A Deep Learning-Based Method for Anomaly Detection in 3D Cadastral Data

Abstract

This invention discloses a deep learning-based method for anomaly detection in 3D cadastral data, comprising the following steps: Step 1, 3D cadastral data preprocessing; Step 2, multimodal feature extraction and fusion; Step 3, deep learning detection model construction and training; Step 4, model inference and initial anomaly detection; Step 5, anomaly result post-processing and optimization. This invention achieves high detection accuracy by employing multimodal feature fusion technology to fully exploit the geometric, topological, and semantic information of 3D cadastral data. Combined with an improved implicit shape network based on pose awareness, it can accurately identify eight common anomalies, including overlapping features, hanging points, and face intersections, achieving a high detection accuracy rate, significantly improved compared to traditional methods. In particular, its ability to identify subtle anomalies (such as small facet warping and attribute encoding errors) is significantly enhanced, with a false negative rate controlled below 5%.

Description

Technical Field

[0001] This invention belongs to the fields of geographic information engineering and deep learning technology, specifically a method for detecting anomalies in three-dimensional cadastral data based on deep learning. Background Technology

[0002] As an important development direction of modern cadastral management, 3D cadastral data can accurately represent the three-dimensional spatial location, ownership boundaries, and attribute information of real estate such as land parcels and buildings, providing core data support for real estate registration, land resource management, and urban planning. 3D cadastral data is mainly acquired through technologies such as oblique photogrammetry and lidar scanning. Data types include point cloud data, image data, topological relationship data, and ownership attribute data, characterized by large data volume, complex structure, and multi-source heterogeneity.

[0003] The quality of 3D cadastral data directly determines its application value; data anomalies can lead to serious problems such as errors in real estate registration and ownership disputes. Common 3D cadastral data anomalies include overlapping features, hanging points, discontinuous lines, unclosed surfaces, surface intersections, self-intersections, contradictions between attributes and spatial information, and inconsistent topological relationships. Existing 3D cadastral data anomaly detection methods are mainly divided into two categories: one is detection methods based on traditional geometric algorithms, which identify anomalies by constructing topological rules and using distance thresholds, such as using distance thresholds to detect hanging points and using topological adjacency relationships to detect surface intersections; the other is detection methods based on simple machine learning, which achieve anomaly identification by extracting single features to train a classifier.

[0004] However, existing technologies have significant drawbacks: traditional geometric algorithms rely on manually set rules, resulting in weak identification capabilities for complex spatial relationships (such as topological conflicts in irregular parcels) and poor adaptability; rule parameters need to be readjusted for different regions and types of cadastral data. Simple machine learning methods often use single features, failing to fully represent the multi-dimensional information of 3D cadastral data, leading to low anomaly detection accuracy, especially for subtle anomalies (such as minor surface warping or attribute encoding errors). Furthermore, existing methods generally suffer from low detection efficiency, making it difficult to meet the quality control requirements of large-scale 3D cadastral data. In addition, existing technologies lack effective handling of data pose variations; when 3D point cloud data exhibits pose differences due to different acquisition angles, it severely impacts the stability and accuracy of anomaly detection.

[0005] Therefore, there is an urgent need for an anomaly detection method that can adapt to different types of 3D cadastral data, make full use of multi-dimensional information, and balance detection accuracy and efficiency, in order to overcome the shortcomings of existing technologies.

[0006] Based on this, a deep learning-based method for anomaly detection in 3D cadastral data is designed. Summary of the Invention

