1、X. Li et al. / Remote Sensing of Environment 166 (2015) 78-9079ELSEVIERCrossMarkContents lists available at ScienceDirectRemote Sensing of Environmentjournal homepage: www.elseviercom/locate/rseA 30-year (1984-2013) record of annual urban dynamics of Beijing City derived from Landsat dataXuecao Lia,

2、 Peng Gong Lu Lianga,ba Ministry of Education Key Laboratory for Earth System Modeling, Center for Earth System Science, Tsinghua University, Tsinghua, Beijing 100084, Chinab Department of Environmental Science. Policy and Management, University of California, Berkeley, CA 94720, USAc Joint Center f

3、or Global Change Studies, Beijing 100875, ChinaARTICLE INFOABSTRACTArticle history:Received 26 June 2014Received in revised form 29 May 2015Accepted 5 June 2015Available online 12 June 2015Keywords:Urban landTemporal consistency checkTime seriesChange detectionAlthough mapping activities of urban la

4、nd change have been widely carried out, detailed information on urban development in time over rapid urbanization areas would have been lost in most studies with multi-year intervals. Here we provide a two-stage framework of long-term mapping of urban areas at an annual frequency in Beijing, China,

5、over the period from 1984 to 2013. Classification for each year was carried out initially based on a number of Landsat scenes within that year using spectral information from a base image plus NDVI time series derived from all scenes. A temporal consistency check involving both temporal filtering an

6、d heuristic reasoning was then applied to the sequence of classified urban maps for further improvement. We assessed this time-series of urban maps based on two schemes. One is change detection in rapidly developing areas over the past three decades, and the other is accuracy assessment over the who

7、le region in four selected years (i.e., 1984, 1990, 2000 and 2013). Based on validation using independent samples, the OAs (overall accuracies) of these four years are 96%, 93%, 92% and 95%, respectively. Meanwhile, the average accuracy of change detection for all years is 83%. In add讓ion, the propo

8、sed temporal consistency check was found to be able to make considerable improvements (about 6%) to the overall accuracies and results of change detection. The resultant urban land sequence revealed that the average growth rates were 47.51 士 4.17 km2/yeart 34.65 2.90 km2/year and 99.48 土 1.3 km2/yea

9、r for 1984-1990,1990-2000 and 2000-2013, respectively. 2015 Elsevier Inc. All rights reserved./! 0.1016/j.rse.2O15.06.007 0034-4257/ 2015 Elsevier Inc. All rights reserved.1. IntroductionInformation on urban areas (or impervious surfaces) and their changes in time has significant imp

10、lications to studies of urban growth modeling (Li & Yeh, 2000; Li, Liu & Gong, 2015), public health (Gong et al 2012; Liang et al 2010), urban climate (Georgescu, Morefield, Bierwagen & Weaver 2014), and land cover change (DeFries, Rudel, Uriarte & Hdnsen, 2010; Seto, Kaufmann & Woodcock, 2000; Yang

11、, Huang, Zhang & Wang, 2014). Although only accounting for a tiny pro- portion (perhaps less than 1%) of the worlds land surface, urban areas serve as a base to more than 90% of the global economy and 50% of the worlds population (Schneider, Friedl & Potere, 2010; Solecki, Seto & Marcotullio, 2013)

12、With the rapid growth of the urban population and intensive human activities, urban extent will undoubtedly continue to grow, and the consequences are even more severe in fast developing regions (e.g.t China), in terms of development intensity and growth rate (Schneider & Meites, 2014; Wang et al.,

13、2012).Temporal mapping of urban extent w讓h acceptable accuracy is crucial to urban planners for better management and development ofCorresponding author.E-mail address: (P. Gong). sustainable cities. Remote sensing is a useful tool to monitor land use/ cover change (LUCC) at

14、large spatial and temporal scales (Turner, Lambin & Reenberg, 2007; Weng, 2012) Various approaches have been proposed for urban or impervious surface extraction, such as the V-I-S (Vegetation-Impervious surface-Soil) model (Ridd, 1995), spectral mixture analysis (Li, Lu, Moran & Hetrick, 2013; Lu &

