Urban green spaces, their spatial pattern, and ecosystem service value The case of Beijing.pdf

Habitat International 56 (2016) 84e95

Contents lists available at ScienceDirect

Habitat International journal homepage: www.elsevier.com/locate/habitatint

Urban green spaces, their spatial pattern, and ecosystem service value: The case of Beijing Liyan Xu a, b, Hong You a, Dihua Li a, Kongjian Yu a, * a b

Peking University, China Massachusetts Institute of Technology, USA

a r t i c l e i n f o

a b s t r a c t

Article history: Received 11 January 2016 Received in revised form 16 April 2016 Accepted 25 April 2016

Green spaces provide various kinds of ecosystem service functions. Though some of them, such as the carbon-sinking and biodiversity preservation functions are of value to everyone, others, especially those related to aesthetic and recreational functions, only benefit people who have direct access to green spaces. In urban settings, where ecosystem services in the second category prevail, this means the spatial dimension of urban green spaces, including their richness, accessibility, shape configuration, and distributional characteristics, may considerably influence the realization of their ecosystem service value, and is therefore subject to scrutiny. In this paper, we study how the spatial pattern of urban green spaces influence the realization of their ecosystem service value by utilizing the Hedonic Price Modeling (HPM) method. Taking Beijing as the case, we use the price and other information in the city's developable land transaction records from 2000 to 2004 to construct the HPM, and use Landscape Ecological Metrics (LEM) as proxies of the spatial characteristics of urban green spaces. Four LEMs are used to measure the above mentioned spatial characteristics of urban green spaces. While subject to certain shortcomings in data quality and quantitative estimations of the magnitude of the spatial effects cannot be made, results show that most spatial characteristics of urban green spaces do influence their ecosystem service value as embedded in land value, except for the shape configuration characteristic for which the study yields no result. Further, specifically for Beijing, results indicate that in order to effectively realize their ecosystem service value, green spaces should occupy between 2.20% and 13.40% of the total urban area, located within a 50e550 m range from other developments, with green space patches so divided that each patch occupies more than 3.00% but less than 62.50% of the total green space area, and the ecosystem service value will be at the optimal level when each patch occupies 20.00% of the total green space area. Lastly, we stress the practical significance of the findings, urging an integration of the spatial patterns aspect of urban green spaces in urban planning practices. © 2016 Elsevier Ltd. All rights reserved.

Keywords: Urban green spaces Ecosystem service Spatial pattern Landscape Ecological Metrics (LEM) Hedonic Price Model (HPM) Beijing

1. Introduction Urban areas across the world have been facing threatens from environmental degradation (McMichael, 2000), and the situation is especially severe in fast urbanizing and industrializing developing countries, such as China (Stern, Common, and Barbier, 1996; Liu & Diamond, 2005; Economy 2011). Environmental degradation not only causes physical harms like air and water pollution (Booth & Jackson, 1997; Shao, Tang, Zhang, & Li, 2006), but also inflicts mental problems to the urban residents (Jiang, Zhang, and Sullivan, 2015), and thus constitutes an urgent issue to address.

* Corresponding author. http://dx.doi.org/10.1016/j.habitatint.2016.04.005 0197-3975/© 2016 Elsevier Ltd. All rights reserved.

Green spaces, including forest, grassland, farmland, etc., play a crucial role in the global ecosystem, and urban green spaces in particular are commonly regarded as a remedy to the urban environmental problems. They help remove air and water pollution (Jim & Chen, 2008; De Ridder et al., 2004), preserve biodiversity €rtberg & Wallentinus, 2000), and create an amenable atmo(Mo sphere which benefits people's physical and mental health (De Vries et al. 2003; Hillsdon, Panter, Foster, & Jones, 2006; Jim & Chen, 2006; Jiang, Larsen, Deal, & Sullivan, 2015). An interesting question, though, is how much value have urban green spaces realized in improving the environment. Technically, a feasible way to answer the question is by evaluating the ecosystem  mez-Baggethun service value generated by urban green spaces (Go & Barton, 2013). However, in the absence of a market for urban

L. Xu et al. / Habitat International 56 (2016) 84e95

green spaces, one needs non-market evaluating methods to estimate the former's value in monetary terms (McConnell & Walls, 2005). Seminal works of this kind can be attributed to Costanza et al. (Costanza et al. 1997), who invented a constructed willingness-to-pay method to estimate the service value of the global ecosystem. This study established a “unit value”-based paradigm for evaluating the value of ecosystem services, which despite some innate drawbacks (Bingham, Bishop, & Brody, 1995; Pearce, 1998), has since been widely applied in ecosystems of various types and scales (Rapport, Gaudet, & Karr, 1998), and green spaces in particular, with a fairly large geographical coverage (Costanza, Stern, & Fisher, 2004; Cilliers, Cilliers, Lubbe, & Siebert, 2012; Bateman et al. 2013). However, the “unit value”-based method is built implicitly on the assumption that the ecosystem service value is independent of the spatial dimension. This may hold true at the global or other very large scales, as in the case of the above mentioned Costanza et al.’s work. But at smaller scales, the assumption is not as sound from a landscape ecology perspective, which states though some ecosystem service functions, such as carbon sinking, are locationinsensitive, others are not. Indeed, as Forman (Forman, 1995) noted, green space patches with different spatial attributes (richness, accessibility, shape configuration, and distributional characteristics) may have different ecological functions in a landscape, and they therefore should convey varied amount of ecosystem service value. For example, Xie et al. (Xie, Xiao, & Lu, 2006) demonstrate that the soil and water preservation service of forests is much less important in plains than in slope terrains, thus is of less ecosystem service value in the former case. Particularly, in urban contexts, where the most prominent ecosystem service functions are regarding the aesthetic and recreational aspects, the spatial dimension matters even more profoundly. It is therefore necessary to study the relationship between the spatial pattern of urban green spaces and their ecosystem service value, a subject the existing literature sheds little light on. The lack of research attention on the issue not only constitutes an academic gap, but has also inflicted negative influences in realworld practices. The argument above suggests that urban planners and managers should pay as much attention to the spatial dimension of urban green spaces as the quantity. However, despite certain degrees of academic coverage, the notion appears not quite commonly appreciated in practice, resulting an overemphasis on the latter and neglect of the former in many occasions (Haase et al. 2014). For example, some large cities in China have adopted an “occupation/compensation balance of urban green spaces” policy,1 requiring developers who destruct urban green spaces to create new ones elsewhere. In practice, however, for quite understandable reasons, such make-up green spaces usually locate in the exurbia or even remoter areas. Such practices, letting alone the obvious political ecological problem they imply (Heynen, Perkins, & Roy, 2006; Wolch, Byrne, & Newell, 2014), cause losses in ecosystem service value, too, as urban green spaces’ ecosystem service functions, such as the city beautification and micro-climate control, are only meaningful where population concentrates, and the loss of such services in the city center cannot be “compensated” by as large, or even larger green spaces in remote areas where their ecosystem service value hardly realizes. In this paper, we study the relationship between urban green spaces’ spatial pattern and the ecosystem service value they convey.

1 For example, Guidelines for the Basic Ecological Control Line Management for the city of Wuhan, see http://www.wuhan.gov.cn/frontpage/pubinfo/PubinfoDetail. action?id¼1201205242201590014; and Regulations on Urban Green Spaces for the city of Qingdao, see http://rules.yuanlin.com/Html/Detail/2012-2/1599_2.html.

