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

net/publication/267822241

Radar and Optical Data Integration for Land-Use/Land-Cover Mapping

Article  in  Photogrammetric Engineering and Remote Sensing · June 2000

CITATIONS READS

43 600

3 authors, including:

Barry Haack
George Mason University
78 PUBLICATIONS   1,837 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Barry Haack on 24 November 2015.

The user has requested enhancement of the downloaded file.

 Radar and Optical Data Integration for
Land-Use/Land-Cover Mapping
Barry N. Haack, Nathaniel D. Herold, and Matthew A. Bechdol

Abstract Agency ERS, the Japanese JERS, and the Canadian RADARSAT sat-
This study evaluated the advantages of combining traditional ellites. The research and applications remote sensing commu-
spaceborne optical data from the visible and infrared wave- nities are still in the early stages of understanding the character-
lengths with the longer wavelengths of radar. East African istics and uses of spaceborne radar. Some applications, includ-
landscapes, including areas of settlements, natural vegetation, ing tropical deforestation and sea ice monitoring, are routinely
and agriculture, were examined. For three study sites, being used.
multisensor data sets were digitally integrated with training The purpose of this study was to evaluate multisensor data
data and ground-truth information derived from field visits. sets of spaceborne multispectral optical data and spaceborne
The primary methodology was standard image processing, radar for the delineation of land uselland cover for sites in East
including spectral signature extraction and the application of Africa. The primary intentions were to ascertain what im-
a statistical decision rule to classify the surface features. The provements may be obtained using a multisensor data set and
relative accuracy of the classifications was established b y how to improve the utility of radar data. East Africa provides a
comparison to ground-truth information. In all sites, the merger range of surface characteristics that are representative of many
of optical and radar sensors improved the ability to map regions of the Earth. Mapping procedures established there
surface features over either sensor independently, although should be spatially extendible.
different manipulations of the radar data were necessary to
obtain the most useful results. Those manipulations included Radar Background
measures of texture, spatial filtering, and despeckling prior to Radar has special properties that make it a viable alternative
texture extraction. and/or partner to traditional optical remote sensing techniques
(Foody, 1988). For instance, microwave energy is capable of
Introduction penetrating atmospheric conditions that render traditional
A basic problem in economic planning, environmental studies, spaceborne optical and multispectral systems useless (Elachi,
or resource management is obtaining and maintaining current, 1988).Radar therefore has the ability to image through rain, fog,
accurate information. The need for basic surface characteristic hail, smoke, and, most importantly, clouds. These characteris-
information, such as land use and land cover, is critical to both tics hold enormous data-collection potential in many coun-
scientific analysis and decision making activities. Without tries around the world, especially those areas such as Central
accurate information, scientists cannot complete valid studies Africa that experience persistent cloud cover. An additional
and decision-makers often fail to make correct decisions. One advantage of radar is that the feature interaction is a function of
significant method for providing current, reliable surface geometry, texture, and dielectric constant which is different
information is satellite remote sensing. Spaceborne remotely from the reflectance interactions of optical systems (Forster,
sensed data may be particularly useful in developing countries 1996).These different interactions have the potential to pro-
such as in Africa (Morain, 1991). vide information beyond that of optical data.
The use of remote sensing for resource assessment and Remote sensing research on radar exists in two primary
basic mapping has a considerable history and increasing use in categories. The first includes methods to improve the capabil-
Africa. There have been regional and national centers for re- ity of radar as an independent sensor. A difficulty with analysis
mote sensing established in many countries. Some of these of radar data as an independent sensor, particularly in auto-
have existed for over 20 years, even prior to the availability of mated classification, is that most spaceborne systems only col-
spaceborne imagery with the launch of Landsat in 1972. A lect data at a single wavelength with a fixed polarization. This
number of these centers were established under international prohibits many classifications. There are several options to pro-
assistance programs directed at remote sensing technology vide more bands for digital classification from single-band
transfer. The United States Agency for International Develop- radar, Those options typically include the use of multitemporal
ment was quite active with these centers, including the estab- data sets or the extraction of texture as unique bands (Luckman
lishment of regional centers in Nairobi, Kenya and Ouaga- et al., 1997; Prasad and Gupta, 1998). Other manipulations of
dougou, Bukina Faso (Paul and Mascarenhas, 1981). radar to improve information extraction include spatial filter-
One of the more promising recent achievements in remote ing such as pre- or post-classification smoothing and despeck-
sensing has been the operational capability to collect radar ling (Durand et al., 1987).
data from space, such as has been established by the United
States Shuttle Imaging Radar (SIR) missions, the European Space

