Defining Urban Boundaries by Characteristic Scales
Yanguang Chen1, Jiejing Wang2, Yuqing Long1, Xiaohu Zhang3, Xiaoping Liu4, Xiaosong Li5 (1.Department of Geography, College of Urban and Environmental Sciences, Peking University, Beijing, China; 2. Department of Urban Planning and Management, School of Public Administration and Policy, Renmin University of China, Beijing, China; 3. Future Urban Mobility IRG, Singapore-MIT Alliance for Research and Technology, Singapore; 4. School of Geography and Planning, Sun Yat-sen University, Guangzhou, Guangdong, China; 5. Institute of Remote Sensing and Digital Earth, Chinese Academy of Sciences, Beijing, China.)
Abstract: Defining an objective boundary for a city is a difficult problem, which remains to be solved by an effective method. Recent years, new methods for identifying urban boundary have been developed by means of spatial search techniques (e.g. CCA). However, the new algorithms are involved with another problem, that is, how to determine the radius of spatial search objectively. This paper proposes new approaches to looking for the most advisable spatial searching radius for determining urban boundary. The key is to find out the characteristic length of spatial search by certain functional relationships. A discovery is that the relationships between the spatial searching radius and the corresponding number of clusters take on an exponential function, in which the scale parameter just represents the characteristic length. Using the characteristic length, we can define the most objective urban boundary. Two sets of China’s cities are employed to test this method, and the results lend support to judgment that the characteristic parameter can serve for the spatial searching radius. This study suggests a new way of determining urban boundary and determining city size in the right perspective.
Key words: urban boundary; urban form; characteristic scale; distance decay law; system of cities; Chinese cities
1 Introduction One of basic measures of a city is its size, which can be evaluated by urban population. Population is one of the central variables in the studies on urban evolution (Dendrinos, 1992), and 1
it represents the first dynamics of city development (Arbesman, 2012). In order to determine the urban population in the proper way, we must determine the urban area. The precondition of determining the urban area is to determine the urban boundary objectively. However, for a long time, the definition of urban boundary is a hard technical problem. In urban geography, there exist three key concepts of cities: city proper (CP), urbanized area (UA), and metropolitan areas (MA) (Davis, 1978; Zhou, 1995). A new trend seems to be that urbanized area will be replaced by urban agglomeration. The well-known urban statistician Davis (1978) once pointed out: “Whenever we speak of the city population, one must remember that our data do not refer to units defined in strictly comparable terms. Some of our cities are Cities Proper (CP’s)—with suburbs, if any, omitted because of lack of information—while others are units variously known as Urbanized Areas (UA’s), Urban Agglomerations, or Metropolitan Areas (MA’s), in which the suburbs are included.” Generally speaking, the definition of “city” refers to the urbanized area instead of city proper (Davis, 1978). However, the concept of “urban” differs from country to country, influencing people’s understanding of urbanized area. If the areal units of cities are improper or noncomparable, the corresponding city population cannot be employed to measure city sizes effectively and test urban regularities such as the rank-size rule, allometric growth law, and gravity law. Urban boundary depends on the way of spatial measurements. The boundary curve of a city is termed urban envelopes (Batty and Longley, 1994; Longley et al, 1991). Based on remote sensing images, at least three approaches have been developed to determining urban envelopes for cities. The first is the city clustering algorithm (CCA) proposed by Rozenfeld et al (2008, 2011), the second is the method of clustering street nodes/blocks advanced by Jiang and Jia (2011), and the third is the fractal-based method presented by Tannier et al (2011). Among these methods, CCA is based on raster data, while street network clustering is based on vector data. The two methods are involved with the technology of automated spatial search based on digital maps. However, how to determine the searching radius is still a difficult problem remaining to be solved. Tannier et al (2011) tried to identify urban boundary using the ideas from fractals and scaling. This is a practicable suggestion that will be discussed in a companion paper. This paper is devoted to exploring the methods of definition of urban boundary using the characteristic scale. The issue is how to find the characteristic scale of the searching radius. In fact, geographical phenomena fall 2