85

Specific spatial characteristics to be examined include the richness, accessibility, distribution, and shape configuration of urban green spaces. Using empirical data from Beijing, China, we construct quantitative models to examine the spatial effects, and also explore their practical implications. We hope our work will not only contribute to the academic literature, but also have practical impacts and could thus help build a better urban environment. The paper is organized as follows. Section two provides a review of relevant studies and methods. Section three discusses the modeling approach, measurement of spatial characteristics, and data issues. Section four presents the results, with detailed demonstration of the spatial effects resulting from different richness, accessibility, distribution, and shape configuration variables. Section five offers further analysis on the quantified estimation of the spatial effects, as well as discussions on the possible reasons for the non-significant results. We conclude the paper in section six with a summary of the study and discussions on the study's practical implications.

2. Evaluating the ecosystem service value of urban green spaces and the influence of their spatial patterns: a literature review As there does not exist a market for urban green spaces in most occasions, people's willingness-to-pay for urban green spaces could only be measured through indirect approaches (McConnell & Walls, 2005). The price of real estate, for example, is a commonly used proxy, which is considered to include a “green space premium” e the willingness-to-pay for accessibility to urban green spaces so as to enjoy the ecosystem service they provide. The Hedonic Price Model (HPM) (Chau, Ma, & Ho, 2001) is a typical method to separate out the “green space premium” from real estate prices. HPM assumes that the price of a commodity, such as an apartment unit or a land parcel, includes the contributions from its various innate and environmental characteristics (Lancaster, 1966). Therefore, one may identify the willingness-to-pay for each feature involved using analytical techniques such as multivariate regression. Specifically, for urban real estates, their prices are usually considered to consist of the contributions from three categories of characteristics: the structural (such as the size of a land parcel or the unit plan type of an apartment), neighborhood (such as the transportation accessibility), and environmental (such as nearby amenity and recreational facilities) variables (Poudyal, Hodges, & Merrett, 2009). A typical Hedonic Price Model is thus formulated as follows:

ln pi ¼ b0 þ

X

bj Sij þ

X

bk Nik þ

X

bl Eil þ εi

(1)

where ln pi denotes the logarithm of the price of the i-th real estate object, Sij denotes its j-th structural variable, Nik denotes its k-th neighborhood variable, and Nik denotes its l-th environmental variable. b0 , bj , bk , bl , and εi are the respective estimates of regression coefficients and the residual term. Therefore, the HPM can be used to evaluate the value of urban green spaces when they serve as the environmental variables in Equation (1). Loads of works of this sort have been done during the past half-century (McConnell & Walls, 2005), covering various green space types such as natural habitats, parks, planted forests, wetlands, and farmlands. The divergent influences of different green space types have also been widely discussed (Bolitzer & Netusil, 2000; Neumann, Boyle, & Bell, 2009). Particularly, regarding the urban green spaces, the subject of this study, detailed studies have been conducted concerning their influence on real estate value (Bolitzer & Netusil, 2000; Poudyal et al.,

86

L. Xu et al. / Habitat International 56 (2016) 84e95

2009). While some studies fail to find a relationship between the two (Morancho, 2003), most empirical evidences show that real estate values are indeed positively correlated to the existence of nearby urban green spaces. Other inquiries have gone further to reveal such relationship's dependency on spatial proximity, as measured by the distance between urban green space locations and real estate sites (Czembrowski & Kronenberg, 2016; Morancho, 2003; Sander & Polasky, 2009; Tajima, 2003; Tyrvainen & Miettinen, 2000). With these progresses, however, the majority of the literature has not addressed other spatial characteristics of urban green spaces, e.g. the richness, distribution, and shape configuration characteristics, and their influence on nearby real estate values. Among the few works that do cover this issue, Kong, Yin, and Nakagoshi (2007) show that real estate values in Jinan, China are indeed influenced by the spatial patterns of urban green spaces as represented by a size-distance index, as well as by the richness of urban green space. Poudyal, Hodges, Tonn, & Cho, 2009, studying a real estate transaction sample in the City of Roanoke, Virginia, have gone further and showed that open space plots with square shape and smooth, straight edges were preferred to those with more complex shapes and irregular edges, and that open spaces in few larger plots are preferred to many smaller scattered pieces. Methodologically, what is worth noting in these works is that they use Landscape Ecological Metrics (LEM) to quantify the spatial characteristics of urban green spaces, and thus to construct the environmental variables in the HPM. We borrow this method in our paper, too, which we elaborate in the next section. Despite these works, research on the spatial aspect of urban green space and their ecosystem service values is quite limited in both types of spatial characteristics involved and real-world cases covered. All in all, to remedy the lack of relevant research, a systematic examination on the spatial effects of urban green spaces on real estate values in a real-world setting is needed. 3. Methodology 3.1. Data 3.1.1. Urban green spaces and other basic geographical features We use an official Geographic Information System (GIS) database of Beijing, produced in 2004, as our basic geographical database in this paper. The database consists the location and shape of all major urban green spaces in the city, as well as other geographical features such as administrative boundaries, streets, roads, and public transit lines and stops which are necessary for computing the neighborhood variables for the HPM. The key geographical features are shown in Fig. 1, and the location and shapes of Beijing's urban green spaces are shown in Fig. 2. According to the data, Beijing's urban green spaces occupy about 10% of the area in the city's Inner Eight Districts,2 which is also the spatial extent covered in this study (i.e. whenever referred, “Beijing” means the Inner Eight Districts of the city). It should be noted, however, that the urban green spaces here do not include all urban green spaces in the general sense. According to China's land use classification system, urban land are classified into certain main categories, including residential, commercial, industrial, transportation, and green spaces, etc. However, within the first-order categories such as residential or commercial land use, there are second-order land use sub-categories, such as green

2 Including: Dongcheng, Xicheng, Chongwen, Xuanwu, Haidian, Chaoyang, Fengtai, Shijingshan. Chongwen and Xuanwu had been annexed to Dongcheng and Xicheng Districts, respectively in 2010.

