Exploring land use functional variance using mobile phone derived human activity data in Shanghai

Abstract

Land use functions can categorize places where people perform different socioeconomic activities. This classification plays an important role in urban management, policy making, and resource allocation. However, due to the rapid changes of built environment and living demands, human activities might vary significantly, in space and time, even within the same land use function as conventionally defined, impeding the formulation of targeted and user-oriented planning policies. This study took the first step to explore land use subcategorization using mobile phone-derived human activities. The study area is the 5,298 census tracts in Shanghai. Sixteen million mobile phone users’ data were collected from Shanghai Mobile Co., Ltd., in 2014. We proposed a multi-dimensional indicator framework to capture collective features of activities and identified land use subcategories using the K-Means clustering method. Analysis of variance (ANOVA) was applied to detect the proportion of activity variances captured by the classification results. Subcategory labelling method was applied to reveal the relationship between land use subcategories and built environment factors. The results show that (1) Conventional land-use functional zones (LFZs) cannot fully capture the activity variances, especially in behavioral regularity and temporal variation; (2) According to the variance analysis, at least four to five subcategories should be identified upon current LFZs to capture the main activity variances; and (3) In the case of Shanghai, land use subcategories presented palpable spatial regularity, which revealed a citywide structure deserves for further study. We concluded that data-derived activity features can provide an innovative perspective complementary to existing land use classification standards and facilitate policymakers with their decision-making processes on urban resource allocation.

Keywords

land use subcategory activity variances mobile phone derived human activity K-means clustering Shanghai

Introduction

Land use functions categorize places where people perform different socioeconomic activities (Batty et al., 2013; White et al., 1997; Zhang et al., 2018), and its classification plays an important role in planning policy-making, urban management, and resource allocation (Guan et al., 2021; Lopez and Ferreira, 2020; Noyman et al., 2019; Ornetsmuller et al., 2018). The conventional Euclidean zoning defines discrete land use functions, such as residential, industrial, commercial, green space, etc. ( Li and Parrott, 2016). As cities expand and economies evolve, the relationship between land use and human activity becomes more complicated. Increasing human mobility is weakening the boundary of land use functional zones (LFZs), thus causing alteration of human activities even within the same LFZ and making the daily regularity more difficult to capture (Cottineau et al., 2019; Sagl et al., 2014). Therefore, to facilitate better future urban planning and address functional variance, land use classification should be reconsidered from a human activity perspective (Pan et al., 2020; Xiao et al., 2019; Yang, 2020).

In recent years, location-based service (LBS) technology has been adopted to provide a lower cost, larger scale, and higher granularity data source for behavioral research (Wang et al., 2018). Existing studies have built a series of human activity indicators based on social media, smart card, and mobile phone data. Some studies linked data-derived activity features to land use classification (Guo et al., 2019; Pei et al., 2014). For example, Cao et al. (2019), Wu et al. (2017), and Zhang et al. (2021) applied supervised machine learning methods to inferring current land use functions to reduce the cost of field investigation. Other studies, Pei et al. (2014) and Gao et al. (2017), focused on exploring new functional zones based on clustering algorithms, as a comparison to conventional ones. Despite a growing base of literature, to what extent conventional LFZs can capture human activity variances remains unclarified. Moreover, how to identify subcategories upon the current land use zoning system remains understudied.

This study used mobile phone data to explore land use subcategories of 5,298 census tracts in Shanghai. The human activity information of 16 million mobile phone users was collected by Shanghai Mobile Co., Ltd., from March 15 to 28 March 2014. Individual trip chains were extracted from the raw data to construct multi-dimensional activity indicators. Land use subcategories were identified using the K-Means clustering method based on these indicators. Then, analysis of variance (ANOVA) was applied to detect the proportion of activity variance explained by the classification results. Finally, a subcategory labelling method was applied to reveal the relationship between land use subcategories and built environment factors. We try to answer the following questions: (1) To what extent can conventional land-use functional zones (LFZs) capture human activity variances? (2) How many subcategories should be identified for each land use function? (3) Whether these land use subcategories have differentiated spatial distributions and built environment characteristics?

The rest of the paper is organized as follows: Section 2 briefly introduces conventional land use classification approaches, human activity studies using mobile phone data, and emerging activity-based classification approaches. Section 3 describes data sources and presents methods used for classification of land use subcategories. Section 4 presents the results of our case study in Shanghai, including activity variances between land use functions and the results of subcategory identification. Section 5 discusses contributions to the existing literature and urban policy implication. The conclusion section presents limitations and future work.

Literature review

Conventional land use classification

As a critical instrument of urban planning, earlier targets of land use classification were to integrate compatible land uses and reserve buffer areas between incompatible land uses. For example, land parcels were labelled as various functions including residential, commercial, industrial, and green space ( Li and Parrott, 2016). Later, land use classifications were not limited to categorizing physical materials at land surface ( buildings, roads, vegetation, water, etc.). Instead, they were also used to categorize places where people perform different socioeconomic activities for regulation (Alberti et al., 2004; Zhang and Foody, 2001). Land use function zones (LFZs), under the Euclidean zoning ordinances, could be spatially aggregated by land parcels and their categories were semantically abstracted from land use functions (commercial zones, residential zones, industrial zones, etc.) (Zhang et al., 2018).

