CN113761971B - Remote sensing image target knowledge graph construction method and device - Google Patents
Remote sensing image target knowledge graph construction method and device Download PDFInfo
- Publication number
- CN113761971B CN113761971B CN202010491409.3A CN202010491409A CN113761971B CN 113761971 B CN113761971 B CN 113761971B CN 202010491409 A CN202010491409 A CN 202010491409A CN 113761971 B CN113761971 B CN 113761971B
- Authority
- CN
- China
- Prior art keywords
- target
- entity
- remote sensing
- sensing image
- entities
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
- G06F16/367—Ontology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Computation (AREA)
- Evolutionary Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Animal Behavior & Ethology (AREA)
- Computational Linguistics (AREA)
- Databases & Information Systems (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention relates to a remote sensing image target knowledge graph construction method and device, and belongs to the technical field of remote sensing image data processing. According to the method, a remote sensing image target knowledge extraction model is respectively constructed aiming at different data types of remote sensing image targets, and then target entity extraction is carried out on the obtained remote sensing image targets based on the similarity between the entity to be detected and the corresponding ontology concept; and extracting the obtained target entities according to a predefined mode, and finally selecting corresponding target entities based on the similarity between the remote sensing image target entities and the entities in the existing knowledge base to realize the link of the remote sensing image target entities. Therefore, the invention realizes the link between the target information of image discrimination and the semantic network of the knowledge graph, enriches the semantic information of the remote sensing image target and provides support for further large-scale target association analysis.
Description
Technical Field
The invention relates to a remote sensing image target knowledge graph construction method and device, and belongs to the technical field of remote sensing image data processing.
Background
The remote sensing image target classification and identification are important components of information extraction and processing of a high-resolution earth observation system and an automatic target identification system, are research hotspots and difficulties in the remote sensing field, and play a very important role in intelligent transportation, smart cities, target dynamic monitoring and positioning and other applications. The classification and identification of the remote sensing image targets aim at improving the accuracy, the intelligent level, the real-time performance and the processing efficiency of algorithm processing. Along with the deep research of the remote sensing image training data set, the accuracy and the recognition efficiency of image target recognition are greatly improved.
However, there is a semantic gap between the recognition of the target of the remote sensing image and the cognition of the remote sensing image. The target classification and recognition of the remote sensing image is essentially realized by using machine learning, deep learning and other methods, for example, the target classification such as 'plane', 'ship' and the like in the target image is rapidly recognized based on a deep learning model trained by a remote sensing image data set with labels, but due to the lack of related target semantic information, further 'cognition' of the remote sensing image target is very difficult to realize.
The remote sensing image knowledge is mainly used for filling the problem of semantic deletion of low-level information of images, and researchers in the related field perform a great deal of research around the concept, classification and application of the remote sensing image knowledge. Knowledge of remote sensing images is known differently for different applications, for example: li Sheng (Li Sheng.2018. Urban earth surface coverage change detection method combining field knowledge and deep learning. Wuhan: university of Wuhan) the remote sensing image knowledge is divided into image knowledge, geographic knowledge and change pattern prior knowledge; peng femto (Peng femto.2018. Research on high spatial resolution remote sensing image segmentation method of knowledge constraint. Beijing: china university of geology (Beijing.)) aiming at the remote sensing image segmentation problem, the knowledge constraint in remote sensing image segmentation is divided into an internal knowledge constraint and an external knowledge constraint; sun Gubo (Sun Gubo.2014. Knowledge-based automatic extraction technique research of high-resolution remote sensing image cultivated land. Beijing: china university of agriculture.) the related knowledge is divided into: knowledge of feature spectra, knowledge of feature textures, knowledge of feature geometries, and the like. Gu Haiyan et al (Gu Haiyan.2015. Object classification technique driven by remote sensing image geographic ontology modeling. Wuhan: university of Wuhan.) divide the knowledge system of geographic entities into four classes: geographic knowledge, remote sensing image features, image object features and expert knowledge. Research and application of knowledge semantics and the like in the remote sensing field can be roughly divided into three types:
(1) Knowledge-based remote sensing image segmentation
The core is to realize hierarchical relation description among ground object types based on a semantic network, realize remote sensing image classification (Zhang Jianting and the like, 2016; li Sheng, 2018; peng fly, 2018 (Zhang Jianting and the like, 2016. Remote sensing image classification based on semantic network. Remote sensing information, 31 (01): 38-42. Li Sheng. 2018. Urban earth surface coverage change detection method combining field knowledge and deep learning. Wuhan: wuhan university, peng fly. 2018. High spatial resolution remote sensing image segmentation method research of knowledge constraint. Beijing: chinese university (Beijing)), and then acquire, express, infer and accumulate knowledge for types such as cultivated land (Sun Gubo, 2014. Automatic extraction technology research of high resolution remote sensing image cultivated land based on knowledge. Beijing: automatic extraction of Chinese agricultural university).
(2) Knowledge-based remote sensing image target identification
Knowledge-based remote sensing target recognition is to recognize related targets in images by mainly utilizing related knowledge such as geometric features, ground feature attributes, contexts and the like of the targets. For example: carrying out remote sensing image port target identification (Chai Honglei.2015. Knowledge-based remote sensing image port target identification; capital: university of electronic technology.); based on heuristic rules, remote sensing image bridge information is identified by utilizing target priori knowledge (Mandal D P, et al 1996.Analysis of IRS imagery for detecting Man-made objects with a multivalued recognition system IEEE Transactions on Systems, man, and Cybernetics-Part A: systems and Humans,26 (2): 241-247); according to the characteristic change of the image characteristics of the road and the spatial relationship, a road target in the remote sensing image is extracted by a matched filtering mode (Haverkamp D S.2002.extraction straight road structure in urban environments using IKONOS satellite image, optical Engineering,41 (9): 2107-2111.).
