Introduction

Urban areas facilitate intensified human activities, although they currently cover less than 1% of the Earth’s surface (Gong et al. 2013). In the wake of rapid economic development, urbanization has been attributed to the accelerating integration of developing areas in East Asia into the world economy, which has involved dramatic changes in land use and urban structure in China (Dick and Rimmer 1998). Urbanization rate of China has increased from 19.4 to 55.74% during 1980–2014 as revealed by the census data from the National Bureau of Statistics of China (http://www.stats.gov.cn).

Urbanization has permanently converted large areas of cropland and natural vegetation to impervious urban land, resulting in the changes to land surface properties including land use/cover, surface albedo, and surface roughness et al. (Kalnay and Cai 2003; Foley et al. 2005; Normile 2008; Hu et al. 2015b). This transformation in urban areas has greatly altered energy and water exchange in land–atmosphere interactions and boundary layer processes over urban canopies and downstream areas, causing further local and regional climate change (Crawford et al. 2001; Ren et al. 2008; Fernando et al. 2012; Fischer et al. 2012; Peng et al. 2012). Global temperatures have increased since pre-industrial times; however, the trends and magnitudes of climate warming induced by urbanization on different spatial scales are still controversial (Stocker et al. 2013).

To investigate the accurate impacts of urbanization on climate change, more and more regional climate models have been implemented by employing different land use change scenarios. However, recent studies indicate that accurate representation of urban land use is important but poorly captured in current climate models (Oleson et al. 2008; Gao and Jia 2013; Jia et al. 2015). The default land cover data sets used in regional climate models, e.g., Weather Research and Forecasting (WRF) model, are at a coarse resolution of 1 km. The urban and built-up areas within those data sets are not very accurate in capturing surface properties due to limits in spatial resolution and classification algorithm, which would result in deviations in model simulation (Ge et al. 2007; Friedl et al. 2010; Sertel et al. 2010). These urban data are not up-to-date and have low Cohen’s kappa values for regions in experiencing rapid urbanization process such as Southeast and East Asia (Potere et al. 2009). Meanwhile, researches showed that the better representations of urban features can directly improve model outputs. For example, Sertel et al. (2010) found that simulations with the new land cover data from Landsat led to more reasonable temperature than with the default land cover data employed in WRF model. Root-mean-square errors for the Landsat-derived data were between 2.1 and 3 °C, whereas for the existing land cover data are between 2.9 and 7.1 °C. Therefore, the deviation of urban land use/cover data declines the results of climate modeling and probably results in the uncertainty of climate warming trends and magnitude by urbanization.

Integration of human and earth system models for climate change impact assessment requires consistent projections of urban extent and accurate representation of urban intensity. The nature of urban built-up areas, such as the size, shape, intensity, and spatial context, is a key parameter of climate models (Jackson et al. 2010). Many researchers have made use of remote sensing imagery to distinguish urban lands from non-urban lands. To interpret remotely sensed data for urban derivations, methods range from pure interpretation, semiautomated and fully automated have been proposed (Sun et al. 2013). Pure interpretation started with conventional multispectral visual classification, which normally has less than 80% accuracy (Xu 2007). Semiautomated methods, such as supervised classification by the maximum likelihood classifier, have accuracy depending on subjective training samples. Due to spectral confusion of the heterogeneous urban built-up land class, classification of a multidate, multiband image would be difficult for the most common classification algorithm (Schneider et al. 2005; Hu et al. 2015a). Many studies combined different methods to improve the urban extraction. Fully automated approaches, e.g., combination of different indices without manual thresholding, have the advantages of quickly and more objectively processing data and thus are easy replicated (He et al. 2010). On the one hand, urban landscapes are typically composed of features, such as buildings, concrete pavements, which are much smaller than the grid resolution of regional climate models. Moreover, urban mapping accuracies increase with increasing spectral sensor resolution (Herold and Roberts 2010). Thus, we used Landsat data to map urban extent, because of their long-term digital archives as well as the fine spatial resolution for capturing urban characteristics in models (Bagan and Yamagata 2012; Weng 2012; He et al. 2013). On the other hand, urban intensity can be represented by fractional urban signals aggregated by the original results of urban classification data from Landsat images. This fractional urban land cover may capture well the urbanization process and provide detailed dynamic information.

