Multi-source heterogeneous data fusion-based intelligent extraction method for urban styles and features

Abstract

The invention discloses a multi-source heterogeneous data fusion-based urban style and appearance element intelligent extraction method, which comprises the following steps: dividing veins, determining veins influence points, calculating veins influence degrees, screening veins nodes, extracting multi-source heterogeneous urban information, carrying out preprocessing, veins mapping and feature processing to obtain apparent urban style and appearance, and clustering the apparent urban style and appearance. Calculating style and feature differences to determine style and feature uniqueness, taking intersections of different apparent city style and features, calculating association degree weights and style fusion degrees among the apparent city style and features, performing weighted fusion to obtain city style and feature, and extracting overall city style and feature elements of a target city in different time periods according to a style and feature node sequence. And obtaining an urban style and feature element set based on the literature development. According to the method, a scientific, systematic and operable quantitative analysis tool can be provided for cognition, evaluation, protection and inheritance of urban styles and features, and urban styles and features management is effectively promoted from experience dominance to data driving and intelligent decision making.

Description

Technical Field

[0001] This invention relates to the field of urban landscape protection and planning technology, and in particular to an intelligent extraction method for urban landscape elements by fusing multi-source heterogeneous data. Background Technology

[0002] As a comprehensive reflection of a city's physical spatial form and cultural connotation, urban landscape is an important carrier of a city's historical context and contemporary development. Scientifically and accurately identifying and extracting urban landscape elements is of vital importance for preserving urban characteristics, guiding protection planning, and implementing precise management and control. At the same time, the development of remote sensing, geographic information systems, big data, and artificial intelligence technologies also provides new directions for urban landscape research.

[0003] The existing technologies have the following shortcomings: (1) Historical dimension is fragmented: Most methods are based on current data for analysis only, without placing the city in the long river of history for temporal deconstruction, resulting in the loss of context and difficulty in distinguishing the historical layering and contemporary shaping of the landscape; (2) Data silo problem is prominent: The existing solutions have a single data source and cannot effectively integrate multi-source heterogeneous data such as textual historical materials, geographic information, image vision, and real-time sensors, which is insufficient for the completeness and complexity of the landscape; (3) Single analysis dimension: Traditional methods lack systematic correlation analysis and comprehensive evaluation of the landscape in multiple dimensions such as natural base and human activities; (4) Subjectivity and staticity of evaluation: The judgment of the landscape characteristics and value lacks a quantifiable and dynamic objective evaluation model, which is difficult to reflect the dynamic evolution process and internal coordination relationship of the landscape. Therefore, this invention proposes an intelligent extraction method for urban landscape elements by fusing multi-source heterogeneous data. By dividing contextual lines and quantifying contextual influencing events to construct a temporal skeleton, it then integrates multi-source data for cross-modal semantic alignment and feature extraction. Furthermore, it innovatively proposes dual quantitative indicators of landscape uniqueness and contextual integration, ultimately forming a dynamic, structured, and interpretable set of urban landscape elements. This provides a scientific, systematic, and operable quantitative analysis tool for the cognition, evaluation, protection, and inheritance of urban landscape, effectively promoting urban landscape management from experience-driven to data-driven and intelligent decision-making. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent extraction method for urban landscape elements by fusing multi-source heterogeneous data.

[0005] To achieve the above objectives, the present invention is implemented according to the following technical solution: This invention includes the following steps: The city's development history is used to divide the cultural context lines, and the cultural context influence events of the target cities within each cultural context line are extracted to determine the cultural context influence points. The cultural context influence degree of each cultural context influence point is calculated based on the cultural context influence events, and cultural context nodes are screened out. Extract multi-source heterogeneous urban information between each context node, preprocess it, perform context mapping to obtain contextual landscape semantic vector, and process the features of the contextual landscape semantic vector to obtain the apparent urban landscape. The apparent urban landscape of the target city is clustered, the landscape differences between each landscape cluster and neighboring landscape clusters are calculated, the landscape uniqueness is determined, and the corresponding landscape clusters are associated. The intersection of different urban landscapes within the same landscape cluster is obtained. The correlation weight between each urban landscape is calculated based on the intersection result. The cultural context integration degree of the corresponding landscape cluster is determined based on the correlation weight, and weighted integration is performed to obtain the urban cultural context landscape. Urban landscape elements are composed of landscape clusters consisting of uniqueness, integration with cultural context, and urban cultural context. The overall urban landscape elements of the target city at different times are extracted according to the order of cultural context nodes to obtain a set of urban landscape elements based on cultural context development. The cultural context includes ancient cultural context, modern cultural context, and contemporary cultural context; the events that influence the cultural context include dynastic changes, administrative regional adjustments, population policies, urban construction policies, and development plans. The multi-source heterogeneous urban information includes text information, image data, GIS data, and sensor data; The apparent urban landscape includes the climatic environment landscape, architectural landscape, and cultural activity landscape; The overall urban landscape elements consist of the landscape elements of all landscape clusters at the same time period.

