Full text of DOI:10.1080/01431160119174

This article was downloaded by: [171.67.34.69]
On: 29 July 2013, At: 07:58
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
International Journal of
Remote Sensing
Publication details, including instructions for
authors and subscription information:
http://www.tandfonline.com/loi/tres20
Vegetation mapping of a
tropical freshwater swamp
in the Northern Territory,
Australia: A comparison of
aerial photography, Landsat
TM and SPOT satellite
imagery
K. R. Harvey & G. J. E. Hill
Published online: 25 Nov 2010.
To cite this article: K. R. Harvey & G. J. E. Hill (2001) Vegetation mapping of a
tropical freshwater swamp in the Northern Territory, Australia: A comparison of
aerial photography, Landsat TM and SPOT satellite imagery, International Journal
of Remote Sensing, 22:15, 2911-2925
To link to this article: http://dx.doi.org/10.1080/01431160119174
PLEASE SCROLL DOWN FOR ARTICLE
Taylor & Francis makes every effort to ensure the accuracy of all
the information (the “Content”) contained in the publications on our
platform. However, Taylor & Francis, our agents, and our licensors
make no representations or warranties whatsoever as to the accuracy,
completeness, or suitability for any purpose of the Content. Any opinions
and views expressed in this publication are the opinions and views of
the authors, and are not the views of or endorsed by Taylor & Francis.
The accuracy of the Content should not be relied upon and should be
independently verified with primary sources of information. Taylor and
Francis shall not be liable for any losses, actions, claims, proceedings,

demands, costs, expenses, damages, and other liabilities whatsoever
or howsoever caused arising directly or indirectly in connection with, in
relation to or arising out of the use of the Content.
This article may be used for research, teaching, and private study
purposes. Any substantial or systematic reproduction, redistribution,
reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access
and use can be found at http://www.tandfonline.com/page/terms-and-
conditions

int. j. remote sensing, 2001, vol. 22, no. 15, 2911–2925
Vegetation mapping of a tropical freshwater swamp in the Northern
Territory, Australia: a comparison of aerial photography, Landsat TM
and SPOT satellite imagery
K. R. HARVEY* and G. J. E. HILL
Faculty of Science, Information Technology and Education and Tropical
Savannas CRC, Northern Territory University, Darwin, NT 0909, Australia
(Received 26 April 1999; in Ž nal form 11 February 2000)
Abstract. The tropical wetland environments of northern Australia have ecolo-
gical, social, cultural and economic values. Additionally, these areas are relatively
pristine compared to the many other wetland environments in Australia, and
around the world, that have been extensively altered by humans. However, as the
remote northern coastline of Australia becomes more populated, environmental
problems are beginning to emerge that highlight the need to manage the tropical
wetland environments. Lack of information is currently considered to be a major
factor restricting the eVective management of many ecosystems and for the expans-
ive wetlands of the Northern Territory, this is especially the case, as these areas
are generally remote and inaccessible. Remote sensing is therefore an attractive
technique for obtaining relevant information on variables such as land cover and
vegetation status. In the current study, Landsat TM, SPOT (XS and PAN) and
large-scale, true-colour aerial photography were evaluated for mapping the vegeta-
tion of a tropical freshwater swamp in Australia’s Top End. Extensive ground
truth data were obtained, using a helicopter survey method. Fourteen cover types
were delineated from 1:15000 air photos (enlarged to 1:5000 in an image processing
system) using manual interpretation techniques, with 89% accuracy. This level of
detail could not be extracted from any of the satellite image data sets, with only
three broad land-cover types identiŽ ed with accuracy above 80%. The Landsat
TM and SPOT XS data provided similar results although superior accuracy was
obtained from Landsat, where the additional spectral information appeared to
compensate in part for the coarser spatial resolution. Two diVerent classiŽ cation
algorithms produced similar results.
1. Introduction
The coastal and sub-coastal areas of the Northern Territory contain numerous
freshwater wetlands, many of which are listed in the Directory of Important Wetlands
in Australia (Jaensch 1993, Whitehead and Chatto 1995). Additionally, 192 000 ha
of wetlands in the Cobourg Aboriginal Land and Wildlife Sanctuary, and 667 000 ha
in Kakadu National Park (Stage 1) are listed as Wetlands of International
Importance under the Ramsar Agreement. The wetland areas of the Northern
Territory provide important habitats for wildlife (e.g. Finlayson et al. 1988, Whitehead
*Now at: EcOz, GPO Box 381, Darwin, NT 0801, Australia;
e-mail: kylieharvey@ecoz.cam.au
International Journal of Remote Sensing
ISSN 0143-1161 print/ISSN 1366-5901 online © 2001 Taylor & Francis Ltd
http://www.tandf.co.uk/journals

