting of spatial topology rules, attribute logic rules, and entity overlap detection rules is established. Data blocks that do not meet the rules are removed to form standardized geographic data blocks.

[0023] As a preferred embodiment of the multi-source heterogeneous geographic data processing method based on deep fusion algorithm described in this invention, generating the candidate cross-modal entity pairs includes:

[0024] The improved KD-Tree spatial index is invoked to retrieve spatially nearest neighbor entities within a first threshold.

[0025] SimHash feature vectors are used to calculate the semantic Hamming distance of the text, and only entities with a Hamming distance less than the second threshold are retained.

[0026] Entities that satisfy the first threshold and the second threshold conditions are combined into candidate cross-modal entity pairs, and their initial matching confidence values ​​are recorded.

[0027] As a preferred embodiment of the multi-source heterogeneous geographic data processing method based on deep fusion algorithm described in this invention, a cross-modal attention network is constructed, comprising raster convolution branches, vector convolution branches, and text encoding branches, including:

[0028] The raster convolution branch uses the ResNet-50 network to extract multi-scale texture features layer by layer;

[0029] The vector graphics convolution branch uses graph convolutional networks to extract the topological structure and attribute features of vector data;

[0030] The text encoding branch uses the BERT-Base model to extract contextual semantic vectors;

[0031] The raster convolution branch, the vector convolution branch, and the text encoding branch interact with each other through a multi-head Transformer attention mechanism and align in the shared latent space to generate a fused feature vector.

[0032] As a preferred embodiment of the multi-source heterogeneous geographic data processing method based on deep fusion algorithm described in this invention, candidate cross-modal entity pairs are input into the cross-modal attention network to extract fusion features and calculate posterior fusion confidence, including:

[0033] The candidate cross-modal entity pairs are respectively input into the cross-modal attention network to obtain raster data feature vectors, vector data feature vectors and text data feature vectors;

[0034] By utilizing a cross-modal attention mechanism, attention weights between different data features are calculated one by one, and cross-modal feature fusion is completed through weighted summation to obtain a cross-modal fused feature representation;

[0035] Based on the fusion feature representation, the maximum a posteriori probability method is introduced to calculate the posterior confidence of the fusion result;

[0036] Output the fusion feature representation and posterior confidence of each candidate cross-modal entity pair for the next step of entity confirmation and fusion identifier generation.

[0037] As a preferred embodiment of the multi-source heterogeneous geographic data processing method based on deep fusion algorithm described in this invention, the calculation of posterior fusion confidence includes:

[0038] The conditional probabilities of different geographic entity categories are calculated based on the fused feature representation of the input using a trained feature fusion network.

[0039] By combining the prior probability of the geographic entity category with the conditional probability, the posterior confidence of the fused entity is determined using the maximum a posteriori probability method.

[0040] As a preferred embodiment of the multi-source heterogeneous geographic data processing method based on deep fusion algorithm described in this invention, generating the unique fusion identifier includes:

[0041] Rank the posterior confidence of each candidate cross-modal entity pair;

[0042] The combination of entities with the highest posterior confidence is selected as the final confirmed target geographic entity.

[0043] A unique identifier generation algorithm is used to generate a non-repeatable fusion identifier, wherein the unique identifier generation algorithm uses a hash algorithm that combines spatial location encoding and timestamp to generate a unique identifier.

[0044] As a preferred embodiment of the multi-source heterogeneous geographic data processing method based on deep fusion algorithm described in this invention, the obtained unified geographic entity dataset includes:

[0045] For confirmed target geographic entities, read their unique fusion identifier, unified coordinate geometry and attribute dictionary. Based on the result layer data model of the fusion database, generate a record to be entered into the database for each target geographic entity. The record fields include key primary key, spatial field, attribute field and version field.

[0046] Concurrent conflict detection is performed at the granularity of the unique fusion identifier. If a record with the same name exists, the version number Ver of the newer record is compared. new Old record version number Ver old and new record confidence Conf new Confidence of old records old When Ver is satisfied new Ver old And confidence level Conf new -Conf old ≥ Confidence difference threshold δ conf If the old record is overwritten, it is overwritten; otherwise, the old record is retained and written to the history table.

[0047] The batch COPY / BULK interface is used to write the fused result layer data in the serialized transaction. If the write fails, the whole process is rolled back. If the write succeeds, the transaction is committed.

