Method for Processing a Land Shape Database Using Composite Embedding Vectors, and System Using the Same
The method addresses the limitations of conventional land shape classification by generating composite embedding vectors from GIS data, incorporating land orientation and environmental factors, ensuring precise and objective evaluations and searches, suitable for large-scale databases and AI integration.
Patent Information
- Application Number
- KR1020250150277
- Authority / Receiving Office
- KR · KR
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2025-10-17
- Publication Date
- 2026-11-11
- Estimated Expiration
- 2045-10-17
AI Technical Summary
Conventional land shape classification methods fail to adequately reflect the diverse shapes and environmental characteristics of actual land, leading to subjective and imprecise evaluations and searches, especially in large-scale database environments.
A land shape database processing method that extracts boundary polygons, calculates azimuth, divides boundary segments, assigns attributes, generates composite embedding vectors, and performs similarity searches using GIS data, incorporating land orientation and environmental factors to ensure objective and precise evaluations.
Enables objective and precise similarity evaluation and search, aligning with actual development conditions, and supports efficient searching and usability evaluation in large-scale databases, with potential expansion into AI learning-based automated land analysis systems.
Smart Images

Figure 112025116021653-PAT00015_ABST
Abstract
Description
Technology Field
[0001] The present invention relates to a database processing method that vectorizes land shape data based on a GIS (Geographic Information System) to perform similarity evaluation and search. More specifically, it belongs to the field of technology that generates a composite embedding combining the outer shape of the land, a boundary sequence, and surrounding environmental factors, and rapidly and precisely searches for similar candidates in large-scale land data through approximate nearest search. Background Technology
[0002] Existing methods for classifying land shapes have relied on very simplistic criteria. Typically, land has been classified based on limited geometric types such as squares, rectangles, and triangles, or land value has often been judged solely on whether it abuts a road. Such simplified approaches fail to adequately reflect the complex shapes of actual land and the diverse surrounding conditions.
[0003] For example, even if land is of the same rectangular shape, its usability and value can vary significantly depending on whether it faces south or north, and whether the adjacent road is a wide main road or a narrow side road. However, existing methods failed to distinguish these differences, frequently resulting in evaluation outcomes that were detached from reality.
[0004] Furthermore, existing classification methods were heavily dependent on the subjective judgment of the inputter, leading to problems where different results could be obtained even with the same data. In particular, it was difficult to reflect the actual distinctiveness between land parcels because they treated rotated shapes uniformly or failed to consider differences between surrounding land and facilities.
[0005] Furthermore, when searching for similar land in a database environment, relying on simple keyword or attribute filtering makes it difficult to strike a balance between result precision and search speed. Narrowing the criteria excessively restricts results, while broadening them includes a large amount of unnecessary data, causing difficulties in post-processing. This search method with a simplistic structure is unsuitable for handling large-scale land data and fails to provide the precise information necessary for actual development or policy formulation.
[0006] Ultimately, existing methods fail to consider important factors such as land orientation, boundary characteristics, and the surrounding environment due to overly simplified classification and search procedures. Consequently, there is a limitation in that it is difficult to obtain analysis results that reflect realistic usability. Prior art literature
[0007] Republic of Korea Registered Patent 10-2795814 The problem to be solved
[0008] The present invention aims to solve the problem that conventional land shape classification methods fail to adequately reflect the diverse shapes and environmental characteristics of actual land by merely considering limited types such as rectangles and triangles or partial conditions such as road frontage, and that results vary depending on the analyst's subjectivity. By calculating the area, perimeter, and azimuth from the land's boundary polygons and quantifying attributes and surrounding environmental factors for each boundary segment to store them as composite embedding vectors, the invention enables objective and precise similarity evaluation free from the inputter's subjectivity. Furthermore, the invention supports efficient searching and usability evaluation that aligns with actual development conditions even in large-scale database environments, and provides a database processing method that can be expanded into an AI learning-based automated land analysis system in the future. means of solving the problem
[0009] The land shape database processing method provided in the present invention includes a boundary polygon extraction step of extracting a boundary polygon of a target land from GIS data by a boundary polygon extraction module, an azimuth calculation step of calculating the main azimuth of the target land by an azimuth calculation module, a boundary segment division step of dividing the boundary polygon into a plurality of boundary segments by a boundary segment division module, a boundary attribute assignment step of assigning boundary attributes to each of the plurality of boundary segments by a boundary attribute assignment module, a shape vector generation step of generating a shape vector including an absolute azimuth shape vector and / or a rotation-invariant shape vector for the shape of the target land by a shape vector conversion unit, a surrounding information extraction step of extracting surrounding information which is information of a target adjacent to the target land by a surrounding information extraction unit, a composite embedding generation step of generating a composite embedding vector by combining the absolute azimuth shape vector, the rotation-invariant shape vector and the surrounding information by an embedding generation unit, a database storage step of storing the composite embedding vector and metadata in a database storage unit, and a similarity search step of evaluating similarity using the stored data.
[0010] In one embodiment, the azimuth calculation step may be characterized by calculating the main azimuth based on the direction of the main side of the land.
[0011] In one embodiment, the azimuth calculation step may be characterized by calculating the main azimuth of the land using principal component analysis (PCA).
[0012] In one embodiment, the boundary attributes assigned in the boundary attribute assignment step may be characterized by including the length, azimuth angle, boundary type, run sequence of the boundary type, and length ratio of each boundary segment.