15、Weng, 2004), structural or contextual classification (Gong & Howarth, 1990,1992), ensemble approaches (Ghimire, Rogan, Galiano, Panday & Neeti, 2012; Li, Liu & Yu, 2014a), object-based approach (Vieira et al., 2012; Wang et al., 2010), and hierarchical mapping method (Im, Lu, Rhee & Quackenbush, 201

16、2).Change detection techniques (e.g., change vector analysis, CVA; principle component analysis, PC A; or post-classification processing) have been widely employed for comparison of paired images (Chen, Gong, He, Pu & Shi, 2003; Lu, Mausel, Brondizio & Moran, 2004; Lunetta, Johnson, Lyon & Crotwell,

17、 2004; Singh, 1989). More recently, following the free release of the Landsat Archive (Roy et al., 2014; Woodcock et al., 2008), it is possible to produce long-term LUCC maps (Hansen et al., 2013; Sexton et al., 2013a; Sexton, Urban, Donohue & Song, 2013b). Many attempts have been carried out for mu

18、ltitemporal classification, and the supervised classification has been the primary approach. To track urban growth,讓 is important to have accurate urban area maps in multiple yearsCurrently, there are two major approaches for multi-temporal urban area extraction. The first is classification based on

19、 “stacked images”, that is, images acquired at different dates are stacked together as combined features for classification (Gao et al., 2012; Schneider, 2012; Schneider & Mertes, 2014; Seto et al, 2002). However, rather than categorical labels, additional information such as type of LUCC and time o

20、f change is required (Seto et al., 2002). The identification of changed samples is usually subjective and labor-intensive. Some approaches, such as a spectral similarity map (Gao et al., 2012), and signature extension (Gray & Song, 2013), have been con ducted to alleviate the work of sample selectio

21、 n via increasing sample size. Another stream of multi-temporal classifica- tion methods is a spatial-temporal consistency check following an initial classification of single-date images (Liu & Cai, 2012; Liu, Kelly & Gong, 2006; Wang et al., 2015) For maps with specific types, a Markov Random Field

22、 (MRF) can be used to model classification probability in space and time (Cai, Liu, Sulla-Menashe & Friedl, 2014; Liu & Cai, 2012). Nevertheless, the initial classification is crucial to subsequent improvement. In add 讓 ion, studies on multi-temporal mapping in elude training one single model from c

23、omposited intra-annual (phe no logical) images and expanding it in time (Sexton, Song et al., 2013a; Sexton, Urban, Donohue & Song, 2013b). The premise of this approach is that corrections on radiometric, atmospheric and topographic effects should be done carefully to ensure that the adopted model i

24、s suitable for images acquired in different years (Chander, Markham & Helder, 2009).Many studies on multi-temporal urban area extraction have already been carried out. Due to their long time availability, Landsat data (from Thematic Mapper, TM; Enhanced Thematic Mapper plus, ETM + ; and Operational

25、Land Imager) are often chosen for this purpose. Gao et al. (2012) used Landsat data to map expansion of impervious surfaces in Yangtze River Delta, China, for 9 periods from 1973 to 2006. Schneider and Mertes (2014) used Landsat TM and ETM+ data to study the expansion of 143 Chinese cities covering

26、7 periods from 1978 to 2010. Wang et al. (2012) completed the mapping of all cities in China at three points in time (i.e., 1990, 2000, and 2010) through human interpretation based on Landsat TM/ETM + data. Xian et al. (2011, 2012) characterized the fraction of impervious surfaces in the United Stat

27、es with a well-known product of the National Land Cover Database (NLCD) every five years from 2001 to 2011. Sexton, Song et al. (2013a) employed the Landsat records to explore the change of im- peivious surfaces in the Washington, D.C.-Baltimore, MD metropolitan region from 1984 to 2010. Castrence,

28、Nong, Tran, Young, and Fox (2014) mapped urban land use change at the Red River Delta, in Vietnam, based on multi-temporal Landsat images (circa 1990, 2000 and 2005).However, although many studies cover a long period in time (e.g, in decades), high (e.g., annual) temporal frequency is needed as well

