See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/248977308

Use of normalized difference built-up index in automatically mapping urban

areas from TM imagery

Article  in  International Journal of Remote Sensing · February 2003

DOI: 10.1080/01431160304987

CITATIONS READS

811 10,020

3 authors, including:

Yong Zha Jingqing Gao

Nanjing Normal University Zhengzhou University
56 PUBLICATIONS   2,327 CITATIONS    139 PUBLICATIONS   4,048 CITATIONS   

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Yong Zha on 04 April 2016.

The user has requested enhancement of the downloaded file.

. .  , 2003, . 24, . 3, 583–594

Use of normalized difference built-up index in automatically mapping

urban areas from TM imagery

Y. ZHA
College of Geographical Sciences, Nanjing Normal University, Nanjing 210097,
China; e-mail: yzha@njnu.edu.cn

J. GAO*
School of Geography and Environmental Science, University of Auckland,
Private Bag 92019, Auckland, New Zealand; e-mail: jg.gao@auckland.ac.nz

and S. NI
College of Geographical Sciences, Nanjing Normal University, Nanjing 210097

(Received 23 January 2001; in final form 3 October 2001)

Abstract. Remotely sensed imagery is ideally used to monitor and detect land
cover changes that occur frequently in urban and peri-urban areas as a con-
sequence of incessant urbanization. It is a lengthy process to convert satellite
imagery into land cover map using the existing methods of manual interpretation
and parametric image classification digitally. In this paper we propose a new
method based on Normalized Difference Built-up Index (NDBI) to automate the
process of mapping built-up areas. It takes advantage of the unique spectral
response of built-up areas and other land covers. Built-up areas are effectively
mapped through arithmetic manipulation of re-coded Normalized Difference
Vegetation Index (NDVI) and NDBI images derived from TM imagery. The
devised NDBI method was applied to map urban land in the city of Nanjing,
eastern China. The mapped results at an accuracy of 92.6% indicate that it
can be used to fulfil the mapping objective reliably. Compared with the max-
imum likelihood classification method, the proposed NDBI is able to serve as a
worthwhile alternative for quickly and objectively mapping built-up areas.

1. Introduction
Land covers in urban areas tend to change more drastically over a short
period of time than elsewhere because of incessant urbanization. Urbanization has
led land covers to change especially frequently in peri-urban areas in China as a
result of rapid economic development. These changes are ideally monitored and
detected from remotely sensed images as they are relatively up-to-date and give a
panoramic view.
Remote sensing materials in the form of aerial photographs and satellite images
are usually converted into useful information such as land cover maps using two

*Corresponding author.
International Journal of Remote Sensing
ISSN 0143-1161 print/ISSN 1366-5901 online © 2003 Taylor & Francis Ltd
http://www.tandf.co.uk/journals
DOI: 10.1080/01431160210144570
584 Y. Zha et al.

