Analyzing Urban Spatial Patterns and Functional Zones Using Sina Weibo POI Data: A Case Study of Beijing
利用新浪微博 POI 数据分析城市空间格局和功能区——以北京为例

Abstract  抽象

With the development of Web2.0 and mobile Internet, urban residents, a new type of “sensor”, provide us with massive amounts of volunteered geographic information (VGI). Quantifying the spatial patterns of VGI plays an increasingly important role in the understanding and development of urban spatial functions. Using VGI and social media activity data, this article developed a method to automatically extract and identify urban spatial patterns and functional zones. The method is put forward based on the case of Beijing, China, and includes the following three steps: (1) Obtain multi-source urban spatial data, such as Weibo data (equivalent to Twitter in Chinese), OpenStreetMap, population data, etc.; (2) Use the hierarchical clustering algorithm, term frequency-inverse document frequency (TF-IDF) method, and improved k-means clustering algorithms to identify functional zones; (3) Compare the identified results with the actual urban land uses and verify its accuracy. The experiment results proved that our method can effectively identify urban functional zones, and the results provide new ideas for the study of urban spatial patterns and have great significance in optimizing urban spatial planning.
随着 Web2.0 和移动互联网的发展，城市居民这种新型的“传感器”为我们提供了大量的自愿地理信息（VGI）。量化 VGI 的空间模式在理解和发展城市空间功能方面发挥着越来越重要的作用。利用 VGI 和社交媒体活动数据，本文开发了一种自动提取和识别城市空间模式和功能区的方法。该方法基于中国北京的案例提出，包括以下三个步骤：（1）获取多源城市空间数据，如微博数据（相当于中文的 Twitter）、OpenStreetMap、人口数据等;（2） 使用分层聚类算法、术语频率-逆文档频率 （TF-IDF） 方法和改进的 k-means 聚类算法来识别功能区;（3） 将确定的结果与实际城市土地用途进行比较，并验证其准确性。实验结果证明，该方法能够有效识别城市功能区，为城市空间格局研究提供了新思路，对优化城市空间规划具有重要意义。

1. Introduction  1. 引言

In the era of Web2.0 and mobile Internet, people often use Weibo (equivalent to Twitter in Chinese), online comments, photo sharing, travel records, and social media to generate, process, and share a large amount of information [1,2,3]. With the popularization of global positioning systems (GPS) and wireless cellular positioning technology in mobile devices, most of the information spontaneously created by users automatically carries spatial information [4]. This kind of spatial information is called volunteered geographic information (VGI) in academia [5]. VGI’s real-time, diversity, and content creativity have huge application potential in the fields of spatio-temporal analysis, urban planning, environmental monitoring, disaster warning, and public information services [6,7,8,9]. These massive data are gradually being mined and analyzed, and thus people have truly entered the era of big data. Goodchild also pointed out that we are rapidly entering an era where ordinary citizens are both consumers and producers of geographic information [1].
在 Web2.0 和移动互联网时代，人们经常使用微博（相当于中文的 Twitter）、在线评论、照片分享、旅行记录和社交媒体来生成、处理和分享大量信息 [ 1， 2， 3]。随着全球定位系统 （GPS） 和无线蜂窝定位技术在移动设备中的普及，用户自发创建的大部分信息都自动携带了空间信息 [ 4]。这种空间信息在学术界被称为自愿地理信息（VGI） [ 5]。VGI 的实时性、多样性和内容创意在时空分析、城市规划、环境监测、灾害预警和公共信息服务等领域具有巨大的应用潜力 [ 6， 7， 8， 9]。这些海量数据正在逐渐被挖掘和分析，人们也因此真正进入了大数据时代。Goodchild 还指出，我们正在迅速进入一个普通公民既是地理信息的消费者又是生产者的时代 [ 1]。

The advent of the big data era has put forward new ideas for the study of urban spatial patterns. Currently, data based on location-based service (LBS) technology are the most widely used data in urban research, such as bus card records, taxi trajectory data, mobile phone call records, and login data based on social media [10,11,12,13,14]. These data can be interpreted as a description of the city, and their mining and analysis can lead to a more people-oriented urban spatial pattern [1,2]. Traditional surveying methods based on visual and statistical data have some limitations in the research process, such as being expensive, not open to everyone, insufficient in real-time, and having data accuracy greatly affected by the environment [15,16]. They usually require more time and cost, but are more accurate, especially for small towns and when done by engaged people studying urban patterns [17,18]. Using easily accessible big data and location information can quickly obtain timely information that can describe the city, and thus is a meaningful research field in urban planning. In particular, the detailed urban functional zone classifications are of great importance for the landscape of urban spatial structure. Fine city function delineation using VGI data could depict the pre-planning layout for future cities, and this yields valuable insights for longstanding city economic development [19]. Taking the areas yet to be developed as an example, currently, they may be well equipped with public transport and industrial sites, however, balancing the provision between jobs and housing to facilitate healthy polycentric growth is crucial for sustainable and efficient urban development [20,21].
大数据时代的到来为城市空间格局的研究提出了新的思路。目前，基于位置服务（LBS）技术的数据是城市研究中应用最广泛的数据，如公交卡记录、出租车轨迹数据、手机通话记录和基于社交媒体的登录数据 [ 10， 11， 12， 13， 14]。这些数据可以被解释为对城市的描述，它们的挖掘和分析可以导致一个更加以人为本的城市空间格局 [ 1， 2]。传统的基于视觉和统计数据的调查方法在研究过程中存在一些局限性，例如成本高、不对所有人开放、实时性不足以及数据准确性受环境影响较大 [ 15， 16]。它们通常需要更多的时间和成本，但更准确，特别是对于小城镇，并且由研究城市模式的积极参与的人完成时 [ 17， 18]。使用易于获取的大数据和位置信息可以快速获得能够描述城市的及时信息，因此是城市规划中一个有意义的研究领域。特别是，详细的城市功能区分类对于城市空间结构景观具有重要意义。使用 VGI 数据进行精细的城市功能勾划可以描绘未来城市的预规划布局，这为城市的长期经济发展提供了有价值的见解 [ 19]。以尚未开发的地区为例，目前，它们可能配备了良好的公共交通和工业用地，但是，平衡就业和住房之间的供应以促进健康的多中心增长对于可持续和高效的城市发展至关重要 [ 20， 21]。

