A city data correlation analysis method based on big data processing

Abstract

This invention discloses a method for urban data association analysis based on big data processing, comprising the following steps: Step 1: collecting multi-source urban data of the target city; Step 2: standardizing the multi-source urban data to form a regional multimodal sample set; Step 3: inputting the regional multimodal sample set into an improved LLaVA network for multimodal feature modeling to generate a regional multimodal feature vector set; Step 4: performing multidimensional prediction modeling on the regional multimodal feature vector set through an MLP scoring module; Step 5: constructing an urban vitality association model using linear regression to analyze the association weight between urban vitality factors and spatial vitality scores; Step 6: generating spatial configuration suggestions for the target city through feature transformation and DBSCAN clustering; and Step 7: constructing an urban health check report. This invention enhances the semantic modeling capability of urban data association analysis through an improved LLaVA network.

Description

Technical Field

[0001] This invention relates to the fields of urban intelligent planning and urban big data analysis technology, and in particular to a method for urban data correlation analysis based on big data processing. Background Technology

[0002] With the increasing demand for smart city construction and refined management of urban spatial structure, big data-driven analysis methods for urban spatial vitality modeling, public space value assessment, and business configuration optimization have received widespread attention. Existing urban data analysis platforms mainly rely on static geographic information system data or single-modal POI density distribution for heat map assessment and functional zoning, which commonly suffers from the following problems in practical applications: The collected urban spatial images, population and traffic statistics, and social text data exhibit significant differences in data type and modal dimensions. Traditional methods struggle to effectively integrate the semantic association features of multimodal data, resulting in one-sided or spatiotemporally inflexible public space heat analysis results. Most existing methods employ shallow statistical models or simple rule matching for factor correlation analysis, failing to reveal the complex causal mechanisms behind spatial vitality and lacking the ability to model nonlinear relationships between multi-source urban elements. When outputting spatial optimization suggestions, they often rely on expert experience or static configuration templates, lacking a dynamic mapping mechanism with actual regional usage scenarios, social sentiment, and potential population behavior. This leads to spatial configuration suggestions lacking specificity and adaptability, failing to meet the needs of refined governance and intelligent urban design development.

[0003] Therefore, how to provide a method for urban data correlation analysis based on big data processing is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a method for urban data association analysis based on big data processing. This invention constructs an improved LLaVA network to achieve joint modeling of multi-scale images, text, and attribute information of urban areas. It describes in detail the process from multimodal feature extraction, spatial vitality scoring, factor association modeling to the generation of spatial configuration suggestions. It has the advantages of strong data fusion capability, rich analysis dimensions, high accuracy of vitality prediction, and intelligent generation of design guidelines.

[0005] A method for urban data correlation analysis based on big data processing according to an embodiment of the present invention includes the following steps: Step 1: Collect multi-source city data for the target city; Step 2: Standardize the multi-source city data to form a regional multimodal sample set; Step 3: Input the regional multimodal sample set into the improved LLaVA network to perform multimodal feature modeling and generate a regional multimodal feature vector set; the improved LLaVA network includes a multi-scale image coding module, a spatial awareness adaptation module, a language coding module, a regional mask guidance module, a visual-language alignment module, and a gating fusion module; Step 4: Input the regional multimodal feature vector set into the MLP scoring module to generate the spatial vitality score sequence, social popularity level sequence, and spatial use potential vector sequence of the target city; Step 5: Based on the regional multimodal feature vector set and spatial vitality score sequence, a linear regression model is constructed to analyze the correlation weight between urban vitality factors and spatial vitality scores, and an urban vitality correlation matrix is ​​generated. Step 6: Based on the spatial vitality score sequence, the urban element correlation matrix, and the spatial use potential vector sequence, perform feature transformation and DBSCAN clustering to generate spatial configuration suggestions for the target city; Step 7: Generate a city health check report based on the spatial vitality score sequence, social popularity level sequence, spatial use potential vector sequence, urban vitality correlation matrix, and spatial configuration suggestions.

[0006] Optionally, the multi-source city data includes latitude and longitude coordinates of each city's spatial elements, image data, several related social texts, and tag data.

[0007] Optionally, step two specifically includes: The latitude and longitude coordinates of each urban spatial element are uniformly projected and gridded to form a regional spatial vector; The image data of each urban spatial element is scaled and cropped to obtain regional image patches, which are then divided into image patch sequences of a set size. For each urban spatial element, several social texts are segmented and classified by sentiment, and thematic features are extracted to obtain regional text vectors. Embedding and encoding are performed on the label data of each urban spatial element to obtain the regional label embedding vector of each urban spatial element; The regional spatial vector, image patch sequence, regional text vector, and regional label embedding vector of each urban spatial element are structurally combined to obtain regional multimodal samples. All regional multimodal samples are then combined to form a regional multimodal sample set.