conventional methods: manual interpretation and computer-assisted digital pro-

cessing. During manual interpretation analogue photographs or satellite images are
visually interpreted and the results delineated directly on the photographs or images
or on tracing paper placed over them. Manual interpretation of Landsat 70 mm film
images could produce small-scale (e.g. 1:250 000) maps with an acceptable semantic
accuracy (89%), especially if a high level of generalization was intended (Lo 1981).
Visual interpretation of space shuttle Large Format Camera (LFC) photographs and
National High Altitude Photographs (NHAP) of Boston resulted in urban land use
and land cover types mapping 65% and 70%, respectively, from LFC and NHAP
photographs correctly at level III of the US Geological Survey (USGS) classification
scheme (Lo and Noble 1990).
Remotely sensed data have become increasingly available in a digital form,
allowing for their computer-assisted interpretation and processing. Irrespective of
the specific form of the remote sensing materials, manual interpretation is tedious,
time-consuming, and the interpreted results highly subjective to the image analyst.
By comparison, supervised classification is much faster and requires far less human
intervention. Lo (1981) found that a computer-assisted method of analysis of Landsat
data permits more detailed urban land use information to be extracted, but at an
accuracy of only 69%. A critical limitation with this method is that only the spectral
information of image pixels is taken advantage of for the classification. Therefore,
other image elements such as location, shape, shadow are ignored. Understandably,
the classification accuracy is rather low. As a matter of fact, automatic classification
of satellite images for urban areas is a difficult task to achieve at a high accuracy
level due to the diverse range of covers. Gao and Skillcorn (1998) achieved an overall
accuracy of 76.2% and 81.4% from winter and summer Système Probatoire de
l’Observation de la Terre (SPOT) sensor data, respectively, in generating detailed
land cover maps at the urban–rural fringe of Auckland, New Zealand, at level II
of the Anderson scheme. Treitz et al. (1992) achieved an overall Kappa coefficient
of 82.2% for training area data and 70.3% for the entire SPOT High Resolution
Visible (HRV) data of the rural–urban fringe of Toronto, Canada. In order to
improve the accuracy, Stuckens et al. (2000) used a hybrid segmentation procedure
to integrate contextual information. Overall accuracy of the optimal classification
technique was 91.4% for a level II classification (10 classes) with a K(e) of 90.5%.
The results automatically classified from satellite data, to some extent, are still
subject to the characteristics of the selected training samples. Their location, size
and representativeness all directly govern the reliability of the subsequently classified
results. The classification is slowed down by the requirement of selecting training
samples for all those covers to be mapped. Therefore, conventional methods of
parametric classification tend to be slow as a result of the need to select quality
training samples.
Considerable efforts have gone into simplifying the process of automatically
mapping land covers, such as using indices. One of the commonly used indices is
the Normalized Difference Vegetation Index (NDVI). This index takes advantage of
the unique shape of the reflectance curve of vegetation, and has been widely used
for mapping vegetation on the global scale from the Advanced Very High Resolution
Radiometer (AVHRR) data. For example, Achard and Estreguil (1995) used multi-
temporal AVHRR mosaics for tropical forest discrimination and mapping. Other
applications of NDVI include mapping of the surfaces affected by large forest
fires (Fernandez et al. 1997), assessment of the status of agricultural lands (Lenney
 Mapping urban areas using NDBI 585

et al. 1996), derivation of snow cover products (Slater et al. 1999), discrimination
of areas affected by volcanic eruption (Kerdiles and Diaz 1996) and so on. Multi-
temporal NDVI data were also classified to characterize phenological responses on
a spatially dissected landscape (Fleischmann and Walsh 1991) and to monitor
dynamic parameters of vegetation (Azzali and Menenti 2000).
In addition to NDVI, Normalized Difference Snow Index (NDSI) has been
devised from Landsat Thematic Mapper (TM) bands 2 and 5 to map glaciers (Sidjak
and Wheate 1999). This index is based on the difference between strong reflection
of visible radiation and near total absorption of middle infrared wavelengths by
snow (Hall et al. 1995). It is effective in distinguishing snow from similarly bright
soil, vegetation and rock, as well as from clouds (Dozier 1989).
Unlike NDSI, Normalized Difference Water Index (NDWI) has been developed
to delineate open water features and enhance their presence in remotely sensed
imagery based on reflected near-infrared radiation and visible green light. NDWI
may allow turbidity of waterbodies to be estimated from remotely sensed data
(McFeeters 1996). NDWI is sensitive to changes in liquid water content of vegetation
canopies. It is complementary to, but not a substitute for NDVI (Gao 1996).
In this study we propose a new and simple method for the rapid and accurate
mapping of urban areas. This method is based on the combinational use of NDBI
and NDVI. The mapping is accomplished through arithmetic manipulations and
recoding of NDBI and NDVI images derived from a 1997 TM image. This method
does not involve any subjective human intervention in the mapping process. The
effectiveness of this method was tested through the mapping of urban areas in the
Chinese City of Nanjing. Comparison of the results obtained using this method
with the manually interpreted ones demonstrates that it is highly reliable. This
method also produces very accurate results more efficiently than the supervised
classification method.

