Fine-grained identification of urban functional areas based on coupling of spatio-temporal features and ensemble learning

Abstract

The application discloses a city functional area fine identification method coupling street block space-time characteristics and integrated learning, namely: firstly, a street block is generated by using an OSM road network, as a basic unit of the city functional area; then, a SALT street block space-time characteristic system composed of shape-attribute-position-texture is constructed by using multi-source space-time big data such as building outlines, POIs, Tencent location data and Google images; then, a deep learning self-encoder Autoencoder is introduced to reduce the dimension of the extracted SALT characteristics and eliminate the information redundancy among the characteristics; finally, a city functional area identification model based on Adaboost integrated learning and the SALT street block space-time characteristics is constructed to realize fine classification of the functional area. The city functional area fine identification method provided by the application guarantees the robustness and accuracy of the identification effect from feature extraction of multi-source space-time big data, core semantic mining of the city functional area to integrated application of the machine learning model, and greatly improves the classification precision of the city functional area.

Description

Technical Field

[0001] This invention belongs to the field of geographic information spatiotemporal big data, specifically a method for refined identification of urban functional areas by coupling the spatiotemporal characteristics of blocks and ensemble learning. Background Technology

[0002] With rapid urbanization, urban land use has become increasingly complex and diverse. Detailed urban land use information provides crucial data support for understanding urban structure and assisting government decision-making. Traditional methods of collecting land use information are primarily led by government agencies, involving field visits and questionnaires—a time-consuming, labor-intensive process with limited public accessibility.

[0003] Existing urban land use classification methods primarily rely on remote sensing imagery as the data source, utilizing spectral and textural information of land features to classify land use types. While these methods are effective at identifying natural categories with prominent spectral characteristics, such as forests, lakes, and oceans, they struggle to further differentiate functional uses within built-up areas, failing to meet the demands for refined urban land use identification. The emergence of crowdsourced geographic information (CGI) has introduced massive amounts of data containing human activities and socioeconomic characteristics, greatly enriching the data sources for urban land use identification. Point of interest (POI) is a typical example of CGI data, possessing rich attribute information and suitable for urban functional zoning mapping. Yao et al. used POI to identify urban functional zoning categories in the Pearl River Delta; Dou et al. used POI to identify and visualize land use functions within the five districts of Jinan City; and Silva et al. used POI to create a map of London's urban functional zoning distribution. With the development of CGI, more data with temporal attributes, including user location data such as GPS trajectories and Tencent location data, are gaining public attention. Combining remote sensing imagery with CGI has become a new trend in characterizing urban functional zoning. Hu Zhongwen et al. extracted urban landscape and human activity information by integrating remote sensing imagery and mobile phone positioning data, and clustered them to generate functional zone categories in Shenzhen. Qiu Guoping et al., based on deep learning methods, fused remote sensing imagery and social perception data to reveal the types of urban functional zones in the study area. Zhong Yanfei et al. used POI, OSM road network, and remote sensing imagery to construct a point-line-area semantic object mapping framework, and achieved urban functional zone mapping for six areas: Beijing, the central area of ​​Wuhan, Hanyang District of Wuhan, Hannan District of Wuhan, Macau, and Wan Chai District of Hong Kong.

[0004] The above research indicates that remote sensing and crowdsourced data both have unique advantages in land information extraction, and the integration of multi-source spatiotemporal big data is a new trend in identifying urban functional zones. However, existing research lacks in-depth exploration and comprehensive analysis of the characteristics of urban functional zones. For example, buildings, as the basic units of urban function, can reflect detailed information about urban functional zones to a certain extent through their geometric features; dynamic changes in population distribution reflect the job-residence pattern of urban land use, and exploring the changing characteristics of human activities over different time periods has great potential in analyzing the structure of urban functions.

[0005] Integrating urban functional area features extracted from multi-source spatiotemporal big data inevitably leads to problems such as high data dimensionality and information redundancy. Effective feature dimensionality reduction plays a positive role in revealing hidden core features and improving the efficiency of urban functional area identification. Principal component analysis (PCA), as a classic data dimensionality reduction method, has been widely applied in various data analysis problems. This method performs well when dealing with feature data with strong linear relationships and balanced sample sizes across categories. However, urban functional area features come from different data sources, and there are complex nonlinear relationships between features. When dealing with such high-dimensional data with weak linear relationships, it is necessary to utilize neural network models to learn complex data structures and extract core features.

[0006] Constructing intelligent classification models is another core issue in urban functional zone identification. Existing functional zone classification methods mainly include unsupervised and supervised classification. Among them, K-means, fuzzy c-means, and self-organizing map networks (SOM) are commonly used unsupervised classification methods for urban functional zone division. Xu Jun et al. used the K-means method to cluster urban functional zones into 5 categories, and then used POI labels to infer the category of the clusters. Pei Tao et al. used fuzzy c-means to divide Singapore into 5 types: residential, business, commercial, open space, and others. Wang Hong et al. used SOM to aggregate functional zones into 9 categories: residential, commercial, commercial / social, shopping mall / residential, etc. From the above studies, it can be seen that unsupervised classification methods have significant uncertainty in the number of clusters. Furthermore, due to the lack of a pre-defined classification system, the classification results differ from national standards (such as GB / T21010-2017, GB50137-2011), which is not conducive to serving urban planning. In contrast, supervised classification can establish a relationship model between features and pre-defined urban categories, and its applications are more widespread. Gong Peng et al. used the random forest method to identify functional zones in major Chinese cities according to the Chinese Land Use Standard (GB / T 21010-2017). Ming Dongping et al. used convolutional neural networks to reveal the functional zone categories of major areas in Hangzhou according to urban and rural land use classification and development land planning standards. However, using a single machine learning method for urban functional zone identification has significant limitations in terms of model generalization ability and classification accuracy. Ensemble learning methods, which combine multiple weak classifiers, can effectively compensate for the shortcomings of a single learner.

[0007] This invention proposes a refined identification method for urban functional zones that couples street block spatiotemporal features with ensemble learning. This method uses street blocks as the basic unit of urban functional zones, further strengthening the integration of remote sensing imagery and multi-source geographic information. It proposes a SALT street block spatiotemporal feature system, which extracts urban functional zone features more comprehensively and deeply. A deep learning autoencoder model is introduced to eliminate SALT feature redundancy. Compared to the classic principal component analysis model, this model can handle complex nonlinear structures and is more effective in mining the core semantic information of urban functional zones. An Adaboost-based ensemble learning-based urban functional zone classification model is constructed, which not only improves the model's generalization ability but also enhances the accuracy of functional zone classification. In summary, this method, from multi-source spatiotemporal big data feature extraction and core semantic mining of urban functional zones to the integrated application of machine learning models, ensures the robustness and accuracy of the identification results, significantly improving the classification accuracy of urban functional zones. Summary of the Invention

[0008] The technical problem this invention aims to solve is: addressing the need for refined classification of urban functional areas, this invention proposes a refined identification method for urban functional areas that couples street-level spatiotemporal features with ensemble learning. This method uses street blocks as the basic unit of urban functional areas, deeply mining information from multi-source spatiotemporal big data from four aspects: building shape, POI tag attributes, mobile user location, and high-resolution Google image texture. It constructs a SALT street-level spatiotemporal feature system, introduces a deep learning autoencoder model to eliminate SALT feature redundancy, and uses Adaboost ensemble learning for urban functional area classification. From multi-source spatiotemporal big data feature extraction and core semantic mining of urban functional areas to the integrated application of machine learning models, the robustness and accuracy of the identification results are ensured, significantly improving the classification accuracy of urban functional areas.