[0007] In order to overcome the shortcomings of the prior art, this invention provides a deep learning-based method for detecting anomalies in three-dimensional cadastral data, which effectively solves the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for anomaly detection in three-dimensional cadastral data based on deep learning, comprising the following steps: Step 1: 3D Cadastral Data Preprocessing: Perform continuous processing on the collected multi-source 3D cadastral raw data, including format standardization, coordinate unification, and noise filtering, to output standardized 3D cadastral data; Step 2: Multimodal feature extraction and fusion: Point cloud geometric features, topological relationship features, and attribute semantic features are extracted from the standardized 3D cadastral data. The three types of features are dynamically fused through the feature fusion module to generate fixed-dimensional multimodal fused features. Step 3: Construction and training of deep learning detection model: Construct an improved implicit shape network with pose alignment function, use the multimodal fusion features as input to train the model, and obtain a converged anomaly detection model; Step 4, Model Inference and Initial Anomaly Detection: The three-dimensional cadastral data to be detected is processed sequentially through Step 1 and Step 2 and then input into the converged anomaly detection model. The model outputs the initial anomaly detection results, which include the three-dimensional coordinates of the anomaly location, the anomaly type, and the anomaly confidence level. Step 5: Post-processing and optimization of abnormal results: Based on the three-dimensional cadastral data specifications, confidence screening and spatial consistency verification are performed on the initial anomaly detection results, and the final anomaly detection results are output.

[0009] Preferably, the multi-source three-dimensional cadastral raw data in step one includes oblique photogrammetry point cloud data, lidar point cloud data, ownership attribute data, and topological relationship data; The format standardization specifically involves converting raw point cloud data from different sources into LAS 1.4 format, converting ownership attribute data into JSON format, and converting topology relationship data into SHP format. The coordinate unification specifically involves using a seven-parameter coordinate transformation model to convert all data to the 2000 National Geodetic Coordinate System, with the transformation accuracy controlled within ±2cm. The noise filtering employs a statistical outlier removal algorithm, which performs the following operations: performing a neighborhood search on the standardized point cloud data, selecting 15-25 nearest points within the neighborhood for each point, calculating the average distance between each point and the selected nearest points, and identifying and deleting points whose average distance is greater than 1.8-2.2 times the average distance of the neighborhood.

[0010] Preferably, the point cloud geometric features mentioned in step two include point cloud density, normal vector, curvature, and FPFH features; The topological relationship features include point-line association, line-plane topological relationship, plane-plane containment relationship, and connectivity features. The topological relationship features are quantitatively represented by constructing a topological adjacency matrix. The attribute semantic features are obtained by semantic segmentation of the image data corresponding to the three-dimensional cadastral data using a YOLOv8 model trained on a custom dataset, and include feature vectors corresponding to land cover type, ownership code and elevation attribute. The feature fusion module adopts a self-attention mechanism fusion architecture, which realizes dynamic weighted fusion by calculating the attention weights of the three types of features, and generates 256-dimensional multimodal fusion features.

[0011] Preferably, the improved implicit shape network with attitude alignment function described in step three is composed of an attitude alignment module, an implicit shape encoding module, and an anomaly detection module connected in series. The pose alignment module adopts a combined architecture of voxel sampling, FPFH feature extraction, and ICP iterative optimization to achieve pose normalization between the data to be detected and the standard cadastral shape template. The termination condition of ICP iterative optimization is that the pose change between two adjacent iterations is less than 0.001 radians. The implicit shape encoding module introduces a sinusoidal position encoding function to map the pose-normalized three-dimensional coordinates into a high-dimensional position feature vector, which is then concatenated with the multimodal fusion feature and input into a stacked fully connected layer to construct a continuous signed distance field SDF representation. The anomaly detection module calculates the difference between the SDF representation of the data to be detected and the SDF representation of the standard cadastral shape template, and outputs the anomaly confidence.

[0012] Preferably, the model training in step three uses a hybrid loss function, which is a weighted sum of truncated L1 loss and topology consistency loss, expressed as: Loss = α × Loss_L1 + β × Loss_Top, where α is 0.6-0.7 and β is 0.3-0.4. The training process uses the Adam optimizer, with an initial learning rate set to 1e-4. Every 50 epochs, the learning rate decays to 0.5 times the current value, and the total number of training iterations is 200-300 epochs. The training terminates when the anomaly detection accuracy on the validation set shows no improvement for 20 consecutive epochs.

[0013] Preferably, the anomaly type mentioned in step four includes at least one of the following: overlapping feature anomaly, hanging point anomaly, line object discontinuity anomaly, surface non-closure anomaly, surface intersection anomaly, self-intersection anomaly, attribute and spatial information contradiction anomaly, and topological relationship inconsistency anomaly. The anomaly confidence level is calculated by weighted product of SDF difference and feature deviation, where SDF difference is the absolute difference between the data to be detected and the SDF value of the standard template, and feature deviation is the Euclidean distance between the multimodal fusion feature and the normal data feature distribution. The weight coefficients of both are set to 0.5.