2. Study area
The study area is largely the urban area of Nanjing City, East China, located at
118° 47∞ E and 32° 04∞ N in the lower reaches of the Yangtze River (figure 1). Most
of the urban areas lie to the south of the river. Situated in the Yangtze River delta,
the city has a mostly gentle topography with little relief. Hilly areas are found in the
outskirts. The tallest mountain of Zijin is 448 m above sea level. Thus, little topo-
graphic shadow is present on the satellite imagery. Apart from the Yangtze River,
another major waterbody is a recreational lake next to the mountain.
As the capital of Jiangsu province, Nanjing has a total area of 976 km2. Land
covers present in this area are mainly urban residential, commercial and industrial.
Woodland is found in the adjacent mountainous areas. Natural vegetation occurs in
the form of mixed coniferous trees. Artificially planted trees are mostly deciduous.
In addition, there is farmland in the surrounding rural areas. Rice, wheat, vegetable
oil seed, as well as vegetables are cultivated throughout the year. Some crop fields
at the early stage of their growth may appear as barren on the TM image. Due to
rapid economic development, farmland adjacent to the urban periphery has been
converted to urban uses over the last two decades. Therefore, it is necessary from
time to time to monitor these changes from satellite imagery.

3. Data and processing

Landsat TM imagery was used in this study because of its finer spectral resolution
than other commonly used images such as SPOT and Multi-Spectral Scanner (MSS).
586 Y. Zha et al.

Figure 1. Location of the study area.

A full-scene TM imagery (5728 rows by 6920 columns) of 18 October 1997 was

acquired with all seven bands. The image quality was rather good with no cloud
cover over the study area. A topographic map of 1:50 000 was also acquired; pub-
lished in 1970, it has the Gausse-Krügel coordinate system with ground coordinates
indicated by 1 km grids.
All image processing and analyses were carried out in ER MapperB in the Windows
NT (version 6.0) environment. A sub-area of 1401 rows by 1408 columns was initially
subset from the raw image. False colour composites were formed using various band
combinations and displayed on the screen to differentiate diverse types of land covers.
In the end the standard composite of bands 2, 3 and 4 (figure 2) was selected. The
subset image was geometrically rectified using eight ground control points. They were
intersections of river channels, turns and intersections of roads with themselves and
with river channels. Their ground coordinates in the Gausse-Krügel coordinate system
were read from the topographic map. The residuals at these control points ranged
from 0.15–1.75 pixels. Once this accuracy was considered accurate, the original image
was projected linearly to the new system. During projection transformation the image
was resampled to the same spatial resolution using the nearest neighbour method.
Afterwards, the image, which did not conform to any regular orientation, was subset
once more to 800 rows by 800 columns (figure 2).

4. NDBI
Figure 2 is the standard false colour composite (TM4—red, TM3—green, TM2—
blue) on which various surface covers (e.g. built-up, woodland, farmland, barren and
water) are clearly distinguishable. Through repeatedly clicking on the representative
 Mapping urban areas using NDBI 587

Figure 2. False colour composite of TM bands 4 (red), 3 (green) and 2 (blue). Typical land
covers are urban (bluish yellow), woodland (dark red), barren (yellow), river (green),
lakes (dark), and farmland (orange). Image size: 800 by 800 pixels.

pixels of each of these covers, their values in all seven bands are averaged and
displayed graphically in figure 3. This profile illustrates that their spectral disparity
is the largest in bands 3, 4 and 5. An examination of the minimum, maximum and
standard deviation of each of the covers in the seven TM bands (table 1) confirms
the same conclusion. Namely, these values are most distinctive from one another for
each cover in bands 3, 4 and 5. Therefore, they are the most useful bands from which
some of the land covers may be potentially differentiated spectrally. Rivers and lakes
have a similar shape of profile. Their Digital Number (DN) value is markedly lower
in the fourth and fifth bands. They experience a sharp rise in reflectance in band 6,
but a low reflectance in band 7. The curve for rivers lies above that for lakes because
they are laden with more silt.
A close scrutiny of figure 3 reveals that except for barren, vegetation (woodland
and farmland) has a higher reflectance on band 4 than other covers. Moreover, its
value on band 4 still exceeds those on band 3. By comparison, all the non-vegetative
categories have a smaller DN on band 4 than 3. Therefore, the subtraction of band
3 from band 4 will result in positive DNs for vegetation pixels only. The afore-
mentioned relationships exist for the minimum and maximum DNs as well (table 1).
This outcome allows broad vegetative covers to be distinguished easily. This
processing is commonly referred to as NDVI (equation 1).
NDVI=(Band 4−band 3)/(band 4+band 3) (1)
In order to facilitate the subsequent processing, the derived NDVI image was
recoded with 254 for all pixels having positive indices (vegetation) and 0 for all
remaining pixels of negative indices (table 2).
588 Y. Zha et al.

