PLEASE SCROLL DOWN FOR ARTICLE
This article was downloaded by:
On: 16 November 2010
Access details: Access Details: Free Access
Publisher Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House, 37-
41 Mortimer Street, London W1T 3JH, UK
Remote Sensing Letters
Publication details, including instructions for authors and subscription information:
http://www.informaworld.com/smpp/title~content=t915281289
Improving the normalized difference built-up index to map urban built-up
areas using a semiautomatic segmentation approach
Chunyang Heab
; Peijun Shia
; Dingyong Xieb
; Yuanyuan Zhaoab
a
State Key Laboratory of Earth Surface Processes and Resource Ecology, Beijing Normal University,
Beijing, China b
College of Resources Science and Technology, Beijing Normal University, Beijing,
China
First published on: 26 April 2010
To cite this Article He, Chunyang , Shi, Peijun , Xie, Dingyong and Zhao, Yuanyuan(2010) 'Improving the normalized
difference built-up index to map urban built-up areas using a semiautomatic segmentation approach', Remote Sensing
Letters, 1: 4, 213 — 221, First published on: 26 April 2010 (iFirst)
To link to this Article: DOI: 10.1080/01431161.2010.481681
URL: http://dx.doi.org/10.1080/01431161.2010.481681
Full terms and conditions of use: http://www.informaworld.com/terms-and-conditions-of-access.pdf
This article may be used for research, teaching and private study purposes. Any substantial or
systematic reproduction, re-distribution, re-selling, loan or sub-licensing, systematic supply or
distribution in any form to anyone is expressly forbidden.
The publisher does not give any warranty express or implied or make any representation that the contents
will be complete or accurate or up to date. The accuracy of any instructions, formulae and drug doses
should be independently verified with primary sources. The publisher shall not be liable for any loss,
actions, claims, proceedings, demand or costs or damages whatsoever or howsoever caused arising directly
or indirectly in connection with or arising out of the use of this material.
Improving the normalized difference built-up index to map urban built-up
areas using a semiautomatic segmentation approach
CHUNYANG HE*†‡, PEIJUN SHI†, DINGYONG XIE‡ and
YUANYUAN ZHAO†‡
†State Key Laboratory of Earth Surface Processes and Resource Ecology, Beijing Normal
University, Beijing 100875, China
‡College of Resources Science and Technology, Beijing Normal University, Beijing
100875, China
(Received 20 December 2009; in final form 26 March 2010)
Remote sensing images are useful for monitoring the spatial distribution and growth
of urban built-up areas because they can provide timely and synoptic views of urban
land cover. Although the normalized difference built-up index (NDBI) is useful to
map urban built-up areas, it still has some limitations. This study sought to improve
the NDBI by using a semiautomatic segmentation approach. The proposed approach
had more than 20% higher overall accuracy than the original method when both were
implemented simultaneously at the National Olympic Park (NOP), Beijing, China.
One reason for the improvement is that the proposed NDBI approach separates
urban areas from barren and bare land to some extent. More importantly, the
proposed method eliminates the original assumption that a positive NDBI value
should indicate built-up areas and a positive normalized difference vegetation index
(NDVI) value should indicate vegetation. The new method has improved universality
and lower commission error compared with the original method.
