A multi-source data fusion real estate ownership intelligent verification method and system

Abstract

The application relates to the technical field of real estate ownership data processing, and discloses a multi-source data fusion real estate ownership intelligent verification method and system. The method collects cadastral, ownership, remote sensing and field data, and constructs a standardized data set through coordinate registration, semantic alignment and Beta distribution reliability evaluation. Spatial topology and logical consistency verification, time sequence perception graph attention network reasoning and twin convolution network remote sensing detection are used to extract three types of verification evidence. Adaptive fusion is carried out by using the D-S evidence theory modified by Jousselme distance, and a comprehensive confidence score and conclusion are output. Key evidence is calibrated through a multi-head attention mechanism, and an explainability report and an abnormal annotation map are generated. The application solves the problems of non-uniform coordinate reference, inconsistent semantics and information loss, realizes automatic intelligent verification of real estate ownership, improves the accuracy and efficiency of verification, and provides reliable data support for real estate registration.

Description

Technical Field

[0001] This invention relates to the field of real estate registration data processing technology, and more specifically, to a method and system for intelligent verification of real estate registration data through multi-source data fusion. Background Technology

[0002] Real estate ownership verification is a fundamental task in real estate registration management. Its core mission is to verify the consistency and legality of multi-source ownership information, including cadastral survey data, ownership registration data, remote sensing image data, and on-site verification data. With the advancement of the unified real estate registration system, a large amount of historical ownership data needs to be incorporated into the unified management system. These data exhibit heterogeneity in coordinate benchmarks, semantic coding, and data formats, posing challenges to batch ownership verification.

[0003] Existing methods for verifying real estate ownership mainly rely on manual comparison or rule-based verification of a single data source. Manual comparison is inefficient and highly subjective, making it difficult to meet the verification needs of large volumes of historical ownership data. While rule-based verification can achieve partial automation, it only performs independent verification on a single data source and lacks the ability to comprehensively analyze the relationships between multiple data sources, thus failing to effectively identify implicit ownership anomalies across data sources.

[0004] Furthermore, existing technologies lack a quantitative processing mechanism for the differences in credibility between different data sources and the conflicting relationships between evidence when handling the fusion of multi-source evidence. When contradictions exist among the multi-source verification evidence, unreasonable fusion conclusions are easily generated. At the same time, the verification conclusions of existing methods lack interpretability, making it difficult for verifiers to trace the supporting evidence for the conclusions, which is not conducive to subsequent manual review and decision-making. Therefore, there is an urgent need for an intelligent verification method that can fuse multi-source heterogeneous property rights data, quantitatively process evidence conflicts, and output interpretable verification conclusions. Summary of the Invention

[0005] This invention provides a method and system for intelligent verification of real estate ownership records by fusing multi-source data, which solves the technical problems in related technologies such as the lack of unified means for fusing and verifying multi-source heterogeneous ownership data, unreasonable fusion conclusions when evidence conflicts, and the lack of interpretability of verification conclusions.

[0006] This invention provides a method for intelligent verification of real estate ownership based on multi-source data fusion, comprising the following steps: S1. Collect multi-source land ownership data, and obtain the multi-source land ownership dataset and data source credibility benchmark vector through control point constraint coordinate registration, ontology semantic alignment and data credibility assessment. S2, apply spatial topological constraint verification and attribute logical consistency rule base joint verification to the multi-source cadastral dataset to obtain the cadastral consistency verification report; S3: Take the multi-source property rights dataset and property rights consistency test report as input, construct a property rights knowledge graph and perform temporal reasoning through graph attention network to obtain the confidence score of property rights legality and the list of abnormal property rights relationships; S4. Acquire multi-temporal remote sensing images and cadastral status maps. Through twin convolutional neural network differential feature extraction and multi-scale semantic segmentation overlay analysis, obtain the remote sensing to cadastral difference map patch set and change confidence matrix. S5 uses the ownership consistency verification report, ownership legality confidence score and change confidence matrix as evidence input, and the credibility benchmark score vector as prior weight. After conflict correction and fusion by DS evidence theory, the comprehensive verification confidence score and verification conclusion label are obtained. S6 takes the comprehensive verification confidence score and verification conclusion label as input, and through multi-head attention key evidence labeling and hierarchical organization of evidence chain, it obtains an interpretable verification report and anomaly space annotation map.

[0007] In a preferred embodiment, S1 includes: The data collection includes cadastral survey data, ownership registration data, multi-temporal remote sensing image data, and on-site verification data, which are then categorized and organized according to the data source institution, data production period, and data format type. Control points of common features are automatically extracted from the original cadastral spatial data. After matching corresponding points, the plane transformation model is estimated by least squares adjustment for two-dimensional coordinate transformation and the seven-parameter Bursa model is estimated for three-dimensional coordinate transformation. Control points with residuals exceeding the first preset threshold are removed and re-estimated. All spatial data are unified to the CGCS2000 coordinate datum. Based on the concept mapping rules of the real estate ownership ontology model, the land category codes in various data sources are uniformly mapped to the standard land category concept, the right type codes are uniformly mapped to the standard right type concept, and the unique identifiers of land parcels are cross-system associated and matched to obtain a standardized ownership dataset. A text recognition model is used to identify and extract key fields from scanned historical paper archives. Fields with a confidence level below a second preset threshold are marked as fields to be verified and recorded in the difference candidate set. The historical verification results of each data source are used as Beta distribution parameters for initialization, and the expected value of the Beta distribution is used as the confidence benchmark score of each data source to form a confidence benchmark score vector.

[0008] In a preferred embodiment, S2 includes: A spatial topology rule base for parcels is constructed, covering rules for non-overlapping parcels, seamless parcels, boundary closure, and geometric validity. Spatial relationships are calculated for each parcel patch to obtain a list of spatial topology errors. A rule library for attribute logic constraints is constructed, covering uniqueness constraints, area consistency constraints, time limit compliance constraints, and usage compliance constraints. The rules are verified one by one for the ownership attribute records of each data source to obtain a list of attribute logic anomalies. Using the unique identifier of the parcel as the association key, cross-modal consistency between spatial data and attribute data is tested. Parcels with area differences exceeding the area difference threshold are marked as area inconsistency anomalies, and parcels with spatial location and administrative region attributes that are inconsistent are marked as location-attribute inconsistency anomalies. Parcels corresponding to low confidence fields in the difference candidate set are included in the summary as attribute doubtful anomalies. All discrepancies are categorized into three levels based on their severity: high, medium, and low. Violations of multiple rules are weighted and accumulated according to the rule independence weights, then normalized and mapped. The weighted average of these sums with the cross-source consistency discrepancy components yields a comprehensive discrepancy confidence score, resulting in a property ownership consistency test report that includes a set of discrepancy patches and a discrepancy confidence matrix.

[0009] In a preferred embodiment, S3 includes: Define the entity type system of the property rights knowledge graph, which includes rights subject entities, real estate unit entities, cadastral map entities, rights certificate entities, and business rule entities, as well as six types of directed relation edges: holding, covering, transferring, dividing, merging, and binding. Historical operation records are extracted from the ownership registration database according to the order of ownership certificate transfer. They are parsed into structured tuples containing operation type, entity identifiers involved, operation date, and registration authority. Entities are instantiated as nodes and relationships as directed edges to construct a cadastral knowledge graph. Land parcels with high severity discrepancies in the cadastral consistency inspection report are injected into the corresponding cadastral map nodes with special marker attributes. A temporally-aware graph attention network is used to learn the representation of knowledge graph nodes. A time encoding vector is introduced for each relation edge. The timestamp is mapped to a time feature vector through a time kernel function to participate in the attention weight calculation. When the timestamp is missing, the average registration time of the historical archives of the local cadastral area is used as the default timestamp and a missing label vector is added. A high-dimensional representation of each node is obtained by stacking three layers of aggregation operation. Reasoning is performed based on node representation in three dimensions: ownership link integrity, ownership uniqueness, and business rule compliance. The confidence scores of the three dimensions are weighted and combined according to preset weights to obtain a list of ownership legality confidence scores and ownership anomalies for each land parcel.

[0010] In a preferred embodiment, S4 includes: Using the current image as a reference, the histogram matching method is used to perform radiometric normalization on the historical images; the feature point matching method is used to extract the same feature points in the two images, and after removing mismatched points by the random sampling consistency algorithm, the residual geometric deviation is eliminated by the affine transformation model. Pixel-level difference features are extracted from dual-temporal remote sensing images using a twin convolutional neural network. Two feature extraction branches with shared weights extract multi-scale semantic feature maps for each phase of the images. The change relationship judgment network generates difference feature maps by calculating element-wise differences and decodes them into pixel-level change probability maps. Connected regions whose change probability exceeds the preset threshold for change detection and whose area exceeds the minimum patch area threshold are extracted as candidate change patches. A semantic segmentation network incorporating a hollow spatial pyramid pooling module is used to classify land cover types in current remote sensing images. The classification results are then spatially overlaid with cadastral status maps. Land cover types that match the remote sensing classification and cadastral registration use types are considered internal morphological changes, while those that do not match are considered use type changes and are marked as high-interest difference areas. The comprehensive change confidence score is calculated from three dimensions: change probability, severity of use difference, and area ratio of difference patches. A change confidence matrix is ​​constructed with the parcel number as the row index and the difference type as the column index. Parcels with change confidence scores exceeding the difference significance threshold are included in the confirmed difference patch set.