[0014] Preferably, the confidence threshold for screening in step five is set to 0.75-0.85, and entries with confidence levels lower than this threshold in the initial anomaly detection results are directly removed. The spatial consistency verification specifically performs the following operations: A spatial association network of adjacent cadastral units is constructed, and abnormal results of adjacent units are extracted. The location, type and confidence level of abnormal results of adjacent units are compared. If there is a logical conflict, the data of the conflict area is re-input into the converged anomaly detection model for secondary reasoning. The abnormal status of the conflict area is determined by combining the manual review rules.

[0015] Preferably, in the statistical outlier removal algorithm, the number of nearest neighbors selected is 20, and the noise threshold is set to 2.0 times the average neighborhood distance. The seven-parameter coordinate transformation model solves for the transformation parameters using at least three evenly distributed control points, and the spatial position deviation of the transformed data does not exceed ±1.5cm.

[0016] Preferably, the custom dataset contains 100,000+ 3D cadastral image samples, covering 8 typical land features including buildings, roads, land parcels, water bodies, forest land, grassland, cultivated land, and structures. The samples are labeled with boundaries and types using the LabelMe tool. The YOLOv8 model, after being trained on the custom dataset, achieves an F1 score of no less than 0.93 for semantic segmentation.

[0017] Preferably, the manual review rules are as follows: for abnormal results that still have logical conflicts after secondary reasoning, a review list containing the three-dimensional coordinates of the abnormal location, image screenshots, conflict types, and the basis for model reasoning is generated, and professional staff judge the contents of the list. The error rate of manual review is controlled within 3%.

[0018] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention boasts high detection accuracy. Employing multimodal feature fusion technology, it fully leverages the geometric, topological, and semantic information of 3D cadastral data. Combined with an improved implicit shape network based on pose perception, it can accurately identify eight common anomalies, including overlapping features, hanging points, and face intersections. The detection accuracy is significantly improved compared to traditional methods. In particular, its ability to identify subtle anomalies (such as minor face warping and attribute coding errors) is significantly enhanced, with a false negative rate controlled below 5%. 2. This invention has high detection efficiency, realizing full automation of the detection process of anomalies in 3D cadastral data without the need for manual rule setting and parameter adjustment, and improving detection efficiency by more than 60% compared with traditional methods; through optimization strategies such as voxel sampling and feature fusion, the computational complexity of the model is reduced, which can meet the real-time detection needs of large-scale 3D cadastral data (such as millions of point cloud data); 3. The present invention has strong adaptability. The pose alignment module of the present invention can effectively solve the pose difference problem of 3D cadastral data and improve the stability of the model under different acquisition scenarios. The adoption of multimodal feature fusion and attention mechanism enables the model to adapt to different regions (urban, rural, industrial land) and different types of 3D cadastral data without retraining the model, thus having strong adaptability. 4. This invention is highly practical. The post-processing workflow for abnormal results combines the three-dimensional cadastral data specifications with manual review rules to ensure the reliability of the detection results. The output abnormal detection report is detailed and intuitive, providing clear processing suggestions for the quality control of three-dimensional cadastral data. It can be directly applied to practical work such as real estate registration and land resource management, and has important practical value. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0020] In the attached diagram: Figure 1 This is a schematic diagram of a deep learning-based three-dimensional cadastral data anomaly detection method according to the present invention. Detailed Implementation

[0021] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0022] Depend on Figure 1 This invention relates to a method for anomaly detection in three-dimensional cadastral data based on deep learning, comprising the following steps: Step 1: 3D Cadastral Data Preprocessing The core objective of this step is to convert the multi-source, heterogeneous raw 3D cadastral data into a standardized, highly reliable data format, laying the foundation for subsequent feature extraction and model training. The specific implementation process is as follows: 1. Raw Data Collection: The collected 3D cadastral raw data includes oblique photogrammetry point cloud data, lidar point cloud data, ownership attribute data (such as rights holder information, land category codes, and area data), and topological relationship data (such as point-line-area relationships and parcel boundary adjacency relationships). Among them, oblique photogrammetry point cloud data is collected by UAV equipped with a multi-lens camera, lidar point cloud data is collected by a ground lidar scanner, ownership attribute data comes from the real estate registration system, and topological relationship data is automatically generated by surveying software.