Photogrammetric Engineering & Remote Sensing

B.N. Haack and N.D. Herold are with the Department of Geogra-
phy and Earth Sciences, MSN 1E2, George Mason University, Vol. 66, No. 6, June 2000, pp. 709-716.
Fairfax, VA 22030 ([email protected].). 0099-1112/00/6606-709$3.00/0
M.A. Bechdol is with Global Science and Technology, 6411 0 2000 American Society for Photogrammetry
Ivy Lane, Greenbelt, MD 20770. and Remote Sensing

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

 The second area for research with spaceborne radar is to information was then converted to a raster-based GIS format and
evaluate its relative value and complementarity with more tra- registered to the TM/SIR-B data set. There were four categories
ditional visible and infrared wavelength sensors. In some situa- of basic land uselland cover incorporated into the GIs layer.
tions it is clear that one system has advantages over the other. In Those categories were urban, agricultural, natural vegetation,
other situations, the integrated use of data from these discrete and backgroundlother. The backgroundlother category was pri-
portions of the electromagnetic spectrum may be advanta- marily the extensive, flat areas of bare soil, which produced a
geous (Welch and Ehlers, 1988; Harris et al., 1990; Raghava- low radar backscatter return. Housing in the area is made of
wamy et al., 1996; Pohl and Van Genderen, 1997). Sensor indigenous materials, primarily clay, which is spectrally simi-
integration may provide a clearer image, may reduce redun- lar to the surrounding bare soil.
dancy of optical bands, and may improve classification Radar data have been effectively used in locating cultural
accuracies. targets, because man-made features provide a large return due
Paris and Kwong (1988) used a combination of SIR-B and to the high dielectric properties of some construction materials
Landsat Thematic Mapper (TM) data to provide quantitative (such as metal) and the geometric shape of many cultural sur-
information on the amounts of herbaceous and woody vegeta- faces. Buildings often act as corner reflectors to the radar signal
tion. They found these data types to be complementary, with and thus have a high return, but radar can also produce a high
the optical TM providing information on green biomass and the backscatter from vegetation. This is due to the texture of the
radar on woody biomass. The synergism of airborne radar and vegetative canopy and the dielectric properties of leaf moisture
Landsat Thematic Mapper (TM) was examined by Brisco and [Richards. 1990:Dobson et al.. 1992: Ranson et al.. 19951. This
Brown (1995) for agricultural fields using both single-date and similarly high radar backscattkr between villages and vegeta-
multi-date data. They obtained the best classification accuracy, tion can be observed in the SIR-Bimage of the Wad Medani area
92 percent, by a multitemporal combination of the two data (Figure 1). This image is oriented such that north is to the
types. upper left and the Blue Nile is very apparent. The city of Wad
In automated classification with multispectral data, often Medani is the large, very high return area at the extreme right
the best band combinations are those with at least one band within a major bend of the Blue Nile.
from each available portion of the electromagnetic spectrum. The agricultural areas are obvious along the lower left por-
By merging optical and radar data, an additional portion of the tion of the scene, due to their rectilinear patterns. These areas
spectrum is available that may improve classification. Radar, include both high and low returns. The high-return agricultural
because it responds more to the structure of surface features features are active fields, generally cotton or sugarcane, while
rather than to their internal characteristics and reflectivity, is the low-return fields are fallow. These fallow fields act as spec-
potentially an interesting and useful addition to optical data. ular features and reflect the radar energy away from the sensor.
The villages and urban areas are not identifiable on the TM
Study Sites and Data image because much of the construction material (vegetation,
A number of sites in East Africa were selected in order to exam- bricks, and adobe) spectrally blends into the surrounding land-
ine different climatic areas and surface features. These sites scapes. Both the bare soil and villages provide similar blue-
include a very dry landscape with settlements and irrigated grey returns on a standard false color composite image. Visu-
agriculture in central Sudan (Wad Medani); a very wet location ally, neither the TM nor Sm-B imagery can independently differ-
of natural forest, plantation agriculture, and intense small scale entiate between the primary surface features in this area of
agriculture in western Kenya (Kericho);and a site of refugee Sudan (Haack and Slonecker, 1994).
camDs in an arid region in eastern Kenva IDadaabl. There is a
temGoral differenciin the analysis of &eie sites. The original
work was in Sudan and, based on its success, the methods have Kerlcho, Kenya
been extended to Kenya. This is a complex site in western Kenya, about 80 km to the east
of Lake Victoria and the port city of Kisumu. This area has a
Wad Medani, Sudan
This analysis was limited to an area along the Blue Nile River,
in central Sudan, where coincident radar and optical data had
been acquired. Specifically,the study area was~pproximately
25 km bv 40 km, and included the second lareest citv in Sudan.
Wad ~ i d a n iThe
. site extended northwest f k m w i d Medani '
along the Blue Nile. Wed Medani is about 160 km southeast of
~ h a 6 o u mand has apopulation near 100,000.It is a service city
for the large Geneid Gezira irrigated agricultural schemes
along the Blue Nile that extends west to the White Nile. This is
an extremely productive area for cotton and sugar cane.
Two primary data sets were obtained for this analysis. The
first was digital radar data from the SIR-B mission. The SIR-B
mission was flown in October 1984 and collected L-band (23.5-
cm) synthetic aperture data at a pixel size of 12.5 m. The sec-
ond data set came from a standard, seven-band digital Landsat
TM image, acquired on 18 November 1984. The TM sensor col-
lects data in six visible and infrared wavelengths at a 30-m spa-
tial resolution, and a thermal wavelength at a 120-mresolution.
The two data sets were geometrically registered and merged to band 8
the 12.5-mpixel size of the original SIR-B using a nearest-neigh- M Rhntr!
bor intensity resampling.
Ground information was obtained during a field visit in Figure 1. SIR-13Image along the Blue Nile in Sudan. Approxi-
1988. Using enlarged SIR-B and Landsat TM prints with limited mate original scale 1:40,000, reduced to approximately
available maps, samples of the various land uselland covers 1:110,000. North is to the upper left.
were documented on overlays to the hardcopy imagery. This