spaces, which in some cases could occupy as much as 35% of the total area. These second-order land uses are not designated at the master planning level. Rather, they are designated ad hoc when it comes to the detailed planning of specific land parcels, and therefore these second-order land uses, such as green spaces, do not count in the statistics of the respective land use class. As a result, the database that we use in this study only consist of the land parcels that are classified into the first-order “green spaces” land use class. This data quality issue may lead to certain estimation errors, as previous studies have otherwise pointed out (Derkzen, van Teeffelen, & Verburg, 2015). In the absence of any better geographical data source of urban green spaces, we use the above mentioned database anyway. However, conceptually, the omission of second-order urban green spaces may almost surely result in estimation errors in the following modeling. We address this problem whenever encountered later in this paper. 3.1.2. Land transaction records We use the complete record of developable residential land transactions in Beijing from 2000 to 2004 as basic dataset for analysis. As noted above, the HPM requires market price data of housing or land transactions. In this study, the main reason for using land rather than housing data is data availability concerns. With reliable official or third-party database for housing transactions unavailable, there seems to be no alternative but to use the land data for which the full record is available. Moreover, there are evidences that the housing market during that period was dominated by speculation and therefore bore a bubble (Hui & Yue, 2006), thus violating the market equilibrium assumption which is essential for the HPM. The land transaction data, however, has its own problems. On the one hand, as noted above, the database consists of all residential land transaction records during the study period, thus constituting a valuable source of land market data. On the other hand, however, whether the land price-formulating mechanism in Beijing during the study period was completely free-market based is questionable, which may result in some problems. We explain the problem as follows. China's urban land is owned by the state by law, and the urban land market (strictly speaking, it is the market for the leasehold rights of urban land) has been evolving since the 1988 real estate reform. Different land price-formulating mechanism dominates in different periods, and these mechanisms form a full spectrum from totally central-planned economy to totally free-market styles. The traditional way, known as the “land acquisition by governmental allocation” approach, is not a market-based mechanism at all. Indeed, land price under such a mechanism is usually 0. Since early 1990s, however, a new market-based mechanism began to prevail, which is known as the “land transfer by agreement” approach. Under this mechanism, land price is based on the agreement between the local government and the entities seeking leasehold rights, and market-based methods, such as the market comparison method and the replacement costs method, have been developed to determine the land price. However, though having some market features, this mechanism is, to say the most, a semi-marketized one (Mak, Choy, & Ho, 2007), as the procedure of land price determination is usually not transparent, and there is usually no free entry to the land market for all market entities. Lastly, first experimented in some cities in 2002 and becoming prevail nationwide after Aug 31, 2004, a new, fully marketized mechanism, known as the “land transfer by bidding, auction, and listing” mechanism was introduced, and land price under this mechanism should represent a totally free-market value. Specifically, in Beijing, the “bidding, auction, and listing” mechanism was first introduced in Dec 8, 2003, and became the

L. Xu et al. / Habitat International 56 (2016) 84e95

87

Fig. 1. Key geographical features of the study area.

only mechanism after Sep 1, 2004. This means that in our land transaction database, which covers the land transaction records between 2000 and 2004, only a fraction of the transactions are based on the “bidding, auction, and listing” mechanism, and the rest are all based on the “agreement” mechanism. As noted above, the latter is not a fully marketized one. Problem is, we do not know how marketized the prices are. Irrational (by a free market standard) land prices may exist in some cases, rational ones may exist in others. In the absence of a better data source, and considering other merits of the database (for example, China's current land regulation system prohibits any secondary land market, which has repressed any chance of speculation and thus prevented a bubble to form, maintaining a relatively rational market land price), we opt to still use the database. However, the above mentioned problems may affect the analysis results, which we address later when encountered. There are in total 3659 land parcels in the dataset, with information on land area size, land use (residential for all cases), approved development intensity (Floor-to-Area Ratio, FAR), and transaction price for each parcel. We have inflation-adjusted all

prices to reflect constant 2000 Yuan value. The location and price of the land transactions are also shown in Fig. 2. 3.2. Measurements of the spatial pattern: Landscape Ecological Metrics The quantification of the various spatial characteristics of urban green spaces is key to the study, for which Landscape Ecological Metrics are the most widely used tool (Wu & Hobbs, 2002). In this paper, we select four indices to measure the richness, accessibility, distribution, and shape configuration characteristics of green space patches, as shown in Table 1. We compute the DIST metric in ESRI ArcGIS, and all others in Fragstats 4.2. As the CAR, LDI, and LSI metrics are only meaningful in a given area of landscape, one must first designate the specific landscape area to calculate these metrics. In this study, we designate a 500  500 m window as the basic landscape area for analysis, as it is roughly the average distance between green space patches in the study area such that at least one patch is included in each window. Moreover, to get the metric values within the whole

88

L. Xu et al. / Habitat International 56 (2016) 84e95

Fig. 2. Green space parcels and land transactions in Beijing, 2000e2004.

landscape (study area), we perform a convolution with a sliding window of the 500  500 m size mentioned above, and calculate the average landscape metric values within the landscape. The results are shown in Fig. 3. It should to be noted here that due to the existence of some “No Data” areas produced in the convolution process, the real estate transaction records which happen to locate in these “holes” cannot get a value of the metrics, and thus must be excluded from the respective model. Hence, though the model with the DIST variable can include all 3659 samples, the ones with the CAR, LDI and LSI variables only have a sample size of 2922. 3.3. Specification of the Hedonic Price Model We specify the HPM in this study following Equation (1). The HPM typically takes a semi-log form, with the dependent variable logarithmic transformed, for two reasons. First, housing or land price samples are usually lognormal distributed (Poudyal et al., 2009). Second, researchers usually focus on the changes rather than the absolute levels of the dependent variable, the semi-log equation form is thus more appropriate. In our case, moreover, preliminary data explorations show that the landscape metrics are also lognormal distributed. Therefore, we use a modified semi-log

OLS regression to estimate the model, with both the dependent variable (land price) and the environmental variables (landscape metrics) logarithmic transformed. Regarding the variables, we first include two variables: land parcel size and approved structure area size in the structural group Sij . As noted above, all land parcels are of the same designated land use (residential), so land use is not included as a variable here. Second, we include five variables: distances to city center, district center, nearest highway entrance, nearest urban street, and nearest bus stop in the neighborhood group Nik . It should be noted that not all the variables would necessarily enter the final model, for issues such as multicollinearity may render some variables statistically insignificant. Lastly and most importantly, variables in the environmental group Eil include the spatial metrics discussed above. However, the specific way in which the variables enter the model is worth noting here. A search of relevant literatures reveals two ways of including environmental variables in the HPM. On one hand, the first approach is to directly use the values of landscape metrics. For example, one can construct a variable indicating a land parcel's accessibility to urban green spaces using the distance from the land parcel to the nearest green space patch, as in the case of Kong, Yin,

L. Xu et al. / Habitat International 56 (2016) 84e95

89

Table 1 Landscape ecological metrics in the study. Spatial characteristics

LEM

Richness

Class Area Ratio (CAR) Nearest Distance (DIST) Landscape Division Index (LDI)

Accessibility Distribution Shape Configuration

Landscape Shape Index (LSI)

LEM definition* P CAR ¼ nj¼1 aij

Notes Ratio between the area of a land use class (urban green space in this case, and the same hereinafter) and the whole landscape area (or the area of the convolution sliding window) (m2) Distance from a real estate site to the nearest urban green space patch (m)

DIST ¼ dnearest LDI ¼ 1  ei LSI ¼ mine

i

Pn

j¼1

 2 aij A

A measurement of the fragmentation level of a land use class in a certain area. Ranging from 0 (Not fragmented at all) to 1 (Perfectly fragmented) A measurement of the shape configuration of patches, LSI equals the total length of edge (or perimeter) involving the corresponding class, divided by the minimum length of class edge (or perimeter) possible for a maximally aggregated class, which is achieved when the class is maximally clumped into a single, compact patch (McGarigal, Cushman, & Ene, 2012). LSI ¼ 1 implies a square patch; as LSI increases, the patch gets more complex in shape.