[0006] Furthermore, the method for filtering out context nodes includes: Based on the city's development history, the cultural context is divided into ancient, modern, and contemporary cultural contexts. The cultural context-influencing events and corresponding event characteristics of the target cities within each cultural context are extracted. The event characteristics include the level of influence, the level of authority, the duration, the intensity of adjustment, and the number of related projects. Taking the time point of the event that influences the context as the context influence point, the context influence degree of the corresponding context influence point is calculated based on the event characteristics, as expressed in the following expression: ; in To influence the cultural context, , , For feature event weights, To determine the level of impact range, As an authority level, For duration, To adjust the intensity, For the number of related projects, The time decay coefficient, For the current assessment point in time, The point in time when the event occurred; Contextual influence points with a contextual influence value greater than the influence value threshold are selected as contextual nodes on the contextual line.

[0007] Furthermore, the method for obtaining the apparent urban landscape includes: Extract multi-source heterogeneous city information of the target city within the time period between context nodes and perform preprocessing; Constructing a contextual knowledge graph based on historical context information. ,in For entity sets, The entity set includes entities such as buildings, streets, public spaces, climate, topography, living conveniences, and residents; the relationship set includes spatiotemporal relationships, derivative relationships, influence relationships, and representational relationships. The preprocessed multi-source heterogeneous city information is input into a cross-modal encoder for feature mapping, and then concatenated with the embedding vector of the contextual knowledge graph to obtain the contextual features semantic vector; the embedding vector of the contextual knowledge graph is learned through a graph attention network. The semantic vector expression for the cultural context is: ; ; in For the first Group Sample No. Class Data Source The semantic vector of the textual context and style, For encoder functions, The encoder parameters are obtained through multimodal pre-training. For entities The context-aware embedding vector is obtained by aggregating information from neighboring entities through a graph attention network. It is a non-linear activation function. For entities The set of neighboring entities, For attention weights, It is a learnable linear transformation matrix. The initial feature vector of the neighboring entities.

[0008] Furthermore, the method for determining the uniqueness of the landscape includes: The apparent urban landscape is obtained by processing the semantic vector features of the cultural context; the feature processing includes feature extraction, SVM landscape classification, and spatiotemporal matching. Extract the urban base map information corresponding to the time periods between the context nodes, set the administrative region boundaries according to the administrative region divisions corresponding to the time periods between the context nodes, and obtain the administrative boundary vector data at the street / township level; A three-level clustering strategy is adopted to divide the city into landscape clusters of different scales. The specific steps are as follows: (1) Divide administrative units according to administrative boundary vector data; (2) Establish regular grids in administrative units with an area greater than a set threshold, calculate the grid density of the regular grids, reduce the grid size of high-density grid areas and maintain the grid size of low-density areas; the grid density is calculated based on building density, population density and land use mixing degree; (3) Evaluate the morphological similarity of adjacent grids, merge highly similar regular grids, and finally output an optimized grid, and take all the apparent urban landscapes in the optimized grid as landscape clusters; The climatic environment landscape differences between each landscape cluster and its neighboring landscape clusters were calculated using weighted Euclidean distance. The architectural landscape differences between each landscape cluster and its neighboring landscape clusters were calculated using a combination of Jaccard distance and cosine distance. The KL divergence was used to measure the distribution differences between each landscape cluster and its neighboring landscape clusters to determine the differences in cultural activities landscape. The average weighted value of the differences in climatic environment landscape, architectural landscape, and cultural activities landscape was taken as the landscape difference between each landscape cluster and its neighboring landscape clusters. Calculate the mean difference between each landscape cluster and all neighboring landscape clusters to quantify the uniqueness of the landscape cluster. The expression is: ; in for The unique features of the landscape cluster, for The mean difference in landscape between a landscape cluster and all neighboring landscape clusters. This is the magnification factor.

[0009] Furthermore, the method for calculating the correlation weights between various apparent urban landscape features includes: For different urban landscapes within the same landscape cluster, the intersection is taken. Based on the intersection results, the degree of intersection of each urban landscape is determined. The mutual information entropy of each urban landscape is calculated. The correlation weight between each urban landscape is determined by the degree of intersection and the mutual information entropy. The expression is as follows: ; ; ; in To showcase the city's appearance and Association weights For the mutual information entropy of the apparent urban landscape, To represent the degree of overlap in urban landscapes, The intersection sensitivity coefficient. , For the appearance and Feature category set, , For the appearance and Feature keyword set, For the appearance and The main time difference of the intersection features is the change in the intersection features. The time decay constant, Two types of characteristic time series and The correlation coefficient, Two types of features and The shared spatial area The total spatial area where the two types of features appear. For the two types of features, the intersection-union ratio of the grid is involved. , , , As the intersection degree weight, Let be the probability distribution of the degree of intersection. , The marginal probability represents the degree of intersection.