[0008] Optionally, step three specifically includes: The multi-scale image coding module includes a shallow coding path, a medium coding path, and a deep coding path; In the shallow encoding path, each image patch in the image patch sequence of each region's multimodal samples is mapped to a visual token through a trainable projection weight matrix and projection bias vector to obtain a shallow visual token sequence. The intermediate coding path is composed of stacked L2 Transformer encoders, and the deep coding path is composed of stacked L3 Transformer encoders, wherein L2 is smaller than L3; The intermediate encoding path models the shallow visual token sequence through a multi-head self-attention mechanism and performs nonlinear feature transformation through a feedforward network to generate the intermediate visual token sequence. The deep encoding path performs global attention modeling on the mid-layer visual token sequence to generate a deep visual token sequence; In the spatial perception adaptation module, the shallow visual token sequence, the middle visual token sequence, and the deep visual token sequence are mapped to the set semantic space through linear projection, respectively, to obtain the shallow projected visual token sequence, the middle projected visual token sequence, and the deep projected visual token sequence. The shallow projection visual token sequence, the middle projection visual token sequence, and the deep projection visual token sequence are respectively concatenated with the region spatial vector to obtain the shallow spatial visual token sequence, the middle spatial visual token sequence, and the deep spatial visual token sequence. In the language encoding module, the region text vector and the region label embedding vector are concatenated to obtain a joint feature vector. The joint feature vector is then mapped to a language embedding vector through a trainable linear mapping matrix and a bias vector. The language embedding vector is rearranged to construct a language token sequence. In the region mask guidance module, by constructing region mask vectors, the shallow spatial visual token sequence, the middle spatial visual token sequence, and the deep spatial visual token sequence are mask-weighted respectively to obtain the shallow mask visual token sequence, the middle mask visual token sequence, and the deep mask visual token sequence. In the visual-language alignment module, the shallow mask visual token sequence, the middle mask visual token sequence, and the deep mask visual token sequence are respectively subjected to cross-modal attention operation with the language token sequence to obtain the shallow fusion feature matrix, the middle fusion feature matrix, and the deep fusion feature matrix. In the gated fusion module, the shallow fusion feature matrix, the middle fusion feature matrix, and the deep fusion feature matrix are averaged in the token dimension to obtain the shallow region fusion vector, the middle region fusion vector, and the deep region fusion vector. The shallow region fusion vector, the middle region fusion vector, and the deep region fusion vector are respectively passed through a parameter-shared linear transformation layer and a Softmax activation function to generate shallow gating weights, middle gating weights, and deep gating weights. Based on shallow gating weights, mid-level gating weights, and deep gating weights, the shallow region fusion vector, mid-level region fusion vector, and deep region fusion vector are weighted and fused to obtain the regional multimodal feature vector; the regional multimodal feature vectors of all regional multimodal samples are combined to form a regional multimodal feature vector set.

[0009] Optionally, the construction process of the region mask vector specifically includes: Obtain the grid cells covered by each urban spatial element and the grid cells to which each image patch belongs in the image patch sequence; If the grid cell to which the current image patch belongs falls within the grid cell covered by urban spatial features, then the mask value corresponding to the current image patch is set to 1; otherwise, it is set to 0. Arrange the mask values ​​in order to form a region mask vector.

[0010] Optionally, in the visual-language alignment module, the shallow mask visual token sequence, the mid-layer mask visual token sequence, and the deep mask visual token sequence are respectively subjected to cross-modal attention operations with the language token sequence to obtain the shallow fusion feature matrix, the mid-layer fusion feature matrix, and the deep fusion feature matrix, specifically including: The shallow mask visual token sequence, the middle mask visual token sequence, and the deep mask visual token sequence are mapped to the shallow query matrix, the middle query matrix, and the deep query matrix respectively through three sets of trainable query projection matrices. The language token sequence is mapped to a language key matrix and a language value matrix through a trainable key projection matrix and value projection matrix; The shallow query matrix, the middle query matrix, and the deep query matrix are respectively subjected to matrix dot product operation and normalization with the language key matrix to obtain the shallow attention weight matrix, the middle attention weight matrix, and the deep attention weight matrix. The shallow attention weight matrix, the middle attention weight matrix, and the deep attention weight matrix are multiplied by the language value matrix to obtain the shallow fusion feature matrix, the middle fusion feature matrix, and the deep fusion feature matrix, respectively.

[0011] Optionally, step four specifically includes: The MLP scoring module includes a vitality score branch, a social popularity level branch, and a spatial potential branch. The multimodal feature vector of each region is input into the vitality score branch, and the spatial vitality score is calculated through a trainable linear transformation vector and a bias term. The multimodal feature vector of each region is input into the social popularity level branch. A social popularity probability vector is generated through linear mapping and the Softmax function. The vector index corresponding to the maximum probability value in the social popularity probability vector is selected as the social popularity level. The multimodal feature vector of each region is input into the spatial potential branch and transformed into a spatial use potential vector through linear mapping; It summarizes the spatial vitality score, social popularity level, and spatial use potential vector of all urban spatial elements, and outputs the spatial vitality score sequence, social popularity level sequence, and spatial use potential vector sequence.

[0012] Optionally, step five specifically includes: The urban vitality factors include population density, traffic flow intensity, business density, green space coverage, and public opinion intensity. The semantic features of each type of urban vitality factor are extracted by dividing the multimodal feature vector of each region into subspaces, resulting in population density feature vector, traffic flow feature vector, business density feature vector, green space coverage feature vector and public opinion heat feature vector. These are then mapped to population density feature value, traffic flow feature value, business density feature value, green space coverage feature value and public opinion heat feature value through linear projection. Using spatial vitality score as the monitoring target, linear regression is performed on the feature vector of urban vitality factor of each urban spatial element to obtain urban vitality correlation model; The urban vitality correlation model is solved by fitting linear regression to obtain the correlation weight of each type of urban vitality factor, and the correlation weight of each type of urban vitality factor is used to form the urban vitality correlation vector of the current urban spatial elements; the urban vitality factor weight vectors of all urban spatial elements are arranged into an urban vitality correlation matrix.