[0023] 2. Format standardization: Due to the differences in data formats output by different acquisition devices (e.g., LiDAR data is in LAZ format, and oblique photogrammetry data is in PLY format), all raw point cloud data are uniformly converted to LAS 1.4 format. This format supports the storage of multi-dimensional information such as point cloud intensity, color, and classification, and conforms to the geographic information data exchange standard; ownership attribute data is converted to JSON format, and topological relationship data is converted to SHP format.

[0024] 3. Coordinate unification: All types of 3D cadastral data are uniformly converted to the 2000 National Geodetic Coordinate System. A seven-parameter coordinate transformation model is used to realize the conversion between different coordinate systems. During the conversion process, at least three evenly distributed control points are selected for parameter solving. The conversion accuracy is controlled within ±2cm to ensure the spatial consistency of data from different sources.

[0025] 4. Noise filtering: A statistical outlier removal algorithm is used to remove noise points (such as isolated points caused by measurement errors and noise points caused by environmental interference) from the point cloud data. The specific steps are as follows: ① Perform a neighborhood search on the standardized point cloud data, and select the 20 nearest points in the neighborhood of each point (k=20). ② Calculate the average distance between each point and its 20 nearest neighbors; ③ Set the noise threshold to 2.0 times the average distance of the neighborhood, and identify and delete points whose average distance is greater than the threshold as noise points; ④ The filtered point cloud data is smoothed, and the moving least squares method is used to optimize the surface accuracy of the point cloud.

[0026] Through the above processing, standardized, noise-free three-dimensional cadastral data are obtained, and the data integrity is improved to over 99%.

[0027] Step 2: Multimodal Feature Extraction and Fusion This step extracts and fuses multi-dimensional features from 3D cadastral data to fully explore the geometric, topological, and semantic information of the data, thereby improving the model's ability to identify different types of anomalies. The specific implementation process is as follows: 1. Point cloud geometric feature extraction: Extracting features that characterize spatial geometry from standardized point cloud data, including: ① Point cloud density is obtained by calculating the number of points per unit volume; ② Normal vector: The normal vector of each point is solved using principal component analysis (PCA) to represent the orientation information of the point cloud surface; ③Curvature, calculated based on the rate of change of the normal vector, reflects the degree of curvature of the point cloud surface; ④FPFH features: By calculating the statistical information of the angle between the normals of a point and its neighboring points, a 33-dimensional global geometric feature vector is generated to represent the local geometric structure.

[0028] By concatenating the above geometric features, a 128-dimensional point cloud geometric feature vector is obtained.

[0029] 2. Topological Relationship Feature Extraction: Construct a topological relationship network from the 3D cadastral data and extract topological relationship features. Specific steps are as follows: ① Construct a topological adjacency matrix based on point, line, and surface entities. The matrix elements represent the relationships between different entities (such as points on lines, lines in surfaces, and surfaces adjacent to each other). ② Calculate the topological connectivity characteristics by analyzing the number of connected components and the length of connected paths in the topological network; ③ Extract topological consistency features to determine whether the relationship between points, lines, and surfaces conforms to cadastral data specifications (e.g., surfaces must be composed of closed lines, and lines must be composed of continuous points).

[0030] The topological adjacency matrix is ​​vectorized, and combined with connectivity and consistency features, to obtain a 64-dimensional topological relationship feature vector.

[0031] 3. Attribute Semantic Feature Extraction: An improved YOLOv8 model is used to perform semantic segmentation on the image data corresponding to the 3D cadastral data to obtain attribute semantic features. The specific steps are as follows: ① Construct a custom training dataset containing 100,000+ 3D cadastral image samples, covering 8 typical land features including buildings, roads, land parcels, water bodies, forest land, grassland, cultivated land, and structures. Use the LabelMe tool to complete the sample annotation, and the annotation information includes land feature boundaries and land feature types. ② The YOLOv8 model was trained, the backbone network structure of the model was optimized, and an attention module was added to improve the ability to recognize small targets. During the training process, data augmentation strategies such as random cropping, flipping, and color gamut transformation were adopted. The F1 score of the model training reached more than 0.93. ③ Input the three-dimensional cadastral image data into the trained YOLOv8 model to obtain the semantic segmentation results of land cover types. Combine the ownership attribute data (such as land category code and elevation attribute) and generate a 64-dimensional attribute semantic feature vector through fully connected layer mapping.