into two groups: one is with characteristic scale (scaleful group), and the other, without characteristic scale (scale-free group). The former can be described with characteristic length such as average value, standard deviation, and eigenvalue, while the latter should be described with scaling exponent such as fractal dimension. Geographical systems comprise both scaleful and scale-free processes and patterns, which are woven into each other. The problem of urban boundary will be readily solved as long as we can find a characteristic length for the searching radius. In order to solve the problem, mathematical modelling is necessary. An interesting finding is that the relationship between the searching radius and the number of clusters of an urban agglomeration follows a negative exponential law. A characteristic length can be found from the exponential distribution. This study is based on two important geographical laws: one is the distance decay law of urban density (e.g., Batty and Longley, 1994; Bussiere and Snichers, 1970; Chen, 2009; Chen, 2013; Clark, 1951; Smeed, 1963; Tobler, 1970; Wilson, 1970), which guarantees the feasibility of spatial search technique; and the other is the allometric scaling law of urban size and shape (e.g., Batty, 2008; Batty et al, 2008; Batty and Longley, 1994; Bettencourt, 2013; Bettencourt et al, 2007; Chen, 2010; Chen, 2014a; Chen, 2014b; Lee, 1989; Lo, 2002; Lo and Welch, 1977), which suggests that urbanized area can be employed to represent city population. The other parts of this article are organized as follows. In Section 2, the principle of spatial search for urban boundary is illuminated, and the mathematical models for characteristic scale analysis are presented. In Section 3, two sets of Chinese cities are employed to make case analyses to show how to find the characteristic radius for spatial search. In Section 4, several related questions are discussed. Finally, the paper is concluded by a summary of the main points.
2 Models and methods 2.1 Mathematical models The CCA and its variants can be employed to determine an urban boundary line, and thus a set of spatial measurements such as area, perimeter, and cluster number can be counted. Changing the searching radius yield different values of spatial measurements. Using the datasets from spatial searching, we can find a functional relationship between the searching radius and the 3
corresponding measures including urban area, street length, and node number. If the relationship follows a power law, we will be unable to find a characteristic length for the searching radius, and thus cannot determine an objective urban boundary. If so, we will never find an objective urban boundary. On the contrary, if the relationship satisfies a function with a characteristic scale, we will be able to find a characteristic length for the searching radius, and thus determine an objective urban envelope. The scale functions include exponential function, normal function, logarithmic function, logistic function, and so on. A scale-free function such as power function follow scaling relation, while a scale function does not follow scaling relation. By trial and error based on observational data, we find that the most possible functional relation is the exponential function, which is a scale function with characteristic length. In fact, the searching radius and the corresponding number of clusters satisfy an exponential relation as below
N (s) N0e s / s0 ,
where s denotes the length of searching radius, N(s) refers to the number of clusters, N0 is a proportionality constant, and s0 is the scale parameter indicating some characteristic length. Another possible function is the normal function
N (s) N 0e s
/(2 s02 )
which is in fact the variant of Gaussian function, and the symbols are the same as in equation (1). In practice, the exponential function and the normal function can be unified into a general form as follows
N (s) N 0e s
/( s0 )
where σ denotes a latent scaling exponent, which comes between 1/2 and 2 (Chen, 2010). If σ=1, we will have an exponential function; and if σ=2, then we will have a normal function. In above equations, the scale parameter s0 represents characteristic length of the searching radius. In theory, the scale parameter s0 is related to the average value of the searching radius. If we find the s0 value, we will find the most appropriate searching radius for defining an urban boundary. As a result, the most objective urban envelopes can be identified from digital maps by means of GIS technology.