(3) Knowledge graph cognitive calculation of remote sensing information
The remote sensing information knowledge graph is constructed to construct a remote sensing information interpretation model combining refinement, quantification, intellectualization and synthesis, so that intelligent remote sensing information extraction and intelligent calculation are realized. The remote sensing information map target is a remote sensing information map calculation method for constructing a picture element-target-pattern as a whole. Based on the remote sensing information map, referring to the visual cognition flow, the method can be used for automatically interpreting remote sensing data; xie Rong and the like construct a knowledge semantic model based on FCA-concept lattices (Xie Rong and the like 2017. Key technology for large-scale knowledge graph construction in the specific field of remote sensing satellites. Radio engineering, 47 (4): 1-6.). In recent studies, two-layer association and combination of "information patterns" and "knowledge patterns" has been noted (Luo Jiancheng et al 2020. Geographical pattern intelligent computing and pattern mining method research. Scientific journal of earth information, 22 (1): 57-75.).
In summary, the application of the remote sensing knowledge in the aspects of remote sensing image classification, target recognition, remote sensing image intelligent cognition and the like is wider and wider.
The knowledge graph is a technology which is developed very rapidly in the current artificial intelligence field, and the core of the knowledge graph is to construct a large-scale semantic net to make up for a 'semantic gap' between human perception and cognition layers, and starts to be in the brand-new corner in the fields of natural language question-answering, machine translation, product recommendation, knowledge mining and the like. The related technology of the knowledge graph provides a reference for filling the semantic gap of the remote sensing image cognition. Knowledge Graph (knowledgegraph) was developed from the original semantic network (semanteme net), and has become an important technical foundation in the field of artificial intelligence. The essence of a knowledge graph is a semantic web in which the nodes of the graph represent entities (entities) or concepts (concepts) and the edges of the graph represent various semantic relationships between entities/concepts. Through the link of the knowledge graph to massive information, a series of information can be quickly searched according to the keywords and the adjacent relations, so that the manual retrieval efficiency is greatly improved, and the core is a structured semantic knowledge base. The knowledge graph describes concepts, entities and relations thereof in the objective world in a structured form, and the information of the Internet is expressed to be more similar to the form of the human cognitive world, so that the capability of better organizing, managing and understanding mass information of the Internet is provided.
The target knowledge graph of the remote sensing image inherits the concept of the knowledge graph proposed by Google, and the target knowledge graph is similar to the remote sensing information graph in word, but different in connotation. The remote sensing information map inherits from the geologic information map proposed by Mr Peng, and the mutual conversion relation between the remote sensing pixel spectrum and the aggregate structure is established by inversion of the spectrum information, objectification extraction and multi-scale expression of the earth surface fine space structure map information, so that a quantitative, intelligent and fine combined map-integrated remote sensing space cognitive model is truly constructed.
The current remote sensing image target knowledge graph is a multi-mode knowledge graph (Xiong Hui, etc. 2019. Research on multi-mode data semantic relativity in cross-media knowledge graph construction. Information theory and practice, 42 (02): 13-18+24; yuxin PENG, etc. 2017.Cross-media analysis and reasoning: steps and directives. Front of Information Technology & Electronic Engineering,18 (01): 44-58.), as shown in fig. 1, the concept, entity, attribute and interrelationship of the remote sensing image target are structurally described through the extraction and association of multi-mode knowledge such as deep learning, natural language processing, remote sensing target recognition, etc. by the intelligent sensing means such as deep learning, encyclopedia knowledge, target, etc. so that the concept, entity and entity are mutually linked to form an image target semantic network. Although concepts and demands of the target knowledge graph of the remote sensing image are given, specific means for constructing the target knowledge graph of the remote sensing image are lacking at present.
Disclosure of Invention
The invention aims to provide a remote sensing image target knowledge graph construction method and device, which are used for solving the problem of semantic attribute information missing in the existing remote sensing image target recognition.
The invention provides a remote sensing image target knowledge graph construction method for solving the technical problems, which comprises the following steps:
1) Acquiring remote sensing image target data, including remote sensing image metadata information, remote sensing image content information, remote sensing image target product information and target encyclopedia information;
2) Extracting target entities from the obtained remote sensing image targets based on the similarity between the entity to be detected and the corresponding ontology concept;
3) Extracting the relation of the obtained target entities according to a predefined mode, and determining the relation among the target entities;
4) And selecting a corresponding target entity according to the identified relation between the image target entity and the extracted target entity through the similarity between the remote sensing image target entity and the entity in the existing knowledge base, and realizing the link of the remote sensing image target entity.
The invention also provides a remote sensing image target knowledge graph construction device, which comprises a processor and a memory, wherein the processor executes a computer program stored by the memory so as to realize the remote sensing image target knowledge graph construction method.
According to the method, a remote sensing image target knowledge extraction model is respectively constructed aiming at different data types of remote sensing image targets, and then target entity extraction is carried out on the obtained remote sensing image targets based on the similarity between the entity to be detected and the corresponding ontology concept; and extracting the obtained target entities according to a predefined mode, and finally selecting corresponding target entities based on the similarity between the remote sensing image target entities and the entities in the existing knowledge base to realize the link of the remote sensing image target entities. Therefore, the invention realizes the link between the target information of image discrimination and the semantic network of the knowledge graph, enriches the semantic information of the remote sensing image target and provides support for further large-scale target association analysis.
Further, in order to quickly and accurately determine the target entity, the target entity extraction process in the step 2) is as follows:
A. locating all concepts C (I) according to the ontology concept to be searched and arranging the concepts C (I) into an entity set
And defines the reference character set refer: { Refer1, refer2, refer3 … } and maximum allowable number of prefix characters T to be taken k ；
B. Increasing the prefix character number in the entity set, and updating the entity set to be:
if the entity->
The new character is different from the original character code, and is considered +.>
Is the entity name;
C. traversing the entity set to judge, and deleting the entity if the first character or the first two characters of the entity are the reference words { re|re epsilon Referet }; record the number of deletion entities as d 1 The update entity set is:
counting the number of entity names in the collection;
D. steps B and C are cyclically performed until the maximum allowable number of prefix characters T is reached k And counting the entity with the highest occurrence frequency, wherein the entity is the extracted target entity.