Therefore, the purpose of this paper is to identify better urban representation than the existing urban data in regional climate models and consequently to provide better inputs for accurate modeling of critical environmental feedbacks over expanding urban clusters. We classified urban and built-up areas from Landsat images based on normalized indices and aggregated the original results of urban classification data from Landsat images as fractional cover information in 1-km grids. Moreover, the urbanization process over the Guangzhou–Foshan–Dongguan area was examined from 2000 to 2009 by change detection of spatial patterns and urban fractions. We also evaluated the accuracy or deviations of existing urban land cover products to examine their limitations in climate model simulation and urbanization process analysis.

Study area

Southeastern coastal region of China has been experiencing rapid urbanization and economic growth by government directives (Seto et al. 2002). Fringe area and satellite towns caused by rapid urbanization process made more mosaic land surface characterization in this region. We chose two representative areas around Shanghai and Guangdong over southeastern coastal region of China (Fig. 1). One was focused on the center of Jiading District (31°23′N, 121°15′E), northwest of Shanghai (Fig. 1a). It was confined to a 9 km × 9 km test site with a round moat and a ring road in the center. The other was located in the Guangzhou–Foshan–Dongguan area (GFD), confined to a 79 km × 93 km window (columns and rows) (Fig. 1b).

Fig. 1
figure 1

Location of this study in southeastern coastal region of China: a the red rectangle is the test site located in the Jiading District, Shanghai; b the case study area for urban fractional change detection, located in the Guangzhou–Foshan–Dongguan area

Full size image

Data and methods

Data sets for urban land cover derivation

Landsat images

Cloud-free and high-quality Landsat TM/ETM + data (Table 1) covering the above areas were collected from the USGS government website (http://earthexplorer.usgs.gov/). The Landsat data were geometrically rectified according to the Universal Transverse Mercator projection at 30 m × 30 m resolution, using second-order polynomial and bilinear interpolation. Ground control points were selected from intersections of river channels and roads, and RMSEs for geo-correction were less than 1.5 pixels. In addition, some auxiliary data were also collected to validate classification results, including a 1:50,000 scale topographic map, historical images from Google Earth, and field investigations.

Table 1 MODIS land cover product and Landsat TM/ETM + images used in this study
Full size table

Existing land cover products employed by climate models

The 500-m C5 MODIS land cover product (MCD12Q1, Friedl et al. 2010) in 2001, 2005, and 2009 (Table 1) was collected from the MODIS land team (http://modis-land.gsfc.nasa.gov/). For comparison with Landsat-derived urban land cover data, MODIS tiles over study areas of these products were resampled by the nearest neighbor method and reprojected under the datum of WGS84 by the MODIS Reprojection Tool (MRT, http://lpdaac.usgs.gov/tool/modis_reprojection_tool). The International Geosphere-Biosphere Programme (IGBP) classification scheme was chosen for urban-class derivation and the analysis of the reprocessed data. The urban and built-up category (class 13) was defined as the area covered by buildings and other man-made structures (Friedl et al. 2002). The MODIS products were reclassified into urban (class 13) and non-urban (other classes) categories to facilitate the subsequent process.

Normalized Indices-based Urban Built-up Areas Mapping

The normalized difference vegetation index (NDVI) and other normalized indices are widely used in environmental and climatic related studies. With respect to urban area classification, Zha et al. (2003) proposed the normalized difference built-up index (NDBI) to map urban built-up areas. First, continuous (c) NDVI and NDBI images were obtained by spectral responses of pixels in the following equations.

$${\text{NDVIc }} = \left( {{\text{NIR}} - {\text{RED}}} \right)/\left( {{\text{NIR}} + {\text{RED}}} \right)$$
(1)
$${\text{NDBIc}} = \left( {{\text{MIR}} - {\text{NIR}}} \right)/\left( {{\text{MIR}} + {\text{NIR}}} \right)$$
(2)

Second, the derived NDVIc and NDBIc were recoded to create binary NDVI and NDBI image with 1 for all pixels having positive value and 0 for all remaining pixels of negative value. Finally, the urban and built-up areas were extracted by taking only positive difference values of binary (b) NDBI and NDVI images.

$${\text{Ub}} = {\text{NDBIb}}{-}{\text{NDVIb}} > 0$$
(3)

However, the binary NDBI and NDVI images were recoded under the assumption that all positive values of continuous NDBI and NDVI images should indicate built-up areas and vegetated areas, respectively. Thus, this mapping method is unable to separate urban areas from patches with low fractions of natural vegetation and bare lands.