[0010] Furthermore, the method for obtaining the urban cultural context includes: The correlation weight matrix is ​​determined based on the correlation weights among the apparent urban features. The contextual integration degree of the corresponding feature cluster is then calculated from the correlation weight matrix, expressed as follows: ; in To ensure the integration of cultural context into the landscape cluster, The trace of the matrix reflects the overall strength of the weights. For the correlation weight matrix, The Frobenius norm for matrix asymmetry; The attention coefficients between the target urban landscape and the neighboring urban landscapes are calculated using graph attention networks and association weights. Based on these attention coefficients, the neighboring urban landscapes are weighted and aggregated into the target urban landscape to obtain the urban context landscape. Urban landscape elements, which consist of landscape clusters composed of unique features, cultural integration, and urban cultural landscape, are extracted from the overall urban landscape elements of the target city at different times according to the order of cultural context nodes, thus obtaining a set of urban landscape elements based on cultural context development.

[0011] The beneficial effects of this invention are: This invention is an intelligent extraction method for urban landscape elements from multi-source heterogeneous data fusion. Compared with existing technologies, this invention has the following technical advantages: This invention enhances data preprocessing capabilities and model adaptability in the intelligent extraction of urban landscape elements by employing steps such as context node selection, context mapping, landscape clustering, calculation of landscape uniqueness, and calculation of landscape association weights. This improves the efficiency and accuracy of intelligent extraction of urban landscape elements, enabling precise temporal deconstruction and semantic-level characterization of urban landscape features. It overcomes the technical barriers to the fusion of multi-source heterogeneous urban information, scientifically assesses landscape value, and constructs a structured and evolvable set of urban landscape elements. This set can be directly applied to multiple scenarios, including urban landscape characteristic assessment, protection priority determination, simulation of updated planning schemes, and prediction of landscape evolution trends, providing dynamic and continuous data support and decision-making basis for urban planning, construction, and management. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating the steps of the intelligent extraction method for urban landscape elements using multi-source heterogeneous data fusion, as described in this invention. Detailed Implementation

[0013] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0014] The intelligent extraction method for urban landscape elements based on multi-source heterogeneous data fusion of this invention includes the following steps: like Figure 1 As shown, this embodiment includes the following steps: The city's development history is used to divide the cultural context lines, and the cultural context influence events of the target cities within each cultural context line are extracted to determine the cultural context influence points. The cultural context influence degree of each cultural context influence point is calculated based on the cultural context influence events, and cultural context nodes are screened out. Extract multi-source heterogeneous urban information between each context node, preprocess it, perform context mapping to obtain contextual landscape semantic vector, and process the features of the contextual landscape semantic vector to obtain the apparent urban landscape. The apparent urban landscape of the target city is clustered, the landscape differences between each landscape cluster and neighboring landscape clusters are calculated, the landscape uniqueness is determined, and the corresponding landscape clusters are associated. The intersection of different urban landscapes within the same landscape cluster is obtained. The correlation weight between each urban landscape is calculated based on the intersection result. The cultural context integration degree of the corresponding landscape cluster is determined based on the correlation weight, and weighted integration is performed to obtain the urban cultural context landscape. Urban landscape elements are composed of landscape clusters consisting of uniqueness, integration with cultural context, and urban cultural context. The overall urban landscape elements of the target city at different times are extracted according to the order of cultural context nodes to obtain a set of urban landscape elements based on cultural context development. The cultural context includes ancient cultural context, modern cultural context, and contemporary cultural context; the events that influence the cultural context include dynastic changes, administrative regional adjustments, population policies, urban construction policies, and development plans. The multi-source heterogeneous urban information includes text information, image data, GIS data, and sensor data; The apparent urban landscape includes the climatic environment landscape, architectural landscape, and cultural activity landscape; The overall urban landscape elements consist of the landscape elements of all landscape clusters at the same time period.