Further, in order to accurately and comprehensively obtain the relationship between the two target entities, the relationship extraction process in the step 3) is as follows:
a. combining all the obtained target entities pairwise to obtain a binary group set of the target entities;
b. traversing a binary group set of a target entity, comparing the concept of the ontology with the predefined relationship between the entities for each combination in the binary group set, and determining the relationship of each combination in the binary group set;
c. traversing the determined relationship, eliminating the zero matching relationship, wherein the eliminated relationship is the relationship of corresponding combination in the binary group set, so as to obtain the relationship between the target entities.
Further, the implementation process of the remote sensing image target entity link in the step 4) is as follows:
and taking the distance of the target entity as a similarity characteristic, calculating the similarity characteristic of the target entity and the candidate entity in the existing knowledge base, weighting each similarity characteristic to form a weighted characteristic value, taking the candidate entity with the largest weighted characteristic value as the best matching entity of the target entity, and realizing the link between the target entity and the existing knowledge base.
Further, in order to accurately determine the weight of the similarity, the weight of each similarity feature is obtained by selecting a geographic knowledge base as a training set and extracting an entity set l= { with sufficient link information therein<m i ，e i >And (3) using a Logistic regression model as training data to calculate.
Further, the calculation method of the target entity distance feature is as follows:
and obtaining the geographic coordinates of the target entity according to a Geographic Information System (GIS) or map data, obtaining the coordinate range of the shadow image block according to the image metadata of the target entity, carrying out barycenterization on the shadow image block to obtain corresponding image coordinates, and converting the corresponding image coordinates into a geographic coordinate system, wherein the distance between the geographic coordinates of the target entity and the image coordinates under the converted geographic coordinates is the distance characteristic of the target entity.
Drawings
FIG. 1 is a diagram of multi-modal knowledge graph correlation of a conventional remote sensing image;
FIG. 2 is a diagram illustrating a target knowledge graph of a remote sensing image according to the present invention;
FIG. 3 is a flow chart of a method for constructing a target knowledge graph of a remote sensing image according to the invention;
FIG. 4 is a schematic diagram of correlation of target knowledge of remote sensing images in an embodiment of the method of the present invention;
FIG. 5 is a schematic diagram of the remote sensing image object ontology and relationship in an embodiment of the method of the present invention;
FIG. 6 is a schematic diagram of a target knowledge extraction process in an embodiment of the method of the present invention;
FIG. 7 is a schematic diagram of the results of a method embodiment of the present invention using prot g software to construct a body sample;
FIG. 8 is a schematic diagram of recognition results of unstructured text data in an embodiment of the method of the present invention;
FIG. 9 is a schematic diagram of a target-image information triplet in an embodiment of the method of the present invention;
FIG. 10 is a schematic diagram of knowledge completion of a target entity in an embodiment of the method of the present invention;
FIG. 11 is a schematic diagram of a candidate entity list ordering result in an embodiment of the method of the present invention
FIG. 12 is a schematic diagram of a remote sensing image target entity linking with a geographic knowledge base according to an embodiment of the method of the present invention;
FIG. 13 is a schematic diagram of the association of an image object with a remote sensing image established in an embodiment of the method of the present invention.
Detailed Description
The following describes the embodiments of the present invention further with reference to the drawings.
Embodiments of the control method
The remote sensing image target knowledge can be expressed in a triplet form by adopting a resource description framework (Resource Description Framework, RDF), namely "< s (subject), p (predicate), o (object) >", and a large-scale directed graph consisting of 'point-edge'. The midpoint represents a target concept, a target entity and an attribute value of the remote sensing image, and the edge represents a relationship between concepts, a relationship between concepts and entities, a relationship between entities and attributes, and a relationship between attributes and attribute values. For example, as shown in FIG. 2, the relationship of entities to concepts: < Luoyang television tower, example, television tower >; concept-to-concept relationship: < television tower, subclass, communication facilities >; relationship of entities to attributes: < Luoyang thermal power plant, topography, flat open >.
According to the invention, knowledge graph and remote sensing image target interpretation are combined, unfolding research is constructed around the remote sensing image target knowledge graph, and the link between the image distinguished target information and the knowledge graph semantic net is realized through knowledge extraction and knowledge link based on multi-source data such as remote sensing image metadata information, target information and geographic knowledge, and the specific flow is shown in figure 3. The method comprises the steps of constructing a remote sensing image target knowledge graph, namely, constructing two core contents by a mode layer and a data layer, wherein the mode layer construction mainly comprises the construction of a remote sensing image target body and a concept set; the data layer construction is mainly to extract specific data of the target instance object. The following describes the construction process of the target knowledge graph of the remote sensing image in detail in connection with a specific example.
1. And acquiring remote sensing image target data comprising remote sensing image metadata information, remote sensing image content information, remote sensing image target product information and target encyclopedia information.
The image product mainly comprises the following elements: 1) The corrected original image; 2) Image sub-blocks (patches); 3) Image metadata information (metadata); 4) A feature data set extracted based on the image; 5) A target information description file; 6) Semantic tags of the image sub-blocks; 7) An image target body covered in the image range. The remote sensing image target knowledge represents knowledge of 7 elements such as concepts, entities, attributes and the like related to an image product, and is formally represented as a form which can be identified by a computer.
(1) Remote sensing image metadata information: the storage form is metadata description file, the formatting is better, and the metadata description file is generally stored in xml or txt format and mainly comprises information such as track number, line number, image description, sensor type, satellite type, image source, longitude and latitude coordinates and the like.
(2) Remote sensing image content information: the remote sensing image content can obtain image content label data through various intelligent segmentation algorithms and target detection algorithms, such as airplane, port and the like labels obtained by the target detection algorithm.
(3) Remote sensing image target product information: the target product data is a product subjected to manual verification, and not only the target type but also specific target attributes are generally determined.
(4) Target encyclopedia information: the encyclopedia information records relatively perfect target attribute information, and is helpful for grasping the target information more comprehensively.
2. And carrying out image target knowledge on the remote sensing image to carry out entity identification.
Image target entity identification mainly aims at different data sources, different entity extractors are constructed on structured, unstructured and semi-structured data (such as metadata information of remote sensing images, basic geographic information data and the like), and extraction of target entities is achieved, as shown in fig. 6.