Figure 3. Spectral profiles of six typical land covers in the study area.

Table 1. Minimum, maximum and standard deviation of DNs of the six covers in the seven
spectral TM bands.

TM spectral band 1 2 3 4 5 6 7

Built-up Minimum 79 33 35 25 29 125 15

Maximum 96 49 61 51 82 134 53
Standard deviation 4.1 2.9 4.7 4.4 9 1.6 6.6
Barren Minimum 82 46 70 66 103 132 65
Maximum 94 51 98 79 154 143 101
Standard deviation 4.3 1.8 11.8 5.6 22.9 4.4 15.7
Farmland Minimum 76 33 34 41 55 128 20
Maximum 91 43 56 79 103 131 57
Standard deviation 2.4 2.1 3.5 8.2 9.5 0.9 6
Woodland Minimum 67 26 25 38 41 124 13
Maximum 76 35 39 59 70 130 34
Standard deviation 1.7 1.4 2.1 3.5 6 1.2 3.4
River Minimum 85 42 54 28 8 119 2
Maximum 91 47 63 35 18 123 9
Standard deviation 1.4 1.1 1.6 1.5 1.6 0.6 1.4
Lake Minimum 71 29 27 17 9 118 2
Maximum 78 33 32 20 13 121 8
Standard deviation 1.6 0.8 1.1 0.8 1.2 0.8 1.3

Built-up areas and barren land experience a drastic increment in their reflectance
from band 4 to band 5 while vegetation has a slightly larger or smaller DN value
on band 5 than on band 4 (figure 3). This pace of increment greatly exceeds that of
any other covers. The minimum and maximum DNs in band 4 are much smaller
than those in band 5 for the same cover. The standardized differentiation of these
 Mapping urban areas using NDBI 589

Table 2. Pixel values of representative land covers after differencing and binary recoding.

Built-up Barren Woodland Farmland Rivers Lakes

NDVI 0 0 254 254 0 0

NDBI 254 254 254 or 0 254 or 0 0 0
NDBI-NDVI 254 254 0 or −254 0 or −254 0 0

two bands (equation 2) will result in close to 0 for woodland and farmland pixels,
negative for waterbodies, but positive values for built-up pixels, enabling the latter
to be separated from the remaining covers.
NDBI=(TM5−TM4)/(TM5+TM4) (2)
The derived NDBI image was then recoded to create a binary image. The
resultant ratio was assigned a new value of 0 if the input pixel had a negative index
or 254 if its input index was larger than 0 (table 2). The spectral profiles in figure 3
suggest that the ratio for vegetative covers can be larger or smaller than 0, depending
upon pixels in the surrounding environs. While many vegetative pixels may have
been coded 0 correctly in the output binary image, this handling cannot effectively
ensure that all vegetative pixels will receive the new value of 0. In order to avoid
mistakenly grouping those vegetative pixels into the built-up category, a further step
of processing is imperative. According to the results in table 2, subtraction of the
recoded NDVI image from the recoded NDBI image will lead to only built-up and
barren pixels having positive values while all other covers have a value of 0 or −254,
thus allowing built-up areas to be mapped automatically. Through three arithmetic
manipulations of TM bands 3, 4 and 5 followed by recoding, it is thus possible to
differentiate urban areas (including barren land). In order to enhance the appearance
of the final difference image, the derived urban built-up image was spatially filtered
using the median filter with a window of 5 pixels by 5 pixels (figure 4). The filtered
image was vectorized and later overlaid with the original colour composite to check
for its spatial accuracy.