Xu (2007) proposed an urban mapping method by a logic calculation, which involved three indices including NDBI, modified normalized difference water index (MNDWI), and soil adjusted vegetation index (SAVI). The continuous images of MNDWI and SAVI can be calculated by the following equations.

$${\text{MNDWIc}} = \left( {{\text{GREEN}} - {\text{MIR}}} \right)/\left( {{\text{GREEN}} + {\text{MIR}}} \right)$$
(4)
$${\text{SAVIc}} = \left( { 1+ {\text{L}}} \right)\left( {{\text{NIR}} - {\text{RED}}} \right)/({\text{NIR}} + {\text{RED}} + {\text{L}})$$
(5)

where L is the correction factor for soil background brilliance, and the value of L is from 0 to 1 and set to 0.5 for urban areas.

The logic calculation stated that if the values of pixels in continuous NDBI images were greater than both values in continuous MNDWI and SAVI images, then the pixels were considered as built-up category.

$${\text{Uc}} = ({\text{NDBIc}} > {\text{SAVIc}}){\text{AND}}({\text{NDBIc}} > {\text{MNDWIc}})$$
(6)

Here, we improved urban areas mapping method based on Uc. By examining the spectral reflectance of 100 representative pixels of urban built-up and non-built-up areas (e.g., vegetation and water), it has been found that built-up areas have an increase in reflectance from green band to red band (band 2 and band 3 in Fig. 2). The red band is used for detection of chlorophyll (i.e., vegetation), which absorbs high in the red wavelengths, as well as for the observation of roads and bare lands, which reflect highly in the red band. The green band is mainly used to distinguish vegetation from background categories, as vegetation reflects highly in the green band. Taking advantage of a normalized calculation, an urban index can be expressed by the standard difference between these two bands.

Fig. 2
figure 2

a Spectral reflectance of urban built-up, vegetation, and water in different bands, taking the Jiading Landsat ETM + image as an example, b spectral features of different normalized indices represented by urban built-up, vegetation, and water

Full size image
$${\text{UIc}} = ({\text{RED}}{-}{\text{GREEN}})/\left( {{\text{RED}} + {\text{GREEN}}} \right)$$
(7)

Based on the above spectral characteristic, the improved method involved four indices including NDBIc, MNDWIc, SAVIc, and UIc. The four indices were used as four bands of a new image. For the new image, in addition to the band characteristics described by Uc, the positive value of UIc could help to further capture weak urban signals, especially in those areas of low fractional vegetation. The Integrated Urban Index (IUI) approach was proposed to segment urban built-up areas from other land types using an improved logic calculation.

$${\text{IUI}} = \left( {\text{NDBIc > SAVIc}} \right){\text{AND}}\left( {\text{NDBIc > MNDWIc}} \right){\text{AND}}\left( {{\text{UIc}} > 0} \right)$$
(8)

The IUI approach and existing mapping methods (Ub and Uc) were applied and evaluated in Jiading District by logic tree operation to identify urban built-up and non-urban classes.

Assessment methods

Accuracies of the IUI against existing methods (Ub and Uc) were performed at both the pixel scale and the model grid scale of 1 km. At the pixel scale, based on historical images from Google Earth and field investigation, 100 random samples for built-up areas and 100 for non-built-up areas for each method were created by Python code to validate our results. Four commonly used measures, namely the overall accuracy, the kappa (κ) coefficient, omission error, and commission error, were calculated from the confusion matrix to reveal the accuracies of the IUI against existing methods. The overall accuracy is computed by dividing the total number of correctly classified pixels by the total number of reference pixels. The κ is calculated by multiplying the total number of pixels in all the ground truth classes (N) by the sum of the confusion matrix diagonals (x kk ), subtracting the sum of the ground truth pixels in a class times the sum of the classified pixels in that class summed over all classes (x kΣ x Σk ), and dividing by the total number of pixels squared minus the sum of the ground truth pixels in that class times the sum of the classified pixels in that class summed over all classes by the following equation.