1. Introduction
Urban land accounts for a small fraction of the Earth’s surface area but has a
disproportionate influence on its surroundings in terms of mass, energy and resource
fluxes (Lambin and Geist 2001). Mapping urban land in a timely and accurate manner
is indispensible for watershed run-off prediction and other planning applications
(Small 2003). Remote sensing images are useful for monitoring the spatial distribution
and growth of urban built-up areas because of their ability to provide timely and
synoptic views of land cover (Guindon et al. 2004, Xu 2008, Bhatta 2009, Griffiths
et al. 2010). Over the past two decades, researchers have become increasingly interested
in using remotely sensed imagery to address urban and suburban problems
(Jacquin et al. 2008). A number of techniques for automatically mapping urban land
cover using satellite imagery have been formulated, applied and evaluated. These
techniques can be broadly grouped into two general types: (1) those based on the
classification of the input data, including pixel- and object-based classifications
(Guindon et al. 2004, Cleve et al. 2008) and (2) those based on directly segmenting
the indices, such as the commonly used normalized difference vegetation index
(NDVI) (Zha et al. 2003, Zhang et al. 2005). However, the spatial and spectral
*Corresponding author. Email: hcy@bnu.edu.cn
Remote Sensing Letters
ISSN 2150-704X print/ISSN 2150-7058 online # 2010 Taylor & Francis
http://www.tandf.co.uk/journals
DOI: 10.1080/01431161.2010.481681
Remote Sensing Letters
Vol. 1, No. 4, December 2010, 213–221
DownloadedAt:05:4016November2010
variabilities of urban environments present fundamental challenges to deriving accurate
remote sensing-based products for urban areas (Powell et al. 2007). Efforts to
improve automatic mapping of urban land use using satellite imagery are thus still
worthwhile.
Zha et al. (2003) proposed the normalized difference built-up index (NDBI) to
automatically map urban built-up areas. The method takes advantage of the unique
spectral responses of built-up areas and other land covers. Built-up areas are effectively
mapped through the arithmetic manipulation of recoded NDVI and NDBI
images derived from Landsat Thematic Mapper (TM) imagery. However, the
approach proposed by Zha et al. (2003) recodes the derived NDBI and NDVI images
to create binary images under the assumption that a positive value of NDBI should
indicate built-up areas and a positive value of NDVI should indicate vegetation. As
discussed by Zha et al. (2003), with this recoding process, their approach is unable to
separate urban areas from barren and bare land. They suggested that the universality
of the approach needed to be tested in other geographic areas because of the actual
complicated spectral response patterns of vegetation.
In the light of the original NDBI approach’s advantages and drawbacks, this study
tried to improve it using a semiautomatic segmentation approach. The proposed
NDBI approach and the original NDBI approach were also simultaneously applied
to the National Olympic Park (NOP), Beijing, and the results were evaluated and
compared with those of Zha et al. (2003).
2. Method
2.1 Original NDBI approach
Based on the analysis of the unique spectral responses of built-up areas and other land
covers in seven Landsat TM bands, the original NDBI approach developed by Zha
et al. (2003) was implemented as three arithmetic manipulations of TM bands 3–5,
followed by recoding. First, a continuous NDVI image NDVIc was obtained by the
following equation:
NDVIc ¼
band4 À band3
band4 þ band3
(1)
To facilitate the subsequent process, the derived NDVIc was recoded to create a
binary NDVI image NDVIb with 245 for all pixels having positive indices (vegetation)
and 0 for all remaining pixels of negative indices.
Second, the continuous NDBI image NDBIc was produced by the following
equation:
NDBIc ¼
band5 À band4
band5 þ band4
(2)
The derived NDBIc was then recoded to create a binary NDBI image NDBIb with 0
for those pixels having a negative value or 254 for those having a positive value.
Third, the built-up areas were extracted by the following equation:
BUb ¼ NDBIb À NDVIb (3)
where BUb is the resultant binary image with only the built-up and barren pixels
having positive value, thus allowing built-up areas to be mapped automatically (Zha
et al. 2003).
214 C. He et al.
DownloadedAt:05:4016November2010
2.2 Proposed NDBI approach
The proposed NDBI approach was implemented as follows: first, the continuous
NDVI image NDVIc and the continuous NDBI image NDBIc were directly derived
from the Landsat TM image following equations (1) and (2), respectively.
Second, the continuous image BUc was produced by the following equation:
BUc ¼ NDBIc À NDVIc (4)
Unlike the binary image BUd obtained with the original NDBI approach, BUc is one
continuous image. The greater the value of a pixel in BUc is, the higher is the
possibility of the pixel being a built-up area.
Third, using one optimal threshold value, the continuous image BUc was segmented
into one binary image with the built-up area being 1 and other areas being 0.
Often, the optimal threshold value to segment one continuous image is determined
according to the empirical strategies or from manual trial-and-error procedures.