[0032] 4. Multimodal Feature Fusion: An attention-based fusion architecture is used to dynamically fuse point cloud geometric features, topological relationship features, and attribute semantic features. The specific steps are as follows: ① Input the three feature vectors into the self-attention module respectively, calculate the attention weight of each feature, and the weight value ranges from 0 to 1, which represents the importance of different features to anomaly detection; ②The three features are weighted and summed according to the attention weights to obtain the initial fused features; ③ The initial fusion features are compressed and optimized by using a 1×1 convolutional layer to generate a unified multimodal fusion feature of 256 dimensions.

[0033] This fusion method can adaptively adjust the weight distribution of each feature under different scenarios, thereby improving the feature discrimination ability.

[0034] Step 3: Construction and Training of Deep Learning Detection Model This step constructs an improved implicit shape network based on pose awareness. Through model training, it learns the feature patterns of normal 3D cadastral data, enabling accurate identification of abnormal data. The specific implementation process is as follows: 1. Model Architecture Design: The pose-aware improved implicit shape network consists of three core parts: a pose alignment module, an implicit shape encoding module, and an anomaly detection module. The functions of each module are as follows: ① Pose Alignment Module: This module addresses the pose differences in 3D cadastral data caused by varying acquisition angles by employing a combination of voxel sampling, FPFH feature extraction, and ICP iterative optimization.

[0035] The specific steps are as follows: voxel sampling is performed on the point cloud data corresponding to the input multimodal fusion features to reduce the amount of data; FPFH features of the sampled point cloud are extracted and matched with FPFH features of the preset standard cadastral shape template to obtain initial corresponding point pairs; RANSAC algorithm is used to screen reliable corresponding point pairs and remove erroneous matches; the pose of the data to be detected is adjusted by ICP iterative optimization algorithm to make the pose of the data to be detected consistent with that of the standard template, and the iteration termination condition is that the pose change between two iterations is less than 0.001 radians.

[0036] ② Implicit Shape Encoding Module: Introduces a sinusoidal position encoding function to map three-dimensional coordinates into high-dimensional feature vectors, improving the model's ability to perceive subtle geometric differences.

[0037] The specific steps are as follows: input the pose-aligned 3D coordinates into the sinusoidal position encoding function γ(x) to generate a high-dimensional position feature vector; concatenate the high-dimensional position feature vector with the multimodal fusion feature, input it into a stacked fully connected layer (6 layers in total, with 512, 512, 256, 256, 128, and 64 neurons in each layer, respectively), and perform a nonlinear transformation through the ReLU activation function to construct a continuous signed distance field (SDF) representation. The SDF value represents the signed distance from a point in 3D space to the cadastral feature surface (positive values ​​indicate outside the surface, negative values ​​indicate inside the surface, and zero values ​​indicate on the surface).

[0038] ③ Anomaly Detection Module: Determines whether the data is abnormal by calculating the difference between the SDF representation of the data to be detected and the SDF representation of the standard template.

[0039] The specific steps are as follows: calculate the absolute difference between the SDF value of the data to be detected and the SDF value of the standard template to obtain the SDF difference value; combine the deviation of the multimodal fusion feature (Euclidean distance from the normal data feature distribution) and calculate the anomaly confidence level by weighted product (the weight coefficients are all set to 0.5). The anomaly confidence level ranges from 0 to 1. The larger the value, the higher the probability of data anomaly.