[0013] In one embodiment, the shape vector generation step may be characterized by calculating a vector based on Hu moments, Fourier descriptors (FD), or Zernike moments.
[0014] In one embodiment, the surrounding information extracted in the surrounding information extraction step may be characterized by including information about adjacent roads, information about adjacent land, and information about adjacent facilities.
[0015] In one embodiment, the database storage step may be characterized by storing together land shape categories and azimuth buckets, which are categorized values obtained by dividing the azimuth angles of the target land into segments.
[0016] In one embodiment, the composite embedding generation step may be characterized by generating a composite embedding by combining a shape vector, a vector composed of boundary attributes, and a vector composed of surrounding information.
[0017] In one embodiment, the similarity search step may be characterized by including a candidate filtering step, an approximate search step, and a re-ranking step.
[0018] In one embodiment, the re-ranking step may be characterized by performing a weighted sum of cosine similarity between a shape vector, a vector composed of boundary attributes, and a vector composed of surrounding information.
[0019] In one embodiment, the method may further include an effective area calculation step that calculates the effective area ratio by reflecting the separation distance and boundary type of the target land based on surrounding information.
[0020] The land shape database processing system provided in the present invention comprises: a land data processing unit including a boundary polygon extraction module for extracting boundary polygons of a target land, an azimuth calculation module for calculating the main azimuth angle of the target land, a boundary segment division module for dividing the boundary polygons into a plurality of boundary segments, and a boundary attribute assignment module for assigning boundary attributes to the boundary segments; a shape vector conversion unit for generating a shape vector including an absolute azimuth shape vector and / or a rotation-invariant shape vector; a surrounding information extraction unit for extracting surrounding information; an embedding generation unit for generating a composite embedding; a database storage unit for storing embeddings and metadata; and a similar land search unit for performing a similarity search.
[0021] In one embodiment, the shape vector transformation unit may be characterized by including a Hu moment calculator, a Fourier descriptor calculator, and a Zernike moment calculator.
[0022] In one embodiment, the surrounding information extraction unit may be characterized by extracting information about adjacent roads, information about adjacent land, and information about adjacent facilities.
[0023] In one embodiment, the embedding generation unit may be characterized by generating a composite embedding by combining a shape vector, a vector composed of boundary attributes, and a vector composed of surrounding information.
[0024] In one embodiment, it may further include an effective area calculation module that calculates an effective area ratio reflecting the separation distance and boundary type of the target land using surrounding information.
[0025] In one embodiment, the similar land search unit may be characterized by including a candidate filtering module, an approximate search module, and a re-ranking module.
[0026] In one embodiment, the similar land search unit may be characterized by applying a weighting function based on the difference in azimuth angle when searching for similarity.
[0027] In one embodiment, it may be characterized by further including a recommendation module that recommends a corresponding building layout pattern based on the search results of a similar land search unit.
[0028] In one embodiment, the weight-based re-ranking module applies a Gaussian series damping function that causes the similarity to gradually decrease as the difference in azimuth angle between the query land and the candidate land increases, and the parameter σ representing the variance of the damping function may be set in the range of 15 degrees or more and 45 degrees or less. Effects of the invention
[0029] According to the present invention, a boundary polygon extraction module (110) extracts a boundary polygon of land from GIS data, and an azimuth calculation module (120) calculates the main azimuth, thereby expressing the shape and orientation of the land as objective numerical values. This has the effect of allowing land to be compared on the same standard without relying on the subjective judgment of the inputter.
[0030] In addition, the boundary segment division module (130) and the boundary attribute assignment module (140) divide the land boundary line into multiple straight sections and assign length, azimuth, boundary type, run sequence, etc. to each section, thereby reflecting in the data not only the shape of the land but also the way it interacts with the surrounding environment. As a result, even with the same external appearance, differentiated similarity evaluations are possible depending on the characteristics of the border, such as roads, buildings, and waterfronts.
[0031] Furthermore, the shape vector transformation unit (200) generates a rotation-invariant or absolute orientation-based shape vector using Hu moments, Fourier descriptors (FD), Zernike moments, etc., thereby enabling the land shape to be perceived differently depending on the direction or to be compared identically regardless of rotation. This ensures both precision and flexibility in shape comparison.
[0032] The surrounding information extraction unit (300) collects data regarding adjacent roads, adjacent land, adjacent facilities, etc., and converts them into an environmental factor vector. Through this, even if the shape is the same, if conditions such as surrounding road width, building density, and coverage rate differ, they can be reflected in the usability evaluation, thereby having the effect of determining the actual development potential more accurately.
[0033] The embedding generation unit (400) generates a composite embedding by combining a shape vector, a boundary attribute vector, and an environmental factor vector, and stores it in the database storage unit (500). As a result, composite features for each land parcel can be managed as a single vector, and fast and accurate searching can be performed in a large-scale database through approximate nearest neighbor search (ANN) and precise re-ranking. This leads to the effect of simultaneously securing search efficiency and precision even in large amounts of land data.