[0011] In a preferred embodiment, S5 includes: Based on the identification framework consisting of three mutually exclusive sets of hypotheses—compliance, doubt, and non-compliance—basic trust functions are constructed for cadastral spatial-attribute consistency evidence sources, ownership inference evidence sources, and remote sensing change evidence sources, according to the comprehensive difference confidence value, ownership legality confidence score, and remote sensing change confidence value of each parcel. When the difference confidence value is higher than the high severity threshold, it is mapped to the support of the non-compliance hypothesis; when it is in the medium range, it is mapped to the support of the doubt hypothesis; and when it is lower than the low severity threshold, it is mapped to the support of the compliance hypothesis. The basic trust functions are discounted using the credibility benchmark of each data source as a discount factor. Calculate the Jousselme distance between the basic trust function pairs of each evidence source and construct the evidence conflict degree matrix; for evidence pairs whose conflict degree exceeds the conflict warning threshold, calculate the discount factor based on the mean Jousselme distance and the credibility benchmark score, and apply discount correction to their basic trust function. The basic trust functions of the three types of evidence sources are gradually integrated according to the Dempster combination rule in the order of cadastral consistency evidence, ownership inference evidence, and remote sensing change evidence. The confidence value and similarity value of each hypothesis are extracted from the comprehensive basic trust function, and the average of the two is used as the comprehensive confidence score. The hypothesis with the highest comprehensive confidence score is used as the verification conclusion label. Land parcels with a comprehensive confidence score difference of less than the conclusion distinction threshold among the three types of conclusions are uniformly marked as questionable.

[0012] In a preferred embodiment, S6 includes: The confidence values ​​of each difference type output by S2, the scores of each dimension of ownership reasoning output by S3, and the components of change confidence output by S4 are encoded into feature vectors and concatenated into an evidence sequence; the concatenated vector of the one-hot encoding of the verification conclusion label and the comprehensive confidence score is used as the attention query vector; the attention score between the query vector and each evidence feature vector is calculated through a multi-head attention mechanism to identify key evidence nodes. The key evidence nodes are structured into three levels: verification conclusion layer, evidence source layer, and evidence detail layer. The verification conclusion layer records the verification conclusion label and the comprehensive verification confidence score. The evidence source layer records the intermediate verification conclusion and actual contribution weight of each evidence source. The evidence detail layer records the specific difference information corresponding to each key evidence node. A structured verification report is automatically generated based on a three-layer structure. Key evidence discrepancies are highlighted, and verification suggestions are provided at the end of the report. Using the cadastral map as the base map, the discrepancy patches detected in S2 to S4 are marked with different colored overlay layers to indicate spatial-attribute consistency differences, ownership anomalies, and remote sensing-cadastral differences. An interpretable verification report and anomaly spatial annotation map are jointly output.

[0013] In a preferred embodiment, S3 further includes: When the neighboring nodes of some land parcel nodes in the land ownership knowledge graph are sparse, before performing time-aware graph attention network aggregation, the node encoder is pre-trained by using the node attribute similarity and structural proximity as the basis for constructing positive sample pairs and the contrastive loss function. The output of the pre-trained node encoder is used as the initial node feature vector for time-aware graph attention network aggregation, replacing the original numerical representation of attribute fields.

[0014] In a preferred embodiment, S5 further includes: When the number of evidence sources exceeds a preset threshold, the Dezert-Smarandache theory is used to replace the DS evidence theory framework, and the super-power set is used to replace the standard power set framework, allowing for the intersection of different evidence supporting the hypothesis. In the construction of the basic trust function, conflict quantification, and stepwise fusion calculation, the standard DS framework is replaced with the super-power set framework and the corresponding proportional conflict redistribution combination rules, and the comprehensive verification confidence score and verification conclusion label of each parcel are output.

[0015] In a preferred embodiment, a multi-source data fusion-based intelligent verification system for real estate ownership is used to perform the steps in the aforementioned multi-source data fusion-based intelligent verification method for real estate ownership, including: The data acquisition and preprocessing module is used to collect multi-source cadastral data. After control point constraint coordinate registration, ontology semantic alignment and data credibility assessment, the multi-source cadastral dataset and data source credibility benchmark vector are obtained. The consistency verification module is used to apply spatial topological constraint verification and attribute logical consistency rule base joint verification to the multi-source cadastral dataset to obtain a cadastral consistency verification report. The knowledge graph reasoning module is used to obtain a multi-source property rights dataset and a property rights consistency verification report as input, construct a property rights knowledge graph through graph attention network temporal reasoning, and obtain the property rights legality confidence score and a list of property rights anomaly relationships; The remote sensing change detection module is used to acquire multi-temporal remote sensing images and cadastral status maps. Through the extraction of differential features by twin convolutional neural networks and multi-scale semantic segmentation and overlay analysis, the remote sensing to cadastral difference map patch set and change confidence matrix are obtained. The evidence fusion module is used to take the ownership consistency verification report, ownership legality confidence score and change confidence matrix as evidence input, and the credibility benchmark score vector as prior weight. After DS evidence theory conflict correction and fusion, a comprehensive verification confidence score and verification conclusion label are obtained. The report generation module takes the comprehensive verification confidence score and verification conclusion label as input, and through multi-head attention key evidence labeling and hierarchical organization of the evidence chain, generates an interpretable verification report and an anomaly space annotation map.

[0016] The beneficial effects of this invention are as follows: The system achieves standardized integration of multi-source heterogeneous data through control point constraint coordinate registration, ontology semantic alignment, and credibility assessment of Beta distribution priors. It combines the knowledge graph reasoning of the title system of the temporal-aware graph attention network and the remote sensing change detection of the Siamese convolutional neural network to obtain multi-dimensional verification evidence. Furthermore, it adopts the DS evidence theory with Jousselme distance conflict correction to adaptively fuse multi-source evidence, and can output reasonable comprehensive verification conclusions under the conditions of differences in the credibility of data sources and conflicts between evidence. The multi-head attention mechanism calculates the importance weight of various intermediate evidence nodes, identifies the key evidence nodes that contribute the most to the verification conclusion, and organizes the evidence chain hierarchically according to three levels: verification conclusion layer, evidence source layer, and evidence detail layer. It automatically generates structured verification reports and anomaly space annotation diagrams, enabling verifiers to trace the supporting evidence for the verification conclusion, which facilitates subsequent manual review and decision-making. Attached Figure Description

[0017] Figure 1 This is a flowchart of a multi-source data fusion intelligent verification method for real estate ownership registration according to the present invention; Figure 2 This is a flowchart of a multi-source data fusion-based intelligent verification method for real estate ownership based on the present invention. Detailed Implementation

[0018] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0019] At least one embodiment of the present invention discloses a multi-source data fusion intelligent verification method for real estate ownership, such as... Figures 1 to 2 As shown, it includes the following steps: S1. Collect multi-source land ownership data, and obtain the multi-source land ownership dataset and data source credibility benchmark vector through control point constraint coordinate registration, ontology semantic alignment and data credibility assessment. S11. Collection, Aggregation, and Classification of Multi-Source Cadastral Data; This step aggregates original cadastral data from four main data sources: cadastral survey data, including parcel coordinate strings, boundary point coordinates, parcel area, and parcel number, stored in vector space file format; ownership registration data, including right holder information, right type, right start and end dates, registered area, and use type, stored in structured database tables; multi-temporal remote sensing image data, including high-resolution satellite remote sensing images and UAV aerial photographs from different imaging periods; and on-site verification data, including on-site photographs, measurement records, and text descriptions recorded by verification personnel. The above four types of original data are classified and organized according to the data source institution, data production period, and data format type to establish an original data catalog for subsequent standardization processing.

[0020] S12. Extraction and Adaptive Coordinate Registration of Public Feature Control Points: Historical data often suffers from missing coordinate transformation parameters. Therefore, highly stable public feature control points are automatically extracted from the original cadastral spatial data obtained in S11. The selection criteria are road intersections, corner points of permanent buildings, and waterway confluences—feature points that are clearly identified in measurement results from different periods. After matching the extracted public control points with corresponding points, different processing is performed for different coordinate system transformation types: For two-dimensional transformations involving only differences in planar coordinate systems (such as transformations between different projection zones under the same ellipsoid), at least three pairs of corresponding points are used to estimate a four-parameter or six-parameter planar transformation model using the least squares adjustment method; for three-dimensional coordinate transformations involving different ellipsoids (such as transformations from the Beijing 54 or Xi'an 80 coordinate system to the CGCS2000 coordinate system), at least three pairs of corresponding control points with three-dimensional coordinates are used to estimate a seven-parameter Bursa model containing three translation parameters, three rotation parameters, and one scale parameter. After the transformation parameters are estimated, their reliability is assessed through residual testing. Control points with residuals exceeding a first preset threshold are removed and re-estimated. Based on the estimated coordinate transformation parameters, coordinate transformation is performed on each batch of historical cadastral spatial data to unify all spatial data to the CGCS2000 coordinate datum, eliminating systematic biases of multi-source data at the spatial datum level, and obtaining a spatial dataset with a unified coordinate datum.

