2.1.

Study Area

Our research focuses on the urban agglomeration area in the Bohai Rim (38°30′N to 41°8′N; 115°17′E to 119°30′E), covering approximately $42,000{\mathrm{km}}^2$. It includes Beijing and Tianjin municipalities, a medium-sized city of Tangshan, and several small cities belonging to Hebei Province in the east of Beijing. The Jing-Jin-Tang urban agglomeration is surrounded by the Yan Mountains in the north and the Taihang Mountains in the west (Fig. 1). Most of the urban areas are located in plain areas between 20 and 60 m. The region has a semihumid warm-temperate monsoon climate with an average yearly temperature of 11.7°C and annual rainfall of 700 mm. Agricultural lands dominate the suburban environment and winter wheat and maize are the main crops.

Fig. 1

Elevation map of the Jing-Jin-Tang urban agglomeration. The area is covered in three TM/ETM+ scenes.

Beijing, the capital city, is among the most developed cities in China. It is the country’s political, cultural, and educational center and also a major transportation hub, with dozens of railways and roads passing through. Tianjin, a municipality under direct administration of the central government, ranked sixth in terms of urban population in China. As a dual-core city, Tianjin is divided into the old city and the “Binhai New Area” (BNA). Tangshan is a prefecture-level city in Hebei Province and an important heavy-industry city in North China. Each city is divided into county-level administrative subdivisions. All the subdivisions can be categorized into three groups: urban district, suburban district, and rural district, by their distance from the city center. Urban districts are located in the inner city. Suburban districts refer to outlying areas of a city and have a lower BU density than the inner city. Rural districts are located far from the inner city and are less urbanized with a low population density. These three cities experienced rapid urbanization in the last 20 years as reported in national census data in 1990, 2000, and 2010 (Table 1). Megacities are defined as urban agglomerations with at least 10 million inhabitants.²⁹ The metropolitan areas of Beijing and Tianjin can thus be treated as megacities (Table 1). The total population of Jing-Jin-Tang reached 40.127 million in 2010.

Table 1

Population growth of the Jing-Jin-Tang urban agglomeration in the past 20 years (in thousands).

2.2.

Dataset

The Landsat multitemporal dataset was used to measure urban development in the study area. The Landsat data have been continuously available since 1972, featuring geometric resolutions of 79 m (multispectral scanner since 1972), 30 m (TM since 1982), and 15 m (panchromatic band of the ETM+ since 1999). These long-term data records provide an extended time series that are superior to other high-resolution sensors.^33,34 Moreover, the relatively high temporal resolution (16 days) and the sensor’s wide swath (185 km) facilitate the depiction of extended areas of urban agglomerations.

The ETM+ and TM data in this study were acquired from the remote sensing data sharing platform of the Center for Earth Observation and Digital Earth, Chinese Academy of Sciences ( http://ids.ceode.ac.cn/) and the global land cover facility, University of Maryland ( ftp://ftp.glcf.umd.edu/glcf/Landsat/). Three Landsat footprints (path/row 122/32, 122/33, and 123/32) cover the entire urban agglomeration region (Fig. 1). We used images acquired in July–September, since bare soils were covered by vigorous vegetation and could be better differentiated from BU areas during this period. Particularly, images under cloud-free conditions were manually selected with scene 1 to 3, 4 to 6, and 7 to 9 representing the development of urban areas in earlier 1990s, 2000s, and 2010s respectively (Table 2). All TM/ETM+ images listed above were geometrically rectified and mosaicked. Atmospheric correction was conducted with atmospheric and topographic correction software.³⁵

Table 2

List of Landsat data used in this study.

3.1.

BU Area Extraction

Urban areas and their decadal extension are the major concern of this study. Here, we categorized the land surface types of the study region into BU and nonurbanized areas, which could be delineated based on the difference of their spectral features. To simplify the process, a water mask was built to eliminate water bodies by visual interpretation of satellite images. Water bodies include rivers, lakes, and ponds. Human settlements located higher than 60 m in elevation were also eliminated for considering the large uncertainty of BU area detection over rugged surfaces and steep terrains. We adopted an enhanced difference BU index to extract BU areas from other land cover types.

During the time of image acquisitions in July–September, nonurbanized areas were natural vegetation or agricultural lands covered with healthy, green plants in their growing season. Therefore, these areas were characterized by high values of the normalized difference vegetation index:^21,22

BU areas are composed of manmade constructions such as buildings and paved roads where green cover is limited. Therefore, these areas often reach low NDVI values. Rather, they are characterized with slightly higher reflectance in the short-wave infrared band than that in near-infrared. Contrarily, vegetated nonurbanized areas often experience a rapid decrease from near-infrared to short-wave infrared, especially for vigorous vegetation with high leaf water content.³⁶ The NDBI could be calculated as²¹

We then define the BU index:

Apparently, BU areas have large BU values while the nonurbanized areas have small ones. An optimal threshold method is applied to assign each pixel into the corresponding class. For each scene, a specific threshold is defined with randomly selected training samples. To simplify the process, NDVI, NDBI, and BU were normalized into (0, 255) using linear stretch.