*Where aij ¼ area (m2) of patch ij; dnearest ¼ distance to the nearest patch; ni ¼ number of patches in the land use class (type) i; A ¼ total landscape area (m2); ei ¼ total length of edge (or perimeter) of class i in terms of number of cell surfaces; min ei ¼ minimum total length of edge (or perimeter) of class i in terms of number of cell surfaces; includes all landscape boundary and background edge segments involving class i. In this paper, there is only one possible value for i e the green spaces. The definitions of the metrics are given by Fragstats(McGarigal, Cushman, and Ene 2012). For detailed description of the indices, refer to the source above.

and Nakagoshi (2007) and Poudyal et al. (2009). The second approach, on the other hand, uses dummy variables. Specifically, it first designate a few sections across the range of a landscape metric, and then construct dummy variables accordingly, thus convey the spatial pattern information in an indirect way. Examples of this approach include Bolitzer and Netusil (2000) and Tyrvainen and Miettinen (2000). The first approach would directly reveal the influence of the spatial effects on the dependent variable (land price), but would also require high-quality data supports. The second approach, on the contrary, is less effective in revealing the specific influence of the spatial effects, but also requires weaker data support. Therefore, we face a trade-off between data quality and modeling efficiency. As discussed above, both the urban green space geographical data and the land price data used in this study could be far from perfect, implying the potential benefit of adopting the second approach. Although this is a compromise we have to make because of data limitations, we argue that imperfect data, rather than high-quality data, is the normal in most real-world settings, especially in developing countries such as China. In this sense, the second approach has its unique merits. It should also be noted that by following the second modeling approach, the main purpose of this study is not to quantify the monetary value “premium” brought about by the environmental amenities (though we try to give some rough estimations later in the paper). Rather, we focus on the very existence (or not) of the spatial effects per se. To this end, we need only to do the following hypothesis test at a given (for example, 5%) significance level:

H0 : bl s0 and H1 : bl ¼ 0

We do the hypothesis tests in a series of models to find the effective range of spatial characteristics in which the spatial effects exist. Taking the DIST variable as an example, we first set a series of cutoff values, ranging from 50 m to 100, 150, ..., 1000 m, and then create a 0/1 dummy variable and assign the value for it based on the cutoff values. For the cutoff value 50 m, for instance, we create a dummy variable DIST 50, which is defined as follows: DIST50 ¼ 1, if DIST <50 m; DIST50 ¼ 0, Otherwise Then, we use the new dummy variable DIST 50 to fit the first model, denoting HPM (DIST 50). If the regression coefficient of the dummy variable is statistically significant, then we can safely conclude that the presence of urban green spaces within 50 m of a land development site does influence its value. Similarly, we create a series of dummy variables with different cutoff values and fit a group of models accordingly (Table 2). Then we only need to focus on the regression coefficients of the dummy variables (DISTCOV), and identify the range of cutoff values within which the coefficients of DISTCOV are statistically significant, thus indicating the effective range of the spatial effects. The criteria for setting the cutoff values is worth noting here. On the one hand, for those landscape metrics with physical units (for example, DIST, measured in meters), we simply arbitrarily set a series of cutoff values such that the entire value range is evenly and reasonably divided, a method similar to other researchers' (Bolitzer & Netusil, 2000). On the other hand, for the standardized metrics

Fig. 3. The spatial pattern (LEM values) of the CAR, LDI and LSI variables.

90

L. Xu et al. / Habitat International 56 (2016) 84e95

Table 2 The group of HPMs for an environmental variable (DIST). Model no.

Model name

DIST cutoff Value

1

HPM (DIST

50)

50

2

HPM (DIST

100)

100

COVn)

… COVn

… n

HPM (DIST

which don't have an explicit physical meaning, we divide their value ranges such that the sample size within each subinterval is the same to avoid possible bias induced by radically skewed distribution of samples among the subintervals. Moreover, in both occasions we make sure there are not too many subintervals that unnecessarily complicates the analysis. In sum, the variables, their descriptive statistics, and expected signs of regression coefficients are shown in Table 3. One major drawback of the above described modeling approach, though, is that it eliminates the possibility of constructing a “complete” model including all spatial-related characteristics. For example, if we designate m dummy variables for spatial characteristic 1 and n dummy variables for spatial characteristic 2, then we would need m * n equations in a “complete” model group. As the number of spatial characteristics examined increase, it would soon become manifest that it is impractical to include all the spatial characteristics in the same model. Besides, including all spatial characteristic variables in the same model may not only be impractical, but also undesirable, as different landscape metrics deal with different aspects of spatial characteristics. For example, it is easy to find that the four landscape metrics used in this paper are not correlated with each other at all, which is manifest as revealed in the definitions of the landscape metrics per se. Thus, interactive terms of different landscape metrics hardly make any practical sense. Therefore, instead of fitting a general but very complicated model with all the environmental variables, we construct a series of HPMs, with all possible variables in Sij and Nik in every model, but only one spatial metric in each model as the Eil variable, such that we focus on one spatial dimension at a time. Overall, each model group would reveal the marginal effective range of the respective spatial pattern, so that one can directly combine the marginal results from each model group to reach a comprehensive conclusion.

4. Results Results show that, for different models with different environmental variables, only three independent structural and neighborhood variables are statistically significant (at the 5% significance level) to enter all models. These are the approved construction area on the land parcel (built_size), the distance to the city center (dist_city), and the distance to the district center (dist_district). Moreover, the regression coefficients for all the three variables remain quite stable not only within each model group, but also across different model groups. The coefficients for the environmental dummy variable, however, vary dramatically between models within each group, with some statistically significant while others not (Table 4). The fact means that the way the dummy variables are constructed does influence their relevance. Putting another way, the spatial dimension does matter.3 We discuss the model groups one by one in this section.

3 To avoid excessive details, we only demonstrate the ranges of R2, regression coefficients (Beta), and T-values in each model group. Detailed information is available upon request.

Regression equation P P P ln pi ¼ b0 þ bj Sij þ bk Nik þ bl DIST50i þ εi P P P ln pi ¼ b0 þ bj Sij þ bk Nik þ bl DIST100i þ εi … P P P ln pi ¼ b0 þ bj Sij þ bk Nik þ bl DISTCOVn i þ εi

4.1. Model group 1: the richness and ecosystem service value of urban green spaces The first model group, with the land use area ratio (CAR) of urban green spaces being the environmental variable, examines whether e and if yes, how the richness of urban green spaces influences their ecosystem service value as reflected in land prices. As shown in Table 5, the answer to the first question depends on how much area is occupied by green spaces in the urban landscape. The dummy variable's regression coefficient is not statistically significant when the cutoff value is small, but becomes significant (at a 5% significance level) as the cutoff value increases to reach a tipping point, when green space patches occupy 2.20% (i.e. exp(3.8073)) of the total area, and then basically remains significant until the cutoff value reaches a high threshold (between 9.20% and 13.40% green space ratio). A notable problem, however, is that the regression coefficients (standard beta) for the CAR variable are negative in all models, and show no trend between models. This is against the intuitive expected sign of the variable, and is also inconsistent with the results from other researchers (McConnell & Walls, 2005). We elaborate on this issue in the discussions section. Nevertheless, despite the negative regression coefficients, the main result of the model group, i.e. that the existence of urban green spaces would contribute to nearby land value when the former's size ratio in the landscape exceeds a certain threshold, still holds.