[0048] After the transaction is committed, asynchronous index maintenance is triggered, the R*-Tree main index and GeoHash bucket cache are incrementally updated, and the memory Skip-List is refreshed to improve real-time query efficiency.

[0049] Based on the unique fusion identifier, multi-source metadata is aggregated, geometric topology is corrected, and a unified geographic entity dataset view is generated by segmenting according to business theme and scale.

[0050] The beneficial effects of this invention are as follows: This invention effectively solves the problems of low data fusion quality, insufficient credibility of the fusion process, and difficulty in cross-modal feature matching in the prior art by going through data collection, preprocessing, matching, deep fusion, fusion decision and final data output. It achieves high-precision, high-efficiency and high-credibility fusion processing of multi-source heterogeneous geographic data as a whole, and significantly improves the fusion accuracy, efficiency and credibility. Attached Figure Description

[0051] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0052] Figure 1 This is a flowchart illustrating the multi-source heterogeneous geographic data processing method based on deep fusion algorithm shown in this invention. Detailed Implementation

[0053] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0054] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort should fall within the scope of protection of this invention.

[0055] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0056] According to an embodiment of the present invention, in combination Figure 1 The flowchart shown illustrates a method for processing multi-source heterogeneous geographic data based on a deep fusion algorithm, which specifically includes the following steps:

[0057] S1. Collect multi-source heterogeneous geographic data and write it into a unified index database according to the collection time-spatial metadata. Note that the following points should be noted in this step:

[0058] The collected multi-source heterogeneous geographic data are stored in the index database based on the metadata information of the unified timestamp and spatial reference system, using a spatial quadtree index and GeoHash (a spatial indexing algorithm that encodes latitude and longitude coordinates into strings) grid encoding method;

[0059] Spatial quadtree indexes are used to manage vector feature data and LiDAR point cloud data;

[0060] GeoHash grid coding is used to manage remote sensing raster image data and POI data.

[0061] In an optional implementation, the method for constructing a spatial quadtree index includes: performing bounding box calculations on vector features and LiDAR point clouds; recursively dividing spatial regions to generate quadtree nodes.

[0062] In an optional implementation, the GeoHash grid encoding method includes: generating a GeoHash string GH based on geographic coordinates (x, y) with a precision of p bits. p =GeoHash(x,y,p); Using GH p Use the key to write the raster slice ID and POI record into the hash bucket.

[0063] In an optional implementation, the method for writing to the unified index library includes: first writing to the spatial quadtree layer, and then obtaining the node path. QT Then write to the GeoHash bucket layer to obtain the bucket key GH. p ; Form a composite primary key Key=T utc ||SRID||Path QT ||GH p Insert into the NoSQL-Spatial engine using the BULK mode (large data transfer mode) to achieve second-level concurrent writes.

[0064] It should be noted that multi-source heterogeneous geographic data includes at least vector feature data, remote sensing raster image data, lidar point cloud data, and POI data.

[0065] As an example, vector feature data includes at least road centerlines, building outlines, and administrative boundaries.

[0066] As an example, remote sensing raster image data includes at least multispectral satellite imagery and UAV orthophotos.

[0067] As an example, lidar point cloud data includes at least airborne LiDAR terrain points and vehicle-mounted LiDAR street points.

[0068] As an example, POI data (point of interest data, which is geospatial information describing a specific location, including location coordinates, name, address, and category attributes) should at least include the point of interest name, category, and latitude and longitude coordinates.

[0069] For example, an airborne LiDAR point cloud was collected at 10:30 UTC on June 1, 2025:

[0070] Timestamp T utc =20250601T103000Z;

[0071] SRID = 4490;

[0072] Meta.sensor = "ALS-60";

[0073] Quadtree path / NE / NE / SW.

[0074] For example, a satellite image slice was acquired at 10:32 UTC on June 1, 2025:

[0075] Image Center (118.773°, 32.065°), GeoHash6 66 =“wtw37p”;

[0076] Store in bucket wtw37p, recording the slice ID and image resolution of 0.5m.

[0077] Preferably, through the above steps, rapid archiving and retrieval of multi-source heterogeneous geographic data such as vector, raster, point cloud and POI are achieved under the framework of unified timestamp + unified coordinate benchmark + two-level spatial index (quadtree / GeoHash), which effectively avoids the phenomenon of data silos, reduces the time consumption of data search and matching during subsequent data fusion, and improves the efficiency and real-time performance of data management and fusion process.