5. Results
5.1. NDBI-derived result and its accuracy
Since this study concentrates on the mapping of built-up areas, all the mapped
land covers are categorized into only two groups, built-up and all others. There are
166 180 built-up pixels in the study area, or an area of 15 403.148 ha (each pixel is
30.445 m by 30.445 m in dimension). Built-up areas account for nearly 26% of the
entire study area. As implied previously, no further attempt was made to differentiate
the specific uses of these areas.
The accuracy of the mapped built-up areas was assessed both spatially and
aspatially. The spatial discrepancy between the mapped and actual boundaries of
built-up areas is illustrated graphically in figure 5. It shows that the vectorized
boundary of the mapped built-up areas matches closely with the actual border of
built-up areas in this part of the image. In order to provide a quantitative assessment
of the accuracy for the entire urban area, 68 pixels were randomly selected from the
mapped results (figure 4). Their genuine identity on the ground was verified under
the guidance of a global positioning system (GPS) receiver in the field according to
their coordinates. It was found that of these checkpoints, 63 were correctly mapped
590 Y. Zha et al.

Figure 4. Results of automatically mapped urban land use after spatial filtering with a
window size of 5 pixels by 5 pixels.

as urban areas, resulting in an accuracy of 92.6%. Of the five misclassified pixels,

three were fallow land, one denudated rock and one sandy beach.
Aspatially, the area of 15 403.148 ha derived from the proposed NDBI method
closely resembles the area of 14 300 ha that was obtained from manual measurement
of a 1995 topographic map at a scale of 1:10 000. These two figures differ from each
other by only 1103 ha, or less than 8% in relative terms. The apparent factor that
may explain the discrepancy is rapid urbanization. Since the photographs used to
generate the topographic map were taken, the urban area in Nanjing has expanded,
resulting in a larger built-up area on the 1997 satellite image. In this sense, the
mapped result is more current than the map-derived one.
The built-up areas have also been measured manually from the false colour
composite. The manual result of 15 061.958 ha is highly similar to the NDBI-derived
15 403.148 ha. The two sets of result have a discrepancy of 341 ha or 2%. This
discrepancy is explained by the fact that barren land is not included in the manual
result. Barren land in the form of beaches, mudflats, rocks (both denudated and
quarries), fallow land and transitional areas were not mapped as separate covers in
the NDBI method. Due to their spectral proximity to built-up areas, they have been
lumped together with built-up areas in this study. Since these covers make up a tiny
portion of the study area, such a handling will not degrade the mapping accuracy
of built-up areas considerably.

5.2. Comparison with maximum likelihood method

In order to assess the performance of the proposed NDBI method, the original
TM imagery with all seven bands was classified using the conventional parametric
 Mapping urban areas using NDBI 591

Figure 5. The agreement between built-up areas and the vectorized boundaries of filtered
built-up areas on the false colour composite (an enlargement). Refer to figure 2 for
the colour codes.

classification method (maximum likelihood) in which five classes of land covers

(woodland, built-up areas, lakes, rivers and farmland) were mapped. Since barren
land in the form of beaches, mudflats, rocks (both denudated and quarry) and fallow
land is so subordinate within the study area, it has been ignored in the mapping. In
total, the image was classified seven times. During each trial both the size of training
samples and their location were varied. The results (table 3) indicate that the built-
up areas mapped with the maximum likelihood method vary enormously from 10
to 319 320 pixels. The area of other covers fluctuates with each trial, as well. There
is no definite relationship between the size of training samples and the classified area
for a given cover category. The worst trial appears to be number 6 in which only
13 built-up pixels are classified, much less than the number of input pixel (112). No
pixels in the other three categories are classified at all. Such an utterly unrealistic
result is due probably to the poor quality of the training samples. These figures at
least confirm that the results derived from supervised classification are not quite
objective.
The classified built-up areas closest to their manually derived counterpart are at
195 519 pixels, or 18 123 ha. This result deviates from the manual one by 3061 ha. In
other trials the supervised method consistently overestimates built-up areas. Therefore,
the proposed NDBI method is superior to supervised classification. It may be argued
that the inferior performance of the maximum likelihood method is partially contrib-
uted by the fact that built-up areas were classified into one single category. Other
non-urban areas that share a similar spectral response with built-up areas, such as
592 Y. Zha et al.