[0013] Optionally, step six specifically includes: The spatial vitality score, spatial use potential vector, and urban vitality correlation vector of each urban spatial element are concatenated to form a fused feature vector; The fused feature vectors are transformed nonlinearly through a multilayer perceptron model to obtain regional customer profile vectors. The multilayer perceptron model includes two fully connected layers, each of which consists of a linear mapping and a GELU activation function. Set up a planning template library, input customer profile vectors for all areas of the target city into the DBSCAN clustering model to perform customer profile clustering, and generate spatial configuration suggestions based on the customer profile clustering results.

[0014] Optionally, step seven specifically includes: The layer data includes regional activity heatmaps, social media sentiment distribution maps, and usage potential distribution maps; The text data includes suggestions for optimizing urban spatial structure and suggestions for spatial configuration. The layer data and text data are structured into a city health check report, which supports visualization and access on the platform.

[0015] The beneficial effects of this invention are: (1) This invention constructs a regional multimodal sample set, integrates latitude and longitude coordinate data, image data, social text and tag data of urban spatial elements, and performs unified projection, scale normalization, text semantic extraction and tag embedding encoding, which significantly improves the structured expression capability of urban multi-source heterogeneous data; (2) This invention introduces an improved LLaVA network to construct a multi-scale image coding module, a spatial perception adaptation module, a language coding module, a region mask guidance module, a visual language alignment module, and a gating fusion module. In the image modeling process, shallow, medium, and deep visual tokens are extracted respectively. The spatial semantic spatial association expression ability is improved through spatial vector fusion. The alignment effect between visual tokens and language tokens is enhanced through region mask guidance and multi-scale cross-modal attention mechanism. At the same time, parameter-sharing linear transformation and Softmax activation function are introduced in the gating fusion module to realize dynamic weighting of multi-scale fusion features. Finally, a region multi-modal feature vector that can simultaneously represent spatial structure, semantic context, and social behavior features is generated, which improves the accuracy and semantic integrity of urban spatial vitality modeling. (3) Based on the regional multimodal feature vector set, this invention generates spatial vitality score sequence, social popularity level sequence and spatial use potential vector sequence through the MLP scoring module, and constructs urban vitality correlation matrix by combining linear regression model to reveal the mapping relationship between urban vitality factors and spatial vitality score; on this basis, the DBSCAN clustering algorithm is used to perform customer profile clustering, and spatial configuration suggestions are generated by combining planning template library, and finally outputs urban health check report including regional vitality heat map, social public opinion distribution map, use potential distribution map and spatial structure optimization suggestions. It realizes the intelligent process of multimodal data fusion, semantic modeling, causal analysis to spatial optimization suggestions, and has the advantages of strong model fusion ability, good factor interpretability and high practicality of configuration scheme. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1This is a schematic diagram of a city data correlation analysis method based on big data processing proposed in this invention; Figure 2 This is a flowchart of the improved LLaVA network structure in a city data correlation analysis method based on big data processing proposed in this invention; Figure 3 This is a flowchart of urban vitality association modeling in an urban data association analysis method based on big data processing proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figures 1-3 A method for urban data correlation analysis based on big data processing includes the following steps: Step 1: Collect multi-source city data for the target city; Step 2: Standardize the multi-source city data to form a regional multimodal sample set; Step 3: Input the regional multimodal sample set into the improved LLaVA network to perform multimodal feature modeling and generate a regional multimodal feature vector set; the improved LLaVA network includes a multi-scale image coding module, a spatial awareness adaptation module, a language coding module, a regional mask guidance module, a visual-language alignment module, and a gating fusion module; Step 4: Input the regional multimodal feature vector set into the MLP scoring module to generate the spatial vitality score sequence, social popularity level sequence, and spatial use potential vector sequence of the target city; Step 5: Based on the regional multimodal feature vector set and spatial vitality score sequence, a linear regression model is constructed to analyze the correlation weight between urban vitality factors and spatial vitality scores, and an urban vitality correlation matrix is ​​generated. Step 6: Based on the spatial vitality score sequence, the urban element correlation matrix, and the spatial use potential vector sequence, perform feature transformation and DBSCAN clustering to generate spatial configuration suggestions for the target city; Step 7: Generate a city health check report based on the spatial vitality score sequence, social popularity level sequence, spatial use potential vector sequence, urban vitality correlation matrix, and spatial configuration suggestions.

[0019] In this embodiment, the multi-source city data includes latitude and longitude coordinates of each city spatial element, image data, several related social texts, and tag data.