[0009] The present invention solves its technical problem by adopting the following technical solution:

[0010] The present invention provides a refined identification method for urban functional areas that couples the spatiotemporal features of street blocks with ensemble learning. Specifically, the method involves: First, preprocessing the OSM road network to generate street block division results within the study area, which are the basic units of urban functional areas; then, extracting the texture and semantic features of each street block unit, including constructing a SALT street block spatiotemporal feature system composed of "shape-attribute-location-texture" using multi-source spatiotemporal big data of building outlines, POIs, location data, and high-resolution imagery; next, using a deep learning autoencoder to reduce the dimensionality of the SALT features to eliminate information redundancy between features; finally, using the dimensionality-reduced SALT features and functional area labels to train an Adaboost ensemble learning model, and using the trained model for urban functional area classification to identify the refined functional categories of each street block.

[0011] The present invention can generate basic units of urban functional areas, referred to as street blocks, using the following method: sequentially perform projection transformation, buffer analysis, centerline extraction, feature conversion and elimination processing on the OSM road network, and then use high-resolution Google imagery as the base map to check the generated surface features to ensure that they match the base map completely, thus obtaining street blocks.

[0012] This invention can construct a SALT street spatiotemporal feature system using the following methods: For shape features, Tianditu images are obtained from a public platform, and buildings in the study area are extracted using a threshold-based image segmentation algorithm. Building shape feature indices are calculated, and the mean, standard deviation, and sum of building shape feature indices within each street are statistically analyzed. For attribute features, POI data are classified according to their labels, and kernel density analysis is performed on each category of POI. The mean kernel density of each type of POI within each street, as well as the POI category corresponding to the maximum kernel density, are statistically analyzed. For location features, Tencent location data of mobile users is segmented into three time attributes: holidays, weekdays, and weekends, based on the time field. The total number of people, the average number of people per hour, and the number of people per unit area under each time attribute within each street are statistically analyzed. Simultaneously, spatiotemporal cubes are created under different time types to conduct emerging spatiotemporal hotspot analysis, and the mode of spatiotemporal hotspots is statistically analyzed to obtain the spatiotemporal patterns of each street for weekdays, holidays, and weekends. For texture features, a gray-level co-occurrence matrix of the first principal component band of the remote sensing image is constructed, and the texture features of the image are extracted. The average value of the texture features within each street is statistically analyzed.

[0013] This invention uses a threshold-based image segmentation algorithm to extract buildings in the study area and calculates the building shape features of each block unit. The process includes:

[0014] (1) Obtain RGB images of Tianditu Level 18 from the open platform, and use mask extraction, projection grid and band extraction to obtain Tianditu single-band images of the study area;

[0015] (2) Determine the gray threshold for extracting buildings by using the gray histogram. If the gray value of a pixel is within the threshold range, the value is set to 1, which means it is a building; otherwise, the value is set to 0.

[0016] (3) The mode filter is used to remove pixels with gray values ​​less than the threshold. Then, the raster to polygon conversion is used to convert the building into a vector format. Finally, the vector building is processed to eliminate the polygon part and simplify the building.

[0017] (4) Overlay the processed buildings with high-resolution Google images to verify the integrity and accuracy of the extracted buildings;

[0018] (5) Calculate the shape characteristics of the buildings, extract the area, perimeter, roundness, number of nodes, rectangularity, aspect ratio, radius shape index and direction of each building, and then use spatial connection to statistically analyze the sum, average and standard deviation of each shape index of each block.

[0019] This invention uses kernel density estimation to extract attribute features of various POIs within the study area. The process includes:

[0020] (1) Based on the POI labels, they are divided into 14 categories: "Government agencies and social organizations, medical services, sports and leisure services, life services, commercial residences, automotive services, science, education and culture services, residential residences, financial and insurance services, transportation facilities services, shopping services, parks and green spaces, corporate and catering services";

[0021] (2) Using the kernel density estimation method, the POI point data in the study area were converted into continuous surfaces to obtain the density distribution of 14 types of POIs, and the mean kernel density of each type of POI in each block was obtained using regional statistics to obtain the mean value of the kernel density of each type of POI in each block; the kernel density calculation formula is as follows:

[0022]

[0023] In the formula, F(x) is the density estimation function at position x; d is the spatial dimension; h represents the bandwidth; N is the number of points whose distance from position x is less than h; and K is the spatial weighting function.

[0024] (3) Export the attribute table to an Excel spreadsheet. Using Excel's pivot table and formula calculation functions, obtain the POI category corresponding to the maximum kernel density of each block.

[0025] This invention uses emerging spatiotemporal hotspot analysis methods to extract the location features of mobile phone users during holidays, weekdays, and weekends. The process includes:

[0026] (1) Tencent location data from June 18 to June 24, 2018, 7:00 to 21:00, with a time resolution of 1 hour and a spatial resolution of 1.5 kilometers, was collected through the API interface provided by Tencent YiChuxing platform. The data was cleaned and deduplicated using Python language. Finally, 11,297,374 records were obtained, each containing four fields: number of people, longitude, latitude and time.

[0027] (2) Based on the time attribute, the processed data is divided into three layers: holidays, weekdays and weekends; the total number of people, the average number of people per hour and the number of people per unit area for each time type in each block are calculated.

[0028] (3) Using longitude as the x-axis, latitude as the y-axis, and time as the z-axis, create spatiotemporal cubes for Tencent Location Data holidays, weekdays, and weekends respectively; then, based on the spatiotemporal cubes, perform emerging spatiotemporal hotspot analysis on the data to identify trends in the clustering of the people field in the spatiotemporal cubes.

[0029] (4) Use spatial connection methods to count the mode of spatiotemporal hotspots in each block during holidays, weekdays and weekends to characterize the location characteristics of residents.

[0030] This invention can create a grayscale co-occurrence matrix of high-resolution Google imagery and extract texture features of impermeable areas in each street block. The process includes:

[0031] (1) Download 18-level RGB Google images from the open platform, and preprocess the images by image stitching, extraction by mask and projection grid.

[0032] (2) Using principal component analysis, the first principal component band of the Google imagery was extracted. Then, a gray-level co-occurrence matrix was created, and eight classic texture features were extracted: mean, variance, homogeneity, contrast, dissimilarity, entropy, second moment, and correlation.

[0033] (3) Intersect the impermeable surface data in the study area with the street block units, and obtain the mean texture feature of the impermeable surface of each street block by region.