Table 3. Results of maximum likelihood classification using various training sample sizes.
Samples were selected from the standard colour composite. Classification was carried
out using all seven bands.

Farmland Lakes Rivers Built-up Woodland

First trial Sample size 640 400 560 576 576

No. of pixels 284 659 3834 42 560 280 447 28 500
Hectare 26 385 355 3945 25 995 2642
Second trial Sample size 368 304 400 336 336
No. of pixels 267 574 2885 42 231 314 727 12 583
Hectare 24 801 267 3914 29 172 1166
Third trial Sample size 176 192 256 160 256
No. of pixels 262 276 3014 38 832 319 320 16 558
Hectare 24 310 279 3599 29 598 1535
Fourth trial Sample size 112 96 144 96 128
No. of pixels 0 7 86 708 10 553 275
Hectare 0 1 8037 1 51 282
Fifth trial Sample size 128 96 128 144 128
No. of pixels 63 604 7 14 384 195 519 366 486
Hectare 5895 1 1333 18 123 33 969
Sixth trial Sample size 80 112 80 112 128
No. of pixels 0 0 0 13 639 987
Hectare 0 0 0 1 59 320
Seventh trial Sample size 48 64 64 64 80
No. of pixels 283 296 4287 43 417 260 271 48 729
Hectare 26 259 397 4024 24 124 4517

Note: training samples are also varied in their locations.

transitional areas and beaches, were not classified as separate groups. Certainly,
the accuracy would be much higher had built-up areas first classified as subgroups
that were merged in a post-classification session. This, undoubtedly, will prolong the
already time-consuming process of supervised classification.

6. Conclusions and discussion

This proposed NDBI method is able to map built-up areas at an accuracy level
of 92.6%. The results mapped using NDBI are highly comparable to those from
manual interpretation in quantity. The two sets of results differ from each other by
2% and closely match each other spatially, as well. In comparison with supervised
classification, NDBI enables built-up areas to be mapped at a higher degree of
accuracy and objectivity. The absence of training samples from the mapping makes
subjective intervention from the human analyst redundant. This means that the same
results can be derived regardless of the analyst or how many times the mapping is
repeated. The redundancy also considerably expedites the mapping process that can
be accomplished by direct subtractions of original spectral bands. Through arithmetic
manipulation of TM bands and simple recoding of the intermediate images, NDBI
does not require complex mathematical computation. It is concluded that the pro-
posed NDBI is much more effective and advantageous in mapping general built-up
areas than the maximum likelihood method. It can serve as a worthwhile alternative
for quickly mapping urban land.
The assumption underlying the NDBI method is the spectral reflectance of urban
areas in TM5 exceeding that in TM4. This method will generate valid results so
 Mapping urban areas using NDBI 593

long as this assumption is not violated. Because the reflectance of urban areas
exhibits little seasonality, this method is not prone to its impact. However, its
performance may be adversely affected indirectly by the presence of other covers
whose reflectance is seasonal, such as forest. This problem may be overcome with
the selection of images recorded when defoliation is minimal or non-existent. The
mixture of built-up areas with barren farmland may be overcome with the use of an
image taken when vegetative cover is at its maximum.
Nevertheless, this proposed method does have a number of limitations. First of
all, it can map only broad urban land covers. For instance, urban industrial, commer-
cial and residential areas are impossible to be separated. This, however, may not
prove to be a liability in most Chinese cities where they are highly intermixed
spatially. They are difficult to be satisfactorily mapped even using the conventional
supervised classification method anyway. Secondly, the NDBI method is unable to
separate urban areas from barren (e.g. sandy beaches) because both of them have a
similar spectral response in all TM bands. This limitation may be overcome with
the use of spatial knowledge, as the latter is located next to water. Another remedial
method is to select an image recorded in a season when the water level is so high
that sand beaches are submerged under water. By comparison, the effect of drought
on the performance of the NDBI method is more difficult to counteract. The loss of
moisture from soil and disappearance of vegetation as a result of drought will
assimilate the spectral characteristics of both urban and agricultural areas, debilitat-
ing the validity of the proposed method. Predictably, the reliability of this method
is lowered in mapping peripheral urban areas where barren or fallow land is wide-
spread. Thirdly, the universality of this proposed method needs to be tested in other
geographic areas. The success of the proposed method lies in the NDVI value of
vegetation being larger than 0. However, the spectral response of vegetation varies
from location to location due to different kinds of species and nature of underlying
soil and moisture conditions. Besides, the response pattern for vegetation varies with
its density. Under these circumstances it is uncertain whether vegetative NDVI value
still exceeds 0. It is speculated that the specific reflectance values may vary with
these conditions, but the general pattern of the spectral response of vegetation in all
seven TM bands will remain identical, ensuring a positive NDVI value and thus
maintaining the validity of the method.

