Assessing spatial pattern of urban thermal environment in Shanghai, China

Stoch Environ Res Risk Assess (2012) 26:899–911 DOI 10.1007/s00477-012-0638-1 ORIGINAL PAPER Assessing spatial pattern of urban thermal environment in Shanghai, China Wenze Yue • Yong Liu • Peilei Fan Xinyue Ye • Cifang Wu • Published online: 22 August 2012  Springer-Verlag 2012 Abstract The aggravating urban thermal environment has considerable adverse effects on urban physical environment, energy consumption, and public health. Due to the complexity of factors contributing to the urban thermal environment, traditional statistical methods are insufficient for acquiring data and analyzing the impacts of human activities on the thermal environment, especially for identifying dominant factors. Based on thermal remote sensing imageries and Geographic Information System analysis, we assessed spatial pattern of urban thermal environment in Shanghai in 2008, and analyzed the factors contributing to the generation of urban heat island (UHI) using principal component analysis (PCA). We found that Shanghai had obvious UHI with uneven spatial pattern in 2008. Further, we identified three most important components leading to the variances of Shanghai’s UHI: the gradient from manmade to natural land cover, landscape configuration, and W. Yue  C. Wu (&) Department of Land Management, Zhejiang University, Hangzhou 310029, China e-mail: wucifang@zju.edu.cn W. Yue e-mail: wzyue@zju.edu.cn Y. Liu College of Resources and Environment, Southwest University, Chongqing 400716, China P. Fan School of Planning, Design and Construction and Center for Global Change and Earth Observation, Michigan State University, East Lansing, MI 48824, USA X. Ye School of Earth, Environment, and Society and Center for Regional Development, Bowling Green State University, Bowling Green, OH 43403, USA anthropogenic heat release. A linear model has thus been successfully constructed, implying that PCA is helpful in identifying major contributors to UHI. The findings are of significance for policy implication to urban thermal environment mitigation. Keywords Urban heat island  Human activities  Principal components analysis  Shanghai 1 Introduction Since the economic reforms of 1978, China has been experiencing an unprecedented development towards marketoriented growth. Rapid economic development (about 10 % annual GDP growth rate during the past 30 years) and urbanization (urban population accounting for 18 % in 1978 but 51 % in 2011) have caused significant changes in land use, with associated degradations in the environment and ecology (Han et al. 2009). One of the ecological consequences of rapid urban expansion is the urban heat island (UHI) phenomenon, which is formed when higher atmospheric and surface temperatures in urbanized areas are observed over the surrounding rural area (Voogt and Oke 1998). UHI, along with urban sprawl and urban spatial restructuring, has become evident for densely populated Chinese cities. UHI has led to various problems, such as deteriorated urban environment, amplified energy consumption for cooling, and growing mortality rates due to high temperature in the city core (Wu et al. 2007; Li et al. 2009). It is important to pay attention to the spatial pattern of UHI and its driving forces for policy implication. Traditionally, UHI was calculated by the air temperature difference, measured by permanent meteorological station or moving observations between urbanized and rural area 123 900 (Carlson and Arthur 2000; Wilson et al. 2003). Recently, more studies used land surface temperature (LST) derived from satellite thermal infrared (TIR) imagery (Li et al. 2011). Apparently, traditional observation data cannot provide a synchronized view of temperature over a city, while remotely sensed data can offer a continuous view of the whole city, which is important to investigate the spatial pattern of urban surface temperature (Weng 2012). Extensive research attentions have been paid to deriving LST by moderate resolution imageries (TM, ETM?, and ASTER). The LST with a 60–120 m resolution significantly facilitated the studies on the spatial pattern of UHI and its relationship with LULC and surface biophysical parameters (Nichol 1996; Weng et al. 2004, 2006; Yuan and Bauer 2007). Therefore it is of great scientific significance to investigate how satellite-derived LSTs can be utilized to characterize UHI phenomenon (Weng 2012). Many scholars have heatedly debated over the trends and underlying forces of UHI. Lo and Quattrochi (2003) argued that the formation of UHI is caused by dense buildings and roads combined with scarce vegetation cover. Chudnovsky and Ben-Dor (2004) pointed out that the physical properties of various types of urban surfaces, their color, the sky view factor, street geometry, traffic loads, and anthropogenic activities significantly contribute to UHI. Zhan et al. (2012) simulated the influences of high buildings in CBD on urban LST, disclosing that during the daytime the buildings increased the surface air temperature by up to 1 C which reinforces the heat island effect. Moreover, UHI can be further intensified by the anthropogenic heat release from automobiles, power plants, air conditioners, as well as pollutants and greenhouse gases that absorb and emit infrared radiation (Chen and Pian 1997). Based on the surface energy balance theory, UHI is mainly generated by the combination of anthropogenic heat discharge due to energy consumption, increased impervious surface area, and decreased vegetation and water area. Although scholars have long agreed over the general causes of UHI, it is not clear how different human activities have contributed to the intensity of UHI (Rizwan et al. 2008). Obviously, UHI has been driven by a coupled human-environment system, involving land use and land cover characteristics, urban construction pattern, population density, and human activities. Urban remote sensing data can be used to explore the interaction between urban environmental process and human activities (Wu and Hobbs 2002). Various methods have been employed to analyze the mechanism of UHI, such as the comparison of statistical index, correlation and regression methods (Yue and Xu 2007; Weng 2012). Most