PHOTOGRAMMETRIC ENGINEERING 81REMOTE SENSING

high average elevation (1500 m) with considerable local varia- pink toned and flat textured tea plantations are distinctly visi-
tions and several steeply sloping features. The area has excel- ble, while at higher elevations on the southeast edge of the sub-
lent volcanic soils and suitable climatic conditions for scene is the natural forest in dark red.
agriculture, and also one of the highest population densities in Areas of settlement are limited within this region. Most
East Africa. housing is associated with the small, dispersed farms or the
Small, family owned and operated farms of mixed crops large-scale agricultural landscapes. The housing for the
cover much of the region. This is very intense agriculture, as employees of the tea estates is quite concentrated. These areas
field sizes are small and the cropping patterns are complex. The are small blue-grey tones among the tea. They are, however,
crops include corn, legumes of various types, mixed vegeta- spectrally and spatially very similar to the cut-back tea fields
bles, fruit including bananas and papaya, and small plots of tea on the TM subscene. The larger city of Kericho, located near the
and coffee, most of which are for family consumption. This center of the subscene between the small-scale agriculture and
crop complexity and the small field sizes make it nearly impos- tea, is difficult to delineate due to the similarity in tone (blue-
sible to map individual crops. grey) to the bare soil.
Large-scale plantations of tea are present in the more ele- Figure 2 is a RADARSAT, C band (5.6 cm) HH polarization,
vated areas. These fields are quite extensive and provide a high image collected on 27 February 1997. There is surprisingly lit-
green vegetation response thoughout the year as tea does not tle differentiation between the primary land covers in this
have an annual dormant season. Each year a small number of scene. The city of Kericho is a bit more evident, as an area of
fields are cut back to a minimum stem and primary branches to very high returns. The limited variation in the scene was con-
promote better growth on about a seven-year rotation. For sev- sistent over a variety of dates and is apparently a result of fre-
eral months followingthis cut back, there is no green vegetation quent rainfall and surfaces being quite moist, providing
in these fields, and optically the fields are spectrally similar to similar backscatter.
bare soil. Another cover type in this region is mature, evergreen
forest which occupies the highest elevations. Dadaab, Kenya
Plate 1is a Landsat TM image of this study site. Three of the This is a refugee area located in the very arid region of eastern
primary cover types (urban areas excluded) are easily seen in Kenya just north of the equator, about 100 km northeast of the
this image based on their tone, texture, and continuity. This is a larger city of Garissa and 70 krn west of the Somali border. This
dry season image and the small, fallow fields are distinctly vis- is a historic area of low population density, based on nomadic
ible in the northwestern part of the image as blue-grey spectral herding with virtually no agricultural cropping. There are three
tones. Small areas of green trees, tea, and coffee are the inter- large refugee camps, each with about 30,000 individuals, mostly
spersed red features among the fields. Moving east the large, from Somalia. These three camps are under the supervision of

