-
As the capital city of Guangdong Province located in the Pearl River Delta, Guangzhou is a representative example of the rapidly growing and large coastal cities of China. It covers an area of 7249 km2 with a total population of over 15 million in 2019. Guangzhou City was selected as the research scope for this study. The selected areas include Liwan, Yuexiu, Tianhe, Haizhu, Baiyun, Huangpu, and Panyu districts, which cover the major urban areas of Guangzhou (Fig. 1). Among them, the central urban area, comprising Liwan, Yuexie, Tianhe, and Haizhu districts, enjoys the advantages of transport accessibility, scientific and technological innovations, and comprehensive service provisions, compared to other districts of Guangzhou City (Wei et al., 2021).
-
As the largest search engine and website in China, Baidu provides a variety of location-based service (e.g., Baidu Search, Baidu Map, Baidu Weather). Since 2011, Baidu has begun to provide access to aggregated information on the spatial distribution of Baidu App users via the public service BHM. According to its official definition, BHM is a digital map in which the geographical location information of Baidu App users at a certain time point are projected and different colors are used to show the user distribution in a region (Li et al., 2019; Wang and Chang, 2020; Yang et al., 2021). As a measure of population density, BHM data have been widely used to measure the movement of people across urban space and urban vibrancy (Wu and Ye, 2016; Yang et al., 2021). Compared with the conventional population density derived from the Census data which does not vary within a day, BHM updates every 15 minutes capturing the real-time dynamic information about crowd distribution. With several hundred million Baidu mobile application users (Wang and Loo, 2019), BHM data have great potential to provide significant information regarding population density across time and space (Li et al., 2019; Wang and Chang, 2020; Yang et al., 2021). Fig. 2 illustrates an example of a BHM of Guangzhou City at 18:00 on December 4, 2019. Totally, 252 BHMs were collected. Adopted from Tan et al. (2016) and Yang et al. (2021), BHMs were loaded into ArcGIS 10.3, and BHI, a measure of population density, was calculated based on the pixel data of each unit (0.1 km × 0.1 km) and the quantitative relationship between color and population density as defined by BHM. At the community level, we also compared the BHI at 11:00 pm (when the vast majority of people are staying at home) to the population density derived from the recent Census data and found that the two variables have a significantly high level of correlation (0.90), suggesting that using BHI is acceptable for measuring urban vibrancy and its spatiotemporal dynamics. Noted that the validity of BHI as a proxy for urban vibrancy depends on our definition of the seminal definition by Jacobs (1961) that urban vibrancy is the presence of street life over a 24-hour period.
-
The built environment characteristics are represented by POIs data, floor area and land use data. POIs data measure both the intensity and diversity of activity destinations (He et al., 2018; Yang et al., 2021). Compared with land-use data, POIs data have much finer statistical granularity and thus show greater flexibility for studying at various scales. Moreover, human activity can be better presented by their interactions with POIs rather than by land-use type (Loo and Wang, 2018). In this study, POIs data obtained from AutoNavi, a popular web-mapping platform and location-based service provider in China, are used to reflect the built environment. A total of 344 829 POIs in Guangzhou City in 2018 were collected. According to the classification of AutoNavi, POIs were classified into 12 types. These 12 types were further grouped into consumption-related POIs (CPOI), housing-related POIs (HPOI), traffic-related POIs (TPOI), and other POIs (OPOI), as shown in Table 1.
Category Original types of POIs CPOI Shopping service, catering service, life service, recreation and entertainment service, accommodation service HPOI Residential district (community names, apartments, residential quarter, etc.) TPOI Traffic service (road, stations, bus stop, subway station, airport, harbor, etc.) OPOI Corporate business, medical and health service, financial service, education service, government and administrations Table 1. Four categories of points of interest (POIs)
Note that POI data are count data that do not differentiate the size of a facility, which affects its capacity to satisfy people’s activity requirements. For example, a shopping mall will provide more shopping opportunities than a convenience store does, thereby influencing people’s choices of shopping destination in a vastly different manner. We expect the introduction of total floor area, as a measurement of 3-dimensional built environment can help mitigate this drawback (Yang et al., 2021). The information on floor area and land use of Guangzhou City were collected from the 2020 Survey of Urbanization Evaluation, which is provided by the local planning administrative department with the land use and floor area map in a vector graphics file format.
