A method and apparatus for geographic entity classification based on knowledge graph neighborhood learning

Abstract

This invention provides a method and apparatus for geographic entity classification based on knowledge graph neighborhood learning. The method includes: generating initial vectors for each entity using an image recognition model, comparing their similarity with standard vectors of entity classes, directly classifying entities with similarity reaching a threshold, and adjusting the vector values ​​based on neighborhood information for entities below the threshold before classification; calculating corresponding weights for different relationships based on distance and dependency; constructing a geographic knowledge graph; adjusting entity vectors by learning neighborhood information from the knowledge graph; calculating the cosine similarity between the adjusted entity vectors and the standard vectors of each category; and selecting the category with the highest similarity as the final classification result. This method can better capture the contextual relationships of entities in geographic scenes, compensate for the shortcomings of traditional methods that ignore neighborhood information, make the final classification results more accurate, and significantly improve the recognition accuracy of various geographic entity targets.

Description

Technical Field

[0001] This invention belongs to the field of geographic image analysis technology, specifically relating to a geographic entity classification method and apparatus based on knowledge graph neighborhood learning. Background Technology

[0002] Geographic entity classification is a key technical link in geographic information management, regional planning, resource operation, and other fields, and its classification accuracy directly affects the application effect in these fields. Currently, the industry's geographic entity classification technologies for geographic imagery mainly fall into three categories, all of which have significant technical shortcomings and are unable to meet the high-precision classification requirements of complex application scenarios. The first category is based on traditional image segmentation and feature extraction methods. These methods first complete image segmentation using traditional algorithms such as threshold segmentation and edge detection, then extract features such as color and texture using manually designed rules, and finally rely on a classifier to achieve entity classification. However, traditional segmentation algorithms have poor adaptability to complex scenes, and the expression and generalization capabilities of manually designed features are limited, failing to capture complex relationships between entities, resulting in low classification accuracy. The second category is based on deep learning single-object detection models. These methods utilize CNNs to construct Faster... Models like R-CNN and YOLO, trained on labeled data, achieve geographic entity detection and classification. While performing well in simple scenarios, they are susceptible to false positives and false negatives in complex scenes due to issues such as entity occlusion, overlap, and similar appearance. Furthermore, these models focus only on individual object features, neglecting neighborhood information and semantic relationships between entities, failing to fully utilize the contextual information of the geographic scene and limiting classification performance improvement. Thirdly, methods based on multi-source data fusion, which combine remote sensing imagery, GIS, LiDAR, and other data sources to compensate for the shortcomings of a single data source, suffer from differences in format, resolution, and accuracy, making data registration difficult. Moreover, the design of feature selection and fusion strategies lacks effective solutions, and improper handling can easily lead to a decline in classification performance. In summary, existing geographic entity classification techniques only learn the features of the current entity, ignoring the learning of neighborhood information. When entity features are similar, the classification effect deteriorates significantly, failing to effectively adapt to the application requirements of complex scenarios. For example, the spectral features of unplanted farmland are extremely similar to those of bare soil, and learning only its own features is insufficient to meet classification requirements. Summary of the Invention

[0003] In view of the shortcomings of the existing methods, this invention proposes a geographic entity classification method and apparatus based on knowledge graph neighborhood learning. This method can better capture the contextual relationships of entities in a geographic scene, resulting in more accurate classification results and significantly improving the recognition accuracy of various geographic entity targets such as buildings, bare soil, farmland, and grassland. The method includes:

[0004] Geographic images of the region to be classified are acquired, entity categories and category standard vectors are determined, and image recognition models are used to extract features from the geographic images to obtain initial vectors for each entity. The cosine similarity between each initial entity vector and each category standard vector is calculated. The maximum cosine similarity between each entity vector and all category standard vectors is extracted. For entity vectors with a maximum cosine similarity greater than or equal to a threshold, the entity is determined to belong to the corresponding category, and the result is used as classification label data. For entity vectors with a maximum cosine similarity less than the threshold, the entity is determined to be an unclassified entity. Classified entities are then extracted from the geographic images. For each unclassified entity, construct "entity-relationship-entity" triples. Based on the spatial distance index and functional dependency between entities, calculate the relationship weights between entities using a linear weighting method. Then, integrate all triples to construct a geographic knowledge graph. Based on the geographic knowledge graph, obtain the classified entities in the neighborhood of the unclassified entity. Adjust the initial vector of the unclassified entity using the classified entities in the neighborhood and the relationship weights between entities to obtain the final vector of the unclassified entity. Then, calculate the cosine similarity between the final vector of each unclassified entity and the standard vectors of all categories. The category with the highest cosine similarity is determined as the category of the unclassified entity.