Classifications of LFZs in empirical studies are greatly impacted by geographic units and algorithms determining the dominant land use function, though the functions of land use parcels have already been labelled in the master plan. As for the geographic units, researchers have aggregated land parcels into administrative units (Liu et al., 2020a), traffic analysis zones (TAZs) (Wang et al., 2016), and self-defined grids (Cao et al., 2019) according to their research scopes. As for determining the dominant land use function, though proportion approach is widely used (Cao et al., 2019; Wang et al., 2016; Wu et al., 2017), some studies choose to apply normalization approaches (Liu et al., 2020b) considering unbalanced land use proportions (e.g., commercial parcels are much fewer than residential parcels and green space parcels in suburbs) and difference between cities (e.g., small cities have fewer commercial parcels compared to large cities). For instance, Liu et al. (2020b) defined Land Use Dominance Index (LUDI) and Shannon Entropy Index (Lin, 1991) to assign five main land use functions to 91 sub-districts, which is quite inspirable to our study.

Human activity studies using mobile phone data

In light of changing urban economic structure, flexible working schedule, and diverse recreational and shopping attractions, some studies argued that how space is actually used cannot be fully captured by the conventional simple and broad zoning classifications (Li et al., 2018; Lopez and Ferreira, 2021; Wu et al., 2017; Yang, 2020). Mobile phone data were deemed suitable for human activity studies by providing individual activity information of massive user numbers (Zhou et al., 2020). We summarized human activity studies using mobile phone data and the proposed indicators. The relevant literatures are listed in Table 1.

Table 1.

Studies on human activities using mobile phone data.

Dimension	Author	Research contribution
Spatiotemporal distribution of human activity
Activity intensity	Sagl et al. (2014)	Used GIS technology to visualize the spatial distribution of crowd activities in different periods
	Zhang et al. (2020)	Proposed machine learning method for the estimation of spatial temporal population density using mobile phone records data
	Chen et al. (2020)	Explored differences in activity densities across those groups of community members at a specific POI revealed by mobile phone records data
Temporal variation	Sevtsuk and Ratti (2010)	Proposed that the density repeats in four cycles: 8 h, 12 h, 24 h, and 3.5 days
	Zhong et al. (2017)	Found that there was a density exchange between city center and suburb residential from day to night
	Metulini and Carpita (2020)	Built spatio-temporal indicators for city users based on mobile phone signals and administrative data
	Guan et al. (2021)	Find notable seasonal variations of visitor volume and park service area existed in the medium-sized urban parks
Location of individual activity category
Spatial coverage	Chen et al. (2016)	Calculated trip distance and activity range at the individual level to measure social equity in different residents
	Wang et al. (2018)	Introduced displacement and radius of gyration to measure the area individual can cover for his/her daily activities
	Paipuri et al. (2020)	Estimated trip lengths and path flow distributions in a multi-region setting using mobile phone data
	Zhao et al. (2020)	Proposed a semi-supervised approach to predict individual socioeconomic status based on users’ range of activities
Behavioral regularity	Duan et al. (2017)	Included activity space, activity points, and daily trip-chain patterns to measure the stability of residents’ activities
	Wang and Ren (2019)	Used day frequency of an individual activity happened at the same place to identify random and daily activities
	Qian et al. (2020)	Proposed a Gaussian mixture model, and measured regularity by calculating the probability density of daily activities

Spatial temporal distribution of collective human activities

Activity intensity and temporal variation are the two key dimensions of collective human activity patterns (Sagl et al., 2014; Sevtsuk and Ratti, 2010). In terms of activity intensity, Sagl et al. (2014) used the Geographic Information Science (GIS) technology to visualize the spatial distribution of crowd activities at the city level. Building on this, Wang and Ren (2019) transformed activity intensity into population density, and found that the distribution of population identified by mobile phone data revealed a stronger spatial aggregation than census population. Also, machine learning models were applied to estimate the activity intensity using mobile users’ trajectories (Zhang et al., 2020).

The temporal variation of activity intensity is also an important issue. Sevtsuk and Ratti (2010) analyzed the time regularity of residents' activity intensity in Rome, and proposed that the intensity repeats in four different time cycles. Zhong et al. (2017) constructed an “activity-time” analysis framework and found that there was a density exchange between the city center and suburban residential areas from day to night. More recently, Metulini and Carpita (2020) combined mobile phone signals and administrative data to build spatiotemporal indicators for city users. Guan et al. (2021) combined mobile phone big data analyses and field observations to reveal seasonal variations of visitor volume and park service area existed in the medium-sized urban parks.

Exploration of individual activity patterns

The application of mobile phone data also enabled the exploration of human activities at the individual level (Yang, 2020). Related indicators can be further concluded into spatial coverage and behavioral regularity.

The spatial coverage range is usually measured by displacement and radius of gyration (Chen et al., 2016; Wang et al., 2018). Displacement refers to the distance between staying points, showing how far individuals travel for an activity. Radius of gyration is an area indicator that shows the area individuals can cover for their daily activities. Moreover, some studies aggregated individual activity information into zones, and further estimated trip lengths and path flow distributions in a multi-region setting (Paipuri et al., 2020; Zhao et al., 2020).