[0040] 2. Training Dataset Construction: A large-scale training dataset containing both normal and abnormal samples was constructed. Normal samples were selected from 3D cadastral data that had been manually verified to be free of anomalies, covering different regions (urban, rural, and industrial land) and different land cover types, totaling over 50,000 normal samples. Abnormal samples were generated by adding manually designed anomalies (such as overlapping features, hanging points, polygon intersections, etc., eight common anomalies) to the normal samples, with at least 3,000 samples of each anomaly type, totaling over 30,000 abnormal samples. The training dataset was divided into training, validation, and test sets in a ratio of 7:2:1.

[0041] 3. Model Training: A hybrid loss function is used for model training. The hybrid loss function is a weighted sum of truncated L1 loss and topology consistency loss, as follows: Loss = α × Loss_L1 + β × Loss_Top, where α is the weight of truncated L1 loss (0.65) and β is the weight of topology consistency loss (0.35). The truncated L1 loss is used to optimize the model's SDF representation accuracy. The truncation threshold for SDF values ​​is set to 0.1m. When the absolute difference of SDF values ​​is greater than 0.1m, the loss value is fixed at 0.1m. The topology consistency loss is used to... The topological relationships output by the constraint model are consistent with those of normal cadastral data, obtained by calculating the difference between the topological adjacency matrix output by the model and the actual topological adjacency matrix. During training, the Adam optimizer is used, with an initial learning rate set to 1e-4. Every 50 epochs, the learning rate decays to 0.5 times the original rate, and the total number of training iterations is 250 epochs. The training batch size is set to 32. During training, the anomaly detection accuracy of the validation set is monitored in real time. When the validation set accuracy does not improve for 20 consecutive epochs, training is stopped and the optimal model parameters are saved.

[0042] Step 4: Model Inference and Initial Anomaly Detection This step involves preprocessing and feature fusion of the 3D cadastral data to be detected, then inputting it into the trained model. The model then uses inference to obtain initial anomaly detection results. The specific implementation process is as follows: 1. Preprocessing of data to be tested: The original three-dimensional cadastral data to be tested is processed by standardizing the format, unifying the coordinates and filtering noise according to the method in step one to obtain standardized data to be tested.

[0043] 2. Feature extraction and fusion of the data to be detected: The standardized data to be detected is used to extract point cloud geometric features, topological relationship features and attribute semantic features according to the method in step 2. The 256-dimensional multimodal fusion features to be detected are obtained through multimodal feature fusion.

[0044] 3. Model Inference: The multimodal fusion features to be detected are input into the trained pose-aware improved implicit shape network. The pose alignment module normalizes the pose with the standard template. The implicit shape encoding module generates the SDF representation of the data to be detected. The anomaly detection module calculates the anomaly confidence.

[0045] 4. Initial Anomaly Result Output: Set the anomaly judgment threshold to 0.5. When the anomaly confidence score is greater than 0.5, it is judged as anomalous data. Combine the SDF difference distribution and feature deviation position output by the model to determine the three-dimensional coordinates of the anomaly location. According to the type of anomaly feature (such as geometric feature anomaly, topological feature anomaly, semantic feature anomaly), match the corresponding anomaly type (such as overlapping ground feature anomaly, hanging point anomaly, etc.). Finally, output the initial anomaly detection results, including the three-dimensional coordinates of the anomaly location, anomaly type, anomaly confidence score, and image screenshot of the anomaly area.

[0046] Step 5: Post-processing and optimization of abnormal results This step optimizes the initial anomaly detection results by incorporating 3D cadastral data standards, eliminating false positives, and improving detection accuracy. The specific implementation process is as follows: 1. Confidence Filtering: Based on the accuracy requirements of 3D cadastral data, the confidence filtering threshold is set to 0.8 (which can be adjusted according to the actual application scenario, with an adjustment range of 0.75-0.85). Abnormal results with a confidence level lower than 0.8 in the initial anomaly detection results are deleted, while high-confidence abnormal results are retained to reduce the false detection rate.

[0047] 2. Spatial consistency verification: Check whether there are logical conflicts in the abnormal results of adjacent cadastral units (such as contradictory boundary anomaly judgments of adjacent parcels, or conflicting anomaly types in different parts of the same building).