[0020] The urban spatial elements include buildings, roads, green spaces, points of interest (POIs), public space nodes, and competing projects; the tag data includes functional area tags, POI type tags, spatial attribute tags, and social semantic tags. The functional area labels include residential area labels, commercial area labels, office area labels, mixed-use area labels, educational area labels, and medical area labels; the point of interest (POI) type labels include dining labels, shopping labels, entertainment labels, park labels, educational labels, and medical labels; the spatial attribute labels include building density level labels, openness level labels, street interface type labels, and pedestrian friendliness level labels; the social semantic labels include nightlife labels, landmark labels, trendy street labels, sense of security labels, and environmental quietness labels.

[0021] In this embodiment, step two specifically includes: The latitude and longitude coordinates of each urban spatial element are uniformly projected and gridded: the latitude and longitude coordinates of each urban spatial element are mapped to planar coordinates; in the planar coordinate space, a set of grid cells is constructed based on the set grid resolution, and the planar coordinates of each urban spatial element are mapped to the corresponding grid cell; the planar coordinates within each grid cell are aggregated to form a regional spatial vector. The image data of each urban spatial element is scaled and cropped: the image data of each urban spatial element is resampled according to a set resolution to obtain a normalized image; the normalized image is cropped based on the spatial boundary of the grid cell to obtain regional image patches, and the regional image patches are divided into image patch sequences of a set size; For each urban spatial element, several social texts are processed for word segmentation and sentiment classification, and topic features are extracted: each social text is segmented using the BERT word segmenter to obtain a word sequence, and the word sequence is input into the TextCNN model to obtain sentiment labels; The sentiment labels include positive labels, neutral labels, and negative labels. The sentiment weight coefficient of the positive label is set to 1, the sentiment weight coefficient of the neutral label is set to 0.5, and the sentiment weight coefficient of the negative label is set to 0.3. The sentiment weight coefficients are then normalized to obtain normalized weight coefficients. Each social text is input into the SBERT model to extract topic semantic vectors. Based on normalized weight coefficients, the topic semantic vectors of each urban spatial element are weighted and aggregated to obtain regional text vectors. Embedding encoding is performed on the label class data of each urban spatial element: each discrete label of each urban spatial element is converted into a continuous label vector through a trainable embedding matrix, and all continuous label vectors are concatenated to obtain the regional label embedding vector of each urban spatial element. The regional spatial vector, image patch sequence, regional text vector, and regional label embedding vector of each urban spatial element are structurally combined to obtain regional multimodal samples. All regional multimodal samples are then combined to form a regional multimodal sample set.

[0022] In this embodiment, step three specifically includes: The multi-scale image coding module includes a shallow coding path, a medium coding path, and a deep coding path; In the shallow encoding path, each image patch in the image patch sequence of each region's multimodal samples is mapped to a visual token through a trainable projection weight matrix and projection bias vector to obtain a shallow visual token sequence. The intermediate coding path is composed of stacked L2 Transformer encoders, and the deep coding path is composed of stacked L3 Transformer encoders, wherein L2 is smaller than L3; The intermediate encoding path models the shallow visual token sequence through a multi-head self-attention mechanism and performs nonlinear feature transformation through a feedforward network to generate the intermediate visual token sequence. The deep encoding path performs global attention modeling on the mid-layer visual token sequence to generate a deep visual token sequence; In the spatial perception adaptation module, the shallow visual token sequence, the middle visual token sequence, and the deep visual token sequence are mapped to the set semantic space through linear projection, respectively, to obtain the shallow projected visual token sequence, the middle projected visual token sequence, and the deep projected visual token sequence. The shallow projection visual token sequence, the middle projection visual token sequence, and the deep projection visual token sequence are respectively concatenated with the region spatial vector to obtain the shallow spatial visual token sequence, the middle spatial visual token sequence, and the deep spatial visual token sequence. In the language encoding module, the region text vector and the region label embedding vector are concatenated to obtain a joint feature vector. The joint feature vector is then mapped to a language embedding vector through a trainable linear mapping matrix and a bias vector. The language embedding vector is rearranged to construct a language token sequence. In the region mask guidance module, by constructing region mask vectors, the shallow spatial visual token sequence, the middle spatial visual token sequence, and the deep spatial visual token sequence are mask-weighted respectively to obtain the shallow mask visual token sequence, the middle mask visual token sequence, and the deep mask visual token sequence. In the visual-language alignment module, the shallow mask visual token sequence, the middle mask visual token sequence, and the deep mask visual token sequence are respectively subjected to cross-modal attention operation with the language token sequence to obtain the shallow fusion feature matrix, the middle fusion feature matrix, and the deep fusion feature matrix. In the gated fusion module, the shallow fusion feature matrix, the middle fusion feature matrix, and the deep fusion feature matrix are averaged in the token dimension to obtain the shallow region fusion vector, the middle region fusion vector, and the deep region fusion vector. The shallow region fusion vector, the middle region fusion vector, and the deep region fusion vector are respectively passed through a parameter-shared linear transformation layer and a Softmax activation function to generate shallow gating weights, middle gating weights, and deep gating weights. Based on shallow gating weights, mid-level gating weights, and deep gating weights, the shallow region fusion vector, mid-level region fusion vector, and deep region fusion vector are weighted and fused to obtain the regional multimodal feature vector; the regional multimodal feature vectors of all regional multimodal samples are combined to form a regional multimodal feature vector set.