[0034] Finally, the effective area calculation module (700) calculates the buildable area by reflecting boundary regulations and setback distance conditions, thereby enabling a direct evaluation of actual construction and development potential beyond simple similarity evaluation. Additionally, this database can be utilized as an AI training dataset in the future, which has the effect of expanding into an automated land evaluation and recommendation system. Brief explanation of the drawing
[0035] FIG. 1 is a flowchart of a processing method according to an embodiment of the present invention. FIG. 2 is a configuration diagram showing the overall configuration of a processing system according to the embodiment of FIG. 1. FIG. 3 is a diagram explaining the process of extracting land shape data and calculating the main orientation angle according to the embodiment of FIG. 1. FIG. 4 is a diagram explaining the configuration of shape feature vectors, boundary vectors, environmental factor vectors, and composite embedding generation according to the embodiment of FIG. 1. FIG. 5 is a diagram explaining the configuration of surrounding information according to the embodiment of FIG. 1. FIG. 6 is a diagram explaining the step-by-step flow of candidate filtering, approximate nearest search, and precise re-ranking according to the embodiment of FIG. 1. Specific details for implementing the invention
[0036] Preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings. The configurations, shapes, sizes, etc., illustrated in the drawings for the purpose of explaining the embodiments may be somewhat exaggerated or schematically depicted for ease of understanding and clear explanation, and may differ from the actual configuration. Accordingly, the details described in the drawings are for illustrative purposes only and are not intended to limit the technical concept of the invention, and the scope of protection of the invention should be determined by the claims.
[0037] Furthermore, the terms used in this specification are defined to reflect, as much as possible, their functional meanings related to the technical content of the invention, and should be interpreted by a person skilled in the art considering the technical context of the invention and the content of the specification as a whole. While the definitions of terms may vary depending on general interpretation by users or in literature, in this specification they should be interpreted based on the technical configuration and effects of the invention.
[0038] Where in this specification any configuration or step is described as being "combined," "connected," "joined," or similarly expressed to another configuration or step, this should be understood to include the possibility that it may be combined or linked directly or indirectly through the other configuration or step. On the other hand, expressions such as "directly combined" or "directly connected" mean a case where they are directly connected without any separate intermediate configuration.
[0039] In addition, expressions such as "comprising" and "comprising" used throughout this specification are used to mean that additional components or ingredients may be included without excluding other components or ingredients, unless otherwise specifically stated. Therefore, even if a specific component or step is described as "comprising," it implies that other components or steps other than said component or step may be included.
[0041] Overview of this system
[0042] FIG. 1 is a flowchart of a processing method according to one embodiment of the present invention. FIG. 2 is a configuration diagram showing the overall configuration of a processing system according to one embodiment of the present invention.
[0043] Referring to FIGS. 1 and 2, the present system is a land shape database construction and analysis system designed to quantitatively analyze land shapes and efficiently search for similar lands. Conventionally, the mainstream approach involved simply classifying land shapes into a few basic types, such as rectangles and triangles, or considering only whether there was a road frontage. However, this method failed to accurately reflect the diverse shapes and environmental factors of actual land, and the results could vary significantly depending on the subjective judgment of the inputter. To overcome these limitations, the present system provides a structure that integrates precise analysis based on GIS data, mathematical feature vectorization, quantification of environmental factors, generation of complex embeddings, and Approximate Nearest Neighbor Search (ANN).
[0044] The basic flow of the system begins in the land data processing unit (100). The input land data is converted into a closed polygon by the boundary polygon extraction module (110), and the absolute azimuth is calculated by the azimuth calculation module (120) based on the principal side direction or principal component analysis (PCA). Subsequently, the boundary segment division module (130) divides the outer edge of the land into multiple straight line segments, assigns boundary type labels, and records detailed attributes such as segment length, azimuth, and length ratio through attribute memos. This process lays the foundation for quantifying not only the external shape of the land but also the actual location conditions.
[0045] The shape vector transformation unit (200) generates rotation-invariant feature vectors, such as Hu moments, Fourier descriptors, and Zernike moments, and direction-sensitive vectors that directly reflect absolute azimuths from the outline data. The surrounding information extraction unit (300) constructs surrounding information vectors by statistically calculating environmental factors such as road density, building coverage ratio, and road direction distribution within a certain radius of the surroundings. Subsequently, the embedding generation unit (400) generates a single composite embedding vector by combining the shape feature vector, the boundary attribute vector based on the boundary attribute assignment module (140), and the surrounding information vector. This vector is recorded in the database storage unit (500) and managed together with metadata such as category values, azimuth bucket values, and effective area ratios.
[0046] The search process consists of three stages: candidate filtering, approximate nearest search, and precise re-ranking. In the candidate filtering stage, suitable candidates are selected first from the entire database by utilizing category values and azimuth buckets. The approximate nearest search module (620) narrows down the candidate group by performing an approximate nearest search based on a composite embedding vector, and the weight-based re-ranking module (630) calculates the final ranking by weighting the similarity between shape, boundary, and environmental factors. The azimuth difference is quantitatively reflected through a damping function g, and the final result is provided to the user along with the top candidate group.
[0047] The system does not stop at simple similarity calculation, but calculates the ratio of buildable area reflecting separation distance and boundary type through the effective area calculation module (700). Additionally, the building layout pattern recommendation module (640) proposes building layout patterns based on search results of similar land, providing guidelines that can be utilized during the actual development or design process. For example, for land where the front of a main road is dominant, it may recommend the placement of the front entrance / exit, and for land with high waterfront accessibility, it may recommend a view-priority placement.