Plate 1. Landsat TM scene of Kericho collected on 21 January 1995. Bands 3,

5, 4, are displayed in BGR. Approximate original scale of 1:140,000, reduced
to approximately 1:185,000. North is at the top. O Space Imaging/Landsat 1995.

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING lune 2000 711

t

Figure 2. RADARSAT scene of Kericho collected on 27 February 1997. Approxi-

mate original scale of 1:140,000 reduced to approximately 1:185000. North
is to the top. ORADARSAT (1997).

the United Nation's High Commission for Refugees (UNHCR) sites, separate from the training areas, which were also derived
and are very compact and densely populated. The areas around from the field efforts. For all accuracy assessments, polygons
the camps are medium height desert shrubs (1to 3 m) with lit- of the primary cover types were used as truth areas and not indi-
tle grass cover. Much of the vegetation within the immediate vidual pixels.
proximity of the camps has been removed by grazing or for fuel The results from this study compare the accuracy from var-
and building materials. ious classifications for individual land-uselland-cover types
Various spaceborne data were received for this site, includ- and for all of these classes combined. A number of data combi-
ing a Landsat TM scene from 17 February 1995. A RADARSAT nations and geospatial manipulations of the data were exam-
image from 15 August 1996 (Figure 3), having a spatial resolu- ined, including comparisons of the original sensor data
tion of 25 by 28 m, was also acquired and co-registered to the independently and in combination, incorporation of texture
optical bands. The camps are very distinct visually due to their measures, and speckle reduction procedures.
high backscatter and geometric shape. Areas around the camps
are dark tones of low backscatter from bare soil, because much Wad Medanl, Sudan -0rlgInal Values
of the vegetation has been removed for fuel or by grazing. The Spectral signatures for the primary surface features (Table 1)
vegetation provides a range of backscatter that is generally quite were extracted from training data, obtained during field visita-
high and similar to the camps in intensity. The classes selected tion. 'Itvo training sites were included for each of these features
for this scene included these camps, the bare soil, and various to provide information on the within-class variations. An
types of vegetation. examination of this table correlates well with what can be
observed in the imagery. Reflectance values for the urban and
Methodology and Results other (bare soil) classes are not significantly different in any of
The basic procedure was to conduct a digital classification the TM bands, but can be differentiated in the SIR-Bdata. By
using standard processing techniques applied to the spatially contrast, the spectral signatures for the agriculture and urban
coregistered set of spaceborne radar and optical bands, all classes are similar in the radar data, but are separable in several
resampled to the same pixel size. Spectral signatures were of the TM bands, particularly bands 5 and 7. Neither sensor can
extracted for the various land uses/covers using supervised independently delineate the primary cover types, either visu-
training site procedures. Field visits were used to identify valid ally or digitally.
training sites. After signature extraction, a decision rule was The classification strategy was to first separate the urban
employed to classify the data set and a contingency table com- and other (bare soil) sites from the agricultural sites using the
piled for accuracy assessment. The contingency tables were TM bands; then, the urban areas could be delineated from the
created from comparison of classifications to a set of truth other (bare soil) areas using the SIR-B data. This process was

712 l u n e 2000 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

 Figure 3. RADARSATscene of Dadaab from 15 August 1996.Approximate origi-
nal scale of 1:180,000 reduced to approximately 1:240,000. North is to the
top. ORADARSAT (1996).