[0023] In this invention, the improved LLaVA network is structurally enhanced to address the characteristics of large scale differences, strong spatial semantics, and complex modal coupling in urban spatial data. A multi-scale image coding module introduces shallow, mid-level, and deep coding paths to model local texture, regional structure, and global spatial semantics respectively, significantly improving the perception capability of elements at different spatial scales. In the spatial perception adaptation module, multi-layer visual tokens and regional spatial vectors are concatenated to give visual features clear geographic semantic constraints, avoiding spatial semantic drift caused by pure visual modeling. A regional mask guidance module suppresses visual tokens unrelated to the target urban spatial elements, enhancing the ability to focus regional features. In the visual-language alignment module, a cross-modal attention mechanism is used to achieve fine alignment of multi-scale visual information and language tokens, strengthening the consistent expression of image, text, and label semantics. In the gating fusion module, an adaptive weighting mechanism for multi-layer features is introduced to automatically balance the contributions of different scales and multi-modal information to the modeling results. In summary, the improved LLaVA network can maintain multimodal representation capabilities, highlight the structural constraints and semantic consistency of urban space, and effectively improve the accuracy of regional multimodal feature vectors in depicting complex relationships such as urban vitality, social popularity, and spatial potential.

[0024] In this embodiment, the construction process of the region mask vector specifically includes: Obtain the grid cells covered by each urban spatial element and the grid cells to which each image patch belongs in the image patch sequence; If the grid cell to which the current image patch belongs falls within the grid cell covered by urban spatial features, then the mask value corresponding to the current image patch is set to 1; otherwise, it is set to 0. Arrange the mask values ​​in order to form a region mask vector.

[0025] In this embodiment, the visual-language alignment module performs cross-modal attention operations with the language token sequence, including the shallow mask visual token sequence, the mid-layer mask visual token sequence, and the deep mask visual token sequence, to obtain the shallow fusion feature matrix, the mid-layer fusion feature matrix, and the deep fusion feature matrix. Specifically, this includes: The shallow mask visual token sequence, the middle mask visual token sequence, and the deep mask visual token sequence are mapped to the shallow query matrix, the middle query matrix, and the deep query matrix respectively through three sets of trainable query projection matrices. The language token sequence is mapped to a language key matrix and a language value matrix through a trainable key projection matrix and value projection matrix; The shallow query matrix, the middle query matrix, and the deep query matrix are respectively subjected to matrix dot product operation and normalization with the language key matrix to obtain the shallow attention weight matrix, the middle attention weight matrix, and the deep attention weight matrix. The shallow attention weight matrix, the middle attention weight matrix, and the deep attention weight matrix are multiplied by the language value matrix to obtain the shallow fusion feature matrix, the middle fusion feature matrix, and the deep fusion feature matrix, respectively.

[0026] In this embodiment, step four specifically includes: The MLP scoring module includes a vitality score branch, a social popularity level branch, and a spatial potential branch. The multimodal feature vector of each region is input into the vitality score branch, and the spatial vitality score is calculated through a trainable linear transformation vector and a bias term. The multimodal feature vector of each region is input into the social popularity level branch. A social popularity probability vector is generated through linear mapping and the Softmax function. The vector index corresponding to the maximum probability value in the social popularity probability vector is selected as the social popularity level. The multimodal feature vector of each region is input into the spatial potential branch and transformed into a spatial use potential vector through linear mapping; It summarizes the spatial vitality score, social popularity level, and spatial use potential vector of all urban spatial elements, and outputs the spatial vitality score sequence, social popularity level sequence, and spatial use potential vector sequence.

[0027] In this embodiment, step five specifically includes: The urban vitality factors include population density, traffic flow intensity, business density, green space coverage, and public opinion intensity. The semantic features of each type of urban vitality factor are extracted by dividing the multimodal feature vector of each region into subspaces, resulting in population density feature vector, traffic flow feature vector, business density feature vector, green space coverage feature vector and public opinion heat feature vector. These are then mapped to population density feature value, traffic flow feature value, business density feature value, green space coverage feature value and public opinion heat feature value through linear projection. Using spatial vitality score as the monitoring target, linear regression is performed on the feature vector of urban vitality factor of each urban spatial element to obtain urban vitality correlation model; In this invention, the multimodal feature vector of each region is divided into sub-spatial features corresponding to urban vitality factors according to the semantic dimension, including five types of semantic feature sub-vectors: population density, traffic flow intensity, business density, green space coverage, and public opinion heat. A linear projection operation is performed on each type of sub-vector to map it into specific factor numerical representations, resulting in the corresponding five urban vitality factor feature values. Using the spatial vitality score as the supervision target, the five vitality factor feature values ​​are combined to form an urban vitality factor feature vector, which is then input into a linear regression model for fitting and training. The regression weights of each factor are solved by fitting using the least squares method, thereby establishing an urban vitality correlation model that reflects the weight relationship between the urban vitality score and various factors, and realizing a quantitative analysis of the contribution of different factors to regional spatial vitality.

[0028] The urban vitality correlation model is solved by fitting linear regression to obtain the correlation weight of each type of urban vitality factor, and the correlation weight of each type of urban vitality factor is used to form the urban vitality correlation vector of the current urban spatial elements; the urban vitality factor weight vectors of all urban spatial elements are arranged into an urban vitality correlation matrix.