[0048] In summary, this system is an integrated platform that quantitatively analyzes land by comprehensively considering three layers—shape, boundaries, and environment—and performs efficient search and recommendation. Through this, it ensures both objectivity and precision, enabling fast and stable searching even in large-scale database environments. Furthermore, the datasets generated by this system can be utilized as training data for AI-based land valuation models, possessing the potential to expand into automated land valuation and urban planning systems. A detailed explanation follows below.
[0050] Extraction of land shape data
[0051] FIG. 3 is a diagram illustrating the process of extracting land shape data and calculating the main orientation angle according to an embodiment of the present invention.
[0052] (a) Extraction of boundary polygons (S110)
[0053] Referring to FIG. 2, the extraction of land shape data is a process that serves as the starting point of the present invention, and is a step of receiving original GIS data (11) and using a boundary polygon extraction module (110) to generate the outer shape of the target land in the form of a polygon (12). The boundary polygon extraction step (S110) utilizes coordinate information embedded in the map data to convert the location of each point into a coordinate pair and connects them to calculate a closed polygon of the land boundary. In this process, a alignment algorithm may be applied to correct errors such as broken or overlapping boundaries, and coordinate correction is performed based on a satellite reference coordinate system or a national standard coordinate system.
[0054] The extracted boundary polygon is used as basic data for calculating the area and perimeter of the land. The area is calculated according to the polygon area formula, and the perimeter is calculated by summing the lengths of all sides. Additionally, the centroid is defined based on the center of gravity of the land polygon, and this is linked to the subsequent azimuth calculation step (S120) to serve as an important reference point for assigning absolute azimuth values. This geometric calculation module is configured to calculate the area, perimeter, and centroid simultaneously.
[0055] Therefore, the land shape data extraction process of the present invention is significant in that it does not merely visualize the outline but secures precise polygonal data capable of numerical analysis. In particular, unlike existing technologies that relied on simple external shape classification, the present invention quantitatively calculates the geometric quantity of the land, thereby providing a basis for subsequent vectorization and similarity evaluation.
[0056] (b) Calculation of azimuth (S120)
[0057] Referring again to FIG. 2, the azimuth calculation is performed using data derived from the boundary polygon extraction module (110) and the geometric quantity calculation module. The azimuth calculation module (120) is responsible for calculating the main azimuth of the land shape, which is defined within an absolute azimuth range from 0° to 360°.
[0058] The first method involves deriving the primary azimuth based on the dominant side. The dominant side refers to the longest side of a land polygon, and the azimuth is defined based on the slope of this side. For example, if the dominant side extends in an east-west direction, the azimuth is calculated as 90° or 270°. This is a simple and intuitive calculation method that offers high reliability for quadrilaterals or regular polygons.
[0059] The second method utilizes Principal Component Analysis (PCA). PCA statistically analyzes the distribution of coordinate points of a land polygon to derive the axis with the largest variance as the principal component, defining it as the principal azimuth of the land shape. This method has the advantage of being able to derive consistent azimuths even for land with irregular polygons or curved outlines. PCA-based calculation is particularly effective when the principal azimuth is unclear or dispersed in complex land shapes.
[0060] The azimuth calculation result is structured to enable rotational transformation based on the center point coordinates. For example, if one intends to align all land features to the same reference axis, rotational transformation based on the absolute azimuth can be performed to generate rotation-invariant features. Additionally, a feature vector reflecting directionality can be calculated by directly utilizing the absolute azimuth value. As such, the azimuth calculation is not merely an auxiliary value, but acts as a key element reflecting directionality in the subsequent shape vector generation step (S200) and similarity evaluation.
[0061] Boundary segment division (S130) and boundary attribute assignment (S140)
[0062] The boundary segment division step (S130) is a process of dividing a boundary polygon into multiple straight-line segments according to a certain standard. The land outline is divided by the boundary segment division module (130), and each segment is defined by the coordinates of the start and end points. Attribute metadata is assigned to each segment. The attributes include the length and azimuth angle of the segment. Additionally, the run-length and length ratio of the boundary type are also recorded. The boundary type label and attribute memo constitute a unique boundary sequence for each land, which is subsequently used as a key input in the boundary attribute vector generation step.
[0064] Shape vector generation step (S200)
[0065] FIG. 4 is a diagram illustrating the configuration of shape feature vectors, boundary vectors, environment factor vectors, and composite embedding generation according to an embodiment of the present invention.
[0066] Shape vector generation is a step performed primarily by the shape vector transformation unit (200), which receives outline information divided by the boundary segment division module (130) as input and represents the land shape as a numerical vector. This step is designed to generate absolute orientation shape vectors and rotation-invariant shape vectors in parallel, thereby reflecting differences in cases where the direction is different even for the same shape, while also enabling comparisons that eliminate the influence of rotation transformation. The shape vector transformation unit (200) performs normalization preprocessing of the input polygon and applies normalization based on a reference length or reference area to ensure scale invariance. Normalization serves as a basis for controlling the excessive influence of specific dimensions during the subsequent combination process in the embedding generation unit (400).
[0067] The rotation-invariant shape vector is calculated through the Hu moment calculation module (210), the Fourier descriptor calculation module (220), and the Zernike moment calculation module (230).