[0021] S13. Multi-source semantic element alignment mapping driven by the real estate ownership ontology model; the spatial dataset with unified coordinate benchmark obtained in S12 and the ownership registration data obtained in S11 still have heterogeneity issues at the semantic level, which need to be uniformly expressed through the real estate ownership ontology model. The ontology model covers three core concept systems: spatial entity concepts including land parcels, cadastral patches, real estate units, and boundary lines; ownership relationship concepts including ownership, right of use, mortgage, and easement; and business rule concepts including land use control rules, right term rules, and area calculation rules. Based on the concept mapping rules defined in the ontology model, the land category codes in each data source (covering historical land use classification, urban cadastral classification, and current real estate registration classification) are uniformly mapped to standard land category concepts, and the right type codes in each data source (historical coding systems such as state-owned land use rights, collective land use rights, and house ownership) are uniformly mapped to standard right type concepts. At the same time, cross-system association matching is performed on the unique identifier of the land parcel to establish the correspondence between the spatial records and ownership attribute records of the same land parcel in multi-source data. After ontology semantic alignment processing, a standardized ownership dataset with unified semantic expression of spatial data and ownership attribute data is obtained.

[0022] S14. Data quality assessment and credibility benchmark score calculation for OCR recognition of historical archives and Beta distribution prior; In the standardized land ownership dataset obtained in S13, digitized scanned copies of historical paper archives require supplementary data extraction and quality assessment. An end-to-end text recognition model is used to identify and extract key fields from scanned copies of land ownership certificates, registration applications, and measurement records. Each recognition result is accompanied by a field-level recognition confidence score. The extracted field values ​​are compared with the corresponding records in the database. Fields with recognition confidence scores lower than the second preset threshold are marked as fields to be verified. Their corresponding land parcel numbers, field names, and recognition confidence scores are recorded in the difference candidate set. This difference candidate set is used as a supplementary source of anomalies in the attribute logical consistency test in S2 for summary processing.

[0023] The credibility benchmark scores for each data source are calculated using the Beta distribution prior parameter estimation method: the number of correct verifications and the total number of verifications in the historical verification results of each data source category are used as two parameters of the Beta distribution for initialization; when the number of historical verification samples is insufficient, the initialization is supplemented with prior values ​​of domain experience from the same type of data source, and the parameter estimation is gradually adjusted as subsequent verification cases accumulate using the Bayesian update method; the expected value of the Beta distribution is used as the credibility benchmark score for each data source, forming a credibility benchmark score vector that includes the credibility scores of cadastral survey data, ownership registration data, remote sensing image data, and on-site verification data, which serves as the weight prior input for BBA initialization in the S5 evidence fusion calculation. After processing in S11 to S14, a standardized multi-source cadastral dataset and the data source credibility benchmark score vector are obtained, which serve as the data basis input for each step in S2, S3, S4, and S5.

[0024] In some embodiments, mountainous areas or areas with dense historical buildings may lack a sufficient number of highly stable ground control points, resulting in an insufficient number of available corresponding point pairs and difficulty in guaranteeing the accuracy of registration parameter estimation. In such cases, a control point densification method based on UAV low-altitude photogrammetry can be adopted: in areas with insufficient ground control points, high-resolution orthophotos of the current period are acquired by UAVs, with a flight overlap of not less than 80%. The geometric distortion in the orthophotos is processed using self-calibrated bundle adjustment, and the coordinates of the adjusted image points are inversely calculated into ground point coordinates. These are then used as supplementary control points to participate in the least squares adjustment and registration process described in S12 together with the original ground control points. By increasing the number of control points, the reliability of transformation parameter estimation is improved. Finally, a unified coordinate reference space dataset with the same format as the main process is obtained, which continues to participate in the semantic alignment processing in S13.

[0025] S2, apply spatial topological constraint verification and attribute logical consistency rule base joint verification to the multi-source cadastral dataset to obtain the cadastral consistency verification report; S21. Construction of a Spatial Topology Rule Base for Land Parcels and Detection of Spatial Topology Errors: Based on the cadastral spatial data in the standardized multi-source cadastral dataset obtained in S1, a spatial topology rule base for land parcels is constructed. This rule base covers the following core constraints: No spatial overlap is allowed between parcels of the same level within the same cadastral area (parcel non-overlap rule); no spatial gaps exceeding the topological tolerance range exist between adjacent parcels (parcel seamless rule); no suspended nodes exist between boundary lines (boundary closure rule); the geometric shape of parcels must satisfy the valid polygon condition and must not have self-intersections (geometric validity rule). Based on the above rule base, spatial relationships are calculated for each parcel, recording the parcel number, coordinate range of the violation location, and area of ​​each rule, resulting in a list of spatial topology errors.

[0026] S22. Construction of the Ownership Attribute Logical Constraint Rule Base and Attribute Logic Anomaly Detection: Based on the ownership attribute data in the standardized multi-source cadastral dataset obtained in S1, an attribute logical constraint rule base is constructed. This rule base covers the following core constraints: The same land parcel number cannot simultaneously possess multiple valid ownership certificates of the same type within the same period (uniqueness constraint); the difference between the registered area on the certificate and the cadastral measurement area cannot exceed the area verification threshold (area consistency constraint); the right start date cannot be later than the right end date, and the right term cannot exceed the statutory upper limit (term compliance constraint); there must be no contradiction between the land parcel use type and the planned use of its location (use compliance constraint). Based on the above rule base, each ownership attribute record from each data source is verified according to the rules. The land parcel number, anomaly field name, anomaly type, and anomaly confidence score corresponding to each attribute anomaly are recorded, resulting in an attribute logical anomaly list.

[0027] S23. Spatial-attribute cross-modal consistency comparison of parcel identifier association: Using the unique parcel identifier as the association key, the spatial records of parcel plots from the cadastral spatial database and the ownership attribute records from the ownership registration database in the standardized cadastral dataset obtained in S1 are associated with fields to systematically test the cross-modal consistency between spatial data and attribute data. The test results are combined and summarized with the spatial topology error list obtained in S21 and the attribute logic anomaly list obtained in S22. For spatial and attribute records corresponding to the same unique identifier of a land parcel, calculate the difference in the area field between the two types of records. Land parcels with a difference exceeding the area difference threshold are marked as area inconsistency anomalies. At the same time, check whether the spatial location of the land parcel plot matches the administrative region scope corresponding to the administrative region code recorded in the ownership attribute. Land parcels with inconsistent spatial location and administrative region attribute are marked as location-attribute inconsistency anomalies. In the case of multiple records of the same land parcel in multi-source data, compare the consistency of key attribute fields (area, right type, right holder code) in each record and record the field differences between records across data sources. OCR identifies land parcels corresponding to low confidence fields in the difference candidate set obtained in S14 and includes them in the summary results of this step as attribute doubt anomalies.

[0028] S24. Difference Item Classification Marking and Title Consistency Verification Report Generation: All differences in the verification results from S21 to S23 are classified into three levels according to their impact on title verification: High-severity differences include land parcel spatial overlap, violation of uniqueness constraints, and area differences exceeding the high-severity threshold; Medium-severity differences include spatial gaps, time limit compliance issues, and area differences within the medium-severity threshold; Low-severity differences include dangling nodes, format anomalies, and minor differences in attribute fields across data sources.

[0029] The comprehensive difference confidence score for each parcel is calculated as follows: The base score corresponding to the highest severity level among all detected differences for that parcel (0.7 for high severity, 0.4 for medium severity, and 0.2 for low severity) is taken as the starting point. If multiple rule violations of different types exist simultaneously for the same parcel, the base scores of each rule are weighted and accumulated according to rule independence weights (0.3 for rules of the same type and 0.8 for rules of different types). The accumulated result is then normalized by min-max and mapped to the 0-1 interval. For cases where multiple records exist for the same parcel in multi-source data and there are field differences between the records, the ratio of the number of field differences between records to the total number of key fields is used as the cross-source consistency difference component. This is then weighted and averaged with the rule violation confidence component mentioned above to obtain the final comprehensive difference confidence score. The numbers, location coordinates, type labels, severity levels, and confidence scores of all discrepancies are summarized to form a title consistency verification report containing a set of discrepancy patches and a confidence matrix. The confidence matrix uses the parcel number as the row index and the discrepancy type as the column index. Each element in the matrix records the confidence score of the corresponding parcel for the corresponding discrepancy type. This report serves as the priority input for the S3 knowledge graph construction and as part of the evidence source input for the S5 evidence fusion.

[0030] In some embodiments, changes in data standards at different historical periods can lead to differences in the definition of topology rules across different batches of data (for example, early historical data allowed for a certain range of overlapping gaps between adjacent parcels, while current standards do not). Directly applying a unified topology rule base may misjudge historically compliant data as topology errors. In such cases, a version-adaptive topology rule base method can be used: each topology rule is marked with its applicable data production period range in the rule base. When performing topology verification, the production period mark of the current verification data is read, and a subset of rules applicable to the current version is selected from the rule base based on this period mark. This subset of rules is then used to replace the general rule base in the subsequent topology verification process, thereby obtaining a list of spatial topology errors that is accurately applicable to data from different historical periods, which can then continue to participate in the spatial-attribute cross-modal consistency comparison processing in S23.