2.2 The method of spatial search for urban envelope The study object of CCA is remote sensing images of urban agglomerations. The general idea of 4
city clustering is to aggregate land patches within a distance threshold s. It should be noted that although the data used in this study are vector files, the algorithm itself can be formulated in raster format. If the distance of two patches is smaller than s, the two patches will be identified as in the same cluster. The larger the distance threshold s, the less the formed clusters. Number of clusters will be reduced to 1 if the threshold is larger than the radius of study area. Since the clustering needs to be done in a set of different radii, iteratively joining nearby polygons could be too computationally intensive, especially in a large study area. We implemented an efficient method to
identify clusters. Firstly, we calculated distances between patches. But instead of computing every pair of them, we filtered out those which distances were larger than the maximum distance threshold using spatial indexing. This generated a sparse matrix M with its element dij denoting the distance of patch i to patch j. We then performed clustering over a series of distance thresholds. For each distance threshold s, we constructed a network which nodes were patches and two nodes (patches) were connected if their distance is smaller than s. The network would have many isolated sub-networks and nodes (patches) in each sub-network were identified as in the same cluster.
3 Empirical results 3.1 Material and methods The datasets are the land-use patches interpreted from Landsat satellite imagery of China. The images were classified into X categories, from which the built-up areas are extracted for our study. First of all, 24 Chinese cities are employed to show how to find a characteristic length of searching radius to define urban boundary. These 24 cities are scattered on all over China (Figure 1). The typical regions consist of National capital (i.e., Beijing), Yangtze River Delta (e.g., Hangzhou, Nanjing, Shanghai, Wuxi), Pearl River Delta (e.g. Guangzhou), Central Plain (e.g., Kaifeng, Luoyang), Central south China (e.g., Changsha), Shandong Peninsula (Rizhao), northeast China (e.g. Changchun, Fuxin, Liaoyuan), northwest China (e.g., Changzhi, Yinchuan), southwest Chian (e.g., Chengdu, Kunming). Where urban hierarchy is concerned, these cities comprise megacities (megalopolis), large cities (metropolis), medium-sized cities, and small cities (Table 1). The original remote sensing images are of vector format and belong to 2000 and 2010, 5
respectively. The spatial datasets were extracted and processed by ArcGIS from the digital maps of urban land use.
Table 1 The 24 cities as a sample from the hierarchy of Chinese cities Types
Figure 1 A spatial sample of 24 Chinese cities which are taken as examples of characteristic radius analysis
A set of searching radii can be given to determine set of boundary curves of a city. Based on an 6
urban boundary, three measurements can be obtained, that is, number of clusters, the area of the largest cluster, and the total area of all the clusters. We can write a computer program of ArcGIS to perform the search and design a cycle to control the searching radius. By repeated tests, the minimum searching radius is taken s=50 meters, and the step length of radius change is set as ∆s=10 meters. The maximum searching radius (smax) depends on urban shape, network pattern, map layout, and so on. For each selected city, we use the method descripted in the above section to perform the spatial searching. Where the national capital of China is concerned, the larger searching radii leads to larger urban clusters (Figure 2). The cases of other cities are similar to Beijing (Figure 3). The cluster number, cluster area, and the area of the largest cluster can be automatically extracted by the computer program during each searching cycle. Thus we will have three datasets for a series of variable urban boundaries of each city.
Figure 2 The clusters of Beijing’s urban agglomeration g based on different searching radius 7
Figure 3 The sketch maps of the spatially identified results of the 24 Chinese cities 8
[Note: The area of these cities depends on the searching radius. The longer the searching radius is, the larger the urban area will be.]
The linear regression analysis based the least squares method can be employed to estimate the values of parameters. Taking natural logarithms on both sides of equation (3) yields
ln N ( s) ln N 0
s =a bs , s0
in which the parameters are a=lnN0 and b=1/(σs0σ), respectively. Thus, the characteristic radius can be given by
s0 ( b)1/ .