studies focused on biophysical and meteorological factors, such as vegetation coverage (Weng et al. 2004), impervious surface area (Weng et al. 2006; Yuan and Bauer 2007), land cover characteristics (Hart and Sailor 2009; Weng 123 Stoch Environ Res Risk Assess (2012) 26:899–911 et al. 2007), and ecological setting (Imhoff et al. 2010). Social-economic factors have also been emphasized, such as pavement area, building area (Weng et al. 2008), building and population density (Chen et al. 2006; Xiao et al. 2008), and building energy use (Zhou et al. 2011). However, it is not enough for the policy implication in urban planning and urban management, since high correlation existed in these multiple driving factors. In other words, we cannot pay attention to the relationship between only one factor and UHI, while ignoring all other factors (Rizwan et al. 2008). A synthesis understanding of the linkage between the multiple driving factors and UHI is needed. For instance, Weng et al. (2008) employed a factor analysis method to reveal the relationship between UHI and multiple factors in Indianapolis. Xiao et al. (2008) used a similar method to explore the causes of urban LST in Beijing, integrating physical and human factors towards policy implications. However, the above synthetic studies paid more attention to physical factors, except for a few social factors, such as population density, pavement and building density. Little attention has been paid to economic factors, such as the manufacturing land ratio and road density. A more comprehensive indicator system is helpful to deepen the understanding of UHI causes. The main goal of this paper is to derive the principle drivers of UHI, and quantify the relationship between UHI and the principle components in the spatial setting. Shanghai, the largest cities in China, is used as the case study. Over the past two decades, Shanghai has experienced rapid urbanization. Accelerated urban expansion, fast industrial and spatial transformations have made the mechanism of UHI more complicated. Hence, understanding the linkage between UHI and driving factors has important implications for urban planning in Shanghai. 2 Data and methodology 2.1 Study area Shanghai, the largest economic center of China, has witnessed a severe urban thermal environment with accelerated urbanization in recent years. This coastal city has a registered population of 23.03 million in 2010 and a total area of 6,340.5 km2 (Fig. 1). Similar to other large Chinese cities, it has experienced rapid urban growth and significant urban land transformation since the economic reform in 1978. With its urbanization rate (the ratio of urban population to total population) increasing from 59 % in 1978 to 86 % in 2007, spatial structure of Shanghai has undergone significant changes. The intensifying human activities of Shanghai have significant impacts on its urban environment. For instance, the UHI effects were first documented Stoch Environ Res Risk Assess (2012) 26:899–911 in the early 1980s and the UHI center was detected in the downtown (Zhou and Zhang 1982). Ding et al. (2002) discovered that in the 1990s, the UHI was located in the areas within the distance of 17–33 km around the downtown. A recent study by Dai et al. (2010) found that LST declines distinctly inside the inner ring road, while it rises obviously outside the central city from 1989 to 2002. While the changed pattern has been identified, determinants for the change have not been studied. We intend to address this gap by analyzing the determinants for the spatial pattern of UHI in Shanghai. 2.2 Data We selected cloud-free Landsat 5 Thematic Mapper (TM) images (April 13, 2008) (Path 118/Row 38, 39). The bands 1–5 and 7 were combined and resampled to 30 m pixel resolution using nearest neighborhood algorithm. The band 6 of Landsat images was resampled to 120 m pixel resolution. A projection of WGS84 NORTH, Zone 51 N is selected for Shanghai city. Land use map in 2009 was processed from Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) Level 1B satellite image data (date: October 20, 2009; resolution: 15 m; projection: UTM WGS84 Zone 51 N). We conducted several supervised classifications for land types. High-resolution images from Google Earth are used as interpretation keys. Further, two intensive field survey trips were carried out in Shanghai for about 2 weeks to facilitate land use classification. We have also used other data sources such as census data, road network maps, master plans (1999–2020), and statistical yearbooks (1995–2010). They were provided by municipal agencies including the Shanghai Municipal Planning Bureau, the Bureau of Land Resource Administration, and the Construction Committee. 2.3 Methods 2.3.1 Retrieval of UHI density LST is an important parameter to modulate the concurrent air temperature of the lower layer of urban atmosphere (Voogt and Oke 1998). Among a series of satellite sensors developed to collect LST data, Landsat TIR sensor image has been most widely used to derive LST, as it provides a basis for continued long-term studies of urban environment without the significant bias (Weng 2009). Based on TIR remote sensing images, a three-step procedure was adopted to derive LST. To derive LST, we first converted the digital number (DN) of Landsat thermal band to spectral radiance (Lk) by using the following equation (Yue and Xu 2007): 901 Lk ¼ gain  DN þ offset ð1Þ The gain and offset values have been obtained from the data header files. Second, we derived the temperature values of a black body using the inverse of the Planck function (Landsat Project Science Office 2002): TB ¼ K2 lnðK1 =Lk þ 1Þ ð2Þ where TB is the effective at-satellite temperature in kelvin (K); Lk is the spectral radiance in W/(m2 ster lm); for Landsat 5 TM images, K1 = 607.76 mW cm-2 ster-1 lm-1, and K2 = 1260.56 K (Landsat Project Science Office 2002). Third, we obtained the LST by applying corrections for emissivity (e) on TB based on the land cover of the surface by Eq. 3. The emissivity corrected LSTs were computed as follows (Artis and Carnahan 1982): LST ¼ TB 1 þ ðk  TB =qÞ ln e ð3Þ where k is the wavelength of the emitted radiance (for which the peak response and the average of the limiting wavelengths (k = 11.5 lm) will be used), q = h 9 c/r (1.438 9 10-2 mK), where h is Planck’s constant (6.626 9 10-34 Js), r is Boltzmann constant (1.38 9 10-23 J/K) (Markham and Barker 1985), and c is the velocity of light (2.998 9 108 m/ s). For emissivity (e), vegetated areas were given a value of 0.95 and non-vegetated areas 0.92 (Weng 2001). Instead of directly using the absolute LST derived from the above, we calculated the UHI intensity of each pixel, which is the temperature difference between LST and the mean temperature of suburb nature land. First, after dividing land use data into urban built-up land and rural land, we did a random sampling for rural land and calculated the average value for all sample points. And we sampled a total of 650 points in rural areas, i.e., one point per km2 as the rural land cover is about 650 km2. Second, to obtain the suburban temperature, we derived the LST value of these sampling points and obtain their mean value of suburban as 22.64 C. We then used the LST to deduct the mean temperature of suburb nature land. Finally, we processed the difference data by assigning zero to pixels with negative values. The final data was used as the main indicators for the intensity of the land surface heat. 2.3.2 Land surface fraction and land use classification information extraction First, we obtained the Impervious Surface Fraction through the Linear Spectral Mixture Analysis (LSMA). In remotely sensed data, many different land covers may be mixed in a pixel. The LSMA assumes that spectrum measured by a 123 902 Stoch Environ Res Risk Assess (2012) 26:899–911 Fig. 1 Location and administrative divisions of Shanghai. Note The research area includes most of the urban area in Shanghai Municipality sensor is a linear combination of the spectra of green vegetation, impervious surface (including high and low albedo surfaces), and exposed soil in varying proportions within the pixel. A typical LSMA model can be expressed as (Small 2001): n X Ri ¼ fk Rik þ ERi ð4Þ k¼1 123 where i is the number of spectral bands. k is the number of endmembers. Ri is the spectral reflectance of band i containing one or more endmembers. fk is the proportion of endmember k within the pixel. Rik is the known spectral reflectance of endmember k within the pixel on band i. ERi is the error for band i. A constrained least-squares solution was used in this research, assuming that the following two conditions are simultaneously satisfied (Small 2001): Stoch Environ Res Risk Assess (2012) 26:899–911 903 Table 1 Index used in PCA method Factors (abbreviation) (unit) Descriptions of the factor and the calculation method Building Density (BUILD) (%) BUILD is the ratio of ground area taken up by buildings to the block area. We map building density at every city block based on land use/cover maps Impervious Surface Fraction (ISF) (%) ISF is the estimated relative amount of impenetrable surface area within a pixel, which is retrieved using LSMA from Landsat images Population Density (POP)(Persons/km2) Using a population growth model*, we projected the population in 2008 based on 2000 census’s population data. We then created population surfaces from discrete living districts with population density values by using an Inverse Distance Weighted (IDW) interpolation Road Density (ROAD) (km/km2) ROAD sums the length of different types of roads in a unit area. It is calculated: [(L1 9 V1) ? (L2 9 V2) ?  (Ln 9 Vn)]/(area), where L1…Ln represent the length of the portion of each line that falls within the circle at a search radius of 200 m**; V1…Vn are weight values of roads, ‘‘3’’ for national highways, ‘‘2’’ for provincial highways, and ‘‘1’’ for minor roads Industrial Land Ratio (INDUS) (%) INDUS is the ratio of industrial land area to the total land area. Based on LULC maps, we obtained the value of the total area of industrial land in a unit of 500 m by 500 m by using Zonal statistics in ArcGIS Normalized Difference Vegetation Index (NDVI) NDVI is a simple numerical indicator of vegetation derived from remote sensing data. It is calculated as: (NIR-VIS)/(NIR ? VIS), where VIS and NIR stand for the spectral reflectance measurements acquired in the visible (red) and near-infrared regions, respectively Normalized Difference Water Index (NDWI) NDWI is an index of vegetation water content derived from remote sensing imageries. It is calculated as: (NIR-MIR)/(NIR ? MIR), where NIR and MIR represent the reflectance of near-infrared and midinfrared regions, respectively SHDI measures the diversity of the land types and generally increases as the number of land cover types increases and/or the proportional distribution of area among land cover types becomes more equitable. We calculated it by using FRAGSTATS software based on land cover maps Shannon’s Diversity Index (SHDI) Contagion index (CI) CI measures the intermixing of different land types and the spatial distribution of a land type. CI is high when the patch types are aggregated and is low vice versa. We also calculated it by using FRAGSTATS software Note: All the data were transformed to the same project system, and resized into 500 m by 500 m by zonal statistics or re-sampled by using neighborhood method * Population data is based on 2000s census at the living districts level (census block of US equivalent) of Shanghai. We used the following equation to estimate the population of 2008: P2008,j = P2000,j 9 (1 ? rj)8 ? Nj 9 8, where P2008,j is the estimated population of the street j in 2008, rj is the annual natural population growth rate of street j, Nj is the average net inflow of population of street j ** Its determination is largely based on the scale of research area. It is assumed that the temperature change may be caused by urban buildings beyond the distance of 200 m to roads n X fk ¼ 1 and 0  fk  1 ð5Þ k¼1 ffiffiffiffiffiffiffiffiffi vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi! u m u X ER2i =m RMS ¼ t ð6Þ i¼1 A lower Root Mean Square (RMS)1 error of the abundance images is desired. When the estimated value of fk is less than 0 or larger than 1, we simply set fk = -0.05 and fk = 1.05, respectively. Four components, i.e., vegetation, bared soil, high albedo and low albedo were selected in the LSMA. We used a sum of high albedo and