[0031] S3: Take the multi-source property rights dataset and property rights consistency test report as input, construct a property rights knowledge graph and perform temporal reasoning through graph attention network to obtain the confidence score of property rights legality and the list of abnormal property rights relationships; S31. Definition of Entity Types and Relationship Types in the Cadastral Knowledge Graph; The entity type system of the cadastral knowledge graph includes 5 core entity nodes: Rights subject entities include legal persons and natural persons, with attributes covering the subject's unique identifier, subject name, subject category, and subject registration date; Real estate unit entity attributes cover the real estate unit number, location address, and unit type; Cadastral patch entity attributes cover the parcel number, coordinate string, area, and cadastral district to which the patch belongs; Rights certificate entity attributes cover the certificate number, rights type, rights start date, rights end date, and registration authority; Business rule entities include land use control rules, rights term rules, and registration procedure rules. The relationship type system defines 6 core directed relationship edges: Holding relationship (rights subject points to rights certificate), Coverage relationship (rights certificate points to cadastral patch), Transfer relationship (source rights subject points to target rights subject via rights certificate), Division relationship (source cadastral patch points to sub-cadastral patch), Merger relationship (source cadastral patch group points to merged cadastral patch), and Constraint relationship (business rule entity points to constrained real estate unit or rights certificate).

[0032] S32. Structured Extraction of Historical Ownership Operation Records and Construction of Knowledge Graph Instances: Based on the knowledge graph pattern defined in S31, entity attribute values ​​and relational connection information required for knowledge graph instantiation are extracted from the standardized cadastral dataset. Historical operation records of ownership transfer, division, merger, and mortgage are extracted from the ownership registration database according to the order of ownership certificate transfer. Each operation record is parsed into a structured tuple containing the operation type, entity identifier involved in the operation, operation date, and operation registration agency. For supplementary fields extracted by OCR of historical archive scans, their recognition confidence scores are marked on the corresponding entity attributes. All entities in the structured tuples are instantiated as nodes. Based on the relational connection information in the operation records, the corresponding types of relational edges are instantiated as directed edges to connect the corresponding nodes, constructing complete instances in the cadastral knowledge graph line by line. At the same time, high-severity difference parcels in the difference patch set output in S2 are injected with special marker attributes into the corresponding cadastral patch nodes, so that the difference information is reflected in the knowledge graph structure, providing difference location priors for subsequent reasoning.

[0033] S33. Knowledge Graph Node Representation Learning Using Temporal Graph Attention Network: In the property rights knowledge graph constructed in S32, the timestamps of historical property rights operation records generally suffer from missing or non-standard issues. To address this challenge, the Temporal Graph Attention Network (TGAT) method is used to learn the representations of various types of nodes in the knowledge graph. While capturing multi-hop property rights relationship structure information, it adaptively reduces the credibility weight of historical property rights relationships with incomplete temporal information. The core mechanism of TGAT is as follows: Based on the neighbor aggregation of standard graph attention networks, a time encoding vector is introduced for each relationship edge. A learnable time kernel function maps the timestamp of the relationship edge to a fixed-dimensional time feature vector. The time feature vector is then concatenated with the feature vectors of neighbor nodes and used together to participate in the calculation of attention weights. When the timestamp information of the relationship edge is complete, the time encoding vector accurately reflects the relative temporal position of the ownership operation in the entire ownership chain, enabling the model to distinguish the ownership status of the same land parcel in different historical periods. When the timestamp information is missing or abnormal, the model reduces the aggregation dependence on the relationship edge through adaptive adjustment of attention weights, and instead relies more on neighbor relationships with complete temporal information, thereby avoiding the misleading of inference conclusions by historical records with chaotic temporal information.

[0034] In terms of model training, the ownership legality annotation results in historically verified ownership cases are used as supervision signals, and the cross-entropy loss function of the node classification task is used to train the TGAT model parameters. When the number of historically verified cases is insufficient, transfer learning is adopted. The feature extraction layer of TGAT is initialized with the graph attention network weights pre-trained on general knowledge graph datasets (such as Wikidata subgraphs or public enterprise relationship graph datasets), and then fine-tuned on a small number of labeled samples in this project to reduce the dependence on a large amount of domain labeled data. The initial node feature vector is formed by the entity type encoding of the node, the numerical representation of key attribute fields, and the S2 difference severity marker. The aggregation layer of TGAT is set to 3 layers, so that the final representation of each node can capture the ownership relationship structure information within its 3-hop neighborhood, thereby covering the depth range of most historical ownership operation chains. For relationship edges with completely missing timestamps, the average registration time of the historical archives of the cadastral area where the parcel is located is used as the default timestamp for time encoding calculation, and a missing marker vector is added in the attention calculation to distinguish it from normal time-series records. By stacking 3 layers of TGAT aggregation operations, the representation vector of each node is updated layer by layer, and finally a high-dimensional representation of each node is obtained. This representation contains both the node's own attribute information and its multi-hop time-series relationship structure information in the ownership chain.

[0035] S34. Multi-dimensional reasoning and scoring of ownership legality under business rule constraints; The node representation vectors obtained in S33 are used to systematically verify the legality of ownership relationships in the knowledge graph. The reasoning and verification are carried out in three dimensions: The first dimension is ownership link integrity reasoning. By performing a depth-first traversal of the holding relationship links of the right holder nodes, it is checked whether each ownership link satisfies the complete transmission path from the initial registration to the current right holder. Cases with broken links or multiple path divergences are marked as ownership link anomalies. For parcels with detected anomalies, the representation vector of the cadastral map node of the parcel is used as input, and a binary classification linear layer is used to output the ownership link integrity confidence score. This linear layer is supervised and trained during the model training phase using the link integrity annotation results of historically verified cases; The second dimension is the ownership link integrity reasoning. The first dimension is uniqueness reasoning, which examines whether there are multiple right certificates with overlapping validity periods and the same right type in the set of right certificate nodes associated with the cadastral map node of the same land parcel. Land parcels with the above situation are marked as ownership uniqueness conflicts, and the conflict confidence score is calculated using the cosine distance between the representation vectors of the overlapping certificate nodes as a difference measure: the larger the cosine distance, the more significant the difference in the ownership relationship structure between the two certificates, and the higher the conflict confidence score. The third dimension is business rule compliance reasoning, which maps the rule content of the business rule entity node to the constrained real estate unit node through the constraint relationship edge, and examines whether the use type, right term and registration procedure of each real estate unit meet the corresponding business rule constraints. For violations, the rule violation confidence score is calculated with the severity of the rule violation as the weight. The confidence scores of the above three dimensions are weighted and combined according to preset weights to calculate the overall ownership legality confidence score of each land parcel node (the higher the score, the higher the legality, and the lower the score, the greater the possibility of anomalies). A list of ownership anomaly relationships containing the ownership anomaly types of each land parcel, the identifiers of the nodes involved in the anomalies, and the anomaly confidence scores is generated as the key evidence input for S5 evidence fusion verification.

[0036] In some embodiments, due to the extremely small number of ownership operation records in some historical cadastral data, the neighbor nodes of some parcel nodes in the knowledge graph are sparse, limiting the aggregation effect of TGAT on sparse graph structures and reducing the quality of node representations. A node representation pre-training method based on graph contrastive learning can be used to enhance the representation of sparse nodes. The aim is to mine prior similarities between nodes from the graph structure through contrastive learning when ownership operation records are scarce, generating higher-quality initial representations for sparse nodes. Before performing TGAT aggregation in S33, the node encoder is pre-trained using the attribute similarity and structural proximity of nodes in the knowledge graph as positive sample pairs. This pre-trains nodes with similar attributes and structural proximity in the representation space. The output of the pre-trained node encoder is used as the initial node feature vector for TGAT aggregation, replacing the original numerical representation of attribute fields, thereby providing more discriminative initial features for sparse nodes and improving the representation quality of subsequent TGAT aggregation. The remaining steps are consistent with the main process, ultimately outputting high-dimensional representation vectors for each node to participate in the ownership legitimacy reasoning and scoring in S34.

[0037] S4. Acquire multi-temporal remote sensing images and cadastral status maps. Through twin convolutional neural network differential feature extraction and multi-scale semantic segmentation overlay analysis, obtain the remote sensing to cadastral difference map patch set and change confidence matrix. S41. Radiometric Normalization and Spatial Fine Registration of Multi-Temporal Remote Sensing Images: The standardized multi-temporal remote sensing image data obtained in S1 needs to eliminate systematic differences between images from different periods to provide consistent dual-temporal image input for subsequent change detection. For radiometric normalization, using the current image as a reference, a histogram matching method is employed to perform radiometric normalization on historical images: Histograms of pixel value distributions in each band of historical and current images are calculated. Using the histogram distribution of the current image as the target, a band-by-band mapping function is established from the pixel values ​​of each band of historical images to the target distribution. This mapping function is then applied to all pixels of the historical images to eliminate radiometric differences caused by sensor type differences and changes in atmospheric conditions. In terms of spatial registration, the orthorectification results under the cadastral coordinate datum are used as a reference to perform orthorectification on the remote sensing images. The feature point matching method is used to extract the same feature points in the two images. After removing mismatched points by the Random Sampling Consensus (RANSAC) algorithm, the residual geometric deviation between the two images is eliminated by the affine transformation model to ensure that the spatial alignment accuracy of the two images under the cadastral coordinate system meets the requirements of subsequent overlay analysis.