Cover TMI TM2 TM3 TM4 TM5 TM6 TM7 SIR-B

Agriculture 1 Mean 87 40 47 86 82 162 35 168
SD 10 7 15 22 17 7 12 42
Agriculture 2 Mean 89 41 51 75 82 165 38 161
SD 10 8 15 14 23 9 14 38
Urban 1 Mean 106 56 82 74 129 183 80 151
SD 5 4 7 4 12 2 9 60
Urban 2 Mean 116 60 85 75 118 182 74 153
SD 7 6 8 6 10 2 8 46
Other 1 Mean 103 56 85 72 129 191 79 77
SD 3 2 4 3 6 1 3 17
Other 2 Mean 100 55 87 72 132 189 79 78
SD 4 4 8 6 14 2 8 17

implemented using a parallelepiped classification procedure.

This is a simple classification logic that draws upon the basic
Kerlcho, Kenya -Variance Texture Measurea
Multiple-spectralsignatures were extracted using training sites
complimentary nature of the fused ‘rh4and SIR-Bdata set. for the surface features in Kericho. Representative spectral signa-
The confusion matrix showing the classification results for tures for the different land uses/covers can be seen in Table 3.
these categories can be seen in Table 2. An overall classification These signatureswere extracted from training data obtained dur-
accuracy of 94 percent was achieved, demonstrating the value ing field visitation. There was little spectral variation within sur-
of sensor integration (Haack and Slonecker, 1994). face features, so only one signatureper class is presented.

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

 ClassifiedITruth Urban Vegetation Other Totals
Urban
Vegetation
Other
Totals
Correct %
Correctly Identified Pixels

(Haralick, 1973;Nuesch, 1982).Many attempts have been made

RSAT to define, characterize, construct, and incorporate quantitative
Cover TM3 TM4 TM5 RSAT Texture texture measures in remote sensing with varied results (Fasler,
1980;Irons and Peterson, 1981;Wang and He, 1990).
Forested Mean 17 74 46 132 39 A previous study has compared different texture measures
SD 1 6 4 45 3 and window sizes for East Africa (Haack and Bechdol, 2000).
Tea Mean 23 114 71 109 30 Four texture algorithms were examined: mean Euclidean dis-
SD 1 7 4 30 2
Urban Mean 49 46 76 193 59 tance, variance, skewness, and kurtosis. Results indicated that
SD 7 8 8 59 4 the second-order measure of texture, variance, was the most
Mixed Agriculture Mean 35 66 94 95 30 useful for feature delineations in East Africa.
SD 4 14 10 31 4 Most texture measures use a moving array of cells with a
variety of mathematical measures to derive texture values for
the center cell of the moving array. The size of the moving win-
dow may influence classification accuracies. Hsu (1978) stated
For the Landsat data, three bands were selected for initial that relatively small moving arrays can cause extensive mis-
examination. Those were bands 3,4, and 5 from the visible red, classification at the boundaries between classes. Blom and
near-infrared, and mid-infrared regions of the electromagnetic Daily (1982) found that window sizes of 15 by 15,31by 31, and
spectrum. These three spectral regions, as recorded in TM, gen- 61 by 61 provided constructive results for larger scale
erally provide good analysis results. A more traditional maxi- applications.
mum-likelihood decision rule was employed over a Texture measures were examined for East African sites
parallelepiped decision as in Wad Medani. There was, there- with many window sizes. Although an increase in overall clas-
fore, a preference to use a smaller number of optical bands in sification accuracy was steady as window sizes became larger,
the merger, so that the radar data would maintain a sufficient the increase was very slight. The point of diminishing returns
influence in the integration. Maximum likelihood also has the for overall classification accuracies due to window size was
advantage of being a simpler decision rule to use. generally at the 13 by 13 window size (Haack and Bechdol,
The TM bands 3,4,5 analysis provides a very good overall 2000).
classification, but poor discrimination of settlements. The Based on these previous studies, classification for Kericho
extremely poor resultant accuracy for the settlement class can used the variance measure of texture at a 13by 13 moving win-
be seen in Table 4, along with a summary of all the classifica- dow to determine if results could be improved over those of the
tion accuracies. A classification based on all seven TM bands original RADARSAT data. Overall classification accuracy was
increased the settlement accuracy to 31 percent, which was improved from 41 to 67 percent. Despite the fact that this is still
still not useful. The potential for TM spectral confusion far below the results achieved with the use of the three TM
between urban and mixed agriculture, primarily bare soil at bands, texture provided an excellent ability to discriminate the
this date, can be recognized in Table 3. Both have very similar settlementlurban areas (99 percent correct). The spectral sig-
spectral responses and relatively high standard deviations, natures for texture (Table 3) indicate this potential. The high
making their separability difficult. texture of the urban areas seems reasonable because the city of
The overall classification results for the original RADARSAT Kericho and the tea estate settlements are structurally complex
data are quite low. The more unique, higher backscatter areas areas with buildings of high backscatter separated by open
of settlements have the best classification accuracy at 67 per- areas of low backscatter and, thus, high texture. In contrast, the
cent. The RADARSAT spectral signatures, seen in Table 3, for the forest, tea, and agriculture have much lower texture values.
three primary covers are not greatly different. Their average val- As was the case in Wad Medani, there is an advantage to
ues are 95,109, and 132 and all have high standard deviations, sensor fusion. In Wad Medani, neither sensor could indepen-
resulting in spectral confusion. dently delineate the primary cover types. In Kericho, each sen-
As both a function of the radar signal and the complexity of sor could independently classify one or more covers, but not
the radar interaction with surface features, radar generally has all covers, and an added textural variable was needed. The
a high texture in comparison to optical data. Increasingly, digi- combined classification, seen in Plate 2, delineates all covers.
tal measures of texture are incorporated in automated classifi- This plate as well as the classification results (Table 4) clearly
cations. Digital texture is the spatial variation of pixel values indicate the advantage of the TM and radar texture fusion.