[0078] S2. Perform coordinate benchmark normalization, time series completion, and joint integrity-correctness verification on the written multi-source heterogeneous geographic data to obtain standardized geographic data blocks. Note the following in this step:

[0079] A dynamic seven-parameter spatial transformation model is used to normalize data from different coordinate datums;

[0080] For raster cells with temporal missing data, bidirectional spatiotemporal interpolation is used to fill in the missing data, and adaptive interpolation based on word vector cosine similarity is used for missing vector attributes.

[0081] A joint verification rule base consisting of spatial topology rules, attribute logic rules, and entity overlap detection rules is established. Data blocks that do not meet the rules are removed to form standardized geographic data blocks.

[0082] For example, for the source coordinates (X) s ,Y s Z s And the target benchmark ITRF-2014, applying the dynamic Helmert seven-parameter model:

[0083]

[0084] Where ΔX, ΔY, ΔZ are translation parameters, and R x ,R y ,R z Here, is the micro-angle rotation parameter, s is the scale parameter, and X is the micro-angle rotation parameter. s,Y s Z s The source coordinate system is rectangular coordinates, X t ,Y t Z t The target coordinate system uses rectangular coordinates.

[0085] In an optional implementation, when a raster cell value V(i,j,t) is missing at time t, the time t of the most recent valid previous frame is defined. f With the time t of the next frame b :

[0086]

[0087] Based on the linear time weighting coefficients, the raster cell completion values ​​are obtained:

[0088]

[0089] Missing attributes for vector features From the set of similar elements A = {a k Search for the one with the highest cosine similarity in the array:

[0090]

[0091] in, For the corresponding word vector, with a * As missing fill values, it achieves semantically consistent attribute completion.

[0092] Furthermore, define a set of joint validation rule bases.

[0093] in, This is a subset of spatial topological rules (such as polygons not self-intersecting and road centerlines not crossing water surfaces). For a subset of attribute logic rules (such as building height greater than 0, point cloud intensity I∈[0,255]), This is a subset of entity overlap detection rules (such as the overlap ratio of similar face features not exceeding 5%).

[0094] The rule base is fixed in the form of JSON-Schema+SQL CHECK and is automatically invoked through triggers.

[0095] Furthermore, the joint verification rules are applied to each normalized and padded data block for evaluation:

[0096] If the data does not meet the validation rules, it is marked as invalid and removed; otherwise, it is marked as a valid record and enters the standardized cache area.

[0097] Records in the standardized cache are sorted according to StdKey=T utc||SRID=4490||GridID is renumbered to form a standardized geographic data block. It should be noted that the standardized geographic data block consists of three types of entities: unified coordinate vector features + semantically complete attributes, continuous multi-temporal raster slices with no missing pixels, and strongly consistent coordinate point clouds + POI records.

[0098] It should be noted that by performing coordinate benchmark normalization, time series completion, and joint integrity and correctness verification on the collected multi-source heterogeneous geographic data, a unified spatial representation of data from different coordinate systems was achieved. This effectively repaired the temporal loss or attribute incompleteness issues that occurred during the data collection process, fundamentally ensuring the integrity and accuracy of the data. This refined preprocessing significantly improved the quality of the data before fusion, reduced the possibility of erroneous matching or low-quality fusion results in the subsequent fusion process, and ensured the reliability of the fusion process.

[0099] S3. Based on the unique identifiers of geographic entities and spatial topological constraints, perform semantic-spatial joint matching on standardized geographic data blocks to generate candidate cross-modal entity pairs. Note that the following points should be noted in this step:

[0100] Build an improved KD-Tree index for each standardized entity;

[0101] Using the entity's center point as the retrieval center, search for spatial neighbors based on a first threshold (e.g., 50 meters);

[0102] If no entity is found within this distance range, a pairing will not be created for that entity at this time.

[0103] Furthermore, for the spatially nearest entities found, the SimHash fingerprint of the semantic label is calculated respectively;

[0104] Compare the Hamming distance of two fingerprints, and retain only entity pairs whose Hamming distance is less than the second threshold (e.g., 12 bits) to form candidate entity pairs that satisfy the dual conditions of "spatial proximity + semantic similarity";

[0105] For each candidate entity pair, an initial matching confidence value is assigned to it based on the principle that the closer the spatial distance, the smaller the semantic distance.