2912
K. R. Harvey and G. J. E. Hill
et al. 1990) and form a key resource of social, economic and cultural importance to
traditional Aboriginal land owners.
A number of threats and management problems can be identiŽ ed in the Northern
Territory’s tropical wetland environments. The physical impacts of feral animals
(including buValo, cattle, horses and pigs) are visually evident and have been implicated
with other problems such as saltwater intrusion (e.g. Finlayson et al. 1988, Whitehead
et al. 1990, Pidsley et al. 1993) and the establishment of noxious weeds (e.g. Finlayson
et al. 1988, Whitehead et al. 1990, Storrs and Finlayson 1997). A new threat to the
integrity of these wetland ecosystems is the cane toad (Bufo marinus) which will have
likely adverse impacts on native fauna as it moves into the higher rainfall areas of the
Northern Territory. The potential and actual environmental problems in the tropical
wetlands of the Northern Territory highlight the need for restoration and preventative
management. Currently such management is hindered by the lack of information
available on which to base and justify management interventions.
The paucity of information on wetland areas worldwide can be attributed to the
fact that the physical characteristics of wetlands make Ž eldwork expensive, time
consuming and often inexact (Lee and Lunetta 1995). These factors limit Ž eldwork
in wetlands to small areas and restrict the usefulness of the data for extrapolation
(Cowardin and Myers 1974, Jensen et al. 1984). Remote sensing, with its utility for
surveying large areas in a time and cost eVective manner, is a possible solution to
diYculties of this type as is illustrated by its successful application to wetland
mapping in general (e.g. Cowardin and Myers 1974, Scarpace et al. 1981, Young and
Dahl 1994, Lee and Lunetta 1995).
Aerial photography has been a widely used remotely sensed data source for
wetland information acquisition both historically and more recently (e.g. Scarpace
et al. 1981, Jensen et al. 1984, Hefner and Storrs 1994, Dale et al. 1996, Ozesmi and
Mitsch 1997, Mumby et al. 1999). The inherent tradeoV between spectral and spatial
resolution in current satellite remote sensing platforms restricts the utility of satellite
imagery for mapping wetlands. Wetlands are often spatially characterised by steep
ecological gradients with vegetation units narrower than the pixel size of current
sensors (e.g. Johnston and Barson 1993, Young and Dahl 1994, Lee and Lunetta
1995). Even where the most superior spatial resolution (10 m) would suYce, the
associated inferior spectral resolution may result in the inability to distinguish
between diVerent vegetation types, which are often spectrally similar in wetland
environments (e.g. Cowardin and Myers 1974, Ernst-Doattavio et al. 1981, Jensen
et al. 1984, Catt and Thirarongnarong 1992, Johnston and Barson 1993, Mattheson
1994, Anderson and Perry 1996, Hacck 1996, Phinn and Hill 1998). The textural
features and superior spatial resolution of aerial photography make it a very useful
data source for detailed wetland mapping, however, this method is not feasible for
collecting regional information or information requiring continual validation,
because of the cost and logistical diYculties associated with capturing the data
(Jennings et al. 1992, Johnston and Barson 1993, Jensen et al. 1993 ).
Examples of the application of remote sensing (airborne and spaceborne) in
tropical wetland environments speciŽ cally are limited. Research is required to develop
remote sensing methodologies that factor in the extreme environmental conditions
experienced in tropical areas. This paper documents a study conducted at Melacca
Swamp in the Northern Territory of Australia, a tropical region in which little
published work on remote sensing of wetlands is available. The aim of this research
was to investigate the utility of remotely sensed data of varying spectral and spatial
resolutions for the spectral discrimination of vegetation types within a tropical

Vegetation mapping of a tropical freshwater swamp
2913
freshwater swamp environment. This was investigated with the ultimate goal of
determining which remotely sensed data source is most suited to mapping these
environments. Melacca Swamp is a highly heterogeneous, freshwater wetland envir-
onment, which provided the opportunity to identify the limits of remote sensing
capabilities for monitoring this type of wetland in the region.
2. The study area
Melacca Swamp is a permanent freshwater wetland adjacent to the Adelaide
River, approximately 50 km east of Darwin, Northern Territory (Ž gure 1). The swamp
drains into Melacca Creek and unlike many of the swamps in the Top End that dry
out in the ‘dry’ season, Melacca Swamp is permanently inundated, with a spring fed
creek traversing the southern section (Webb et al. 1983). The area has a high
conservation value, as it is one of the most highly utilised nesting habitats of the
previously endangered saltwater crocodile (Crocodylus porosus) in the Northern
Territory. Melacca Swamp is managed by the Parks and Wildlife Commission of
the Northern Territory as a reserve speciŽ cally for the conservation of the saltwater
crocodile nesting habitat that occurs there (PWCNT 1996).
The Melacca Swamp Conservation Reserve encompasses an area of 23 km . The
2
southern section of the reserve area was included in this study. The main cover types
present are woodlands and open forests of Melaleuca spp. with an understory of
T horacostachyum sumatranum and Phragmites karka sedges. Additionally, some large
areas of sedges, namely T. sumatranum, P. karka and T ypha sp., without a canopy
of Melaleuca spp. are also present. Open woodland comprising Acacia, Eucalyptus
and Melaleuca spp. intermixed with open grassland areas surrounds the swamp,
which is bordered by a dense fringe of Melaleuca and monsoon vine-forest species.
A large area of  oodplain, dominated by numerous grass and sedge species, extends
from the eastern extremes of Melacca Swamp to the banks of the Adelaide River. A
Figure 1. The location of Melacca Swamp in the Northern Territory, Australia.