TABLE
4. CLASSIFICATION ACCURACIESCOMPARING
SENSORS
AND SENSOR
COMBINATIONS
FOR KERICHO

Forested Tea Urban Mixed Agriculture Overall

Landsat TM bands 345 - 21 Jan,1995 99.44 99.93 00.88 98.68 98.62
Radarsat - 27 Feb, 1997 30.96 39.62 66.67 57.70 40.67
Radarsat Variance Texture - 13 X 13 Window 91.72 55.41 99.12 28.46 66.78
TM 345 and RSAT Variance Texture 99.95 99.46 98.25 99.96 99.65

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

 the relationship between speckle reduction, measures of tex-
ture, and smoothing methods. Speckle is an unavoidable result
of using coherent monochromatic light/energy such as in
radar, and appears as a noisy, high textured image. Speckle is
perhaps the most difficult problem to address when employing
radar imagery because even homogeneous surfaces may have a
grainy appearance and high backscatter standard deviations
(Durand et a]., 1987).
It is difficult to separate which radar tonal variations are
due to coherent speckle and which are due to actual ground
variations. Because an automatic classifier expects a given sur-
face to be represented by a certain image tone or brightness, the
incorrect values caused by speckle make automated classifica-
tions difficult. As a result, the presence of speckle often pre-
cludes a pixel-by-pixel classification unless the data are
smoothed (Niiesch, 1982).
While speckle cannot be removed from data, its impact can
be significantly suppressed through speckle reduction algo-
rithms. The application of a window that averages pixels will
reduce the effect of speckle (Blom and Daily, 1982),but will
also visibly degrade edges. This is especially important when
dealing with smaller agricultural fields or urban studies where
boundaries play a critical role (Jensen, 1979;Niiesch, 1982).
Generally, there is no value in using a speckle reduction
algorithm and extracting texture measures from the same data,
Plate 2. Sensor integration classification of Kericho for because the speckle reduction will remove most of the texture.
small-scaleagriculture,tea, settlements, and forest. Approxi- The exception to this is if the speckle reduction and texture
mate original scale of 1:140,000reduced to approximately measures are at different window sizes. For Dadaab, a median
1:250,000.Forest in red, tea in yellow, small scale agricul- speckle reduction method was applied using a 5- by 5-pixel
ture in green, and settlements in purple. window. This was a compromise between limiting the speckle
and still retaining image detail including texture. A variance
texture extraction and classification was then applied using a
much larger 21- by 21-pixel window. The results are an
Dadaab, Kenya -
Speckle Reduction and Texture improvement for overall classification, and especially for
camps, in comparison to the two other radar classifications.
The problem of mapping settlements also occurs in Dadaab, as
a significant spectral confusion exists between the refugee The TM bands 3,4,5and despeckled variance texture com-
camps and the areas of bare soil and dry vegetation. From visual bination results are a considerable improvement in overall
inspection of either the optical TM or radar data, the settle- classification (84 percent) and in individual class discrimina-
ments are quite apparent based on their geometry. However, tion. Further refinement of the manipulations of the radar and
spectrally the discrimination is not easy. Many of the housing inclusion of different dates of the optical data such as a green
units are constructed of natural brush, thus creating confusion season image (similar to the strategy used in Wad Medani),
with dry vegetation. In addition, some of the housing uses may further improve upon these results.
metal or plastic roofs that are spectrally similar to bare soil in
these wavelengths. Summary and Future Directions
Classification accuracies (Table 5) for the TM bands 3,4,5 These results show the potential of optical/radar merger for
combination is limited, at 73 percent overall. The results for mapping basic land-uselland-cover patterns in different envi-
the camps are reasonable, at 65 percent, but there are many ronments in East Africa. Given the range of landscapes exam-
errors of commission where the scattered vegetation and bare ined, the results of this study should have applicability beyond
soil are incorrectly delineated as urban. East Africa. The results support the value of integrating optical
Both the classifications of the original RADARSATvalues and radar data. In all sites examined in this study, sensor fusion
and variance texture provide low overall results. Texture mea- improves the classification accuracies, particularly for
sures failed to delineate camp areas even though, intuitively, it settlements.
would seem as if the camps should have a texture different The radar results can be improved by various manipula-
from the surrounding landscapes. The housing units are so tions, such as derived texture measures and filtering. The strate-
dense that there may be little texture at the spatial resolution of gies for using radar that provided the best results were not
the sensor. consistent. As with optical data, the most productive pro-
One of the problems associated with radar is speckle and cessing strategies may be site and data specific. More case stud-