[0029] In this embodiment, step six specifically includes: The spatial vitality score, spatial use potential vector, and urban vitality correlation vector of each urban spatial element are concatenated to form a fused feature vector; The fused feature vectors are transformed nonlinearly through a multilayer perceptron model to obtain regional customer profile vectors. The multilayer perceptron model includes two fully connected layers, each of which consists of a linear mapping and a GELU activation function. Set up a planning template library, input customer profile vectors for all areas of the target city into the DBSCAN clustering model to perform customer profile clustering, and generate spatial configuration suggestions based on the customer profile clustering results.

[0030] For example, in a certain city, a commercial district has a spatial vitality score of 0.82. Its spatial use potential vector reflects its high development potential, and its urban vitality correlation vector shows that the commercial district's vitality is mainly affected by "traffic flow intensity" and "business density." These features are concatenated and input into a multilayer perceptron model to generate a customer profile vector for the commercial district, represented as "nightlife-oriented young people." After performing DBSCAN clustering on the customer profile vectors of all areas in the city, the commercial district is classified into the "nighttime consumption cluster" category. Subsequently, it is matched with the corresponding planning template library to automatically generate spatial configuration suggestions including "setting up 24-hour convenience services, open plaza spaces, and strengthening nighttime lighting and social signage systems," supporting the city's nighttime economic development goals.

[0031] In this embodiment, step seven specifically includes: The layer data includes regional activity heatmaps, social media sentiment distribution maps, and usage potential distribution maps; The text data includes suggestions for optimizing urban spatial structure and suggestions for spatial configuration. The layer data and text data are structured into a city health check report, which supports visualization and access on the platform.

[0032] Example To verify the feasibility of this invention in practice, the method was applied to a core urban planning project in a city. This urban area is an old, mixed-use zone covering approximately 16 square kilometers, including historical districts, open commercial areas, green spaces, and several undeveloped plots. While the area has experienced rapid urban renewal in recent years, it still faces challenges such as insufficient vitality in public spaces, sparse nighttime pedestrian traffic, and a lack of diversity in business types. Traditional assessment methods based on POI density or traffic heat maps are insufficient to accurately reveal spatial potential and business suitability, and the proposed solutions lack specificity and effectiveness, hindering the progress of scientific planning and efficient resource allocation.

[0033] During implementation, multi-source urban data was first collected from the target area through an urban big data platform. This included seven data sources: building outline images, pedestrian heat maps, bus stop distribution, regional POI information, social media geo-check-in images and text, government planning tags, and public comment text, totaling approximately 470,000 data points. After unified projection and grid division, a regional multimodal sample set containing 6,400 grid units was constructed. Using a multi-scale image coding module of an improved LLaVA network, detailed structures, spatial layouts, and environmental atmospheres in street images were extracted from different semantic levels. A regional mask guidance mechanism was used to enhance semantic focus on core functional nodes such as small squares and green space intersections. By combining social text sentiment and functional tags, a spatial state vector was constructed, and the spatial vitality score, social popularity level, and spatial use potential were output in the MLP scoring module. Based on linear regression analysis, a weight vector of urban vitality factors is generated, revealing a strong correlation between population density, green coverage, and vitality. Through DBSCAN clustering, areas with outstanding vitality and potential are classified, and based on the clustering results, a planning template library is matched to automatically generate spatial configuration suggestions, including the setting of nighttime leisure nodes, the supplementation of convenient service points, and the optimization of areas with integrated business formats.

[0034] To compare and verify the performance differences between the present invention and existing methods, the method of the present invention was compared and analyzed with the POI density analysis method and the static pedestrian flow heat map analysis method. Specifically, the POI density analysis method uses the distribution of POIs in urban space as the main data source, statistically analyzes the number and type density of POIs within a unit area, and estimates the activity level of the area based on the number of POIs. The static pedestrian flow heat map analysis method, specifically, uses existing historical pedestrian flow statistics in the city to form a static pedestrian flow distribution map through time aggregation, serving as a basis for measuring the heat of urban areas. The analysis results are shown in Table 1.

[0035] Table 1. Performance Comparison of Different Methods in Spatial Planning Scenarios in Core Urban Areas

[0036] As shown in Table 1, the method of this invention significantly outperforms the POI density analysis method and the static pedestrian flow heat map analysis method in multiple comparative indicators, demonstrating stronger spatial modeling and planning decision support capabilities. In terms of spatial vitality scoring, the mean squared error of the method of this invention is only 0.096, which is 59.7% and 44.8% lower than the POI density analysis method and the static pedestrian flow heat map analysis method, respectively, indicating that the method of this invention can more accurately capture the actual vitality status of a region. In terms of spatial potential identification, the accuracy of the method of this invention reaches 87.3%, which is 26.1 percentage points and 17.8 percentage points higher than the two comparative methods, respectively, significantly enhancing the ability to identify areas with future potential. In terms of social popularity level prediction, the accuracy of the method of this invention is 89.1%, which is 30.7% and 18.9% higher than the POI density analysis method's 58.4% and the static pedestrian flow heat map analysis method's 70.2%, respectively, indicating that the method, after integrating image, text, and tag semantics, can more effectively perceive user behavioral preferences.