[0034] This invention can eliminate feature redundancy using the following method: Using the SALT street spatiotemporal feature system as the original input, a deep learning autoencoder model is constructed to reduce the dimensionality of the original features. The process includes:

[0035] (1) Use the range standardization method to standardize the SALT features and unify the feature values ​​in the range of 0 to 1 to eliminate the data dimensions;

[0036] (2) Based on neural networks, a deep learning autoencoder model consisting of an encoder and a decoder is constructed. The encoder compresses the original feature SALT by progressively reducing the number of neurons layer by layer. The decoder increases the number of neurons layer by layer based on the compressed data to obtain the reconstructed feature SALT'. During training, the weights and biases of each layer of the neural network are adjusted by minimizing the error between the original feature SALT and the reconstructed feature SALT'. The encoder and decoder processes are as follows:

[0037] g(SALT)=W(SALT)+b

[0038] f(SALT')=f(g(SALT))=W T (SALT)+b'

[0039] In the formula, SALT represents the input feature, SALT' represents the reconstructed feature; g(*) and f(*) are multilayer networks; W and b represent the weights and bias matrices of the encoder, respectively; W T b' and b' represent the weights and bias matrices of the decoder, respectively;

[0040] (3) Train the Autoencoder model using the standardized SALT features. Calculate the loss between the original and reconstructed features using the mean squared error (MSE) metric. Update the error using gradient descent. Stop training when the error converges. The expression for the MSE is as follows:

[0041]

[0042] In the formula: Loss(f,g) represents the mean squared error between the original features and the reconstructed features; N represents the number of samples; SALT i and SALT' i Let represent the input features and reconstructed features of the i-th sample, respectively.

[0043] This invention can identify the specific categories of urban functional areas using the following method: A test set and a validation set are constructed based on SALT-reduced features and visually interpreted functional area category labels. An Adaboost ensemble learning model is then trained to classify urban functional areas. The process includes:

[0044] (1) Referring to the "National Urban Land Use and Planning Standard (GB50137-2011)," the functions of the study area are divided into 7 categories: administrative and public services; commercial facilities; residential; industrial, manufacturing and warehousing; mixed commercial and residential facilities; green space and squares; roads, streets and transportation; and then, with the help of online high-definition Google Maps, the categories of 30%-50% of the block plots are visually interpreted.

[0045] (2) Initialize the training sample weights, that is, assign the same weight to each sample;

[0046] (3) The dimensionality-reduced SALT street spatiotemporal feature system is used as the input feature, the category of the functional area sample is used as the label, the number of weak classifier decision trees is set, and the Adaboost model is trained to learn the mapping relationship between SALT features and functional area labels, and the error of each round is output. During the training process, the weight of the sample correctly classified by the weak classifier decision tree in this round will be reduced in the next round of training, and conversely, the weight of the sample misclassified will be increased in the next iteration.

[0047] (4) When all weak classifiers have completed their evaluation, a strong classifier is formed and used to classify all urban functional areas.

[0048] The present invention provides a refined identification method for urban functional areas that combines spatiotemporal features of urban blocks with integrated learning, which is used to optimize traditional urban functional area identification methods and achieve high-precision classification of functional areas.

[0049] Compared with the prior art, the present invention has the following main technical advantages:

[0050] (1) When constructing a feature system for urban functional zones, traditional methods often combine the spectral features of remote sensing images with the attribute features of POIs. Although POI tags can compensate for the semantic deficiencies of remote sensing images, uneven spatial distribution and uncertain data quality may lead to biases in urban function identification. This invention further introduces building shape features that can reflect information within functional zones and location features that reveal the activities of the main urban functional entities. It constructs a SALT spatiotemporal feature system for urban blocks, consisting of "building shape, POI attribute, mobile phone user location, and remote sensing image texture." This system is conducive to strengthening the combination of remote sensing images and crowdsourced geographic information, and more deeply and comprehensively reflecting the categories of urban functional zones.

[0051] (2) When eliminating information redundancy between features, classic dimensionality reduction methods such as principal component analysis (PCA) are only suitable for extracting the core information of linear features. When dealing with nonlinear features of functional areas from different data sources, the dimensionality reduction effect is often poor. This invention uses a deep learning autoencoder model to reduce the dimensionality of features and employs a nonlinear activation function to handle complex nonlinear relationships, which can more effectively reveal the potential data structure of the original features and help improve the efficiency and accuracy of urban functional area identification. When using PCA dimensionality reduction, the dimensionality-reduced features from the Autoencoder model, and the original features to classify urban functional areas, the Autoencoder model shows a significant advantage, with overall accuracy improved by 4.8% and 5.9% compared to PCA and the original features, respectively.

[0052] (3) When predicting the categories of urban functional areas, single machine learning methods have obvious limitations in terms of generalization ability and classification accuracy. This invention uses the Adaboost ensemble learning model composed of multiple weak classifiers to divide urban functional areas, which overcomes the shortcomings of single learners such as strong sensitivity and large error, has stronger generalization ability, and shows superior performance when dealing with high-dimensional, multi-class, and imbalanced data. Attached Figure Description

[0053] Figure 1 This is a flowchart of the method of the present invention.

[0054] Figure 2 This is a map showing the results of a detailed identification of urban functional zones. Detailed Implementation

[0055] The invention will be further described below with reference to application examples and accompanying drawings, but it is not limited to the content described below.

[0056] The present invention provides a refined identification method for urban functional areas that couples the spatiotemporal features of street blocks with ensemble learning. Specifically, the method involves: First, preprocessing the OSM road network to generate street block division results within the study area, which serve as the basic units of urban functional areas. Then, extracting the texture and semantic features of each street block unit, i.e., using multi-source spatiotemporal big data such as building outlines, POIs, Tencent location data, and high-resolution Google imagery to construct a SALT street block spatiotemporal feature system composed of "shape-attribute-location-texture". Next, using a deep learning autoencoder, the SALT features are dimensionality reduced to eliminate information redundancy between features. Finally, the dimensionality-reduced SALT features and functional area labels are used to train an Adaboost ensemble learning model, which is then used for urban functional area classification to identify the refined functional categories of each street block.

[0057] The aforementioned coupled spatiotemporal feature system of urban blocks and the integrated learning-based refined identification method for urban functional areas include the following steps:

[0058] 1. Generate street block units:

[0059] First, the OSM road network undergoes a projection transformation, converting its coordinate system to a projected coordinate system. Based on the width of dual (multi-line) roads, buffer analysis is used to expand the roads, generating buffers of 30-50 meters. Then, using a centerline extraction method, the centerlines of the buffers are extracted, completing the transformation from dual (multi-line) to single-line roads. Subsequently, feature conversion and elimination processes are used to convert road line features into polygon layers, and polygon features with areas smaller than a threshold are merged with adjacent features. The fill color of the generated polygon layers is set to no color. Using a high-resolution Google image as the base map, the generated polygon features are checked to ensure a perfect match with the base map, yielding the street block division results within the study area, serving as the basic units for urban functional zones.