The behavioral regularity has been used to measure the frequency of an activity, in other words, whether an activity of a resident follows a fixed routine. Existing studies focus on trip-chain-derived features and use classification methods to find individual mobility patterns (Qian et al., 2020). For instance, Duan et al. (2017) included activity space, activity points, and daily trip-chain patterns to measure the stability of residents’ activities. Wang and Ren (2019) classified resident’s activities using daily frequency occurring at the same place and found notable variance between the distribution of daily activities and random activities. Qian et al. (2020) proposed a Gaussian Mixture Model to measure regularity by calculating the probability density of daily activities.

Activity-based land use classification

The methods of land use classification using mobile phone data can be divided into supervised and unsupervised. The supervised classification aims to accurately infer current land use functions such as commercial, industrial, residential, and green space (Cao et al., 2019; Wang et al., 2016; Wu et al., 2017). The goal is to substitute onsite surveys and field observations, which can be labor intensive. However, the supervised classification can be problematic when applied to large metropolitan areas where land uses are highly-mixed, since it assumes that human activities are similar in the same land use function (Wu et al., 2017).

The unsupervised classification aims to explore latent land use categories, which can be beyond the current land use classification system. Instead of land use function, research on unsupervised land use classification emphasizes on human activity variances among proposed categories. For instance, Pei et al. (2014) constructed a vector composed of the total call volume and normalized hourly call volume, and applied the FCM method to infer land-use types in Singapore. Gao et al. (2017) combined POIs data and social network data to extract urban functional regions and applied the K-Means clustering method to find places with similar activity features. Zhong et al. (2017) built “Time—Activity Density” curves to classify land use functions according to day/night and workday/weekend differences in Shanghai.

Although earlier efforts have proved the feasibility of applying activity indicators to land use classification, they either focused on inferring current land use functions or building a new classification system ignoring the existing one. To what extent can conventional land-use functional zones (LFZs) capture human activity variances remains unclarified and how to identify subcategories upon current land use zoning system remains understudied. Hence, our study adopted unsupervised classification methods to identify subcategories within the current land use function.

Methodology

Data collection

The mobile phone data used in this study was collected from Shanghai Mobile Co., Ltd between March 15th and 28th in 2014. The two-week dataset recorded mobile phone activities from over 16 million phone users. The detailed information is in Supplementary Material Section 1. Supplementary Table S1 displays an example of the original mobile phone data. Supplementary Figure S3 shows the spatial distribution of the 37,461 mobile phone towers in Shanghai.

The land use data were collected and processed by the Natural Resources Bureau of Shanghai in 2014, we acquired access through institutional associations. In total, there were over 22,000 land use parcels in Shanghai, which can be divided into six main land use categories and 20 subcategories. The six main categories are Residential (R), Commercial and public (C), Industrial (I), Green space (G), Transportation (T), and Non-urban (N), as shown in Figure S1 in Supplementary Material.

Analytical methods

The analytical methods include three parts: (1) indicator calculation, (2) category classification, and (3) pattern labelling. Figure 1 shows the general framework.

Figure 1.

Flow chart of the proposed framework.

Calculation of the activity indicators

We adopted four aspects of human activity measures from existing studies: activity intensity, behavioral regularity, spatial coverage, and temporal variation. Activity intensity reflects the vitality of human activity, including activity density and activity/resident ratio. Behavioral regularity reflects the composition of human activity in each census tract, including proportion of random activity and stable activity identified through the type of mobile phone subscribers. Spatial coverage reflects the catchment area of census tract activities, including average trip distance and coverage ratio. Temporal variation reflects the variance of activity density in a census tract by time, including the day/night ratio and workday/weekend ratio. The eight activity indicators are listed in Table 2 and the basic method of calculations can be found in Supplementary Material Section 2 (Table 3).

Table 2.

Multi-dimensional human activity indicators.

	Indicator	Unit	Explanation
Activity intensity	Activity density (Ac_dens)	Person/km²	The number of activities in a census tract per day divided by its shape area
Activity intensity	Activity/resident ratio (Ac_ratio)	/	The number of activities in a census tract per day divided by its residents
Behavioral Regularity	Proportion of random activity (P_out)	%	The percentage of activities generated by outsiders in a census tract
Behavioral Regularity	Proportion of stable activity (P_daily)	%	The percentage of local residents’ daily activities in a census tract
Spatial coverage	Average trip distance (Tr_dist)	km	Average trip distance of all activities in a census tract
Spatial coverage	Coverage ratio (Cov_rate)	%	Percentage of origin census tracts linked by a destination census tract with more than 50 activities per day
Temporal variation	Day/night ratio (DN_ratio)	/	The ratio of hourly activity volume in the daytime and that at night
Temporal variation	Weekday/weekend ratio (WW_ratio)	/	The ratio of activity volume per weekday and that per weekend

Table 3.

Human activity indicators of each dominant land use functions.