These procedures usually require a more experienced image analyst and a long trial
time (Chen et al. 2003). Chen et al. (2003) presented the double-window flexible pace
search (DFPS) approach to determine the optimal threshold value between change
and non-change pixels when detecting land use/cover change. This semiautomatic
method only requires the involvement of an image analyst during the selection of
typical training samples and thus has an advantage of effectively determining the
optimal threshold value to segment one continuous image to detect land use/cover
change (Chen et al. 2003). Using the idea of the DFPS, we developed a semiautomatic
approach to determine the optimal threshold value to segment BUc to extract built-up
areas. The basic idea of the semiautomatic approach is to select a threshold value from
the training samples, assuming that the threshold value leading to the maximum
accuracy in extracting the built-up area within the training samples is also optimal
for the entire BUc. The process mainly involves the following four steps.
First, some typical built-up areas are chosen as the sample areas by visual interpretation
of the image. The criteria for selecting sample areas are (1) training samples
should include only built-up pixels and (2) training samples should be ‘islands’
encircled by non-built-up pixels.
Second, the search range and pace are determined according to the analysis of the
histogram of BUc. The search range can be set as a difference between the minimum
value a and the maximum value b of BUc. The first search pace (increment) P1 may be
calculated according to the following formula:
P1 ¼
b À a
m
(5)
where m is a positive integer that determines the number of potential thresholds in a
search process and can be set manually. The potential thresholds to detect built-up
pixels from the training samples in the search process are given within the range of (a,
b) as b – P1, b – 2P1, . . .. It should be noted that the size of the manually set m does not
affect the search efficiency and the final results. A large m increases the number of
potential thresholds during one search, but it decreases the number of searches.
Third, the success rate of built-up-area extraction is defined to evaluate the performance
of each potential threshold value during one search process for identifying
built-up/non-built-up pixels. The success rate Lk is calculated for a potential threshold
value of k as
Improving the normalized difference built-up index 215
DownloadedAt:05:4016November2010
Lk ¼
Ak1 À Ak2ð Þ
A
 100% (6)
where Ak1 is the number of built-up pixels detected inside all training patches, Ak2 is
the number of built-up pixels that are detected incorrectly and A is the total number of
pixels within all training patches. After all the Lk for all the m thresholds in one search
are obtained, the maximum and minimum values of Lk can be found and designated
as Lmax and Lmin during this search process. If the two parameters do not satisfy the
exit conditions described in the fourth step, a new search begins. The new search range
is set in the range (kmax - P1, kmax þ P1), and a new smaller search pace is set based on
the modified search range with equation (5). Here, kmax is the potential threshold
value corresponding to Lmax in the search process.
Finally, the second and third steps involve an iterative process, which will be
terminated when the following equation is satisfied:
Lmax À Lmin  (7)
where Lmax and Lmin are the maximum and minimum values of the success rate in one
search process, and  is an acceptable error constant. The threshold value corresponding
to Lmax is considered to be an optimal threshold value to extract the built-up areas
from BUc when the change of search pace has little influence on the result of built-up
area extraction.
3. Case study: mapping built-up areas in the National Olympic Park (NOP), Beijing
3.1 Study area and data
The NOP, located in northern Beijing, was selected as the study area to evaluate the
performance of the proposed approach (figure 1). The NOP covers 12.15 km2
.
Massive construction began at the NOP in 2000 in preparation for the 2008
Olympic Games in Beijing. A Level 1R Landsat Enhanced Thematic Mapper
(ETMþ) image acquired on 19 May 2001 was used as the test image for radiometric
correction. This image had high quality and fully covered the study area. Some
auxiliary data were collected, including a 1:50,000-scale topographic map for 1972
and a 1:100,000-scale land use map for 2000. In addition, one IKONOS image
acquired on 26 April 2001 was also used to assist the accuracy assessment because
of its fine resolution compared with the Landsat ETMþ image.