[0060] 2. Constructing a spatiotemporal feature system for SALT neighborhoods:

[0061] The SALT (Spatial-Terminal-Localization) street feature system includes four features: shape, attribute, location, and tag, which are used to characterize the functions of urban functional areas.

[0062] For shape features, Tianditu images were obtained from public platforms. A threshold-based image segmentation algorithm was used to extract buildings in the study area. Building shape feature indicators such as perimeter, area, roundness, and number of nodes were calculated. The mean, standard deviation, and sum of building shape feature indicators in each block were also obtained.

[0063] For attribute characteristics, POI data are divided into 14 categories according to their labels, including government agencies and social organizations, medical services, sports and leisure services, life services, and commercial and residential. Then, kernel density analysis is performed on each category of POI to calculate the mean kernel density of the 14 categories of POI in each block, as well as the POI category corresponding to the maximum kernel density.

[0064] For location features, based on the time field, the Tencent location data of mobile users is divided into three time attributes: holidays, weekdays and weekends. The total number of people, the average number of people per hour and the number of people per unit area under each time attribute in each block are counted. At the same time, spatiotemporal cubes under different time types are created to conduct emerging spatiotemporal hotspot analysis, and the mode of spatiotemporal hotspots is counted to obtain the spatiotemporal patterns of each block for weekdays, holidays and weekends.

[0065] For texture features, a gray-level co-occurrence matrix of the first principal component band of the remote sensing image is constructed, texture features of the image are extracted, and the average value of texture features within each block is calculated.

[0066] In step 2, a threshold-based image segmentation algorithm is used to extract buildings in the study area, and the building shape features of each block unit are calculated. The process mainly includes:

[0067] (1) Obtain RGB images of Tianditu at level 18 from the open platform, and extract Tianditu images of the study area by masking; convert the coordinate system of the data into projected coordinates through the projection grid; and then use the band extraction function to obtain Tianditu single-band images of the study area.

[0068] (2) Determine the gray threshold for extracting buildings by using the gray histogram. If the gray value of a pixel is within the threshold range, the value is set to 1, which means it is a building; otherwise, the value is set to 0.

[0069] (3) The mode filter is used to remove pixels with gray values ​​less than the threshold. Then, the raster to surface conversion is used to convert the building into a vector format. Finally, the vector building is processed to eliminate the surface part and simplify the building.

[0070] (4) Overlay the processed buildings with high-resolution Google images to verify the integrity and accuracy of the extracted buildings.

[0071] (5) Calculate the shape characteristics of the buildings, extract the area, perimeter, roundness, number of nodes, rectangularity, aspect ratio, radius shape index and direction of each building, and then use spatial connection to statistically analyze the sum, average and standard deviation of each shape index of each block.

[0072] In step 2, the kernel density estimation method is used to extract attribute features of 14 types of POIs within the study area. The process mainly includes:

[0073] (1) Based on the POI label, it is divided into 14 categories: "Government agencies and social organizations, medical services, sports and leisure services, life services, commercial residences, automotive services, science, education and culture services, residential residences, financial and insurance services, transportation facilities services, shopping services, parks and green spaces, corporate and catering services".

[0074] (2) Using the kernel density estimation method, the POI point data in the study area were converted into continuous surfaces to obtain the density distribution of 14 types of POIs. Regional statistics were then used to obtain the mean kernel density of each type of POI in each block. The kernel density calculation formula is as follows:

[0075]

[0076] In the formula, F(x) is the density estimation function at position x; d is the spatial dimension; h represents the bandwidth; N is the number of points whose distance from position x is less than h; and K is the spatial weight function.

[0077] (3) Export the attribute table to an Excel spreadsheet. Using Excel's pivot table and formula calculation functions, obtain the POI category corresponding to the maximum kernel density of each block.

[0078] Step 2 uses emerging spatiotemporal hotspot analysis to extract the location features of mobile phone users during holidays, weekdays, and weekends. The process mainly includes:

[0079] (1) Tencent location data for the period from June 18 to June 24, 2018, between 7:00 and 21:00, with a time resolution of 1 hour and a spatial resolution of 1.5 kilometers, was collected through the API interface provided by the Tencent YiChuxing platform. The data was cleaned and deduplicated using Python. A total of 11,297,374 records were obtained, each containing four fields: number of people, longitude, latitude, and time.

[0080] (2) Based on the time attribute, the processed data is segmented into three layers: holidays, weekdays, and weekends. The total number of people, the average number of people per hour, and the number of people per unit area are calculated for each time type in each block.

[0081] (3) Using longitude as the x-axis, latitude as the y-axis, and time as the z-axis, create spatiotemporal cubes for Tencent Location Data for holidays, weekdays, and weekends respectively. Based on the spatiotemporal cubes, perform emerging spatiotemporal hotspot analysis on the data to identify trends in the clustering of the people field in the spatiotemporal cubes.

[0082] (4) Use spatial connection methods to count the mode of spatiotemporal hotspots in each block during holidays, weekdays and weekends to characterize the location characteristics of residents.

[0083] In step 2, a grayscale co-occurrence matrix of high-resolution Google imagery is created to extract texture features of impermeable areas in each block. This process mainly includes:

[0084] (1) Download 18-level RGB Google images from the open platform and preprocess the images by image stitching, extraction by mask and projection raster.

[0085] (2) Using principal component analysis, extract the first principal component band of the remote sensing image. Then create a gray-level co-occurrence matrix and extract eight classic texture features: mean, variance, homogeneity, contrast, dissimilarity, entropy, second moment and correlation.

[0086] (3) Intersect the impermeable surface data in the study area with the street block units, and obtain the mean texture feature of the impermeable surface of each street block by region.

[0087] 3. Eliminate information redundancy in the spatiotemporal characteristic system of SALT blocks:

[0088] Using the SALT street spatiotemporal feature system as the original input, a deep learning autoencoder model is constructed to reduce the dimensionality of the original features and eliminate SALT spatiotemporal features.

[0089] In step 3, a deep learning autoencoder model is constructed for feature dimensionality reduction. This process mainly includes:

[0090] (1) Use the range standardization method to standardize SALT features and unify the feature values ​​within the range of 0 to 1 to eliminate data dimensions.

[0091] (2) Based on neural networks, a deep learning autoencoder model consisting of an encoder and a decoder is constructed. The encoder compresses the original feature SALT by reducing the number of neurons layer by layer, and the decoder increases the number of neurons layer by layer based on the compressed data to obtain the reconstructed feature SALT'. During training, the weights and biases of each layer of the neural network are adjusted by minimizing the error between the original feature SALT and the reconstructed feature SALT'.

[0092] The encoder and decoder processes are as follows:

[0093] g(SALT)=W(SALT)+b (2)

[0094] f(SALT')=f(g(SALT))=W T (SALT)+b' (3)

[0095] Where: SALT represents the input feature, SALT' represents the reconstructed feature; g(*) and f(*) are multilayer networks; W and b represent the weights and bias matrices of the encoder, respectively; W T b' and b' represent the weight and bias matrices of the decoder, respectively.