3.2.

Spatial Metric Analysis

Urban areas are characterized with complex spatial heterogeneity with a mixture of various land cover types. Urban growth in different cities may follow different sprawling patterns. It can be linear along traffic axes or punctual. It may have one center or multiple centers. It may increase the density of constructions in an area or sprawl to surrounding areas. In this study, we try to characterize urban growth at regional and city levels using a set of quantitative measures of spatial metrics. The BU areas extracted from Landsat images were put into FRAGSTATS to generate spatial metrics.

In FRAGSTATS, we calculated six parameters to quantify the urban footprint at each time: class area (CA) or absolute urban area, percentage of landscape (PLAND), number of patches (NP), largest patch index (LPI), landscape shape index (LSI), and normalized landscape shape index (NLSI). The CA metric is the total BU area and the PLAND is its percent cover in a specific area. They describe the growth of urban areas at different time frames. The NP is a measure of fragmentation of urban areas. The eight-neighbor rule was used in delineating patches. An increase of NP indicates the conversion of nondeveloped land into BU areas in suburbs and rural areas, where BU areas occur as isolated patches. The LPI represents the percentage of an urban core area as a proportion of total urban land. It quantifies the percentage of total landscape area comprising the largest patch. It increases when urban areas become more aggregated into urban cores. The LSI provides a measure of class aggregation or clumsiness, which increases as the patch type becomes more disaggregated. The NLSI rescales LSI for each class to give a simplified measure of class aggregation. These metrics are further described in Table 3.

Table 3

Spatial metrics used in this study.

In addition to the analysis using the complete spatial extent in each city, spatial patterns at local scales can be illuminated via zonal analysis. A zonal-based parameter, BU density (BD), is also employed to characterize spatiotemporal urban sprawl.²³ It is the ratio of the BU area to the total area of the particular zone. BD values of each zone are calculated using ArcGIS 9.3.

4.1.

BU Index and Urban Development Mapping

To demonstrate the results of BU area detection, Fig. 2 shows an example subset of the ETM+ data for the metropolitan area of Tangshan in 2000. The urbanized city’s extent, the linear axial growth, and the existing satellite towns are clear in the subset image [Fig. 2(a)] and the build-up index [Fig. 2(b)]. The extracted results are illustrated in Fig. 2(c). Due to the spectral similarity of bare soil and BU areas in the TM/ETM+ data, it is difficult to delineate the two land cover types. High classification confusion error also exists under complex topographic conditions. To achieve a precise depiction, we constrained our analysis to plain areas.

Fig. 2

Built-up (BU) area extraction from an example Landsat image. (a) Subset of Landsat ETM+ Bands 5, 4, and 3 displayed in red, green, and blue, respectively, (b) BU index, and (c) BU area extracted.

We implemented a comparative analysis of BU area extraction results in different periods to monitor and analyze land cover changes in the Jing-Jin-Tang area. The urban land map in Fig. 3 reveals the urban footprints and their spatiotemporal evolution since the 1990s, and shows the evolution patterns of the Jing-Jin-Tang urban agglomerations in the past two decades. From this figure, we can visualize the spatial expansion of each city, the increase of satellite towns, and the growth of traffic infrastructure all over the study area. The traffic network grew rapidly during this period. The Jing-Jin-Tang highway, the first highway approved by the Chinese government, was built between 1984 and 1995. The eight-lane Jing-Jin highway (135-km long) and high-speed railway (120-km long) between Beijing and Tianjin began to construct in 2005 and was finished in 2008. These new infrastructures dramatically promoted regional development and urbanization in this area.

Fig. 3

Expansion of BU areas in the Jing-Jin-Tang urban agglomeration between 1990 and 2010.

For each city, a dense urban core was located in the center in the early 1990s, surrounded by smaller urban centers of counties and towns under each city municipal administration. Beijing is the most populous city of the three and has the largest urban center. New BU areas, settlements, and infrastructure were developed in the east between 1990 and 2000 and the north and south between 2000 and 2010. It is related to Beijing’s “new city building” policy in the corresponding districts and counties. The extension is limited in the west due to the limited land resources in mountainous areas. The metropolitan Tianjin extends along a southeastern axis defined by the Jing-Jin-Tang highway. The east of Tianjin is located in the centric position of Jing-Jin-Tang and also the coastal area. With great geographic advantages, the BNA has been under development as a major economic development zone under the opening-up policies since 1994. Industrial districts, commercial districts, airports, and seaports have been built and extended. The gross domestic product grew at a speed of 20.6% between 1994 and 2010.³⁷ The counties in northwestern and southeastern Tangshan also sprawled quickly between 2000 and 2010.

