Satellite Monitoring of Urbanization in China for Sustainable Development: The Dragon ‘Urbanization’ Project

EarthzineArticles, Earth Observation, Original, Urban Monitoring Theme

Image of the study areas from Google Earth.

Yifang Ban
Division of Geoinformatics, Department of Urban Planning & Environment, Royal Institute of Technology – KTH, Stockholm, Sweden, Email:
Paolo Gamba
Remote Sensing Group, Department of Electronics,University of Pavia, Pavia, Italy, Email:
Peng Gong
State Key Lab of Remote Sensing Science, Chinese Academy of Sciences & Beijing Normal University, Beijing, China, Email:
Peijun Du
Department of Geoinformatics, Nanjing University, Nanjing, Jiangsu, China, Email:

Image of the study areas from Google Earth.

Figure 1: Study areas. Image Source: Google Earth.

China, the most populous country on Earth, has experienced rapid urbanization due to an unprecedented combination of economic and population growth since the early 1980s. Two decades ago, fewer than 20% of China’s people lived in urban areas; today it is approximately 50%; and by 2030, it’s estimated that another 400 million people will move to urban areas. Urbanization and the impact of human settlements are two of the main causes of global environmental degradation. Therefore, the monitoring of urbanization and its impact on the environment is of critical importance for urban planning and sustainable development in China [1].
Most of the efforts in China for urbanization monitoring are based on optical remote sensing. Due to frequent cloud cover, smog, haze and sand storms, however, optical data may not be available during critical monitoring cycles. With its all-weather capability, Synthetic Aperture Radar (SAR) is an attractive data source for monitoring urbanization. The overall objective of this research is to investigate spaceborne SAR data and fusion of SAR and optical data for monitoring urbanization in China, and to assess the impact of urbanization on the environment for sustainable urban development. The specific objectives are to:

– Develop effective methodologies to extract urban land-cover information from spaceborne SAR data;
– Develop effective change detection methods for urbanization monitoring using ENVISAT ASAR and ERS SAR data;
– Investigate the fusion of C-, X- and L-band SAR data, and fusion of SAR and optical data for urban land-cover mapping and change detection;
– Determine which satellite data at which resolution/scale and during which seasons are suitable for identifying which urban land covers and monitoring what types of changes;
– Assess the impacts of these changes on the environment for sustainable urban development.

image of different LU/LC maps of Shanghai

Figure 2: Different LU/LC Maps of Shanghai Test Site: using BIS segmentation algorithm (left); using Marpu segmentation algorithm (right).

Three focus regions, Jing-Jin-Ji (Beijing, Tianjin, Hebei), Yangtze River Delta, and the Pearl River Delta, were selected for this research (Figure 1). These areas represent the fastest growing regions in China, with various driving forces for urbanization, such as economic development, policy and culture. ENVISAT ASAR C-band AP data were acquired in Beijing, Shanghai and Guangzhou during 2008 and in the Yangtze River Delta during 2009. ERS-1/2 SAR C-HH SAR data acquired in the late 1990s were collected from the ESA archive for change detection. Two scenes of ALOS PALSAR L-band, dual-polarization SAR data and two scenes of TerraSAR-X data were acquired over Shanghai in 2008.
Several scenes of Beijing-1 MSI data, HJ-1B multispectral and hyperspectral data and CBERS-02B panchromatic and multispectral data were acquired over Beijing, Shanghai and Guangzhou in 2009. Landsat data and SPOT data acquired in late 1980s through late 1990s were also collected for change detection.
Original HJ-1B image of Bejoing, and LU/LC classifications of HJ-1B (UR), ASAR (BL) and fusion of ASAR & HJ-1B data (BR).

Figure 3: The original HJ-1B image (UL of Beijing), and LU/LC classifications of HJ-1B (UR), ASAR (BL) and fusion of ASAR & HJ-1B data (BR).