Many urban researchers have done related work using multi-source VGI data. Patrick Lüscher et al. [22] used an iterative online questionnaire to construct a city center cognitive model, and divided the city into different functional zones using the kernel density analysis method for the points of interests (POI) data obtained from the British Army Survey. Jameson L. et al. [23] used mobile phone data to analyze the spatio-temporal dynamics of the population, and analyzed land use patterns based on machine learning classification algorithms. John Steenbruggen et al. [24] comprehensively reviewed the research on mobile phone data and emphasized the feasibility of using digital data to optimize city management. Vincent Blondel [25] used more than 200 million communication data to study the corresponding areas and boundaries, and finally proposed geographic mobile communication based on the communication frequency and its average duration. Yang used [26] Baidu POI data to analyze the spatial composition of the urban network and divided it into 12 functional zones to analyze its aggregation mode. Liu et al. [27] used one-week taxi trajectory data in Shanghai, and used the source-sink model to quantify daily traffic characteristics and then discover the urban land use functions. Long et al. [28] used OpenStreetMap road network data and crowdsourced POI data to distinguish urban residential areas, and then they used census data to integrate population attributes. Rao et al. [29] used Shenzhen mobile phone data for one week to analyze user’s spatio-temporal attributes and proposed a model to identify different zones of employment in Shenzhen.
许多城市研究人员使用多源 VGI 数据完成了相关工作。Patrick Lüscher et al. [ 22] 使用迭代在线问卷构建了城市中心认知模型，并使用核密度分析方法对从英国陆军调查中获得的兴趣点 （POI） 数据将城市划分为不同的功能区。Jameson L. et al. [ 23] 使用手机数据分析了人口的时空动态，并基于机器学习分类算法分析了土地利用模式。John Steenbruggen 等 [ 24] 全面回顾了对移动电话数据的研究，强调了利用数字数据优化城市管理的可行性。Vincent Blondel [ 25] 使用了超过 2 亿条通信数据来研究相应的区域和边界，最终根据通信频率及其平均持续时间提出了地理移动通信。Yang 使用 [ 26] 百度 POI 数据分析了城市网络的空间构成，并将其划分为 12 个功能区来分析其聚合模式。Liu 等[27]利用上海一周的出租车轨迹数据，使用源-汇模型量化了日常交通特征，然后发现了城市土地利用功能。Long et al. [ 28] 使用 OpenStreetMap 路网数据和众包 POI 数据来区分城市居民区，然后他们使用人口普查数据来整合人口属性。Rao et al. [ 29] 使用深圳手机数据为期一周来分析用户的时空属性，并提出了一个模型来识别深圳不同的就业区域。

It can be seen that a lot of research has been conducted on urban spatial patterns and functional zones using mobile phone data, taxi data, and other POI data. This kind of VGI data are large-volume, easily obtainable, time-saving, and more people-oriented than traditional datasets, and their application in delineating city functional zones could provide more detailed information. Therefore, taking Beijing (the capital of China) as the study area, the focus of this study was to automatically extract and identify the urban spatial patterns and functional zones in Beijing using Sina Weibo data, OpenStreetMap, and other data. Through the theoretical and experimental research in this article, the following contributions were made: Firstly, the feasibility of using user-generated social media data on investigating urban spatial structures was verified; Secondly, by dividing research units using the road network, we obtained the natural areas in Beijing; Finally, the automatically-identified urban functional zones using social media data provided more information than did generally-defined residential or employment areas. The results of this study can help people better understand the spatial composition of large cities and megacities, and can also assist urban planners in planning urban functional zones using POI data. Our study has new reference value for urban planning, and can also provide reference for the development and improvement of different urban functional zones.
可以看出，利用手机数据、出租车数据和其他 POI 数据，对城市空间格局和功能区进行了大量研究。与传统数据集相比，这类 VGI 数据体积大、易获取、省时且更以人为本，在城市功能区划定中的应用可以提供更详细的信息。因此，以北京（中国首都）为研究区域，本研究的重点是利用新浪微博数据、OpenStreetMap 等数据自动提取和识别北京的城市空间格局和功能带。通过本文的理论和实验研究，做出了以下贡献：首先，验证了使用用户生成的社交媒体数据研究城市空间结构的可行性;其次，利用路网划分研究单位，得到北京的自然区域;最后，使用社交媒体数据自动识别的城市功能区提供了比一般定义的住宅或就业区更多的信息。本研究的结果可以帮助人们更好地理解大城市和特大城市的空间构成，也可以帮助城市规划者使用 POI 数据规划城市功能区。本研究对城市规划具有新的参考价值，也可为不同城市功能区的开发和改进提供参考。

The reminder of this paper is organized as follows: The Materials and Methods section describes the study area, data we collected, and the method used for analyzing urban spatial structure. The Results and Discussion section presents the experiments and results, and discusses next steps. Finally, the paper ends with the Conclusions section.
本文的提醒组织如下：材料和方法部分描述了研究区域、我们收集的数据以及用于分析城市空间结构的方法。Results and Discussion 部分介绍了实验和结果，并讨论了后续步骤。最后，本文以结论部分结束。

2. Materials and Methods  2. 材料和方法

In this section, we describe our study area and present the data we collected (including data pre-processing). Then, we show how to conduct hotspot analysis based on these data and use clustering methods to identify urban functional areas. The specific research process is shown in Figure 1.
在本节中，我们将介绍我们的研究领域并介绍我们收集的数据（包括数据预处理）。然后，我们展示了如何基于这些数据进行热点分析，并使用聚类方法来识别城市功能区。具体的研究过程如图 1 所示。

Figure 1. Flowchart of our study. (API is short for Application Programming Interface; POI is short for Point of Interest; OSM is short for OpenStreetMap; TF-IDF is short for Term Frequency–Inverse Document Frequency).
图 1.我们的研究流程图。（API 是应用程序编程接口的缩写;POI 是 Point of Interest 的缩写;OSM 是 OpenStreetMap 的缩写;TF-IDF 是 Term Frequency-Inverse Document Frequency 的缩写。

2.1. Study Area  2.1. 研究区域

The study area was Beijing, China (115°24′39″–117°30′37″ E, 39°26′9″–41°3′32″ N, as shown in Figure 2). Beijing is the capital and also a typical megacity in China. With the rapid urbanization, its urban scale has been expanded 12 times in 55 years [30]. The total area is 16,410.54 km², and the permanent population is 21.53 million. Considering the complexity of urban space, the large population (who can act as sensors), and even the increasingly prominent problem of big city diseases, Beijing is an ideal study area.
研究区域为中国北京（115°24′39“–117°30′37” E，39°26′9“–41°3′32” N，如图 2 所示）。北京是中国的首都，也是典型的特大城市。随着城市化进程的快速发展，其城市规模在 55 年间扩大了 12 倍[ 30]。总面积 16,410.54 公里 ² ，常住人口 2153 万。考虑到城市空间的复杂性、庞大的人口（可以充当传感器），甚至日益突出的大城市疾病问题，北京是一个理想的研究区域。

Figure 2. The location of the study area in China: Beijing.
图 2.研究区域在中国的位置：北京。

2.2. Data Collection  2.2. 数据收集

2.2.1. Sina Weibo POIs and Data Categorization
2.2.1. 新浪微博 POI 和数据分类

As one of the most popular social media platforms in China, Sina Weibo has the characteristics of fast updates, a large number of participants, and widely distributed users [3,9]. Most of the information on Sina Weibo is closely related to urban life. Since the content and types of Weibo POI are very rich, it is best to determine its category before acquisition. Considering the research content and the special background of Beijing, we then divided Weibo POI data into 15 categories (Table 1, modified according to [31]).
作为中国最受欢迎的社交媒体平台之一，新浪微博具有更新速度快、参与者人数多、用户分布广等特点 [ 3， 9]。新浪微博上的大部分信息都与城市生活息息相关。由于微博 POI 的内容和类型非常丰富，因此最好在获取之前确定其类别。结合研究内容和北京的特殊背景，我们将微博 POI 数据分为 15 类 （ 表 1，根据 [ 31 ] 修改）。

Table 1. POI categories and their descriptions.