If σ=1 as given, then equation (4) will return to an ordinary linear relation such as
ln N ( s ) ln N 0
s a bs , s0
Thus we have s0=1/b. If the relation between the searching radius and cluster number follow the exponential decay law, we can use the linear regression analysis based on equation (6) to evaluate the characteristic radius s0. If the relation follow the fractional exponential law, we can find the value of the latent scaling exponent σ by means of the cut-and-try method. As soon as the σ value is determined, we can use equation (4) to make a linear regression analysis and find the characteristic radius s0.
3.2 Results The exponential function and the fractional exponential function can applied to the spatial searching processes of Chinese cities. Concretely speaking, equations (4) and (6) can be fitted to the datasets of the 24 Chinese cities. For example, for Beijing in 2010, the relation between searching radii and cluster numbers can be well described by the common exponential function. By the least squares computation, the model can be built as follows
Nˆ (s) 2416.8043e0.0017 s ,
where the hat “^” indicates that the result is a predicted value rather than an observed value. The coefficient of determination is about R2=0.9992. The decay coefficient is b≈0.00173992, thus the characteristic radius of spatial search is estimated as s0=1/b≈574.7391 (Figure 4). Where the 9
spatial searching is concerned, all the 24 cities can be approximately modeled by the negative exponential function. Thus the characteristic searching radius can evaluated by the least squares regression based on equation (1). According to the exponential model, for the 24 cities in 2000, the average characteristic searching radius is about ŝ0(2000)=479.4751; for these cities in 2010, the average value is about ŝ0(2010)=463.0777 (Table 2). Here ŝ0 refers to the mean of the characteristic searching radii of different cities.
Table 2 The regression coefficients, characteristic searching radii and the goodness of fit of the 24 Chinese cities based on exponential function City
N(s) = 3057.6176e-0.0021s R²= 0.9944 2000
Cluster Number N(s)
Cluster Number N(s)
N(s)= 2416.8043e-0.0017s R²= 0.9992
2000 1500 1000
Searching radius s
Searching radius s
a. Beijing 2000
b. Beijing 2010
Figure 4 The exponential decay relations between the searching radii and the numbers of clusters of Beijing city (Note: For Beijing in 2000, the negative exponential function is not the best model, the most probable model is a fractional exponential model with a latent scaling exponent σ=1.25. However, in 2010, the common negative exponential function is the most advisable model for Beijing city.)
However, not all cities can be well described with the common exponential decay function at all times. For example, the cities of Changchun and Kaifeng satisfy the normal function. Great majority of cities follow the fractional exponential decay law. Where Beijing in 2000 is concerned, it should be described with the generalized exponential function such as 1.25 Nˆ ( s) 2610.4595e0.0004 s ,
where the latent scaling exponent is about σ=1.25. The goodness of fit is about R2=0.9992. The decay coefficient is b≈0.0003507, thus the characteristic radius of spatial search is estimated as s0 = (1.25*0.0003507)^(-1/1.25) ≈ 485.8737. Luoyang city in 2010 is similar to Beijing city in 2000 (Figure 5). According to the fractional exponential model, for the 24 cities in 2000, the average characteristic searching radius is about ŝ0(2000)=528.7216; for them in 2010, the average value of typical searching radii is about ŝ0(2010)=569.6380 (Table 3).