[0106] In an optional implementation, the confidence score is a weighted sum of spatial distance contribution and semantic similarity contribution, with the weighting parameter β. d and β h Defined as β d =0.4,β h =0.6;

[0107] Candidate pairs corresponding to the same source entity are sorted from high to low according to their initial confidence values, and the top 5 pairs are retained.

[0108] Record the source entity, candidate entities and their initial confidence values ​​to form a list of candidate cross-modal entity pairs, which will serve as input for subsequent deep fusion processing.

[0109] It should be noted that by accurately corresponding and initially fusing different modal data in terms of spatial location and semantic attributes, the problem of large modal differences and matching difficulties between heterogeneous data is effectively solved. This significantly improves the fusionability and matching accuracy between multi-source heterogeneous data, provides high-quality candidate inputs for deep fusion, and thus improves the overall fusion accuracy.

[0110] S4. Construct a cross-modal attention network containing raster convolution branches, vector convolution branches, and text encoding branches. Input candidate cross-modal entity pairs into the cross-modal attention network, extract fusion features, and calculate posterior fusion confidence. Note that the following should be noted in this step:

[0111] The raster convolution branch uses the ResNet-50 network to extract multi-scale texture tensors from remote sensing slices of candidate entity pairs step by step;

[0112] The vector convolution branch uses graph convolutional networks to perform graph embedding encoding on vector topologies such as roads and buildings;

[0113] The text encoding branch uses the BERT-Base model to generate contextual semantic vectors from entity labels and descriptive text;

[0114] The above three types of features are fed into the multi-head Transformer attention layer to adaptively learn the complementary correlations between the modalities, complete feature alignment in the shared latent space, and output the aligned sequence attention weight curve.

[0115] Global pooling is performed on the alignment weight curve in the normalized time dimension to obtain three attention functions: raster, vector, and text. Then, the three features are converged into a single fusion feature representation through a cross-modal weighted gating module.

[0116] Using the maximum a posteriori probability paradigm, the fused features are fed into the trained discrimination model. The fusion confidence is inferred by combining the entity category prior and the conditional likelihood, and a tuple of fused feature vector + confidence is output for the next step of entity confirmation and fusion label generation.

[0117] In an optional implementation, candidate cross-modal entity pairs are input into a cross-modal attention network, fusion features are extracted, and posterior fusion confidence is calculated, including:

[0118] Candidate cross-modal entity pairs are input into a cross-modal attention network to obtain raster data feature vectors, vector data feature vectors, and text data feature vectors.

[0119] By utilizing a cross-modal attention mechanism, attention weights between different data features are calculated one by one, and cross-modal feature fusion is completed through weighted summation to obtain a cross-modal fused feature representation;

[0120] Based on the fusion feature representation, the maximum a posteriori probability method is introduced to calculate the posterior confidence of the fusion result:

[0121] The conditional probabilities of different geographic entity categories are calculated based on the fused feature representation of the input using a trained feature fusion network.

[0122] By combining the prior and conditional probabilities of geographic entity categories, the posterior confidence of the merged entity is determined using the maximum a posteriori probability method.

[0123] Finally, the fusion feature representation and posterior confidence of each candidate cross-modal entity pair are output for the next step of entity confirmation and fusion identifier generation.

[0124] As an example, the trained discriminative model is used to compute cross-modal attention output and posterior confidence, expressed mathematically as follows:

[0125]

[0126] Among them, C ij Let σ(·) be the posterior fusion confidence of candidate entity pair (i, j), σ(·) be the Sigmoid normalization function, κ be the texture difference penalty factor, and a r (τ), a v (τ), a t (τ) represents the raster, vector, and text attention weight vectors at the normalized time index τ, respectively; N(·) is the zero-mean Gaussian normalization function; and L is the number of Transformer layers. Let be the l-th layer raster, vector, and text depth embedding vector, ReLU(·) be the linear rectified function, and ξ be the text sparsity adjustment coefficient. Let ||·||1 represent the 2-norm and the 1-norm, respectively;

[0127] C ij ∈(0,1), when C ij When the value is close to 1, it indicates that the raster, vector, and text modalities are highly consistent, the spatial differences are minimal, and the semantics are highly matched, resulting in the highest fusion reliability.

[0128] When C ij When the value is close to 0, it indicates that the differences in the three modal features are significant or that the attention cannot be aligned, resulting in the lowest fusion reliability.