4.2. Model group 2: the accessibility and ecosystem service value of urban green spaces The second model group, with the distance between land development parcels and the nearest urban green space patch (DIST) being the environmental variable, examines whether e and if yes, how the accessibility of urban green spaces influences their ecosystem service value as reflected in land prices. As shown in Table 6, when a green space patch is within 50 m of a land parcel, its influence on the land price is statistically insignificant. However, when a green space patch is within the 50e550 m range of distance from a land development parcel, its influence on the land price is statistically significant in all models. Lastly, when the distance between a green space patch and a land development parcel is larger than 600 m, the former's influence on the land price fades away again as the respective regression coefficients again become statistically insignificant. Overall, these results are in line with common sense, except for the 50-m lower threshold of the spatial effects. We argue that this lower threshold is very likely an estimation error, and the error comes from a special data limitation. As indicated above, we only include first-order urban green spaces. These green space patches usually occupy entire urban blocks, and are separated from other land parcels by major urban streets, which, in the specific case of Beijing, are seldom narrower than 50 m. Thus, in principle, the number of land transaction sites that are within 50 m of nearby urban green space parcels should be very small. In fact, there are only 64 such cases in the land transactions database, and even

L. Xu et al. / Habitat International 56 (2016) 84e95

91

Table 3 Definitions, descriptive statistics, and expected signs of independent variables.a Variable

Definition/Unit

Mean

Common variables in all models Land_size Area of parcel/m2 Built_size Approved construction area on parcel/m2 DIST_city Distance to city center/m DIST_district Distance to district center/m DIST_highway Distance to nearest highway entrance/m DIST_urbanstreet Distance to nearest urban street/m DIST_busstop Distance to nearest bus stop/m Dummy variable groups in different model groups CARk See Table 1. DISTl LDIp LSIq a b c

Std. deviation

Expected sign

Sample 1b

Sample 2c

Sample1

Sample 2

16499.97 48474.85 9455.55 5559.41 751.21 768.46 402.00

15476.98 47531.69 9229.29 5145.43 661.73 792.48 380.18

26974.71 64747.60 4697.87 2778.54 811.29 934.30 276.43

25771.64 65426.55 4571.87 2553.09 671.98 979.38 253.09

N/A

þ þ e e e e e þ þ - when LDI is too large or too small, þ otherwise Not clear

N/A

Dependent variable ¼ Land price. Observations ¼ 3659. Observations ¼ 2922.

Table 4 Summary of regression results. Model group No Observations Environmental Variable R2

Structural Variable(s) entered Built_size

1

2922

CARk

2

3659

DISTl

3 4

2922 2922

LDIp LSIq

High Low High Low High Low High Low

0.33 0.32 0.35 0.34 0.33 0.32 0.33 0.32

Neighborhood Variable(s) entered

Environmental Variable

Dist_city

Std. Beta

T Value

0.01 0.05

0.54 3.26

Dist_district

Std.Beta

T Value

Std. Beta T Value Std.Beta T Value

0.54 0.53 0.55 0.54 0.54 0.53 0.54 0.53

35.19 35.04 40.75 40.70 35.15 34.99 35.18 35.10

0.15 0.16 0.17 0.17 0.15 0.16 0.16 0.16

these 64 cases are possibly produced by mapping errors. Putting another way, the entire “less than 50 m” category makes little sense in the specific study context, so is the insignificant regression coefficient. Like in the previous model group, the DIST variables in this model group show negative regression coefficients, too, still with no trend between models. Similarly, we discuss possible explanations in the next section. 4.3. Model group 3: the distribution and ecosystem service value of urban green spaces The third model group, with the degree of division (LDI) of green spaces being the environmental variable, examines whether e and if yes, how the fragmentation status of urban green spaces

9.79 10.49 12.47 12.64 9.72 10.38 10.46 10.60

0.06 0.07 0.07 0.08 0.06 0.07 0.07 0.07

4.05 4.29 5.01 5.47 4.05 4.26 4.22 4.34

0.01

0.05

0.61

3.39

0.04

2.84

0.00

0.10

0.01

0.03

0.83

1.83

influences their ecosystem service value as reflected in land prices. By definition, given total urban green space area, the larger the LDI value is, the more fragmented the green space patches are; therefore, the LDI is in some sense a comprehensive indicator for the spatial distribution of urban green space in a landscape. As shown in Table 7, results show both an upper and a lower threshold in the ln LDI value range, and only the models with LDI cutoff values in between yields statistically significant results in which the positive regression coefficients indicate a positive influence on nearby land values when the degree of division of urban green spaces is within the range, an instinctive result that is in line with our expectation. Other models have insignificant environmental variable coefficients, thus yielding no meaningful results. We elaborate on the physical meaning of the ln LDI range in the next section.

Table 5 Regression results for the class area ratio (CAR) model group. Sub-model no.

Environmental dummy Variable name

Ln CAR

1 2 3 4 5 6 7 8 9 10

CAR_1 CAR_2 CAR_3 CAR_4 CAR_5 CAR_6 CAR_7 CAR_8 CAR_9 CAR_10

5.452 4.759 4.232 3.807 3.525 3.258 2.977 2.702 2.385 2.006

The bold, italics figures indicate significance levels that are below 0.05.

cutoff

Std. Beta

Sig.

0.008 0.010 0.027 0.032 0.027 0.030 0.042 0.050 0.036 0.023

0.591 0.506 0.081 0.038 0.082 0.049 0.007 0.001 0.018 0.128

92

L. Xu et al. / Habitat International 56 (2016) 84e95

4.4. Model group 4: the shape configuration and ecosystem service value of urban green spaces The last model group, with the LSI of green spaces being the environmental variable, examines whether e and if yes, how the shape configuration of urban green spaces influences their ecosystem service value as reflected in land prices. By definition, the LSI is an indicator that measures “how much” a shape is different from a square, with the value of 1 when the shape is a square, and larger values when the shape is “deviated” from a square. As shown in Table 8, no model in this group yields a statistically significant regression coefficient for the environmental variable (LSI), thus giving no meaningful results. We discuss on possible explanations for these results in the next section. 5. Analysis and discussions Among the three model groups yielding statistically significant results, on the one hand, the first and second one (on CAR and DIST, respectively) give results with explicit physical meanings (in terms of green space ratios and distances, respectively), though the results are in some cases against the common sense. We discussion the possible causes to this problem. The model group for LDI, on the other hand, does not give results with explicit physical meanings, as the LDI is a standardized metric that does not have any explicit physical meaning. However, based on the regression results, we can “interpret” the results by quantitatively calculating the range for urban green spaces’ degree of fragmentation in which they are influential on real estate values, as well as the optimal degree of fragmentation in which the most influential effects are realized. Also, we discuss on possible causes for the non-significant results in the Landscape Shape Index (LSI) model group, and give an explanation on it. 5.1. On the counter-intuitive regression coefficients in the richness and accessibility model groups In the regression results of the richness and accessibility model groups, a most notable problem is that the regression coefficients for the CAR and DIST variables are negative in all models that are statistically significant, and show no trend between models. These

results are clearly counter-intuitive, as they imply that increased richness and accessibility to urban green spaces would decrease nearby real property values whenever the spatial effects are effective. We attribute the cause of the problem to data quality issues. As discussed above, both the land price data and the geographical data on urban green spaces may be problematic, and thus constitute sources of errors. On the one hand, with regard to the geographical urban green space data, the dataset may not be complete due to administrative reasons, and the omission of certain second-order green spaces may affect the model estimations. A refined geographic data on urban green spaces, therefore, may help eliminate such errors. On the other hand, with regard to the land price data, it should be noted that Beijing's land market during the study period was not a totally free-market one, and that the possibly irrational land prices may result in the strange negative regression coefficients. The problem, therefore, may disappear in a real free-market context. We admit that these problems are the major limitations of this study, and they are subject to future inquiries. Nevertheless, despite the unexpected signs of the regression coefficients, the results concerning the effective range of the spatial effects are robust. In sum, green spaces should occupy between 2.20% and 13.40% of the total urban landscape area, and within a 100e550 m range from nearby real estate development sites, such that the spatial effects are effectively realized.