The metadata information, the target product information and the encyclopedia information are well structured, belong to important target achievements of the image products, and can be related with target attributes from a semantic level through target entity association. The remote sensing image content information can be associated with a mode (concept) layer of a remote sensing image target map through a target classification label, as shown in fig. 4.
According to the target knowledge classification of the remote sensing image, a remote sensing image target ontology model is built mainly based on the geographic ontology and the target ontology of the remote sensing image according to a classification method.
(1) Define the Concept set ON_Concept: { C }
On_accept: { C } is defined as: { C1: image characteristics, C2: object feature, C3: geographic ontology, C4: target body, C5: encyclopedia ontology }.
(1) The remote sensing image characteristic body comprises: thematic characteristics, sensor type, resolution, location, spectrum, layer characteristics, etc.;
(2) the remote sensing image object feature body comprises: geometric features, object classification, location features, texture features, scene features, thematic indices, etc.;
(3) the geographic ontology includes: cities, highways, rivers, lakes, buildings, weather, etc.;
(4) the target body includes: traffic facilities, economic facilities, manual facilities, social mankind, natural elements, information facilities, etc.;
(5) the encyclopedia ontology includes: geography, culture, life, character, sports, etc.;
(2) Define the Relationship on_relationship in the ontology:
{ R } = { R (C1, C2), R (C2, C5), R (C2, C6), R (C6, C7) }, wherein: (1) r (C1, C2) is the membership of the original image and the image sub-block; (2) r (C2, C6) is the relation 'tag semantics' between the image sub-block and the semantic tag; (3) r (C2, C5) is the relation 'file annotation' between the image subblocks and the target annotation file; (4) r (C6, C7) is the relation 'association' between the semantic tag and the image target concept. Only semantic relations among five types of ontologies are defined, and the relation among the various ontologies still adopts the ontological relation in the self field.
The structure of the target body of the remote sensing image is shown in fig. 5, and it can be seen that from the remote sensing image, three possible links are linked to the target body: (1) remote sensing image ontology-object feature-geographic ontology-target ontology; (2) remote sensing image body-object feature-target body; (3) remote sensing image ontology, image characteristics, open link knowledge and target ontology.
The core of remote sensing image target entity identification is to calculate the similarity between the entity to be detected and the corresponding ontology concept, and mainly depends on two key rules:
(1) Similarity rules for concept names and target entity names
The similarity rules of the ontology concept and the target entity determine how the entity is located. Such rules are often "inclusion rules" defined as:
where I represents an entity, C (I) represents its corresponding concept category, string () represents a String operation, including rules that are widely present among ontology concepts and instances, because for named entities of the same category, the way to distinguish uniqueness is typically to add characters. For example, the concept "airport", an example of which is "Xinzheng airport", and the character string of the concept "airport" are part of the character string of the example "Xinzheng airport".
(2) Dissimilarity rule of concept names and target entity names
The rules of dissimilarity of ontology names with target entities determine how an entity determines a boundary. In the actual corpus, besides explicitly including rules, semantic components and positions thereof, which are lacking in concept names compared with entity names, are required to be explicitly determined, namely, dissimilarity rules of the concept names and target entities are mainly prefix rules through analyzing a large number of corpora, and can be defined as:
where prefixEntity (k) denotes a prefix entity name where k characters exist before the concept name. For most concepts, 1.ltoreq.k.ltoreq.T k ，T k Representing the maximum allowable number of prefix characters taken, since 15.ltoreq.T is typically taken k ≤20。
The steps for entity identification based on inclusion rules and prefix rules are as follows:
(1) locating all concepts C (I) according to the concepts to be searched by the ontology, and listing into a set
(2) The definition refers to the character set refer set: { Refer1, refer2, refer3. First, T k ；
(2) The number k of the prefix characters is increased by 1, and the entity set is updated as follows:
if the entity is
The new character is different from the original character code, and is considered +.>
Is the entity name.
(4) Traversing the set to judge, if the first character or the first two characters of the entity are the reference words { re|re epsilon Referet }, deleting the entity; record the number of deletion entities as d 1 The update set is:
(5) counting the number of entity name occurrences in the collection:
for 1.ltoreq.k.ltoreq.T k Circularly executing the steps (2) to (5);
(6) counting the entity with the highest occurrence frequency:
where t represents the entity name generated after the t-th iteration.
3. And extracting the relation of the identified target entity based on a predefined model.
The relationship between target entities is determined based primarily on predicates between two entity references, if the relationship between entities is explicitly revealed in the data, such as < for..offer >, indicating that there is a < provision relationship > between the two entities. If the relation between the entities is implicitly displayed in the data, all semantic sequences of predicates need to be called out, and the optimal semantics are solved.
The entity relation extraction in the remote sensing image target body belongs to the category of the limit domain relation extraction. In the extraction of the defined domain relationship, since the relationship is predefined, a labeling corpus is automatically constructed by adopting heuristic rules for training. Let the relation set defined in the current ontology be { R ] 1 ，R 2 ....Ri...R n In the actual corpus, each relation R is greater than or equal to 1 and less than or equal to n i Will appear r i The species expression forms, the collection of which is recorded as
Extraction based on front and rear entities and predefined patternsThe method comprises the following implementation processes:
(1) and detecting a section of remote sensing image target description by using an entity detection module, and returning to an entity list:
{E 1 ，E 2 ...E i ....E N }，1≤i≤n (8)
(2) will { E 1 ，E 2 ...Ei....E N All entities in the } are combined. Obtaining a binary group set:
{(E 1 ，E 2 )...(E i ，E j )，....(E N-1 ，E N )}，1≤i＜j≤N (9)
(3) traversing the binary set, comparing the predefined relationship between concept entities of the ontology, finding
A group relationship;
(4) traversing the above relation, eliminating zero matches (marked as z) to obtain a relation combination:
(5) combining the two-tuple entities and relationships to generate a triplet
Wherein, the liquid crystal display device comprises a liquid crystal display device,
representation and E r Paired entities.
4. And carrying out image target entity link according to the relationship between the identified image target entity and the extracted target entity.