	Activity density	Activity/resident ratio	Prop. of random activity, %	Prop. of stable activity, %	Average trip distance	Coverage ratio, %	Day/night ratio	Weekday/weekend ratio
R	24,628	2.02	2.75	44.33	5.57	5.14	1.31	2.18
C	31,597	2.56	3.50	44.12	5.71	6.03	1.46	2.48
I	5,013	1.93	3.05	44.01	7.15	3.21	1.50	2.56
G	14,181	2.14	3.30	43.02	6.80	4.64	1.46	2.50
T	7,483	2.03	4.32	41.08	7.78	3.83	1.46	2.52
N	590	1.64	4.00	39.91	9.38	1.50	1.56	2.72

Note: The unit of activity density is person/km² per day, the unit of average trip distance is km.

Land use subcategory classification

In general, a census tract contains multiple land parcels. Hence, we assigned a dominant land use function for each census tract to obtain conventional land-use functional zones (LFZs). Considering both the unbalanced proportions of land use functions in Shanghai and the feasibility to apply our method to other cities, we chose the entropy approach, similar to Liu et al. (2020b)’s study, to determine the dominant functions. A comparison between the results of the proportion approach and the entropy approach can be found in Supplementary Material Section 3. For the six main land use functions, we calculated the Shannon Entropy Index for each function by equation (1)

S E I_{i j} = \frac{F S_{i j} * \sum_{i = 1} S_{i}}{\sum_{i = 1} F S_{i j} * S_{i}}

(1)

where

S E I_{i j}

represents the Shannon Entropy Index of function j in census tract i,

F S_{i j}

represents the total area of parcels with function j in census tract i,

S_{i}

represents the area of census tract i. A large value of

S E I_{i j}

means the proportion of function j is relatively large in census tract i (compared to other census tracts). Finally, we standardized the value and define a Type Dominance Index (TDI) for each land use function in equation (2)

T D I_{i j} = \frac{S E I_{i j}}{\sum_{j = 1} S E I_{i j}}

(2)

where

T D I_{i j}

represents the TDI of function j in census tract i. The land use function with the largest TDI was identified as the dominant function.

Next, we chose the K-Means clustering algorithm to classify subcategories within each land use function. The advantage of the K-Means algorithm is that it treats every activity indicator evenly and that it is an unsupervised method revealing latent categories. Typically, Elbow Curve and Silhouette analysis are applied to determine the optimal number of clusters. We applied the K-Means method to the four main land use functions: Residential (R), Commercial and public (C), Industrial (I), and Green space (G).

Finally, we conducted an analysis of variance (ANOVA) to calculate the proportion of activity variance captured by the classification and determine the proper number of clusters. Analysis of variance (ANOVA) is a statistical approach used to compare variances across the means (or average) of different groups (Lars and Wold, 1989), in which the total variance, sum of squares in total (SST), is divided into sum of squares between groups (SSB) and sum of squares within groups (SSW). Typically, the outcome of ANOVA is the “F statistic,” the ratio of SSB and SSW. In this study, the ratio of SSB and SST was reported because our focus is on the proportion of activity variance captured by land use classification. For each activity indicator k, the proportion of captured variance $P C V_{k}$ is defined as equation (3)

P C V_{k} = S S B_{k} / S S T_{k} = \sum_{c = 1}^{C} {({\bar{x}}_{k, c} - {\bar{x}}_{k})}^{2} / {\sum (x_{k, i} - {\bar{x}}_{k})}_{i = 1}^{I}^{2}

(3)

where

x_{k, i}

denotes the value of activity indicator k in census tract i,

{\bar{x}}_{k}

denotes the mean value of the indicator in all census tracts;

{\bar{x}}_{k, c}

denotes the mean value the indicator in census tracts belonging to category c. A high PCV value indicates a large proportion of activity variance captured by the classification. For consistency, our study used PCV value to determine the optimal number of subcategories identified by K-Means algorithm with two standards: (1) The mean value of

P C V_{k}

should be larger than 50%; (2) The mean value of

P C V_{k}

starts to grow slowly as the number of clusters increases. The advantages of using our standards are to track the extent to which the variance of each indicator is captured by the classification. These standards are basically in line with Elbow Curve. A comparison between the methods can be found in Supplementary Material Section 3.

Subcategory land use labeling

We integrated three features to label identified subcategories: human activity indicators, spatial distribution, and the built environment factors: (1) The mean value of eight activity indicators of each subcategory were calculated and shown in radar charts. Each angle of the radar chart represented an activity indicator, normalized to 0–1; (2) Spatial distributions of subcategories were drawn in ArcGIS to reveal their location, and; (3) Built environment factors of census tracts were calculated using land use, POI, Gaode Map, and Shanghai Census data, as shown in Supplementary Material Table S3. The Percentage Difference Index (PDI) is used to reflect the difference of a subcategory compared to others within the same land use function. The PDI is calculated as equation (4)

P D I_{k, c} = \frac{{\bar{x}}_{k, c} - {\bar{x}}_{k, f}}{{\bar{x}}_{k, f}} * 100 %

(4)

where

P D I_{c, f}

denotes the PDI of built environment factor k of subcategory c,

{\bar{x}}_{k, c}

denotes the mean value of factor k in census tracts belonging to subcategory c,

{\bar{x}}_{k, f}

mean value of factor k in census tracts belonging to main land use function f, c is a subcategory of function f. A large value of PDI indicates that a subcategory has a noticeable difference with other subcategories in built environment factors.