29、, to represent high-order complexity (e.g., acceleration) (Sexton, Song etal., 2013a). Spontaneous growth of urban land use has become a commonly seen phenomenon in China (Chen, Li, Liu & Ai, 2014), which requires more frequent observations. From the perspective of urban growth modeling, such as Cel

30、lular Automata (CA) (Wh讓e & Engelen, 1993), the intermediate outputs of model performance should be validated by dense observations, which is ignored in most existing studies (Batty & Torrens, 2005; Li, Liu & Yu, 2014b). Hence, it is important to monitor changes of urban land at a higher frequency (

31、e.g., annual) (Hansen & Loveland, 2012). Using 10 classified images, Seto and Fragkias (2005) created a continuous sequence from 1988 to 1999 in Pearl River Delta. In their work, the Tasseled Cap transformation technique was implemented with each single Landsat image derived for different years Sext

32、on, Song et al. (2013a) completed a series of annual maps of impervious surfaces anging from 1984 to 2010 recently with outputs of continuous-field proportion based on Landsat TM records. However, the inconsistencies cannot be avoided where areas of higher impeivious proportions may decrease in subs

33、equent years at some locations.An urban area (or an impervious surface) is a relatively stable land type over a long period of time, i.e., conversion from urban area to non-urban is usually not realistic (Mertes, Schneider, Sulla-Menashe, Tatem & Tan, 2015). Therefore, classification with stable pix

34、els (i.e., training sample urdts of urban that can be assumed not changed in time thereafter) will benefit from the consistency of this class at different times (Gray & Song, 2013). Therefore, for categorical classification, a heuristic approach employing temporal con texts (Latifovic & Rouliot,2005

35、) for a consistency check in the sequence of urban areas is crucial to obtain more realistic time-series urban maps.In this research, we attempt to map an urban area on an annual basis from 1984 to 2013 using Landsat data over Beijing City. For each year, almost all the available data of a Landsat T

36、M image including extracted NDVI time series were employed in urban land classification. The initial classification performance of urban land is estimated with an accuracy calculated by withholding a subset of the training samples (Breiman, 2001). After a temporal consistency check against the class

37、ified urban- area sequence, we assessed the performance of change detection in rapid urbanization areas over the past three decades and obtained overall accuracy for four selected years (i.e., 1984,1990,2000 and 2013). Finally, we characterized the expansi on pattern of the urba n area in Beijing fr

38、om 1984 to 2013.2. Methods2.1. Study areaBeijing City covers an area of 16,808 km2. The administrative boundary of Beijing is located within two scenes (path/row: 123/32 and 123/ 33; World Referenee System-2, WRS-2), with most of the municipality falling in path/row 123/32 (Fig. 1). Since the purpos

39、e of this study is to test the efficiency and feasibility for mapping urban areas, we only used the main tile (path/row 123/032), which covers more than 95% of the total Beijing City. The expansion of the urban area in Beijing has drawn a wide range of attention (Chen et al., 2003; Du, Thill, Peiser

40、 & Feng, 2014; He, Okada, Zhang, Shi & Li, 2008; Jia et al., 2014; Wu et al.,2006) . Since the economic reform in 1978, considerable urban expansion has been witnessed (Wu et al., 2006).2.2. Landsat time seriesWe chose the Landsat Surface Reflectance Climate Data Record (Landsat CDR) of L4-5 TM as o

41、ur primary data source (obtained from U.S. Geological Survey: /). Due to missing data or termination of satellites, other data sources (e.g., Landsat CDR ETM+ or Landsat 8 OLI) were also used to replenish the temporal sequence. All images had been processed to LIT level (

42、e.g, passed the corrections of topography and radiation). The Landsat CDR datasets (i.e., L4-5 TM and L7 ETM + ) were processed uniformly with the Landsat Ecosystem Disturbance Adaptive Processing System (Masek et al., 2006), whereas for Landsat 8 OLI datasets, their atmospheric corrections were con

43、ducted by ourselves with the Fast Line-of-Sight Atmospheric Analysis of Spectral Hypercubes algorithm (Adler-Golden et al, 1999; Gong et al., 2013). The temporal distribution of adopted Landsat images for this tile (path/row-123/032) is shown in Fig. 2. A total of 123 scenes of Landsat images were i