[0068] (a) Hu moment calculation module (210)
[0069] The Hu moment calculation module (210) calculates feature vectors that are invariant to rotation, translation, and scale changes from the input land outer shape. This module calculates geometric moments based on polygon data from the boundary segment division module (130) and combines them into seven invariants to stably express structural characteristics such as symmetry and aspect ratio. In the present invention, directionality can be distinguished by reflecting that even identical shapes have different values depending on the difference in the main orientation angle. For example, it is possible to distinguish cases where rectangular land is arranged in the north-south direction and the east-west direction. The Hu moment provides simple yet stable feature quantities and is utilized as a base vector in the search for similar ratio structured land within the database and in the composite embedding process.
[0070] (b) Fourier descriptor calculation module (220)
[0071] The Fourier descriptor calculation module (220) represents the macroscopic outline of the shape by converting the land outline into a complex coordinate signal and then extracting frequency components through a Fourier transform. While low-frequency components reflect the overall shape and high-frequency components reflect detailed curvature, the present invention primarily utilizes low frequencies to reduce noise and stably represent only the main outline. This allows for the distinction of similar shapes, such as L-shapes and U-shapes, and compensates for outline differences that Hu moments cannot capture. Therefore, this module concisely quantifies the land outline and combines it with other vectors to increase the efficiency of searching for similar land.
[0072] (c) Zernike moment calculation module (230)
[0073] The Zernike moment calculation module (230) calculates fine features by normalizing the boundary polygon to a unit circle and projecting it onto a Zernike orthogonal polynomial. This module provides independent and non-redundant feature quantities through orthogonality and maintains the same vector values even for rotational transformations. In this invention, shapes can be compared regardless of direction, and at the same time, it reflects minute differences in curvature or irregularities that might be overlooked by Hu moments or Fourier technologists. For example, even for identical trapezoidal land, differences in corner angles or curvature can be precisely distinguished, contributing to securing fine structural differentiation within the database.
[0074] In summary, Hu moments utilize a combination of low-order moments to produce feature values that reflect information regarding symmetry and aspect ratio. Fourier descriptors convert the complex form signal of the boundary curve into the frequency domain and extract coefficients in the low-frequency band to stably represent the macroscopic contour of the shape. Zernike moments precisely represent fine differences in texture and shape based on orthonormal polynomials within the unit circle, while ensuring invariance under rotational transformation. Consequently, since these three methods are mutually complementary, the shape vector transformation unit (200) produces vectors through an optional or integrated configuration.
[0075] The absolute orientation shape vector focuses on generating a vector containing absolute orientation information. The absolute orientation vector generation module (240) receives as input an outline sample that has not undergone rotation correction based on the main orientation angle calculated by the azimuth angle calculation module (120), and calculates a direction-sensitive descriptor. At this time, to distinguish it from the vector generated by the rotation-invariant vector generation module (250), the absolute orientation vector includes expressions with inherent directionality, such as a direction histogram or directional Fourier coefficients.
[0076] In addition, for smoothing and noise suppression, outline resampling is performed based on a uniform arc length, and the number of sample points can be matched to the standard schema of the database storage unit (500).
[0077] The output of the shape vector generation step (S200) is tagged with metadata after undergoing quality verification regarding the number of dimensions, normalization method, and stability. The quality verification consists of reviewing correlations with area, perimeter, and number of segments, self-similarity testing, and sensitivity analysis for micro-deformation. As a result, the feature vector generation module ensures stable vector representation even in environments with various shapes and resolutions.
[0079] Surrounding information extraction step (S300)
[0080] FIG. 5 is a diagram for explaining the configuration of surrounding information according to the embodiment of FIG. 1.
[0081] Referring to FIG. 5, surrounding information extraction is performed around the surrounding information extraction unit (300), and is a step of acquiring and vectorizing surrounding information such as roads, buildings, waterfronts, and green spaces around the land. This step is primarily intended to extract surrounding information that is directly adjacent.
[0082] As metadata, the analysis reference date, the type of adjacent land, and the legal classification of adjacent land can be extracted based on each boundary segment. For example, based on Fig. 5, the subject land is divided by polygonal boundaries, which consist of a total of five segments: a, b, c, d, and e. Boundary segment a is adjacent to a road, and metadata such as the type, width, and legal classification of the adjacent road is extracted along with it. This information serves as important data for evaluating the accessibility and potential uses of the subject land. Boundary segment b abuts a green space, and information such as the legal classification, area, and vegetation status of the green space is recorded as metadata. This information is utilized not only for evaluating comfort and landscape but also for analyses related to environmental regulations.
[0083] Boundary segment c is adjacent to a waterfront, and in this case, the type of water body (such as rivers, lakes, or reservoirs), whether it is legally designated as a riparian zone, and water quality management standards are extracted. This riparian information plays a crucial role in analyzing riparian regulations and flood risks, in addition to its scenic value. Boundary segment d abuts land or building sites intended for the same purpose, and metadata such as the land use classification, building status, and whether a district unit plan exists for the adjacent site is stored. This allows for the evaluation of the development potential and suitability of the target land for use. Finally, boundary segment e is adjacent to another road, and information such as road width, traffic volume, and frontage length is extracted. This road proximity information can be utilized as data to evaluate traffic accessibility or visibility.