[0129] It should be noted that by constructing a cross-modal attention network consisting of raster convolution branches, vector convolution branches, and text encoding branches, and using this network to extract fusion features and calculate posterior fusion confidence, deep fusion and feature interaction of data in the cross-modal feature space are achieved. The cross-modal attention mechanism can accurately capture the potential correlations and feature complementarity between data from different modalities, making the fused features more comprehensive and accurate. The calculated posterior confidence can effectively quantify the credibility of the fusion results, significantly improving the reliability and interpretability of the fusion decision-making process.

[0130] S5. In each candidate cross-modal entity pair, select the fusion result with the highest confidence as the target geographic entity, and generate a unique fusion identifier for the target geographic entity. Note that the following should be noted in this step:

[0131] Rank the posterior confidence of each candidate cross-modal entity pair;

[0132] The combination of entities with the highest posterior confidence is selected as the final confirmed target geographic entity.

[0133] A unique identifier generation algorithm is used to generate a non-repeatable fusion identifier. The unique identifier generation algorithm uses a hash algorithm that combines spatial location encoding and timestamp to generate a unique identifier.

[0134] It should be noted that by selecting the fusion result with the highest fusion confidence as the target geographic entity in each candidate cross-modal entity pair and generating a unique fusion identifier, automatic screening and efficient confirmation of fused entities are achieved. This step can automatically eliminate fuzzy or erroneous matches that may occur during the fusion process, improve the accuracy of fused data, avoid the subjective problems caused by traditional manual judgment or single rule selection methods, and ensure the consistency and accuracy of fusion results.

[0135] S6. Write the target geographic entities with unique fusion identifiers into the fusion database, and output the unified geographic entity dataset. Note the following in this step:

[0136] For confirmed target geographic entities, read their unique fusion identifier, unified coordinate geometry and attribute dictionary. Based on the result layer data model of the fusion database, generate a record to be entered into the database for each target geographic entity. The record fields include key primary key, spatial field, attribute field and version field.

[0137] Concurrency conflict detection is performed at the granularity of a unique fusion identifier. If a record with the same name exists, the version number Ver of the newer record is compared. new Old record version number Ver old and new record confidence Conf new Confidence of old records oldWhen Ver is satisfied new Ver old And confidence level Conf new -Conf old ≥ Confidence difference threshold δ conf If the old record is overwritten, it is overwritten; otherwise, the old record is retained and written to the history table.

[0138] The batch COPY / BULK interface is used to write the fused result layer data in the serialized transaction. If the write fails, the whole process is rolled back. If the write succeeds, the transaction is committed.

[0139] After the transaction is committed, asynchronous index maintenance is triggered, incrementally updating the R*-Tree main index and GeoHash bucket cache, and refreshing the memory Skip-List to improve real-time query efficiency;

[0140] Based on a unique fusion identifier, multi-source metadata is aggregated, geometric topology is corrected, and a unified geographic entity dataset view is generated by segmenting according to business theme and scale.

[0141] It should be noted that by writing target geographic entities with unique fusion identifiers into the fusion database and outputting a unified geographic entity dataset, unified storage and efficient sharing of multi-source heterogeneous data are achieved. This effectively reduces the complexity of downstream applications calling fusion data, significantly improves the sharing and ease of use of the unified dataset, and provides high-quality unified data support for data analysis and business applications of geographic information systems.

[0142] It should be further noted that, in the acquisition phase, the processing method provided in this embodiment constructs a unified index library based on acquisition time-space metadata by uniformly acquiring and organizing multi-source heterogeneous geographic data such as vector feature data, remote sensing raster image data, lidar point cloud data, and POI data. This significantly improves the retrieval efficiency and management convenience of multi-source data, provides an efficient data foundation for subsequent cross-modal data matching and fusion, avoids data silos, and significantly reduces the time and computing resource costs of subsequent fusion calculations.

[0143] Secondly, this invention achieves standardized data processing through a combination of coordinate benchmark normalization, time series completion, and joint verification of integrity and correctness. By comprehensively applying a dynamic seven-parameter spatial transformation model, a bidirectional spatiotemporal interpolation algorithm, and an attribute logic verification rule base, it solves the problems of chaotic, missing, and logically incorrect data benchmarks in existing technologies, improves the intrinsic quality and consistency of data, and significantly reduces the problems of incorrect matching or low fusion quality in the fusion stage.