3.2 Image preprocessing
High-precision geometric registration is the basic requirement of image preprocessing
(Guindon et al. 2004). First, the Landsat ETMþ image was geometrically corrected
according to the Universal Transverse Mercator projection at 30 Â 30-m resolution,
using second-order polynomial and bilinear interpolation. Forty ground control
points were collected from the 1:50,000 topographic map. The root mean square
error was less than 1 pixel. Then, the boundary of the NOP, digitized into a GIS
database, was used to clip the study area from images (figure 2).
3.3 Mapping built-up area
Following equations (1), (2) and (4), the continuous NDVI image NDVIc, the continuous
NDBI image NDBIc and the continuous image BUc were directly derived
216 C. He et al.
DownloadedAt:05:4016November2010
from the Landsat ETMþ image. To assist the process of determining the optimal
threshold value, NDVIc, NDBIc and BUc were first normalized into (0, 255) with the
following equation (figure 3):
N ¼
T À Tminð Þ
Tmax À Tminð Þ
 255 (8)
where N is the pixel value of the normalized image, T is the pixel value of the inputted
image and Tmax and Tmin represent the maximum and minimum pixel values of the
inputted image, respectively.
From the analysis of the histogram of BUc, the search range was set (0, 226) with the
initial pace of 20. The semiautomatic segmentation approach was then used to
determine the optimal threshold value as 102; pixels with values over 102 were
extracted from BUc as built-up areas (figure 4(c)).
3.4 Accuracy assessment
To evaluate the performance of the proposed NDBI method, a stratified random
sampling method was used for accuracy assessment with 80 samples each of built-up
areas and of non-built-up areas (Congalton 1991). The results showed that the total
Figure 1. The study area.
Improving the normalized difference built-up index 217
DownloadedAt:05:4016November2010
accuracy of the proposed NDBI approach was 86.30% (table 1). In addition, the
original NDBI approach was also used to produce binary built-up-area imagery
(figure 4(b)), and the same accuracy assessment was implemented to compare the
accuracies of the two approaches. The results showed that the overall accuracy of the
original NDBI approach was 64.38%. The proposed approach improved the overall
accuracy by over 20%. Interpretation of the original imagery (figure 4(a)) and the
resultant images (figure 4(b) and (c)) showed that the original approach had obvious
commission error, showing some vegetation as built-up area (see the area in the yellow
rectangle in figures 4(a)–(c)). The proposed approach separated urban areas from
barren and bare land to some extent and reduced the commission error of the original
approach.
4. Conclusion and discussion
The original NDBI approach is an effective method for automatically mapping urban
built-up areas using the Landsat TM imagery, offering easy operation and independence
of sample selection by manual operation. However, this approach also has
some limitations associated with recoding the derived NDBI imagery and NDVI
imagery to create a binary image and assuming that positive NDBI value should
indicate built-up areas and positive NDVI value should indicate vegetation. Because
Figure 2. The used Landsat ETMþ image of the study area. Note: The image is a standard
false colour composite: band 4, red; band 3, green; and band 2, blue.
218 C. He et al.
DownloadedAt:05:4016November2010
of this recoding process, the original approach was unable to separate urban areas
from barren and bare land. In this study, the original approach also suffered from
commission error, showing some vegetation as built-up area because of the complicated
spectral response patterns of vegetation.
In this letter, the original NDBI approach was improved by directly segmenting
the resultant continuous image with an optimal threshold value determined by a
semiautomatic segmentation approach based on the idea of the DFPS. In this case
study, our proposed approach had over 20% higher overall accuracy than the
original approach when implemented simultaneously for the NOP, Beijing. One
reason for the improvement is that the proposed approach separated urban areas
from barren and bare land to some extent. More importantly, however, the proposed
approach abolishes the assumption that a positive NDBI value should indicate
a built-up area whereas a positive NDVI value should indicate vegetation. The
proposed approach improves the universality and reduces the commission error of
the original NDBI approach.
Figure 3. The obtained continuous NDVI, NDBI and built-up-area images of the study area.
Note: (a) is the continuous NDVI image, (b) is the continuous NDBI image and (c) is the
continuous built-up area image.