low albedo fraction to obtain the fraction of impervious surface. Second, in order to obtain a detailed urban land use classification data, we employed a supervised classification 1 RMS error images appear as noise, which determine the overall error of all of the endmember abundance values for each pixel. The areas with high RMS indicate low accuracy of spectral unmixing. with the maximum likelihood algorithm to classify the ASTER images. Six land use categories were identified, including residential land, industrial land, mixed urban land, urban green land, farmland and water. Mixed urban land consists of urban commercial land, educational land, and land for public administration and service. Accuracy of the classification was verified by field checking. We conducted sampling and Kappa analysis to assess the accuracy of our classification by selecting 60 sample points for each land use type. The Kappa coefficient also called KHAT statistic is a measure of agreement or accuracy, computed as: K¼ N P  ri¼1 ðxiþ  xþi Þ i¼1 xiiP N 2  ri¼1 ðxiþ  xþi Þ Pr ð7Þ where r is the number of rows in the error matrix, xii is the number of observations in row i and column i, xi? and x?i are the marginal totals of row i and column i, respectively N is the total number of observations (Stehman 1996). 123 904 Stoch Environ Res Risk Assess (2012) 26:899–911 The overall kappa value of land use classification was 0.8452 and the kappa value for the industrial land reached 0.9034. Based on this land classification data, several landscape configuration indexes were calculated by the software of Fragstats 3.3 developed by McGarigal. 2.3.3 Principal component analysis (PCA) To analyze complex and correlated factors contributing to UHI, we use the PCA to compress the data by eliminating redundancy, make the data more interpretable, and determine the weights of the factors (Smith et al. 2007). PCA, a mathematical procedure that transforms a number of correlated factors into a smaller number of uncorrelated orthogonal factors, can remove data redundancy with a minimal loss of information (Li and Yeh 2001). The first principal component accounts for as much of the variability in the data as possible, and each successive component accounts for as much of the remaining variability as possible in a descending order. The weights of the factors are determined in accordance with their contribution rate using PCA. The PCA transformation is given by (Jolliffe 2002): n X pcij ¼ Xik Ekj ð8Þ k¼1 where pcij is the component score of the jth principle component for cell i, Xik is the value of the kth layer for cell i, and Ekj is the element of the eigenvector matrix at row k and column j. Fig. 2 Spatial pattern of UHI intensity in 2008.  Central Business District; ` New Jiangwan City Park; ´ Baoshan Iron and Steel Plant; ˆ Minhang Industrial Park; ˜ Waigaoqiao Free Trade Zone; Þ Century Park; þ Jiading Industrial Park 123 The eigenvectors and eigenvalues for the linear transformation are mathematically derived from the covariance matrix by the following equation (Jolliffe 2002): ECovET ¼ V ð9Þ where Cov is the covariance matrix, V is the diagonal matrix of eigenvalues, E is the matrix of eigenvectors, T is the transposition function. We adopted the following steps to implement the PCA method: (1) Select driving factors and calculating the values of all factors In previous studies, inadequate attention has been paid to human factors (Xiao et al. 2008). The human impact on the UHI is directly through anthropogenic heat release, and indirectly through the land use configuration. Therefore, the factors from human side should receive similar amount of attention. Hence, we selected factors such as building density, Impervious Surface Fraction, Population Density, Road Density, Industrial Land Ratio, Normalized Difference Vegetation, Normalized Difference Water Index, Shannon’s Diversity Index and Contagion index. The means and calculations were explained in Table 1. We chose the 500 m 9 500 m spatial extent as the unit of analysis. The up-scaling method, while losing data accuracy, can notably reduce the impacts of spatial autocorrelation. After numerous tests by using different size pixels, we identified that the resolution of 500 m as the best Stoch Environ Res Risk Assess (2012) 26:899–911 905 size to balance the data accuracy and spatial autocorrelation as it significantly decreased the spatial autocorrelation of independent variables compared to the original data (Yue and Xu 2008). We then transformed the original data into the standard scores in order to allow comparison of factors from different sources by the following equation: ½Xik  minðXik Þ  100 ½maxðXik Þ  minðXik Þ ð10Þ where Xik is a raw score to be standardized for the kth layer and cell i; max (Xik) and min (Xik) are the maximum and minimum value of the population, respectively. (2) We identified principal components using PCA in the Multivariate Analysis of ArcGIS 9.3 (ESRI, Redlands) and examined the component loadings for the nine spatial variables, based on the data of 2008. The objective of this step was to choose as few principal components as possible while achieving a reasonably high value of cumulative percentage of variance above a certain threshold, such as 75 %. (3) To validate our results, we performed a linear regression model to assess the relationship between the intensity of UHI and the standardized scores of the selected principal components (e.g. the first three components), using Ordinary Least Squares in ArcGIS 9.3. We adopted regression method to calculate the contributing rate (Li and Weng 2007). The simulated formula is as follows: Tsimu ¼ b0 þ 3 X bj  pcij þ ei ð11Þ j¼1 where Tsimu is the simulated value of the LST; pcij (j = 1, 2, 3) are the factor scores of the first three components, respectively; bj (j = 0, 1, 2, 3) is the estimated coefficient; ei is the residual errors. By doing so, we can map the simulated UHI intensity and error distribution using Raster Calculation in ArcGIS. Then we used a buffer analysis from the Central Business District (CBD) to outskirts in the western and eastern direction, in order to assess the similarity of the real and simulated values. 