-
Choosing an appropriate spatial unit is important due to the uncertain geographic context problem (Kwan, 2012). When the unit of analysis is too big or small, the geographical disparity of urban vibrancy may be masked (Wu et al., 2018; Wang et al., 2020a). In previous studies, a grid size of 0.5–2.0 km is usually used at the city scale (Wang et al., 2020a). Considering that a positioning error of about 0.1–1.0 km may exist when the phone signal or WiFi is not good, a 1 km × 1 km grid is used as the spatial unit to reflect urban vibrancy on a fine scale. Accordingly, Guangzhou City can be divided into 6565 grids.
For the temporal analysis, dividing a day into twelve 2-h time periods have been widely used in previous behavior studies (Zheng and Zhou, 2017; Wu et al., 2018). We also adopted this approach and examined the temporal dynamics of urban vibrancy in 2-h intervals within a day.
The average BHI in each spatiotemporal unit (i.e., a grid during 2-h interval) were used as the dependent variable (Table 2). The independent variables refer to the built environment of each grid, including two categories, namely, POIs-related variables and other variables. POIs-related variables include CPOI, HPOI, TPOI, and OPOI, while other variables consisting of location, land use mix, and building intensity.
Variable name Variable definition Mean SD Dependent variable (urban vibrancy) Weekday vibrancy Average Baidu Heat Index (BHI) in a grid during 2-h interval on a weekday 107.68 227.32 Weekend vibrancy Average BHI in a grid during 2-h interval on a weekend 149.49 352.79 Independent variable (built environment) Location Straight line distance from grid center to the CBD (Zhujiang New Town) / km 18.43 7.80 Land use mix The proportion of the six major land use types (i.e., commercial, residential, industrial, municipal administration, education, and public open space), calculated with the adapted entropy method by Song et al. (2013) 0.61 0.16 Building intensity Total floor area in the grid / million m2 0.17 0.18 CPOI Number of consumption-related POIs in a grid / (counts / km2) 57.99 115.50 HPOI Number of housing-related POIs in a grid / (counts / km2) 17.23 35.58 TPOI Number of traffic-related POIs in a grid / (counts / km2) 19.26 40.81 OPOI Number of other POIs in a grid / (counts / km2) 85.40 159.58 Table 2. Variable definition and sample statistics in Guangzhou City
-
To understand the spatially and temporally heterogeneous effects of the built environment on urban vibrancy, geographically and temporally weighted regression (GTWR) is adopted in this study. As a temporal extension of geographically weighted regression (GWR), GTWR examines the local relationship between independent and dependent variables in the time-space dimension (Huang et al., 2010). Compared to GWR which captures spatial nonstationarity (Brunsdon et al., 1996), GTWR provides excellent advantages in simultaneously addressing spatial and temporal heterogeneity. Given the nature of spatiotemporal dynamics, as discussed above, GTWR is chosen to model the relationship between the built environment and urban vibrancy. Specifically, the GWTR model can be defined as follows:
$$ \begin{split} \;{y_i} =& {\beta _0}\left( {{u_i},{v_i},{t_i}} \right){\rm{ }} + {\beta _1}\left( {{u_i},{v_i},{t_i}} \right){x_{i1}} + {\beta _2}\left( {{u_i},{v_i},{t_i}} \right){x_i}_2 + {\rm{ }} \ldots {\rm{ }} +\\ &{\beta _k}\left( {{u_i},{v_i},{t_i}} \right){x_{ik}} + {\varepsilon _i} \end{split}$$ (1) where the dependent variable yi refers to the urban vibrancy in spatiotemporal unit i on a weekday (or a weekend). (ui, vi, ti) is the coordinates of unit i in the time-space dimension (ui, vi and ti are the longitude, latitude and time, respectively). xik denotes the kth variable (k = 7) for unit i, and βk (ui, vi, ti) represents a set of parameter values at unit i. β0 (ui, vi, ti) is the intercept value, while εi is the unobservable disturbance term of unit i. Noted that the dependent variable and independent variables are transformed by logarithm to conform to the normality assumption. Specifically, ln (x+1) was adopted in the logarithmic transformation because some variables have values below 1. The estimated parameter can be explained as ‘elasticity’, a measurement of the percentage changes of one variable in response to a change in another (Yang et al., 2018).