[0048] The specific steps are as follows: Construct a spatial association network of adjacent cadastral units and extract abnormal results from adjacent units; compare the location, type, and confidence level of abnormal results from adjacent units. If there is a conflict (e.g., unit A determines that a boundary is an overlap anomaly, while unit B determines that the boundary is normal), then re-input the relevant data of the conflict area into the trained anomaly detection model for secondary inference; based on the spatial relationship between the secondary inference results and adjacent units, and combined with the three-dimensional cadastral data specifications (e.g., DB12T+1419-2025 Technical Specification for Surveying and Mapping Three-Dimensional Cadastral Property Rights), make logical judgments to determine the final abnormal state of the conflict area.

[0049] 3. Manual Review and Result Output: For abnormal results that still conflict in the secondary inference, a manual review list is generated. The list includes the 3D coordinates of the abnormal location, image screenshots, conflict type, and model inference basis. Professional staff manually judge the review list, and the error rate of manual review is controlled within 3%. The results of confidence screening, spatial consistency verification, and manual review are integrated to generate a final anomaly detection report. The report includes anomaly summary statistics (anomaly type, quantity, distribution area), detailed anomaly information (location, confidence level, handling suggestions), and visualization charts (anomaly distribution heat map, 3D location annotation map).

[0050] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0051] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for anomaly detection in three-dimensional cadastral data based on deep learning, comprising the following steps: Step 1: 3D Cadastral Data Preprocessing: Perform continuous processing on the collected multi-source 3D cadastral raw data, including format standardization, coordinate unification, and noise filtering, to output standardized 3D cadastral data; Step 2: Multimodal feature extraction and fusion: Point cloud geometric features, topological relationship features, and attribute semantic features are extracted from the standardized 3D cadastral data. The three types of features are dynamically fused through the feature fusion module to generate fixed-dimensional multimodal fused features. Step 3: Construction and training of deep learning detection model: Construct an improved implicit shape network with pose alignment function, use the multimodal fusion features as input to train the model, and obtain a converged anomaly detection model; Step 4, Model Inference and Initial Anomaly Detection: The three-dimensional cadastral data to be detected is processed sequentially through Step 1 and Step 2 and then input into the converged anomaly detection model. The model outputs the initial anomaly detection results, which include the three-dimensional coordinates of the anomaly location, the anomaly type, and the anomaly confidence level. Step 5: Post-processing and optimization of abnormal results: Based on the three-dimensional cadastral data specifications, confidence screening and spatial consistency verification are performed on the initial anomaly detection results, and the final anomaly detection results are output.

2. The method for anomaly detection of three-dimensional cadastral data based on deep learning according to claim 1, characterized in that: The multi-source 3D cadastral raw data mentioned in step one includes oblique photogrammetry point cloud data, lidar point cloud data, ownership attribute data, and topological relationship data; The format standardization specifically involves uniformly converting raw point cloud data from different sources into LAS. 1.4 Format: Ownership attribute data is converted to JSON format, and topology relationship data is converted to SHP format; The coordinate unification specifically involves using a seven-parameter coordinate transformation model to convert all data to the 2000 National Geodetic Coordinate System, with the transformation accuracy controlled within ±2cm. The noise filtering employs a statistical outlier removal algorithm, which performs the following operations: performing a neighborhood search on the standardized point cloud data, selecting 15-25 nearest points within the neighborhood for each point, calculating the average distance between each point and the selected nearest points, and identifying and deleting points whose average distance is greater than 1.8-2.2 times the average distance of the neighborhood.

3. The method for anomaly detection of three-dimensional cadastral data based on deep learning according to claim 1, characterized in that: The point cloud geometric features mentioned in step two include point cloud density, normal vector, curvature, and FPFH features; The topological relationship features include point-line association, line-plane topological relationship, plane-plane containment relationship, and connectivity features. The topological relationship features are quantitatively represented by constructing a topological adjacency matrix. The attribute semantic features are obtained by semantic segmentation of the image data corresponding to the three-dimensional cadastral data using a YOLOv8 model trained on a custom dataset, and include feature vectors corresponding to land cover type, ownership code and elevation attribute. The feature fusion module adopts a self-attention mechanism fusion architecture, which realizes dynamic weighted fusion by calculating the attention weights of the three types of features, and generates 256-dimensional multimodal fusion features.