Table 3 The regression coefficients, characteristic searching radii and the goodness of fit of the 24 Chinese cities based on fractional exponential function City
N(s) = 2610.4595e-0.0004(s^1.25) R²= 0.9992
Cluster Number N(s)
Cluster Number N(s)
N(s)= 3228.2028e-0.0004(s^1.25) R²= 0.9983
2000 1500 1000 500
2500 2000 1500 1000 500
Searching radius s
Searching radius s
a. Beijing 2000
b. Luoyang 2010
Figure 5 The fractional exponential decay relations between the searching radii and the numbers of clusters of Beijing city (σ=1.25) (Note: For both Beijing in 2000 and Luoyang city in 2010, the most probable latent scaling exponent is σ=1.25.) 12
There are two ways of defining boundaries for cities. One is for an individual city, and the other is for all cities in a region (country). If we want to find the most advisable boundary for a particular city, we can use the fractional exponential function to find the characteristic searching radius of the city at given time (e.g., Table 3). However, for different cities at different times, the values of the characteristic searching radius are always different. Thus it is difficult to define comparable urban boundaries for a set of cities. In this instance, we can calculate the average value of characteristic searching radii of a sample of cities. Unfortunately, another problem arises: the average value is not often stable. A solution to this problem is to remove the outliers in a sample using the double standard derivation before computing the average value. Based on the significance level α=0.05, an acceptable characteristic searching radius s0 should fall between average value plus double standard deviation and average value minus double standard deviation, that is
sˆ0 2sd s0 sˆ0 2sd , where ŝ0 denotes the average value (mean) of the characteristic searching radius, and sd refers to the standard deviation. Otherwise, we have 95% of confidence level to treat the s0 value as an outlier. After removing the outliers step by step, we can find an acceptable ŝ0 value. The effective average value of the characteristic searching radii can be approached step by step. Taking the dataset of the 24 cities in 2000 based on exponential model as an example, the procedure of finding a stable average value is as follows. Step 1, calculate the average value ŝ0 and the standard deviation sd. If s0(k)< ŝ0-sd or s0(k)> ŝ0+sd, remove it (k=1,2,…,24). In this step, the characteristic searching radius of Mengzi city is an outlier and should be removed (Figure 6(a), Table 4). Step 2, recalculate the average value ŝ0 and the standard deviation sd. If s0(k)< ŝ0-sd or s0(k)> ŝ0+sd, remove it (k=1,2,…,23). In this step, the characteristic radius of Kunming city is an outlier and should be deleted (Figure 6(b), Table 4). Step 3, recalculate the average value ŝ0 and the standard deviation sd once again. If s0(k)< ŝ0-sd or s0(k)> ŝ0+sd, remove it (k=1,2,…,22). In this step, the characteristic radius of Changzhi city is an outlier and should be eliminated (Figure 6(c), Table 4). Step 4, repeat above calculation. If s0(k)< ŝ0-sd or s0(k)> ŝ0+sd, remove it (k=1,2,…,21). In this step, all the characteristic radii of the 21 remaining cities fall into the range from ŝ0-sd to 13
ŝ0+sd, and the final average value and standard deviation are 405.9491 and 170.3990 (Figure 6(d), Table 4).
Table 4 An example of removing outliers by average values and double standard deviations of the 24 cities in 2010 (common exponential model)
Lower limit (mean-2*sd)
Upper limit (mean+2*sd)
No excepted city
1000 800 600 400 200
700 600 500
400 300 200
a. Step 1: Mengzi
b. Step 2: Kunming
500 400 300
500 400 300
c. Step 3: Changzhi
d. Step 4: No outlier
Figure 6 The process and patterns of removing outliers by average values and double standard deviations of the 24 cities in 2000 (Note: The model is common exponential decay function. In step 1, the city of Mengzi is an outlier; in step 2, Kunming is an outlier; in step 3, Changzhi is an outlier; in step 4, no outlier, and the process is over.) 14
Based on the two models and two years, we have four datasets. Using the similar method, we can address all these datasets (Table 5). The other results are as below. For the dataset of 2010 based on exponential decay model, the average value and standard deviation are 407.2503 and 170.3565 (the outliers are Mengzi and Kunming); For the dataset of 2000 based on fractional exponential decay model, the average value and standard deviation are 415.6138 and 176.5345 (the outliers are Mengzi, Kunming and Changsha); For the dataset of 2010 based on fractional exponential decay model, the average value and standard deviation are 404.9980 and 166.6497 (the outliers are Mengzi, Kunming and Changsha). The most advisable characteristic searching radius can be regarded as about 410, and the corresponding standard deviation is about 170.