The entity link core of the image target is used for calculating the similarity between the remote sensing image target entity and the entity in the existing knowledge base, and selecting the corresponding target entity based on the similarity so as to realize the link between the remote sensing image and the target entity. The method is realized based on a Logistic regression model by adopting a plurality of features, aiming at the characteristic that a remote sensing image target has a geographic position, adding the distance of the target as the feature of the model, firstly establishing a mapping set of entity mention and candidate entities, and then calculating the feature values of all the entities to evaluate the similarity between the entities.
The core of the method is that the computing entity m and the candidate entity E E m The weighted characteristic value F between the two is calculated as follows:
F＝ω 1 ×f 1 (m，e)+ω 2 ×f 2 (m，e)+ω 3 ×f 3 (m，e)+...+ω n ×f n (m，e) (12)
in the candidate entity matching set, the entity with the highest weighted characteristic value is calculated according to the formula and becomes the final matching pair of the entity m. Wherein ω is the weight of each feature value, f 1 、f 2 、f 3 、f 4 The method comprises the steps of respectively representing entity existence, attribute correlation, topic correlation and distance correlation characteristics, extracting similarity characteristics, and learning different similarity characteristics from training corpus through a model.
In the candidate entity matching set, the entity with the highest weighted characteristic value is calculated according to the formula and becomes the final matching pair of the entity m. The specific method is to select a geographic knowledge base as a training set and extract an entity set l= { with sufficient link information (i.e. a large number of links exist)<m i ，e i >And taking the characteristic values as training data, and then carrying out weight calculation of each characteristic value by using a Logistic regression model. For the entity m and the corresponding candidate entity e, the weights ω of the respective feature values need to satisfy the following relationship:
ω·(f(m，e * )-f(m，e))＞0，(e∈E m ，e≠e * ) (13)
wherein ω=<ω 1 ，ω 2 ，ω 3 ，...，ω n >，f＝<f 1 ，f 2 ，f 3 ，...，f n >。
For entity e 1 And entity e 2 Chain-oriented realThe probability of a volume m can be calculated using a sigmoid function, the calculation formula of which is as follows:
if s (m, e) 1 )＞s(m，e 2 ) Then P ((e) 1 ＞e 2 ) =true) > 0.5, otherwise P ((e) 1 ＞e 2 ) =true) < 0.5. The final weight can be determined by maximum likelihood estimation and Logistic regression model, and then the weight omega is substituted into the solution weighted eigenvalue, so as to find and predict new entity links.
After the candidate entity matching set is established, the entity needing to be optimally matched is selected, and the main steps of the method adopted in the text are as follows: firstly, defining 4 similarity characteristic values (f 1-f 4) to represent semantic association degrees between an entity m and a candidate entity e, so as to select an entity with the largest semantic similarity; weighting each feature value to form a weighted feature value F (so that 4 feature values better cooperatively represent semantic similarity); finally, the size of F is calculated and the maximum value is selected as the best match. The algorithm principle is as follows:
(1) inputting entity m to be linked with candidate set e<m i ，e i >；
(2) Initializing a threshold, referring to m for all entities i E M and candidate entity e i E, calculating characteristic values (f 1-f 4);
(3) executing the Logistic model to calculate the weight value:
F＝ω 1 ×f 1 (m，e)+ω 2 ×f 2 (m，e)+ω 3 ×f 3 (m，e)+ω 4 ×f 4 (m，e)
(3) selecting the maximum value of F and the corresponding e i ；
(4) If the maximum value of F is greater than the initial threshold, then e * ←e i ；
(5) And outputting F and the best matching entity after the calculation is finished.
After the calculation of the weighted eigenvalues F of all the entities in the candidate entity matching set is completed, adding weighted eigenvalue items into the candidate entity matching set to form a new candidate set as follows:
m 1 ＝{<e 11 ，F 11 >，<e 11 ，F 11 >，…，<P 1p ，F 1p >}
the candidate sets are ordered according to the size of the F value, namely, the candidate sets are rewritten into:
m 1 ＝{<e 1i ，F max >，…，<e 1j ，F 1min >}
finally, selecting the entity corresponding to the maximum F value as the final best matching entity, namely finding out the entity with the association relation and establishing the corresponding entity link<m 1 ，e 1i >。
The distance characteristic calculation method of the target entity comprises the following steps: and acquiring a target geographic coordinate by using GIS or map data, acquiring a coordinate range of an image sub-block by using image metadata, converting the coordinates, comparing, calculating Euclidean distance or block distance, and judging the similarity. The method comprises the following steps:
(1) building a name field mapping M:
{M(F e ，F d )|F e →F d ，F e ∈EntityList，F d ∈GISDatabase} (15)
wherein EntityList represents an entity list, F e Representing the entity name. The map specifies the GIS database field in which the entity to be processed is located.
(2) And (5) taking out the physical geographic coordinates:
(3) taking out coordinates of the image sub-blocks and barycentering:
wherein, the liquid crystal display device comprises a liquid crystal display device,
x i and y i Representing the coordinates of boundary points of the sub-blocks, wherein i is more than or equal to 1 and n is more than or equal to n.
(4) Coordinate conversion, converting the accurate positioning point coordinates of the target entity into image coordinates GIL (x.y), wherein: { GIL (x.y) |GIL (x.y) to spatial reference_image }, "to" indicates compliance with spatial references.
(5) Judging whether the target is in the image sub-block range, and reducing the search space is as follows:
patchSet：{P 1 ，P 2 …P k …P p }，1≤k≤p (18)
(6) calculating Euclidean distance between the gravity centers of the target and the image sub-block:
(6) taking out the image sub-block with the minimum distance as a match:
min{D 1 ，D 2 ，…D k …D p }→P min ，1≤k，min≤p (20)
5. knowledge representation of remote sensing image target by OWL language
Knowledge representation refers to formalizing (formatting) facts, experiences and knowledge in the field, which is convenient for computer acceptance and operation. Knowledge representation methods must be able to accommodate the representation of the various types of knowledge acquired by the knowledge, organize the knowledge with a unified, simple and intuitive logic, and facilitate retrieval and utilization of the knowledge by the machine. The formal representation research of the geographic knowledge is developed from predicate logic, ontology language and description logic to knowledge framework, and the related research of describing the geographic knowledge by RDF, RDFS and OWL languages appears, so that the formal representation research of the geographic knowledge is knowledge representation of the geographic semantic network in the Internet age. According to the invention, the OWL knowledge representation framework is adopted to realize knowledge representation of the remote sensing image target knowledge graph. The OWL network ontology language (Ontology Web Language) is recommended by W3C, is an important RDF description language for networks, and is a powerful tool for ontology construction. In the OWL description language, there are three types of basic information, namely, individual (attribute), attribute (Property), class (Class), respectively. Each type of basic information has a predefined function F in describing the hierarchy and specifying the relationship and nature of the entities.