[0015] In this embodiment, the method for filtering out context nodes includes: Based on the city's development history, the cultural context is divided into ancient, modern, and contemporary cultural contexts. The cultural context-influencing events and corresponding event characteristics of the target cities within each cultural context are extracted. The event characteristics include the level of influence, the level of authority, the duration, the intensity of adjustment, and the number of related projects. Taking the time point of the event that influences the context as the context influence point, the context influence degree of the corresponding context influence point is calculated based on the event characteristics, as expressed in the following expression: ; in To influence the cultural context, , , For feature event weights, To determine the level of impact range, As an authority level, For duration, To adjust the intensity, For the number of related projects, The time decay coefficient, For the current assessment point in time, The point in time when the event occurred; Contextual influence points with a contextual influence value greater than the influence value threshold are selected as contextual nodes on the contextual line; In actual assessments, among the event characteristics, the scope of impact level refers to the administrative and geographical level of the event's effect, assigned values ​​at four levels: national, provincial, municipal, and county (district), corresponding to quantifications of 4 / 3 / 2 / 1; the authority level refers to the level of official documents or decisions upon which the event is based, such as laws, national plans, provincial plans, and municipal documents, corresponding to quantifications of continuous values ​​(0,1); the duration refers to the actual number of years that the relevant policies or projects have been implemented, taking a value greater than 0; the adjustment intensity refers to the degree of change in urban spatial structure and function caused by the event, determined comprehensively based on factors such as the proportion of land use changes and the ratio of demolition to reconstruction area, corresponding to quantifications of continuous values ​​(0,1); and the number of associated projects refers to the number of key construction projects or cultural activities at the municipal level or above directly associated with the event, corresponding to quantifications of integers greater than or equal to 0. Taking the "modern and contemporary cultural context" of City A as an example, this paper selects the completion of the "Overall Construction Plan of City A" in 1953 as the event that influenced the cultural context, and the evaluation time is 2025 to demonstrate the complete process of screening the points of influence of the cultural context. This event, which impacts the historical context, belongs to the modern historical context and is classified as an urban construction policy. The characteristic event weights are 0.35 / 0.3 / 0.35, the time decay coefficient is 0.002, the impact scope level is calculated to be 4, the authority level to be 0.95, the duration to be 20 years (planning guidance period to 1973), the adjustment intensity to be 0.85, and the number of related projects to be 15. The historical context impact degree of this event is calculated to be 5.112, which is greater than the impact degree threshold of 2. The year 1953, corresponding to this event, is taken as the historical context node. The impact degree threshold is randomly selected from the 50th percentile of the impact degree of 500 major historical events in cities across the country from 1950 to 2000.

[0016] In this embodiment, the method for obtaining apparent urban landscape includes: Extract multi-source heterogeneous city information of the target city within the time period between context nodes and perform preprocessing; Constructing a contextual knowledge graph based on historical context information. ,in For entity sets, The entity set includes entities such as buildings, streets, public spaces, climate, topography, living conveniences, and residents; the relationship set includes spatiotemporal relationships, derivative relationships, influence relationships, and representational relationships. The preprocessed multi-source heterogeneous city information is input into a cross-modal encoder for feature mapping, and then concatenated with the embedding vector of the contextual knowledge graph to obtain the contextual features semantic vector; the embedding vector of the contextual knowledge graph is learned through a graph attention network. The semantic vector expression for the cultural context is: ; ; in For the first Group Sample No. Class Data Source The semantic vector of the textual context and style, For encoder functions, The encoder parameters are obtained through multimodal pre-training. For entities The context-aware embedding vector is obtained by aggregating information from neighboring entities through a graph attention network. It is a non-linear activation function. For entities The set of neighboring entities, For attention weights, It is a learnable linear transformation matrix. The initial feature vectors of the neighboring entities; In practical evaluation, taking the application for World Heritage status for the central axis of City A (initiated in 2012) as an example, this demonstrates how to extract semantic vectors of cultural context from multi-source heterogeneous urban information: Textual information was extracted from city gazetteers (such as "A City Planning Gazetteer" and "A Central Axis World Heritage Application Text"), authoritative reports (such as "A Central Axis Protection and Management Plan"), academic literature (papers related to Palace Museum studies and architectural history), and online information (official website of Beijing Municipal Bureau of Cultural Relics and World Heritage Centre bulletins). Preprocessing was performed (structured parsing: using BERT-BiLSTM-CRF model to extract entities and relationships; feature vectorization: using Sentence-BERT to encode text paragraphs into 384-dimensional semantic vectors, and special embedding of professional terms (such as "central axis symmetry" and "ritual order"); spatiotemporal annotation: extracting and standardizing time information (dynasty, year) from the text, and extracting spatial references ("south of Gate B", "north of Mountain C") and mapping them to coordinates) to obtain structured triples (such as [Gate D, built in 1553], [Gate B, functional, National Day parade], [Central Axis, embodies the idea of ​​harmony between man and nature], etc.), semantic feature vectors, and spatiotemporal labels). Image data was extracted from satellite remote sensing (Landsat-8 multispectral imagery, 30m resolution), aerial photography (UAV oblique photography, 5cm resolution), street view images (continuous acquisition from Baidu Street View and Tencent Street View), and historical photographs (digitized historical images from 1900-2020). Preprocessing was performed (global feature extraction: using ResNet-50 to extract 2048-dimensional depth features; local feature extraction: using Mask R-CNN to segment building facades and roof forms; style feature extraction: using Gram matrices to calculate architectural styles; spatial registration: converting pixel coordinates to the WGS84 coordinate system and unifying them to a vertical overhead view and a 45° human-view angle; denoising and color correction of historical photographs; spatiotemporal alignment of multiple temporal images at the same location) to obtain image feature vectors, georeferenced matrices, and time series data. Extract topographic data (SRTM 30m DEM, slope and aspect map), land use (Globeland30 30m resolution classification data), building vectors (OpenStreetMap building outlines), road network data (Beijing road network vectors (expressways, main roads, and branch roads), POI data (Gaode Map 2 million points of interest), and other GIS data. Preprocess the data (spatial index construction: establish an R-tree spatial index to accelerate neighborhood queries, generate 500m×500m regular grids as analysis units; topology calculation: calculate the proportion of building base area within each grid to obtain building density, take the alpha index = (actual number of loops / maximum possible number of loops) as road network connectivity, determine sight corridors based on DEM-based view area analysis; multi-source fusion: unify the coordinate system to CGCS2000, resample raster data to 10m resolution, and perform topology checks and repairs on vector data) to obtain a spatial feature matrix, topology diagram, and sight corridor set. Data from sensors such as weather stations, environmental monitoring stations, pedestrian flow monitoring, traffic flow, and social media (covering meteorology, environment, population, pedestrian flow, and traffic) are acquired and preprocessed (spatiotemporal interpolation: Kriging interpolation converts point data into a 500m grid continuous field; LSTM is used to predict missing values ​​for time series imputation; outlier detection: the 3σ principle is used to remove abnormal meteorological data; DBSCAN clustering is used to detect abnormal pedestrian flow hotspots; multimodal fusion: meteorological, pedestrian flow, and traffic data are aligned to the same spatiotemporal grid, and hourly, daily, and monthly statistical features are calculated) to obtain a spatiotemporal cube (containing three-dimensional data of time, space, and channels) and a time series feature vector. Modern multi-source heterogeneous urban information includes text information, image data, GIS data, and sensor data. Modern multi-source heterogeneous urban information mainly includes text information, image data, and a small amount of GIS data and sensor data. Ancient multi-source heterogeneous urban information mainly includes text information and image data (mainly obtained through text book scanning, geometric correction, and digitization). The preprocessed multi-source heterogeneous urban information is input into the cross-modal encoder for feature mapping, and concatenated with the embedding vector of the context knowledge graph to obtain the contextual features semantic vector; the cross-modal encoder includes a text encoder branch, an image encoder branch, a GIS encoder branch, and a sensor encoder branch, which are obtained through multimodal pre-training (using contrastive learning loss, positive samples are multi-source data from the same spatiotemporal location, and negative samples are data from random spatiotemporal locations).