5.2. On the degree of fragmentation of urban green spaces: the effective range and optimal level The LDI measures the degree of fragmentation of a land use class (i.e. urban green spaces) in a landscape, and is given in the form of the ratio between green space patches and the landscape area. Therefore, to calculate the thresholds and optimal levels of fragmentation of urban green spaces as measured by the area ratio between the patches and the total green space area, one needs additional information regarding the ratio of area occupied by green spaces in the urban landscape. In this paper's particular context of Beijing, as mentioned earlier, the figure is about 10% within the study area. We base the following calculations on this figure. As shown in Table 7, the lower and upper thresholds for statistically significant regression coefficients for the ln LDI variable

Table 6 Regression results for the Accessibility (DIST) model group. Sub-model no.

Environmental dummy variable name

DIST

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

DIST_1 DIST_2 DIST_3 DIST_4 DIST_5 DIST_6 DIST_7 DIST_8 DIST_9 DIST_10 DIST_11 DIST_12 DIST_13 DIST_14 DIST_15 DIST_16 DIST_17 DIST_18 DIST_19 DIST_20

50 100 150 200 250 300 350 400 450 500 550 600 650 700 750 800 850 900 950 1000

The bold, italics figures indicate significance levels that are below 0.05.

cutoff

Std. Beta

Sig.

0.014 0.032 0.032 0.029 0.032 0.044 0.047 0.043 0.032 0.034 0.034 0.022 0.013 0.007 0.003 0.003 0.000 0.008 0.001 0.000

0.308 0.017 0.018 0.031 0.016 0.001 0.001 0.001 0.018 0.012 0.012 0.111 0.340 0.589 0.802 0.843 0.983 0.557 0.933 0.992

L. Xu et al. / Habitat International 56 (2016) 84e95

93

Table 7 Regression results for the landscape division index (LDI) model group. Sub-model no.

Environmental dummy variable name

Ln LDI

Std. Beta

Sig.

1 2 3 4 5 6 7 8 9 10

LDI_1 LDI_2 LDI_3 LDI_4 LDI_5 LDI_6 LDI_7 LDI_8 LDI_9 LDI_10

0.0278 0.0118 0.0064 0.0036 0.0020 0.0010 0.0006 0.0003 0.0001 0.00003

0.023 0.025 0.033 0.039 0.044 0.043 0.036 0.031 0.007 0.002

0.129 0.102 0.034 0.013 0.005 0.006 0.018 0.045 0.654 0.917

Std. Beta

Sig.

0.007 0.013 0.002 0.015 0.017 0.007 0.000 0.000 0.007 0.023 0.026 0.021 0.021 0.024 0.007 0.009 0.006 0.020 0.028

0.643 0.409 0.871 0.336 0.258 0.659 0.995 0.982 0.651 0.131 0.092 0.162 0.172 0.111 0.650 0.556 0.715 0.189 0.067

cutoff

The bold, italics figures indicate significance levels that are below 0.05.

Table 8 Regression results for the landscape shape index (LSI) model group. Sub-model no.

Environmental dummy variable name

Ln LSI

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19

LSI_1 LSI_2 LSI_3 LSI_4 LSI_5 LSI_6 LSI_7 LSI_8 LSI_9 LSI_10 LSI_11 LSI_12 LSI_13 LSI_14 LSI_15 LSI_16 LSI_17 LSI_18 LSI_19

0.1335 0.2231 0.3064 0.3795 0.4418 0.5108 0.5718 0.6286 0.6931 0.7481 0.7885 0.8313 0.8928 0.9372 0.9860 1.0427 1.0935 1.1477 1.2303

are 0.0064 and 0.0003, respectively, and the value of ln LDI when the regression coefficient is at the maximum (i.e. realizing the most ecosystem service value) is 0.0020. Thus:

LDILower ¼ expð0:0064Þ ¼ 0:9936 LDIUpper ¼ expð0:0003Þ ¼ 0:9997 LDIOptimal ¼ expð0:0020Þ ¼ 0:9980 Without loss of generality, assuming all green space patches in the landscape are evenly divided, and there are n patches in total. By the definition of LDI:

1  LDI ¼ n

 2 aj A

(2)

Let g denote the overall green space ratio in the landscape:

naj ¼g A

(3)

Insert Equation (3) into Equation (2), and apply g ¼ 10%, we can calculate that the lower and upper thresholds for n is between 1.60 and 33.00, respectively, and the optimal value is 5.00. Putting in terms of area ratios, the ratio each green space patch should occupy in total green space area is 3.00%, 62.50%, and 20.00% at the lower, upper, and optimal levels, respectively. This is to say, in order to effectively realize the ecosystem service value embedded in urban green spaces, the green space patches should be so distributed that

cutoff

every patch occupies more than 3.00% but less than 62.50% of the total green space area, and when each patch occupies 20.00% of the total green space area, the ecosystem service value is optimally realized. It should be noted, however, that these figures are only effective for the Beijing case as they are calculated based on city's land market data, and also assuming a 10% total green space ratio which is also specific of Beijing's context. Nevertheless, the analysis and calculation methods are universally applicable. We thus expect these methods be tested, applied, and further developed in other contexts by cohorts in the academic community. 5.3. An explanation for the non-significant results in the LSI model group The models with LSI as the environmental variable have yielded no statistically significant regression coefficients for the LSI variable. We attribute these results to the innate ambiguousness in the very definition of the LSI. As stated earlier, the LSI by definition only measures how much a shape is “deviated” from a square, but this does not mean that a monotonously and continuously changing LSI series necessarily implies a “continuously” changing shape series as well. For example, it is easy to construct a shape series as shown in Fig. 4. The series is actually composed of two alternating sub-series, with the “square” series becoming more and more convex, and the “half-circle” series becoming more and more concave, but the LSI values for the composed series monotonously increase. In fact, this example is of specific meaning from a landscape ecological perspective, which states that the concave and convex patches in a

94

L. Xu et al. / Habitat International 56 (2016) 84e95

Fig. 4. Two alternating shape series with monotonously increasing LSI values.