3 Results 3.1 The spatial pattern of the UHI intensity According to the definition of UHI, Fig. 2 revealed an apparent UHI effect in Shanghai. The temperature in urban area was much higher than that in suburbs or exurbs, especially the area of Puxi urban area (the west of Huangpu river). We derived the temperature change pattern through a profile line from northwest corner to southeast corner. The temperature profile detected an UHI between urban area and suburban area (Fig. 3). However, due to the different impacts of driving factors, UHI did not exhibit an Distance to NW Corner (Pixle number) Fig. 3 The pattern of UHI intensity along the NW–SE profile 123 906 Stoch Environ Res Risk Assess (2012) 26:899–911 equivalent pattern from urban center to outer area. We also identified several spots with extra higher temperature in urban area, such as Yangpu District, Zhabei District and part of Putuo District where a number of old residential and industrial buildings mixed and crowded. In addition to the above-mentioned spots, some other spots appeared to have distinct higher temperature, such as Wusong Port where Baoshan Iron and Steel Plant is located, Minhang Industrial Park in the southwest of outer ring road, and Jiading Industrial Park in the northwest of outer ring road, as well as Waigaoqiao Free Trade Zone in Pudong District. Unsurprisingly, urban surfaces characterized with natural land cover, such as the New Jiangwan City Park, Century Park and most areas in Pudong District had a lower temperature. 3.2 PCA: factors contributing to the UHI Based on the data of 2008, we identified principal components through using PCA in ArcGIS (Table 2) and examined the component loadings for the nine spatial variables (Table 3). The first three principal components collectively account for 76.58 % of the variance. The spatial distribution of the first three components is mapped in Fig. 4. We classified three major components responsible for the variances of UHI. The first component, ‘‘the gradient from man-made to natural land cover’’ (abbreviated as ‘urban–rural gradient’), comprises information of land surface modification, as it pertains mainly to variables such as building density, while having a negative correlation with the vegetation cover (Fig. 4a). The second component, ‘‘landscape configuration’’, has heavy loading with variables indicating the landscape composition as it is positively correlated with SHDI and negatively correlated with the Contagion Index (Fig. 4b). The third component, ‘‘anthropogenic heat release’’, has heavy loading of Table 2 Principal components and their variances Principal components Eigenvalues Percentage of variance (%) Cumulative percentage of variance (%) 1 970.15 48.16 48.16 2 310.31 15.41 63.57 3 262.16 13.01 76.58 4 178.30 8.85 85.43 5 119.56 5.94 91.37 6 69.47 3.45 94.82 7 45.17 2.24 97.06 8 43.33 2.15 99.21 9 15.88 0.79 100.00 123 Table 3 Component loading of the spatial variables Variables Factor 1 Factor 2 Factor 3 BUILD 0.73 -0.12 -0.17 ISF 0.27 0.10 -0.05 POP 0.28 -0.07 -0.38 ROAD 0.24 0.06 -0.33 INDUS 0.30 -0.32 0.79 NDVI -0.29 0.32 0.01 NDWI -0.05 -0.23 0.01 SHDI 0.27 0.65 0.27 -0.13 -0.54 -0.10 CI industrial heat generation as it is positively correlated with spatial variable of industrial land ratio and is negatively correlated with population density (Fig. 4c). 3.3 Simulation of urban thermal environment based on PCA results Based on Eq. 11, we obtained the following linear model to simulate the UHI effect: Tsimu ¼ 19:7481 þ 0:0400  pci1 þ 0:0033  pci2 þ 0:0045  pci3 ð172:94Þ ð86:23Þ ð4:06Þ ð5:03Þ ð12Þ where Tsimu is the simulated value of LST; pci1, pci2, pci3 are standardized values of the first three components; and the values in the second line are T-statistics. The correlation coefficient was 0.729, significant at the 0.01 level. According to Eq. 8, the principal components was the linear function of the original variables, so we transformed Eq. 12 into the final model based on Eq. 8 (loadings are listed in Table 3) as follows: Tsimu ¼ 0:028  BUILD þ 0:011  IMPER þ 0:009  POP þ 0:008  ROAD þ 0:015  INDUS  0:011  NDVI  0:003  NDWI þ 0:014  SHDI  0:007  CI ð13Þ The model indicated that the intensity of UHI was positively correlated with building density, industrial land, impervious surface, SHDI, population and road density, but was negatively correlated with NDVI, CI and NDWI. Figure 5 revealed the distribution of the simulated UHI, similar to the real world. For example, both of them had high temperature in industrial sites/urban center, and low temperature in Pudong New Area/suburban areas with large portions of greenbelts and farmland. Especially, locations with abundant vegetation regardless in urban center or suburban area, had low temperature, such as New Stoch Environ Res Risk Assess (2012) 26:899–911 907 Fig. 4 Resulting raster of the first three principal components. Note The value of pixels decrease with the transition from red areas to green areas Jiangwan City Park, Century Park and Gongqing Forest Park. 4 Discussion A number of factors have been found to be associated with UHI intensity, such as building density (Yue and Xu 2008), industrial areas (Kato and Yamaguchi 2005), impervious surface fraction (Imhoff et al. 2010), population concentration (Huang et al. 2005), and road density (Hart and Sailor 2009). This study confirmed the relationship between these factors and UHI as illustrated by Eq. 11. In addition, we found that NDVI and NDWI were associated with mitigating UHI responses, as expected in related reports (e.g. Arnfield 2003). Furthermore, this research demonstrated that PCA can provide concise and coherent descriptions of major components underlying patterns of UHI intensity. Most of the factors can be transformed into one of the three major components, that is, urban–rural gradient, landscape configuration, or anthropogenic heat release. 