Results

Activity variance captured by conventional land use functions

Table 2 summarizes the average value of the eight human activity indicators of census tracts within the six land use functions. The spatial distribution of land use-functional zones (LFZs) can be found in Supplementary Material Figure S6. We found several interesting relationships between conventional land use functions and human activity: (a) Activity intensity varied greatly among land use functions. Commercial and Public census tracts had an average activity density of 31,597 person/km² per day, while the density was below 15,000 person/km² per day in Industrial, Green Space, Transportation, and Non-urban census tracts. (b) Spatial coverage of activity also presented differences among land use functions. A Commercial and Public census tract had connections with 6.03% of all census tracts in the city, while the number for a Residential census tract was only 1.50%. (c) However, indicators of behavioral regularity and temporal variation changed slightly among land use functions, indicating obvious differences between human activity dimensions. Figure 2(a) presented the normalized value of eight indicators, which revealed that some activity dimensions were differentiated by conventional LFZs while others were not.

Figure 2.

Activity variation within six land use functions.

To quantitatively measure the proportion of activity variance captured by conventional land use functions, we calculated the values of PCV for eight activity indicators, as shown in Figure 2(b). The mean value of PCV equals to 15.42%, indicating the current land use function can only explain a small portion of the human activity variance. Although plenty of literature has argued that current land use zoning cannot fully capture human activity features (Jiang et al., 2017; Lopez and Ferreira, 2021; Yang, 2020), our study is one of the first attempts to provide quantitative measures. Among the eight indicators, the PCV values of activity density and average trip distance were relatively higher (28.06% and 23.95%, respectively) than proportion of random and stable activity (5.75% and 5.52%, respectively). This indicates that land use functions could be helpful in inferring collective density and trip demands. However, new categories are necessary to capture the diversity of individual mobility patterns.

Results of subcategory land use classification

We classified Residential, Commercial and Public, Industrial, and Green Space into subcategories using human activity indicators as mentioned above (classification results of Industrial, and Green Space can be found in Supplementary Material).

Classification results of residential (R) function

For the Residential census tracts, Figure 3 shows the values of PCA given the number of clusters from 2 to 10. According to the rules, four subcategories should be identified for Residential function. The PCV values of temporal variation increase quickly as the number of clusters grows, with nearly 70% of the variance of day/night ratio and weekday/weekend ratio captured by four subcategories. Figure 4 shows the spatial distribution, activity radar chart, and the built environment PDI of the four subcategories of residential land use: (a) Subcategory 1 can be labeled as Residential area in South Central City. Census tracts in this subcategory were mainly distributed in the southern part of central Shanghai inside the outer ring road, most of them were high-density residential areas, with higher activity intensity, short trip distance but large spatial coverage ratio. As for their built environment factors, the average distance to the city center was shorter than other census tracts, and the number of amenities and the area of Commercial and Public parcels were higher than others. (b) Subcategory 2 can be labeled as Residential in North Central City. Census tracts in this subcategory were located in the northern part of the central city, most of them were work unit neighborhoods or workers’ villages. The proportion of daily activities in these census tracts was much higher than others, however, the average trip distance was much shorter. The number of amenities in subcategory 2 was slightly higher than the average of the total residential census tracts and the proportion of Green Space land parcels was lower. (c) Subcategory 3 can be labeled as Large Residential in Suburbs. Census tracts in this subcategory mainly distributed in the suburbs surrounding the central city, many of them were large residential communities. Activity intensity, spatial coverage, day/night ratio, and weekday/weekend ratio were low, meaning there were more activities during non-work hours. Though they had fewer urban amenities than the former two subcategories, the amount of green space was larger. (d) Subcategory 4 can be labeled as Rural Residential. Census tracts belonging to this subcategory were far from the city center, with low activity intensity, low coverage rate, and longer trip distance. As for the built environment factors, the proportion of Industrial land parcels was much higher than the average and amenity volume was much lower.

Figure 3.

Values of PCV for the given number of clusters (residential).

Figure 4.

The spatial distribution, activity radar chart, and built-environment PDI (residential).

Classification results of commercial and public (C) function

For the Commercial and Public census tracts, Figure S12 in the Supplementary Material shows the values of PCA given the number of clusters from two to ten. According to the rules, four subcategories should be identified for Commercial and Public function. Except for the behavior regularity, all other indicators have PCV values between 0.6 and 0.7, indicating that a major part of the activity variance can be captured by the four subcategories. Figure S13 in the Supplementary Material shows the spatial distribution, activity radar chart, and the built environment PDI of the four subcategories: (a) Subcategory 1 can be labeled as City Center. This subcategory had the highest activity intensity, coverage rate, and proportion of outsiders. Business areas inside the Inner Ring Road and several employment centers were in this subcategory. After examining the built environment factors, we found they were close to the city center with good transit accessibility and sufficient amenities. (b) Subcategory 2 can be labeled as Community Center. Compared with the former one, subcategory 2 had relatively lower activity intensity and higher proportion of daily activities. There were more residents in subcategory 2 than in the other three subcategories. Furthermore, census tracts in subcategory 2 were closer to the city center. (c) Subcategory 3 can be labeled as Industrial and Service Center. We added “Industrial” because the proportion of the Industrial land parcels in subcategory 3 was way above average. Moreover, there were fewer amenities in this category compared with the other three. (d) Subcategory 4 can be labeled as Industrial and Service Center in the Outskirts. This subcategory was similar to subcategory 3, but it was farther away from the city center with fewer amenities.