GTWR model estimates the local regression coefficients based on a weighting matrix built upon space-time distances between observed unit i and other observations (Huang et al., 2010):
$$\;\beta \left( {{u_i},{v_i},{t_i}} \right){\rm{ }} = {\rm{ }}{\left[ {{x^T}W\left( {{u_i},{\rm{ }}{v_i},{\rm{ }}{t_i}} \right)x} \right]^{-1}}{x^T}W\left( {{u_i},{\rm{ }}{v_i},{\rm{ }}{t_i}} \right)y $$ (2) where the weighting matrix W (ui, vi, ti) is an n × n diagonal matrix, i.e., diag (Wi1, Wi2,…Wij,…Win). Wij (1 ≤ j ≤ n) refers to the space-time distance decay function, determined by the space-time distance (dst) and bandwidth h. The main assumption is that the closer measurements to unit i in the space-time coordinate system have higher weight in predicting βk. By contrast, the GWR model only considers the spatial distance and models the variety of spatial relationship (Brunsdon et al., 1996). According to Huang et al (2010), the space-time distance
$d_{ij}^{st}$ is defined as:$$\; d_{ij}^{st} = \sqrt {\gamma {{\left[ {({u_i} - {u_j}} \right)}^2} + {{({v_i} - {v_j})}^2}] + \delta {{({t_i} - {t_j})}^2}} $$ (3) where
$ \gamma $ and$ \delta $ are the weights for harmonizing the influence of differing units between space and time. In this study, a common Gaussian distance decay functions and Euclidean distance are adopted to calculate the weighting matrix with the greatest efficiency:$$ \;{W_{ij}} = {\rm{exp}}\left[ { - \frac{{{{\left( {d_{ij}^{st}} \right)}^2}}}{{{h^2}}}} \right] $$ (4) where h denotes a nonnegative parameter named the space-time bandwidth, which can be acquired via the use of Akaike information criterion (AIC) (Hurvich et al., 1998):
$$\; AIC = 2k + n {\rm{ln}}(RSS) $$ (5) where k is the number of estimated parameters in the model (k = 7), n refers to the number of units, and RSS is the Root-Sum-Squares. AIC deals with the trade-off between the goodness of fit and the simplicity of the model. For model comparison, the lower the value for AIC, the better the fit of the model (Hurvich et al., 1998).
-
Prior to estimating each regression model, a Pearson correlation analysis and a variance inflation factor test were conducted. Results suggest that correlation between independent variables are low (below 0.40) and/or statistically insignificant and multicollinearity is not a problem in this study. Table 3 summarizes the performance statistics of OLS, GWR, and GTWR models for explaining the variations in urban vibrancy on weekdays and weekends, respectively. In all three models, all built environment variables listed in Table 2 are significant at the 1% level and exhibit the expected signs. Generally, being closer to the CBD, having higher levels of building intensity and land use mix, and concentrating more POIs of various functions contribute to urban vibrancy.