4.1 Determinants of UHI The first component, the urban–rural gradient component, explained 48.16 % of the variation. The large values appeared in the zones extending from the urban core to the west of Huangpu River, whereas the values decreased 123 908 Stoch Environ Res Risk Assess (2012) 26:899–911 Fig. 5 Simulated urban thermal environment of Shanghai in 2008 along various urban–rural transects. This is due to building materials in urbanized areas with low albedo and high temperatures in urban canyons. In contrast, the vegetated areas had lower temperatures. The second component, the landscape configuration component, exhibited a very high value in the emerging urban area near the inner ring road of Pudong New Area. Landscape in this area was highly fragmented, mixed with the residential and agricultural uses. In general, UHI was intensified by the complexity of the urban environment and intense human activities in the emerging urban area with high SHDI and low Contagion Index (Weng et al. 2007). Our results confirmed that the area, shape, and especially the composition of underlying 123 surfaces affected the thermal environment significantly. The third component, the anthropogenic heat release, contained information on industrial sites, which created large amounts of waste heat. In addition, the industrial sites were often featured with low-density population, resulting in the negative correlation of the third component with the population. Consequently, the effort of the urban renewal by local governments has resulted in the conversion of the old industrial land to the vegetation land or office in the city center, leading to a decreased intensity of UHI. However, more intense industrial producers flooding into the outskirts of the city lead to the high values of anthropogenic heat release, thus contributing to the formation of UHI. Stoch Environ Res Risk Assess (2012) 26:899–911 909 40000 35000 Scores of PC3 30000 25000 20000 15000 10000 5000 0 0 20000 40000 60000 80000 100000 120000 Gross output of manufacturing sector (100,000 yuan) Fig. 6 The scatter plot between gross output of industrial sector and scores of PC3 4.2 Validation for the anthropogenic heat release As a non-parametric model it is difficult to do the validation for PCA. Among the three principal components of this study, the anthropogenic heat release is the most complicated factor. As a synthetic factor the anthropogenic heat release relates to all human activities, such as population density, traffic, industrial production. Here we selected the gross output of industrial sector as a reference indictor to validate PC3, namely anthropogenic heat release. We collected the gross output of industrial sector data of all districts, and then calculated the area proportion of each district relative to total area. We also computed the total scores of PC for each district. The relationship is expressed in the scatter plot (Fig. 6). Finally we did a Pearson correlation analysis. The correlation index is 0.94 (significant at 0.01 level). The correlation analysis revealed that the spatial pattern of anthropogenic heat release has a close relationship with the distribution of industrial activities. 4.3 Disparities between simulation and real world Nevertheless, some disparities existed between the simulated LST and the real world. While the highest temperature in the reality is at the industrial sites such as Wusong Port and Minhang Industrial Parks, the highest temperature in the simulated LST appeared to be at the intersection of Suzhou River and Huangpu River inside the inner ringroad. Air pollution in the industrial parks might not be fully captured by our model, exhibiting large deviation as shown in Fig. 5c. To assess the differences of sub-regions, we used zonal statistics over the eastern/western gradient to explore the issue of spatial heterogeneity (Fig. 7). Through the buffer analysis, the simulated values were smoother than the observed value of UHI, with higher values in water zones and vegetated areas. Although simulated LST provided a general trend of UHI, it cannot fully reflect the real world as the impact of other biophysical and human factors are difficult to be assessed using the current method. Fig. 7 The zonal statistics of real and simulated UHI intensity from the CBD to outskirts, in the western and eastern areas of Hangpu River (Puxi and Pudong) 4.4 Research highlights and limitations We would like to highlight three particular contributions of our studies. First, our study underlined the contribution of landscape configuration to UHI, which was seldom reported in previous work (Yue and Xu 2007). Landscape configuration had a coefficient of 0.0033 in Eq. 12, and was an important factor, especially in emerging urban areas where land became more fragmented due to intensive human activities. For instance, despite small impervious surface fraction, the urban–rural transects usually had high UHI in Shanghai. Second, although the industrial locations were positively correlated with UHI in general as expected by other studies (Chrysoulakis and Cartalis 2002; Hart and Sailor 2009), our study indicated that different types of industrial firms would have diverse impacts on UHI. For instance, hot spots with maximum temperature differences appeared in old industrial sites, such as Baoshan Iron and Steel Plant, but not in others such as high-tech industrial parks in the Pudong New Area (Fig. 2). This implied that heavy industries with severe pollution contributed to the surface temperature rise whereas high-tech industries such as information technology, micro-electronics, and pharmacy, had lower surface temperature due to the little pollution. This can help us to explain the relatively high deviation in industrial sites between real and simulated LST (Fig. 7). Third, in contrast to previous studies (e.g. Oke 1973; Tran et al. 2006; Xiao et al. 2008), population density and road density did not have overwhelming influences on UHI compared with other factors in Eq. 13. For example, Tran et al. (2006) observed maximum intensity of UHI with a high population density in metropolitan areas such as Tokyo and Manila. However, we found that the intensity of UHI declined in the old city center with high population density (Fig. 2), which could be explained by massive 123 910 residential relocation from the inner city (e.g. the urban renewal prepared for World Expo 2010 in Shanghai), and huge inflow of migrant population in urban periphery (Sheng and Wang 2000). Roads also did not exhibit highest temperature as expected with other studies (Hart and Sailor 2009). This may be due to the properties of roadside trees and adjacent