[0037] Furthermore, regarding the explanatory power of factor correlation, the R² of the linear regression model in this invention is 0.87, significantly higher than the 0.52 and 0.66 of the comparative methods, indicating that the method of this invention is better able to reveal the true mapping relationship between urban vitality and multiple spatial factors. In practical applications, the spatial configuration suggestions generated by this invention achieved an adoption rate of 83.9% in planning expert reviews, while the adoption rates of the two comparative methods were only 36.7% and 54.1%. In terms of public satisfaction, the configuration schemes generated by the method of this invention achieved a satisfaction rate of 91.5%, higher than the 63.2% and 71.0% of the comparative methods. This invention has achieved a performance leap in spatial semantic modeling, potential identification, decision recommendation, and user perception, possessing stronger engineering feasibility and promotional application value.

[0038] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for urban data correlation analysis based on big data processing, characterized in that, Includes the following steps: Step 1: Collect multi-source city data for the target city; Step 2: Standardize the multi-source city data to form a regional multimodal sample set; Step 3: Input the regional multimodal sample set into the improved LLaVA network to perform multimodal feature modeling and generate a regional multimodal feature vector set; the improved LLaVA network includes a multi-scale image coding module, a spatial awareness adaptation module, a language coding module, a regional mask guidance module, a visual-language alignment module, and a gating fusion module; Step 4: Input the regional multimodal feature vector set into the MLP scoring module to generate the spatial vitality score sequence, social popularity level sequence, and spatial use potential vector sequence of the target city; Step 5: Based on the regional multimodal feature vector set and spatial vitality score sequence, a linear regression model is constructed to analyze the correlation weight between urban vitality factors and spatial vitality scores, and an urban vitality correlation matrix is ​​generated. Step 6: Based on the spatial vitality score sequence, the urban element correlation matrix, and the spatial use potential vector sequence, perform feature transformation and DBSCAN clustering to generate spatial configuration suggestions for the target city; Step 7: Generate a city health check report based on the spatial vitality score sequence, social popularity level sequence, spatial use potential vector sequence, urban vitality correlation matrix, and spatial configuration suggestions.

2. The urban data correlation analysis method based on big data processing according to claim 1, characterized in that, The multi-source city data includes latitude and longitude coordinates, image data, and several related social texts and tag data for each city's spatial elements.

3. The urban data correlation analysis method based on big data processing according to claim 1, characterized in that, Step two specifically includes: The latitude and longitude coordinates of each urban spatial element are uniformly projected and gridded to form a regional spatial vector; The image data of each urban spatial element is scaled and cropped to obtain regional image patches, which are then divided into image patch sequences of a set size. For each urban spatial element, several social texts are segmented and classified by sentiment, and thematic features are extracted to obtain regional text vectors. Embedding and encoding are performed on the label data of each urban spatial element to obtain the regional label embedding vector of each urban spatial element; The regional spatial vector, image patch sequence, regional text vector, and regional label embedding vector of each urban spatial element are structurally combined to obtain regional multimodal samples. All regional multimodal samples are then combined to form a regional multimodal sample set.

4. The urban data correlation analysis method based on big data processing according to claim 1, characterized in that, Step three specifically includes: The multi-scale image coding module includes a shallow coding path, a medium coding path, and a deep coding path; In the shallow encoding path, each image patch in the image patch sequence of each region's multimodal samples is mapped to a visual token through a trainable projection weight matrix and projection bias vector to obtain a shallow visual token sequence. The intermediate coding path is composed of stacked L2 Transformer encoders, and the deep coding path is composed of stacked L3 Transformer encoders, wherein L2 is smaller than L3; The intermediate encoding path models the shallow visual token sequence through a multi-head self-attention mechanism and performs nonlinear feature transformation through a feedforward network to generate the intermediate visual token sequence. The deep encoding path performs global attention modeling on the mid-layer visual token sequence to generate a deep visual token sequence; In the spatial perception adaptation module, the shallow visual token sequence, the middle visual token sequence, and the deep visual token sequence are mapped to the set semantic space through linear projection, respectively, to obtain the shallow projected visual token sequence, the middle projected visual token sequence, and the deep projected visual token sequence. The shallow projection visual token sequence, the middle projection visual token sequence, and the deep projection visual token sequence are respectively concatenated with the region spatial vector to obtain the shallow spatial visual token sequence, the middle spatial visual token sequence, and the deep spatial visual token sequence. In the language encoding module, the region text vector and the region label embedding vector are concatenated to obtain a joint feature vector. The joint feature vector is then mapped to a language embedding vector through a trainable linear mapping matrix and a bias vector. The language embedding vector is rearranged to construct a language token sequence. In the region mask guidance module, by constructing region mask vectors, the shallow spatial visual token sequence, the middle spatial visual token sequence, and the deep spatial visual token sequence are mask-weighted respectively to obtain the shallow mask visual token sequence, the middle mask visual token sequence, and the deep mask visual token sequence. In the visual-language alignment module, the shallow mask visual token sequence, the middle mask visual token sequence, and the deep mask visual token sequence are respectively subjected to cross-modal attention operation with the language token sequence to obtain the shallow fusion feature matrix, the middle fusion feature matrix, and the deep fusion feature matrix. In the gated fusion module, the shallow fusion feature matrix, the middle fusion feature matrix, and the deep fusion feature matrix are averaged in the token dimension to obtain the shallow region fusion vector, the middle region fusion vector, and the deep region fusion vector. The shallow region fusion vector, the middle region fusion vector, and the deep region fusion vector are respectively passed through a parameter-shared linear transformation layer and a Softmax activation function to generate shallow gating weights, middle gating weights, and deep gating weights. Based on shallow gating weights, mid-level gating weights, and deep gating weights, the shallow region fusion vector, mid-level region fusion vector, and deep region fusion vector are weighted and fused to obtain the regional multimodal feature vector; the regional multimodal feature vectors of all regional multimodal samples are combined to form a regional multimodal feature vector set.