landscape have different edge effects, and therefore such two so evolving series may lead to opposite ecological effects, such as ecosystem service value (Forman, 1995), a result that would not otherwise have been revealed by the seemingly monotonously changing LSI series. Therefore, the LSI alone does not provide enough information that can reflect all necessary shape configuration characteristics of a green space patch. Therefore, a model solely based on it hardly makes sense, thus naturally the non-significant results. In contrast, indicators such as DIST and LDI convey clear physical meanings (either explicitly or implicitly), thus the definite results. For a better measurement of the shape configuration characteristics of green space patches, one needs to introduce additional indicators, which is beyond the scope of this paper and we leave it for further studies. 6. Conclusion The study clearly shows that the spatial pattern of urban green spaces does matter when it comes to the realization of their ecosystem service value. Though the direction and magnitude of the spatial effects are in some cases unclear, or cannot be determined in other cases as limited by the quality of data, the existence of the spatial effects per se is manifest. Specifically, the distance from a green space patch to a nearby land development parcel determines whether the green space patch's ecosystem service value can contribute to the land parcel's market value, and green space patches with different sizes and distributional characteristics may have varying degrees of value-adding effect. Furthermore, an optimal level of spatial fragmentation of urban green spaces that maximizes the value-adding effect also exists. To our knowledge, this is the first piece of work that systematically examine the different aspects of the spatial pattern of urban green spaces and their influence on urban green spaces' ecosystem service value, thus these findings constitute the main contribution of this study to the academic literature. Moreover, with the additional information of overall green space ratio, quantitative estimations on the conditions on which the spatial effects exist can be made specifically for the Beijing case. It is suggested that green spaces should occupy between 2.20% and 13.40% of the urban landscape area, and within a 100e550 m range from nearby real estate development sites. Also, green space patches should be so distributed that every patch occupies more than 3.00% and less than 62.50% of the total green space area to ensure effective ecosystem service value-adding effect, and that the effect is optimized when every green space patch occupies 20.00% of the total green space area. Again, it should be noted that these quantitative results are context-dependent and are therefore not to be simply generalized. The major limitation of the study is that due to data quality issues, we are not able to find the specific magnitude of the spatial effects. The implication of this limitation is twofold. On the one hand, since the lack of high-quality data is the normal condition in most cases in a develop country context, the analysis method that

we develop in this study can be applied in such conditions when only imperfect data are available, and they can still yield results that are meaningful to some extent. This illustrates the universal applicability of our modeling and estimation methods. We thus expect other researchers apply the method in various contexts so as to yield more empirical results. On the other hand, however, the specific magnitude of the spatial effects are indeed an important aspect of the inquiry. Though the original database we use in the study has little potential of being improved, we do expect that we analyze the same study area in a different time period, hopefully with fully marketized price information and refined geographical data of urban green spaces, such that estimates on the magnitude of the spatial effects can be made, and that comparative studies can be done. Also, the analysis on the shape configuration characteristics of urban green spaces yields no results. As discussed above, this would require the construction of more effective measurements of the shape configuration characteristics of urban green spaces. We also leave this methodological study for future inquires. Further, in practical terms, a direct practical implication of the study is that urban green space systems can, and should be spatially optimized such that they realize the most ecosystem service value. Such optimization should be carried out in a systematical manner, which means not only a certain aspects of the spatial characteristics of urban green spaces are at concern; rather, all relevant aspects of the spatial characteristics should be included. Items to optimize thus include the specific way green spaces and other built environment elements are spatially arranged, the size of the green space patches, the way the green space patches are divided, and also perhaps the shape configuration of the green space patches. Given the prevailing urban environment problems across the globe, we suggest such optimization task be explicitly included in urban planning codes for all cities, and that the results from a number of empirical research projects may help contribute to the making of a general quantitative guideline for the spatial planning of green spaces in the urban environment in a way to optimize their ecosystem service value. Acknowledgments We would like to thank the Beijing Municipal Administration of Land Resources for permission of use of the land transaction records data and the Beijing GIS database. References Bateman, I. J., Harwood, A. R., Mace, G. M., Watson, R. T., Abson, D. J., Andrews, B., et al. (2013). Bringing ecosystem services into economic decision-making: land use in the United Kingdom. Science, 341(6141), 45e50. http://science. sciencemag.org/content/341/6141/45.abstract. Bingham, G., Bishop, R., & Brody, M. (1995). Issues in ecosystem Valuation: Improving information for decision making. Bolitzer, B., & Netusil, N. R. (2000). The impact of open spaces on property values in Portland, Oregon. Journal Of Environmental Management, 59(3), 185e193. http:// gateway.isiknowledge.com/gateway/Gateway.cgi?GWVersion¼2&SrcAuth¼ AegeanSoftware&SrcApp¼NoteExpress&DestLinkType¼FullRecord&DestApp¼

L. Xu et al. / Habitat International 56 (2016) 84e95 WOS&KeyUT¼000089385700002. Booth, D. B., & Jackson, C. R. (1997). Urbanization of aquatic systems: degradation thresholds, stormwater detection, and the limits of mitigation. Journal of the American Water Resources Association, 33(5), 1077e1090. Chau, K. W., Ma, V. V., & Ho, D. C. W. (2001). The pricing of‘Luckiness’in the apartment market. Cilliers, S., Cilliers, J., Lubbe, R., & Siebert, S. (2012). Ecosystem services of urban green spaces in african countriese-perspectives and challenges. Urban Ecosystems, 16(4), 681e702. http://dx.doi.org/10.1007/s11252-012-0254-3. Costanza, R., Arge, R., Rudolf de, G., Farber, S., Grasso, M., Hannon, B., et al. (1997). The value of the world's ecosystem services and natural capital. Nature, 387(6630), 253e260. http://dx.doi.org/10.1038/387253a0. Costanza, R., Stern, D., & Fisher, B. (2004). Influential publications in ecological economicas: A citation analysis. Czembrowski, P., & Kronenberg, J. (2016). Hedonic pricing and different urban green space types and sizes: insights into the discussion on valuing ecosystem services. Landscape and Urban Planning, 146(February), 11e19. http://dx.doi.org/ 10.1016/j.landurbplan.2015.10.005. ~ uelos, A., Bruse, M., Bürger, M., et al. (2004). An De Ridder, Kon, Adamec, V., Ban integrated methodology to assess the benefits of urban green space. Science of the Total Environment, 334, 489e497. Elsevier. De Vries, Sjerp, Verheij, R. A., Groenewegen, P. P., & Spreeuwenberg, P. (2003). Natural environments-healthy environments? an exploratory analysis of the relationship between greenspace and health. Environment and Planning A, 35(10), 1717e1732. Derkzen, Marthe, L., van Teeffelen, A. J. A., & Verburg, P. H. (2015). Review: quantifying urban ecosystem services based on high-resolution data of urban green space: an assessment for Rotterdam, the Netherlands. Journal of Applied Ecology, 52(4), 1020e1032. http://dx.doi.org/10.1111/1365-2664.12469. Economy, & Elizabeth, C. (2011). The River Runs Black: The Environmental Challenge to China's Future. Cornell University Press. Forman, R. T. T. (1995). Land Mosaics: The Ecology of Landscapes and Regions. Cambridge university press.  mez-Baggethun, E., & Barton, D. N. (2013). Classifying and valuing ecosystem Go services for urban planning. Ecological Economics, 86(February), 235e245. http://dx.doi.org/10.1016/j.ecolecon.2012.08.019. €m, S., Breuste, J., et al. Haase, D., Larondelle, N., Andersson, E., Artmann, M., Borgstro (2014). A quantitative review of urban ecosystem service assessments: concepts, models, and implementation. AMBIO, 43(4), 413e433. http://dx.doi.org/ 10.1007/s13280-014-0504-0. Heynen, N., Perkins, H. A., & Roy, P. (2006). The political ecology of uneven urban green space the impact of political economy on race and ethnicity in producing environmental inequality in Milwaukee. Urban Affairs Review, 42(1), 3e25. Sage Publications. Hillsdon, M., Panter, J., Foster, C., & Jones, A. (2006). The relationship between access and quality of urban green space with population physical activity. Public Health, 120(12), 1127e1132. Elsevier. Hui, E. C. M., & Yue, S. (2006). Housing price bubbles in Hong Kong, Beijing and Shanghai: a comparative study. The Journal of Real Estate Finance and Economics, 33(4), 299e327. Springer. Jiang, B., Larsen, L., Deal, B., & Sullivan, W. C. (2015). A doseeresponse curve describing the relationship between tree cover density and landscape preference. Landscape and Urban Planning, 139(July), 16e25. http://dx.doi.org/10.1016/ j.landurbplan.2015.02.018. Jiang, B., Zhang, T., & Sullivan, W. C. (2015). Healthy cities: mechanisms and research questions regarding the impacts of urban green landscapes on public health and well-being. Landscape Architecture Frontiers, 3(1), 24e35. Jim, C. Y., & Chen, W. Y. (2006). Recreation-amenity use and contingent valuation of urban greenspaces in Guangzhou, China. Landscape and Urban Planning, 75(1e2), 81e96. http://gateway.isiknowledge.com/gateway/Gateway.cgi? GWVersion¼2&SrcAuth¼AegeanSoftware&SrcApp¼NoteExpress& DestLinkType¼FullRecord&DestApp¼WOS&KeyUT¼000234771600008. Jim, C. Y., & Chen, W. Y. (2008). Assessing the ecosystem service of air pollutant removal by urban trees in Guangzhou (China). Journal of Environmental Management, 88(4), 665e676. Elsevier. Kong, F. H., Yin, H. W., & Nakagoshi, N. (2007). Using GIS and landscape metrics in the Hedonic Price Modeling of the amenity value of urban green space: a case study in jinan city, China. Landscape and Urban Planning, 79(3e4), 240e252. http://gateway.isiknowledge.com/gateway/Gateway.cgi?GWVersion¼ 2&SrcAuth¼AegeanSoftware&SrcApp¼NoteExpress&DestLinkType¼ FullRecord&DestApp¼WOS&KeyUT¼000244584600005. Lancaster, K. J. (1966). A new approach to consumer theory. The Journal of Political