$$\kappa = \frac{{N\sum\limits_{k} {x_{kk} } - \sum\limits_{k} {x_{k\varSigma } x_{\varSigma k} } }}{{N^{2} - \sum\limits_{k} {x_{k\varSigma } x_{\varSigma k} } }}$$
(9)

The omission error represents pixels that belong to the ground truth class, but the classification technique has failed to classify them into the proper category. The commission error represents pixels that belong to another class that are labeled as belonging to the class of interest.

At the model grid scale, urban fractions were calculated by the algebraic proportions of urban and built-up areas within the given landscape unit. Firstly, we aggregated the original results of urban classification data from Landsat images as fractional cover information in 1-km grids. Then, we chose three grids as exemplars with 990 m × 990 m window to represent urban fractions at the model grid scale for different development levels, i.e., from rural, to urban–rural fringe, and to urban. The 990 m is chosen because it can be divided by 30 m (the resolution of Landsat data) and comparable to 1 km. Next, the urban and built-up areas for corresponding locations of grids from QuickBird imagery were delineated by manual interpretation and on-screen digitization. The proportions of these outlined urban areas within grids were computed as urban fractions for evaluation of reference. Thus, urban fractions aggregated from Landsat and QuickBird images were compared and examined.

Detection of urban expansion

A framework for change detection of urban expansion at the model grid resolution of 1 km based on IUI was created. Firstly, the IUI approach was applied to Guangzhou–Foshan–Dongguan area to map urban areas based on Landsat images in 2000, 2005, and 2009, respectively. The 30-m IUI-based urban areas were derived for the three periods. Second, square grids at 1-km spatial resolution were created to cover the boundaries of the study area. Next, urban land classification results were aggregated to a 1-km resolution by calculating proportions of urban areas within grids. Thus, fractional covers of urban land in 1-km grids were derived for these years. We examined urban expansion by change detection of spatial pattern and urban fractions. The interval of 0.2 was taken as an example to analyze development of urban intensity from 2000 to 2009. At last, we overlaid and compared with MODIS urban land cover to examine accuracy of existing land cover data in climate models.

Results

Comparison of urban and built-up areas classifications

Urban areas were derived by Ub, Uc, and IUI approaches from Landsat ETM + image taken on April 27, 2000, located in Jiading District, Shanghai (Fig. 3). The IUI result for the test site was in high agreement with the expected urban morphology in the original Landsat imagery. The Ub result only delineated the outline of urban and built-up area, neglecting details within the city. The Uc result represented some information of interior urban, but it described little in area of low vegetation density within the city. Thus, the IUI approach is a better alternative than existing methods for patchy urban area detection, because it not only delineated the urban outline but also classified urban land from the inner mosaic land surface.

Fig. 3
figure 3

Comparison of results by IUI and existing methods: a the false color composition (red:NIR, green:red and blue:green) of the Landsat ETM + image taken on April 27, 2000, located in Jiading District, Shanghai; b–d classification images of urban and built-up areas by methods of Ub, Uc, and IUI, respectively; the red rectangle in each panel shows where existing methods of Ub and Uc produced obvious commission error, incorrectly assigning some non-built-up areas as urban class

Full size image

Accuracy

Pixel scale

The overall accuracy of IUI approach was 88% (Table 2), greater than Ub (70.5%) and Uc (83.5%). The values of the κ coefficient for Ub, Uc, and IUI were 0.41, 0.67, and 0.76, respectively. Moreover, Ub and Uc methods had higher commission errors than the IUI approach, incorrectly assigning some non-urban areas to urban class (see the area in the red rectangle in Fig. 3). The results demonstrated that the IUI was able to capture detailed urban land use with higher accuracy than existing methods when applied simultaneously in the same context.