[0084] The surrounding information extraction unit (300) does not merely identify the existence of adjacent objects, but combines and stores spatial vector data and metadata related to each boundary segment. In particular, based on the analysis reference date, it can record changes such as road expansion, new building construction, and damage to green spaces, which can be usefully utilized for land valuation or analysis of future development possibilities. Accordingly, the surrounding information extraction according to the embodiment of FIG. 5 is characterized by being implemented as a composite data structure combining spatial information precisely analyzed by boundary unit and time-series metadata. The surrounding information vector is combined with shape vectors and boundary attribute vectors in the embedding generation unit (400) to increase the precision of the similarity evaluation.
[0086] Composite embedding vector generation step (S400)
[0087] The embedding generation unit (400) can generate a single composite embedding vector by combining a shape feature vector, a boundary attribute vector, and a surrounding information vector. The combination method can support multiple strategies, such as simple combination, weighted combination, normalization followed by projection, and dimensionality reduction followed by combination.
[0088] Each of the above vectors can be standardized so that the mean is 0 and the variance is 1, and robust scaling can be applied if there are outliers in the data. Then, weights can be assigned to the dimensions of each vector, and said weights can be set to be consistent with the weights (w_s, w_b, w_e) applied in the similarity search unit (600).
[0089] The embedding generation unit (400) can perform dimensionality reduction as needed. As a dimensionality reduction technique, an approximation preservation method based on random projection or a reduction method based on principal component analysis (PCA) may be used. The reduced embedding vector can ensure the stability of the cosine similarity calculation by normalizing the L2 norm to 1. In addition, some components of the vector may be maintained as sparse representations, which can contribute to increasing the computational efficiency of the approximate nearest neighbor search module (620).
[0090] The above combination process may include some cross features that weakly reflect the correlation term between shape information and environment information. This allows for adjustment so that even if the shapes are identical, they can be evaluated with a higher degree of similarity if they have similar environmental conditions.
[0091] The generated composite embedding vector can be stored in a composite embedding storage module (540). When storing, not only the vector but also the identifier of the land, the analysis time, the version, and the pipeline setting values can be recorded together. The version management function enables comparison with past results when the model is changed or the feature definition is changed, which is advantageous for ensuring reproducibility and verifiability.
[0092] The above composite embedding vector can finally be used as an input to the approximate nearest search module (620) and the weight-based re-ranking module (630) in the similarity search step (S600). The similarity value S calculated at this time can be defined as shown in the following mathematical formula 1.
[0093]
[0094] Here, is the shape feature vector of the land to be compared, is a boundary attribute vector, represents the surrounding information vectors, respectively. Also, are weights corresponding to shape, boundary, and environment information, respectively, and is the difference in the main orientation angle between the two plots of land It is a function that corrects similarity according to
[0095] Therefore, mathematical formula 1 defines a similarity calculation formula that comprehensively reflects shape, boundary attributes, and surrounding information, and each weight is applied consistently with the value set in the embedding generation unit (400), thereby ensuring consistency of the evaluation criteria.
[0097] Database storage step (S500)
[0098] Referring again to FIGS. 1 and 2, database storage is performed through a database storage unit (500). The data schema consists of land basic metadata, a shape category storage module (510), an azimuth bucket storage module (520), an effective area calculation and storage module (530), and a composite embedding storage module (540). The land basic metadata includes an identifier, coordinate boundaries, area, perimeter, center point, and analysis reference date. The shape category storage module (510) stores classification values such as rectangle, triangle, and composite shape, and the azimuth bucket storage module (520) records bucket values that divide the absolute azimuth into regular intervals.
[0099] The composite embedding storage module (540) records fixed-length vectors in a column-based storage and synchronizes with selected index structures such as HNSW, IVF, and PQ for linkage with the ANN search module. To ensure efficiency in large-scale data environments, tokenized categories, bucket partitioning, and time-based partitioning are configured together. Additionally, a cache layer is provided to store embeddings of recently viewed land and similar candidates for short-term storage. To ensure data integrity, pipeline versions and feature definition versions are stored together, and parameters of each stage are stored in a log for reproducible analysis.
[0100] The database storage unit (500) also provides access control and auditing functions. Permissions for vector access, metadata viewing, and result downloading are separated according to user roles, and all search requests and result returns are accumulated as audit indicators. Through this, the database storage unit (500) satisfies operational security along with analytical accuracy.
[0102] Similarity search step (S600)
[0103] Referring again to FIGS. 1 and FIGS. 2, the similarity search step (S600) may consist of detailed steps of candidate filtering, approximate nearest search, and precise re-ranking.
[0104] First, the candidate filtering module (610) can select an initial group of candidates using category values and azimuth buckets to reduce unnecessary comparison operations in the entire database. For example, primary filtering can be performed based on the land use classification (residential, commercial, industrial, etc.) or cadastral classification of the target land, and land with significantly different main azimuths can be excluded to increase search efficiency. This step is a preprocessing step that reduces the search space and has the effect of reducing the amount of computation in the subsequent approximate search step.
[0105] Next, the approximate nearest neighbor search module (620) can perform an Approximate Nearest Neighbor Search (ANN) using the composite embedding vector generated by the embedding generation unit (400) as input. The algorithm used at this time may be a technique such as HNSW (Hierarchical Navigable Small World), LSH (Locality Sensitive Hashing), or IVF (Inverted File Index), and rapidly reduces the candidate pool based on the L2 normalization results of the composite embedding vector. In this process, the composite characteristics of the shape, boundary, and environment vectors are reflected, thereby ensuring higher precision than simple attribute comparison.