2914
K. R. Harvey and G. J. E. Hill
band of mangroves follows Melacca Creek from the Adelaide River and at the
transition between the saline mangroves and the freshwater swamp vegetation, T ypha
sp. and Flagellaria indica are dominant (Webb et al. 1983).
The climate of the study area is typical of the monsoonal weather regime that
dominates the northern coastal belt of Australia. Two seasons occur, the ‘wet’ and
the ‘dry’. These seasons are distinct with large diVerences in the amount of rainfall
and humidity levels. Mean annual rainfall is approximately 1380 mm. Ninety-Ž ve
per cent of the yearly rainfall occurs during the wet season (November to April) and
during these months humidity rises to over 80% daily. The dry season produces
milder weather conditions with a low probability of rainfall and maximum humidity
of 67–70% daily. Temperature does not vary greatly between the seasons with the
maximum daily temperatures ranging between 31 C (July) and 35 C (October).
°
°
3. Methods
Four remotely sensed data sets of Melacca Swamp were obtained. These were: a
Landsat TM image acquired on 20 May 1995; a SPOT XS image acquired on
24 October 1995; a SPOT Panchromatic image acquired on 24 October 1995; and
1: 15 000 scale, true colour aerial photography acquired on 4 July 1982. Landsat
TM spectral bands two, three, four and Ž ve were chosen for this study based on the
proven suitability of these bands for wetland studies (e.g. Dottavio and Dottavio
1984, Walsh et al. 1990, Johnson and Barson 1993, Melack et al. 1994). Additionally,
bands two, three and four have been identiŽ ed as useful for determining the presence
of a vegetated understory (Congalton et al. 1993), which is characteristic of some
cover types in the swamp. All bands of SPOT imagery were purchased.
Processing of satellite imagery and aerial photography was undertaken in the
ERDAS Imagine (version 8.3.1) digital image processing system and ArcView GIS
(version 3.0). Tools in the microBRIAN digital image processing system were also
utilised where the other systems did not provide the functions required.
3.1. Pre-processing
Aerial photography was obtained in hardcopy and converted to digital format
for computer aided analysis. Each photograph was scanned using a UMAX Mirage
D-16L scanner. As increasing the resolution of the scanning procedure did not
appreciably improve the visual quality of the scanned photographs, a resolution of
200 dots per inch was used, achieving the desired clarity whilst minimising Ž le
storage space.
All data sets were georeferenced to 1:50 000 topographic maps of the study area,
which were the largest scale available. Ground control points (GCPs) were selected
at locations distinct on both the imagery and topographic maps. Whilst, ideally,
GCPs should be as small as a single pixel (Campbell 1996) the nature of the study
area, which is mainly comprised of undeveloped open woodland and  oodplain
areas, made the selection of the appropriate number and distribution of GCPs a
challenging task.
A photo basemap was constructed in the form of a continuous mosaic of each
individual aerial photo. Each photo was referenced to an arbitrary coordinate system
(row/column values). This method was chosen as the area of each photo provided
insuYcient features for the location of ground control points needed to reference to
a standard projection. Once the Ž rst photo in each run was referenced to the row/
column coordinate system, each subsequent photo was referenced to the base photo

Vegetation mapping of a tropical freshwater swamp
2915
through the selection of GCPs in the 60% area of overlap. Each run was referenced
to the adjacent run using the same method, however, the smaller 20% overlap made
the selection of reference points more diYcult. Each photo was resampled to the
new reference system using the nearest neighbour resampling method.
The referenced photos were mosaiced using the MOSAIC module in Imagine.
Afek and Brand (1998) found that the geometric distortions and radiometric diVer-
ences between adjacent photographs must be corrected to create a seamless mosaic.
To achieve this in the aerial photo mosaic of Melacca Swamp, image matching was
applied in the overlap area, using a feather function (ERDAS 1997). This matches
images by recalculating the pixel values in the overlap area and resulted in a visually
smooth transition at the boundaries between each photo and each run in the Ž nal
image mosaic.
A total of 20 GCPs were established on the Landsat TM image subset and 15
on each of the SPOT scenes. As suggested by Franklin (1993) a Ž rst order polynomial
equation was adopted to achieve a higher accuracy in the absence of a number of
highly accurate GCPs. The RMS error for each of the established GCPs was
evaluated and those points with errors representing a ground error of greater than
30 m were replaced. Resampling was performed using the nearest neighbour algo-
rithm to ensure that the geometric integrity of the data was maintained (Rutchey
and Vilcheck 1994, Campbell 1996, ERDAS 1997). Following resampling, examina-
tion of the registered imagery sets indicated that registration errors were small,
however, the potential for confusion during the Ž eldwork phase of the project was
noted because of the lack of readily identiŽ able ground points.
Subsets of the Melacca Swamp study area were extracted from each of the
registered image data sets. Each subset contained large areas of open woodland and
mangrove forest not pertinent to this study and these were masked in an eVort to
reduce the chances of class confusion in the classiŽ cation process and to decrease
processing time. Various enhancement models were tested on each image to determine
the best technique to apply to increase the overall contrast between all cover types.
Histogram equalisation provided a suitable method for enhancing Landsat TM and
SPOT PAN data, whilst the SPOT XS image was best enhanced using a linear
technique.
3.2. Interpretation of aerial photography
The referenced mosaic was used as a base map from which cover types present
in the swamp were digitised on-screen using the ArcView GIS. A number of display
scales were evaluated for the digitising process with a 1:5000 display providing for
clearest visual distinction of cover type boundaries. This display scale was maintained
for the duration of the digitising process. Vegetation cover types comprising the
swamp were manually digitised using tone and texture as the main interpretation
cues to deŽ ne the boundaries between cover types. Each delineated vegetation type
was labelled based on the knowledge of an experienced Ž eld researcher (B. Ottley).
Polygons representing the same vegetation types on the annotated map were merged,
having a single entry in the associated attribute table. The resulting coverage was
assessed to establish more speciŽ c class labels and to identify where misinterpretation
of the boundaries may have occurred.
3.3. Interpretation of satellite imagery
An unsupervised classiŽ cation technique was adopted in the analysis of satellite
imagery because of the suitability of this approach to areas with complex land cover