[0038] S42. Siamese Convolutional Neural Network for Dual-Temporal Image Differential Feature Extraction and Change Probability Map Generation: The radiometrically normalized and spatially precisely registered dual-temporal remote sensing images obtained in S41 are used to extract pixel-level differential features through a Siamese Convolutional Neural Network (Siamese CNN) to generate a change probability map. The Siamese CNN consists of two feature extraction branches with shared weights and one change relationship judgment network: the two feature extraction branches have identical structures and independently receive historical and current images as inputs, extracting multi-scale deep semantic feature maps of each period's images through multi-layer convolutional operations; the change relationship judgment network receives the feature maps output from the two branches, generates a differential feature map by calculating element-wise differences, and decodes the differential feature map into a pixel-level change probability map with the same spatial resolution as the input images through subsequent fully convolutional layers. The change probability value of each pixel reflects the probability that the land cover type at that pixel location changes between two periods. For model training, the Siamese convolutional neural network was pre-trained using publicly available remote sensing change detection benchmark datasets (such as LEVIR-CD and WHU-CD), and then fine-tuned using manually labeled change patches from historical and current remote sensing images of this project. The model parameters were optimized using a binary cross-entropy loss function. When labeled samples were insufficient, data augmentation methods (including random flipping, random cropping, and radiometric perturbation) were used to expand the training samples. Connectivity analysis was performed on all pixel regions in the generated change probability map whose change probability values ​​exceeded the first threshold for change detection. Connected regions with areas exceeding the minimum patch area threshold were extracted as candidate change patches, which served as input for subsequent overlay analysis with the cadastral status map.

[0039] S43. Fine-grained classification of current land cover types using a multi-scale semantic segmentation network and overlay analysis with cadastral status maps; In the candidate change patches obtained in S42 and the current remote sensing image, historical industrial park scenes suffer from dense buildings, complex land cover types, and difficulties in directly judging changes in the internal functions of industrial land from their appearance. Therefore, a semantic segmentation network integrating multi-scale contextual features is used to perform fine-grained land cover type classification on the current remote sensing image, combined with spatial overlay analysis of the cadastral status map to identify inconsistencies between the current land cover status and the cadastral record. The semantic segmentation network adopts an encoder-decoder structure, introducing Atrous Spatial Pyramid Pooling (ASPP) between the high-level feature map output by the encoder and the decoder. The pooling module serves as the bottleneck layer for feature enhancement. It extracts contextual features at different receptive field scales through multiple dilated convolutions with varying dilation rates (1, 6, 12, and 18). These multi-scale features are then concatenated and fused using a 1×1 convolution before being input into the decoder for progressive upsampling to restore spatial resolution. The final output is a pixel-by-pixel land cover type classification result. Based on the main categories of construction land, cultivated land, forest land, water bodies, roads, and other unused land, the land cover type system is further refined for industrial park scenarios. Land use subcategories are distinguished into industrial plants, warehouse land, office buildings, and paved ground to support the identification of changes in the internal function of industrial land. The expansion rate parameters of the ASPP module are set to 1, 6, 12, and 18, and the corresponding receptive field ranges are approximately 3×3, 15×15, 27×27, and 39×39 pixels, respectively. For the high-resolution remote sensing imagery used in this scenario (spatial resolution 0.5m to 2m), it can effectively cover the area from individual buildings (approximately 10m to 30m) to the entire industrial plant area (approximately 50m to 10m). The system captures multi-scale land cover ranges (0m) to simultaneously capture building outline details and contextual semantic information of land use functional blocks in the same feature extraction layer. The semantic segmentation network is pre-trained using publicly available remote sensing semantic segmentation datasets and then fine-tuned using manually labeled land cover type samples from the current remote sensing images of this project. The semantic segmentation results are spatially overlaid with the cadastral status map (containing the registered use types of each parcel) under a unified coordinate reference. For each candidate change patch area, the degree of conformity between the remote sensing classification results and the cadastral record use type is calculated: if the main land cover type in the remote sensing classification of the candidate change patch area is consistent with the corresponding cadastral registered use type of the area, the change is determined to be an internal morphological change with a low impact on the verification conclusion; if the main land cover type in the remote sensing classification of the candidate change patch area is inconsistent with the corresponding cadastral registered use type of the area, the change is determined to be a use type change and marked as a high-concern difference area. At the same time, the cadastral parcel number, area of ​​the difference area, remote sensing classification land cover type, and cadastral registered use type of the difference area are recorded as difference information tuples.

[0040] S44. Screening and Grading of Differential Patches and Construction of Change Confidence Matrix: The set of differential information tuples obtained in S43 is screened and graded using a comprehensive differential confidence assessment method to construct a change confidence matrix. For each differential information tuple, its change confidence is calculated from the following three dimensions: The first dimension is the change probability dimension, using the mean of the change probability of all pixels within the corresponding patch area in S42 as the image layer change confidence component of the patch; the second dimension is the use difference dimension, assigning a difference severity weight based on the severity of the difference between the remote sensing classified land cover type and the cadastral registered use type; the third dimension is the area dimension, using the ratio of the differential patch area to the total area of ​​the land parcel as a quantitative indicator of the degree of area difference, and grading and scoring according to the area difference threshold. The scores from the three dimensions mentioned above are weighted and summed according to preset weights to calculate the comprehensive change confidence score for each differential patch. A change confidence matrix is ​​constructed using the parcel number as the row index and the difference type as the column index. Each element in the matrix records the comprehensive change confidence score of the corresponding parcel for the corresponding difference type. Parcels with change confidence scores exceeding the difference significance threshold are included in the confirmed differential patch set. Finally, the confirmed differential patch set and the change confidence matrix are used as the output of step S4 to transmit remote sensing verification evidence to step S5.

[0041] In some embodiments, due to the wide geographical coverage of certain real estate registration verification scenarios, a single Siamese convolutional neural network may have reduced detection capability for cross-scale land cover changes when processing large-scale imagery due to the limitation of the global feature receptive field. This makes it difficult to effectively distinguish between overall changes in large-scale industrial land and localized renovations of small-scale buildings. Therefore, a Siamese Transformer network with multi-scale feature fusion can be used to replace the Siamese convolutional neural network for differential feature extraction. Its purpose is to capture the global contextual relationships of cross-scale land cover changes in large-scale remote sensing imagery through the global self-attention mechanism of the transformer architecture, thereby improving the ability to distinguish between different spatial scale change types. The Siamese Transformer network uses a Vision Transformer (ViT) as the backbone network for feature extraction. It divides the input image into a sequence of image patches of fixed size and models the global relationship between all image patches through a multi-head self-attention mechanism to extract image features containing global contextual information. After the global feature maps of the two-phase images are extracted by the two Siamese branches respectively, the change relationship judgment network performs cross-phase attention calculation on the two feature maps. The attention weights reflect the significance of the change of each pixel position between phases and generate a change probability map. The difference feature extraction process of the Siamese convolutional neural network in S42 is replaced by the Siamese Transformer network mentioned above. The remaining steps S43 and S44 are consistent with the main process. Finally, the remote sensing-cadastral difference patch set and change confidence matrix are still output to continue to participate in the multi-source evidence fusion verification in S5.