44、n eluded. Among these tiles, most of them (113) were acquired w讓h Landsat 4 and 5. For data missing years of 2006, 2008 and 2012, the Landsat 7 ETM + data was employed as a subst讓ute, although their drop-of-scan lines could lower the image quality. For 2013, Landsat 8 OLI datasets were utilized due

45、to the termination of Landsat 5 in 2011. Overall, available time series were relatively adequate for the first two decades (1984-2004), whereas for the last decade the usable images are insufficient in some years, except for 2013. Besides, most images were acquired in summer (44) or winter (47), and

46、 the remaining (32) were either in spring or in autumn. Although these images are unevenly distributed in time, it is generally feasible for long-term mapping of urban lands with an annual frequency.81X. Li et aL / Remote Sensing of Environment 166 (2015) 78-90?o.?ob116o0f0nE1170(TE1180f0tfE1170f0nE

47、ii8(r(rEFig. 1. Study area Beijing, China The box represents the normal WRS-2 boundaiy (path/row: 123/032).XLOO2.P116o0f0ME2.3. Classification frameworkIn this study, we refer the defin讓ion of urban land as in Schneider and Mertes (2014): s讓es that are dominated by built environment, including all n

48、on-vegetative, humanconstructed elements (e.g., roads and buildings). The “dominated” indicates a greater than 50% coverage of built environment within each pixel. Based on this definition, we designed a classification scheme with collected samples in different years. The NDVI time series within eac

49、h year are also incorporated in the classification. Once the initial classifications were obtained, a spatial-temporal filtering was sequentially applied for consistency checking. Finally, the whole process was completed after slight but necessary man ual editing (Fig. 3). A detailed description of

50、each individual component in the entire process is presented in the following sections.2.3.1. Classification within a yearFor each year, we did an annual classification based on the spectral bands of the base scene, as well as NDVI time series acquired from all available scenes in that year as input

51、s for image classification. To distinguish bareland from urban, we defined bareland here as a natural environment w讓h less or no vegetation, but dominated by exposed soil, sand, gravel or rock backgrounds (Gong et al., 2013). Meanwhile, the selection criteria of the base scene consider both cloud co

52、ver (derived from Fmask, Zhu & Woodcock, 2012) and seasonal phenology. Images acquired in summer with less cloud cover are preferred. An urban area is a relatively stable land type, and a base scene acquired in summer can maximize its spectral separability w讓h other land cover types (e.g, vegetation

53、) (Sexton, Song et al., 2013a). In addition, apart from the spectral bands obtained from the base seene, NDVI time series5040()35032030027325020015012910080 Landsat 4-5 TM .Lcmdsal 7 ETM+ Landsat 8 OL1WinterAutumitSummerSpringWinter1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 199

54、7 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013Fig. 2. Temporal distribution (DOYt day of year) of Landsat images (tile, pl23r032) used in this study.X. Li etaL / Remote Sensing of Environment 166 (2015) 78-9083Fig. 3. The classification framework for multitempora1

55、urban-land extraction.acquired within each specific year were also employed to improve the classification performance, particularly for cropland (Yu. Wang & Gong. 2013; Zhong. Gong & Biging, 2014).(1) Collection of stable training samples. Training data were collected through visual interpretation o

56、n Landsat images aided by higher resolution images in Google Earth (Cong et al. 2013; Schneider. 2012; Zhao et al.t in press). Besides urban land, four additional land cover categories (i.e. forest, cropland, bareland and water) were included.【t should be noted that there are overlaps between cropla

57、nd and bareland due to cropland rotation, causing annually alternating cover classes for certain locations. Considering that the base scenes were acquired from different seasons (i.e. corresponding to differentcsidual atmospheric or phenological variations) (Sexton, Song et aL 2013a). we imple mente

58、d a hierarchical sampling scheme to collect stable training data. First, certain years from 1984 to 2013 with adequate annual time series were chosen preferenrially. With the support of NDVI time series and Google Earth, training data were collected with higher confide nee. Thereafter, they were expanded to other years as follows: training samples that were temporally adjacent to a specific year will be loaded in advance, then rechecked throughout the whole set of the training sample to identify whether their rypes have