2916
K. R. Harvey and G. J. E. Hill
patterns and where Ž eldwork is diYcult (e.g. Stenback and Congalton 1990, Rutchey
and Vilcheck 1994, ERDAS 1996, Brady et al. 1997, Menges et al. 1998). In addition
to investigating the application of an unsupervised classiŽ cation technique in a
wetland environment, the impact of the algorithm used for the classiŽ cation was
also investigated through the adoption of both minimum distance to mean and
isodata classiŽ cations.
The minimum distance to mean algorithm assesses the euclidean distance between
each pixel value and the means for each class, assigning the pixel to the class from
which it has the minimum distance measurement (Harrison and Jupp 1993). A 15%
tolerance level was found to be most suitable for the SPOT XS and Landsat TM
imagery, whilst a narrower tolerance of 2% was required to derive a suYcient
number of classes from the SPOT PAN imagery. These tolerance levels were applied
to ensure that pixels unrelated to any of the classes in the image remain unclasssiŽ ed
(Harrison and Jupp 1993).
A maximum number of 50 classes was speciŽ ed for the classiŽ cation. Following
the initial classiŽ cation the statistics for all classes were viewed. Classes containing
only a small number of pixels were deleted, with the criteria for deletion deŽ ned by
the level of aggregation in the spatial distribution of pixels. The classiŽ cation was
iterated using the new class statistics from the initial classiŽ cation, minus the small
classes. The tolerence level was increased by 1% in order to ensure that the deleted
classes were assigned to their nearest neighbours (Ahmad and Neil 1994).
For the isodata algorithm 50 arbitrary cluster means were initially chosen by
default and allowed to shift through an iterative process until a maximum number
(95%) of unchanged pixels was reached between the Ž nal two iterations. The 50
classes produced by the classiŽ cation represented spectral variation within the
imagery but not all classes represented a distinct cover type. Spectrally similar classes
exhibiting spatial relationships were merged in an eVort to replicate the distribution
of cover types at Melacca Swamp. Canonical Variates Analysis (CVA), a separabilty
tool provided in the microBRIAN system (CANVAR module), was used to assist in
amalgamating classes. The canonical variates transformation matrix calculated
during this process was used to create a minimum spanning tree plot which provided
a graphical and numerical representation of class separability based on Mahalanobis
distance, a measure of euclidean distance between classes, in canonical variate space
(Harrison and Jupp 1990, Harrison and Jupp 1993).
Minimum Spanning Trees were prepared for each image classiŽ cation and used
in establishing spectral relationships between classes. This method was chosen as the
use of spectral as well as spatial characteristics for class aggregation has been
successfully adopted in other studies where classes are spatially contiguous as well
as spectrally similar (e.g. Congalton et al. 1993, Ahmad and Hill 1994).
3.4. Field sampling
Initial surveys were conducted during the 1997/1998 wet season. Accuracy assess-
ment for the vegetation mapping was completed during the late dry season in August
1998. In this phase of the project, due to the extreme diYculty of moving through
the swamp on foot or by boat, and because of the danger posed by crocodiles, a
Ž eld sampling scheme based on the use of a helicopter was devised. A Kawasaki
KH4 model helicopter was used, which dictated that the observers sit on opposite
sides, behind the pilot. The use of the helicopter determined the design of the Ž eld

Vegetation mapping of a tropical freshwater swamp
2917
program where locational precision, and sampling eYciency had to be balanced
against the cost of relying on this form of reconnaisance.
Four east–west transects were placed 500 m apart covering the saltwater crocodile
nesting area within the swamp. The position of the Ž rst transect was determined by
the position of the boundary between the swamp and open woodland areas. Five
sample points were placed at 1 km intervals along each transect from the starting
position at the boundary between the swamp vegetation and open woodland areas.
The east–west extent of the swamp was covered by the sample points.
Each point was labelled and the locational information recorded on data collec-
tion sheets for use in the Ž eld. In addition, the location of each sample site was
recorded in a look-up table, and a point vector Ž le containing each sample site was
plotted on the air photo mosaic to aid in the identiŽ cation of sample sites in the
Ž eld. This ensured that each site could be found in the event of a GPS malfunction
and that disagreements between the programmed GPS reading and the recorded site
locations could be easily rectiŽ ed. At each site, the helicopter was positioned at a
height of 20 m over the sample point (or the nearest identiŽ able point) and then
aligned to record observations to the north, south, east and west. The observers
recorded vegetation type at estimated distances of 20 m, 50 m and 100 m from a line
of sight through the helicopter strut. To ensure the correct cover types were recorded,
both photographic and written records were collected at each site. This procedure
provided 12 observations per sample point and a total sample size of 240 observations
for the entire swamp.
3.5. Accuracy assessment
Information recorded at each site was incorporated into the Ž eld vector Ž le of
sampling sites created prior to data collection. The distribution of these ground
samples ranged from 25–109 in the generalised cover types derived from satellite
data. In the more detailed classes derived from aerial photography the sample sizes
ranged from 1–7 for nine of the more restricted cover types and 15–35 for the
remaining Ž ve classes. The Ž eld sample points were overlayed on each of the land-
cover maps for comparison of classiŽ cation derived classes and actual ground classes.
Points that fell in areas not annotated during the classiŽ cation process were omitted
and points for which the geographic location recorded in the Ž eld was dubious were
also deleted, resulting in a small variation in sample size in the evaluation of each
classiŽ cation. The correspondance between the classiŽ ed and ground veriŽ ed cover
types was calculated based on the confusion matrix approach. For the purpose of
determining accuracy using this approach, the 240 observations from the multi-stage
sampling design were treated as independent observations. Therefore, the conŽ dence
intervals derived from this analysis are not absolute.
4. Results
Fourteen land cover classes, which adequately described the vegetation types at
Melacca Swamp, were derived from interpretation of the aerial photography
(Ž gure 2). The overall mapping accuracy, calculated from comparisons between
ground truth and mapped data, was 89% 5% (x 95% CI). Individual class accu-
±
±
racies ranged from 82%, for the open sedge class, to 100% for six of the other
classes. Errors of commission and omission were low for virtually all classes. When
the classes were aggregated into the three more generic land cover types discernable
on the satellite imagery, the overall mapping accuracy increased to 90.6% 4%.
±