Code	POI Category	Description
01	Hotel	Hotels, guesthouses, inns, etc.
02	Restaurants and drinking	Restaurants, KFCs, McDonald’s, Pizza Huts, cafes, etc.
03	Shopping	Shopping malls, shopping centers, shops, convenience stores, supermarkets, specialty stores, pedestrian streets, etc.
04	Tourist attraction	Scenic spots, resorts, parks, squares, zoos, botanical gardens, churches, etc.
05	Healthcare	Hospitals, clinics, emergency centers, pharmacies, etc.
06	Building (including but not limited to companies)	Office buildings, villas, industrial parks, enterprises, companies, etc.
07	Financial and insurance	Banks, ATMs (Automated Teller Machine), insurance offices, security offices, finance offices, etc.
08	Residential	Residential, bathing, laundry, beauty salons, car washes, business halls, express services, etc.
09	Public facility	Newsstands, public telephones, public toilets, post offices, etc.
10	Government agency	Government agencies, embassies, institutions, procuratorates, courts, offices, etc.
11	Industrial site	Factories, farms, fisheries, forest farms, pastures, etc.
12	Public transport	Airports, railway stations, bus stations, subway stations, parking lots, etc.
13	Highway	Expressways, toll stations, gas stations, service areas, etc.
14	Sport and entertainment	Stadiums, football fields, tennis courts, basketball courts, badminton courts, fitness centers, entertainment centers, KTV (Karaoke TV), discotheques, bars, chess rooms, Internet cafes, movie theaters, etc.
15	Science and education	Universities, schools, libraries, research institutes, science and technology museums, historical museums, exhibition halls, conference centers, art galleries, cultural palaces, archives, television stations, newspapers, publishing houses, magazines, theaters, etc.

According to coding classifications, different categories of data were collected. The collection time was from 13 April to 17 April 2015, and a total of 335,234 pieces of data were collected. Among them, the data volumes for codes 01 through 15 were 12,491; 12,152; 29,001; 13,405; 60,863; 113,206; 10,725; 34,175; 658; 5341; 4017; 14,036; 1696; 6690; and 16,778, respectively (shown in Figure 3). However, there were problems such as duplication of data records and ambiguity in place names, which required further data cleaning. Next, we deleted duplicate records and deleted records that did not meet a specific classification.
根据编码分类，收集了不同类别的数据。收集时间为 2015 年 4 月 13 日至 4 月 17 日，共收集了 335,234 条数据。其中，代码 01 到 15 的数据量为 12,491 个;12,152;29,001;13,405;60,863;113,206;10,725;34,175;658;5341;4017;14,036;1696;6690;和 16,778 例（如图 3 所示）。但是，存在数据记录重复和地名歧义等问题，需要进一步清理数据。接下来，我们删除了重复记录和不符合特定分类的记录。

Figure 3. Statistics of the POIs.
图 3.POI 的统计信息。

It can be seen that in the process of categorization, some parks were also classified as tourist attractions. This is because Beijing has a large number of tourists, and they often visit parks. Therefore, the parks were classified as tourist attractions rather than public facilities. The number of companies in Beijing is high (73,224 companies out of 113,206 buildings). Therefore, in the following research, companies were not considered in the category of building, but the distribution of companies was studied separately. After data cleaning and processing, 51,916 company data points and 115,616 classified data points were finally obtained (company was categorized as 06*). The POI data of each category is shown in Figure 3.
可以看出，在分类过程中，一些公园也被归类为旅游景区。这是因为北京游客众多，他们经常去公园。因此，这些公园被归类为旅游景点而不是公共设施。北京的公司数量很多（113,206 栋建筑中有 73,224 家公司）。因此，在下面的研究中，公司没有被考虑在建筑类别中，而是单独研究了公司的分布。经过数据清洗和处理，最终获得 51,916 个公司数据点和 115,616 个分类数据点（公司被归类为 06*）。每个类别的 POI 数据如图 3 所示。

2.2.2. OpenStreetMap and Map Segmentation
2.2.2. OpenStreetMap 和地图分割

We collected the road network data of Beijing on 14 April 2015 from OpenStreetMap (https://www.openstreetmap.org). The road data set included 50,816 roads with a total length of 24,877,717 m, and a railway with 5121 sections and a total length of 3,699,118 m. Then, we combined OpenStreetMap’s road network classification and selected three levels: highway (motorway_link), trunk (trunk_link), and primary (primary_link) as the research objects. There were 9655 lines with a total length of 6,077,240 m, as shown in Figure 4a. These three different road types constitute the natural division of Beijing. It can be seen intuitively that the outline of Beijing’s road network meets the experimental requirements.
我们从 OpenStreetMap （ https://www.openstreetmap.org） 收集了 2015 年 4 月 14 日北京市的道路网络数据。道路数据集包括 50,816 条道路，总长度为 24,877,717 m，以及一条铁路，共有 5121 个路段，总长度为 3,699,118 m。然后，结合 OpenStreetMap 的路网分类，选取高速公路 （motorway_link）、主干 （trunk_link） 和主要 （primary_link） 三个级别作为研究对象。共有 9655 条线路，总长度为 6,077,240 m，如图 4a 所示。这三种不同的道路类型构成了北京的自然划分。直观地可以看出，北京路网大纲符合实验要求。

Figure 4. OpenStreetMap and population data of Beijing: (a) Selected OSM data; (b) Map segmentation results; (c) The spatial distribution density of the population in Beijing.
图 4.OpenStreetMap 和北京人口数据：（a） 选定的 OSM 数据;（b） 地图分割结果;（c） 北京市人口空间分布密度。

In order to better divide the research area into different zones, we needed to remove unnecessary details and ensure the topological relationships of the roads, including multi-lane merging, two-lane road centerline extraction, overpass deletion, and topological relationship correction. After checking the data, we segmented regions according to the center line of the road network (Figure 4b).
为了更好地将研究区域划分为不同的区域，我们需要去除不必要的细节并确保道路的拓扑关系，包括多车道合并、双车道道路中心线提取、立交桥删除和拓扑关系校正。检查数据后，我们根据路网的中心线对区域进行分割（图 4b）。