(1) Class (Class): the class represents a collection of concepts, and OWL describes the conditions of members in the class using formalized methods. A class definition consists of two parts: a name is introduced or referenced, and a restriction table. rdf: the abaut attribute provides a name or reference to the ontology. The relationships between category concepts are: the subclass attributes may be passed on with the same (Equi-valentTo), subclasses (subtlassof), instances (Instances), dissimilarities (distointWith), and the presence of a generative axioms (general ClassAxioms) between the subclasses and parent.
(2) Attribute (Property): the attributes possess a Domain (Domain) and a range (Ranges). Wherein, the definition field represents "the object list with the current attribute", and the value field represents "the range of the value of the attribute". In the OWL language, attributes are divided into two classes: object Properties (objects) and Data Properties (Data Properties). The object attribute is mainly used for describing dynamic information of the object, and is used for describing the relationship between different object attributes, such as 'intersecting A and B'. The "intersection" is an object attribute of a, which defines a domain as a, and the value domain is "all entity lists intersecting a". The data attribute is mainly used to describe static information of the object, such as "x number of articles owned by a", where "x number of articles" is a "data attribute", which is defined as a and the value field is a general data type xsd: number. In each OWL language, the first layer attribute is represented by < OWL: topObjectProperties >, the sub-attribute is represented by < subtypeof >, and the relationship between the attributes is represented by the same (EquivalentTo) and different (distointwidth). (3) Individual (indivisual): that is, the description of the instance includes the type (Types) to which the entity belongs, the attribute assertions (Property assertions) that the instance possesses. Relationships between entities include differences (diffmentindividals), the same (sameindividadals, etc.). In addition, in the OWL language, < OWL: rising rdf: ID/> represents a specific entity, < rdf: type rdf: resource= "#"/> to associate an entity and its class.
Device embodiment
The device provided by the embodiment comprises a processor and a memory, wherein the memory stores a computer program capable of running on the processor, and the processor realizes the method of the embodiment of the method when executing the computer program.
That is, the method in the above method embodiment should be understood that the remote sensing image target knowledge graph construction method may be implemented by computer program instructions. These computer program instructions may be provided to a processor such that execution of the instructions by the processor results in the implementation of the functions specified in the method flow described above.
The processor in this embodiment refers to a microprocessor MCU or a processing device such as a programmable logic device FPGA;
the memory referred to in this embodiment includes physical means for storing information, typically by digitizing the information and then storing the information in an electrical, magnetic, or optical medium. For example: various memories, RAM, ROM and the like for storing information by utilizing an electric energy mode; various memories for storing information by utilizing a magnetic energy mode, such as a hard disk, a floppy disk, a magnetic tape, a magnetic core memory, a bubble memory and a U disk; various memories, CDs or DVDs, which store information optically. Of course, there are other ways of storing, such as quantum storing, graphene storing, etc.
The device formed by the memory, the processor and the computer program is implemented in the computer by executing corresponding program instructions by the processor, and the processor can be loaded with various operating systems, such as windows operating systems, linux systems, android, iOS systems and the like.
As other embodiments, the apparatus may further comprise a display for displaying the established knowledge graph for reference by the staff.
In order to comprehensively evaluate the remote sensing image target knowledge graph construction method and device provided by the invention, the feasibility of the invention is evaluated by carrying out experiments on the invention through specific experimental data.
Experimental data selected for this experiment included: (1) The remote sensing image product is taken as basic data, and semantic tags are attached after target identification and detection; (2) image target data: 200 image target experimental data with attribute information; (3) geographical knowledge graph: contains 8000 thousands of entities, 1.2 hundred million relations. Experiment platform: the development platform is Windows 10, the development language adopts Python 3.7, and the development tool adopts JetBrains PyCharm.
(1) Remote sensing image target knowledge extraction result experiment
Taking the target data of "Luoyang thermal power plant", "Luoyang North train station", "Luoyang television tower" as an example, the data belong to three categories of economic facilities, traffic facilities and information facilities in a target ontology, and according to the target classification system and the attribute thereof, the ontology model is constructed by utilizing Prot e software, and the result is shown in figure 7.
For unstructured text data, entity recognition is achieved by using a similarity method, and a recognition result is shown in fig. 8. The relation recognition is realized through a predefined model, and the condition that the Rocyang television tower is positioned at Rocyang is mainly recognized; < Luoyang television tower located in Luo Pu park >; < Luoyang television tower belongs to building facilities > and other triad entity relations.
For the metadata file of the target, the file structure description is shown in table 1, the D2R method is utilized to realize knowledge extraction from the structured text to the target triplet, and the extraction result is shown in fig. 9.
TABLE 1
The triplets are expressed (locally) in the OWL language, which describes the association by storing the image path in order to establish the association link of the entity and the remote sensing image.
< North railway station of Luoyang rdf: ID= "North railway station of Luoyang" >
< Country rdf: resource= "# Chinese"/>
< rdf: type rdf: resource= "# transportation facilities"/>
< surrounding Environment rdf: resource= "# plains at ground, flat open"/>
< strategic value rdf: resource= "# in"/>
< passenger flow rdf: resource= "#200000"/>
< image storage Path rdf: resource= "# …/…/LYPowerFactoryPatch. Tif ]"
North railway station of Luoyang
(2) Remote sensing image target entity link experiment
(1) The knowledge base is complemented by the attribute information of the existing target entity. 200 target entities are adopted as experiments, and related knowledge of the existing target entities is supplemented. As shown in fig. 10, the geographic knowledge base has an entity of "luoyang railway station", the related attribute information of which comes from hundred degrees encyclopedia, and the framed part is part of attribute information which supplements the target knowledge of the remote sensing image.