[0005] Furthermore, the initial vector of the unclassified entity is adjusted using the classified entities in the neighborhood to obtain the final vector of the unclassified entity. The adjustment formula is as follows:

[0006] ,

[0007] in, Let be the initial vector of entity i. Let j be the feature vector of the classified entity j in the neighborhood. Let i be the final vector of entity i. Weights for relationships between entities. Let i be the neighborhood set of entity i. This is the global adjustment coefficient.

[0008] Furthermore, based on the spatial distance index and functional dependency between entities, a linear weighted method is used to calculate the relationship weights between entities. The specific calculation formula is as follows:

[0009] ,

[0010] Where W is the weight value of the relationship between entities, ranging from [0,1]; It is an indicator of spatial distance between entities. Functional dependency between entities , These are the weighting coefficients for spatial distance and functional dependence, respectively. .

[0011] Furthermore, the spatial distance index between entities is specifically calculated as follows: the spatial distance between entities is calculated using the Euclidean distance formula. ,when Greater than the distance threshold At that time, the spatial distance index is set to 0, and the calculated distance is... Less than the threshold At that time, the formula for calculating the spatial distance index is:

[0012] ,

[0013] in, It is a spatial distance indicator.

[0014] Furthermore, the functional dependency relationship is constructed by expert scoring or based on domain knowledge to form a functional dependency matrix with a value range of [0, 1], where 0 represents no functional dependency and 1 represents high functional dependency.

[0015] Furthermore, the image recognition model used is the VGGNet model.

[0016] The present invention also provides a geographic entity classification device based on knowledge graph neighborhood learning, comprising at least a microprocessor and a memory, wherein the microprocessor is programmed or configured to perform the steps of the above method, or the memory stores a computer program programmed or configured to perform the above method.

[0017] The present invention also provides a computer-readable storage medium storing a computer program programmed or configured to perform the above-described methods.

[0018] The geographical entity classification method and apparatus based on knowledge graph neighborhood learning proposed in this invention first uses an image recognition model to generate initial vectors for each geographical entity and performs initial classification, fully utilizing the advantages of image recognition models in feature extraction and classification. Then, by constructing a geographical knowledge graph, the neighborhood information of the entities is learned, and the entity vectors are adjusted. The knowledge graph can clearly express the semantic relationships and neighborhood information between entities. By learning this information, the contextual associations of entities in the geographical scene can be better captured, thereby making up for the shortcomings of traditional methods that ignore neighborhood information, making the final classification results more accurate and significantly improving the recognition accuracy of various geographical entity targets such as buildings, bare soil, cultivated land, and grassland. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the geographic entity classification method based on knowledge graph neighborhood learning involved in the embodiment.

[0020] Figure 2 This is a geographical knowledge graph involved in the example. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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] Figure 1 This is a flowchart of the geographic entity classification method based on knowledge graph neighborhood learning involved in this embodiment. The method described in the diagram includes:

[0023] The first step is to acquire geographic images of the region to be classified, determine the entity categories and category standard vectors, and use an image recognition model to extract features from the geographic images to obtain the initial vectors of each entity in the geographic images.

[0024] Determining entity categories involves categorizing various objects within an image to identify the entities of interest. In this example, park imagery was used. After analysis and organization, the following entity categories were identified: farmland, woodland, gardens, grassland, bare soil, buildings, roads, water bodies, vehicles, field ridges, embankments, rest pavilions, drainage ditches, building material piles, parking lots, etc. These entities encompass different aspects of the geographic landscape, including natural landscapes, man-made facilities, and transportation vehicles, laying the foundation for establishing subsequent relational triplets.

[0025] Simultaneously, this step utilizes the VGGNet model to generate initial vectors for each entity and perform initial classification. The VGGNet model is a classic convolutional neural network model, adept at extracting image features. Geographical imagery is input into the pre-trained VGGNet model, which gradually extracts key information through a combination of multiple convolutional and pooling layers. For example, convolutional layers identify local features, while pooling layers downsample to retain the main features. As the number of layers increases, it learns more abstract and advanced semantic features, laying the foundation for generating initial entity vectors. After the VGGNet model completes learning the entity image features, the output of the last fully connected layer is extracted as the initial vector for that entity. Each dimension may correspond to a specific attribute or feature combination of the image. In this way, the image information of the entity is transformed into a mathematical expression that the computer can understand and process.

[0026] The second step is to calculate the cosine similarity between the initial vector of each entity and the standard vector of each category, and extract the maximum cosine similarity between each entity vector and all standard vectors of all categories. For entity vectors with a maximum cosine similarity greater than or equal to the threshold, the entity is determined to belong to the corresponding category, and the result is used as the classification label data. For entity vectors with a maximum cosine similarity less than the threshold, the entity is determined to be an unclassified entity.