Discussion

Delineating human activity with mobile phone data

Our study contributes to the existing literature in the following ways: (1) summarized mobile phone-derived human activity features into four dimensions and eight indicators, and (2) proposed a general framework to quantify human activity variances between and within conventional land use functions.

First, mobile phone data were deemed suitable for human activity studies by providing relatively complete (compared to smart card data and social media data) and up-to-date (compared to survey data and census data) activity information of massive user numbers (Wang et al., 2018; Wang and Ren, 2019; Zhou et al., 2020). However, the original purpose of mobile phone data collection was not for activity research, bringing uncertainties in data processing and privacy issues (Chen et al., 2014; Smith et al., 2012). Our study depicted mobile phone-derived human activity features with eight indicators, which can be calculated from raw mobile phone data with standardized data-processing steps proposed in Supplementary Material Section 1. Although we extracted individual trip information from mobile phone users, the indicators only reveal collective information at the aggregated level (census tract). Moreover, all indicators can be calculated based on aggregated-level mobile phone data, since the related variables are indexed by census tract instead of mobile phone users. To this end, our method still works when raw mobile phone data is unavailable due to data privacy.

Second, our study is one of the first attempts to quantify human activity variances between and within conventional land use functions. Based on the ANOVA analysis, two interesting points deserving further research were found. On the one hand, four activity dimensions varied differently among land use functions. Conventional land use functions could partly capture the variance of activity intensity and spatial coverage while almost failed to capture behavioral regularity and temporal variation, necessitating the classification of land use subcategories. On the other hand, land use subcategories identified in this study significantly differentiated temporal variation (with PCV values between 65% and 75%), while partly differentiated behavioral regularity (with PCV values between 30% and 40%). Why behavioral regularity cannot be captured by land use classifications leaves new research opportunities and questions to be addressed.

Towards activity-derived land use regulation

In light of changing urban spatial structure, flexible working schedule, and diverse recreational and shopping attractions, how space is actually used cannot be fully captured by the conventional broad zoning classifications (Jiang et al., 2017; Lopez and Ferreira, 2021; Yang, 2020). Conventionally, urban planners inferred activity demands according to land use functions. In the era of digitalization and urban big data, the available dynamic human activity information enables us to establish an interactive connection between human activity and land use function. For example, we can formulate the subcategories of land use into the processes of policy making. To this end, this empirical study of Shanghai contributes to the goal of achieving a more activity-based land use regulation.

The proposed four-dimensional activity indicators provide a new vision: (1) The strength of land use is usually regulated through building density and floor area ratio (FAR) in urban planning, while activity intensity reflects the actual strength of land use from the user’s perspective. Hence, activity intensity can be a supplementary indicator in determining the intensity of land development; (2) Behavioral regularity generally reflects what kind of people are using the space, which is a crucial basis of land use regulation. A higher proportion of daily activity indicates that the land use mainly serves local residents, while a higher proportion of random activity indicates a regional land use function; (3) Spatial coverage reflects the “connectivity” of land use with respect to human activities. A well-connected area which has the potential to become the urban center or sub-center can be figured out by higher spatial coverage. Accordingly, management strategies should focus on adding diverse transportation alternatives and improving level of service along main corridors; (4) Temporal variation depicts the daily/weekly routine of land usage, which supports a dynamic allocation of urban resources. For example, police force deployment, tidal traffic signals, and operating time of amenities can all be adjusted accordingly in order to relieve both congestion and low-utilization.

For the case of Shanghai, our study found that some identified subcategories combine numbers of census tracts, join multi-types of communities, or even cross several districts. For example, subcategories of Residential function formed four regional clusters in Shanghai: south of the central city, north of the central city, suburbs, and outskirts. Similarly, previous studies also found the spatial division in Shanghai, caused by distance to city center and year built of residential areas (Pan et al., 2009). Moreover, what is unique about our study is that it reveals a clear difference of activity features between the south and north central city: activity intensity in the central south city was higher than the north, while the proportion of daily activity in the north central city was higher than the south. The clear difference might be due to a comprehensive set of factors such as economic activity, built environment, industry structure, and historical issues. Further studies are needed to validate such a difference and figure out proper spatial policies.

Conclusion

As the relationship between land use and human activity is becoming increasingly complex, understanding intra-land use functional variance from a human activity perspective helps to achieve a more sustainable and human-oriented land use regulation. Our study summarized mobile phone-derived activity features into four dimensions and eight indicators, conducted analysis of variance (ANOVA) to calculate the activity variances within and between conventional land use functions, and presented the relationship between land use subcategories and built environment factors.