[0106] The subsequent weight-based re-ranking module (630) calculates a final similarity score for the candidate group and calculates the ranking. At this time, the similarity score S can be defined as in Equation 2.
[0107]
[0108] Here, is the shape feature vector of the land to be compared, is a boundary attribute vector, represents the surrounding information vectors, respectively. Also, are weights corresponding to shape, boundary, and environment information, respectively, and is the difference in the main orientation angle between the two plots of land It is a function that corrects similarity according to
[0109] Therefore, Equation 2 enables a more precise similarity evaluation that reflects actual spatial meaning by performing quantitative correction based on azimuth differences in addition to simply summing the similarities of shape, boundary, and environmental factors.
[0110] Finally, the output result of the weight-based re-ranking module (630) can be provided to the user along with a ranking list of top candidates. For example, a list of the top 10 or 20 candidate lands is returned, and each candidate is provided with a similarity score so that the user can compare them intuitively. In addition, if necessary, the results are provided along with metadata (analysis reference date, adjacent land information, surrounding facility information, etc.) to enable comprehensive decision support beyond simple numerical comparison.
[0111] Meanwhile, the weight-based re-ranking module (630) may apply a damping function to quantitatively reflect the azimuth difference between the query land and the candidate land. The damping function is a Gaussian-based function and has the characteristic that the similarity gradually decreases as the azimuth difference increases. At this time, the parameter σ representing the variance of the damping function can be set in the range of 15 degrees or more and 45 degrees or less. This setting has the effect of adjusting to reflect the actual topographical meaning while ensuring that the azimuth difference does not excessively affect the search results.
[0113] Effective Area Calculation Step
[0114] The effective area calculation step (S700) is a process of estimating the ratio of the actual buildable area by combining boundary attributes and environmental factors. The effective area calculation module (700) receives the setback distance, road width, number of lanes, and boundary type recorded in the boundary type label and attribute memo as input, and performs internal buffering on the land outline. Internal buffering differentially reflects setback lines for road frontage, building frontage, and waterfront frontage by applying different setback distance values for each boundary type.
[0115] The surrounding information vector provides distribution information on average road width and road class, which is utilized as a correction factor in calculating setback distances. For example, for main road frontage segments, the setback width is expanded by multiplying the base value of the internal buffer distance by an additional factor based on road width. For waterfront frontages, the accessibility index is used to set a lower limit for the setback distance. If the result of internal buffering leads to the disintegration of polygons or the creation of holes, the area calculation is defined as the sum of the areas of the remaining polygons.
[0116] The effective area ratio is calculated and stored as a ratio of the area to the original area after internal buffering. This step is defined in a generalized manner that quantifies the relative setback through boundary attributes and environmental indicators, without recording the numerical value of the specific regulation itself. The calculated effective area ratio is recorded in the effective area calculation and storage module (530) of the database storage unit (500) and is used as a constraint for the interpretation of similarity search results and the building layout pattern recommendation module (640).
[0118] Building layout recommendation module (640)
[0119] The building layout pattern recommendation module (640) receives the results of the similarity search unit (600) as input and presents a building layout pattern suitable for the land characteristics. This module extracts common patterns from the top candidates of the final results, and the patterns consist of the direction of the main building, the location of road access, the parking circulation path, and the priority of lighting and views. Pattern extraction is performed by analyzing the boundary attribute vector similarity items between the query land and the candidate land and the key indicators of the environmental factor vector in the form of an intersection.
[0120] The building layout pattern recommendation module (640) provides results in the form of guidelines and parameterized layout templates. For example, if the north-facing road frontage is dominant, it proposes a layout that places the entrance at the northern boundary and secures lighting from the south. If the road width is large, it suggests a strategy that concentrates vehicle access at the front and disperses pedestrian access to the sides or rear. If waterfront accessibility is high, it proposes a view-priority layout and a noise-blocking layout in parallel. All recommendations use internal buffering results as constraints to avoid conflicts with the effective area ratio.
[0121] Recommendation results are provided in an explainable form, and each suggestion item includes an identifier of the candidate case serving as the basis and a partial score summary. The building layout pattern recommendation module (640) also receives custom weights and provides a regeneration function that reflects preferences.