Table 2 Accuracy assessments of IUI and existing methods at both the pixel scale and the model grid scale of 1 km
Full size table

Model grid scale

Grids at around 1 km were chosen to evaluate the ability of IUI method to capture urban signals in various cases. Grid I, II, and III, located in the northwest of the Jiading District, were taken as exemplars to represent urban fractions of different development levels in this region (Fig. 4), i.e., from rural, to urban–rural fringe, and to urban. Results showed that urban extent delineated by Ub was in the least agreement with the expected urban morphology from high-resolution QuickBird and pan-sharpened images. Urban fractions of each grid using Ub, Uc, and IUI methods, as well as corresponding locations from the QuickBird image, are compared (Table 2). The values of urban fractions calculated from the IUI approach were more comparable to those from the QuickBird image than Ub and Uc methods. For grid III, urban fraction using Ub was 100%; however, there were obviously non-urban areas in both the QuickBird and pan-sharpened images. The results indicated that the urban fractions from the IUI had a higher accuracy than existing methods when compared at the 1-km scale. IUI can effectively capture accurate fractional urban signals for different development levels of urbanization and therefore better estimate urban fractions at coarse resolution for climate models.

Fig. 4
figure 4

Grids for accuracy assessments of IUI and existing methods at 1-km scale. a The false color composition (red:NIR, green:red and blue:green) of the Landsat ETM + image taken on April 27, 2000, located in Jiading District, Shanghai. Grid I, II, and III were chosen as exemplars in a to represent different level of urban development. Urban extents in column b, c, e were identified by methods of Ub, Uc, and IUI separately. Red outlines in column d were visual interpretation of urban areas in QuickBird image. Column f was false color composition of pan-sharpened imagery

Full size image

Changes in fractional urban lands under urbanization in the GFD area

Spatial pattern

The spatiotemporal patterns of urban fractions over the Guangzhou–Foshan–Dongguan area at the 1-km scale in 2000, 2005, and 2009 are shown in Fig. 5. From 2000 to 2005, the development of satellite towns was prevalent in the north and southeast, such as Qingyuan and Huadu. From 2005 to 2009, apart from satellite towns in Zengcheng, the main change in urban fraction resulted from increasing intensity of core areas, i.e., Guangzhou, Foshan, and Dongguan. Results showed that urban development in this region has been directed out of the core areas to nearby towns, leading to a poly-nucleated urban form. Multidirectional satellite towns emerged and located in the periphery from the core city. The connections between the core area and satellite towns are covered by low intensity of urban fractions.

Fig. 5
figure 5

Spatiotemporal change patterns of urban fractions in GFD area from 2000 to 2009. Urban fractions were calculated at an interval of 0.2

Full size image

Urban fractions

Change detection of urban fractions at the model grid resolution was applied by IUI for detailed urban expansion in GFD area. Urban fractions were analyzed and summarized at intensity of 0.2 (Table 3). Urban fractions in the intensity of 0.6–1 increased from 2000 to 2009. The 0.6–0.8 intensity increased twice more in 2000–2005 (5.10%) than in 2005–2009 (1.71%), whereas the 0.8–1 intensity increased comparatively in 2005–2009 (1.67%) and in 2000–2005 (1.92%). In contrast, urban fractions in the intensity of 0–0.4 decreased from 2000 to 2009, indicating denser urban canopy and reduced gaps among urban clusters in past decade. The 0–0.2 intensity decreased much more in 2000–2005 (7.73%) than in 2005–2009 (1.27%). It is noteworthy that urban fractions in the intensity of 0.4–0.6 increased 3.63% in 2000–2005 and then decreased 0.41% in 2005–2009. From 2000 to 2005, we found that the maximum increase in urban fractions in this region was in the intensity of 0.6–0.8, whereas the maximum decrease was in 0–0.2. From 2005 to 2009, results showed that the maximum decrease in urban fractions was in the intensity of 0.2–0.4. Overall, the total fractional urban change in 2000–2005 (10.65%) was approximately three times greater than that in 2005–2009 (3.38%).