Improving the normalized difference built-up index 219
DownloadedAt:05:4016November2010
However, like the original NDBI approach, the proposed approach also has some
limitations because it is a semiautomatic approach depending on training sample
selection by manual operation. It, therefore, relies to some degree on the experience
and skill of the image analyst. In addition, the proposed approach cannot entirely
separate bare land and built-up areas because of their similar spectral character in
Landsat ETMþ imagery. Users should consider the trade-offs between efficiency and
accuracy when choosing between the original and proposed NDBI approaches.
Figure 4. Comparison of the results by the original and proposed approaches. Note: (a) is the
Landsat ETMþ image used in this study, (b) is the resultant binary image of the built-up and nonbuilt-up
areas by the original approach and (c) is the resultant binary image of the built-up and nonbuilt-up
areas by the proposed approach; the area in the yellow rectangle in each panel is where the
original approach produced obvious commission error, giving vegetation as a built-up area.
Table 1. Accuracy assessment of the resultant images.
Overall accuracy
(%)
Omission error
(%)
Commission error
(%) k coefficient
NDBI approach 64.38 3.45 47.17 0.35
Improved NDBI
approach
86.30 0 18.87 0.70
220 C. He et al.
DownloadedAt:05:4016November2010
Acknowledgements
This research has been supported by the Natural Scientific Foundation of China
(Grant No. 40971059) and the State Key Laboratory of Earth Surface Processes and
Resource Ecology, Beijing Normal University, China (Grant No. 2008-ZY-06).
References
BHATTA, B., 2009, Analysis of urban growth pattern using remote sensing and GIS: a case study
of Kolkata, India. International Journal of Remote Sensing, 30, pp. 4733–4746.
CHEN, J., GONG, P., HE, C., PU, R. and SHI, P., 2003, Land use/cover change detection using
improved change vector analysis. Photogrammetric Engineering and Remote Sensing,
69, pp. 369–379.
CLEVE, C., KELLY, M., KEARNS, F. and MORITZ, M., 2008, Classification of the wildland–urban
interface: a comparison of pixel- and object-based classifications using high-resolution
aerial photography. Computers, Environment and Urban Systems, 32, pp. 317–326.
CONGALTON, R., 1991, A review of assessing the accuracy of classifications of remotely sensed
data. Remote Sensing of Environment, 37, pp. 35–46.
GRIFFITHS, P., HOSTERT, P., GRUEBNER, O. and VAN DER LINDEN, S., 2010, Mapping megacity
growth with multi-sensor data. Remote Sensing of Environment, 114, pp. 426–439.
GUINDON, B., ZHANG, Y. and DILLABAUGH, C., 2004, Landsat urban mapping based on a
combined spectral–spatial methodology. Remote Sensing of Environment, 92, pp.
218–232.
JACQUIN, A., MISAOVA, L. and GAY, M., 2008, A hybrid object-based classification approach for
mapping urban sprawl in periurban environment. Landscape and Urban Planning, 84,
pp. 152–165.
LAMBIN, E.F. and GEIST, H., 2001, Global land use and land cover change: what have we learned
so far? Global Change News Letter, 46, pp. 27–30.
POWELL, R., ROBERTS, D., DENNISON, P. and HESS, L., 2007, Sub-pixel mapping of urban land
cover using multiple endmember spectral mixture analysis: Manaus, Brazil. Remote
Sensing of Environment, 106, pp. 253–267.
SMALL, C., 2003, High spatial resolution spectral mixture analysis of urban reflectance. Remote
Sensing of Environment, 88, pp. 170–186.
XU, H., 2008, A new index for delineating built-up land features in satellite imagery.
International Journal of Remote Sensing, 29, pp. 4269–4276.
ZHA, Y., GAO, Y. and NI, S., 2003, Use of normalized difference built-up index in automatically
mapping urban areas from TM imagery. International Journal of Remote Sensing, 24,
pp. 583–594.
ZHANG, Q., PAVLIC, G., CHEN, W., FRASER, R., LEBLANC, S. and CIHLAR, J., 2005, A semiautomatic
segmentation procedure for feature extraction in remotely sensed imagery.
Computers and Geosciences, 31, pp. 289–296.
Improving the normalized difference built-up index 221
DownloadedAt:05:4016November2010