Acknowledgments
The comments made by two anonymous referees helped to improve the quality
of the manuscript considerably.

References
A, F., and E, C., 1995, Forest classification of Southeast Asia using NOAA
AVHRR data. Remote Sensing of Environment, 54, 198–208.
A, S., and M, M., 2000, Mapping vegetation–soil–climate complexes in southern
Africa using temporal Fourier analysis of NOAA AVHRR NDVI data. International
Journal of Remote Sensing, 21, 973–996.
D, J., 1989, Spectral signature of Alpine snow cover from the Landsat Thematic Mapper.
Remote Sensing of Environment, 28, 9–22.
F, A., I, P., and C, J. L., 1997, Automatic mapping of surfaces affected
by forest fires in Spain using AVHRR NDVI composite image data. Remote Sensing
of Environment, 60, 153–162.
F, C. G., and W, S. J., 1991, Multi-temporal AVHRR digital data: an approach
for landcover mapping of heterogeneous landscapes. Geocarto International, 6, 5–20.
 594 Mapping urban areas using NDBI

G, B. C., 1996, NDWI—a normalized difference water index for remote sensing of vegetation
liquid water from space. Remote Sensing of Environment, 58, 257–266.
G, J., and S, D., 1998, Capability of SPOT XS data in producing detailed land
cover maps at the urban–rural periphery. International Journal of Remote Sensing, 19,
2877–2891.
H, D. K., R, G. A., and S, V. V., 1995, Development of methods for
mapping global snow cover using Moderate Resolution Imaging Spectroradiometer
data. Remote Sensing of Environment, 54, 127–140.
K, H., and D, R., 1996, Mapping of the volcanic ashes from the 1991 Hudson
eruption using NOAA–AVHRR data. International Journal of Remote Sensing, 17,
1981–1995.
L, M. P., W, C. E., C, J. B., and H, H., 1996, The status of
agricultural lands in Egypt: the use of multitemporal NDVI features derived from
Landsat TM. Remote Sensing of Environment, 56, 8–20.
L, C. P., 1981, Land use mapping of Hong Kong from Landsat images. An evaluation.
International Journal of Remote Sensing, 2, 231–252.
L, C. P., and N, W. E. Jr, 1990, Detailed urban land-use and land-cover mapping using
Large Format Camera photographs: an evaluation. Photogrammetric Engineering and
Remote Sensing, 56, 197–206.
MF, S. K., 1996, The use of the Normalized Difference Water Index (NDWI) in the
delineation of open water features. International Journal of Remote Sensing, 17,
1425–1432.
S, R. W., and W, R. D., 1999, Glacier mapping of the Illecillewaet icefield, British
Columbia, Canada, using Landsat TM and digital elevation data. International Journal
of Remote Sensing, 20, 273–284.
S, M. T., S, D. R., R, W. G., and S, A., 1999, Potential operational
multi-satellite sensor mapping of snow cover in maritime sub-polar regions.
International Journal of Remote Sensing, 20, 3019–3030.
S, J., C, P. R., and B, M. E., 2000, Integrating contextual information with
per-pixel classification for improved land cover classification. Remote Sensing of
Environment, 71, 282–296.
T, P. M., H, P. J., and G, P., 1992, Application of satellite and GIS technolo-
gies for land-cover and land-use mapping at the rural–urban fringe: a case study.
Photogrammetric Engineering and Remote Sensing, 58, 439–448.

View publication stats