[0144] This invention performs semantic-spatial joint matching on standardized geographic data blocks based on unique geographic entity identifiers and spatial topological constraints. Through improved KD-Tree spatial retrieval and SimHash semantic matching technology, it effectively achieves accurate semantic-spatial correspondence between data of different modalities, solving the problem of insufficient matching accuracy caused by large modal differences in traditional technologies. Through accurate preliminary matching, it provides higher quality candidate entity pairs, ensuring the accuracy and efficiency of subsequent deep fusion.

[0145] It should be noted again that the embodiments of the present invention construct a cross-modal attention network. This network achieves deep interaction and precise fusion of features between cross-modal data through the deep integration of grid convolutional neural networks, graph convolutional networks and text encoding models. The cross-modal attention mechanism effectively captures the inherent correlation features and complementary information between different data modalities through adaptive weight allocation, significantly improving the expressive power of fused features. In addition, the fusion confidence calculated by the maximum a posteriori probability method provides a quantitative reliability basis for fusion decision-making, effectively solving the problem that the fusion results are difficult to objectively evaluate in the prior art, making the fusion process more scientific, credible and interpretable.

[0146] Regarding the confirmation of fusion results, the embodiments of the present invention automatically select the fusion result with the highest confidence level from the candidate entity pairs as the final target geographic entity based on the calculated posterior confidence level, and automatically generate a globally unique fusion identifier. This avoids the subjectivity and uncertainty problems caused by traditional manual intervention or single rule judgment methods, and greatly improves the consistency, objectivity and accuracy of fusion results.

[0147] Ultimately, this invention achieves unified storage management and efficient sharing and access of heterogeneous data sources by writing confirmed target geographic entities into a fusion database and outputting a unified geographic entity dataset. By adopting the OGC standard service interface to open data resources to external systems, it significantly improves the shareability, usability, and interoperability of fusion data, and enhances the practical application value of multi-source heterogeneous geographic data.

[0148] The aforementioned methods for extracting feature vectors and generating fusion identifiers can be implemented using existing technologies and methods, and will not be elaborated upon in this example.

[0149] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for processing multi-source heterogeneous geographic data based on a deep fusion algorithm, characterized in that, include: Collect multi-source heterogeneous geographic data and write it into a unified index library according to the collection time-spatial metadata; The written multi-source heterogeneous geographic data is subjected to coordinate benchmark normalization, time series completion, and joint integrity-correctness verification to obtain standardized geographic data blocks. Based on the unique identifiers of geographic entities and spatial topological constraints, semantic-spatial joint matching is performed on the standardized geographic data blocks to generate candidate cross-modal entity pairs. A cross-modal attention network is constructed, which includes a raster convolution branch, a vector convolution branch, and a text encoding branch. The candidate cross-modal entity pairs are input into the cross-modal attention network to extract fusion features and calculate the posterior fusion confidence. In each candidate cross-modal entity pair, the fusion result with the highest confidence is selected as the target geographic entity, and a unique fusion identifier is generated for the target geographic entity; The target geographic entity holding a unique fusion identifier is written into the fusion database, and a unified geographic entity dataset is output.

2. The method for processing multi-source heterogeneous geographic data based on a deep fusion algorithm according to claim 1, characterized in that, The multi-source heterogeneous geographic data includes at least vector feature data, remote sensing raster image data, lidar point cloud data, and POI data.

3. The method for processing multi-source heterogeneous geographic data based on a deep fusion algorithm according to claim 2, characterized in that, The multi-source heterogeneous geographic data is written into a unified index library according to the collection time-spatial metadata, including: The collected multi-source heterogeneous geographic data is stored in an index database using a spatial quadtree index and GeoHash grid encoding method, based on metadata information from a unified timestamp and spatial reference system. The spatial quadtree index is used to manage the vector feature data and the lidar point cloud data; The GeoHash grid encoding is used to manage the remote sensing raster image data and the POI data.

4. The method for processing multi-source heterogeneous geographic data based on a deep fusion algorithm according to claim 3, characterized in that, Obtaining the standardized geographic data block includes: A dynamic seven-parameter spatial transformation model is used to normalize data from different coordinate datums; For raster cells with temporal missing data, bidirectional spatiotemporal interpolation is used to fill in the missing data, and adaptive interpolation based on word vector cosine similarity is used for missing vector attributes. A joint verification rule base consisting of spatial topology rules, attribute logic rules, and entity overlap detection rules is established. Data blocks that do not meet the rules are removed to form standardized geographic data blocks.