2918
K. R. Harvey and G. J. E. Hill
Figure 2. Vegetation map of Melacca Swamp derived from manual interpretation of colour
aerial photography.
However, the overall increase in accuracy achieved (5%), was found to be insigniŽ c-
ant (p 0.5) when the accuracies of the aerial photo classiŽ cation both before and
after amalgamation were statistically compared.
The detailed information derived from the aerial photography could not be
extracted from the Landsat TM, SPOT XS, or SPOT PAN image data analysed.
Spectral overlap between quite diVerent vegetation cover types was evident in the
classiŽ cations derived from the SPOT data but reduced in those derived from the
Landsat TM data. An evident restriction on the utility of the Landsat TM satellite
imagery for detailed mapping in tropical freshwater wetlands was illustrated by the
inability to distinguish between Melaleuca forest with a sedge understory and
Melaleuca open woodland with a sedge understory. These vegetation types were
successfully delineated on the aerial photography and are a large component of most

Vegetation mapping of a tropical freshwater swamp
2919
tropical freshwater wetlands in the Northern Territory. To obtain acceptable accurac-
ies ( 80%) from the satellite image data, the vegetation communities had to be
reduced to a broader classiŽ cation consisting of three land-cover classes as illustrated
in Ž gure 3.
The aerial photo classiŽ cation has a higher level of detail, however, amalgam-
ation of the classes results in a similar classiŽ cation to that obtained from Landsat
TM satellite imagery. Classes 1–8 (Ž gure 2) collectively are similar to the woodland
class (class 1) derived from classiŽ cation of the Landsat TM satellite image. The
swamp class dominated by sedges depicted in the Landsat TM classiŽ cation (class
2) most closely matches the distribution of sedges depicted in the aerial photo
classiŽ cation, however, some misclassiŽ cations of sedges and areas comprised of
a mixture of grasses and sedges did occur (class 3). The spatial distribution of
the areas dominated by grasses and sedges in the Landsat TM and SPOT XS
classiŽ cations (Ž gure 3) appears to show some correlation with classes 13 and 14
derived from aerial photography (Ž gure 2).
The variation in overall mapping accuracy exhibited by the two diVerent classi-
Ž cation algorithms adopted was statistically signiŽ cant (p 0.05). ClassiŽ cation of
Landsat TM imagery produced the most accurate result with associated overall
Figure 3. Vegetation associations derived from classiŽ cation of satellite imagery.

2920
K. R. Harvey and G. J. E. Hill
mapping accuracies of 90% and 86% for the minimum distance to mean and isodata
algorithms respectively. ClassiŽ cation of the SPOT XS image resulted in a higher
mapping accuracy of 85% using the isodata algorithm and 82% using the minimum
distance to mean algorithm, while analysis of the SPOT PAN image resulted in
mapping accuracies of 71% and 75% for the minimum distance to mean and isodata
algorithms respectively. The results for each image classiŽ cation are summarised
in table 1.
5. Discussion
Aerial photography was clearly superior to satellite imagery for detailed mapping
of vegetation communities in the tropical wetland environment studied. The superior-
ity of large-scale aerial photography was not an unexpected result and was
undoubtedly enhanced through the enlargement of photography (300%) within the
geographic information system prior to visual interpretation. In addition to the
large-scale, it is also noted that the utilisation of interpretation cues, such as context
and texture, enhance the class delineation and identiŽ cation capabilities of aerial
photography. These are not extensively utilised in digital image processing (e.g.
Ryerson 1989, Franklin 1993, Ahmad and Hill 1994). The use of contextual and
textural characteristics to delineate vegetation communities is desirable in wetland
environments, especially those with a highly heterogeneous structural composition
where similar vegetation communities occur in diVerent densities.
Comparisons of satellite image and aerial photo classiŽ cations illustrated that at
a structural level (three classes only) aerial photography does not facilitate a signiŽ c-
antly higher level of mapping accuracy. Whilst there is no dispute that aerial photo-
graphy allows for a far more detailed classiŽ cation, the usefulness of this will depend
on the application. When aiming to identify the key structural components of
Melacca Swamp, the resolution of the remotely sensed datasets used did not have a
large impact on the accuracy of the Ž nal classiŽ cation. The results suggest that either
Landsat TM or SPOT XS imagery is adequate for mapping these generalised
landcover classes.
With regard to the satellite imagery, the broader spectral range of the Landsat
TM data appeared to more than compensate for the superior spatial resolution of
the SPOT imagery. Spectral overlap between wetland cover types and wetland and
upland cover types is a problem frequently identiŽ ed in the application of remote
sensing techniques to wetland environments (e.g. Johnston and Barson 1993, Sader
et al. 1995). Whilst spectral overlap did have an adverse impact on mapping accuracy
in this study the impact of this overlap was most signiŽ cant in the classiŽ cations
derived from SPOT satellite imagery with inferior spectral resolution. The superior
classiŽ cation accuracies achieved in analysis of Landsat TM imagery data therefore
appear to be a function of the data being collected in a wider range of the electromag-
netic spectrum, especially in the longer wavelengths which facilitate improved vegeta-
tion discrimination (Lee and Lunetta 1994). In particular, the utilization of the
middle infrared band (1.55–1.75 nm), which was not available in the SPOT data,
provides improved discrimination of vegetation types, especially in wetland environ-
ments (e.g. Cowardin and Myers 1974, Harrison and Jupp 1993, Sader et al. 1995).
An unsupervised classiŽ cation technique was adopted in this study, as theoretic-
ally it was most suited to application in a highly heterogeneous wetland environment.
Supervised classiŽ cation often yields maps with a higher mapping accuracy (e.g.
Johnston and Barson 1993, Lyon 1983) however, this technique is not suited to some