[0017] In this embodiment, the method for determining the uniqueness of the landscape includes: The apparent urban landscape is obtained by processing the semantic vector features of the cultural context; the feature processing includes feature extraction, SVM landscape classification, and spatiotemporal matching. Extract the urban base map information corresponding to the time periods between the context nodes, set the administrative region boundaries according to the administrative region divisions corresponding to the time periods between the context nodes, and obtain the administrative boundary vector data at the street / township level; A three-level clustering strategy is adopted to divide the city into landscape clusters of different scales. The specific steps are as follows: (1) Divide administrative units according to administrative boundary vector data; (2) Establish regular grids in administrative units with an area greater than a set threshold, calculate the grid density of the regular grids, reduce the grid size of high-density grid areas and maintain the grid size of low-density areas; the grid density is calculated based on building density, population density and land use mixing degree; (3) Evaluate the morphological similarity of adjacent grids, merge highly similar regular grids, and finally output an optimized grid, and take all the apparent urban landscapes in the optimized grid as landscape clusters; The climatic environment landscape differences between each landscape cluster and its neighboring landscape clusters were calculated using weighted Euclidean distance. The architectural landscape differences between each landscape cluster and its neighboring landscape clusters were calculated using a combination of Jaccard distance and cosine distance. The KL divergence was used to measure the distribution differences between each landscape cluster and its neighboring landscape clusters to determine the differences in cultural activities landscape. The average weighted value of the differences in climatic environment landscape, architectural landscape, and cultural activities landscape was taken as the landscape difference between each landscape cluster and its neighboring landscape clusters. Calculate the mean difference between each landscape cluster and all neighboring landscape clusters to quantify the uniqueness of the landscape cluster. The expression is: ; in for The unique features of the landscape cluster, for The mean difference in landscape between a landscape cluster and all neighboring landscape clusters. This is the magnification factor; In actual assessment, when extracting urban basic map information corresponding to the time periods between the cultural context nodes, the urban basic map is determined according to the development history. Modern urban basic maps are mainly based on satellite maps and remote sensing images, while modern urban basic maps are mainly based on paper historical maps (which need to be digitized, including scanning, geometric correction, and digitization). Ancient urban basic maps are also mainly based on paper historical maps (mostly hand-drawn, with lower accuracy). When using a three-level clustering strategy to divide the city into landscape clusters of different scales, administrative boundary data within the target context time period is obtained. Administrative units are used as initial clustering units (taking City A as an example, including 16 district-level administrative units such as Dongcheng District, Xicheng District, and Chaoyang District). Regular grids (1km*1km) are established within large administrative units. The average values ​​of building density (building base area / grid area), population density (road length / grid area), and land use mix (green space area / grid area) of each regular grid are calculated as the grid density of the regular grid. The regular grid size of high-density grid areas (mainly the urban center) is reduced to 750m*750m, while the grid size of low-density areas (mainly the suburbs) is maintained. The morphological feature vectors (distribution of building age, building height, and architectural style) of each regular grid are extracted. The morphological similarity of adjacent regular grids is calculated. Adjacent regular grids with a morphological similarity greater than 0.85 are merged into optimized grids. Finally, all apparent urban landscapes within the optimized grids are taken as landscape clusters. Each landscape cluster includes cluster information (cluster number, administrative unit to which it belongs, spatial range, area statistics, core features (dominant function, building age, population density level)), and the landscape differences between each landscape cluster and neighboring landscape clusters are calculated based on the clustering results. The natural environment features include topographic features (average elevation, slope, and aspect distribution), water features (river length, lake area, and shoreline length), green space features (green space ratio, vegetation index / NDVI, and tree coverage), and climatic microenvironment (heat island intensity and ventilation corridor index). Architectural features include the distribution of building age (the proportion of building area divided by dynasty / period), the distribution of building height (the proportion of low-rise (1-3 floors), multi-story (4-7 floors), and high-rise (8 floors and above), the distribution of architectural style (style categories and proportions, decorative elements, and color matching), building density (building footprint ratio and floor area ratio), and spatial texture (street network density, block scale, and consistency of building orientation). The cultural and activity features include population structure (age distribution, household registration ratio, and education level distribution), activity characteristics (daytime population density, nighttime population density, and commuting ratio), socio-cultural features (dialect usage rate, preservation of traditional customs, and community organization activity), and economic vitality (per capita income level, employment ratio in the tertiary industry, and density of commercial facilities). Based on the comprehensive difference between each landscape cluster and all its neighboring clusters, the landscape uniqueness index of the cluster is calculated. The average difference is converted into a uniqueness index using the logarithmic normalization method and associated with the corresponding landscape cluster.