2.2.3. Population Data  2.2.3. 人口数据

China’s 1 km grid population data set was based on land use type data and demographic data obtained from remote sensing data. The data set was used to establish a population spatial distribution model by using the spatial analysis function of the geographic information system to spatialize the statistical population data [32]. We extracted the population distribution data within the border of Beijing from China’s 1 km grid population data set (2010). The generated population density distribution map is shown in Figure 4c.
中国的 1 km 网格人口数据集基于土地利用类型数据和遥感数据获得的人口统计数据。利用该数据集，利用地理信息系统的空间分析功能对统计人口数据进行空间化，建立种群空间分布模型 [ 32]。我们从中国的 1 公里网格人口数据集 （2010） 中提取了北京境内的人口分布数据。生成的人口密度分布图如图 4c 所示。

It can be seen from the above analysis that the population density is the highest in central Beijing. As the urban center expands, the population density gradually decreases. However, in the suburbs, the population density shows a high-density distribution center in a small area, showing obvious characteristics of suburbanization. In addition, the population density in the core areas of suburban counties remains high. Generally speaking, the population density in the east is higher than that of the west, especially in areas where the population density distribution is expanding in the southeast and Langfang.
从以上分析可以看出，北京市中心的人口密度最高。随着城市中心的扩大，人口密度逐渐降低。而郊区人口密度呈现小区域高密度配送中心，表现出明显的郊区化特征。此外，郊区县城核心区的人口密度仍然很高。一般来说，东部地区的人口密度高于西部地区，尤其是在东南部和廊坊地区人口密度分布不断扩大的地区。

2.3. Analysis of Urban Spatial Structure
2.3. 城市空间结构分析

2.3.1. Analyzing Urban Hot Spots Based on Weibo POI Data
2.3.1. 基于微博 POI 数据分析城市热点

Weibo POI data can better describe the distribution of people in a city through the location information of volunteers. In order to further analyze the distribution characteristics, we selected a large number of checked-in POIs in Weibo for analysis, and used the checkin_num of each POI point as a weight to analyze the kernel density.
微博 POI 数据可以通过志愿者的位置信息更好地描述一个城市中的人口分布。为了进一步分析分布特征，我们在微博中选取了大量签入的 POI 进行分析，并使用每个 POI 点的 checkin_num 作为权重来分析核密度。

The kernel density estimation (KDE) algorithm mainly uses a moving unit (equivalent to a window) to estimate the density of a point or line pattern [33]. It is defined as x₁ … x_n and is an independent and identically distributed sample drawn from the population of the distribution density function f (). To estimate the value of f () at a certain x, the Rosenblatt-Parzen kernel estimation is usually used:
核密度估计 （KDE） 算法主要使用移动单元（相当于一个窗口）来估计点或线型的密度 [ 33]。它被定义为 x ₁ ...x _n 和 是从分布密度函数 （） 的总体中提取的独立且同分布的样本。为了估计 （） 在某个 x 处的值，通常使用 Rosenblatt-Parzen 核估计：

𝑓𝑛(𝑥)=1𝑛ℎ∑𝑖=1𝑛𝑘(𝑥−𝑥𝑖ℎ)$$f_n\left(x\right)=\frac{1}{nh}{\displaystyle \sum }_{i=1}^{n}k\left(\frac{x-x_i}{h}\right)$$

(1)

where k () is the kernel function; h > 0 is the variable; and (x − x_i) represents the distance from the estimated point to the sample x_i. In KDE estimation, the determination or choice of variable h has a great influence on the calculation result. When h increases, the point density changes more smoothly in space, but it will hide the density structure; When h decreases, the estimated point density changes suddenly and unevenly [34].
其中 k （） 是内核函数;h > 0 是变量;和 （x − x） 表示从估计点到样本 x 的距离。在 KDE 估计中，变量 h 的确定或选择对计算结果有很大影响。当 h 增加时，点密度在空间上变化更平滑，但会隐藏密度结构;当 h 减小时，估计的点密度会突然且不均匀地变化 [ 34]。

In the KDE module of ArcGIS, the default bandwidth is automatically generated. The larger the search radius value, the smoother is the density grid generated and the higher is the generalization degree; therefore, the smaller the value, the more detailed is the information displayed in the generated grid. In order to obtain more detailed results, we changed the default search radius to 1500 m and the output cell size of the raster image to 100 m.
在 ArcGIS 的 KDE 模块中，默认带宽是自动生成的。搜索半径值越大，生成的密度格网越平滑，泛化程度越高;因此，值越小，生成的网格中显示的信息就越详细。为了获得更详细的结果，我们将默认搜索半径更改为 1500 m，并将栅格图像的输出像元大小更改为 100 m。

2.3.2. Identifying Urban Functional Zones
2.3.2. 识别城市功能区

In this section, we used Sina Weibo POI data to analyze urban functional zones.
在本节中，我们使用新浪微博 POI 数据来分析城市功能区。

Cluster analysis is a statistical analysis method for studying classification problems, and it is also an important algorithm for data mining. In this research, we mainly used the k-means algorithm and hierarchical clustering algorithm.
聚类分析是研究分类问题的一种统计分析方法，也是数据挖掘的重要算法。在这项研究中，我们主要使用了 k-means 算法和分层聚类算法。

• K-means: For a given data set, we made the following provisions: the set of n d-dimensional points was X = {x_i}, i = 1, …, n; the set of k clusters was C = {c_k}, k = 1, …, k; the mean value of c_k was μ_k; and the squared error was J(𝑐𝑘)=∑𝑥𝑖∈𝑐𝑘||𝑥𝑖−𝜇𝑘||2$\mathrm{J}\left(c_k\right)={\displaystyle \sum }_{x_{i\in c_k}}{\left|\left|x_i-{\mu }_k\right|\right|}^2$. Therefore, the goal of K-means can be understood as a solution that minimizes J(𝑐𝑘)=∑𝑥𝑖∈𝑐𝑘||𝑥𝑖−𝜇𝑘||2$\mathrm{J}\left(c_k\right)={\displaystyle \sum }_{x_{i\in c_k}}{\left|\left|x_i-{\mu }_k\right|\right|}^2$.
  K-means：对于给定的数据集，我们做出了以下规定：n 个 d 维点的集合是 X = {x}， = 1， ...， n;k 个集群的集合是 C = {c _k }， k = 1， ...， k;c _k 的平均值为 μ _k ;平方误差为 J(𝑐𝑘)=∑𝑥𝑖∈𝑐𝑘||𝑥𝑖−𝜇𝑘||2$\mathrm{J}\left(c_k\right)={\displaystyle \sum }_{x_{i\in c_k}}{\left|\left|x_i-{\mu }_k\right|\right|}^2$ 。因此，K-means 的目标可以理解为最小化 J(𝑐𝑘)=∑𝑥𝑖∈𝑐𝑘||𝑥𝑖−𝜇𝑘||2$\mathrm{J}\left(c_k\right)={\displaystyle \sum }_{x_{i\in c_k}}{\left|\left|x_i-{\mu }_k\right|\right|}^2$ 的解。
• Hierarchical clustering algorithm: A hierarchical clustering method is used to construct and maintain a clustering tree formed by clusters and sub-clusters according to a given distance measurement criterion between clusters until a certain end condition is met. Hierarchical clustering algorithm is divided into condensed and split, from bottom-up and top-down, according to hierarchical decomposition. The default discussed in this article is cohesive.
  分层聚类算法：使用分层聚类方法，根据聚类之间给定的距离测量标准，构建和维护由聚类和子聚类形成的聚类树，直到满足一定的结束条件。分层聚类算法根据分层分解分为压缩和拆分，从下到上和自上而下。本文讨论的默认值是 cohesive。