Vegetation mapping of a tropical freshwater swamp
2921

2922
K. R. Harvey and G. J. E. Hill
applications. The most useful classiŽ cation technique will depend on the application
and is closely related to the physical characteristics of the features being mapped.
Melacca Swamp and most wetland environments in general, consist of a highly
heterogeneous arrangement of cover types. This makes supervised classiŽ cation
techniques unsuited to mapping these areas, as statistically homogenous training
areas can not be deŽ ned. A supervised classiŽ cation was attempted in this study
using aerial photography as a guide for the selection of training areas. However, the
technique was abandoned when the training areas selected for each of the cover
types showed a signiŽ cant amount of spectral overlap.
One possible source of error in this study is related to the image acquisition
characteristics. A number of remote sensing studies in wetland environments have
concluded that spectral overlap may reduce the utility of imagery collected in certain
seasons but spectral separability may be improved by optimizing image acquisition
(Jensen et al. 1984, Catt and Thirarongnarong 1992, Johnston and Barson 1993).
This is especially an issue in tropical wetland environments, which exhibit large
annual variation due to the highly seasonal nature of rainfall. These factors may
have in uenced the accuracy of the vegetation classiŽ cations derived in this study.
All satellite imagery utilised in this study was acquired in the ‘dry’ season, which
theoretically increases spectral separability due to the reduced in uence of water on
the spectral response of vegetation cover types. However, the Landsat TM imagery
was acquired early in a ‘dry’ season, which followed an above average wet season
rainfall recording for the Top End. In theory this should disadvantage the classiŽ ca-
tion based on the Landsat TM satellite imagery (Harvey et al. 1999), however,
superior results were attained over SPOT, which was recorded near the end of the
‘dry’ season (October). The use of Landsat TM imagery acquired when the landscape
is drier may further enhance the utility of these data.
Limitations on data availability are also an issue when mapping remote tropical
wetland environments. In this study the aerial photography utilised, despite being
the most recent available, was captured 15 years before the Ž eld data were gathered.
The low frequency of air photo coverage for the region highlights a major problem
for mapping and monitoring of tropical wetlands. In the current study the impact
was considered to be minimal as the team involved in the annual monitoring of
crocodile nesting activity within the swamp report that  oristic structure has not
altered signiŽ cantly since the 1980s. The high level of agreement between the map
produced from the 1982 photography and the current Ž eld data support this observa-
tion. However, this is not always the case and often the best available photographic
data are panchromatic and at scales of 1:25 000 or 1:50 000. Very poor discrimination
of vegetation communities is available from these data, as is borne out by the results
obtained for the SPOT PAN imagery. Given the cost of commissioning and analyzing
task-speciŽ c air photo coverage of the Top End wetlands, the utility of satellite
imagery is therefore a key issue.
The main aim of this research was to determine the utility of remote sensing for
mapping vegetation cover types within a tropical freshwater swamp. Aerial photo-
graphy provided the level of resolution required for a detailed  oristic study. This
level of detail will be necessary for speciŽ c applications, for example, mapping the
habitat of waterbirds like the magpie goose (Anseranus semipalmate) which depends
on speciŽ c vegetation types (grasslands) during the early dry season and diVerent
vegetation types (sedgelands) later in the dry season. At a more general level, however,
the high levels of accuracy achieved for classiŽ cation of Landsat TM satellite imagery