Table 5 The final average values, standard deviations, and the lower and upper limits of the characteristic searching radii
Fractional exponential decay
Lower limit (mean-sd)
Upper limit (mean+sd)
3.3 Further application to an system of cities The above calculations and analyses are based on a sample of 24 Chinese cities. These cities have no significant relation to each other. Now, the characteristic length method can be applied to a system of cities in China, Jing-Jin-Ji, from which we can obtain new insight in urban definition. Jing-Jin-Ji region is also termed Beijing-Tianjin-Hebei, including Beijing Municipality (ab. “Jing”), Tianjin Municipality (ab. “Jin”), and part of Hebei Province (ab. “Ji”). The Jing-Jin-Ji urban system comprises 35 cities in 2010 (Figure 7). The original images are of raster format, differing from the previous example. Using variable searching radii such as 20, 25, and 30, we can abstract different urban agglomerations for each city in 2000, 2005, and 2010. The relationships between searching radii and cluster numbers of all these cities follow exponential decay law, and can be modeled or approximately modeled by the negative exponential function (Table 6). The 15
scale parameter suggests the characteristic searching radius, s0. A finding is that the characteristic searching radius becomes shorter and shorter along with urban growth. The average searching radii are 539.9886 m for 2000, 510.9694 m for 2005, and 476.9993 m for 2010, respectively. After removing the outliers from the datasets by average values and double standard deviations, the average values become 490.368 m for 2000, 471.9630 m for 2005, and 443.0722 m for 2010, respectively. Urban growth implies space filling, and thus urban density increases over time. For a city, the higher the urban density is, the shorter the characteristic searching radius will be.
Table 6 The decay coefficients, characteristic searching radii, and goodness of fit of Jing-Jin-Ji system of cities City
Note: The notation is as follows: b—decay coefficient, s0—characteristic searching radius, R2—goodness of fit. Stdev means “standard deviation”. The average values and standard deviations with asterisk “*” are based on the datasets from which the outliers are removed.
Figure 7 A sketch map of the systems of cities and towns in Jing-Jin-Ji region of China 17
Another finding is that the characteristic radius depends on the format of remote sensing image and the method of data processing. In both the sample of 24 Chinese cities and the population (universe) of 35 Jing-Jin-Ji cities, the patterns of Beijing city in 2000 and 2010 are taken into consideration. However, the results for 2010 are significantly different. The former result is 574.7391 (increase), while the latter result is 366.3731 (decrease). The first results are based on vector data (Table 2), while the second results are based on raster data (Table 6). For the first results, the searching radii are s=50, 60, 70, …, 1000, and the step length is ∆s=10; For the second results, the searching radii are s=20, 25, 30,…, 500, and the step length is ∆s=5. This suggests that only based on the same data format and the same data processing method, comparable results can be gained for the definition of urban agglomerations and urban boundaries.
4 Discussion The key to define an objective urban boundary is to find a characteristic searching radius. In fact, scientific research should proceed first by describing a system and later by understanding the mechanism (Gordon, 2005; Henry, 2002). The precondition of effective description is to find the characteristic length of a thing. If an urban pattern follows power laws, it has no characteristic scale, and cannot be described with the traditional method; if an urban pattern satisfies an exponential distribution, it has a characteristic length, and can be effectively described in a simple way. As indicated above, the scale parameter of an exponential distribution model is just the characteristic length indicating the average value. In this paper, we reveal a generalized exponential relation between the spatial searching radii and the numbers of clusters. In this case, the scale parameter s0 give the characteristic searching radius. Based on the characteristic radius, we can define a relatively objective urban boundary. Thus, urban area and city size can be objectively determined by the objective urban envelope. For individual cities, the characteristic radii can be calculated one by one; for a system of cities, the average characteristic radius can be computed step by step using average values and double standard deviations. The significant shortcoming of this study lies in that only exponential decay and fractional exponential decay are taken into account. Geographical systems are different from classical 18
physical systems. Geographical laws are not iron laws, and a geographical process or pattern can be modeled with more than one mathematical equation. Several functions can be employed to describe the relationships between the searching radii and cluster numbers, but the exponential function is the basic one and the most probable one. There are many types of spatial and probability distributions with characteristic lengths such as exponential distribution, logarithmic distribution, normal distribution, lognormal distribution, and so on. The scale-free distributions are mainly power-law distributions and latent power-law distributions. Among various distribution functions with characteristic scales, the exponential function is the simplest one. In fact, many distribution functions with characteristic scales can be transformed into an exponential function (Table 7). Normal distribution can be treated as generalized exponential distribution. The inverse function of logarithmic function is just an exponential function. It is easy to convert a sigmoid function into an exponential function. If the spatial distribution follows a power law indicative of the scaling process without breaking, we will be unable to find an objective urban agglomeration or urban boundary. However, if the scaling breaks, we will find the typical searching radius by the breaking points of scaling. An interesting study is made by Tannier et al (2011), who use scaling breaking point to find effective searching radius. In fact, if a power-law distribution bears scaling break, it implies a self-affine process and pattern and thus takes on a pseudo exponential distribution (Chen and Feng, 2012). In this case, we can use both exponential distribution and scaling breaking point to find characteristic searching radius. This problem will be solved and the related questions will be illustrated in a companion paper.