TF-IDF (term frequency-inverse document frequency) is a statistical method used to evaluate the importance of a word to one of the documents in a document set or corpus [35]. The importance of a word is proportional to the number of times it appears in the document, but it decreases inversely proportional to the frequency of its appearance in the corpus. TF-IDF is TF × IDF, where TF is term frequency (term frequency) and IDF is inverse document frequency (inverse document frequency). In a given document, TF refers to the frequency of a given word in the document,
TF-IDF（词频-逆文档频率）是一种统计方法，用于评估一个词对文档集或语料库中的某个文档的重要性 [ 35]。单词的重要性与它在文档中出现的次数成正比，但它的减少与其在语料库中出现的频率成反比。TF-IDF 是 TF × IDF，其中 TF 是词频（词频），IDF 是逆向文档频率（逆向文档频率）。在给定文档中，TF 是指给定单词在文档中的频率，

𝑡𝑓𝑖𝑗=𝑛𝑖,𝑗∑𝑘𝑛𝑘,𝑗$$t{f}_{ij}=\frac{n_{i,j}}{{{\displaystyle \sum }}_k{n}_{k,j}}$$

(2)

where n_i,_j is the number of occurrences of the word in the file d_j, and the denominator is the sum of the number of occurrences of all words in the file d_j.
其中 n 是单词在文件 d 中出现的次数，分母是文件 d 中所有单词的出现次数之和。

IDF is used to measure the universal importance of a word. The IDF of a specific word can be obtained by dividing the total number of documents in the research by the number of documents containing the word, and then taking the logarithm of the obtained quotient,
IDF 用于衡量单词的普遍重要性。一个特定词的 IDF 可以通过将研究中的文档总数除以包含该词的文档数，然后取所得商的对数来获得，

𝑖𝑑𝑓𝑖=𝑙𝑜𝑔|𝐷||{𝑗:𝑡𝑖∈𝑑𝑗}|$$id{f}_i=log\frac{\left|D\right|}{\left|\left\{j:t_i\in d_j\right\}\right|}$$

(3)

where |𝐷|$\left|D\right|$ is the total number of documents in the corpus, and |{𝑗:𝑡𝑖∈𝑑𝑗}|$\left|\left\{j:t_i\in d_j\right\}\right|$ is the number of documents containing word.
其中 |𝐷|$\left|D\right|$ 是语料库中的文档总数， |{𝑗:𝑡𝑖∈𝑑𝑗}|$\left|\left\{j:t_i\in d_j\right\}\right|$ 是包含 Word 的文档数。

Then, according to 𝑡𝑓𝑖𝑑𝑓𝑖𝑗=𝑡𝑓𝑖𝑗×𝑖𝑑𝑓𝑖$tfid{f}_{ij}=t{f}_{ij}\times id{f}_i$, with a high word frequency in a particular file, and a low file frequency of the word in the entire file set, a high-weight TF-IDF can be generated.
然后，根据 𝑡𝑓𝑖𝑑𝑓𝑖𝑗=𝑡𝑓𝑖𝑗×𝑖𝑑𝑓𝑖$tfid{f}_{ij}=t{f}_{ij}\times id{f}_i$ ，在特定文件中具有较高的字频，而该词在整个文件集中的出现频率较低，可以生成高权重的 TF-IDF。

3. Results and Discussion
3. 结果与讨论

3.1. Weibo Hot Spots Analysis Results
3.1. 微博热点分析结果

We selected a large number of Weibo POIs (with each category) for analysis, and used the checkin_num kernel density analysis weight of each POI point to perform kernel density analysis, and obtained the following results.
我们选取了大量的微博 POI（每个类别）进行分析，并使用每个 POI 点的 checkin_num 核密度分析权重进行核密度分析，得到以下结果。

It can be seen from Figure 5 that the spatial distribution of Weibo users in Beijing is large. In the urban area, it is mainly concentrated in science and education areas, commercial and entertainment areas, and diplomatic and political areas. It is not difficult to understand that there are a large number of universities in science and education areas. College students are an active group of Weibo users. At the same time, office people also like to use Weibo when commuting. In diplomatic and political areas, hot spots are mainly concentrated in tourist attractions of political significance with Tiananmen Square as the center. In commercial and entertainment areas, people mainly use Weibo to share information during leisure or entertainment activities. In addition to the urban area, the Capital International Airport District and Changping District are also hot spots for Weibo user activities. In Changping District, the campuses of some colleges and universities are relatively concentrated, and it is also the area where the Great Wall (Badaling Great Wall) is located. It is a hot spot for people to sign in on Weibo. For the results of kernel density analysis, the data can be used for further interpretation. We selected the POI points with high check-ins numbers to display, as shown in the Table 2.
从图 5 可以看出，北京的微博用户空间分布较大。在市区，主要集中在科学和教育领域、商业和娱乐领域以及外交和政治领域。不难理解，科学和教育领域有大量的大学。大学生是 Weibo 用户的活跃群体。同时，上班族在通勤时也喜欢使用微博。在外交和政治领域，热点主要集中在以天安门广场为中心的具有政治意义的旅游景区。在商业和娱乐领域，人们主要在休闲娱乐活动中使用微博分享信息。除了市区，首都国际机场区和昌平区也是微博用户活动的热点。在昌平区，一些高校的校园比较集中，也是长城（八达岭长城）所在的区域。它是人们在微博上登录的热点。对于核密度分析的结果，数据可用于进一步解释。我们选择了签到数较高的 POI 点进行显示，如表 2 所示。

Figure 5. Kernel density analysis results of Weibo POI data.
图 5.Weibo POI 数据的核密度分析结果。

Table 2. POI points with high check-in times.