5. The urban data correlation analysis method based on big data processing according to claim 4, characterized in that, The construction process of the region mask vector specifically includes: Obtain the grid cells covered by each urban spatial element and the grid cells to which each image patch belongs in the image patch sequence; If the grid cell to which the current image patch belongs falls within the grid cell covered by urban spatial features, then the mask value corresponding to the current image patch is set to 1; otherwise, it is set to 0. Arrange the mask values ​​in order to form a region mask vector.

6. The urban data correlation analysis method based on big data processing according to claim 4, characterized in that, In the visual-language alignment module, the shallow mask visual token sequence, the middle mask visual token sequence, and the deep mask visual token sequence are respectively subjected to cross-modal attention operations with the language token sequence to obtain the shallow fusion feature matrix, the middle fusion feature matrix, and the deep fusion feature matrix, specifically including: The shallow mask visual token sequence, the middle mask visual token sequence, and the deep mask visual token sequence are mapped to the shallow query matrix, the middle query matrix, and the deep query matrix respectively through three sets of trainable query projection matrices. The language token sequence is mapped to a language key matrix and a language value matrix through a trainable key projection matrix and value projection matrix; The shallow query matrix, the middle query matrix, and the deep query matrix are respectively subjected to matrix dot product operation and normalization with the language key matrix to obtain the shallow attention weight matrix, the middle attention weight matrix, and the deep attention weight matrix. The shallow attention weight matrix, the middle attention weight matrix, and the deep attention weight matrix are multiplied by the language value matrix to obtain the shallow fusion feature matrix, the middle fusion feature matrix, and the deep fusion feature matrix, respectively.

7. The urban data correlation analysis method based on big data processing according to claim 1, characterized in that, Step four specifically includes: The MLP scoring module includes a vitality score branch, a social popularity level branch, and a spatial potential branch. The multimodal feature vector of each region is input into the vitality score branch, and the spatial vitality score is calculated through a trainable linear transformation vector and a bias term. The multimodal feature vector of each region is input into the social popularity level branch. A social popularity probability vector is generated through linear mapping and the Softmax function. The vector index corresponding to the maximum probability value in the social popularity probability vector is selected as the social popularity level. The multimodal feature vector of each region is input into the spatial potential branch and transformed into a spatial use potential vector through linear mapping; It summarizes the spatial vitality score, social popularity level, and spatial use potential vector of all urban spatial elements, and outputs the spatial vitality score sequence, social popularity level sequence, and spatial use potential vector sequence.

8. The urban data correlation analysis method based on big data processing according to claim 1, characterized in that, Step five specifically includes: The urban vitality factors include population density, traffic flow intensity, business density, green space coverage, and public opinion intensity. The semantic features of each type of urban vitality factor are extracted by dividing the multimodal feature vector of each region into subspaces, resulting in population density feature vector, traffic flow feature vector, business density feature vector, green space coverage feature vector and public opinion heat feature vector. These are then mapped to population density feature value, traffic flow feature value, business density feature value, green space coverage feature value and public opinion heat feature value through linear projection. Using spatial vitality score as the monitoring target, linear regression is performed on the feature vector of urban vitality factor of each urban spatial element to obtain urban vitality correlation model; The urban vitality correlation model is solved by fitting linear regression to obtain the correlation weight of each type of urban vitality factor, and the correlation weight of each type of urban vitality factor is used to form the urban vitality correlation vector of the current urban spatial elements; the urban vitality factor weight vectors of all urban spatial elements are arranged into an urban vitality correlation matrix.

9. The urban data correlation analysis method based on big data processing according to claim 1, characterized in that, Step six specifically includes: The spatial vitality score, spatial use potential vector, and urban vitality correlation vector of each urban spatial element are concatenated to form a fused feature vector; The fused feature vectors are transformed nonlinearly through a multilayer perceptron model to obtain regional customer profile vectors. The multilayer perceptron model includes two fully connected layers, each of which consists of a linear mapping and a GELU activation function. Set up a planning template library, input customer profile vectors for all areas of the target city into the DBSCAN clustering model to perform customer profile clustering, and generate spatial configuration suggestions based on the customer profile clustering results.

10. A method for urban data correlation analysis based on big data processing according to claim 1, characterized in that, Step seven specifically includes: The layer data includes regional activity heatmaps, social media sentiment distribution maps, and usage potential distribution maps; The text data includes suggestions for optimizing urban spatial structure and suggestions for spatial configuration. The layer data and text data are structured into a city health check report, which supports visualization and access on the platform.

Images