Table 7 Three functions and the corresponding characteristic searching radii (examples)
N (s) a b ln(s)
s sb e N ( s )/b
s0 ea / b1
a, b, sb=exp(a/b)
N max 1 aebs N (s)
s0 1/ b
a, b, Nmax
N (s)sc aebs
s0 1/ b
a, b, c
N max 1 aebs
N (s) aebs sc
5 Conclusions The basic algorithm of defining urban boundary has been developed, but how to find an objective searching radius is a pending problem. Fractal approach proved to be suitable for determining urban envelope for bi-scaling patterns of urban form. In this paper, the characteristic length of spatial search is proposed to define urban boundary. The main conclusions can be reached as follows. First, the scale parameter of the exponential model represents the characteristic searching radius for definition of urban boundary. Generally speaking, the relation between spatial searching radii and the numbers of clusters follows exponential decay law. The reciprocal of the decay coefficient is regarded as the scale parameter, which is just the characteristic length of spatial search. Using the characteristic radius, we can determine an urban agglomeration or define urban boundary more effectively and objectively. Second, the average value and double standard deviation can be employed to find average characteristic searching radius for a system of cities. Different cities at different times have different characteristic searching radius. For individual cities, we can use the characteristic radius of a concrete city. For a system of cities in a region, however, we must find a common characteristic radius by averaging after removing outlier. If a characteristic radius of a cities is less than the average value minus double standard deviation or greater than the average value plus double deviation, it should be deleted as an outlier. Step by step, we can eliminate all the exceptional values and calculate the final average characteristic radius. Third, other models of spatial search with characteristic length can be transformed into exponential model to give the characteristic searching radius. Not all the cities follow the exponential decay law in the process of city clustering. If the relation between searching radii and cluster numbers follow the power law with single scaling process, the urban boundary cannot be found objectively. If the relation satisfy a function with characteristic scale, e.g., normal function, logarithmic function, lognormal function, gamma function, and so on, we can convert the function into the form of an exponential function and find the characteristic searching radius. As soon as the characteristic radius is worked out, an urban agglomeration or urban boundary can be determined in a proper way. 20
Acknowledgements This research was sponsored by the National Natural Science Foundations of China (Grant No. 41590843 &41671167). The supports are gratefully acknowledged.