[0123] Although the embodiments have been described above with reference to the limited drawings, those skilled in the art can apply various technical modifications and variations based on the above. For example, appropriate results may be achieved even if the described techniques are performed in a different order than described, and / or if the components of the described system, structure, device, circuit, etc. are combined or assembled in a form different from the described method, or are replaced or substituted by other components or equivalents. Therefore, other implementations, other embodiments, and equivalents to the claims are also included within the scope of the claims set forth below. Explanation of the symbols
[0126] 100 : Land Data Processing Unit 200 : Shape vector transformation unit 300 : Surrounding Information Extraction Unit 400 : Embedding generation unit 500 : Database storage 600 : Similar Land Search Department 700 : Effective Area Calculation Unit
Claims
Claim 1 A boundary polygon extraction step (S110) for extracting a boundary polygon of a target land from GIS data by a boundary polygon extraction module (110); an azimuth calculation step (S120) for calculating the main azimuth of the target land by an azimuth calculation module (120); a boundary segment division step (S130) for dividing the boundary polygon into a plurality of boundary segments by a boundary segment division module (130); a boundary attribute assignment step (S140) for assigning boundary attributes to each of the plurality of boundary segments by a boundary attribute assignment module (140); a shape vector generation step (S200) for generating a shape vector including an absolute azimuth shape vector and / or a rotation-invariant shape vector for the shape of the target land by a shape vector conversion unit (200); a surrounding information extraction step (S300) for extracting surrounding information, which is information of a target adjacent to the target land by a surrounding information extraction unit (300); and an embedding generation unit (400) for the absolute azimuth A method for processing a land shape database, comprising: a composite embedding generation step (S400) for generating a composite embedding vector by combining a vector, the rotation-invariant shape vector and the surrounding information; a database storage step (S500) for storing the composite embedding vector and metadata in a database storage unit (500); and a similarity search step (S600) for evaluating similarity using the stored data. Claim 2 A land shape database processing method according to claim 1, wherein the azimuth calculation step (S120) calculates the main azimuth based on the direction of the main side of the land. Claim 3 A land shape database processing method according to claim 1, wherein the azimuth calculation step (S120) calculates the main azimuth of the land using principal component analysis (PCA). Claim 4 A land shape database processing method according to claim 1, wherein the boundary attribute assigned in the boundary attribute assignment step (S140) includes the length, azimuth angle, boundary type, run sequence of the boundary type, and length ratio of each of the boundary segments. Claim 5 A land shape database processing method according to claim 1, wherein the shape vector generation step (S200) calculates a vector based on a Hu moment, a Fourier descriptor (FD), or a Zernike moment. Claim 6 A method for processing a land shape database according to claim 1, wherein the surrounding information extracted in the surrounding information extraction step (S300) includes information about adjacent roads, information about adjacent land, and information about adjacent facilities. Claim 7 A land shape database processing method according to claim 1, wherein the database storage step (S500) stores together an azimuth bucket, which is a value categorized by dividing the land shape category and azimuth angle for the target land into certain intervals. Claim 8 A land shape database processing method according to claim 1, wherein the composite embedding generation step (S400) generates a composite embedding by combining the shape vector, the vector composed of the boundary attributes, and the vector composed of the surrounding information. Claim 9 A method for processing a land shape database according to claim 1, comprising: a candidate filtering step (S610) for selecting an initial group of candidates using category values and azimuth buckets; an approximate search step (S620) for reducing the group of candidates by performing an approximate nearest neighbor search (ANN) based on a composite embedding vector; and a re-ranking step (S630) for calculating a final ranking by weighted summing the cosine similarity between a shape vector, a boundary attribute vector, and a surrounding information vector. Claim 10 A land shape database processing method according to claim 9, wherein the re-ranking step (S630) is performed by weighted summing the cosine similarity between the shape vector, the vector composed of the boundary attributes, and the vector composed of the surrounding information. Claim 11 A land shape database processing method according to claim 1, further comprising an effective area calculation step (S700) for calculating an effective area ratio by reflecting the separation distance and boundary type of the target land based on the surrounding information. Claim 12 A land shape database processing system comprising: a boundary polygon extraction module (110) for extracting a boundary polygon of a target land from GIS data; an azimuth calculation module (120) for calculating the main azimuth of the target land; a boundary segment division module (130) for dividing the boundary polygon into a plurality of boundary segments; a boundary attribute assignment module (140) for assigning boundary attributes to the boundary segments; a shape vector conversion unit (200) for generating a shape vector including an absolute azimuth shape vector and / or a rotation-invariant shape vector; a surrounding information extraction unit (300) for extracting surrounding information, which is information of a target adjacent to the target land; an embedding generation unit (400) for generating a composite embedding vector; a database storage unit (500) for storing the composite embedding vector and metadata; and a similar land search unit (600) for performing a similarity search by evaluating the similarity with a query land using the composite embedding vector stored in the database storage unit (500). Claim 13 A land shape database processing system according to claim 12, wherein the shape vector transformation unit (200) comprises a Hu moment calculator, a Fourier descriptor calculator, and a Zernike moment calculator. Claim 14 A land shape database processing system according to claim 12, wherein the surrounding information extraction unit (300) extracts information about adjacent roads, information about adjacent land, and information about adjacent facilities. Claim 15 A land shape database processing system according to claim 12, wherein the embedding generation unit (400) generates a composite embedding vector by combining the shape vector, the vector composed of the boundary attributes, and the vector composed of the surrounding information. Claim 16 A land shape database processing system according to claim 12, further comprising an effective area calculation module (700) that calculates an effective area ratio reflecting the separation distance and boundary type of the target land using the surrounding information. Claim 17 delete Claim 18 A land shape database processing system according to claim 12, wherein the similar land search unit (600) applies a weighting function based on the azimuth angle difference during the similarity search. Claim 19 A land shape database processing system according to claim 12, further comprising a recommendation module (640) that recommends a corresponding building layout pattern based on the search results of the similar land search unit (600). Claim 20 A land shape database processing system according to claim 18, wherein the weight-based re-ranking module (630) applies a Gaussian series damping function that causes the similarity to gradually decrease as the azimuth difference (Δθ) between the query land and the candidate land increases, and the parameter σ representing the variance of the damping function is set in the range of 15 degrees or more and 45 degrees or less.
Citation Information
Patent Citations
Land resource management system
CN117114627A
Method of land shape recognition
KR102795814B1