Vegetation mapping of a tropical freshwater swamp
2923
suggest that satellite data provide an adequate description of the major landcover
categories characterising this tropical freshwater wetland. The major shortcoming of
satellite imagery in this study was its inability to distinguish between diVerent
understory types and canopy densities within the wetland environment. This may
restrict the usefulness of the satellite data for some applications, for example, delineat-
ing hollow trees to provide an assessment of suitability for birds that nest in tree
cavities, but not others.
While aerial photography can provide information that cannot be extracted from
satellite data, its use is restricted based on the cost of acquiring data over large
areas, the time involved in analyzing large datasets and, in many countries, its
restricted availability. Therefore, where the required information can be extracted
from satellite imagery it appears to be the most feasible technology to adopt. The
utility of aerial photography in detailed  oristic studies of wetland environments,
however, should not be entirely discounted as it remains the most suitable data for
small-scale studies and often provides additional information to aid in the analyses
of satellite data. As wetland areas come under increasing pressure from human
utilization, the ability to map and monitor the status of wetland vegetation commu-
nities will be crucial to the conservation of biodiversity. The results of this research,
conducted in northern Australia, suggest that remote sensing can aid in management
and conservation. Both satellite imagery and aerial photography have a role to play
depending on the speciŽ c requirements of the research problem. As has been sug-
gested by many previous studies, a multi-level approach where resolution of the
imagery is logically linked to the scale of landscape features will provide the most
eYcient and cost eVective methodology.
Acknowledgments
This research was conducted under sponsorship from the Australian Research
Council, Strategic Partnerships with Industry Research and Training scheme. The
industry partner, Wildlife Management International Pty Ltd, provided Ž eld support
and expertise. Brett Ottley, Bryan Baker and Charles Manolis were involved in most
facets of the research. Dr Peter Whitehead is thanked for his comments and sugges-
tions on the manuscript. Results will contribute to the program of the Tropical
Savannas Cooperative Research Centre and the Northern Territory University,
Centre for Tropical Wetlands Management.
References
Afek, Y., and Brand, A., 1998, Mosaicking of orthorectiŽ ed aerial images. Photogrammetric
Engineering and Remote Sensing, 64, 115–125.
Ahmad, W., and Hill, G. J. E., 1994, A classiŽ cation strategy for mapping trochus shell
habitat in Torres Strait, Australia. Geocarto International, 3, 39–47.
Ahmad, W., and Neil, D. T., 1994, An evaluation of Landsat Thematic Mapper (TM) digital
data for discriminating coral reef zonation: Herron Reef (GBR). International Journal
of Remote Sensing, 15, 2583–2597.
Anderson, J. E., and Perry, J. E., 1996, Characterization of wetland plant stress using leaf
spectral re ectance: Implications for wetland remote sensing. WET L ANDS, 16,
477–487.
Brady, A., Shaikh, M., King, A., and Sharma, P., 1997, Remote sensing and the Great
Cumbung Swamp. Wetlands Australia, 7, 6–7.
Campbell, J. B., 1996, Introduction to Remote Sensing, 2nd edn (London: Taylor and
Francis Ltd).

2924
K. R. Harvey and G. J. E. Hill
Catt, P., and Thirarongnarong, K., 1992, An evaluation of remote sensing techniques for
the detection, mapping, and monitoring of invasive plant species in coastal wetlands:
a case study of Para Grass (Brachiaria mutica). Proceedings of the 6th Australasian
Remote Sensing Conference, Wellington, New Zealand, 2–6 November 1992 (Committee
of the 6th Australasian Remote Sensing Conference Incorporated: Wellington, New
Zealand), pp. 200–203.
Congalton, R. G., Stenback, J. M., and Barrett, R. H., 1993, Mapping deer habitat
suitability using remote sensing and geographic information systems. Geocarto
International, 3, 23–33.
Cowardin, L. M., and Myers, V. I., 1974, Remote sensing for the identiŽ cation and classiŽ ca-
tion of wetland vegetation. Journal of W ildlife Management, 38, 308–314.
Dale, P. E. R., Chandica, A. L., and Evans, M., 1996, Using image subtraction and
classiŽ cation to evaluate change in sub-tropical intertidal wetlands. International
Journal of Remote Sensing, 17, 703–719.
Dottavio, L. C., and Dottavio, D. F., 1984, Potential beneŽ ts of new satellite sensors to
wetland mapping. Photogrammetric Engineering and Remote Sensing, 50, 599–606.
ERDAS, 1997, ERDAS Field Guide (Atlanta: ERDAS Inc.).
Ernst-Dottavio, C. L., Hoffer, R. M., and Mroczynski, R. P., 1981, Spectral characteristics
of wetland habitats. Photogrammetric Engineering and Remote Sensing, 47, 223–227.
Finlayson, C. M., Bailey, B. J., Freeland, W. J., and Fleming, M. R., 1988, In T he
Conservation of Australian Wetlands, edited by A. McComb and P. Lake (Sydney:
Surrey Beatty and Sons), pp. 103–126.
Franklin, J., 1993, Discrimination of tropical vegetation types using SPOT multispectral
data. Geocarto International, 2, 57–63.
Haack, B., 1996, Environmental auditing: monitoring wetland changes with remote sensing:
an east African example. Environmental Mangement, 20, 411–419.
Harrison, B. A., and Jupp, D. L. B., 1993, MicroBrian Resource Manual (Canberra: CSIRO,
Water Resources).
Harrison, B. A., and Jupp, D. L. B., 1990, Introduction to Image Processing—Part T wo of
the MicroBrian Resource Manual (Canberra: CSIRO Publications).
Harvey, K. R., Duffett, M. L., and Hill, G. J. E., 1999, Identifying and mapping saltwater
crocodile (Crocodylus porosus) nesting habitats in the Northern Territory using remote
sensing and GIS. Proceedings of the 4th North Australian Remote Sensing and GIS
Conference. Darwin, Northern T erritory, 28–30 June 1999 (Northern Territory
#
University), Paper 40 (CD ROM).
Hefner, J. M., and Storrs, C. G., 1994, ClassiŽ cation and inventory of wetlands in the
southern Appalachian region. Water, Air and Soil Pollution, 77, 209–216.
Jaensch, R. P., 1993, Section 5—Northern Territory—Adelaide River  oodplain system. In
A Directory of Important Wetlands in Australia, 1st edn, edited by S. Ushback and
R. James (Canberra: Australian Nature Conservation Agency), pp. 5–8.
Jennnings, C. A., Vohs, P. A., and Dewey, M. R., 1992, ClassiŽ cation of a wetland area
along the upper Mississippi River with aerial videography. WET L ANDS, 12, 163–170.
Jensen, J. R., Narumalani, S., Weatherbee, O., and Halkard, E. M., 1993, Measure-
ment of seasonal and yearly cattail and waterlily changes using multidate SPOT
panchromatic data. Photogrammetric Engineering and Remote Sensing, 59, 519–525.
Jensen, J. R., Christensen, E. J., and Sharitz, R., 1984, Nontidal wetland mapping in South
Carolina using airborne multispectral scanner data. Remote Sensing of Environment,
16, 1–12.
Johnston, R. M., and Barson, M. M., 1993, Remote sensing of Australian wetlands: an
evaluation of Landsat TM data for inventory and classiŽ cation. Australian Journal of
Marine and Freshwater Research, 44, 223–232.
Lee, K. H., and Lunetta, R. S., 1995, Wetlands detection methods. In Wetlands and
Environmental Applications of GIS, edited by J. G. Lyon and J. McCarthy (USA: CRC
Press. Inc), pp. 249–283.
Lyon, J. G., 1983, Landsat derived land-cover classiŽ cations for locating potential Kestrel
nesting habitat. Photogrammetric Engineering and Remote Sensing, 49, 245–250.
Mattheson, W., 1994, The use of Landsat MSS imagery to determine plant characteristics
and community spectral separability from sub-tropical to arid zones in northern
Australia. Geocarto International, 2, 5–13.