4. The method for anomaly detection of three-dimensional cadastral data based on deep learning according to claim 1, characterized in that: The improved implicit shape network with pose alignment function described in step three consists of a pose alignment module, an implicit shape encoding module, and an anomaly detection module connected in series. The pose alignment module adopts a combined architecture of voxel sampling, FPFH feature extraction, and ICP iterative optimization to achieve pose normalization between the data to be detected and the standard cadastral shape template. The termination condition of ICP iterative optimization is that the pose change between two adjacent iterations is less than 0.001 radians. The implicit shape encoding module introduces a sinusoidal position encoding function to map the pose-normalized three-dimensional coordinates into a high-dimensional position feature vector, which is then concatenated with the multimodal fusion feature and input into a stacked fully connected layer to construct a continuous signed distance field SDF representation. The anomaly detection module calculates the difference between the SDF representation of the data to be detected and the SDF representation of the standard cadastral shape template, and outputs the anomaly confidence.

5. The method for anomaly detection of three-dimensional cadastral data based on deep learning according to claim 4, characterized in that: The model training described in step three uses a hybrid loss function, which is a weighted sum of truncated L1 loss and topological consistency loss, expressed as: Loss = α × Loss_L1 + β × Loss_Top, where α is 0.6-0.7 and β is 0.3-0.

4. The Adam optimizer was used during training. The initial learning rate was set to 1e-4. Every 50 epochs, the learning rate was reduced to 0.5 times the current value. The total number of training iterations was 200-300 epochs. The training was terminated when the anomaly detection accuracy on the validation set did not improve for 20 consecutive epochs.

6. The method for anomaly detection of three-dimensional cadastral data based on deep learning according to claim 1, characterized in that: The anomaly types mentioned in step four include at least one of the following: overlapping feature anomalies, hanging point anomalies, line object discontinuity anomalies, surface non-closure anomalies, surface intersection anomalies, self-intersection anomalies, attribute and spatial information contradiction anomalies, and topological relationship inconsistency anomalies. The anomaly confidence level is calculated by weighted product of SDF difference and feature deviation, where SDF difference is the absolute difference between the data to be detected and the SDF value of the standard template, and feature deviation is the Euclidean distance between the multimodal fusion feature and the normal data feature distribution. The weight coefficients of both are set to 0.

5.

7. The method for anomaly detection of three-dimensional cadastral data based on deep learning according to claim 1, characterized in that: The confidence threshold for screening in step five is set to 0.75-0.85, and entries with confidence levels below this threshold in the initial anomaly detection results are directly removed. The spatial consistency verification specifically performs the following operations: A spatial association network of adjacent cadastral units is constructed, and abnormal results of adjacent units are extracted. The location, type and confidence level of abnormal results of adjacent units are compared. If there is a logical conflict, the data of the conflict area is re-input into the converged anomaly detection model for secondary reasoning. The abnormal status of the conflict area is determined by combining the manual review rules.

8. The method for anomaly detection of three-dimensional cadastral data based on deep learning according to claim 2, characterized in that: In the statistical outlier removal algorithm, the number of nearest neighbors selected is 20, and the noise threshold is set to 2.0 times the average distance of the neighborhood. The seven-parameter coordinate transformation model solves for the transformation parameters using at least three evenly distributed control points, and the spatial position deviation of the transformed data does not exceed ±1.5cm.

9. The method for anomaly detection of three-dimensional cadastral data based on deep learning according to claim 3, characterized in that: The custom dataset contains over 100,000 3D cadastral image samples, covering eight typical land features: buildings, roads, land parcels, water bodies, forest land, grassland, cultivated land, and structures. The samples are labeled with boundaries and types using the LabelMe tool. The YOLOv8 model, after being trained on the custom dataset, achieves an F1 score of no less than 0.93 for semantic segmentation.

10. The method for anomaly detection of three-dimensional cadastral data based on deep learning according to claim 7, characterized in that: The manual review rules are as follows: For abnormal results that still have logical conflicts after secondary reasoning, a review list is generated, which includes the three-dimensional coordinates of the abnormal location, image screenshots, conflict types, and the basis for model reasoning. Professional staff will judge the contents of the list, and the error rate of manual review is controlled within 3%.

Images