References Arbesman S (2012). The Half-Life of Facts: Why Everything We Know Has An Expiration Date. New York: Penguin Group Batty M (2008). The size, scale, and shape of cities. Science, 319: 769-771 Batty M, Carvalho R, Hudson-Smith A, Milton R, Smith D, Steadman P (2008). Scaling and allometry in the building geometries of Greater London. The European Physical Journal B - Condensed Matter and Complex Systems, 63(3): 303-314 Batty M, Longley PA (1994). Fractal Cities: A Geometry of Form and Function. London: Academic Press Bettencourt LMA (2013). The origins of scaling in cities. Science, 340: 1438-1441 Bettencourt LMA, Lobo J, Helbing D, Kühnert C, West GB (2007). Growth, innovation, scaling, and the pace of life in cities. Proceeding of the National Academy of Sciences of the United States of America, 104(17): 7301-7306 Bussiere R, Snickers F (1970). Derivation of the negative exponential model by an entropy maximizing method. Environment and Planning A, 2(3): 295-301 Chen YG (2009). Urban gravity model based on cross-correlation function and Fourier analyses of spatio-temporal process. Chaos, Solitons & Fractals, 41(2): 603-614 Chen YG (2010). A new model of urban population density indicating latent fractal structure. International Journal of Urban Sustainable Development, 1(1-2): 89-110 Chen YG (2010). Characterizing growth and form of fractal cities with allometric scaling exponents. Discrete Dynamics in Nature and Society, Volume 2010, Article ID 194715, 22 pages Chen YG (2013). New approaches for calculating Moran’s index of spatial autocorrelation. PLoS ONE, 8(7): e68336 Chen YG (2014a). An allometric scaling relation based on logistic growth of cities. Chaos, Solitons & Fractals, 65: 65-77 21
Chen YG (2014c). The spatial meaning of Pareto’s scaling exponent of city-size distributions. Fractals, 22(1-2):1450001 Chen YG, Feng J (2012). Fractal-based exponential distribution of urban density and self-affine fractal forms of cities. Chaos, Solitons & Fractals, 45(11): 1404-1416 Clark C (1951). Urban population densities. Journal of Royal Statistical Society, 114(4): 490-496 Davis K (1978). World urbanization: 1950-1970. In: I.S. Bourne, J.W. Simons (eds). Systems of Cities: Readings on Structure, Growth, and Policy. New York: Oxford University Press, pp92-100 Dendrinos DS (1992). The Dynamics of Cities: Ecological Determinism, Dualism and Chaos. London and New York: Routledge Gordon K (2005). The mysteries of mass. Scientific American, 293(1):40-46/48 Henry J (2002). The Scientific Revolution and the Origins of Modern Science (2nd Edition). New York: Palgrave Jiang B, Jia T (2011). Zipf's law for all the natural cities in the United States: a geospatial perspective. International Journal of Geographical Information Science, 25(8): 1269-1281 Lee Y (1989). An allmetric analysis of the US urban system: 1960–80. Environment and Planning A, 21(4): 463–476 Lo CP (2002). Urban indicators of China from radiance-calibrated digital DMSP-OLS nighttime images. Annals of the Association of American Geographers, 92(2): 225-240 Lo CP, Welch R (1977). Chinese urban population estimates. Annals of the Association of American Geographers, 67(2): 246-253 Longley PA, Batty M, Shepherd J (1991). The size, shape and dimension of urban settlements. Transactions of the Institute of British Geographers (New Series), 16(1): 75-94 Rozenfeld HD, Rybski D, Andrade Jr. DS, Batty M, Stanley HE, Makse HA (2008). Laws of population growth. PNAS, 2008, 105(48): 18702-18707 Rozenfeld HD, Rybski D, Gabaix X, Makse HA (2011). The area and population of cities: New insights from a different perspective on cities. American Economic Review, 101(5): 2205–2225 Smeed RJ (1963). Road development in urban area. Journal of the Institution of Highway Engineers, 10(1): 5-30 Tannier C, Thomas I, Vuidel G, Frankhauser P (2011). A fractal approach to identifying urban boundaries. Geographical Analysis, 43(2): 211-227 22
Tobler W (1970). A computer movie simulating urban growth in the Detroit region. Economic Geography, 46(2): 234-240 Wilson AG (1970). Entropy in Urban and Regional Modelling. London: Pion Press Zhou YX (1995). Urban Geography. Beijing: The Commercial Press [In Chinese]
Appendix—Beijing clusters based on different searching radii (2010)