[0096] (3) Train the Autoencoder model using the standardized SALT features. Calculate the loss between the original and reconstructed features using the mean squared error (MSE) metric. Update the error using gradient descent. Stop training when the error converges. The expression for the MSE is as follows:

[0097]

[0098] In the formula: Loss(f,g) represents the mean squared error between the original features and the reconstructed features; N represents the number of samples; SALT i and SALT' i Let represent the input features and reconstructed features of the i-th sample, respectively.

[0099] 4. Identify the specific categories of urban functional zones:

[0100] Test and validation sets were constructed based on SALT dimensionality-reduced features and visually interpreted functional area category labels. An Adaboost ensemble learning model was then trained for the refined classification of urban functional areas.

[0101] In step 4, the Adaboost ensemble learning model is trained to classify urban functional zones. The process mainly includes:

[0102] (1) Referring to the "National Urban Land Use and Planning Standard (GB50137-2011)," the functions of the study area were divided into 7 categories: administrative and public services; commercial facilities; residential; industrial, manufacturing and warehousing; mixed commercial and residential facilities; green space and squares; roads, streets and transportation. Then, with the help of online high-definition Google Maps, the categories of 30%-50% of the block plots were visually interpreted.

[0103] (2) Initialize the training sample weights, that is, assign the same weight to each sample.

[0104] (3) Using the dimensionality-reduced SALT features as input features and the category of the functional area samples as labels, the number of weak classifier decision trees is set, and the Adaboost model is trained to learn the mapping relationship between SALT features and functional area labels, outputting the error of each round. During training, the weight of samples correctly classified by the weak classifier decision trees in this round will decrease in the next round of training, while the weight of samples misclassified will increase in the next iteration.

[0105] (4) When all weak classifiers have completed their evaluation, a strong classifier is formed and used to classify all urban functional areas.

[0106] The above steps are used to optimize the traditional urban functional area identification process and achieve high-precision classification of functional areas.

[0107] This invention provides a refined identification method for urban functional areas that couples spatiotemporal features of urban blocks with ensemble learning. This method further strengthens the integration of remote sensing imagery and socially perceived big data, enabling a more comprehensive and in-depth mining of the characteristics of urban functional areas. It introduces a deep learning autoencoder model to effectively extract core feature information, and then constructs an Adaboost ensemble learning model for functional area classification. This method, from multi-source spatiotemporal big data feature extraction and core semantic mining of urban functional areas to the integrated application of machine learning models, ensures the robustness and accuracy of the identification results, significantly improving the classification accuracy of urban functional areas.

[0108] Application examples:

[0109] This case study takes the most urbanized area within the Third Ring Road of Nanchang City as the research area. It uses the proposed method of refined identification of urban functional areas by coupling the spatiotemporal characteristics of blocks and integrating learning to distinguish the functional types of the research area. The invention is further explained in conjunction with the accompanying drawings.

[0110] Specific processing steps ( Figure 1 )as follows:

[0111] Step 1 involves sequentially performing projection transformation, buffer analysis, centerline extraction, feature conversion to polygons, and feature elimination on the OSM road network within the study area. Then, using Google imagery as the base map, the generated polygon features are verified to ensure a perfect match with the base map, yielding the street block division results within the study area, which serve as the basic units for identifying urban functional zones. Specifically, this includes:

[0112] (1) Download the OSM road network of the study area from the OpenStreetMap open-source platform and the level 18 Google RGB imagery from the GGGIS platform. Import the OSM road network and Google imagery into ArcGIS Pro 2.5 software, and use the Project tool in Tools to convert the coordinates of the OSM road network and Google imagery into consistent projected coordinates: WGS 1984 UTM Zone 50N.

[0113] (2) Using the Buffer tool in the Tool, take the projected road network as input, generate a 30-meter buffer on both sides of the road, and merge all output features into one feature.

[0114] (3) Use the Polygon To Centerline tool in the Tool to extract the centerline of the buffer zone from the road buffer zone as the input feature to obtain a single-line road.

[0115] (4) Use the Feature to Polygon tool in the Tool to take the single-line road as input, convert the road line layer into a vector polygon layer, add the Area field, and use Calculate Geometry Attributes in the attribute table to generate the area attributes of each polygon feature.

[0116] (5) Using the Eliminate tool in the Tool menu, with the area layer as input, check "Eliminate by boundary" and set the SQL exclusion expression: Area>200, to exclude areas less than 200m². 2 The surface elements are merged and eliminated.

[0117] (6) Using Google Images as the base map, start editing, merge and segment the eliminated surface layers to make the layers completely match the base map, and finally obtain 1944 block units.

[0118] Step 2: Extract the "building shape, POI attribute, mobile user location, and remote sensing image texture" features of each block unit, that is, construct the SALT feature system.

[0119] Step 2.1: Obtain Tianditu imagery from a public platform, extract buildings in the study area using a threshold-based image segmentation algorithm, calculate building shape features using eight indicators, and statistically obtain the mean, standard deviation, and sum of building shapes for each block. Specifically, this includes:

[0120] (1) Download the Level 18 Tianditu RGB image from the GGGIS public platform. Load the first band of the remote sensing image into ArcGIS Pro 2.5. Use the Extract by Mask tool in Tools to extract the Tianditu image within the study area, using the street blocks as a mask layer. Then use the Project tool in Tools to convert the image coordinates to: WGS 1984UTM Zone 50N.

[0121] (2) Using the Reclassify tool in the Tool, with the projected single-band image as input, the gray values ​​between 246 and 252 are reclassified as 1, and other values ​​are reclassified as NoData, thus initially extracting the building layer.

[0122] (3) Using the Majority Filter tool in the Tool, the reclassified image is used as input. The number of pixels directly adjacent to the current pixel in the filter kernel is set to 4, and the mode is selected as the threshold replacement method to obtain the building layer after removing small units such as roads.

[0123] (4) Using the Feature to Polygon tool in the Tool menu, with the buildings after mode filtering as input features, generate a building vector layer. Add an Area field and use the Calculate Geometry Attributes tool in the attribute table to generate the area attributes of each building.

[0124] (5) Using the Feature Class to Feature Class tool in the Tool menu, with vector buildings as input features, add the SQL expression: Area>50 to remove small-area features and select features with areas greater than 50m². 2 Exporting the building.

[0125] (6) Use the Eliminate Polygon Part tool in the Tool menu, take the exported building as input, and ensure that all features in the input layer are selected. Set the elimination condition to an area less than 50m². 2 Uncheck "Remove only included parts" to fill in any holes in the building.

[0126] (7) Use the Simplify Building tool in the Tool to take the building layer after removing the face as input, set the simplification tolerance to 5 Meters, set the minimum area to 10 Square Meters, remove the redundant nodes, and generate the simplified building layer.

[0127] (8) Add Google Images as a reference, turn on editing, check and adjust the building layers, and finally get 85093 buildings.

[0128] (9) Open the building layer attribute table, add the Area and Perimeter fields, and use CalculateGeometry Attributes to generate the area and perimeter of each building. Further add a Roundness field, and use Calculate Field to calculate the roundness index of each building according to the following formula:

[0129]

[0130] In the formula: Area represents the building's area; Perimeter represents the building's perimeter.