Urban/ Scattered Natural

Thicket Bare Soil Camps Vegetation Agriculture Overall
Landsat TM bands 345 96.42 69.05 64.57 51.94 99.04 72.77
Radarsat- 15 Aug. 1996 26.00 14.62 00.00 31.02 54.71 26.49
Radarsat Variance Texture - 21 X 21 Window 14.52 23.88 54.91 39.26 14.41 24.49
Despeckled Radarsat Variance Texture - 21 X 21 Window 36.67 00.00 64.40 56.01 17.72 39.57
TM 345 and Despeckled RSAT Variance Texture 95.60 67.71 72.29 78.45 98.96 83.78
- - -

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

ies will contribute to an improved understanding of useful Haack, B., and E. Slonecker, 1994. Merged spaceborne radar and The-
analysis techniques. matic Mapper digital data for locating villages in Sudan, Photo-
Future applications of this project will include a compari- grammetric Engineering & Remote Sensing, 60(10):1253-1257.
son of the parallelepiped accuracy with that of a maximum- Haralick, R.M., 1973. Textural features in image classification, IEEE
likelihood and other classifiers, including a hierarchical 7hnsactions on Systems, Man and Cybernetics, SMC-3510-621.
approach, and an extension of basic land uselcover to more Harris, J.R., R. Murray, and T. Hirose, 1990. IHS transform for the
complex land uselcover classification schemes. Most impor- integration of radar imagery and other remotely sensed data, Pho-
togrammetric Engineering & Remote Sensing, 56(12):1631-1641.
tantly, further evaluation of the contribution of the individual
Hsu, S., 1978. Texture-tone analysis for automated land-use mapping,
bands and selection of band subsets for classification will be Photogrammetric Engineering 6 Remote Sensing,
undertaken. Another extension of this study will be a more 44(11):1393-1404.
complete evaluation and comparison of texture and spatial fil- Irons, J.R., and G.W. Peterson, 1981. Texture transformations of remote
ters to improve accuracy. sensing data, Remote Sensing of Environment, 11:359-370.
Jensen, J.R., 1979. Spectral and textural features to classify elusive
Acknowledgments land cover at the urban fiinge, The Professional Geographer,
These results are based upon projects supported by the United 31(4):400-409.
States National Science Foundation International Programs Luckman, A.J., A.C. Frery, C.C.F. Yanasse, and G.B. Groom, 1997. Tex-
Division under Grant Number INT-8519776 and the ~ G t e d ture in airborne SAR imagery of tropical forest and its relationship
States National Aeronautical and Space Administration to forest regeneration stage, International Journal of Remote Sens-
Application Development and Research Opportunity program ing, 18(6):1333-1349.
under Grant Numbers NAGW-4970 and NAG5-4140. The coopera- Lo, C.P., 1986. Settlement population and land use analysis of the
tion of the Regional Centre for Services in Mapping, Surveying north China plain using Shuttle Imaging Radar-A data, The Profes-
and Remote Sensing in Nairobi, Kenya, particularly Luka sional Geographer, 38(2):141-149.
Isavwa, was necessary for this project and greatly appreciated. Morain, S.A., 1991. Observations on transferring earth observing tech-
nology to the developing world, Technical Papers of the ACSM-
In addition, the Sudan National Remote Sensing Centre in ASPRS Annual Convention, Baltimore, Maryland, 3:282-293.