Checkin_num (Number of Checkin Points)	Title
150255	Capital Airport T3 Terminal
90515	Capital Airport T2 Terminal
76175	Weigong Village
69227	Beijing Normal University
67681	Beijing University
64146	Wangfujing
64146	Beijing University of Aeronautics and Astronautics
63287	Beijing Jiaotong University
62810	Tsinghua University
58960	Xidan
58136	University of Science and Technology
56575	Tiananmen Square
51035	Changxindian District
49570	Capital Airport
47521	Communication University of China

3.2. Identifying Urban Functional Zones
3.2. 识别城市功能区

For the 15 categories of POI data we classified, we first used the spatial connection tool in ArcGIS to calculate the number of POI points in each divided area. Furthermore, the hot spot discovery tool was used to detect cluster centers. We selected eight typical categories of POI data to determine the clustering centers (as shown in Figure 6), and the distribution of hot spots obtained by ArcGIS.
对于我们分类的 15 类 POI 数据，我们首先使用 ArcGIS 中的空间连接工具来计算每个分割区域中的 POI 点数量。此外，热点发现工具用于检测集群中心。我们选择了八类典型的 POI 数据来确定聚类中心（如图 6 所示）以及 ArcGIS 获得的热点分布。

Figure 6. Eight categories of POI hot spots: (a) government agencies; (b) science and education; (c) buildings; (d) public transport; (e) shopping; (f) residential; (g) tourist attractions; (h) sports and entertainments.
图 6.八类 POI 热点：（a） 政府机构;（b） 科学和教育;（c） 建筑物;（d） 公共交通;（e） 购物;（f） 住宅;（g） 旅游景点;（h） 体育和娱乐。

In specific experiments, we mainly used three methods, the hierarchical clustering method (Figure 7a), the TD-IDF method (Figure 7b), and the improved k-means clustering method (Figure 7c). The improved k-means method takes the aforementioned hotspot analysis results as the initial clustering center, thus expecting a better clustering result. The TF-IDF method compares the urban function exploring it as text-topic discovery, and this urban function similarity is further explored using a plain k-means method. We analyzed and compared these clustering results.
在具体的实验中，我们主要使用了三种方法，分层聚类方法（图 7a）、TD-IDF 方法（图 7b）和改进的 k-means 聚类方法（图 7c）。改进的 k-means 方法以上述热点分析结果作为初始聚类中心，从而期待更好的聚类结果。TF-IDF 方法将其探索的城市功能作为文本主题发现进行比较，并使用普通的 k-means 方法进一步探索这种城市功能相似性。我们分析并比较了这些聚类结果。

Figure 7. Clustering results: (a) hierarchical clustering; (b) TF-IDF; (c) custom k-means clustering.
图 7.聚类结果： （a） 层次聚类;（b） TF-IDF;（c） 自定义 K-Means 聚类。

By counting the POI data of each functional zone, we sorted the number of various categories of POIs in the functional zone, as shown in Table 3. We then comprehensively analyzed the three clustering results and statistical data, and finally determined eight categories of functional zones, including diplomatic and political centers, science and education areas, mature residential areas, new residential areas, commercial and entertainment areas, tourist attraction areas, areas to be developed, and unclassified areas.
通过对每个功能区的 POI 数据进行计数，我们对功能区中各种类别的 POIs 数量进行排序，如表 3 所示。然后，我们对 3 个聚类结果和统计数据进行综合分析，最终确定了 8 类功能区，包括外交政治中心、科教区、成熟住宅区、新建住宅区、商业娱乐区、旅游景区、待开发区和未分类区。

Table 3. POI category ranking value in each functional zone.

POI Category	1 (Functional Zone)	2	3	4	5	6	7
Restaurants and drinking	3	7	1	2	5	4	6
Highway	6	5	4	1	7	2	3
Industrial site	6	7	4	1	3	5	2
Public transport	5	7	3	1	6	4	2
Public facility	6	4	1	3	5	2	7
Shopping	3	4	1	2	7	5	6
Financial and insurance	3	6	1	2	5	4	7
Residential	4	7	1	2	5	3	6
Science and education	6	1	2	3	5	4	7
Tourist attraction	2	3	1	4	5	6	7
Sport and entertainment	2	6	1	3	5	4	7
Healthcare	3	7	1	2	5	4	6
Government agency	4	7	2	1	6	3	5
Hotel	3	6	1	2	5	4	7
Buildings (including but not limited to companies)	2	3	1	5	4	6	7

• Diplomatic and political zone
  外交和政治区

In these areas, a large number of embassies are gathered, and the number of POIs in tourist attractions, sports and entertainment, and buildings is large. Combining the fact that Beijing is also the capital, this area is not only the gathering place of embassies, but also the location of Tiananmen Square, the Forbidden City, and the Great Hall of the People, etc.
在这些地区聚集了大量大使馆，旅游景点、体育和娱乐以及建筑物的 POI 数量很大。结合北京也是首都的事实，该地区不仅是大使馆的聚集地，也是天安门广场、紫禁城和人民大会堂等的所在地。

• Science and education zone
  科教区

In these areas, POI data for science, education, culture, and publicity are the highest, and combined with the location of the area, it can be seen that there are a large number of universities in this area, such as Peking University and Tsinghua University. At the same time, Zhongguancun, China’s earliest high-tech development center, has a large number of high-tech companies and scientific research institutes in these areas. Therefore, there are a large number of building and companies in these areas.
在这些区域中，科教、文化、宣传的 POI 数据最高，结合该地区的位置，可以看出该区域有大量的高校，比如北京大学和清华大学。同时，中关村作为中国最早的高新技术开发中心，在这些地区拥有大量的高科技公司和科研院所。因此，这些地区有大量的建筑和公司。

• Mature residential zone  成熟住宅区

In these areas, the number of residential POIs is the largest, and the number of restaurants, public facilities, shopping centers, financial and insurance, tourist attractions, sports and entertainments, and healthcare POIs are also the highest. It can be seen that in mature residential areas, all types of service facilities are the most complete. They are distributed around the core functional areas of the city. At the same time, very few areas in the suburbs have developed into mature residential areas.
在这些地区，住宅 POI 的数量最多，餐厅、公共设施、购物中心、金融和保险、旅游景点、体育和娱乐以及医疗保健 POI 的数量也最高。由此可见，在成熟的住宅小区，各类服务设施最为齐全。它们分布在城市的核心功能区域。与此同时，郊区极少数地区已发展成为成熟的住宅区。