[0018] In this embodiment, the method for calculating the correlation weights between various apparent urban landscapes includes: For different urban landscapes within the same landscape cluster, the intersection is taken. Based on the intersection results, the degree of intersection of each urban landscape is determined. The mutual information entropy of each urban landscape is calculated. The correlation weight between each urban landscape is determined by the degree of intersection and the mutual information entropy. The expression is as follows: ; ; ; in To showcase the city's appearance and Association weights For the mutual information entropy of the apparent urban landscape, To represent the degree of overlap in urban landscapes, The intersection sensitivity coefficient. , For the appearance and Feature category set, , For the appearance and Feature keyword set, For the appearance and The main time difference of the intersection features is the change in the intersection features. The time decay constant, Two types of characteristic time series and The correlation coefficient, Two types of features and The shared spatial area The total spatial area where the two types of features appear. For the two types of features, the intersection-union ratio of the grid is involved. , , , As the intersection degree weight, Let be the probability distribution of the degree of intersection. , The marginal probability of the degree of intersection; In actual assessments, when taking the intersection of different urban landscapes within the same landscape cluster, the following are considered to indicate an intersection: the two landscape features involve the same keyword (intersection of nature and architecture / e.g., "sloping roof" (architecture) and "drainage needs" (climate); intersection of architecture and people / e.g., "arcade" (architecture) and "commercial activities" (people); intersection of people and nature / e.g., "morning exercise habits" (people) and "parks and green spaces" (nature)); the same spatial unit (based on a 500m×500m grid); the same time segment (divided by day / week / month / season); or the same cultural concept (establishing semantic connections through a contextual knowledge graph, e.g., "courtyard house" (architecture) associated with "family concept" (people) and "introverted layout" (adaptation to northern climate)). When calculating the degree of overlap among various apparent urban features, the number of overlap categories is used as the basis. For example, natural landscapes can be categorized into three types: {climate, topography, ecology}, while architectural styles can be categorized into three types: {form, space, style}. If the two overlap in the "climate-form" and "ecology-style" categories, then... ; based on the number of intersection features For example, if the keywords for natural landscape are {rainfall, slope, green space}, and the keywords for architectural landscape are {sloping roof, steps, courtyard}, and their intersection is {rainfall - sloping roof, slope - steps}, then... ; with time consistency For example, if the architectural style changes (1980-2000) and population structure changes (1985-2005) of a certain architectural style cluster have a time difference of 5 years and a correlation coefficient of 0.8, then... ; using spatial similarity terms For example, if the ratio of the area of ​​space where the two types of features coexist to the total area is 0.55, and the intersection-union ratio of the grid is 0.65, then... Therefore, the correlation weights of each apparent urban landscape within each landscape cluster are calculated.