green spaces, as well as high thermal admittance (Xiao et al. 2008). However, several limitations should be noted. First, the strength that PCA is a non-parametric analysis can also be viewed as a weakness (Jolliffe 2002). There are no parameters to tweak and no coefficients to adjust based on our experiences, even if we may have a prior knowledge with selected parameters. For example, we theoretically classified the heat released by population and road systems to anthropogenic heat, but we could not contribute the data of these variables to the third polar component in our calculations (Table 3). We plan to incorporate prior knowledge in the selection of the optimal structure of the PCA model in the follow-up study. Second, we did not consider the effects of other factors (e.g. the sea wind speed in the coastal city, the albedo of building materials, and the Floor Area Ratio) on the generation of UHI. The task is challenging because of high spatial heterogeneity even within a block area (Oke 1982; Hart and Sailor 2009). Moreover, the anthropogenic heat on UHI is very complicated to be quantified straightforward, which changes over diurnal and seasonal cycles (Hart and Sailor 2009). We plan to explore some of these issues in our future research. 5 Conclusions Based on LST data obtained from thermal remote sensing imageries, this paper assessed the spatial pattern of urban thermal environment in Shanghai in 2008 and analyzed factors contributing to the generation of UHI by PCA. The following major findings are highlighted here: First, Shanghai had a distinct UHI effect in 2008, regarding the temperature dynamic along the profile from northwest suburb to downtown, then to southeast suburb. UHI is not evenly distributed. This spatial pattern of urban thermal environment is determined by a complex system. Second, our PCA results identified three most important components responsible for the variances of Shanghai’s UHI: urban–rural gradient (building density and vegetation index), landscape configuration (SHDI and Contagion Index), and industrial sites. They together accounted for 76.58 % of the variance. A simple linear model was constructed to simulate the UHI effect by using these three components. The similarity between simulated and real 123 Stoch Environ Res Risk Assess (2012) 26:899–911 UHI intensity indicated a strong explanatory power of the PCA model. In summary, vegetation plantation and the control of anthropogenic heat release should be effective in reducing the intensity of UHI in Shanghai. Hence, city government can mitigate UHI by expanding the city’s green coverage and relocating of industrial sites from the city center. Acknowledgments We acknowledge the project supported by Zhejiang Provincial Natural Science Foundation of China (Y5110009), the National Natural Science Foundation of China (NNSFC) (41101568), the National Aeronautics and Space Administration (NASA)’s Land Cover and Land Use Program through the grant to Michigan State University (NNX09AI32G), Zhi-Jiang Young Scholars Program through Zhejiang Philosophy and Social Science Planning Project (11ZJQN042YB), Qian-Jiang Talent Program (C), and the Fundamental Research Funds for the Central Universities (2012XZZX012). Any opinions, findings, and conclusions or recommendations expressed in this paper are those of the authors and do not necessarily reflect the views of NNSFC or NASA. References Arnfield AJ (2003) Two decades of urban climate research: a review of turbulence, exchanges of energy and water, and the urban heat island. Int J Climatol 23(1):1–26. doi:10.1002/joc.859 Artis DA, Carnahan WH (1982) Survey of emissivity variability in thermography of urban areas. Remote Sens Environ 12(4): 313–329. doi:10.1016/0034-4257(82)90043-8 Carlson TN, Arthur ST (2000) The impact of land use—land cover changes due to urbanization on surface microclimate and hydrology: a satellite perspective. Glob Planet Change 25(1–2): 49–65. doi:10.1016/S0921-8181(00)00021-7 Chen S, Pian L (1997) Effects of urbanization on the annual mean temperature of Beijing. Acta Geogr Sin 52(1):27–36 (in Chinese) Chen XL, Zhao HM, Li PX, Yin ZY (2006) Remote sensing imagebased analysis of the relationship between urban heat island and land use/cover changes. Remote Sens Environ 104(2):133–146. doi:10.1016/j.rse.2005.11.016 Chrysoulakis N, Cartalis C (2002) Thermal detection of plumes produced by industrial accidents in urban areas based on the presence of the heat island. Int J Remote Sens 23(14):2909– 2916. doi:10.1080/01431160110111973 Chudnovsky AE, Ben-Dor Saaroni H (2004) Diurnal thermal behavior of selected urban objects using remote sensing measurements. Energy Build 36:1063–1074. doi:10.1016/j.enbuild.2004.01.052 Dai X, Guo Z, Zhang L, Li D (2010) Spatio-temporal exploratory analysis of urban surface temperature field in Shanghai, China. Stoch Environ Res Risk Assess 24:247–257. doi:10.1007/ s00477-009-0314-2 Ding JC, Zhang ZK, Xi H, Zhou HM (2002) A study of the high temperature distribution and the heat island effect in the summer of the Shanghai area. Chin J Atmos Sci 26(3):412–420 (in Chinese) Han J, Hayashi Y, Cao X, Imura H (2009) Evaluating land-use change in rapidly urbanizing China: case study of Shanghai. J Urban Plan Dev 135(4):166–171. doi:10.1061/(ASCE)0733-9488 Hart M, Sailor D (2009) Quantifying the influence of land-use and surface characteristics on spatial variability in the urban heat island. Theor Appl Climatol 95(3):397–406. doi:10.1007/ s00704-008-0017-5 Stoch Environ Res Risk Assess (2012) 26:899–911 Huang H, Ooka R, Kato S (2005) Urban thermal environment measurements and numerical simulation for an actual complex urban area covering a large district heating and cooling system in summer. Atmos Environ 39(34):6362–6375. doi:10.1016/j.atmo senv.2005.07.018 Imhoff ML, Zhang P, Wolfe RE, Bounoua L (2010) Remote sensing of the urban heat island effect across biomes in the continental USA. Remote Sens Environ 114(3):504–513. doi:10.1016/j.rse. 2009.10.008 Jolliffe IT (2002) Principal component analysis. In: Bickel P, Diggle PJ, Fienberg SE, Gather U, Olkin I, Zeger S (eds) Springer series in statistics. Springer, NY, pp 10–28 Kato S, Yamaguchi Y (2005) Analysis of urban heat-island effect using ASTER and ETM? data: separation of anthropogenic heat discharge and natural heat radiation from sensible heat flux. Remote Sens Environ 99(1–2):44–54. doi:10.1016/j.rse.2005. 