Performance statistics Weekday vibrancy Weekend vibrancy OLS GWR GTWR OLS GWR GTWR R2 0.396 0.513 0.754 0.348 0.474 0.736 Adjusted R2 0.394 0.460 0.731 0.346 0.416 0.711 AIC –6154.953 –6009.452 –5311.210 –7067.941 –6925.521 –4186.150 Notes: Ordinary least squares (OLS), Geographically weighted regression (GWR), Geographically and temporally weighted regression (GTWR), Akaike information criterion (AIC) Table 3. Performance of OLS, GWR, and GTWR models in Guangzhou City
However, it is important to note that the performance statistics vary significantly among the OLS, GWR, and GTWR models. Specifically, GTWR models have the highest explanatory power, explaining 73.1% and 71.1% of the variations in urban vibrancy on weekdays and weekends, respectively. On the contrary, OLS models explain the lowest percentages of the variations (39.4% and 34.6% respectively). Besides, GTWR models also have the lowest values of AIC. The comparisons confirm our hypothesis that the evolution of urban vibrancy is influenced by built environment variables that are heterogeneous across both time and space. Moreover, the Moran’s I values of urban vibrancy for weekdays and weekends are 0.495 and 0.459 (P-value < 0.001), respectively, suggesting that urban vibrancy has positive spatial autocorrelation and noticeable features of spatial clustering. Moran’s I index has been widely adopted as a measure of spatial autocorrelation (Huang et al., 2010; Wu et al., 2018a).Based on this, it is reasonable to believe that ignoring the spatial and temporal effects on urban vibrancy would lead to biased estimates of the associated environmental correlates at the local level. Therefore, it is important for urban planners and city governments to have a better understanding of the local geography on the spatiotemporal relationships between the built environment and urban vibrancy, which enables the formulation of more pertinent, targeted and effective strategies/actions in fostering and maintaining urban vibrancy.
-
Tables 4 and 5 summarize the GTWR results for urban vibrancy on weekdays and weekends, respectively. Obviously, the regression coefficients vary in the time-space dimension, as shown in their quartile distribution. Generally, the variation trends, signs, and degrees of built environment variables are roughly the same on weekdays and weekends. We compared the estimation and interpretation of coefficients based on their median values here. Compared to the mean, the median is proved to be more robust to extremely large or small values.
Variable Min. Lower quartile Median Upper quartile Max. Location –9.924 –3.054 –0.234 0.013 0.928 Land use mix –3.677 –0.138 0.314 2.965 5.636 Building intensity –1.611 –0.009 1.425 4.864 7.275 CPOI –3.257 –0.027 0.182 2.533 4.466 HPOI –2.990 –0.019 0.078 1.621 1.628 TPOI –4.279 –0.024 0.189 2.895 5.309 OPOI –2.006 –0.035 0.016 0.053 1.231 Note: The meaning of CPOI, HPOI, TPOl and OPOl are same as in Table 2 Table 4. Geographically and temporally weighted regression (GTWR) results on associations between built environment and urban vibrancy on weekdays in Guangzhou City
Variable Min. Lower quartile Median Upper quartile Max. Location –9.133 –3.591 –0.132 0.360 1.900 Land use mix –3.104 –0.213 0.561 3.060 5.815 Building intensity –1.825 –0.003 1.768 5.602 7.869 CPOI –3.031 –0.034 0.198 2.063 4.045 HPOI –2.024 –0.026 0.129 1.753 3.187 TPOI –3.288 –0.029 0.173 2.655 3.505 OPOI –2.825 –0.044 0.022 0.057 1.756 Note: The meaning of CPOI, HPOI, TPOl and OPOl are same as in Table 2 Table 5. Geographically and temporally weighted regression (GTWR) results on associations between built environment and urban vibrancy on weekends
The total floor area plays a dominant role in contributing to urban vibrancy, suggesting that increasing building intensity is an effective tool in attracting people and their associated activities. This result is consistent with the findings of earlier studies that the concentration of activity opportunities makes a place more attractive (Jacobs, 1961; Ye et al., 2018). Land use mix is the second most statically positive factor in accounting for urban vibrancy, further verifying that diversity of activity opportunities is highly associated with urban vibrancy (Jacobs, 1961; Ta et al., 2020). High level of land use mix refers to a combination of commercial, residential, institutional, or industrial use, which can offer more attractions to people. Thus, great potential exists to improve urban vibrancy through enhancing land use mix. With negative signs, location is the third most statistically significant factor in influencing urban vibrancy. As distance to the city center increase, the urban vibrancy decrease, all else being equal. This finding coincides with the observations in earlier studies of Guangzhou’s strong urban center in its monocentric spatial structure (Xu and Yeh, 2003). The positive signs of the densities of POIs with various functions suggest that the concentration of either four types of activity opportunities contributes to urban vibrancy. Specifically, CPOI and TPOI have greater impacts on urban vibrancy compared to HPOI and OPOI, suggesting that entertainment and transportation facilities are of importance for Guangzhou people’s daily lives. The local differences of these key associated environmental correlates (i.e., location, land use mix, building intensity, CPOI and TPOI) in the time-space dimension are visualized and analyzed in the following section.