[0042] S5 uses the ownership consistency verification report, ownership legality confidence score and change confidence matrix as evidence input, and the credibility benchmark score vector as prior weight. After conflict correction and fusion by DS evidence theory, the comprehensive verification confidence score and verification conclusion label are obtained. S51. Construction of an adaptive basic trust function driven by the data source credibility benchmark score; The data source credibility benchmark score vector obtained in S1, together with the verification conclusions of steps S2 to S4, are used to initialize the trust allocation of various types of evidence sources on the verification conclusion framework. The core mechanism of DS evidence theory is: Given an identification framework (in this scheme, the identification framework is defined as a set of three mutually exclusive assumptions of the verification conclusion: compliance, doubt, and violation), a basic trust function is assigned to each type of evidence source. This function maps each non-empty subset of the identification framework to a trust value between 0 and 1, and the sum of the trust values ​​of all subsets equals 1; the trust value reflects the degree to which the evidence supports the corresponding conclusion subset. The amount of trust allocation to the entire set of the identification framework is called the uncertainty component, reflecting that the degree of support of the evidence for a specific conclusion is still unclear. For each parcel, a BBA is constructed for the cadastral spatial-attribute consistency evidence source based on the comprehensive difference confidence value of the corresponding row in the difference confidence matrix output by S2: When the difference confidence is higher than the high severity threshold, the difference confidence is mapped proportionally to the support of the violation hypothesis, a certain proportion of the remaining part is allocated to the questionable hypothesis, and the rest is allocated to the uncertainty component of the whole set, with the support of the compliance hypothesis set to 0; when the difference confidence is in the medium range, the difference confidence is mapped to the support of the questionable hypothesis, with a small support each for the violation and compliance hypotheses, and the remainder allocated to the uncertainty component; when the difference confidence is lower than the low severity threshold, 1 minus a certain proportion of the difference confidence is used as the support of the compliance hypothesis, and the remainder is allocated to the uncertainty component; finally, the initial BBA is discounted using the confidence benchmark of cadastral survey data and ownership registration data in S1 as a discount factor. The discounting process reduces the confidence value of each specific hypothesis subset proportionally according to the discount factor, and the reduced part is transferred to the uncertainty component of the whole set. For the ownership legitimacy confidence score output by S3, it is converted into ownership anomaly confidence score (i.e., ownership anomaly confidence score is obtained by subtracting ownership legitimacy confidence score from 1). Then, a BBA (Best-Balanced Assessment) for ownership inference evidence sources is constructed using a similar segmented mapping method: high ownership anomaly confidence scores are mapped to support for the violation hypothesis, medium scores to support for the questionable hypothesis, and low scores to support for the compliance hypothesis. The estimated inference accuracy of the TGAT model on historical verification samples is used as a discount factor. For the comprehensive change confidence value corresponding to the parcel in the remote sensing-cadastral change confidence matrix output by S4, a BBA for remote sensing change evidence sources is constructed, and the credibility benchmark of remote sensing image data is used as a discount factor. By mapping the difference type classification of remote sensing change confidence scores, the support allocation of the corresponding verification conclusion hypothesis is obtained: changes in land use type correspond to the violation hypothesis, changes in internal morphology correspond to the questionable hypothesis, and no change corresponds to the compliance hypothesis. Through the above steps, an initial BBA set from three types of evidence sources is generated for each parcel.

[0043] S52, Jousselme Distance Conflict Quantification and High-Conflict Evidence Correction: The initial BBA sets of the three types of evidence sources obtained in S51 are used to identify high-conflict evidence pairs through the Jousselme distance conflict quantification method, and the BBAs of high-conflict evidence are corrected to prevent counterintuitive fusion conclusions from arising in high-conflict scenarios using the traditional Dempster combination rule. The Jousselme distance is a method that measures the vectorized differences between BBAs on the identification framework. By constructing a Jaccard similarity matrix between each subset of the identification framework, the Euclidean distance between two BBA vectors in the inner product space defined by the similarity matrix is ​​calculated. The Jousselme distance value ranges from 0 to 1; a larger value indicates a higher degree of difference between the two BBAs, i.e., a higher degree of evidence conflict. For all BBA pairs of evidence sources for each parcel, the Jousselme distance is calculated for each pair to construct an evidence conflict degree matrix. Evidence pairs with a conflict degree exceeding the conflict warning threshold are identified as high-conflict evidence pairs. A discount factor is applied to each BBA in the high-conflict evidence pair. The formula for calculating the discount factor α is: α = w_c × (1 - d_J) + (1 - w_c) × r_s, where d_J is the mean Jousselme distance between the evidence source and other evidence sources, r_s is the credibility baseline score of the evidence source, and w_c is the conflict weight coefficient (ranging from 0 to 1, with a default value of 0.6, which can be adjusted according to the actual scenario); the smaller the discount factor α, the greater the discount the evidence source receives. When an evidence source is highly conflicted with multiple other evidence sources, its mean d_J is larger, resulting in a smaller α, and the evidence source receives a greater discount; the confidence values ​​of each specific hypothesis subset in its BBA are reduced proportionally by the discount factor α, and the reduced part is transferred to the uncertainty component of the whole set, thereby reducing the influence weight of the highly conflicting evidence in subsequent fusion and reducing the probability of counterintuitive fusion results.

[0044] S53, Dempster combination rule progressive fusion and verification conclusion label generation; The conflict-corrected BBA sets of each evidence source obtained in S52 are progressively fused through Dempster combination rule to generate a comprehensive verification confidence score and verification conclusion label. The fusion process of the Dempster combination rule is as follows: For any two modified BBAs, the joint product of the two BBAs on each subset of the identification framework is calculated and normalized (the normalization coefficient is 1 minus the sum of the products of all mutually exclusive subsets, i.e., the conflict quality), to obtain the fused new BBA; the BBAs of the three types of evidence sources are gradually fused in pairs in the order of cadastral consistency evidence → ownership reasoning evidence → remote sensing change evidence, finally obtaining the comprehensive BBA for each parcel; from the comprehensive BBA, the confidence Bel value and the corresponding plausibility Pl value of the three single hypothesis subsets of compliance, doubt, and violation are extracted, and (Bel+Pl) / 2 is used as the comprehensive confidence score of each verification conclusion hypothesis; the verification conclusion hypothesis with the highest comprehensive confidence score is selected as the final verification conclusion label, and the comprehensive confidence score itself is output as the quantitative reliability index of the verification conclusion; for parcels where the difference in the comprehensive confidence scores of the three types of verification conclusions is lower than the conclusion distinction threshold, they are uniformly marked as doubtful and trigger the subsequent manual review process.

[0045] In some embodiments, since some historical land ownership data contain evidence from more than five sources (e.g., additional land price assessment data, construction project archive data, etc.), the computational complexity of traditional Dezert-Smarandache (DS) evidence theory increases significantly as the number of focal elements in the identification framework grows exponentially with the increase in evidence sources, leading to decreased processing efficiency when verifying a large number of land parcels in batches. The Dezert-Smarandache (DSmT) theory can be used to replace the DS evidence theory framework to handle high-dimensional multi-source evidence fusion problems. Its aim is to maintain the comprehensive reasoning ability of multi-source evidence while effectively reducing the computational scale of the focal element space through the compact representation of the super-power set framework using DSmT, thereby improving processing efficiency in large-scale verification scenarios. DSmT replaces the power set framework of standard DS theory with a super-power set (a complete set family generated by the union and intersection operations of all subsets in the identification framework). This allows for the intersection of hypotheses supported by different evidence, thus avoiding the problem of computational instability caused by the normalization denominator approaching 0 in the case of strong conflicts in standard DS theory. In the BBA construction, conflict quantification, and stepwise fusion calculations in S51 to S53, the standard DS framework is replaced with the DSmT super-power set framework and the corresponding PCR5 combination rule (proportional conflict redistribution rule). The remaining steps are consistent with the main process, and the final output is still the comprehensive verification confidence score and verification conclusion label for each parcel, which continues to participate in the generation of the interpretable report in S6.

[0046] S6 takes the comprehensive verification confidence score and verification conclusion label as input, and through multi-head attention key evidence labeling and hierarchical organization of evidence chain, it obtains an interpretable verification report and anomaly space annotation map. S61. Calculation of Importance Weights for Key Evidence Nodes Using a Multi-Head Attention Mechanism: The comprehensive verification confidence score and verification conclusion label for each parcel obtained in S5, as well as the intermediate verification conclusions output from steps S2 to S4, are used to calculate the importance weights of various intermediate evidence nodes through a multi-head attention mechanism, identifying the key evidence nodes that contribute the most to the final verification conclusion. Compared to directly using the BBA contribution weights during the S5 fusion process, the multi-head attention mechanism can capture the association patterns between evidence and verification conclusions from different semantic perspectives through multiple attention heads with different parameters, identifying evidence combinations with differentiated importance under specific verification conclusion types (compliant / questionable / illegal), rather than simply reflecting the overall credibility weight of the evidence; for example, for a violation conclusion, the combined correlation between evidence of changes in remote sensing use and evidence of broken ownership links may be much higher than the sum of their individual BBA contribution weights, and the multi-head attention mechanism can capture this synergistic association effect between evidence. The verification evidence for each parcel is organized into a sequence input for attention calculation: the difference confidence values ​​of each difference type output by S2, the ownership inference scores of each dimension output by S3, and the change confidence components output by S4 are encoded into fixed-dimensional feature vectors, and these feature vectors are concatenated into an evidence sequence; the one-hot encoded vector of the verification conclusion label and the concatenated vector of the comprehensive confidence score are used as the attention query vector; the multi-head attention mechanism calculates the scaled dot product attention score between the query vector and each evidence feature vector in the evidence sequence through multiple attention heads with different parameters, and the outputs of each attention head are concatenated and linearly projected to obtain the final attention weight distribution; the parameters of the multi-head attention mechanism are passed through... The attention parameters are determined as follows: key evidence nodes manually annotated in historical verified cases are used as supervision signals, and the KL divergence between the attention weight distribution and the distribution of manually annotated key evidence is used as the loss function to train the attention parameters. When there are insufficient historical annotated samples, the actual contribution weight of each evidence source during the S5 fusion process (i.e., the BBA confidence allocation after discount correction) is used as the initial value of the attention weight. On this basis, it is fine-tuned with a small number of annotated samples so that the attention mechanism can further learn the synergistic correlation effect between evidence based on the BBA contribution weight. The evidence nodes with higher attention weights are the key evidence that contributes the most to the final verification conclusion and are marked as key evidence nodes for output.