04.026 Landsat Project Science Office (2002) Landsat 7 Science Data User’s Handbook. Coddard Space Flight Center, NASA, Washington Li G, Weng Q (2007) Measuring the quality of life in city of Indianapolis by integration of remote sensing and census data. Int J Remote Sens 28(2):249–267. doi:10.1080/0143116060073 5624 Li X, Yeh AGO (2001) Application for spatial decision and urban simulation of principle component analysis and cellular automata. Sci China Ser D 31(8):683–690 Li J, Wang X, Wang X, Ma W, Zhang H (2009) Remote sensing evaluation of urban heat island and its spatial pattern of the Shanghai metropolitan area, China. Ecol Complex 6(4):413–420. doi:10.1016/j.ecocom.2009.02.002 Li J, Song C, Cao L, Zhu F, Meng X, Wu J (2011) Impacts of landscape structure on surface urban heat islands: a case study of Shanghai, China. Remote Sens Environ 115(12):3249–3263. doi:10.1016/j.rse.2011.07.008 Lo CP, Quattrochi DA (2003) Land-use and land-cover change, urban heat island phenomenon, and health implications: a remote sensing approach. Photogramm Eng Remote Sens 69(9):1053– 1063 Markham BL, Barker JL (1985) Spectral characterization of the LANDSAT Thematic Mapper sensors. Int J Remote Sens 6(5):697–716. doi:10.1080/01431168508948492 Nichol JE (1996) High-resolution surface temperature patterns related to urban morphology in a tropical city: A satellite-based study. J Appl Meteorol 35 (1):135–146. doi:10.1175/1520-0450(1996) 035\0135:HRSTPR[2.0.CO;2 Oke TR (1973) City size and the urban heat island. Atmos Environ 7 (8):769–779. doi:10.1016/0004-6981(73)90140-6 Oke TR (1982) The energetic basis of the urban heat island. Q J R Meterol Soc 108(455):1–24. doi:10.1002/qj.49710845502 Rizwan AM, Dennis LYC, Liu C (2008) A review on the generation, determination and mitigation of Urban Heat Island. J Environ Sci 20(1):120–128. doi:10.1016/S1001-0742(08)60019-4 Sheng JF, Wang GX (2000) On the model of population distribution, changes and tendency of central city of Shanghai in the 1990s. Popul Sci China 5:45–52 (in Chinese) Small C (2001) Estimation of urban vegetation abundance by spectral mixture analysis. Int J Remote Sens 22:1305–1334. doi:10.1080/ 01431160151144369 Smith MJd, Goodchild MF, Longley PA (2007) Geospatial analysis: a comprehensive guide to principles, techniques and software tools. The Winchelsea Press, Leicester, UK Stehman SV (1996) Estimating the kappa coefficient and its variance under stratified random sampling. Photogramm Eng Rem Sens 62(4):401–402 Tran H, Uchihama D, Ochi S, Yasuoka Y (2006) Assessment with satellite data of the urban heat island effects in Asian mega 911 cities. Int J Appl Earth Obs Geoinf 8(1):34–48. doi:10.1016/ j.jag.2005.05.003 Voogt JA, Oke TR (1998) Effects of urban surface geometry on remotely-sensed surface temperature. Int J Remote Sens 19(5):895–920. doi:10.1080/014311698215784 Weng Q (2001) A remote sensing-GIS evaluation of urban expansion and its impact on surface temperature in the Zhujiang Delta, China. Int J Remote Sens 22(10):1999–2014. doi:10.1080/ 01431160152043676 Weng Q (2009) Thermal infrared remote sensing for urban climate and environmental studies: methods, applications, and trends. ISPRS J Photogramm Remote Sens 64(4):335–344. doi:10.1016/ j.isprsjprs.2009.03.007 Weng Q (2012) Remote sensing of impervious surfaces in the urban areas: requirements, methods, and trends. Remote Sens Environ 117:34–49. doi:10.1016/j.rse.2011.02.030 Weng Q, Lu D, Schubring J (2004) Estimation of land surface temperature-vegetation abundance relationship for urban heat island studies. Remote Sens Environ 89(4):467–483. doi:10.1016/j.rse.2003.11.005 Weng Q, Lu D, Liang B (2006) Urban surface biophysical descriptors and land surface temperature variations. Photogramm Remote Sens 72(11):1275–1286 Weng Q, Liu H, Lu D (2007) Assessing the effects of land use and land cover patterns on thermal conditions using landscape metrics in city of Indianapolis, United States. Urban Ecosyst 10(2):203–219. doi:10.1007/s11252-007-0020-0 Weng Q, Liu H, Liang B, Lu D (2008) The spatial variations of urban land surface temperatures: pertinent factors, zoning effect, and seasonal variability. IEEE Geosci Remote Sens Soc 1(2): 154–166. doi:10.1109/JSTARS.2008.917869 Wilson JS, Clay M, Martin E, Stuckey D, Vedder-Risch K (2003) Evaluating environmental influences of zoning in urban ecosystems with remote sensing. Remote Sens Environ 86(3):303–321. doi:10.1016/S0034-4257(03)00084-1 Wu JG, Hobbs R (2002) Key issues and research priorities in landscape ecology: an idiosyncratic synthesis. Landsc Ecol 17(4):355–365. doi:10.1023/A:1020561630963 Wu J, Xu J, Tan W (2007) Study on the relationship of urban heat island and vegetation abundance in Shanghai City. Remote Sens Technol Appl 22(1):26–30 (in Chinese) Xiao RB, Weng QH, Ouyang ZY, Li WF, Erich WS, Zhang ZM (2008) Land surface temperature variation and major factors in Beijing. China. Photogramm Eng Remote Sens 74(4):451–461 Yuan F, Bauer ME (2007) Comparison of impervious surface area and normalized difference vegetation index as indicators of surface urban heat island effects in Landsat imagery. Remote Sens Environ 106(3):375–386. doi:10.1016/j.rse.2006.09.003 Yue WZ, Xu LH (2007) Thermal environment effect of urban land use type and pattern: a case study of central area of Shanghai City. Sci Geogr Sin 27(2):243–248 (in Chinese) Yue WZ, Xu JH (2008) Impact of human activities on urban thermal environment in Shanghai. Acta Geogr Sin 63(3):247–256 (in Chinese) Zhan W, Zhang Y, Ma W, Yu Q, Chen L (2012) Estimating influences of urbanizations on meteorology and air quality of a Central Business District in Shanghai, China. Stoch Environ Res Risk Assess 26. doi:10.1007/s00477-012-0603-z Zhou SZ, Zhang C (1982) On the Shanghai urban heat island effect. Acta Geogr Sin 37(4):372–381 (in Chinese) Zhou Y, Weng Q, Gurney RK, Shuai Y, Hu X (2011) Estimation of the relationship between remotely sensed anthropogenic heat discharge and building energy use. ISPRS J Photogramm Remote Sens 67:65–72. doi:10.1016/j.isprsjprs.2011.10.007 123 View publication stats