The results of the case study in Shanghai show that (1) Conventional land-use functional zones (LFZs) cannot fully capture the activity variances, especially in behavioral regularity and temporal variation; (2) At least four to five subcategories should be identified upon current LFZs to capture the main part of activity variances; and (3) In the case of Shanghai, land use subcategories presented palpable spatial regularity and revealed a citywide structure worth further study.

The contribution of our study is twofold. Theoretically, our study proposed standardized data-processing steps to calculate eight activity indicators based on raw mobile phone data and an ANOVA-based framework to quantify human activity variances between and within conventional land use functions. The proposed indicators system can be calculated even with aggregated-level dataset and the ANOVA-based framework is suitable for cities of different characteristics. Practically, our empirical study of Shanghai could help formulate activity-based spatial policies and identify a clear difference between the south and north central city, which deserves further investigations.

Several limitations exist and should be addressed in future studies. First, more experiments should be conducted to validate our indicator system, such as the effectiveness in reflecting activity features and the consistency given different periods of mobile phone data. Second, more case studies should be applied to test the feasibility and sensitivity of the ANOVA-based approach in various urban contexts, considering the quality of mobile phone data and social context varies among cities. Third, fieldwork is necessary to link human activity with policymaking. Ignorance on the development stages, periodical targets, and other domestic issues could yield misguidance in the policymaking process.

Supplemental Material

Supplemental Material—Exploring land use functional variance using mobile phone derived human activity data in Shanghai

Supplemental Material for Exploring land use functional variance using mobile phone derived human activity data in Shanghai by Xiyuan Ren, ChengHe Guan, De Wang, Junyan Yang, Bo Zhang and Michael Keith in Environment and Planning B: Urban Analytics and City Science

Footnotes

Author contributions

XR: conceptualization, investigation, methodology, writing original draft and data curation. CG: conceptualization, supervision, writing––review and editing and resources. DW: supervision. JY, MK and BZ: review and editing.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was sponsored by the Zaanheh Project and Center for Data Science and Artificial Intelligence at New York University (Shanghai); Fujian Urban Investment and Technology Institute’s Research Fund (Grant No. 20210201 FJCT); the PEAK-Urban Programme, which is funded by UK Research and Innovation’s Global Challenge Research Fund (Grant Ref: ES/P01105 5/1).

ORCID iDs

ChengHe Guan

Junyan Yang

Supplemental Material

Supplemental material for this article is available online.

References

Alberti

Weeks

Coe

(2004) Urban land-cover change analysis in central Puget Sound. Photogrammetric Engineering Remote Sensing 70: 1043–1052.

Batty

Vargas

Smith

, et al. (2013) SIMULACRA: fast land-use—transportation models for the rapid assessment of urban futures. Environment and Planning B: Planning and Design 40(6): 987–1002.

Cao

Guo

Zhang

(2019) Comparison of approaches for urban functional zones classification based on multi-source geospatial data: a case study in Yuzhong district, Chongqing, China. Sustainability 11(3): 660.

Chen

Bian

(2014) From traces to trajectories: how well can we guess activity locations from mobile phone traces? Transportation Research Part C: Emerging Technologies 46: 326–337.

Chen

Susilo

, et al. (2016) The promises of big data and small data for travel behavior (aka human mobility) analysis. Transportation Research Part C: Emerging Technologies 68, 285–299.

Chen

Song

, et al. (2020) Using mobile phone data to examine point-of-interest urban mobility. Journal of Urban Technology 27(4): 43–58.

Cottineau

Finance

Hatna

, et al. (2019) Defining urban clusters to detect agglomeration economies. Environment and Planning B: Urban Analytics and City Science 46(9): 1611–1626.

Duan

Wang

Zhang

, et al. (2017) Using longitudinal mobile phone data to understand the stability of individual travel patterns: case study of three communities in Shanghai, China. Transportation Research Record 2643(1): 166–177.

Gao

Krzysztof

Helen

(2017) Extracting urban functional regions from points of interest and human activities on location-based social networks. Transactions in GIS 21(3): 446–467.

10.

Guan

Song

Keith

, et al. (2021) Seasonal variations of park visitor volume and park service area in Tokyo: a mixed-method approach combining big data and field observations. Urban Forestry and Urban Greening 58: 126973.

11.

Guo

Song

Pei

, et al. (2019) Accessibility to urban parks for elderly residents: perspectives from mobile phone data. Landscape and Urban Planning 191: 103642.

12.

Jiang

Ferreira

González

(2017) Activity-based human mobility patterns inferred from mobile phone data: a case study of Singapore. IEEE Transactions on Big Data 3(2): 208–219.

13.

Lars

Wold

(1989) Analysis of variance (ANOVA). Chemometrics and Intelligent Laboratory Systems 6(4): 259–272.

14.

Kim

Sang

(2018) Exploring impacts of land use characteristics in residential neighborhood and activity space on non-work travel behaviors. Journal of Transport Geography 70: 141–147.

15.

Parrott

(2016) An improved genetic algorithm for spatial optimization of multi-objective and multi-site land use allocation. Computers, Environment and Urban Systems 59: 184–194.

16.

Lin

(1991) Divergence measures based on the Shannon entropy. IEEE Transactions on Information Theory, 37(1): 145–151.

17.