Khartoum provided assistance. Niiesch, D.R., 1982. Classification of SAR imagery from an agricultural
region using digital textural analysis, International Society for
Photogrammetry and Remote Sensing's International Archives.
References 24(11/1):231-240.
Blom, R.G., and M. Daily, 1982. Radar image processing for rock-type Paris, J.E, and H.H. Kwong, 1988. Characterization of vegetation with
discrimination, ZEEE Transactions on Geoscience and Remote combined Thematic Mapper (TM) and Shuttle Imaging Radar
Sensing, GE-20(3):343-351. (SIR-B) image data, Photogrammetric Engineering b Remote Sens-
Brisco, B., and R.J. Brown, 1995. Multidate SARITM synergism for ing, 54(8):1187-1193.
crop classification in western Canada, Photogrammetric Engi- Paul, C.K., and A.C. Mascarenhas, 1981. Remote sensing in develop-
neering 9 Remote Sensing, 61(8):1009-1014. ment, Science, 214(4517):139-145,
Dobson, M.C., F.T. Ulaby, T.L. Toan, A. Beadoin, E.S. Kasischke, and Pohl, C., and J.L. Van Genderen, 1997. Multisensor image fusion in
N. Christensen, 1992. Dependence of radar backscatter on conifer- remote sensing: concepts, methods and applications, Interna-
ous forest biomass, LEEE ll-ansactions on Geoscience and Remote tional Journal of Remote Sensing, 19(5):823-854.
Sensing, 30(2):412-415. P~asad,T.S., and R.K. Gupta, 1998. Texture based classification of
Durand, J.M., B.J. Gimonet, and J.R. Derbos, 1987. SAR data filtering multidate SAR images - A case study, Geocarto International,
for classification, IEEE Transactions on Geoscience and Remote 13(3):53-62.
Sensing, GE-25(5):629-637. Raghavawamy, V., N.C. Gautam, M. Padmavathi, and K.V.S. Badarinath,
Elachi, C.T., 1988. Spaceborne Radar Remote Sensing: Applications 1996. Studies on microwave remote sensing data in conjunction
and Techniques, IEEE Press, New York, 255 p. with optical data for land uselland cover mapping and assessment,
Geocarto International, 11(4):25-31.
Fasler, F., 1980. Texture measurement for Seasat-SAR images for urban Ranson, K.J., S. Saatchi, and G. Sun, 1995. Boreal forest ecosystem
land use interpretation, Proceedings of the Fall Technical Meeting, characterization with SIR-C/X-SAR, IEEE Transactions on Geosci-
ASCM-ASP, Niagara Falls, New York, pp. PS-2-B-1 to PS-2-B-10. ence and Remote Sensing, 33(4):867-876.
Foody, G.M., 1988. Crop classification from airborne synthetic aperture Richards, J.A., 1990. Radar backscatter modeling of forests: a review
radar data, International Journal of Remote Sensing, of current trends, International Journal of Remote Sensing,
9(4):655-668. 11(7):1299-1312.
Forster, B., 1996. Current advances in radar remote sensing research Wang, L., and D.C. He, 1990. A new statistical approach for texture
and its application in south east Asia, Geocarto International, analysis, Photogrammetric Engineering & Remote Sensing,
11(4):3-10. 56(1):61-66.
Haack, B., and M. Bechdol, 2000. Integrating multisensor data and Welch, R., and M. Ehlers, 1988. Cartographic feature extraction with
RADAR texture measures for land cover mapping, Computers and integrated SIR-B and Landsat TM images, International Journal
Geosciences, in Press. of Remote Sensing, 9(5):873-889.

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

View publication stats