-
Figs. 3 and 4 show the average temporal and spatial change tendencies of the coefficient, using their median values of each temporal and spatial unit respectively. The differences of these key associated environmental correlates between weekdays and weekends are also illustrated. For the temporal variations, the solid and dashed lines represent weekdays and weekends respectively (Fig. 3). For the spatial variation, on the basis of Jenks natural breaks classification we manually set zero as a threshold to distinguish between the positive and negative effects (Fig. 4).
Figure 3. Temporal trends of key built environment correlates of urban vibrancy on weekdays and weekends. The meaning of CPOI and TPOI are same as in Table 2
Figure 4. Spatial distribution of key built environment correlates of urban vibrancy on weekdays and weekends in Guangzhou City. The meaning of CPOI and TPOI are same as in Table 2
Temporally, the variation trends of location for weekdays and weekends are similar. However, the degrees of location are different. Specifically, during 8:00–22:00, location has greater negative elasticity on weekdays than on weekends as the dashed line is located above the solid line. By contrast, the dashed line is located below the solid line in other 2-h intervals. This result can be related to the urban spatial structure development of Guangzhou City. Based on the spatial policy of ‘expansion in the south, optimization in the north, advance in the east, and linkage in the west’ since 2000, Guangzhou City has experienced rapid urban expansion (Xu and Yeh, 2003). As expected, a polycentric spatial structure will be formed to accommodate the growing urban population and their activities. However, other than a relatively dispersion of residence, there is still a high concentration of activity opportunities in relation to employment, leisure and recreation, and various civil obligations in the central urban area. Therefore, on weekdays, most people are busy at work and their employment activities mainly happened in the central urban area. And their non-employment activities are more likely to be organized around the workplace during the daytime. This fact also explains a ‘dip’ of location appearing during nighttime hours. Spatially, the relatively monocentric structure leads to the mostly negative effects of location on urban vibrancy. It is interesting to note ‘pockets’ of positive values at each district center on both weekdays and weekends. These areas with relatively higher concentrations of activity opportunities in the district are attractive to residents nearby. However, these areas are still far from the CBD and central urban area regarding urban vibrancy. Urban spatial structure, typically characterized by monocentric or polycentric development patterns, has been confirmed by some previous studies to have significant effects on urban vibrancy (Chen et al., 2019; Wang, 2021). The existence of multiple urban subcenters can lead to a greater diversity of urban amenities and hence can attract people to dine out, do shopping, or recreate, thus generating persistent vibrancy. Therefore, to achieve a high level of polycentric spatial structure that is beneficial to urban vibrancy, the city government should pay more attention to the comprehensive development of these areas through well-equipped urban amenities and facilities.