[0047] S62. Hierarchical Organization of Key Evidence Chains and Extraction of Logical Reasoning Paths: The key evidence node set obtained in S61 is structured and organized using a hierarchical evidence chain organization method to extract the logical reasoning paths supporting the verification conclusions. Key evidence is organized according to the following three levels: Level 1 is the verification conclusion level, recording the final verification conclusion label (compliant / questionable / non-compliant) and comprehensive verification confidence score; Level 2 is the evidence source level, grouping key evidence nodes according to the sources in steps S2 to S4, recording the intermediate verification conclusions of each evidence source and its actual contribution weight in the final fusion; Level 3 is the evidence detail level, recording the specific difference information corresponding to each key evidence node, including: for evidence from S2 sources, recording the specific difference type label, difference location coordinate range, and difference severity level; for evidence from S3 sources, recording the specific ownership anomaly type, the identifier of the involved rights holder node, and the anomaly confidence level; for evidence from S4 sources, recording the specific changed map location, the difference description between remote sensing classification land cover type and cadastral registration use type, and the change confidence score. Through the above three-layer structure, the verification conclusions are linked with the contributions and details of each source of evidence to form a complete logical reasoning path.

[0048] S63. Generation of Structured Verification Report and Output of Anomaly Spatial Annotation Map: Based on the hierarchical organization of the verification conclusions obtained in S62, an interpretable verification report and anomaly spatial annotation map are output using a structured report generation method. Specifically, in terms of report content, the structured verification report is automatically generated according to the following format: The report header includes basic information of the verified parcel (parcel number, location description, verification time) and a summary of the verification conclusions (verification conclusion labels and comprehensive confidence score); the report body unfolds layer by layer according to the three-layer structure described in S62, using natural language templates to transform the technical indicators of each layer of evidence into descriptive text that can be understood by non-technical personnel, highlighting the main supporting evidence for the verification conclusions; for discrepancies marked as key evidence, they are highlighted in the report; the report footer provides verification recommendations for the corresponding parcel (including specific matters to be verified on-site and types of supplementary materials to be retrieved). Regarding the anomaly spatial annotation map, using the cadastral map as the base map, all discrepancies detected in S2 to S4 are visualized and annotated with overlay layers of different colors: green areas indicate spatial-attribute consistency differences, blue areas indicate the scope of parcels associated with ownership anomalies, and red areas indicate areas with significant remote sensing-cadastral differences. When multiple types of differences are superimposed in the same parcel area, a mixed color scheme is used for display, and key difference information is annotated with text bubbles in the annotation map. The structured verification report and the anomaly spatial annotation map are jointly output to form the final interpretable verification report, completing the entire intelligent verification process for real estate ownership.

[0049] In some embodiments, since some verifiers need to further understand how the verification conclusion would change if a certain piece of evidence were absent in order to assess the substantial impact of each piece of evidence on the conclusion, relying solely on the interpretation of attention weight distribution may not be sufficient to meet such counterfactual reasoning needs. An interpretability enhancement method based on counterfactual perturbations can be used to quantify the substantial impact of key evidence. The aim is to provide verifiers with a quantified causal impact of each piece of key evidence on the verification conclusion, supporting deeper verification decisions. Specifically, for each identified key evidence node, the evidence node is temporarily removed during the S5 fusion calculation (its BBA is replaced with the average). (Uniform uncertainty allocation) and re-execute the fusion calculation of S52 and S53 to obtain the counterfactual verification conclusion and counterfactual comprehensive confidence score when the evidence node is missing; the difference between the true comprehensive confidence score and the counterfactual comprehensive confidence score is defined as the causal contribution of the evidence node, and the key evidence is sorted in descending order of causal contribution; the counterfactual causal contribution index is added to the evidence detail layer of the structured report in S63, and the substantive impact of each key piece of evidence is quantitatively explained in the format that the confidence of the verification conclusion will decrease by X points if the evidence is missing, so that the explanatory power of the verification report is further enhanced on the basis of attention weight.

[0050] This invention takes the historical property rights verification project in the urban area of ​​a municipality as the application scenario, focusing on the application field of batch verification of real estate property rights in an old industrial park built from the 1990s to the early 2000s. There are about 2,000 construction land parcels in the park, involving multiple administrative division adjustments and historical renewal of real estate property certificates. The data standards span two systems: Beijing 54 coordinate system and CGCS2000 coordinate system.

[0051] Through steps S1 to S4, the system aggregates multi-source cadastral data from the cadastral survey database, the real estate registration database, three periods of high-resolution remote sensing imagery from 2022 to 2024, and field verification records, and completes standardization processing. Tables 1 and 2 show examples of the system's multi-source data acquisition results on typical land parcel samples: Table 1. Examples of Multi-Source Data Collection Results for Typical Land Parcels

[0052] Table 2. Examples of Intermediate and Comprehensive Verification Results for Land Parcel Verification

[0053] The results from the above application examples show that: Parcel YQ2024-0102 has multiple discrepancies, including area inconsistencies, incomplete ownership links, and discrepancies between the current status of remotely sensed features and the registered use. After integrating multiple sources of evidence using the S5 Jousselme distance-corrected DS fusion algorithm, it was determined to be non-compliant, with a high overall confidence score, indicating good consistency among the three types of evidence sources. Parcel YQ2024-0104 has the dual problems of overlapping ownership periods and changes in remotely sensed use, and was also determined to be non-compliant. Parcels YQ2024-0101 and YQ2024-0105 have only minor differences of a single type; after fusion, they were determined to be questionable and triggered manual review, meeting the actual need for manual verification of low-confidence conclusions. Parcel YQ2024-0103 showed no abnormalities in any dimension of verification and was determined to be compliant, with the highest confidence score in this batch, demonstrating the reliability of the fusion conclusion when multiple sources of evidence are highly consistent. This application example verifies the practicality and effectiveness of the method described in this invention in complex historical property rights verification scenarios.

[0054] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for intelligent verification of real estate ownership certificates through multi-source data fusion, characterized in that, Includes the following steps: S1. Collect multi-source land ownership data, and obtain the multi-source land ownership dataset and data source credibility benchmark vector through control point constraint coordinate registration, ontology semantic alignment and data credibility assessment. S2, apply spatial topological constraint verification and attribute logical consistency rule base joint verification to the multi-source cadastral dataset to obtain the cadastral consistency verification report; S3: Take the multi-source property rights dataset and property rights consistency test report as input, construct a property rights knowledge graph and perform temporal reasoning through graph attention network to obtain the confidence score of property rights legality and the list of abnormal property rights relationships; S4. Acquire multi-temporal remote sensing images and cadastral status maps. Through twin convolutional neural network differential feature extraction and multi-scale semantic segmentation overlay analysis, obtain the remote sensing to cadastral difference map patch set and change confidence matrix. S5 uses the ownership consistency verification report, ownership legality confidence score and change confidence matrix as evidence input, and the credibility benchmark score vector as prior weight. After conflict correction and fusion by DS evidence theory, the comprehensive verification confidence score and verification conclusion label are obtained. S6 takes the comprehensive verification confidence score and verification conclusion label as input, and through multi-head attention key evidence labeling and hierarchical organization of evidence chain, it obtains an interpretable verification report and anomaly space annotation map.

2. The intelligent verification method for real estate ownership certificates based on multi-source data fusion according to claim 1, characterized in that, S1 includes: The data collection includes cadastral survey data, ownership registration data, multi-temporal remote sensing image data, and on-site verification data, which are then categorized and organized according to the data source institution, data production period, and data format type. Control points of common features are automatically extracted from the original cadastral spatial data. After matching corresponding points, the plane transformation model is estimated by least squares adjustment for two-dimensional coordinate transformation and the seven-parameter Bursa model is estimated for three-dimensional coordinate transformation. Control points with residuals exceeding the first preset threshold are removed and re-estimated. All spatial data are unified to the CGCS2000 coordinate datum. Based on the concept mapping rules of the real estate ownership ontology model, the land category codes in various data sources are uniformly mapped to the standard land category concept, the right type codes are uniformly mapped to the standard right type concept, and the unique identifiers of land parcels are cross-system associated and matched to obtain a standardized ownership dataset. A text recognition model is used to identify and extract key fields from scanned historical paper archives. Fields with a confidence level below a second preset threshold are marked as fields to be verified and recorded in the difference candidate set. The historical verification results of each data source are used as Beta distribution parameters for initialization, and the expected value of the Beta distribution is used as the confidence benchmark score of each data source to form a confidence benchmark score vector.