Field data were collected in selected areas in the three focus regions during major new acquisitions for calibration and validation.
Recent literature shows an increasing interest in using SAR image for human settlements expansion monitoring, especially in those countries where cloud cover, haze and smog may prevent consistent optical images acquisition. Even though there is a general consensus that SAR provides the solution for this issue, and that spatial analysis is usually preferred to single pixel analysis, there is less agreement on general methodologies for extracting human settlement extent in large geographical areas. Moreover, only a few algorithms are devoted to land-use land-cover (LU/LC) classification using SAR. In most cases they are based on polarization signatures, a feature which relies more on the physics of surface scatterers than on effective land use. The works we propose are aimed at studying the Chinese environment in terms of LU/LC using SAR images at high resolution and based on two different phases.
At first, human settlement extents are extracted using an algorithm called BuiltArea[2], based on three Local Indicators of Spatial Association and two textural features. The algorithm is, to some extent, sensor and resolution independent and provides a binary image where built-up areas are represented as white pixels on a dark background. The second phase is devoted to the classification of urban areas in terms of three different land uses according to CORINE [3] nomenclature as applied to the Chinese environment. The classification procedure is primarily based on image segmentation techniques in order to encode statistically homogeneous areas as independent regions. The aim is to classify each region as a single land-use based on the similar statistical behavior characterizing each land-cover. The followed criterion is the a-majority one, selecting as the region representative class the one to which the majority of the pixels in that region belong.
image of Change detection: a binary classification problem.

Figure 4: Change detection: a binary classification problem.

Further research continued in two different but complementary ways. At first, three different segmentation techniques, based on different methodologies, were compared in order to establish which segmentation approach was more suited for SAR images. The Canny edge detector and region merging techniques used in the first analysis were then compared with other two algorithms specific for remote sensing data segmentation: the Berkeley University “BIS” [4]and the Marpu [5]algorithms (Figure 2). The first is an object-based image analysis algorithm, where compactness, shape and scale parameters may be adjusted in order to obtain the desired level of segmentation. The other algorithm is based on a graph theoretic approach together with a region growing technique, where the graph is used to guide the merging process. The conducted analysis, based on ENVISAT/ASAR images of Shanghai and Beijing, has shown that higher accuracies are obtained using the Marpu algorithm, followed by the ImgSeg and Canny edge approaches.
The second part of this research is mainly focused on the fusion of SAR and optical information, exploiting optical multispectral data provided by sensors onboard of Chinese Beijing-1 and HJ-1 satellites (launched in 2005 and 2008, respectively)through application of Earth observation, disaster monitoring and urban development tasks. In particular, the idea is to perform the segmentation task on the
The detected positive and negative changes in yellow and blue, respectively, overlaid with a SAR image composite using (a) GG, (b) LN, (c) NR and (d) WR.

Figure 5: The detected positive and negative changes in yellow and blue, respectively, overlaid with a SAR image composite using (a) GG, (b) LN, (c) NR and (d) WR.

optical images, taking advantage of the characteristics of the segmentation algorithms originally conceived for optical data. Moreover, even if optical images acquired for the same area as the SAR sensor are not always available(and considering the fact that segmentation algorithms provide a block-level subdivision of the image, an urban feature which almost never changes), dated optical images also may be employed for this task, easing the entire procedure. Results show that the segmentation provided by optical data, together with a minimum distance supervised classification applied to the SAR image, lead to a better discrimination between classes in terms of accuracies evaluated with the confusion matrix. It is also worth considering that all the analyses performed have been validated with ground truth extracted from optical images by means of shapefile extraction, subsequently superimposed on SAR images.
The synergistic effects of multitemporal ENVISAT ASAR and HJ-1B data also were investigated for urban land cover classification using object-based support vector machines. The study demonstrates that improved urban land cover classifications can be achieved by fusion of data from radar and optical sensors in comparison to each sensor alone (Figure 3).
SVM Classification Results, LL: CBERS-2B, UR: BEIJING-1, LR: HJ-1B.

Figure 6: SVM Classification Results, LL: CBERS-2B, UR: BEIJING-1, LR: HJ-1B.

Change detection is considered as a binary classification problem (Figure 4). A generalized version of the Kittler-Illingworth minimum-error thresholding algorithm, taking into account the non-Gaussian nature of SAR images, was tested to automatically classify the change variable derived from SAR multitemporal images into two classes(change and no change). The modified ratio operator was developed to take into account both positive (i.e., backscatter increase such as new built-up areas) and negative (i.e., backscatter decrease such as new airports or new golf courses) changes. Various probability density functions such as Generalized Gaussian (GG), Log normal (LN), Nakagami ratio (NR), and Weibull ratio (WR) models were tested to model the distribution of the change and no change classes.
The results show that the Kittler-Illingworth algorithm applied to the modified ratio image is very effective in detecting temporal changes in urban areas using SAR images (Figure 5).Log normal and Nakagami density models achieved the best results. The Kappa coefficients of these
 Decision Tree Classification Results, LL: CBERS-2B, UR: BEIJING-1, LR: HJ-1B.