With higher degrees of effects on weekends than on weekdays, the temporally variation trends of land use mix, building intensity, and CPOI for weekdays and weekends are similar. The results further suggest that greater potential exists to improve the capacity of these built environment variables for urban vibrancy on weekends than on weekdays. Specifically, the peak values appear during nighttime hours (i.e., 18:00–22:00) when people are typically engaging in leisure, entertainment, and dinner parties (Wang et al., 2015b; 2020). Therefore, places with higher level of land use mix, building intensity, and consumption-related facilities can better satisfy people’s various activity requirements and thus become more attractive. The fact that higher coefficients of land use mix, building intensity, and CPOI on the weekends than those on weekdays is also consistent with people’s everyday lifestyle. Typically, most people face fewer time-space constraints and thus have more time to conduct non-employment activities on weekends. In contrast, on weekdays, most people typically commute between their home and workplace and do little other than work. Spatially, the effects of land use mix, building intensity, and CPOI on urban vibrancy are mostly positive on both weekdays and weekends. It is also interesting to highlight that some grids in the old city area (i.e., Yuexiu, Liwan, and the southern Baiyun), which has almost the highest mixed function, tend to have the greatest elasticity of land use mix on urban vibrancy. In addition, the high elasticity of CPOI on urban vibrancy are mainly distributed in the CBD and other district center. This uneven spatial distribution of the parameter estimates also suggest that the associations between the built environment and urban vibrancy may be non-linear, which deserves further investigation.
For TPOI, the average temporal and spatial change tendencies and degrees of effects on urban vibrancy for weekdays and weekends are different. On weekdays, the peak values appear at 8:00–10:00 and 16:00–20:00 when people commute between home and workplace. The consistence between this temporal variation trend and the time characteristics of commuting is highly related to the increasingly prominent home-work separation during the formation of multi-center spatial structure in Guangzhou City. Because people usually face higher level of time-space constraints of trying to get to work on time during the morning peak hours than that during the evening peak hours, the highest value is at 8:00–10:00. In contrast, on weekends, people face relatively low level of time-space constraints during the morning peak hours and usually have non-employment activities near their residence; thus, TPOI has relatively low, although positive, elasticity on urban vibrancy, especially during the morning peak hour. Spatially, the effects of TPOI on urban vibrancy are also mostly positive on both weekdays and weekends. However, note that the effects are greater in grids with better public transport accessibility (particularly metro stations), confirming the important role of public transport in contributing to urban vibrancy (Wang et al., 2015b).
Elaborating Spatiotemporal Associations Between the Built Environment and Urban Vibrancy: A Case of Guangzhou City, China
doi: 10.1007/s11769-022-1272-6
- Received Date: 2021-03-15
- Accepted Date: 2021-08-20
- Available Online: 2022-04-28
- Publish Date: 2022-05-05
-
Key words:
- urban vibrancy /
- Baidu Heat Map (BHM) /
- built environment /
- heterogeneity /
- geographically and temporally weighted regression
Abstract: This study applies multi-source datasets (i.e., Baidu Heat Map data, points of interest (POIs) data, and floor area and land use data) and geographically and temporally weighted regression (GTWR) models to elaborate the spatiotemporal relationships between the built environment and urban vibrancy on both weekdays and weekends, using Guangzhou City as a case. First, we verified the spatially and temporally nonstationary nature of the built environment correlates, which have been largely ignored in previous studies based on local regression techniques. The spatially and temporally heterogeneous effects of the built environment on urban vibrancy are then presented and visualized, based on the GTWR results. We found that the elasticity of location (i.e., distance), land use mix (i.e., diversity), building intensity and numbers of POIs with various functions (i.e., density) are different across time (2-h intervals within a day) and space (grids), due to people’s everyday lifestyle, time-space constraints, and geographical context (e.g., spatial structure). The findings highlight the importance of a better understanding of the local geography on the spatiotemporal relationships for urban planners and local governments so as to put forward decision-making support for fostering and maintaining urban vibrancy.
Citation: | WANG Bo, LEI Yaqin, XUE Desheng, LIU Jixiang, WEI Chunzhu, 2022. Elaborating Spatiotemporal Associations Between the Built Environment and Urban Vibrancy: A Case of Guangzhou City, China. Chinese Geographical Science, 32(3): 480−492 doi: 10.1007/s11769-022-1272-6 |