Vegetation mapping of a tropical freshwater swamp
2925
Melack, J. M., Hess, L. L., and Sippel, S., 1994, Remote sensing of lakes and  oodplains in
the Amazon Basin. Remote Sensing Reviews, 10, 127–142.
Menges, C. H., Hill, G. J. E., and Ahmad, W., 1998, Landsat TM data and potential feeding
grounds for threatened marine turtle species in northern Australia. Int. J. Remote
Sensing, 19, 1207–1221.
Mumby, P. J., Green, E. P., Edwards, A. J., and Clark, C. D., 1999, The cost-eVectiveness
of remote sensing for tropical coastal resources assessment and management. Journal
of Environmental Management, 55, 157–166.
Ozesmi, U., and Mitsch, W. J., 1997, A spatial habitat model for the marsh-breeding Red-
winged Blackbird (Agelaius phieniceus) in coastal Lake Erie wetlands. Ecological
Modelling, 101, 139–152.
Parks and Wildlife Commission of the Northern Territory, 1996, A Management
Program for Crocodylus porosus and Crocodylus johnstoni—in the Northern Territory
of Australia (Darwin: Parks and Wildlife Commission of the Northern Territory).
Phinn, S. R., and Hill, G. J. E., 1998, An integrated multiscale approach for assessing
th
structure and condition of wetland environments in Australia. Proceedings of the 9
Australasian Remote Sensing and Photogrammetry Conference, Sydney, Australia, 20–24
July 1998 (Causal Productions: Sydney), Paper number 40 (CD ROM).
Pidsley, D., Chin, D., and Sanders, D., 1993, Use of SPOT imagery to map the extent of
tidal intrusion into the Mary River  oodplain. NARGIS 93—Proceedings of the North
Australian Remote Sensing and GIS Forum, Darwin, Australia, 9–11 August 1993
(Canberra: Australian Government Publishing Service), pp. 222–229.
Rutchey, K., and Vilcheck, L., 1994, Development of an Everglades vegetation map using
a SPOT image and the Global Positioning System. Photogrammetric Engineering and
Remote Sensing, 60, 767–775.
Ryerson, R., 1989, Image interpretation concerns for the 1990s and lessons from the past.
Photogrammetric Engineering and Remote Sensing, 55, 1427–1430.
Sader, S. A., Ahl, D., and Wen-Shu, L., 1995, Accuracy of Landsat TM and GIS rule-based
methods for forest wetland classiŽ cation in Maine. Remote Sens. Environ, 53, 133–144.
Scarpace, F. L., Quirk, B. K., Keifer, R. W., and Wynn, S. L., 1981, Wetland mapping
from digitized aerial photography. Photogrammetric Engineering and Remote Sensing,
47, 829–838.
Stenback, J. M., and Congalton, R. G., 1990 Using Thematic Mapper imagery to examine
forest understory. Photogrammetric Engineering and Remote Sensing, 56, 1285–1290.
Storrs, M. A., and Finlayson, M., 1997, Overview of the conservation status of wetlands of
the Northern Territory. Supervising Scientist Report Number 116 (Supervising
Scientist: Barton).
Walsh, S. J., Cooper, J. W., Von Essen, I. E., and Gallager, K. R., 1990, Image enhancement
of Landsat Thematic Mapper data and GIS data integration for evaluation of resource
characteristics. Photogrammetric Engineering and Remote Sensing, 56, 1135–1141.
Webb, G. J. W., Sack, G. C., Buckworth, R., and Manolis, C., 1983, An examination of
Crocodylus porosus nests in two northern Australian freshwater swamps, with an
analysis of embryo mortality. Australian W ildlife Research, 10, 571–605.
Whitehead P., and Chatto, R., 1995, Northern Territory—Adelaide River  oodplain system.
In A Directory of important wetlands in Australia—2nd edition, edited by R. Blackley,
S. Usback and K. Langford (Canberra: Australian Nature Conservation Agency),
pp. 151–154.
Whitehead P. J., Wilson B. A., and Bowman, D. J. M., 1990, The conservation of coastal
wetlands of the Northern Territory of Australia. Biodiversity Conservation, 52, 85–111.
Young, R. D., and Dahl, T. E., 1994, Use of GIS in assessing areas of rapid wetland change.
http://wwwsgi.ursus.maine.edu/gisweb/spatdb/gis-lis/gi94105.html