3. The intelligent verification method for real estate ownership certificates based on multi-source data fusion according to claim 1, characterized in that, S2 includes: A spatial topology rule base for parcels is constructed, covering rules for non-overlapping parcels, seamless parcels, boundary closure, and geometric validity. Spatial relationships are calculated for each parcel patch to obtain a list of spatial topology errors. A rule library for attribute logic constraints is constructed, covering uniqueness constraints, area consistency constraints, time limit compliance constraints, and usage compliance constraints. The rules are verified one by one for the ownership attribute records of each data source to obtain a list of attribute logic anomalies. Using the unique identifier of the parcel as the association key, cross-modal consistency between spatial data and attribute data is tested. Parcels with area differences exceeding the area difference threshold are marked as area inconsistency anomalies, and parcels with spatial location and administrative region attributes that are inconsistent are marked as location-attribute inconsistency anomalies. Parcels corresponding to low confidence fields in the difference candidate set are included in the summary as attribute doubtful anomalies. All discrepancies are categorized into three levels based on their severity: high, medium, and low. Violations of multiple rules are weighted and accumulated according to the rule independence weights, then normalized and mapped. The weighted average of these sums with the cross-source consistency discrepancy components yields a comprehensive discrepancy confidence score, resulting in a property ownership consistency test report that includes a set of discrepancy patches and a discrepancy confidence matrix.

4. The intelligent verification method for real estate ownership certificates based on multi-source data fusion according to claim 1, characterized in that, S3 includes: Define the entity type system of the property rights knowledge graph, which includes rights subject entities, real estate unit entities, cadastral map entities, rights certificate entities, and business rule entities, as well as six types of directed relation edges: holding, covering, transferring, dividing, merging, and binding. Historical operation records are extracted from the ownership registration database according to the order of ownership certificate transfer. They are parsed into structured tuples containing operation type, entity identifiers involved, operation date, and registration authority. Entities are instantiated as nodes and relationships as directed edges to construct a cadastral knowledge graph. Land parcels with high severity discrepancies in the cadastral consistency inspection report are injected into the corresponding cadastral map nodes with special marker attributes. A temporally-aware graph attention network is used to learn the representation of knowledge graph nodes. A time encoding vector is introduced for each relation edge. The timestamp is mapped to a time feature vector through a time kernel function to participate in the attention weight calculation. When the timestamp is missing, the average registration time of the historical archives of the local cadastral area is used as the default timestamp and a missing label vector is added. A high-dimensional representation of each node is obtained by stacking three layers of aggregation operation. Reasoning is performed based on node representation in three dimensions: ownership link integrity, ownership uniqueness, and business rule compliance. The confidence scores of the three dimensions are weighted and combined according to preset weights to obtain a list of ownership legality confidence scores and ownership anomalies for each land parcel.

5. The intelligent verification method for real estate ownership certificates based on multi-source data fusion according to claim 1, characterized in that, S4 includes: Using the current image as a reference, the histogram matching method is used to perform radiometric normalization on the historical images; the feature point matching method is used to extract the same feature points in the two images, and after removing mismatched points by the random sampling consistency algorithm, the residual geometric deviation is eliminated by the affine transformation model. Pixel-level difference features are extracted from dual-temporal remote sensing images using a twin convolutional neural network. Two feature extraction branches with shared weights extract multi-scale semantic feature maps for each phase of the images. The change relationship judgment network generates difference feature maps by calculating element-wise differences and decodes them into pixel-level change probability maps. Connected regions whose change probability exceeds the preset threshold for change detection and whose area exceeds the minimum patch area threshold are extracted as candidate change patches. A semantic segmentation network incorporating a hollow spatial pyramid pooling module is used to classify land cover types in current remote sensing images. The classification results are then spatially overlaid with cadastral status maps. Land cover types that match the remote sensing classification and cadastral registration use types are considered internal morphological changes, while those that do not match are considered use type changes and are marked as high-interest difference areas. The comprehensive change confidence score is calculated from three dimensions: change probability, severity of use difference, and area ratio of difference patches. A change confidence matrix is ​​constructed with the parcel number as the row index and the difference type as the column index. Parcels with change confidence scores exceeding the difference significance threshold are included in the confirmed difference patch set.

6. The intelligent verification method for real estate ownership certificates based on multi-source data fusion according to claim 1, characterized in that, S5 includes: Based on the identification framework consisting of three mutually exclusive sets of hypotheses—compliance, doubt, and non-compliance—basic trust functions are constructed for cadastral spatial-attribute consistency evidence sources, ownership inference evidence sources, and remote sensing change evidence sources, according to the comprehensive difference confidence value, ownership legality confidence score, and remote sensing change confidence value of each parcel. When the difference confidence value is higher than the high severity threshold, it is mapped to the support of the non-compliance hypothesis; when it is in the medium range, it is mapped to the support of the doubt hypothesis; and when it is lower than the low severity threshold, it is mapped to the support of the compliance hypothesis. The basic trust functions are discounted using the credibility benchmark of each data source as a discount factor. Calculate the Jousselme distance between the basic trust function pairs of each evidence source and construct the evidence conflict degree matrix; for evidence pairs whose conflict degree exceeds the conflict warning threshold, calculate the discount factor based on the mean Jousselme distance and the credibility benchmark score, and apply discount correction to their basic trust function. The basic trust functions of the three types of evidence sources are gradually integrated according to the Dempster combination rule in the order of cadastral consistency evidence, ownership inference evidence, and remote sensing change evidence. The confidence value and similarity value of each hypothesis are extracted from the comprehensive basic trust function, and the average of the two is used as the comprehensive confidence score. The hypothesis with the highest comprehensive confidence score is used as the verification conclusion label. Land parcels with a comprehensive confidence score difference of less than the conclusion distinction threshold among the three types of conclusions are uniformly marked as questionable.

7. The intelligent verification method for real estate ownership certificates based on multi-source data fusion according to claim 1, characterized in that, S6 includes: The confidence values ​​of each difference type output by S2, the scores of each dimension of ownership reasoning output by S3, and the components of change confidence output by S4 are encoded into feature vectors and concatenated into an evidence sequence; the concatenated vector of the one-hot encoding of the verification conclusion label and the comprehensive confidence score is used as the attention query vector; the attention score between the query vector and each evidence feature vector is calculated through a multi-head attention mechanism to identify key evidence nodes. The key evidence nodes are structured into three levels: verification conclusion layer, evidence source layer, and evidence detail layer. The verification conclusion layer records the verification conclusion label and the comprehensive verification confidence score. The evidence source layer records the intermediate verification conclusion and actual contribution weight of each evidence source. The evidence detail layer records the specific difference information corresponding to each key evidence node. A structured verification report is automatically generated based on a three-layer structure. Key evidence discrepancies are highlighted, and verification suggestions are provided at the end of the report. Using the cadastral map as the base map, the discrepancy patches detected in S2 to S4 are marked with different colored overlay layers to indicate spatial-attribute consistency differences, ownership anomalies, and remote sensing-cadastral differences. An interpretable verification report and anomaly spatial annotation map are jointly output.

8. The intelligent verification method for real estate ownership certificates based on multi-source data fusion according to claim 4, characterized in that, S3 further includes: When the neighboring nodes of some land parcel nodes in the land ownership knowledge graph are sparse, before performing time-aware graph attention network aggregation, the node encoder is pre-trained by using the node attribute similarity and structural proximity as the basis for constructing positive sample pairs and the contrastive loss function. The output of the pre-trained node encoder is used as the initial node feature vector for time-aware graph attention network aggregation, replacing the original numerical representation of attribute fields.

9. The intelligent verification method for real estate ownership certificates based on multi-source data fusion according to claim 6, characterized in that, The S5 also includes: When the number of evidence sources exceeds a preset threshold, the Dezert-Smarandache theory is used to replace the DS evidence theory framework, and the super-power set is used to replace the standard power set framework, allowing for the intersection of different evidence supporting the hypothesis. In the construction of the basic trust function, conflict quantification, and stepwise fusion calculation, the standard DS framework is replaced with the super-power set framework and the corresponding proportional conflict redistribution combination rules, and the comprehensive verification confidence score and verification conclusion label of each parcel are output.

10. A multi-source data fusion-based intelligent verification system for real estate ownership, used to execute the steps in the multi-source data fusion-based intelligent verification method for real estate ownership as described in any one of claims 1-9, characterized in that, include: The data acquisition and preprocessing module is used to collect multi-source cadastral data. After control point constraint coordinate registration, ontology semantic alignment and data credibility assessment, the multi-source cadastral dataset and data source credibility benchmark vector are obtained. The consistency verification module is used to apply spatial topological constraint verification and attribute logical consistency rule base joint verification to the multi-source cadastral dataset to obtain a cadastral consistency verification report. The knowledge graph reasoning module is used to obtain a multi-source property rights dataset and a property rights consistency verification report as input, construct a property rights knowledge graph through graph attention network temporal reasoning, and obtain the property rights legality confidence score and a list of property rights anomaly relationships; The remote sensing change detection module is used to acquire multi-temporal remote sensing images and cadastral status maps. Through the extraction of differential features by twin convolutional neural networks and multi-scale semantic segmentation and overlay analysis, the remote sensing to cadastral difference map patch set and change confidence matrix are obtained. The evidence fusion module is used to take the ownership consistency verification report, ownership legality confidence score and change confidence matrix as evidence input, and the credibility benchmark score vector as prior weight. After DS evidence theory conflict correction and fusion, a comprehensive verification confidence score and verification conclusion label are obtained. The report generation module takes the comprehensive verification confidence score and verification conclusion label as input, and through multi-head attention key evidence labeling and hierarchical organization of the evidence chain, generates an interpretable verification report and an anomaly space annotation map.

Images