[0027] For cosine similarity, the closer the value is to 1, the more similar the vector directions and the more similar the entity image features. The formula for calculating cosine similarity is as follows:

[0028] ,

[0029] in, Let be the entity vector to be classified. For the j-th class standard vector, in this embodiment, the similarity threshold is set to 0.9.

[0030] The third step is to extract classified and unclassified entities from the geographic imagery, construct "entity-relationship-entity" triples, calculate the relationship weights between entities using a linear weighting method based on the spatial distance index and functional dependency between entities, and then integrate all triples to construct a geographic knowledge graph.

[0031] Based on spatial distance metrics and dependencies between entities, corresponding triples are established to describe the relationships between them. For example... Figure 2 The geographical knowledge graph shown has the following relationships:

[0032] Adjacent relationships: (farmland, adjacent, field ridge), farmland and field ridge are often closely connected, and the field ridge serves to separate and facilitate walking; (farmland 1, adjacent, farmland 2), in geography, different farmland plots may be adjacent to each other.

[0033] Locational relationships: (Building, located on, bare soil) Some temporary buildings or buildings with specific functions may be built on bare soil; (Rest pavilion, located in, garden) Rest pavilions are generally set up in gardens to provide rest places for tourists.

[0034] Surrounding relationship: (water body, surrounding, embankment). Embankments are usually built around water bodies to protect and define the boundaries of the water body.

[0035] Stacking Relationship: (Building material pile, piled on, bare soil) On construction sites, building material piles are often placed in bare soil areas.

[0036] Parking Relationship: (Vehicle, parked in, parking lot), a parking lot is a designated area for parking vehicles.

[0037] Weight calculation method:

[0038] The spatial distance index sets a distance threshold based on the actual geographical scenario and research needs. In this embodiment, the distance threshold is set to 100 meters in the park entity classification; when the distance between two geographic entities exceeds this threshold, their association weight due to spatial distance is directly set to 0. If the geographic entity is represented in a planar coordinate system, the Euclidean distance formula is used:

[0039] ,

[0040] When the distance between two geographic entities When calculating distances, normalization is required to map the distances to the [0,1] interval, with closer distances resulting in larger values. The normalization formula is as follows:

[0041] ,

[0042] in, It is a normalized spatial distance indicator.

[0043] Functional dependency analyzes the degree of interdependence between two entities in terms of their functions. A functional dependency matrix is ​​constructed using expert scoring or domain knowledge, with a value range of [0,1], where 0 represents no functional dependency and 1 represents high functional dependency. In this embodiment, strong correlations include those between buildings and roads, bare soil and piles of building materials, and cultivated land and field ridges. These relationships play a crucial role in the geographical layout and functional realization, and are given a large weight, set to 0.8; medium correlations include those between rest pavilions and gardens, parking lots and roads, etc., which play a certain role in the use and function of the geography, but are relatively weaker than strong correlations, and are weighted at 0.5; weak correlations include those with relatively low correlation, and are weighted at 0.2.

[0044] In this embodiment, a linear weighting method is used to construct the weight calculation model. The formula is as follows:

[0045] ,

[0046] Where W is the final calculated weight value, ranging from [0, 1]; It is a normalized spatial distance indicator; Functional dependency. , These are the weighting coefficients of the two feature indicators, and These coefficients can be adjusted based on the actual application scenario and expert experience. For example, if spatial location has a significant impact on the correlation, they can be increased. The value of .

[0047] The fourth step involves obtaining classified entities within the neighborhood of an unclassified entity based on the geographic knowledge graph. The initial vector of the unclassified entity is then adjusted using the weights of the classified entities within the neighborhood and the relationships between entities to obtain the final vector of the unclassified entity. Finally, the cosine similarity between the final vector of each unclassified entity and the standard vectors of all categories is calculated, and the category with the highest cosine similarity is determined as the category of the unclassified entity.

[0048] For an unclassified entity, it may have multiple neighboring entities, and the vector of this unclassified entity needs to be adjusted based on these neighboring entities. For unclassified entity i, the adjustment is based on a weighted aggregation of classified entities j within its neighborhood. The specific adjustment formula is as follows:

[0049] ,

[0050] in, Let be the initial vector of entity i. Let j be the feature vector of the neighboring entity j. For the adjusted new vector, Weights of relationships between entities (from Module 2). Let i be the neighborhood set of entity i. The global adjustment factor (recommended 0.05–0.2) is used for the normalized denominator. Ensure that the adjustment range is not affected by the size of the neighborhood to avoid excessive offset.