95

Economy, 74(2), 132e157. Liu, J., & Diamond, J. (2005). China's environment in a globalizing world. Nature, 435(7046), 1179e1186. Nature publishing group. Mak, S. W. K., Choy, L. H. T., & Ho, W. K. O. (2007). Privatization, housing conditions and affordability in the People's Republic of China. Habitat International, 31(2), 177e192. Elsevier. McConnell, Virginia, & Walls, Margaret A. (2005). The value of open space: Evidence from studies of nonmarket benefits. Washington, DC: Resources for the Future. McGarigal, K., Cushman, S. A., & Ene, E. (2012). FRAGSTATS v4: spatial pattern analysis program for categorical and continuous maps. University of Massachusetts, Amherst. In Computer software program produced by the authors at the Uni-Versity of Massachusetts. McMichael, A. J. (2000). The urban environment and health in a world of increasing globalization: issues for developing countries. Bulletin of the World Health Organization, 78(9), 1117e1126. SciELO Public Health. Morancho, A. B. (2003). A hedonic valuation of urban green areas. Landscape and Urban Planning, 66(1), 35e41. http://gateway.isiknowledge.com/gateway/ Gateway.cgi?GWVersion¼2&SrcAuth¼AegeanSoftware&SrcApp¼ NoteExpress&DestLinkType¼FullRecord&DestApp¼WOS&KeyUT¼ 000187753900004. €rtberg, U., & Wallentinus, H.-G. (2000). Red-listed forest bird species in an urban Mo environmentdassessment of green space corridors. Landscape and Urban Planning, 50(4), 215e226. Elsevier. Neumann, B. C., J Boyle, K., & R Bell, K. (2009). Property price effects of a national wildlife refuge: great Meadows National Wildlife Refuge in Massachusetts. Land Use Policy, 26(4), 1011e1019. http://gateway.isiknowledge.com/gateway/ Gateway.cgi?GWVersion¼2&SrcAuth¼AegeanSoftware&SrcApp¼ NoteExpress&DestLinkType¼FullRecord&DestApp¼WOS&KeyUT¼ 000266197500015. Pearce, D. (1998). Cost-benefit analysis and environmental policy. Poudyal, N. C., Hodges, D. G., & Merrett, C. D. (2009). A hedonic analysis of the demand for and benefits of urban recreation parks. Land Use Policy, 26(4), 975e983. http://gateway.isiknowledge.com/gateway/Gateway.cgi? GWVersion¼2&SrcAuth¼AegeanSoftware&SrcApp¼NoteExpress& DestLinkType¼FullRecord&DestApp¼WOS&KeyUT¼000266197500012. Poudyal, N. C., Hodges, D. G., Tonn, B., & Cho, S. H. (2009). Valuing diversity and spatial pattern of open space plots in urban neighborhoods. Forest Policy and Economics, 11(3), 194e201. http://gateway.isiknowledge.com/gateway/Gateway. cgi?GWVersion¼2&SrcAuth¼AegeanSoftware&SrcApp¼NoteExpress& DestLinkType¼FullRecord&DestApp¼WOS&KeyUT¼000266155200006. Rapport, D. J., Gaudet, C., & Karr, J. R. (1998). Evaluating landscape health: Integrating societal goals and biophysical process. Sander, H. A., & Polasky, S. (2009). The value of views and open space: estimates from a hedonic pricing model for Ramsey County, Minnesota, USA. Land Use Policy, 26(3), 837e845. http://gateway.isiknowledge.com/gateway/Gateway.cgi? GWVersion¼2&SrcAuth¼AegeanSoftware&SrcApp¼NoteExpress& DestLinkType¼FullRecord&DestApp¼WOS&KeyUT¼000264647500034. Shao, M., Tang, X., Zhang, Y., & Li, W. (2006). City clusters in China: air and surface water pollution. Frontiers in Ecology and the Environment, 4(7), 353e361. Eco Soc America. Stern, D. I., Common, Michael S., & Barbier, E. B. (1996). Economic growth and environmental degradation: the environmental Kuznets curve and sustainable development. World Development, 24(7), 1151e1160. Elsevier. Tajima, K. (2003). New estimates of the demand for urban green space: implications for valuing the environmental benefits of Boston's Big Dig Project. Journal of Urban Affairs, 25(5), 641e655. http://gateway.isiknowledge.com/gateway/ Gateway.cgi?GWVersion¼2&SrcAuth¼AegeanSoftware&SrcApp¼ NoteExpress&DestLinkType¼FullRecord&DestApp¼WOS&KeyUT¼ 000187291200006. Tyrvainen, L., & Miettinen, A. (2000). Property prices and urban forest amenities. Journal of Environmental Economics and Management, 39(2), 205e223. http:// gateway.isiknowledge.com/gateway/Gateway.cgi?GWVersion¼2&SrcAuth¼ AegeanSoftware&SrcApp¼NoteExpress&DestLinkType¼FullRecord& DestApp¼WOS&KeyUT¼000086021400005. Wolch, J. R., Byrne, J., & Newell, J. P. (2014). Urban green space, public health, and environmental justice: the challenge of making cities ‘just green enough. Landscape and Urban Planning, 125(May), 234e244. http://dx.doi.org/10.1016/ j.landurbplan.2014.01.017. Wu, J., & Hobbs, R. (2002). Key issues and research priorities in landscape ecology: an idiosyncratic synthesis. Landscape Ecology, 17(4), 355e365. Springer. Xie, G., Xiao, Y., & Lu, C. (2006). Study on ecosystem services: progress, limitation, and basic paradigm (In Chinese). Journal of Plant Ecology, 30(2), 191e199.