• New residential zone  新住宅区

As seen in the mature residential areas, the residential POI of new residential areas is the largest of all categories, but the number of other categories in this functional zone is mostly lower than that of mature residential zones. Highway services, industrial sites, public transportation, and government agencies have the highest number of POIs. This is because this area is composed of many sub-regions with a large number of government agencies. In addition, it is located in the suburbs, and has more highway services and public transportation.
从成熟住宅区可以看出，新建住宅区的住宅 POI 是所有类别中最大的，但该功能区其他类别的数量大多低于成熟住宅区。公路服务、工业场所、公共交通和政府机构的 POI 数量最多。这是因为该区域由许多子区域组成，拥有大量的政府机构。此外，它位于郊区，拥有更多的高速公路服务和公共交通。

• Commercial and entertainment zone
  商业和娱乐区

This functional zone is located near the diplomatic and political zone, and next to the mature residential zone, which shows people’s shopping habits. However, the number of various categories in this zone is balanced, and the number in the same category is not high, which is mainly due to the small number of sub-regions.
这个功能区位于外交和政治区附近，紧邻成熟住宅区，展示了人们的购物习惯。但是，该区域的各个品类数量均衡，同品类数量不多，这主要是由于子区域数量较少。

• Tourist attractions zone  旅游景点区

In this functional zone, there are many public transportation POIs. It can be seen from the distribution of sub-regions that the functional zone is basically distributed in the suburbs, but the number of tourist attractions POIs is not particularly large.
在这个功能区，有许多公共交通 POI。从子区域分布可以看出，功能区基本分布在郊区，但旅游景点 POI 数量并不是特别大。

• Area to be developed  待开发区域

In these areas, all types of Checkin_num are small, but the number of public transportation and industrial sites is large. At the same time, it can be seen from the distribution that they are adjacent to new residential areas and located in remote counties.
在这些地区，所有类型的 Checkin_num 都很小，但公共交通和工业场所的数量很大。同时，从分布中可以看出，他们毗邻新建住宅区，位于偏远县城。

• Unclassified area  未分类区域

Since Weibo POI check-in data are essentially volunteer geographic information, the number of POIs in some areas is not high enough and they are not classified.
由于微博 POI 签到数据本质上是志愿者地理信息，因此部分地区的 POI 数量不够多，没有分类。

3.3. Verifying the Results
3.3. 验证结果

For the results of functional zones obtained by clustering, we evaluated them using the following three measures.
对于通过聚类获得的功能区结果，我们使用以下三个指标对其进行评估。

Firstly, we compared the clustering results with the Beijing City Master Plan (2004–2020, shown in Figure 8a), mainly comparing the downtown area. It can be clearly seen from the planning map that area A is land for commercial and financial use, and area B is land for science, teaching, and research, which is in full agreement with the results of this article. In addition, it can be seen from the planning map that the downtown area is a residential area, which is not inconsistent with functional zones such as the diplomatic and political area that we derived. Because residential areas have dominant functions in cities, and functions are established by human activities in the residential environment.
首先，我们将聚类结果与北京市总体规划（2004-2020 年，如图 8a 所示）进行了比较，主要比较了市中心区域。从规划图中可以清楚地看出，A 区是商业和金融用途的土地，B 区是科学、教学和研究用地，这与本文的结果完全吻合。此外，从规划图中可以看出，市区是居民区，与我们衍生的外交、政治区等功能区并不矛盾。因为住宅区在城市中具有主导功能，而功能是由人类在居住环境中的活动建立的。

Figure 8. Verifying the results: (a) Beijing City Master Plan (2004–2020); (b) Verification of initial clustering centers; (c) Verification of typical areas.
图 8.验证结果：（a） 北京市城市总体规划（2004-2020 年）;（b） 验证初始集群中心;（c） 典型区域的验证。

Secondly, we compared the clustering results with the initial cluster centers of k-means (shown in Figure 8b). Because the clustering result of the k-means method itself depends on the initial clustering center, we started from the clustering method for comparative verification. It can be seen from the above that the eight initial clusters selected here fell within the corresponding local area, which can be seen in the selection of the initial cluster center. The method here is effective and the results are also reliable.
其次，我们将聚类结果与 k-means 的初始聚类中心进行了比较（如图 8b 所示）。因为 k-means 方法本身的聚类结果取决于初始聚类中心，所以我们从聚类方法开始进行比较验证。从上面可以看出，这里选择的 8 个初始聚类都落在了对应的 local 区域内，这可以从初始聚类中心的选择中看出。这里的方法是有效的，结果也是可靠的。

Finally, we selected some typical areas to verify the results (outer areas were not selected for comparison because they were mainly unclassified areas and areas to be developed). The results (shown in Figure 8c) show that Xiangshan Park is a tourist attraction in several typical areas selected at random. Yongle District is located in a mature residential area; Peking University and the Zhongguancun campus of the Chinese Academy of Sciences are located in the science, education, and cultural district. The French Embassy and Tiananmen Square are located in the diplomatic and political center. Sanlitun Bar Street is in the commercial and entertainment district.
最后，我们选择了一些典型区域来验证结果（没有选择外部区域进行比较，因为它们主要是未分类区域和待开发区域）。结果（如图 8c 所示）显示，象山公园是随机选择的几个典型区域的旅游景点。永乐区位于成熟住宅区;北京大学和中国科学院中关村校区位于科教文化区。法国大使馆和天安门广场位于外交和政治中心。三里屯酒吧街位于商业和娱乐区。

Combining the above three verification methods, and considering the current situation of Beijing’s highly mixed land use, it can be seen that the results of Beijing’s functional zoning obtained by this method had great accuracy.
结合上述三种验证方法，并考虑到北京市土地利用高度混合的现状，可以看出该方法得到的北京市功能分区结果具有很高的准确率。

3.4. Discussion  3.4. 讨论

The significance of this work is the development a method to automatically identify detailed spatial functional zones. The fine distinction between mature and new residential zones, and the delineation of areas to be developed are of greater importance, except for the easily distinguishable zones (diplomatic and political, science and education, commercial and entertainment). Taking mature residential zones as reference, the new residential zones need to place more effort into promoting service-related facilities, including shopping, financial and insurance, sport and entertainment, and healthcare facilities. As for the areas to be developed, they distinguish themselves by highly ranked public transport and industrial sites, and they also next to the new residential zones in the suburban areas. This kind of functional zone is of great potential and places near center zones would enjoy better development if the infrastructures there were gradually improved.
这项工作的意义在于开发一种自动识别详细空间功能区的方法。成熟住宅区和新住宅区之间的细微区分以及待开发区域的划分更为重要，除了易于区分的区域（外交和政治、科学和教育、商业和娱乐）。以成熟的住宅区为参考，新住宅区需要加大力度推广与服务相关的设施，包括购物、金融和保险、体育和娱乐以及医疗保健设施。至于待开发的地区，它们以排名靠前的公共交通和工业用地而著称，而且它们也毗邻郊区的新住宅区。这种功能区具有巨大的潜力，如果那里的基础设施逐渐得到改善，靠近中心区的地方将得到更好的发展。