Figure 7: Decision Tree Classification Results, LL: CBERS-2B, UR: BEIJING-1, LR: HJ-1B.

solutions were 0.82, while the false alarm rates were 2.7%. The positive change detection accuracies were very good in general, except for the Weibull ratio model where negative change accuracies were relatively poor. This is mainly due to the low intensity of change in these areas compared with the very high intensity of change found in positive change areas (urban growth).
In this study, HJ-1 multi-spectral data with 30m resolution and hyperspectral images with 100m resolution were used to classify the land cover into six types: water body, public green space, agricultural land, built-up areas, non-use land, and clouds. Thermal images with 300m resolution were used to evaluate the urban heat island effect. Fieldwork was conducted on Sept. 17, 2009, and was used to evaluate the accuracy of classification results. SVM with RBF kernel and SAM classifiers were used to classify the HJ multi- and hyper-spectral images. In order to quantify the degree along the rural-urban gradient, Moran’s I index and semi-variance were used to assess the spatial autocorrection and describe the scale and pattern of spatial variability.
The preliminary results show that, among the three types of the Chinese Earth observation data acquired in Guangzhou, BEIJING-1 data has achieved the highest classification accuracy, even though CBERS-02B multispectral data has a higher resolution of 19.5m compared to BEIJING-1 and HJ-1B data. The reason for the slightly higher classification accuracies of BEIJING-1 data over CBERS-02B data area result of fewer shadows of mountains and buildings. Clouds on the HJ-1B data were the main cause for HJ-1B’s lowest classification.
Left: HJ-1B multi-spectralSVM Classification Results in Shanghai;  Right: LST in Shanghai.

Figure 8: Left: HJ-1B multi-spectralSVM Classification Results in Shanghai; Right: LST in Shanghai.

In terms of classifiers, MLC and SVMs could achieve good results for CBERS-02B data, and Decision Trees and SVMs could achieve good results for BEIJING-1 data, while SVMs and MLC could achieve good results for HJ-1B data. The classification results using SVM and are shown in Figures 6 and 7, respectively.
The objective of this research was to map land use/land-cover using HJ-1B multi-spectral and hyperspectral images and to analyze the spatial relationship between LU/LC and land surface temperature (LST).The preliminary results (Figures 8and 9) show that the land cover map using HJ-1B multi-spectral and hyperspectral image achieved satisfactory accuracy. However, it is difficult to distinguish public green space and agricultural land cover due to their similar spectral reflectance.
left and right: HJ-1B hyperspectral SAM classification results in different region.

Figure 9: Left and right: HJ-1B hyperspectral SAM classification results in different region.

Therefore, future research is needed to improve the classification of these classes. By analyzing the relationship between multi-spectral land classification and land surface temperature, it is found that clouds and built-up area have the higher LST, with the lowest LST from non-use land, vegetation, and water. Based on Moran’s I index and semi-variance, the preliminary results indicate that a spatial pattern of homogeneous patches exists on scales smaller than 36km, meso-scales between 36 and 81km and large scales greater than 81km.
This research investigated multisensor and multitemporal spaceborne SAR data and Chinese Earth observation data for urban land cover mapping and change detection in China. The approach was developed in the framework of the Dragon II program and preliminary results are very promising. Effective classifications and change detection methods were developed and will be implemented in the final stage of the project, which has already started and will last one year, until the final workshop of Dragon II, scheduled for May 2012 in Beijing.
Monitoring China urban development is a big challenge, and Earth observation data are helping to understand fundamental interactions between urbanized areas and their surroundings. Change detection and land use maps are the first stage of correlating other environmental variables, such as land surface temperature, air quality and air pollution, to the characteristics of each urban area. The fusion of this information with environmental models can provide useful hints toward sustainable urban development.
[1] Ban, Y. and O. A. Yousif, 2010. MultitemporalSpaceborne SAR Data for Urbanization Monitoring in China: Preliminary Result. Proceedings, ESA Dragon II Midterm Symposium, Guilin, China, ESA SP-684.
[2] P. Gamba, M. Aldrighi, M. Stasolla, “Robust extraction of urban area extents in HR and VHR SAR images”, IEEE Journal on Selected Topics in Applied Earth Observation and Remote Sensing, Vol. 4, no. 1, pp. 27-34, March 2011
[3] CORINE Land Cover. Available on-line at
[4] Berkeley Image Segmentation. Available on-line at
[5] P. R. Marpu, M. Neubert, H. Herold, and I. Niemeyer, “Enhanced evaluation of image segmentation results”, Journal of Spatial Science, Vol. 55, No. 1, June 2010, 55–68.
This project is one of the projects within the Dragon 2 program, a collaboration between the European Space Agency (ESA) and the Chinese Ministry of Science and Technology (MOST). The authors would like thank ESA and MOST for providing the satellite data.