[0051] The similarity measurement formula between the adjusted vector and the standard vectors of each category is as follows:

[0052] ,

[0053] in, Let be the entity vector to be classified. It is a standard vector of class c.

[0054] The formula for calculating the final category of an unclassified entity is as follows:

[0055] ,

[0056] This embodiment first uses an image recognition model to generate initial vectors for each geographic entity and performs initial classification, fully leveraging the advantages of image recognition models in feature extraction and classification. Then, by constructing a geographic knowledge graph, it learns the neighborhood information of the entities and adjusts the entity vectors. The knowledge graph can clearly express the semantic relationships and neighborhood information between entities. By learning this information, it can better capture the contextual associations of entities in the geographic scene, thereby making up for the shortcomings of traditional methods that ignore neighborhood information, making the final classification results more accurate and significantly improving the recognition accuracy of various geographic entity targets such as buildings, bare soil, cultivated land, and grassland.

[0057] In another embodiment, a geographic entity classification apparatus based on knowledge graph neighborhood learning is also provided, comprising at least a microprocessor and a memory, wherein the microprocessor is programmed or configured to perform steps of a geographic entity classification method based on knowledge graph neighborhood learning, or the memory stores a computer program programmed or configured to perform a geographic entity classification method based on knowledge graph neighborhood learning.

[0058] In another embodiment, a computer-readable storage medium is also provided, in which a computer program is stored that is programmed or configured to perform a geographic entity classification method based on knowledge graph neighborhood learning.

Claims

1. A method for geographic entity classification based on knowledge graph neighborhood learning, characterized in that, The method includes: Geographic images of the region to be classified are acquired, entity categories and category standard vectors are determined, and image recognition models are used to extract features from the geographic images to obtain initial vectors for each entity in the geographic images. Calculate the cosine similarity between the initial vector of each entity and the standard vector of each category. Extract the maximum cosine similarity between each entity vector and all standard vectors of all categories. For entity vectors with a maximum cosine similarity greater than or equal to the threshold, the entity is determined to belong to the corresponding category, and the result is used as the classification label data. For entity vectors with a maximum cosine similarity less than the threshold, the entity is determined to be an unclassified entity. The system extracts classified and unclassified entities from geographic images, constructs "entity-relationship-entity" triples, calculates the relationship weights between entities using a linear weighting method based on spatial distance indicators and functional dependencies between entities, and then integrates all triples to construct a geographic knowledge graph. Based on the geographic knowledge graph, classified entities within the neighborhood of an unclassified entity are obtained. The initial vector of the unclassified entity is adjusted using the classified entities within the neighborhood and the relationship weights between entities to obtain the final vector of the unclassified entity. Then, the cosine similarity between the final vector of each unclassified entity and the standard vectors of all categories is calculated, and the category with the highest cosine similarity is determined as the category of the unclassified entity.

2. The method according to claim 1, characterized in that, The initial vector of the unclassified entity is adjusted using the classified entities in the neighborhood to obtain the final vector of the unclassified entity. The adjustment formula is as follows: ， in, Let be the initial vector of entity i. Let j be the feature vector of the classified entity j in the neighborhood. Let i be the final vector of entity i. Weights for relationships between entities. Let i be the neighborhood set of entity i. This is the global adjustment coefficient.

3. The method according to claim 1, characterized in that, Based on the spatial distance index and functional dependency between entities, a linear weighted method is used to calculate the relationship weight between entities. The specific calculation formula is as follows: ， Where W is the weight value of the relationship between entities, ranging from [0,1]; It is an indicator of spatial distance between entities. Functional dependency between entities , These are the weighting coefficients for spatial distance and functional dependence, respectively. .

4. The method according to claim 3, characterized in that, The spatial distance index between entities is specifically calculated using the Euclidean distance formula. ,when Greater than the distance threshold At that time, the spatial distance index is set to 0, and the calculated distance is... Less than the threshold At that time, the formula for calculating the spatial distance index is: ， in, It is a spatial distance indicator.

5. The method according to claim 3, characterized in that, The functional dependencies are defined by expert scoring or by constructing a functional dependency matrix based on domain knowledge. The value range is [0, 1], where 0 represents no functional dependency and 1 represents high functional dependency.

6. The method according to any one of claims 1 to 5, characterized in that, The image recognition model used is the VGGNet model.

7. A geographic entity classification device based on knowledge graph neighborhood learning, comprising at least a microprocessor and a memory, characterized in that, The microprocessor is programmed or configured to perform the steps of the method according to any one of claims 1 to 5, or the memory stores a computer program programmed or configured to perform the method according to any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that is programmed or configured to perform the method according to any one of claims 1 to 5.

Images