[0019] In this embodiment, the method for obtaining the urban cultural context includes: The correlation weight matrix is ​​determined based on the correlation weights among the apparent urban features. The contextual integration degree of the corresponding feature cluster is then calculated from the correlation weight matrix, expressed as follows: ; in To ensure the integration of cultural context into the landscape cluster, The trace of the matrix reflects the overall strength of the weights. For the correlation weight matrix, The Frobenius norm for matrix asymmetry; The attention coefficients between the target urban landscape and the neighboring urban landscapes are calculated using graph attention networks and association weights. Based on these attention coefficients, the neighboring urban landscapes are weighted and aggregated into the target urban landscape to obtain the urban context landscape. Urban landscape elements are composed of landscape clusters consisting of uniqueness, integration with cultural context, and urban cultural context. The overall urban landscape elements of the target city at different times are extracted according to the order of cultural context nodes to obtain a set of urban landscape elements based on cultural context development. In practical evaluation, the attention coefficient between the target apparent urban landscape and the apparent urban landscape of its neighbors is calculated using a graph attention network and association weights. The expression is as follows: ; ; in To represent the urban landscape Urban landscape with neighbors Attention coefficient The normalized attention coefficient. To represent the urban landscape eigenvectors, To showcase the city's appearance to neighbors eigenvectors, This is the attention vector; Contextual integration measures the degree of harmony and unity among the three types of landscape within a specific cluster, as expressed in the calculation formula. It is a measure of the strength of the association. It is a measure of bidirectional correlation symmetry (the higher the symmetry, the better the fusion); Urban landscape elements are composed of landscape clusters consisting of uniqueness of appearance, integration with cultural context, and urban cultural context. A target city contains multiple landscape clusters at the same time (between two cultural context nodes) and is arranged according to geographical location. The correlation between urban landscape and cultural context development can be determined by the uniqueness of appearance and integration with cultural context. The set of urban landscape elements of the target city at different times along the cultural context varies (differences in characteristics and geographical arrangement). By analyzing these differences, we can explore the changing trends of urban landscape. By combining the differences in the uniqueness of the landscape and the degree of integration with the cultural context along the same cultural context, we can explore the impact of cultural context changes on urban landscape.

[0020] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent extraction of urban landscape elements from multi-source heterogeneous data fusion, characterized in that, Includes the following steps: S1. Divide the cultural context lines according to the urban development history, extract the cultural context influence events of the target cities within each cultural context line to determine the cultural context influence points, calculate the cultural context influence degree of each cultural context influence point based on the cultural context influence events, and screen out the cultural context nodes. S2. Extract multi-source heterogeneous urban information between each context node, preprocess it, perform context mapping to obtain contextual landscape semantic vector, and process the contextual landscape semantic vector features to obtain apparent urban landscape. S3. Cluster the apparent urban landscape of the target city, calculate the landscape differences between each landscape cluster and neighboring landscape clusters, determine the landscape uniqueness and associate the corresponding landscape clusters; S4. Take the intersection of different urban landscapes within the same landscape cluster, calculate the correlation weight between each urban landscape based on the intersection result, determine the cultural context integration degree of the corresponding landscape cluster based on the correlation weight, and perform weighted fusion to obtain the urban cultural context landscape. S5. Urban landscape elements composed of landscape clusters consisting of uniqueness of landscape, integration with cultural context, and urban cultural context landscape are extracted according to the order of cultural context nodes to obtain the overall urban landscape elements of the target city at different times and obtain a set of urban landscape elements based on cultural context development. The cultural context includes ancient cultural context, modern cultural context, and contemporary cultural context; the events that influence the cultural context include dynastic changes, administrative regional adjustments, population policies, urban construction policies, and development plans. The multi-source heterogeneous urban information includes text information, image data, GIS data, and sensor data; The apparent urban landscape includes climatic and environmental landscape, architectural landscape, and cultural and social landscape. The overall urban landscape elements consist of the landscape elements of all landscape clusters at the same time period.

2. The intelligent extraction method for urban landscape elements based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The method for filtering out context nodes includes: Based on the city's development history, the cultural context is divided into ancient, modern, and contemporary cultural contexts. The cultural context-influencing events and corresponding event characteristics of the target cities within each cultural context are extracted. The event characteristics include the level of influence, the level of authority, the duration, the intensity of adjustment, and the number of related projects. Taking the time point of the event that influences the context as the context influence point, the context influence degree of the corresponding context influence point is calculated based on the event characteristics, as expressed in the following expression: ; in For the degree of influence of cultural context, , , For feature event weights, To determine the level of impact range, As an authority level, For duration, To adjust the intensity, For the number of related projects, The time decay coefficient, For the current assessment point in time, The point in time when the event occurred; Contextual influence points with a contextual influence value greater than the influence value threshold are selected as contextual nodes on the contextual line.