Liu

Thakuriah

, et al. (2020a) The geography of human activity and land use: a big data approach. Cities 97: 102523.

18.

Liu

Fang

Jing

(2020b) How urban land use influences commuting flows in Wuhan, Central China: a mobile phone signaling data perspective. Sustainable Cities and Society 53: 101914.

19.

Lopez

Ferreira

(2020) Identifying spatio-temporal hotspots of human activity that are popular non-work destinations. Environment and Planning B: Urban Analytics and City Science 48(3): 433–448.

20.

Metulini

Carpita

(2020) A spatio-temporal indicator for city users based on mobile phone signals and administrative data. Social Indicators Research 2020(3): 1–21.

21.

Noyman

Doorley

Xiong

, et al. (2019) Reversed urbanism: inferring urban performance through behavioral patterns in temporal telecom data. Environment and Planning B: Urban Analytics and City Science 46(8): 1480–1498.

22.

Ornetsmuller

Heinimann

Verburg

(2018) Operationalizing a land systems classification for Laos. Landscape and Urban Planning 169: 229–240.

23.

Paipuri

Gonzalez

, et al. (2020) Estimating MFDs, trip lengths and path flow distributions in a multi-region setting using mobile phone data. Transportation Research Part C: Emerging Technologies 118: 102709.

24.

Pan

Chen

Gao

, et al. (2020) An urban informatics approach to understanding residential mobility in Metro Chicago. Environment and Planning B: Urban Analytics and City Science 47(8): 1456–1473.

25.

Pan

Shen

Zhang

(2009) Influence of urban form on travel behaviour in four neighbourhoods of Shanghai. Urban Studies 46(2): 275–294.

26.

Pei

Sobolevsky

Ratti

, et al. (2014) A new insight into land use classification based on aggregated mobile phone data. International Journal of Geographical Information Science 28(9): 1988–2007.

27.

Qian

Lauri

Gechter

(2020) A probabilistic approach for discovering daily human mobility patterns with mobile data. In International Conference on Information Processing and Management of Uncertainty in Knowledge-Based Systems, Cham: Springer.

28.

Sagl

Delmelle

(2014) Mapping collective human activity in an urban environment based on mobile phone data. Cartography and Geographic Information Science 41(3): 272–285.

29.

Sevtsuk

Ratti

(2010) Does urban mobility have a daily routine? Learning from the aggregate data of mobile networks. Journal of Urban Technology 17(1): 41–60.

30.

Smith

Szongott

Henne

, et al. (2012) Big data privacy issues in public social media. In 2012 6th IEEE International Conference on Digital Ecosystems and Technologies IEEE.

31.

Wang

Ren

(2019) The distribution and structure of actual population in Shanghai under daily activity perspective. Urban Planning Forum 2: 36–43.

32.

Wang

Tsou

, et al. (2016) Mapping dynamic urban land use patterns with crowdsourced geo-tagged social media (Sina-Weibo) and commercial points of interest collections in Beijing, China. Sustainability 8(11): 1202.

33.

Wang

Leung

(2018) Applying mobile phone data to travel behaviour research: a literature review. Travel Behaviour and Society 11: 141–155.

34.

White

Engelen

Uljee

(1997) The use of constrained cellular automata for high-resolution modelling of urban land-use dynamics. Environment and Planning B: Planning and Design 24(3): 323–343.

35.

Zhu

(2017) Measuring comprehensive accessibility to community facilities for elderly people based on characteristics of activity spaces. In: Transportation Research Board 96th Annual Meeting.

36.

Xiao

Wang

Fang

(2019) Exploring the disparities in park access through mobile phone data: evidence from Shanghai, China. Landscape and Urban Planning 181: 80–91.

37.

Yang

(2020) Understanding commuting patterns and changes: counterfactual analysis in a planning support framework. Environment and Planning B: Urban Analytics and City Science 47(8): 1440–1455.

38.

Zhang

Rui

Poslad

, et al. (2020) A method for the estimation of finely-grained temporal spatial human population density distributions based on cell phone call detail records. Remote Sensing 12(16): 2572.

39.

Zhang

Foody

(2001) Fully-fuzzy supervised classification of sub-urban land cover from remotely sensed imagery: statistical and artificial neural network approaches. International Journal Remote Sensing 22: 615–628.

40.

Zhang

Yao

, et al. (2021) The Traj2Vec model to quantify residents’ spatial trajectories and estimate the proportions of urban land-use types. International Journal of Geographical Information Science 35(1): 193–211.

41.

Zhang

Wang

, et al. (2018) Multiscale geoscene segmentation for extracting urban functional zones from VHR satellite images. Remote Sensing 10: 281.

42.

Zhao

Huang

Yao

, et al. (2020) Predicting individual socioeconomic status from mobile phone data: a semi-supervised hypergraph-based factor graph approach. International Journal of Data Science and Analytics 9(3): 361–372.

43.

Zhong

Wang

Yan

, et al. (2017) The dynamic distribution of population and activity space in Shanghai: a exploration based on cell phone data. Geographical Research 36(5): 972–984.

44.

Zhou

Yang

Webster

(2020) Using big and open data to analyze transit-oriented development. Journal of the American Planning Association 86(3) 364–376.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

2.13 MB