We assessed the accuracy of each BU map with randomly distributed points. Real-time ground data collection is not possible in such a large study area. Alternatively, we collected validation data by visually interpreting the Landsat images since build-up, and nonurbanized areas (mostly agricultural lands in this region) can be easily identified in these 30-m resolution datasets (as shown in Fig. 2). Validation points were also collected from high-resolution imagery such as SPOT5, Ikonos, and QuickBird that were collected in previous studies. For each scene, approximately 1% of the total pixels, in at least 80 polygons including both “BU” and “not BU” classes, were obtained with 70% used for threshold optimization and 30% for accuracy assessment.

The error matrix approach was applied to assess the overall accuracies and two types of classification errors.³⁸ The commission error measures the proportion of samples that are classified as a given class but are not actually from that class, and the omission error measures the proportion of samples that are not identified as a given class but are actually from that class. Table 4 shows the estimation of the commission and omission errors, the overall accuracy, and $\kappa$ coefficient for the same year. Since there are only two classes (BU areas and nonurban areas), only the omission and commission errors of BU areas are shown in Table 4. The overall accuracies range from 75% to 91% correctly classified pixels. Higher omission errors are observed in small villages and the urban-fringe areas, which are often mixed with nonurbanized areas and could not show the obvious spectral signature of urban areas. High commission errors exist in barren lands in which bare soil has a spectral signature similar to BU areas.

Table 4

Accuracy assessment of the resultant BU maps.

4.2.

Spatial Pattern Analysis at the City Level

The changes of CA and PLAND indicate that all the cities have experienced rapid urban area expansion since 1990 (Fig. 4). This is driven by social and economic impacts such as population growth, economic development, and infrastructure construction. The CA of Beijing is lower than Tianjin and Tangshan, since only metropolitan areas in southeastern Beijing were urban lands of the Jing-Jin-Tang urban agglomeration in our study (Fig. 3). Meanwhile, the PLAND of Beijing is higher than the other two cities, which indicates the percentage of BU areas in Beijing is highest across the study area. The NP increased in the 1990s and decreased in the 2000s in Beijing [Fig. 4(a)], indicating the amalgamation between satellite towns (and the urban core as well) and individual urbanized patches in the 2000s. The continuous NP increased in Tianjin [Fig. 4(b)] and Tangshan [Fig. 4(c)] in 2000 and 2010, revealing the growth of new towns in suburban areas. The slight increase of LPI in Beijing and Tianjin implies that their urban cores have continuously comprised more urban land in the past three decades. In Tangshan, the LPI is comparatively stable. The urban growth occurs not only in the urban core but also in counties at similar speeds. The slight decrease of LSI and NLSI indicates that Beijing is becoming more aggregated and a homogeneous urban patch has formed from 2000 to 2010. On the contrary, the increase of LSI in Tianjin and Tangshan implies that they are becoming more decentralized.

Fig. 4

Fluctuation of six spatial metrics between 1990 and 2010 in Beijing (a), Tianjin (b), and Tangshan (c).

4.3.

Zonal Analysis of Spatial Patterns at the County Level

A BU density analysis of each subdivision was carried out for the period of 1990 to 2010. In each city, the highest BU density exists in the metropolitan area or the center of the city (Fig. 5). The BDs of the metropolitan areas of Beijing, Tianjin, and Tangshan have different growth rates. For Beijing and Tianjin, the BD increased from 0.88 and 0.77 to higher than 0.9 in the past 20 years. For Tangshan, the BD only increased from 0.28 to 0.55. Tianjin grew more rapidly in the 1990s and Tangshan in the 2000s. As municipalities, the construction and development of the metropolitan Beijing and Tianjin started in the early 1980s and became almost fully developed at the end of 1990. Tangshan, like many medium-sized cities in China, is still enduring a development process with an accelerating pace.

Fig. 5

BU density (BD) variation in metropolitan area of Beijing, Tianjin, and Tangshan between 1990 and 2010.

Three of the four urban districts in Beijing—Chaoyang, Fengtai, and Haidian—had BD values above 0.4 in 2010. The BDs of the four urban districts of Tianjin were between 0.3 and 0.4. Although located in the outskirts, Tanggu and Hangu had a higher BD than other suburban districts. They are near the coastal zone in the east of Tianjin and are part of the BNA. The counties between Beijing and Tianjin (Xianghe, Sanhe, and Dachang counties) also had a BD above 0.4. Population and BU area increased in these counties, partly caused by the residential migration from Beijing. Many people who work in Beijing choose to live in these counties to avoid the high living expenses. Among all the counties in Tangshan, Yutian, Luan, Luannan, and Fengrun were the densest BU areas, with a BD between 0.2 and 0.3. They are located in the plain area.

The combination of remote sensing and spatial metrics analysis provides a promising method for studying urban change. However, because of the spectral mixture problem and visually selected thresholds between urban and nonurban areas in mapping process, the accuracy of the proposed method needs to be improved.³⁹ Classification techniques combining spectral, spatial, and temporal information of remote sensing images can be developed to improve the classification accuracy. Moreover, satellite data at different spatial scales captured the urban growth footprints with different details. With the increasing availability of satellite observations, the integration of multisource remote sensing data is promising for better delineation of urban extent.