Table 3 Percentage of fractional urban change in the GFD area from 2000 to 2009
Full size table

Comparison of urban expansion using MODIS and Landsat data sets

Urban land spatial patterns in GFD from MODIS land cover product were compared with IUI classification results from Landsat imagery (Fig. 6). All these data were reprocessed to 1-km resolution and the same projection. We aggregated the original results of urban classification data from Landsat images as fractional cover information in 1-km grids. The results from Landsat images and socioeconomic data sets revealed rapid urbanization during 2000–2009, with dramatic changes in urban fringe areas. By contrast, minor fractional urban signals were detected from MODIS land cover data sets, in which the core urban areas were overestimated and rapid changes in the urban fringe were largely ignored.

Fig. 6
figure 6

Comparison of the urban land pattern in GFD area from MODIS land cover a, Landsat images b in 2009 at 1-km resolution

Full size image

To examine whether land cover products effectively captured urban expansion process, urban area derived from Landsat, MODIS land cover data sets, and census data were used to retrieve the urbanization rate, respectively (Table 4). Urbanization rate was defined as the rate of urban area growth per year. The rate was 0% from the MODIS data sets for GFD area, which showed that MODIS failed to capture recent urban expansion in the region. The rate for 2000–2005 was approximately three times as much as the rate for 2005–2009 from both Landsat data and census data. Census results were larger than Landsat results. It is possibly because census data were from the whole administrative divisions of Guangzhou, Foshan, and Dongguan cities. However, the rectangular grid of GFD area is not exactly the administrative boundary of census data.

Table 4 Urban expansion comparison using MODIS and Landsat data in GFD area
Full size table

Discussion

First, in commonly used climate models, the dominant category and urban fractions in a given pixel are used to characterize urban land cover. For dominant category, a coverage of built-up environment greater than 50% of a given landscape unit is defined as the threshold for an urban grid cell (Schneider et al. 2009). That is, those developing urban patches that cover 50% of the grid were taken as dominant underlying composition. For urban fractions, the proportions of urban areas within model grids are calculated to capture the detailed urban density and inner connections. Thus, both the dominant category and urban fractions need the information of fractional urban at the model grid resolution. In the accurate assessment at the 1-km scale, the grid II has a moderate level of development, with urban fractions approximately 50%. Urban fractions using the IUI and the QuickBird image were both <50%, whereas urban fractions using Ub and Uc were much higher (Table 2). Therefore, both dominant category and urban fractions calculated by Ub and Uc had caused biases in urban underlying properties when applied in modeling. Moreover, in the comparison of urban expansion from MODIS and Landsat data sets, MODIS ignored fractional urban signals in the urban fringe and overestimated urban patches that cover 50% of the modeling grid.

Second, the total fractional urban change in 2000–2005 was almost three times greater than that in 2005–2009 (Table 3), suggesting that this area experienced high rates of urban expansion in 2000–2005, but the rates decelerated in 2005–2009. This indicates that the area has reached a mature stage of urbanization during the later five years. These urban dynamics are influenced by policy shifts, economic development, and demographic patterns (Seto et al. 2002). During the decade, China was implementing the 10th and 11th Five-Year Plans. The former plan gave China the first blueprint for the new century with main targets for economics and urbanization levels. The GFD area was developed during this period, with rapidly emerging satellite towns and expansion of core areas. These towns were not well connected to existing central urban areas (Zhang and Seto 2011). The latter plan, started from 2006, had a goal to enable disadvantaged groups and less developed regions in sharing the fruits of economic growth (Fan 2006). The government highlighted the role of urban clusters in fostering urbanization for the first time. These policy shifts might have contributed to the slowing down of urban development in this area, with extension in the core areas to satellite towns for developing an urban cluster. The high density (0.8–1) of fractional urban change indicates that the GFD area has become denser and packed with more continuous urban canopy in the decade. Such changing patterns of urban density and inner connections may potentially alter interactions between urban canopy and atmosphere and urban climate effects.

Conclusions

Reliable information on urban areas and change detection plays an important role in urban studies. The IUI can effectively capture better urban signals for different levels of urban fractions. The case study area in GFD experienced urban expansion during the decade, with fractional urban growth three times greater in 2000–2005 than in 2005–2009. The spatiotemporal patterns of urban expansion were presented with remarkable growth of satellite towns and core areas. We expect that better estimates of changing patterns of urban density and inner connections will help improve performance of models in capturing critical environmental feedbacks over expanding urban clusters.