3. The intelligent extraction method for urban landscape elements based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The method for obtaining the apparent urban landscape includes: Extract multi-source heterogeneous city information of the target city within the time period between context nodes and perform preprocessing; Constructing a contextual knowledge graph based on historical context information. ,in For entity sets, The entity set includes entities such as buildings, streets, public spaces, climate, topography, living conveniences, and residents; the relationship set includes spatiotemporal relationships, derivative relationships, influence relationships, and representational relationships. The preprocessed multi-source heterogeneous city information is input into a cross-modal encoder for feature mapping, and then concatenated with the embedding vector of the contextual knowledge graph to obtain the contextual features semantic vector; the embedding vector of the contextual knowledge graph is learned through a graph attention network. The semantic vector expression for the cultural context is: ； ； in For the first Group Sample No. Class Data Source The semantic vector of the textual context and style, For encoder functions, The encoder parameters are obtained through multimodal pre-training. For entities The context-aware embedding vector is obtained by aggregating information from neighboring entities through a graph attention network. It is a non-linear activation function. For entities The set of neighboring entities, For attention weights, It is a learnable linear transformation matrix. The initial feature vectors of the neighboring entities.

4. The intelligent extraction method for urban landscape elements based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The method for determining the uniqueness of the landscape includes: The apparent urban landscape is obtained by processing the semantic vector features of the cultural context; the feature processing includes feature extraction, SVM landscape classification, and spatiotemporal matching. Extract the urban base map information corresponding to the time periods between the context nodes, set the administrative region boundaries according to the administrative region divisions corresponding to the time periods between the context nodes, and obtain the administrative boundary vector data at the street / township level; A three-level clustering strategy is adopted to divide the city into landscape clusters of different scales. The specific steps are as follows: (1) Divide administrative units according to administrative boundary vector data; (2) Establish regular grids in administrative units with an area greater than a set threshold, calculate the grid density of the regular grids, reduce the grid size of high-density grid areas and maintain the grid size of low-density areas; the grid density is calculated based on building density, population density and land use mixing degree; (3) Evaluate the morphological similarity of adjacent grids, merge highly similar regular grids, and finally output an optimized grid, and take all the apparent urban landscapes in the optimized grid as landscape clusters; The climatic environment landscape differences between each landscape cluster and its neighboring landscape clusters were calculated using weighted Euclidean distance. The architectural landscape differences between each landscape cluster and its neighboring landscape clusters were calculated using a combination of Jaccard distance and cosine distance. The KL divergence was used to measure the distribution differences between each landscape cluster and its neighboring landscape clusters to determine the differences in cultural activities landscape. The average weighted value of the differences in climatic environment landscape, architectural landscape, and cultural activities landscape was taken as the landscape difference between each landscape cluster and its neighboring landscape clusters. Calculate the mean difference between each landscape cluster and all neighboring landscape clusters to quantify the uniqueness of the landscape cluster. The expression is: ; in for The unique features of the landscape cluster, for The mean difference in landscape between a landscape cluster and all neighboring landscape clusters. This is the magnification factor.

5. The intelligent extraction method for urban landscape elements based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The method for calculating the correlation weights between various apparent urban landscape features includes: For different urban landscapes within the same landscape cluster, the intersection is taken. Based on the intersection results, the degree of intersection of each urban landscape is determined. The mutual information entropy of each urban landscape is calculated. The correlation weight between each urban landscape is determined by the degree of intersection and the mutual information entropy. The expression is as follows: ; ; ; in To showcase the city's appearance and Association weights, For the mutual information entropy of the apparent urban landscape, To represent the degree of overlap in urban landscapes, The intersection sensitivity coefficient. , For the appearance and Feature category set, , For the appearance and Feature keyword set, For the appearance and The main time difference of the intersection features is the change in the intersection features. The time decay constant, Two types of characteristic time series and The correlation coefficient, Two types of features and The shared spatial area The total spatial area where the two types of features appear. For the two types of features, the intersection-union ratio of the grid is involved. , , , As the intersection degree weight, Let be the probability distribution of the degree of intersection. , The marginal probability represents the degree of intersection.

6. The intelligent extraction method for urban landscape elements based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The method for obtaining the urban cultural context includes: The correlation weight matrix is ​​determined based on the correlation weights among the apparent urban features. The contextual integration degree of the corresponding feature cluster is then calculated from the correlation weight matrix, expressed as follows: ; in To ensure the integration of cultural context into the landscape cluster, The trace of the matrix reflects the overall strength of the weights. For the correlation weight matrix, The Frobenius norm for matrix asymmetry; The attention coefficients between the target urban landscape and the neighboring urban landscapes are calculated using graph attention networks and association weights. Based on these attention coefficients, the neighboring urban landscapes are weighted and aggregated into the target urban landscape to obtain the urban context landscape. Urban landscape elements, which consist of landscape clusters composed of unique features, cultural integration, and urban cultural landscape, are extracted from the overall urban landscape elements of the target city at different times according to the order of cultural context nodes, thus obtaining a set of urban landscape elements based on cultural context development.

Images