We then combined the characteristics of the study area and the research results for further analysis.
然后，我们将研究区域的特征和研究结果结合起来进行进一步分析。

Firstly, the central city of Beijing is showing a trend of suburbanization, and the spatial distribution structure presents a three-level structure of main center-sub center-town. However, despite the significant increase in population density in the suburbs, the construction of various infrastructures in the area is not yet complete, and the level of urbanization needs to be improved.
（1）北京中心城区呈现郊区化趋势，空间分布结构呈现主中心-副中心-城镇三级结构。然而，尽管郊区人口密度显著增加，但该地区各种基础设施的建设尚未完成，城市化水平有待提高。

Secondly, Beijing is developing towards the southeast. It can be seen that the connection area between Beijing and Tianjin (Langfang, located in the southeast of Beijing) has a large population and spatial distribution density. At the same time, the distribution of Weibo POI also shows that the regional distribution density in the southeast direction is large.
其次，北京正在向东南方向发展。由此可见，京津连接区（廊坊，位于北京东南部）人口和空间分布密度较大。同时，微博 POI 的分布也表明，东南方向的区域分布密度较大。

Finally, diplomacy and politics; business and entertainment; and science, education, and cultural are the main service functions of the major urban areas. Mature residential areas are located near the city center. In the suburbs and counties of Beijing, there are new residential areas and areas to be developed. Commercial and entertainment areas are less distributed in suburban counties.
最后，外交和政治;商业和娱乐;以及科学、教育和文化是主要城市地区的主要服务功能。成熟的住宅区位于市中心附近。在北京的郊区和县城，有新的住宅区和待开发的区域。商业和娱乐区在郊区县的分布较少。

With the process of urbanization, the built-up area of Beijing has become larger and larger, and more and more people live in the suburbs. On the one hand, the process of suburbanization has eased the pressure on population, traffic, and housing in the major city, but at the same time many new problems have emerged. For example, many people live in the suburbs but work in the city center, which requires a long time to commute. On the other hand, it can be seen from the results of the analysis that the infrastructure in the suburbs is still not sound, so that the schooling and medical problems of children cannot be well addressed.
随着城市化进程，北京的建成区越来越大，越来越多的人居住在郊区。一方面，郊区化进程缓解了大城市的人口、交通和住房压力，但与此同时也出现了许多新问题。例如，许多人住在郊区，但在市中心工作，这需要很长时间才能通勤。另一方面，从分析结果可以看出，郊区的基础设施仍然不健全，以至于儿童的学业和医疗问题无法得到很好的解决。

What should be done? In the process of ensuring the stable development of central urban areas, the development of emerging urban areas should also be balanced and attention should be paid to the equal distribution of resources, such as education, healthcare, and other supporting facilities. In addition, while optimizing the internal structure of the city, it is necessary to integrate Beijing’s overall resources for external development, actively drive the surrounding areas, and strive to achieve the coordinated development of the Beijing-Tianjin-Hebei metropolitan area.
应该怎么做？在保证中心城区稳定发展的过程中，新兴城区的发展也要均衡，注重教育、医疗和其他配套设施等资源的均衡分配。此外，在优化城市内部结构的同时，要整合北京整体资源进行外部发展，积极带动周边地区，努力实现京津冀都市圈的协调发展。

4. Conclusions  4. 结论

Currently, urban residents provide massive VGI, and understanding of the urban spatial pattern plays an increasingly important role in promoting urban spatial development. Using VGI and social media activity data, this article developed a method to automatically extract and identify urban spatial patterns and functional zones. We obtained a total of 167,532 Weibo POI data points in Beijing from 13 April to 17 April 2015, OpenStreetMap road network data on 14 April 2015, and China’s 1 km grid population data set. Then, we used the hierarchical clustering algorithm, TF-IDF method, and improved k-means clustering algorithms and identified eight functional zones. The functional zones included the diplomatic and political zone, science and education zone, mature residential zone, new residential zone, commercial and entertainment zone, tourist attractions zone, areas to be developed, and unclassified areas. Finally, we verified the results of the study with the Beijing city master plan and typical areas, and the comparison shows that the clustering results had high accuracy.
当前，城市居民提供了大量的 VGI，对城市空间格局的理解在推动城市空间发展方面发挥着越来越重要的作用。利用 VGI 和社交媒体活动数据，本文开发了一种自动提取和识别城市空间模式和功能区的方法。我们获得了 2015 年 4 月 13 日至 4 月 17 日北京共 167,532 个微博 POI 数据点、2015 年 4 月 14 日的 OpenStreetMap 路网数据以及中国 1 公里网格人口数据集。然后，我们使用分层聚类算法、 TF-IDF 方法和改进的 k-means 聚类算法并确定了 8 个功能区。功能区包括外交政治区、科教区、成熟住宅区、新住宅区、商业娱乐区、旅游景区、待开发区和非分类区。最后，结合北京市城市总体规划和典型区域对研究结果进行验证，对比结果表明聚类结果具有较高的准确率。

The contributions of this work lie in three aspects. Firstly, the feasibility of using user-generated social media data on investigating urban spatial structures was verified. This kind of VGI data are large-volume, easily obtainable, more time-saving, and more people-oriented than traditional datasets, and their application in delineating city functional zones could provide more detailed information. Secondly, by dividing research units using the road network, we obtained the natural areas in Beijing. This street map segmenting method was more consistent with urban function division and was more effective in depicting city heterogeneities than was the urban uniform grid. Lastly, the automatically-identified urban functional zones using social media data provided more information than did generally-defined residential or employment areas. The advantage of mature residential zones over new residential zones provides us with useful information for the future planning of the newly developed areas and areas to be developed, so that sustainable development might utilized for the creation of well-developed center zones. In general, the use of Weibo POI data and OpenStreetMap road network data combined with spatial clustering methods to analyze the urban spatial structure and explore functional areas, provides new ideas for the study of urban spatial structure.
这项工作的贡献在于三个方面。首先，验证了使用用户生成的社交媒体数据调查城市空间结构的可行性。与传统数据集相比，这类 VGI 数据体积大、易获取、更省时、更以人为本，在城市功能区划定中的应用可以提供更详细的信息。其次，通过使用路网划分研究单位，我们得到了北京的自然区域。这种街道地图分割方法与城市功能划分更一致，并且在描绘城市异质性方面比城市统一格网更有效。最后，使用社交媒体数据自动识别的城市功能区比一般定义的住宅或就业区提供了更多的信息。成熟住宅区相对于新住宅区的优势为我们新开发区和待开发区域的未来规划提供了有用的信息，以便利用可持续发展来创建发达的中心区。总体上，利用微博 POI 数据和 OpenStreetMap 路网数据结合空间聚类方法分析城市空间结构，探索功能区，为城市空间结构研究提供了新思路。