[0131] (10) Using the Feature Vertices To Points tool, with building facets as input and ALL selected for the output point creation location, the mapping relationship between building IDs and all vertices is obtained, where the start and end points of each building are counted repeatedly. Then, using the Spatial Join tool, with buildings as the target layer and vertices as the connection features, a count is performed based on the building IDs to obtain the number of nodes after the start and end points of each building are counted repeatedly. Next, a Nodes field is added, and the Calculate Field is used to subtract 1 from the original count of nodes to obtain the number of nodes for each building.

[0132] (11) Using the Minimum Bounding Geometry tool, with the building as input, select CONVEX_HULL as the minimum boundary geometry output type and check MBG_FIELDS. Add geometric attributes to the output feature class. In the output results, the direction of the minimum boundary geometry is the shape of the building. Add Regularity and LW Ratio fields, and use CalculateField to calculate the building's rectangularity and aspect ratio using the following formula:

[0133] Regularity = Area / Area MBG (2)

[0134] LW Ratio = L / W (3)

[0135] In the formula, Area represents the building's area. MBG The area represents the minimum boundary geometry of the building, where L and W represent the length and width of the minimum boundary geometry, respectively.

[0136] (12) Using the Python language, import the Arcpy package and calculate the radius shape index of each building according to the following formula:

[0137]

[0138] In the formula: n is the number of radii radiating outward from the center of the building, and ri is the i-th radius radiating outward from the center point of the building.

[0139] (13) Use the Join Field tool in the Tool to connect the 8 shape indicators to the building layer according to the ID of each building.

[0140] (14) Using the Spatial Join tool in the Tool, with the block as the target element and the results of each shape index as the connecting element, the sum, mean and standard deviation of each index are calculated in turn.

[0141] Step 2.2 involves dividing the POI data into 14 categories, then performing kernel density analysis on each category to extract the mean kernel density of the 14 POI categories within each block, as well as the category of the POI corresponding to the maximum kernel density. Specifically, this includes:

[0142] (1) Based on the POI's label, it is divided into 14 categories: "Government agencies and social organizations, medical services, sports and leisure services, life services, commercial residences, automotive services, science, education and culture services, residential residences, financial and insurance services, transportation facilities services, shopping services, parks and green spaces, corporate and catering services", and the POIs are exported as 14 layers according to the categories.

[0143] (2) Using the Kernel Density tool in the Tool menu, the POI data is used as input features to convert various types of POI point data into continuous surfaces, thus obtaining the density distribution of each type of POI. The kernel density calculation formula is as follows:

[0144]

[0145] In the formula, F(x) is the density estimation function at position x; d is the spatial dimension; h represents the bandwidth; N is the number of points whose distance from position x is less than h; and K is the spatial weight function.

[0146] (3) Using the Zonal Statistics as Table tool in the Tool, calculate the mean kernel density of various POIs based on the ID field of the blocks in the study area. Then export the statistical results to an Excel table using Table to Excel.

[0147] (4) Using the ID of each block as the first column, copy the average POI of each type to the same table according to the corresponding block number, use the MAX function to calculate the maximum value of the kernel density of each block, and then use the IF function to obtain the category corresponding to the maximum kernel density of each block, and then convert the category format to numerical type.

[0148] Step 2.3: Based on the time field, the Tencent location data of mobile users is segmented into three time types: holidays, weekdays, and weekends. The total number of people, the average number of people per hour, and the number of people per unit area are calculated for each block within each of the three time types. Spatiotemporal cubes are then created for each type to conduct emerging spatiotemporal hotspot analysis. The mode of these hotspots is calculated to obtain the spatiotemporal patterns for each block during weekdays, holidays, and weekends. Specifically, this includes:

[0149] (1) Tencent location data for the period from June 18 to June 24, 2018, 7:00 to 21:00, with a time resolution of 1 hour and a spatial resolution of 1.5 kilometers, was collected through the API interface provided by the Tencent YiChuxing platform. The data was cleaned and deduplicated using Python. A total of 11,297,374 records were obtained, each containing four fields: population, longitude, latitude, and time.

[0150] (2) Based on the time attribute, the cleaned data was segmented into three types: holidays (June 18, 2018), weekdays (June 19-22, 2018), and weekends (June 23-24, 2018). Next, the data was loaded into ArcGIS Pro 2.5, and the Project tool in Tools was used to convert the data coordinates to: WGS 1984 UTM Zone 50N.

[0151] (3) Using the Spatial Join tool in the Tool, with the street as the target element, connect the data of holidays, weekdays and weekends in sequence, and calculate the total number of people, the average number of people per hour and the number of people per unit area for each street.

[0152] (4) Use the Create Space Time Cube By Aggregating Points tool in the Tool menu to create a space-time cube for travel data under three time types, with longitude as the x-axis, latitude as the y-axis, and time as the z-axis. Switch the map mode to 3D mode to view the space-time cube results.

[0153] (5) Using the Emerging Hot Spot Analysis tool in the Tool, based on the spatiotemporal cube, the trend in the clustering of the count field in the spatiotemporal cube is identified, and finally 17 clustering results are obtained, including new, continuous, strengthening, persistent, gradually decreasing, scattered, oscillating, historical hot spots and cold spots, and undetected patterns.

[0154] (6) Use the Intersect tool in the Tool menu to intersect the target features and blocks, and then use the Table to Excel tool to export the intersection results to an Excel spreadsheet. Based on the pivot table function, use the block ID as the row and the spatiotemporal pattern category as the column to count the various spatiotemporal patterns in each block. Then use the Max function to get the mode of the spatiotemporal patterns in each block, and then use the IF function to return the spatiotemporal pattern category corresponding to each mode. Finally, match the hotspot patterns of each block and convert them into numerical types.

[0155] Step 2.4: Create a grayscale co-occurrence matrix to extract texture features from remote sensing images of impermeable areas in each block, specifically including:

[0156] (1) Using the Forward PCA Rotation New Statistics and Rotate tool in ENVI 5.3 software, with RGB remote sensing images of the study area at level 18 as input and the number of output bands selected as 1, the first principal component band of the image was extracted.

[0157] (2) Using the Co-occurrence Measures tool in ENVI 5.3 software, with the first principal component band of the image as input, based on the gray-level co-occurrence matrix, eight classic texture features were extracted: mean, variance, homogeneity, contrast, dissimilarity, entropy, second moment and correlation.

[0158] (3) Use the Extract by Mask tool in ArcGIS Pro 2.5 software to extract the impermeable surface in the study area, use Raster to Polygon to convert the impermeable surface into vector format, use Google Map as the base map, open the editing, and adjust the vector layer according to the base map.

[0159] (4) Use the Intersect tool in Tool to intersect the impermeable surface vector data with the block unit to obtain the impermeable surface area containing the block ID field.

[0160] (5) Using the Zonal Statistics as Table tool in the Tool, assign the ID field to the impermeable surface according to the block, calculate the mean value of each texture within the impermeable surface range of each block by region, and export the mean value results to an Excel table.