(2) And linking the remote sensing image target entity with the existing knowledge base. As shown in fig. 12, the left part is newly added remote sensing image target entity data, and the "entity-attribute" relationship is the result of knowledge extraction based on the target product data. The links are realized through the relationships between the entity links and the original geographic knowledge base through the entities. Related information of the geographic entity of Roche can be obtained through primary knowledge expansion, and corresponding target information of Roche east station, roche station and the like can be obtained through secondary expansion.
Taking a remote sensing image target of 'Luoyang television tower' (entity ID is 200) as an example, obtaining candidate pairs of entity links by matching with entities in an existing knowledge base, generating feature vectors through feature values, forming a candidate entity list according to the feature vectors, sorting the candidate entity list by using a ListNet sorting algorithm to obtain sorting scores of the entity list, sorting the candidate entity list, wherein the sorting result is shown in figure 11 (the second column is the candidate entity ID), and selecting the corresponding candidate entity with the score of 1 as the link entity.
(3) Correlation of remote sensing image target and remote sensing image. Through target entity searching, the link visualization of target entities and remote sensing images in the knowledge base is shown in the upper left corner of fig. 13, and remote sensing images associated with each target entity are displayed in nodes; and the right side is the visual mapping of the remote sensing image in the global three-dimensional environment obtained through the association query of the target entity.
Based on the experiments carried out by the remote sensing image with the label, the target entity and the existing knowledge base, the knowledge extraction based on the image metadata is realized, and the feasibility of the invention is verified by linking the remote sensing target knowledge with the existing knowledge base through entity links.
Aiming at different data types of remote sensing image targets, a remote sensing image target knowledge extraction model is firstly constructed, and then the relationship extraction of target entity identification and a predefined mode based on similarity is carried out; and finally, linking the image target entity based on the Logistic model considering the distance characteristics, so as to realize the knowledge linking of the remote sensing image target knowledge and the encyclopedia knowledge base.
Claims (5)
1. The construction method of the remote sensing image target knowledge graph is characterized by comprising the following steps of:
1) Acquiring remote sensing image target data, wherein the remote sensing image target data comprises remote sensing image element data information, remote sensing image content information, remote sensing image target product information and target encyclopedia information;
2) Extracting target entities from the obtained remote sensing image targets based on the similarity between the entity to be detected and the corresponding ontology concept;
3) Extracting the relation of the obtained target entities according to a predefined mode, and determining the relation among the target entities;
4) According to the relationship between the identified image target entity and the extracted target entity, selecting a corresponding target entity through the similarity between the remote sensing image target entity and the entity in the existing knowledge base, and realizing the link of the remote sensing image target entity;
the target entity extraction process in the step 2) is as follows:
A. locating all concepts C (I) according to the ontology concept to be searched and arranging the concepts C (I) into an entity set
And defines the reference character set refer: { Refer1, refer2, refer3 … } and maximum allowable number of prefix characters T to be taken k ；
B. Increasing the prefix character number in the entity set, and updating the entity set to be:
if the entity is
The new character is different from the original character code, and is considered +.>
Is the entity name;
C. traversing the entity set to judge, and deleting the entity if the first character or the first two characters of the entity are the reference words { re|re epsilon Referet }; record the number of deletion entities as d 1 The update entity set is:
counting the number of entity names in the collection;
D. steps B and C are cyclically performed until the maximum allowable number of prefix characters T is reached k Counting the entity with the highest occurrence frequency, wherein the entity is the extracted target entity;
the relation extraction process of the step 3) is as follows:
a. combining all the obtained target entities pairwise to obtain a binary group set of the target entities;
b. traversing a binary group set of a target entity, comparing the concept of the ontology with the predefined relationship between the entities for each combination in the binary group set, and determining the relationship of each combination in the binary group set;
c. traversing the determined relationship, eliminating the zero matching relationship, wherein the eliminated relationship is the relationship of corresponding combination in the binary group set, so as to obtain the relationship between the target entities.
2. The method for constructing a target knowledge graph of a remote sensing image according to claim 1, wherein the implementation process of the target entity link of the remote sensing image in the step 4) is as follows:
and taking the distance of the target entity as a similarity characteristic, calculating the similarity characteristic of the target entity and the candidate entity in the existing knowledge base, weighting each similarity characteristic to form a weighted characteristic value, taking the candidate entity with the largest weighted characteristic value as the best matching entity of the target entity, and realizing the link between the target entity and the existing knowledge base.
3. The method of claim 2, wherein the weight of each similarity feature is obtained by selecting a geographic knowledge base as a training set and extracting a set of entities l= { with sufficient link information<m i ,e i >And (3) using a Logistic regression model as training data to calculate.
4. The method for constructing a target knowledge graph of a remote sensing image according to claim 2, wherein the method for calculating the distance characteristics of the target entity is as follows:
and obtaining the geographic coordinates of the target entity according to a Geographic Information System (GIS) or map data, obtaining the coordinate range of the shadow image block according to the image metadata of the target entity, carrying out barycenterization on the shadow image block to obtain corresponding image coordinates, and converting the corresponding image coordinates into a geographic coordinate system, wherein the distance between the geographic coordinates of the target entity and the image coordinates under the converted geographic coordinates is the distance characteristic of the target entity.