5. The method for processing multi-source heterogeneous geographic data based on a deep fusion algorithm according to claim 3, characterized in that, Generating the candidate cross-modal entity pairs includes: The improved KD-Tree spatial index is invoked to retrieve spatially nearest neighbor entities within a first threshold. SimHash feature vectors are used to calculate the semantic Hamming distance of the text, and only entities with a Hamming distance less than the second threshold are retained. Entities that satisfy the first threshold and the second threshold conditions are combined into candidate cross-modal entity pairs, and their initial matching confidence values ​​are recorded.

6. The method for processing multi-source heterogeneous geographic data based on a deep fusion algorithm according to claim 1, characterized in that, Construct a cross-modal attention network that includes raster convolution branches, vector convolution branches, and text encoding branches, including: The raster convolution branch uses the ResNet-50 network to extract multi-scale texture features layer by layer; The vector graphics convolution branch uses graph convolutional networks to extract the topological structure and attribute features of vector data; The text encoding branch uses the BERT-Base model to extract contextual semantic vectors; The raster convolution branch, the vector convolution branch, and the text encoding branch interact with each other through a multi-head Transformer attention mechanism and align in the shared latent space to generate a fused feature vector.

7. The method for processing multi-source heterogeneous geographic data based on a deep fusion algorithm according to claim 6, characterized in that, The candidate cross-modal entity pairs are input into the cross-modal attention network to extract fusion features and calculate posterior fusion confidence, including: The candidate cross-modal entity pairs are respectively input into the cross-modal attention network to obtain raster data feature vectors, vector data feature vectors and text data feature vectors; By utilizing a cross-modal attention mechanism, attention weights between different data features are calculated one by one, and cross-modal feature fusion is completed through weighted summation to obtain a cross-modal fused feature representation; Based on the fusion feature representation, the maximum a posteriori probability method is introduced to calculate the posterior confidence of the fusion result; Output the fusion feature representation and posterior confidence of each candidate cross-modal entity pair for the next step of entity confirmation and fusion identifier generation.

8. The method for processing multi-source heterogeneous geographic data based on a deep fusion algorithm according to claim 7, characterized in that, The calculation of the posterior fusion confidence includes: The conditional probabilities of different geographic entity categories are calculated based on the fused feature representation of the input using a trained feature fusion network. By combining the prior probability of the geographic entity category with the conditional probability, the posterior confidence of the fused entity is determined using the maximum a posteriori probability method.

9. The method for processing multi-source heterogeneous geographic data based on a deep fusion algorithm according to claim 7, characterized in that, Generating the unique fusion identifier includes: Rank the posterior confidence of each candidate cross-modal entity pair; The combination of entities with the highest posterior confidence is selected as the final confirmed target geographic entity. A unique identifier generation algorithm is used to generate a non-repeatable fusion identifier, wherein the unique identifier generation algorithm uses a hash algorithm that combines spatial location encoding and timestamp to generate a unique identifier.

10. The method for processing multi-source heterogeneous geographic data based on a deep fusion algorithm according to claim 1, characterized in that, The obtained unified geographic entity dataset includes: For confirmed target geographic entities, read their unique fusion identifier, unified coordinate geometry and attribute dictionary. Based on the result layer data model of the fusion database, generate a record to be entered into the database for each target geographic entity. The record fields include key primary key, spatial field, attribute field and version field. Concurrent conflict detection is performed at the granularity of the unique fusion identifier. If a record with the same name exists, the version number Ver of the newer record is compared. new Old record version number Ver old and new record confidence Conf new Confidence of old records old When Ver is satisfied new Ver old And confidence level Conf new -Conf old ≥ Confidence difference threshold δ conf If the old record is overwritten, it is overwritten; otherwise, the old record is retained and written to the history table. The batch COPY / BULK interface is used to write the fused result layer data in the serialized transaction. If the write fails, the whole process is rolled back. If the write succeeds, the transaction is committed. After the transaction is committed, asynchronous index maintenance is triggered, the R*-Tree main index and GeoHash bucket cache are incrementally updated, and the memory Skip-List is refreshed to improve real-time query efficiency. Based on the unique fusion identifier, multi-source metadata is aggregated, geometric topology is corrected, and a unified geographic entity dataset view is generated by segmenting according to business theme and scale.

Images