[0161] Step 2.5: Based on the ID field of the block unit, connect the 65 features extracted above into the same table.

[0162] Step 3: Using the SALT feature system as the original input, construct a deep learning autoencoder model to reduce the dimensionality of the features. Specifically, this includes:

[0163] (1) Standardize the range of the 65-dimensional features in each block to remove the data dimensions.

[0164] (2) Based on neural networks, a deep learning autoencoder model containing an encoder and a decoder is constructed. The encoder model consists of an input layer, a hidden layer, and an output layer. The input layer contains 65 neurons, the hidden layer and the output layer contain 45 neurons and 30 neurons, respectively, with the ReLU activation function. The decoder model is symmetrical to the encoder model. The input layer is a fully connected layer with 30 hidden neurons and the ReLU activation function, while the hidden layer has 45 neurons and the sigmoid activation function. The output is reconstructed as a 65-dimensional variable. The encoder and decoder processes are as follows:

[0165] g(SALT)=W(SALT)+b (6)

[0166] f(SALT')=f(g(SALT))=W T (SALT)+b' (7)

[0167] Where: SALT represents the input feature, SALT' represents the reconstructed feature; g(*) and f(*) are multilayer networks; W and b represent the weights and bias matrices of the encoder, respectively; W T b' and b' represent the weight and bias matrices of the decoder, respectively.

[0168] (3) Divide all block unit features into training and testing samples in a 7:3 ratio, iteratively train the constructed Autoencoder model, use the Adam optimizer and mean squared error to evaluate the loss between the original and reconstructed data, and set the iteration parameters. The expression for mean squared error is as follows:

[0169]

[0170] In the formula: Loss(f,g) represents the mean squared error between the original features and the reconstructed features; N represents the number of samples; SALT i and SALT' i Let represent the input features and reconstructed features of the i-th sample, respectively.

[0171] (4) Input all the street features into the trained model and export the 30-dimensional encoding result of the intermediate layer to obtain the compressed representation of the original data.

[0172] (5) To verify the effectiveness of the Autoencoder dimensionality reduction method, the classification accuracy of urban functional areas was compared with that of the original data, principal component analysis, and the features after Autoencoder dimensionality reduction. The results are shown in Table 1. The results show that the Autoencoder model exhibits a significant advantage in classification, with overall accuracy improved by 4.8% and 5.9% compared to principal component analysis and the original features, respectively.

[0173] Table 1

[0174]

[0175] Step 4: Based on the dimensionality-reduced features and visually interpreted labels, construct test and validation sets, train the ensemble learning Adaboost model, and classify the functions of each block in the research area. Specifically, this includes:

[0176] (1) Referring to the "National Urban Land Use and Planning Standard (GB50137-2011)," the functions of the study area were divided into 7 categories: administrative and public services (A); commercial facilities (B); residential (R); industrial, manufacturing and warehousing (MW); mixed commercial and residential facilities (BR); green space and squares (G); and roads, streets and traffic (S). Then, using online high-definition Google Maps, the categories of 30%-50% of the block plots were visually interpreted.

[0177] (2) Initialize the test sample weights, that is, assign the same weight to each sample.

[0178] (3) Using the compressed data as input features and the sample category as the label, set the number of weak classifiers, train the Adaboost model to learn the mapping relationship between input features and labels, and output the error of each round. During training, the weight of samples correctly classified by the weak classifier in this round will decrease in the next round of training, while the weight of samples misclassified will increase in the next iteration.

[0179] (4) When all weak classifiers have been evaluated, a strong classifier is formed and used to predict all urban functional zones, thus completing the functional zone classification of the study area. Figure 2 ).

[0180] The refined urban functional area identification method provided by this invention, which combines spatiotemporal features of urban blocks with ensemble learning, has the following characteristics: it further strengthens the integration of remote sensing imagery and multi-source geographic information, proposing a SALT spatiotemporal feature system for urban blocks; it introduces a deep learning autoencoder model to effectively extract the core information of SALT features; and it constructs an Adaboost ensemble learning model to achieve high-precision classification of urban functional areas. This method, from multi-source spatiotemporal big data feature extraction and core semantic mining of urban functional areas to the integrated application of machine learning models, ensures the robustness and accuracy of the identification results, significantly improving the classification accuracy of urban functional areas.

Claims

1. A refined identification method for urban functional areas that couples the spatiotemporal features of urban blocks with ensemble learning, characterized by: First, the OSM road network is preprocessed to generate the street block division results within the study area, which are the basic units of urban functional zones. Then, the texture and semantic features of each street block unit are extracted, including the construction of a SALT street block spatiotemporal feature system composed of "shape-attribute-location-texture" using multi-source spatiotemporal big data of building outlines, POIs, location data, and high-resolution imagery. Next, a deep learning autoencoder is used to reduce the dimensionality of the SALT features to eliminate information redundancy between features. Finally, the Adaboost ensemble learning model is trained using the dimensionality-reduced SALT features and functional zone labels. The trained model is then used for urban functional zone classification to identify the refined functional categories of each street block. The basic units of urban functional areas, referred to as street blocks, are generated using the following method: the OSM road network is sequentially subjected to projection transformation, buffer analysis, centerline extraction, feature conversion to polygons and elimination processing. Then, using high-resolution Google imagery as the base map, the generated polygon features are checked to ensure that they match the base map completely, thus obtaining the street blocks. The following methods were used to construct the spatiotemporal feature system of SALT blocks: For shape features, Tianditu imagery was obtained from a public platform, and a threshold-based image segmentation algorithm was used to extract buildings in the study area. Building shape feature indices were calculated, and the mean, standard deviation, and sum of the building shape feature indices within each block were statistically obtained; for For attribute features, POI data is categorized by its labels, and kernel density analysis is performed on each category of POIs to calculate the mean kernel density of each type of POI in each block, as well as the POI category corresponding to the maximum kernel density. For location features, based on the time field, Tencent location data of mobile users is divided into three time attributes: holidays, weekdays, and weekends. The total number of people, the average number of people per hour, and the number of people per unit area under each time attribute in each block are calculated. At the same time, spatiotemporal cubes under different time types are created to conduct emerging spatiotemporal hotspot analysis, and the mode of spatiotemporal hotspots is calculated to obtain the spatiotemporal patterns of each block for weekdays, holidays, and weekends. For texture features, a gray-level co-occurrence matrix of the first principal component band of the remote sensing image is constructed to extract the texture features of the image, and the average value of the texture features in each block is calculated.

2. The method for refined identification of urban functional areas by coupling the spatiotemporal features of blocks and ensemble learning according to claim 1, characterized in that, Buildings in the study area were extracted using a threshold-based image segmentation algorithm, and the building shape features of each block unit were calculated. The process included: (1) Obtain RGB images of Tianditu at level 18 from the open platform, and use mask extraction, projection grid and band extraction to obtain Tianditu single-band images of the study area; (2) Determine the gray threshold for extracting buildings by using the gray histogram. If the gray value of a pixel is within the threshold range, the value is set to 1, which means it is a building; otherwise, the value is set to 0. (3) The mode filter is used to remove pixels with gray values ​​less than the threshold. Then, the raster to polygon conversion is used to convert the building into a vector format. Finally, the vector building is processed to remove the polygon part and simplify the building. (4) Overlay the processed buildings with high-resolution Google images to verify the integrity and accuracy of the extracted buildings; (5) Calculate the shape characteristics of the buildings, extract the area, perimeter, roundness, number of nodes, rectangularity, aspect ratio, radius shape index and direction of each building, and then use spatial connection to statistically analyze the sum, average and standard deviation of each shape index of each block.

3. The method for refined identification of urban functional areas by coupling the spatiotemporal features of blocks and ensemble learning according to claim 1, characterized in that, The kernel density estimation method was used to extract the attribute features of various POIs in the study area. The process included: (1) Based on the POI labels, they are divided into 14 categories: "Government agencies and social organizations, medical services, sports and leisure services, life services, commercial residences, automotive services, science, education and culture services, residential residences, financial and insurance services, transportation facilities services, shopping services, parks and green spaces, corporate and catering services"; (2) Using the kernel density estimation method, the POI point data in the study area were converted into continuous surfaces to obtain the density distribution of 14 types of POIs, and the mean kernel density of each type of POI in each block was obtained using regional statistics to obtain the mean value of the kernel density of each type of POI in each block; the kernel density calculation formula is as follows: ， In the formula, F(x) is the density estimation function at position x; d is the spatial dimension; h represents the bandwidth; N is the number of points whose distance from position x is less than h; and K is the spatial weighting function. (3) Export the attribute table to an Excel spreadsheet, and use Excel's pivot table and formula calculation functions to obtain the POI category corresponding to the maximum kernel density of each block.

4. The method for refined identification of urban functional areas by coupling the spatiotemporal features of blocks and ensemble learning according to claim 1, characterized in that, Emerging spatiotemporal hotspot analysis methods were used to extract the location features of mobile phone users during holidays, weekdays, and weekends. The process included: (1) Tencent location data from 7:00 to 21:00 on June 18-24, 2018, with a time resolution of 1 hour and a spatial resolution of 1.5 kilometers, was collected through the API interface provided by Tencent YiChuxing platform. The data was cleaned and deduplicated using Python language. Finally, 11,297,374 records were obtained, each containing four fields: number of people, longitude, latitude and time. (2) Based on the time attribute, the processed data is divided into three layers: holidays, weekdays and weekends; the total number of people, the average number of people per hour and the number of people per unit area for each time type in each block are calculated. (3) Using longitude as the x-axis, latitude as the y-axis, and time as the z-axis, create spatiotemporal cubes for Tencent Location Data holidays, weekdays, and weekends respectively; then, based on the spatiotemporal cubes, perform emerging spatiotemporal hotspot analysis on the data to identify trends in the clustering of the people field in the spatiotemporal cubes. (4) Use spatial connection methods to count the mode of spatiotemporal hotspots in each block during holidays, weekdays and weekends to characterize the location characteristics of residents.

5. The method for refined identification of urban functional areas by coupling the spatiotemporal features of blocks and ensemble learning according to claim 1, characterized in that, The process of creating a grayscale co-occurrence matrix from high-resolution Google imagery and extracting texture features from impermeable areas in each block includes: (1) Download 18-level RGB Google images from the open platform, and preprocess the images by image stitching, extraction by mask and projection grid. (2) Using principal component analysis, the first principal component band of the Google image was extracted; then a gray-level co-occurrence matrix was created, and eight classic texture features were extracted: mean, variance, homogeneity, contrast, dissimilarity, entropy, second moment and correlation. (3) Intersect the impermeable surface data in the study area with the street block units, and obtain the mean texture feature of the impermeable surface of each street block by region.

6. The method for refined identification of urban functional areas by coupling the spatiotemporal features of blocks and ensemble learning according to claim 1, characterized in that, The following method is used to eliminate feature redundancy: using the SALT street spatiotemporal feature system as the original input, a deep learning autoencoder model is constructed to reduce the dimensionality of the original features. The process includes: (1) Use the range standardization method to standardize the SALT features and unify the feature values ​​in the range of 0 to 1 to eliminate the data dimensions; (2) Based on the neural network, a deep learning autoencoder model consisting of an encoder and a decoder is constructed. The encoder compresses the original feature SALT by reducing the number of neurons layer by layer, and the decoder increases the number of neurons layer by layer based on the compressed data to obtain the reconstructed feature SALT'. During training, the weights and biases of each layer of the neural network are adjusted by minimizing the error between the original feature SALT and the reconstructed feature SALT'. The process of the encoder and decoder is as follows: g(SALT) = W(SALT) + b f(SALT’) = f(g(SALT)) = W T (SALT)+b’ In the formula, SALT represents the input feature, SALT' represents the reconstructed feature; g(*) and f(*) are multilayer networks; W and b represent the weights and bias matrices of the encoder, respectively; W T b' and b' represent the weights and bias matrices of the decoder, respectively; (3) The Autoencoder model is trained using the standardized SALT features. The mean squared error (MSE) metric is used to calculate the loss between the original features and the reconstructed features. The error is updated using gradient descent. Training is stopped when the error converges. The expression for the MSE is as follows: Loss(f, g) = = In the formula: Loss(f, g) represents the mean squared error between the original features and the reconstructed features; N represents the number of samples; and Let represent the input features and reconstructed features of the i-th sample, respectively.

7. The method for refined identification of urban functional areas by coupling the spatiotemporal features of blocks and ensemble learning according to claim 1, characterized in that, The following method is used to identify the specific categories of urban functional zones: Test and validation sets are constructed based on SALT dimensionality-reduced features and visually interpreted functional zone category labels. An Adaboost ensemble learning model is then trained to classify urban functional zones. The process includes: (1) Referring to the "National Urban Land Use and Planning Standard (GB50137-2011)," the functions of the study area are divided into 7 categories: administrative and public services; commercial facilities; residential; industrial, manufacturing and warehousing; mixed commercial and residential facilities; green space and squares; roads, streets and transportation; and then, with the help of online high-definition Google Maps, the categories of 30%-50% of the block plots are visually interpreted. (2) Initialize the training sample weights, that is, assign the same weight to each sample; (3) The dimensionality-reduced SALT street spatiotemporal feature system is used as the input feature, the category of the functional area sample is used as the label, the number of weak classifier decision trees is set, and the Adaboost model is trained to learn the mapping relationship between SALT features and functional area labels, and the error of each round is output. During the training process, the weight of the sample correctly classified by the weak classifier decision tree in this round will be reduced in the next round of training, and conversely, the weight of the sample misclassified will be increased in the next iteration. (4) When all weak classifiers have completed their evaluation, a strong classifier is formed and used to classify all urban functional areas.

8. The method according to any one of claims 1 to 7, characterized in that, This method is used to optimize traditional urban functional area identification methods and achieve high-precision classification of functional areas.

Images