5. A remote sensing image target knowledge graph construction device, characterized in that the device comprises a processor and a memory, the processor executes a computer program stored by the memory to implement the remote sensing image target knowledge graph construction method according to any one of the preceding claims 1-4.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010491409.3A CN113761971B (en) | 2020-06-02 | 2020-06-02 | Remote sensing image target knowledge graph construction method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010491409.3A CN113761971B (en) | 2020-06-02 | 2020-06-02 | Remote sensing image target knowledge graph construction method and device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113761971A CN113761971A (en) | 2021-12-07 |
| CN113761971B true CN113761971B (en) | 2023-06-20 |
Family
ID=78782967
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010491409.3A Active CN113761971B (en) | 2020-06-02 | 2020-06-02 | Remote sensing image target knowledge graph construction method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113761971B (en) |
Families Citing this family (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114638160B (en) * | 2022-05-11 | 2023-04-07 | 西南交通大学 | Knowledge service method for complex equipment digital twin model |
| CN115100643B (en) * | 2022-08-26 | 2022-11-11 | 潍坊现代农业与生态环境研究院 | Monocular vision positioning enhancement method and equipment fusing three-dimensional scene semantics |
| CN115795075B (en) * | 2022-11-29 | 2023-08-11 | 自然资源部国土卫星遥感应用中心 | Method for constructing universal model of remote sensing image product |
| CN115858840B (en) * | 2023-02-28 | 2023-05-05 | 北京数慧时空信息技术有限公司 | Remote sensing image mosaic method based on scene |
| CN117763169A (en) * | 2023-12-29 | 2024-03-26 | 航天科工(北京)空间信息应用股份有限公司 | Knowledge extraction method, device, equipment and storage medium in situation analysis field |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106056609A (en) * | 2016-06-02 | 2016-10-26 | 上海海洋大学 | Method based on DBNMI model for realizing automatic annotation of remote sensing image |
| WO2018036239A1 (en) * | 2016-08-24 | 2018-03-01 | 慧科讯业有限公司 | Method, apparatus and system for monitoring internet media events based on industry knowledge mapping database |
| CN109271530A (en) * | 2018-10-17 | 2019-01-25 | 长沙瀚云信息科技有限公司 | A kind of disease knowledge map construction method and plateform system, equipment, storage medium |
| CN109800671A (en) * | 2018-12-28 | 2019-05-24 | 北京市遥感信息研究所 | The multi-source remote sensing information knowledge map construction method and system of object-oriented interpretation |
| CN110209839A (en) * | 2019-06-18 | 2019-09-06 | 卓尔智联(武汉)研究院有限公司 | Agricultural knowledge map construction device, method and computer readable storage medium |
| CN110472066A (en) * | 2019-08-07 | 2019-11-19 | 北京大学 | A kind of construction method of urban geography semantic knowledge map |
-
2020
- 2020-06-02 CN CN202010491409.3A patent/CN113761971B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106056609A (en) * | 2016-06-02 | 2016-10-26 | 上海海洋大学 | Method based on DBNMI model for realizing automatic annotation of remote sensing image |
| WO2018036239A1 (en) * | 2016-08-24 | 2018-03-01 | 慧科讯业有限公司 | Method, apparatus and system for monitoring internet media events based on industry knowledge mapping database |
| CN109271530A (en) * | 2018-10-17 | 2019-01-25 | 长沙瀚云信息科技有限公司 | A kind of disease knowledge map construction method and plateform system, equipment, storage medium |
| CN109800671A (en) * | 2018-12-28 | 2019-05-24 | 北京市遥感信息研究所 | The multi-source remote sensing information knowledge map construction method and system of object-oriented interpretation |
| CN110209839A (en) * | 2019-06-18 | 2019-09-06 | 卓尔智联(武汉)研究院有限公司 | Agricultural knowledge map construction device, method and computer readable storage medium |
| CN110472066A (en) * | 2019-08-07 | 2019-11-19 | 北京大学 | A kind of construction method of urban geography semantic knowledge map |
Non-Patent Citations (3)
| Title |
|---|
| 地理空间情报知识图谱构建方法概述;陈晓慧 等;《信息工程大学学报》;第21卷(第1期);第101-107页 * |
| 多源异构数据的大规模地理知识图谱构建;蒋秉川 等;《测绘学报》;第47卷(第8期);第1051-1059页 * |
| 遥感卫星特定领域大规模知识图谱构建关键技术;谢榕 等;《无线电工程》;第47卷(第4期);第1-6页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113761971A (en) | 2021-12-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113761971B (en) | Remote sensing image target knowledge graph construction method and device | |
| CN109378053B (en) | Knowledge graph construction method for medical image | |
| Wu et al. | Dynamic graph convolutional network for multi-video summarization | |
| Yin et al. | Region search based on hybrid convolutional neural network in optical remote sensing images | |
| CN114996488B (en) | Skynet big data decision-level fusion method | |
| Deng et al. | GAKG: A multimodal geoscience academic knowledge graph | |
| CN114896408B (en) | Construction method of material knowledge graph, material knowledge graph and application | |
| Gao et al. | Semantic trajectory representation and retrieval via hierarchical embedding | |
| Sun et al. | Remote sensing image interpretation with semantic graph-based methods: A survey | |
| CN115269899A (en) | Remote sensing image overall planning system based on remote sensing knowledge map | |
| Zhuang et al. | A multi-granularity knowledge association model of geological text based on hypernetwork | |
| Sheeren et al. | A data‐mining approach for assessing consistency between multiple representations in spatial databases | |
| Chandra et al. | Human cognition based framework for detecting roads from remote sensing images | |
| CN109583371A (en) | Landmark information based on deep learning extracts and matching process | |
| Fan et al. | DuMapper: Towards Automatic Verification of Large-Scale POIs with Street Views at Baidu Maps | |
| Sun et al. | Conflating point of interest (POI) data: A systematic review of matching methods | |
| CN115965085A (en) | Ship static attribute reasoning method and system based on knowledge graph technology | |
| Deng et al. | Covidia: COVID-19 Interdisciplinary Academic Knowledge Graph | |
| Guesmi et al. | Community detection in multi-relational bibliographic networks | |
| CN114925203A (en) | Conflict text prediction method based on graph embedding method | |
| Lin et al. | SRSG and S2SG: a model and a dataset for scene graph generation of remote sensing images from segmentation results | |
| CN113377739A (en) | Knowledge graph application method, knowledge graph application platform, electronic equipment and storage medium | |
| Gao et al. | Association and Recomendation for Geosciences Data Attributes Based on Semantic Similarity Measurement | |
| Liu et al. | Lekan: Extracting long-tail relations via layer-enhanced knowledge-aggregation networks | |
| Cao et al. | A survey on visual data mining techniques and applications |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |