You are on page 1of 22

Wetlands Ecology and Management 10: 381–402, 2002.

381
 2002 Kluwer Academic Publishers. Printed in the Netherlands.

Satellite remote sensing of wetlands


1,2,
Stacy L. Ozesmi * and Marvin E. Bauer 1
1 2
Department of Forest Resources, University of Minnesota, St Paul, Minnesota 55108, USA; Current
address: Environmental Engineering Department, Environmental Science Branch, Erciyes University School
of Engineering, Kayseri 38039, Turkey; * Author for correspondence (e-mail: stacy@ erciyes.edu.tr; phone:
90 (0.352) 437 4937 ext. 32800; fax: 90 (0.352) 437 4404)

Received 17 January 2001; accepted in revised form 5 June 2002

Key words: Classification techniques, Comparison of methods, Remote sensing, Satellite imagery, Wetland
classification, Wetland identification, Wetlands inventory, Wetlands

Abstract

To conserve and manage wetland resources, it is important to inventory and monitor wetlands and their adjacent
uplands. Satellite remote sensing has several advantages for monitoring wetland resources, especially for large
geographic areas. This review summarizes the literature on satellite remote sensing of wetlands, including what
classification techniques were most successful in identifying wetlands and separating them from other land cover
types. All types of wetlands have been studied with satellite remote sensing. Landsat MSS, Landsat TM, and
SPOT are the major satellite systems that have been used to study wetlands; other systems are NOAA AVHRR,
IRS-1B LISS-II and radar systems, including JERS-1, ERS-1 and RADARSAT. Early work with satellite imagery
used visual interpretation for classification. The most commonly used computer classification method to map
wetlands is unsupervised classification or clustering. Maximum likelihood is the most common supervised
classification method. Wetland classification is difficult because of spectral confusion with other landcover classes
and among different types of wetlands. However, multi-temporal data usually improves the classification of
wetlands, as does ancillary data such as soil data, elevation or topography data. Classified satellite imagery and
maps derived from aerial photography have been compared with the conclusion that they offer different but
complimentary information. Change detection studies have taken advantage of the repeat coverage and archival
data available with satellite remote sensing. Detailed wetland maps can be updated using satellite imagery. Given
the spatial resolution of satellite remote sensing systems, fuzzy classification, subpixel classification, spectral
mixture analysis, and mixtures estimation may provide more detailed information on wetlands. A layered, hybrid
or rule-based approach may give better results than more traditional methods. The combination of radar and
optical data provide the most promise for improving wetland classification.

Introduction grazing, and recreation. Wetlands provide habitat for


fish and wildlife, supporting a rich biodiversity, in-
Extensive loss of wetlands has occurred in many cluding many threatened and endangered species.
countries throughout the world (Mitsch and Gosselink Wetlands are also valued for their recreational oppor-
1993). As the value of wetlands to society has become tunities and aesthetics (Barbier et al. 1994). To pre-
recognized, it is now important to conserve these vent further loss of wetlands, and conserve existing
valuable resources. Wetlands are valued for their wetland ecosystems for biodiversity and ecosystem
ability to store floodwaters, protect shorelines, im- services and goods, it is important to inventory and
prove water quality, and recharge groundwater aquif- monitor wetlands and their adjacent uplands.
ers (Daily 1997). In many countries, local economies For inventorying and monitoring wetlands, satellite
depend on wetlands for fisheries, reed harvesting, remote sensing has many advantages. Satellite data
382

has repeat coverage so that wetlands can be monitored


seasonally or yearly. Satellite remote sensing can also
provide information on surrounding land uses and
their changes over time. Current information on the
uplands surrounding wetlands is important because
land use practices in uplands cause loss of wetland
functions, goods, services and values. Satellite data
are in digital format and relatively easy to integrate
into a geographic information system (GIS). Using
satellite remotely sensed data for land cover classifi-
cation is less costly and less time-consuming than
aerial photography for large geographic areas. Satel-
lite remote sensing can be especially appropriate for
wetland inventories and monitoring in developing
countries, where funds are limited and where little
information is available on wetland areas, surround-
ing land uses, and wetland losses over time.
However, satellite imagery also has limitations
compared to aerial photography. Because of the spa-
tial resolution of most satellite imagery (20 – 30 m), it
is difficult to identify small or long, narrow wetlands. Figure 1. Example of the variability in spectral signatures of
different wetland types plotted for the near infrared and red bands.
Also fewer types of wetlands can be identified com- Note how Phragmites marsh, Juncus marsh, and wet meadow have
pared to aerial photography. It is difficult to separate overlapping signatures. Open water, however, is spectrally very
different wetland types from one another because of separable from the others.
the overlap in their spectral signatures (Gluck et al.
1996). Spectral signatures are measurements of the
spectral response of different features in the bands of
the remote sensor. Figure 1 shows the spectral signa-
tures of different types of wetland vegetation. Wet-
lands can also be confused with upland landcover
classes because of overlapping spectral signatures.
Figure 2 shows an example of the spectral signatures
of wetlands and other landcover classes. The overlap
in spectral signatures between wetlands and other
classes such as agricultural crops and upland forests
limit accuracy in mapping with satellite imagery. In
particular, it is hard to detect wetland types with drier
water regimes, such as forested wetlands, scrub-
shrub, and emergent wetlands, because they are easily
confused with upland vegetation (FGDC 1992). In
order to separate wetlands from uplands, it is usually
advantageous to use satellite images from dates when
the wetlands are at their highest water levels. How-
ever, many wetlands are only flooded at certain times
during the year. Given a satellite’s fixed orbit and
return interval, it is difficult to capture the optimum
water conditions for wetland detection. In addition,
lack of cloud free data often prevents the use of
Figure 2. Example of the variability in spectral signatures of
optimal dates for wetland mapping. different classes plotted for the red and near infrared bands. Note
Because of these disadvantages, the National Wet- that water is spectrally distinct but there is overlap between wetland
land Inventory (NWI) project, began in 1975 by the and agricultural and forest classes.
383

United States Fish and Wildlife Service for conserv- inventory and maintaining a database. MacDonald
ing and managing wetlands, chose high altitude aerial (1999) discussed the current and potential uses of
photography over satellite imagery (Tiner 1990). aerial photography and satellite imagery in wetland
Rutchey and Vilchek (1999) also concluded that rehabilitation projects including monitoring change
photo interpretation was preferred for mapping vege- following rehabilitation projects, understanding
tation in the Florida Everglades, even though they causes of wetland degradation, prioritizing potential
were able to classify SPOT satellite data into 12 rehabilitation sites, weed management, permit com-
wetland classes with 80.9% accuracy (Rutchey and pliance, and site design. Klemas (2001) discussed
Vilcheck 1994). Satellite image classification was current and future satellite remote sensing systems for
limited by four factors. The first factor was fluctuating monitoring environmental indicators in coastal areas.
water levels, which change the spectral reflectance of This review summarizes the literature on satellite
the vegetation. Second, fires left visible scars that remote sensing of wetlands, with an emphasis on the
could be observed on both satellite imagery and aerial different classification techniques that have been used
photography for up to two years after a fire occurred. to identify wetlands. As described above, wetland
These burned areas were often misclassified as open classification with satellite imagery is difficult. How-
water / slough community on satellite imagery. Third, ever, this paper will give the reader insight into the
periphyton formed large floating masses around wet- best classification techniques to use. This review
land vegetation in late summer, causing confusion in examines recent improvements that have been made
the satellite image classification. Fourth, Typha spp. in wetland classification using satellite imagery. Wet-
had variations in growth pattern, which complicate land managers and conservationists will find this
satellite imagery classification. review a helpful resource for understanding how
Thus aerial photography is generally preferred for satellite remote sensing has been used to study wet-
detailed mapping of wetlands, especially if many lands in the past and what classification techniques
different vegetation types must be mapped. However, hold the most promise for the future.
satellite imagery has advantages over aerial photo- Organization of this review is as follows. Almost
graphy for continued monitoring of wetland re- every type of wetland has been studied using satellite
sources. Satellite data are especially appropriate for remote sensing and these wetlands are listed. The
large geographic areas. In addition, using satellite satellite sensors, Landsat, SPOT, AVHRR, Indian
remote sensing data together with more detailed maps Remote Sensing Satellite (IRS) and radar systems,
or aerial photography can give greater information that have been used to study wetlands and their
than either source alone. Together they can be used to advantages and disadvantages are described. Studies
monitor changes in water levels in wetlands and land of the spectral reflectance of various wetland vegeta-
cover change on adjacent uplands. Satellite data can tion types are included because most computer classi-
be used to determine where frequent changes are fication methods are dependent on different spectral
occurring and where more detailed maps should be responses of wetland vegetation types for classifica-
updated frequently. tion. Classification techniques, with selected exam-
Previous reviews on remote sensing of wetlands ples from the literature, are investigated to determine
include Lulla (1983), who reviewed the use of Land- in general which were most successful in identifying
sat MSS satellite imagery for surveying wetlands, wetlands and separating them from other land cover
coastal ecosystems, and aquatic environments. Hard- types. The accuracy for wetland classification is gen-
isky et al. (1986) reviewed different remote sensing erally included to give an idea of the method’s suc-
platforms, including aerial photography, SPOT, Land- cess, however the reader should exercise caution with
sat MSS, Landsat TM and AVHRR, for mapping these figures. There is not a standard accuracy assess-
coastal wetlands and estimating biomass. The FGDC ment procedure, and different studies use different
(1992) summarized the experience of several agen- procedures, thus comparing among studies is difficult.
cies on the use of satellite imagery for mapping and However, generally it was found that multi-temporal
monitoring wetlands. Lee and Lunetta (1996) re- imagery and ancillary information such as soils, ele-
viewed methods to inventory or detect changes in vation, or NWI data, improved wetland classification.
wetlands, including airborne and satellite sensors and Change detection studies of wetlands using satellite
their ability to detect wetlands and monitor changes, imagery are also reviewed because continued moni-
and the costs associated with creating a wetlands toring of wetland resources is important to ensure that
384

wetland functions are maintained. Finally the classifi- temporal data to improve classification. The combina-
cation techniques with the most promise for identify- tion of two dates of imagery allowed separation of
ing and classifying wetlands in the future are summa- emergent and floating vegetation (winter and spring)
rized. These included techniques that can provide and flooded emergent vegetation and open water (fall
detail on wetland vegetation types at the subpixel and winter).
level, a rule-based change detection method for de- Mangrove and other coastal wetlands were studied
termining changes to wetlands, and the combination in West Bengal India by Kushwaha et al. (2000) and
of radar and optical data for wetland mapping. by Butera (1983) in the southern US. Given appro-
priate tidal conditions, mangroves were relatively
easy to classify because they had a unique spectral
Different types of wetlands studied using reflectance (Butera 1983). The optimum time for
satellite remote sensing detecting mangroves was considered to be late sum-
mer to early fall, when the vegetation was mature and
Satellite remote sensing has been used to study all cloud cover was less frequent than in the winter.
types of wetlands. Generally the easiest wetland Bogs and fens, or northern peatlands, have also
classes to detect are permanently flooded or intermit- been mapped with satellite remote sensing (Franklin
tently exposed open water ponds (palustrine uncon- et al. (1994), Gluck et al. (1996) (fens); Palylyk et al.
solidated bottoms) (FGDC 1992). The order of identi- (1987), Pietroniro et al. (1996), Wickware and
fication difficulty from easiest to hardest is water, Howarth (1981)). Landsat MSS data from September
marshes (wettest to driest), deciduous forested wet- and May for a peatland inventory in Alberta allowed a
lands, evergreen forested wetlands, and scrub-shrub separation of deciduous and coniferous cover types
wetlands, respectively. Multi-temporal imagery often (Palylyk et al. 1987).
aids in classification of wetlands and their separation Inland freshwater marshes have been another fre-
from other landcover classes. The dates of imagery, quently studied wetland type (i.e. Ernst and Hoffer
which are most helpful, depend on the types of (1979), Ernst-Dottavio et al. (1981), Forgette and
wetland studied and the surrounding landcover classes Shuey (1997), Lunetta and Barlogh (1999)). In the
as described below. southeastern US, spring imagery was optimal for
Satellite remote sensing has been used to map large wetland discrimination (Jensen et al. 1984). SPOT
wetland ecosystems, including the marshes and multi-spectral images from early spring leaf off con-
swamps of Harike wetland ecosystem in Punjab, ditions (May), for locating wet soils and standing
India, (Chopra et al. 2001) and Sango Bay on Lake water, and summer leaf on conditions (June), for
Victoria in Uganda, which includes papyrus swamps, identifying different vegetation types, were used to
tall grass swamps, riverine and swamp forests (Fuller map wetland areas in a small watershed in northern
et al. 1998). Satellite remote sensing was used to Michigan (Forgette and Shuey 1997). Lunetta and
study energy fluxes of the internationally important, Barlogh (1999) also found that leaf off (April) and
Esteros del Ibera in Argentina, one of South leaf on (June) image dates improved classification of
America’s largest wetlands which consists of a wetlands in Maryland and Delaware. Wet meadows
mosaic of marshes, swamps and lagoons (Louiselle et were studied by Kindscher et al. (1998).
al. 2001). The Great Dismal Swamp (Carter et al. Other inland wetlands studied with satellite remote
1977; Gammon et al. 1979), the Okefenokee Swamp sensing include the Prairie Pothole region of the US
(Lo and Watson 1998), and the Florida Everglades in and Canada (Best and Moore 1979; Gilmer et al.
the United States have been studied using satellite 1980; Work and Gilmer 1976) and the Nebraska
remote sensing (Doren et al. 1999; Hines et al. 1993; Sandhills (Goodin 1995). In Prairie Pothole region,
Rutchey and Vilcheck 1994; Rutchey and Vilchek late April or very early May images in normal precipi-
1999; Welch et al. 1999, 1995). tation years were best for delineating wetlands. How-
Coastal tidal marshes are the type of wetlands that ever at these times it was difficult to identify the
have been most frequently studied using satellite wetland’s water regime. Later in the year it might
remote sensing (i.e. Hardisky et al. (1986), Hinson et have been easier to identify the water regime but it
al. (1994), Jensen et al. (1993a), Lee and Park (1992), was harder to detect the wetlands. Mid-May to June
Nayak and Sahai (1985), Ramsey and Laine (1997)). data was helpful for separation of wetland types (Best
Ramsey and Laine (1997) took advantage of multi- and Moore 1979; Gilmer et al. 1980; Work and
385

Gilmer 1976). A winter scene with snow on the Indian Remote Sensing Satellite (IRS-1B) Linear
ground was helpful to identify stands of emergent Imaging Self-scanning Sensor (LISS-II) are satellite
vegetation (Best and Moore 1979). remote sensing systems that have been used to study
Studied areas included forested wetlands, or wetlands. These sensors detect and measure electro-
swamps (Ernst and Hoffer 1979; Ernst-Dottavio et al. magnetic radiation in the optical portion of the spec-
1981; Llewellyn et al. 1996; Sader et al. 1995) includ- trum. Table 1 compares the characteristics of these
ing those in the Brazilian Amazon (Mertes et al. sensors. Radar systems, which transmit and receive
1995). Southern US deepwater swamps, which are radiation in the microwave portion of the electro-
dominated by cypress, Taxodium, and gum-tupelo, magnetic spectrum, have also been used to study
Nyssa, were studied (Butera 1983; Huguenin et al. wetlands. Table 2 lists the characteristics of radar
1997; Ramsey et al. 1998). October and January sensors.
Landsat MSS images were used to map Savannah
River wetlands, which included cypress dominant and Landsat MSS
tupelo gum dominant forested wetlands, tidal marshes
and freshwater marshes (Butera 1983). The multi- NASA launched the first Landsat, originally known as
temporal images were good for discriminating be- Earth Resources Technology Satellite (ERTS), in
tween agricultural and marsh classes while the winter 1972. Landsat-1, -2, and -3 had a Multi-spectral
image was useful for discriminating between ever- Scanner (MSS) with four bands (Table 1). Many
green and deciduous forest species. Riparian areas in researchers used Landsat MSS to study wetlands (i.e.
arid regions have also been mapped with satellite Ackleson and Klemas (1987), Ackleson et al. (1985),
remote sensing (Hewitt 1990; Lee and Marsh 1995). Butera (1983), Gomarasca et al. (1992), Haack
In open water areas, submerged aquatic vegetation (1996), Haack and Messina (1997), Harper and Ross
was mapped with remotely sensed satellite imagery (1982), Hutton and Dincer (1979), Jensen et al.
(Ackleson and Klemas 1987; Luczkovich et al. 1993; (1986), Lee and Marsh (1995), Nayak and Sahai
Macleod and Congalton 1998; Zainal et al. 1993). The (1985), Palylyk et al. (1987), Wickware and Howarth
ability of satellite imagery to identify submerged (1981)). Most research has concluded that Landsat
aquatic vegetation was influenced by wind conditions, MSS data are useful for spectral discrimination of
depth of water and turbidity. Best results were ob- large vegetated wetlands (Jensen et al. 1984). How-
tained at low tide with no wind and no turbidity. ever, since the launch of Landsat TM and other
satellites with improved capabilities, the use of Land-
sat MSS for mapping wetlands has been mostly
Satellite sensors used in wetland identification limited to historical or change detection studies where
and classification earlier dates of imagery are needed and for studies of
large wetlands where cost-effective imagery is
Landsat MSS, Landsat TM, SPOT, AVHRR, and needed.

Table 1. Comparison of spectral, temporal and spatial resolution of Landsat MSS, Landsat TM, SPOT, AVHRR, and IRS-1B LISS-II imagery.
Resolution Landsat MSS Landsat 4, 5 TM SPOT-1,2,3,4 AVHRR IRS-1B LISS-II
Spectral (mm)
Band 1 0.45–0.52 0.51–0.73 PAN 0.58–0.68 0.45–0.52
0.50–0.59 HRV
Band 2 0.52–0.60 0.61–0.68 HRV 0.72–1.10 0.52–0.59
Band 3 0.63–0.69 0.79–0.89 HRV 3.55–3.93 0.62–0.68
Band 4 0.5–0.6 0.76–0.90 10.3–11.3 0.77–0.86
Band 5 0.6–0.7 1.55–1.75 1.58–1.75 SPOT-4 only 11.5–12.5
Band 6 0.7–0.8 10.4–12.5
Band 7 0.8–1.1 2.08–2.35
Radiometric (bits) 64 256 256 1024 128
Temporal 18 days 16 days 26 days 12 hours 22 days
Spatial 80 m 30 m, 120 m for band 6 20 m HRV, MIR 10 m PAN 1.1 km at nadir 36.25 m
First Launched 1972 1982 1986 1979 1991
386

Landsat TM which was launched in 1998, also has a middle


infrared band in addition to the other bands. Many
The Landsat Thematic Mapper (TM) sensor, first wetland studies used SPOT data (i.e. Cairns et al.
launched on Landsat 4 in 1982, improved spectral, (1997), Forgette and Shuey (1997), Houhoulis and
radiometric, temporal and spatial resolution over Michener (2000), Jensen et al. (1993c), Lee and Park
Landsat MSS. These improvements made Landsat (1992), McCarthy et al. (1993), Narumalani et al.
TM more useful for identifying wetlands as well as (1997), Rutchey and Vilcheck (1994), Rutchey and
other land cover types. Many studies used Landsat Vilchek (1999)). Although not used for wetland
TM imagery for wetland mapping (i.e. Ackleson and identification per se, SPOT panchromatic images
Klemas (1987), Franklin et al. (1994), Gluck et al. provided a coordinate reference layer for registering
(1996), Hewitt (1990), Jensen et al. (1993a, 1986), aerial photographs in a GIS for South Florida’s Na-
Mertes et al. (1995), Pietroniro et al. (1996), Pope et tional Parks and Preserves in the US (Welch et al.
al. (1994), Ramsey and Laine (1997), Wilkinson and 1995).
Shepherd (1995), Zainal et al. (1993))
The most important Landsat TM band for wetland
AVHRR
identification is band 5 because of its ability to
discriminate vegetation and soil moisture levels.
The spatial resolution of NOAA Advanced Very High
Landsat TM bands 3, 4, and 5 are usually the best
Resolution Radiometer (AVHRR) makes it unsuitable
combination of bands for wetland detection. Band 1
for most wetland studies. However, some researchers
should be added if water quality is of interest (FGDC
have used AVHRR to study forested wetlands. Ram-
1992). The middle infrared Landsat TM bands pro-
sey et al. (1997, 1998) used NOAA AVHRR images
vide for much of the separability between wetland
to examine hurricane impact and recovery of forested
types (Jensen et al. 1993a).
wetlands in Louisiana. Llewellyn et al. (1996) created
Although the spatial resolution of Landsat TM
a decision-support system for prioritizing restoration
imagery is 30 m, in practice it takes a ‘‘window’’ of
sites on the Mississippi River alluvial plain. For the
nine pixels or approximately 0.9 ha to consistently
entire Mississippi River alluvial plain, wetland forest
identify an object. However, if the object is not
was estimated using AVHRR satellite images (1.1
square, or if adjacent land covers have similar spectral
km 2 resolution). AVHRR estimates were remarkably
reflectances, more pixels are needed and some sci-
accurate for estimating total percentages of gross
entists believe that as many as 25 pixels are needed to
land-cover types (within 5% of ground truth for the
be confident of certain classification units (FGDC
three state area and within 1% for Louisiana). To
1992).
estimate the proportion of small forest patches that
In 1999, the Landsat 7 ETM1 (enhanced thematic
were misclassified because of the course resolution of
mapper) was launched. This satellite has a panch-
AVHRR data, a Landsat TM scene of the Tensas basin
romatic band with 15 m resolution as well as in-
classified as forest and non-forest was resampled to 1
creased spatial resolution for band 6 (thermal in-
km resolution. The TM scene comparable to the
frared) from 120 m to 60 m. In addition cost of the
AVHRR resolution differed from the original TM
imagery has been substantially reduced.
scene by only 2.4%.
SPOT
IRS-1 B LISS-II
Systeme Pour l’Observation de la Terre (SPOT) was
launched by the French government in 1986. SPOT-2 India began development of an indigenous Indian
was launched in 1990 and SPOT-3 was launched in Remote Sensing Satellite (IRS) program in the early
1993. SPOT was the first earth resource satellite to 1980’s to obtain information on agriculture water
have pointable optics which increases the oppor- resources, forestry and ecology, geology, water sheds,
tunities for imaging an area. SPOT also has stereo- marine fisheries and coastal management. The IRS-
scopic imaging capability. SPOT high resolution vis- 1B Linear Imaging Self-scanning Sensor (LISS-II)
ible (HRV) has green, red and near infrared spectral was launched in 1991. The four bands of the LISS-II
bands with 20 m spatial resolution. SPOT panch- are similar to Landsat TM bands 1, 2, 3 and 4. IRS-1B
romatic (Pan) has 10 m spatial resolution. SPOT-4, LISS-II data were used to study wetlands by
387

Kindscher et al. (1998), Kushwaha et al. (2000), flooding in a forested wetland was studied by Tow-
Chopra et al. (2001). nsend and Walsh (1998). The study area was the
lower Roanoke River floodplain, North Carolina, in
Radar an eastern bottomland forest. Landsat TM images
were limited by their inability to detect standing water
Satellite radar data is available from ERS-1, launched under canopies during leaf out conditions. SAR im-
by the European Space Agency in 1991, JERS-1, ages used were from the JERS-1 and ERS-1. The
launched by the National Space Development Agency wavelength and depression angle of a radar transmis-
of Japan in 1992, and RADARSAT, launched by the sion determined the depth the signal would penetrate
Canadian Space Agency in 1995 (Table 2). Hess et al. below the surface. Longer wavelengths generally had
(1990), Kasischke and Bourgeau-Chavez (1997), greater penetration in forest canopies, especially when
Townsend and Walsh (1998), Kushwaha et al. (2000) the wavelengths were much longer than leaf size.
have used satellite radar remote sensing systems to ERS-1 uses the shorter C-band sensors and JERS-1
study wetlands. Radar has advantages for remote uses the longer L-band sensors. The difference be-
sensing for two reasons: 1) radar systems can collect tween flooded and non-flooded areas was clearly
data at any time of day and under almost any weather distinguished on JERS-1 images. ERS-1 imagery was
conditions; 2) radar reflections (backscatter) provide also useful for detecting the difference between
different information than optical sensors. flooded and non-flooded areas, although the differ-
Flooded forests with a variety of different struc- ences were not as pronounced as on the JERS-1
tures in locations from the tropics to boreal regions imagery.
were detected using radar imagery (Hess et al. 1990). Kushwaha et al. (2000) evaluated ERS-1 SAR and
Flooded forests showed up as bright tones on radar IRS-1B LISS-II data for discrimination of mangrove
imagery because of the strong corner reflections be- wetlands in the coastal region of West Bengal, India,
tween specular water surfaces and tree trunks. L-band an area with persistent cloud cover especially during
data generally provided good distinction between the monsoon season. The main objectives were to
flooded and non-flooded forest and between forest and evaluate filtering algorithms and image enhancement
marsh vegetation, although sometimes marsh vegeta- techniques for wetland detection. The detection of
tion had high returns that could be confused with terrain features in SAR imagery is limited by the
flooded forest. presence of speckles. There are different techniques to
Kasischke and Bourgeau-Chavez (1997) used reduce speckles in SAR imagery but the most com-
ERS-1 synthetic aperture radar (SAR) data for moni- mon is the use of a median filter. This study evaluated
toring the presence or absence of water in wetlands in various filters for the best combination of speckle
Big Cypress National Preserve, Fakahatchee Strand reduction and terrain detail and the median filter with
Preserve, and Collier Seminole State Park in south- a 5 pixel by 5 pixel window gave the best results.
western Florida, US. The results showed that various Because SAR data is achromatic and is limited in its
vegetation communities could be distinguished based ability to detect terrain features, typically false color
on canopy structure, soil moisture, and presence or composite images are generated by either using multi-
absence of flooding. Generally the presence of water temporal SAR data or by fusing SAR data with multi-
under a plant canopy increased the radar backscatter spectral optical sensor data. These false color com-
for wetlands with woody vegetation and decreased posites were generated and intensity, hue and satura-
backscatter for wetlands with herbaceous vegetation. tion (IHS) transformations of the data were done and
The ability of Landsat TM and SAR data to detect the resulting images were compared visually. The IHS

Table 2. Characteristics of satellite synthetic aperture radar (SAR) systems.


ERS-1 JERS-1 RADARSAT
Look angle, deg 23 35 20–59
Wavelength, cm 3.75–7.5 C band 23 L band 5.6 C band
Polarization VV HH HH
Temporal Resolution 18 days 16 days 26 days
Spatial Resolution 30 m 18 m 10–100 m
First Launched 1991 1992 1995
388

transformation allows for an image to be made which Remote sensing classification techniques used for
has enhanced spatial and spectral details because each wetland identification
IHS component can be varied separately without
affecting the others. A false color composite of the Visual interpretation
pre-monsoon, monsoon, and post-monsoon multi-
temporal SAR data improved wetland discrimination Early work with satellite imagery used visual interpre-
compared to any single SAR image date. The false tation to identify wetlands (Best and Moore 1979;
color composite of SAR and IRS-1B LISS-II data, Hutton and Dincer 1979; Nayak and Sahai 1985).
with the IRS near infrared band (band 4) replaced by More recently, Chopra et al. (2001) studied water
SAR data, provided the best wetland discrimination. turbidity, seasonal water fluctuations, and vegetation
Kushwaha et al. (2000) caution that these results are status of Harike wetland in Punjab, India by visual
study area dependent but note that speckle reduction analysis of false color composite images of IRS LISS-
of radar data and image enhancement were necessary II bands 2,3, and 4. In addition, Johnston and Barson
for effective wetland discrimination. (1993), in their study of Australian wetlands, empha-
size how useful visual analysis of hard-copy images
can be for an overview and reconnaissance mapping
of wetlands, especially for those not trained in remote
Spectral reflectance of wetland vegetation sensing. Although visual interpretation of satellite
images is useful for identifying wetlands, most recent
Most computer classification techniques, including work has emphasized computerized classification
maximum likelihood, minimum distance to means, methods, because of the reduction in analyst time.
unsupervised clustering, and parallelpiped methods,
are dependent on different spectral responses of wet- Unsupervised classification or clustering
land vegetation types for classification. Spectral re-
flectance studies have been useful for determining Classification methods can be broken into three main
which wetland vegetation types are spectrally separ- categories: unsupervised, supervised, and hybrid ap-
able and which bands and dates are best for wetland proaches which combine unsupervised and supervised
discrimination. For example, Ernst and Hoffer (1979) methods. Unsupervised classification, or clustering,
measured the spectral reflectances of different wet- groups together pixels with similar spectral values.
land types in northeastern Indiana. Hardwood swamp, Based on ancillary information, the analyst then gives
shrub swamp, and shallow marsh had nearly identical the clusters information class labels. The advantages
values. Therefore it was difficult to classify these of this method are that the time-consuming training
wetland types based solely on spectral characteristics. phase is eliminated and the classes are distinct units.
Conifers could not be distinguished from deep mar- Disadvantages are that the clusters may not corre-
shes with a mix of vegetation and water. Open water spond to desired information classes. Supervised clas-
was spectrally very distinct. sification uses training data, areas of pixels with
Most spectral reflectance studies have been done in known class type to train the computer to recognize
tidal marshes and have shown that biomass could be the various classes. Advantages include the ability to
estimated from reflectance values (Bartlett and specify desired information classes. Thus supervised
Klemas 1980, 1981; Budd and Milton 1982; Gross et approaches are often preferred over unsupervised
al. 1988, 1987, 1993; Hardisky et al. 1983; Reimold et because the land cover classes of interest are chosen a
al. 1973). Spectral reflectances of wetland species priori. Disadvantages are that desired classes may not
have also been studied in the prairie pothole region of correspond to spectrally unique or homogeneous
the Dakotas (Best et al. 1981), in a central California classes and that training data acquisition is time-con-
lake (Penuelas et al. 1993), and in a California subal- suming and expensive. More information on unsuper-
pine marsh (Spanglet et al. 1998). Anderson and vised, supervised, and hybrid classification techniques
Perry (1996) measured differences in leaf spectral can be found in remote sensing texts such as Jensen
reflectance of red maple, Acer rubrum, on semi- (1996), Schowengerdt (1997).
permanently and temporarily flooded sites by the Various computer classification methods have been
Rappahannock River in Virginia. used to identify and classify wetland habitat types and
389

distinguish them from other land cover types. Review to create fifty initial clusters that were combined into
of satellite remote sensing of wetlands literature re- non-forested vegetation classes using aerial photo-
vealed that the most commonly used method is un- graphs and knowledge of the study area. Six meadow
supervised classification or clustering (Butera 1983; types along a xeric to hydric gradient were identified.
Gluck et al. 1996; Goodin 1995; Hinson et al. 1994; Vegetation in these six meadow types was surveyed
Huguenin et al. 1997; Kempka et al. 1992; Kindscher and average wetland values were calculated based on
et al. 1998; Lee and Marsh 1995; Macleod and the percent cover of each plant species and on its
Congalton 1998; Palylyk et al. 1987; Park et al. 1993, wetland prevalence index (Reed 1988), which is a
1991; Pope et al. 1994; Ramsey et al. 1998; Ramsey number from 1 to 5 depending on whether the plant
and Laine 1997). species occurs in wetlands. Field survey of the vegeta-
Unsupervised classification is most successful tion resulted in a 70% accuracy rate of finding wet-
when a large number of clusters are used. For exam- land vegetation at predicted locations. Kindscher et al.
ple, Ducks Unlimited typically used more than 230 (1998) concluded that this technique could be used to
clusters per Landsat TM scene for identifying wet- identify areas of potential wetlands.
lands in California’s Central Valley (Kempka et al.
1992). In another example, Macleod and Congalton Principal component analysis
(1998) used an unsupervised ISODATA classification
of Landsat TM data as one component of a change Another approach for unsupervised classification is to
detection study on eelgrass meadows, Zostera marina use multi-temporal imagery and principal component
L., in Great Bay, New Hampshire. The unsupervised analysis (PCA) to reduce the number of bands, and
ISODATA classification was done initially with 100 then apply clustering to the first few principal com-
clusters but there was still some confusion among the ponents. The approach described below also used
three classes; 255 clusters were used with much better ancillary data on wetlands to isolate wetlands from
results. other land uses and then classified the isolated wet-
‘‘Cluster busting’’, or progressive separation of lands into different habitat types.
mixed clusters until no further spectral separation is Gluck et al. (1996) evaluated photo interpretation
possible, is another technique to improve separation of color infrared aerial photography at three scales
between classes (Jensen et al. 1987). The clusters that and digital classification of Landsat TM data for
are not easily labeled are separated from the rest of the mapping wetlands in northwestern Ontario. Wetlands
image and then the clustering algorithm is applied in northwestern Ontario include marshes, open and
again to obtain additional clusters. In an unsupervised treed fens, and forested wetlands. The green, red, near
classification on Landsat TM images of South and middle IR bands of spring (April) and summer
Carolina Jensen et al. (1993a) performed this ‘‘cluster (August) TM imagery were used. Some bands that
busting’’ technique three times. were confusing because of ice were removed. A PCA
The following example of an unsupervised classifi- was done on the remaining bands and the first three
cation also illustrates that many clusters are necessary principal components were used in an ISODATA
for successful mapping. An unsupervised classifica- unsupervised classification with 250 clusters. The first
tion of August IRS-1B LISS-II satellite data and principal component (PC1) highlighted vegetation
average wetland values were used to identify areas of differences, PC2 highlighted wetness differences, and
potential wet meadows in Grand Teton National Park, PC3 distinguished wetlands from uplands. The band
US (Kindscher et al. 1998). There are two types of intercept and slopes of the spectral profiles of each of
wet meadows in this area. The wettest are hydric these 250 clusters were plotted. According to the
willow flats along rivers and lakes dominated by Salix slopes of the spectral profiles the most similar clusters
spp. and Carex spp., which typically have standing were combined reducing the number of classes to 40.
water throughout the growing season. The other type Using ground reference data and the spectral profile
are wet meadows along foothill slopes near lakes and plots each of these 40 classes were then put into one
rivers and in depressions dominated by Carex spp., of eleven different wetland information classes. In the
which only have periodic flooding although the soil is initial classification using Landsat TM data alone
saturated during a substantial part of the growing there was confusion between gramminoid wetland
season. The ISODATA clustering algorithm was used classes and recent cutovers and burns. Therefore
390

forest inventory polygons were used to separate the ods are minimum distance to means, parallelpiped,
wetlands from uplands. The forest inventory polygons also known as box decision rule or level slice pro-
did not have detailed information about the type of cedure, and maximum likelihood classification. These
wetland so the Landsat TM data was used to provide techniques are described in remote sensing texts such
the wetland type. The overall accuracy level at the as Jensen (1996) or Schowengerdt (1997). Minimum
pixel level was 72% when comparing classified wet- distance to means classifiers compute the centroid of
land types to the ground reference map. By combining the training data classes and assign unknown pixels to
physiognomically similar wetland classes, for exam- the class with the nearest centroid. Minimum distance
ple by combining deep marsh and shallow marsh into to means has been used map wetlands (Forgette and
marsh, the overall accuracy increased to 81%. The Shuey 1997; Hodgson et al. 1987; Huguenin et al.
advantage of using an unsupervised classification 1997).
approach is that the spatial resolution of the image is Parallelpiped, also known as box decision rule or
preserved. If a supervised approach had been used, level slice procedure, uses the range of spectral values
wetland classes might have been omitted or aggre- in the training data to define a region in data space.
gated because the analyst might not have been able to Pixels that fall into the defined data space are classi-
connect finer scale ground reference data to the satel- fied into that class. The parallelpiped classification
lite image data. method was used with Landsat TM images in the
Unsupervised classification techniques have been Florida Everglades (Hines et al. 1993).
very popular for wetland classification because of Maximum likelihood classification uses the means
their suitability for natural areas, which generally and variances of the training data to estimate the
have spectral variability and gradual transitions be- probability that a pixel is a member of a class. The
tween vegetation types. Unsupervised clustering is pixel is then placed in the class with the highest
most successful when a large number of clusters are probability of membership. Maximum likelihood gen-
used or when ‘‘cluster busting’’ is performed. How- erally gives better results than the minimum distance
ever, labeling a large number of clusters with the to means or parallel piped classifiers because the
correct landcover class and procedures such as ‘‘clus- covariance of the data is taken into account. Because
ter busting’’ increase analyst time. Multi-temporal of its performance, maximum likelihood is the most
imagery improves the results obtained. When multi- commonly used supervised classification technique
temporal imagery is used, PCA is often done to for mapping wetlands (Butera 1983; Ernst and Hoffer
reduce the number of bands of data. If reliable ancil- 1979; Ernst-Dottavio et al. 1981; Forgette and Shuey
lary data is available, wetlands should be separated 1997; Hewitt 1990; Huguenin et al. 1997; Lee and
from other land cover types before clustering to Park 1992; Lo and Watson 1998; Macleod and Con-
improve classification results. galton 1998; Palylyk et al. 1987; Yi et al. 1994; Zainal
et al. 1993).
Supervised classification The following example of a supervised maximum
likelihood classification illustrates how satellite imag-
There are many different supervised classification ery can be used to map riparian areas. This technique
techniques. For example, Luczkovich et al. (1993) is useful in arid regions where the riparian zone stands
used multivariate analysis of variance and covariance out from drier terrestrial areas. March Landsat data
to distinguish between sites of predominantly seagrass were used in a supervised maximum likelihood classi-
meadows, Thalassia testudinum or coral and unvege- fication to map riparian habitats along the Yakima
tated sand bottom on the north coast of the Dominican River in the shrub-steppe ecosystem of east central
Republic. Harper and Ross (1982) used discriminant Washington (Hewitt 1990). Landsat TM bands 2,4,
function analysis with Landsat MSS imagery to study and 7 were selected. TM7 was chosen to distinguish
part of the Peace-Athabasca Delta in northeastern water and locate the riparian zone. TM2 was chosen
Alberta, one of the largest freshwater deltas in the because it is uncorrelated with TM7 and would pro-
world. Discriminant function analysis and September vide additional information for the classification.
Landsat TM imagery were used to detect Kalmia TM4 (near infrared) was chosen because the peak
angustifolia L., an ericaceous shrub, on disturbed reflectance of green vegetation occurs in this band.
sites in Newfoundland (Franklin et al. 1994). Training data were obtained from field transects.
The most common supervised classification meth- Transects were located on the imagery to obtain
391

spectral signatures related to vegetation communities. transformation reduced the six TM bands to three
Sixteen distinct training classes were identified and spectral features, which correspond to brightness,
the study area was classified using the maximum greenness, and wetness. Then they used a three-step
likelihood algorithm. Because most of the classes ‘‘guided clustering’’ procedure to classify the bright-
were cropland and only the riparian areas were of ness, greenness, and wetness image: 1) identified
interest, the original 16 classes were aggregated into multi-spectral clusters of wood stork foraging habitat,
three classes of water, riparian, and other. Accuracy 2) determined signatures of other land cover classes,
assessment data were randomly selected and com- 3) merged the two sets of signatures together to
pared to aerial photographs. The final map had an classify the entire study area. They used 32 clusters in
accuracy of 81%. Hewitt (1990) concluded that ac- a minimum distance classification of 12 subscenes of
curacy could be improved by using multi-temporal the image. In the classification of these subscenes it
imagery, one image from winter or spring when water was apparent that marshes were sometimes confused
levels are high and one from summer, when vegeta- with forests. Therefore new signatures were created
tion is in full leaf. In addition, a buffer could be for confused classes resulting in a final set of 37
employed to mask out pixels greater than n distance clusters. Then the entire study area was classified
from the water class. This example shows the impor- using a minimum distance to means classifier. The
tance of imagery date. It also shows how simple accuracy was assessed with low altitude oblique aerial
procedures, such as a buffer around the water class, photography and the recorded locations of 39 known
can be used to improve the classification of wetland wood stork foraging sites. The overall classification
maps. accuracy was 74% for wood stork foraging sites.
Errors of omission were caused by the low spatial
Hybrid classifications resolution of TM imagery, which did not allow for
identification of smaller wood stork foraging sites.
Hybrid classifications involve a combination of super- This study highlights that classification is often an
vised and unsupervised classification techniques. One iterative procedure. For example, when marsh was
hybrid approach is to input descriptive statistics from confused with forest, new signatures were created to
a clustering algorithm into a maximum likelihood reduce the confusion. This study also highlights the
classifier. In this approach, unsupervised classification benefits of a hybrid approach for first clustering
or clustering is done on only a portion of the study portions of a scene where more detailed information is
area. After the clusters are assigned information known from maps or aerial photography, resolving
classes, cluster statistics are generated and input into a confusion between classes, and then classifying the
maximum likelihood classifier so that the entire study whole scene.
area can be classified. This method was used by Hybrid classification methods combine the
Gammon et al. (1979) in the Great Dismal Swamp of strengths of both supervised and unsupervised ap-
the United States, Pietroniro et al. (1995), Pope et al. proaches. They can be valuable for wetland studies
(1994) in Canadian arctic and subarctic wetlands, because of the complex variability in spectral re-
Gomarasca et al. (1992) in the Niger River interior sponses of wetland vegetation types.
delta in Mali, and Lee and Marsh (1995) in Arizona.
Hinson et al. (1994) used a hybrid classification Regression analysis with aerial photography
approach with Landsat TM images from December to
classify Texas coastal wetlands. In the prairie pothole region of North Dakota, Gilmer
Hodgson et al. (1987) used a hybrid classification et al. (1980) used aerial photography together with
technique known as ‘‘guided clustering.’’ A late Landsat MSS data to estimate wetland numbers.
spring (May) Landsat TM image was used to identify Studies indicate that between 73 and 88% of wetland
wood stork habitat (shallow water or marsh) in north basins in the glaciated prairie pothole region are less
central Georgia. Spring imagery was chosen because than 0.4 ha, which is the spatial resolution of Landsat
for the southeastern US spring imagery provides MSS data. Therefore a method was needed for es-
important wetland discrimination (Jensen et al. 1984, timating wetlands smaller than the sensor’s resolu-
1986). First they did a Kauth-Thomas (tasseled cap) tion. First wetlands were counted using Landsat data
transformation (Kauth and Thomas 1976) on the six and then the Landsat results were adjusted based on
TM bands, excluding the thermal infrared band. This samples derived from high resolution aircraft data.
392

Ponds as small as 5 meters across were delineated Several vegetation indices and principal compo-
with 1:20000 aerial photography. Aircraft wetland nents were calculated with October SPOT multi-spec-
counts for areas sampled by the aircraft and paired tral image data in a study of one of Africa’s largest
Landsat counts were used to make a linear regression wetlands, the Okavango Delta in Botswana (McCar-
analysis. The linear regression was then used to thy et al. 1993). This large wetland consists of perma-
estimate wetland numbers over the whole region nent and seasonal swamps and most of the area is
covered by Landsat. The value of the Landsat data inaccessible, thus satellite remote sensing is very
was that it gave a representation of the highly variable useful for reconnaissance investigations. The vegeta-
characteristics of the entire region, even though many tion indices included NIR / red, a transformed vegeta-
small ponds were not detected. The aircraft data were tion index (TVI) h100 3 sqrt [(NIR2red) /(NIR1
useful for detection of small wetlands. The double red)]j, and a leaf area index (41.325 3 green / red 2
sample plots could provide additional information 42.45 3 green / NIR). TVI provided the most detail.
such as wetland types and land-use practices on TVI was related to biomass, but not necessarily the
surrounding uplands that would make the regional species composition of the vegetation. Since plant
assessment of waterfowl habitat more useful. This growth is sensitive to hydrology, the TVI image
technique would be advantageous when an estimate of indirectly provided hydrological information. Con-
wetland areas is desired (but mapping these wetlands ventional aerial photographs have superior resolution
is not required) over a large geographic area, but time to SPOT imagery but the TVI calculated from SPOT
and money limitations preclude the use of aerial imagery provides information on vegetation and hy-
photography for the entire region. drological conditions, which aerial photography does
not. Thus SPOT images and aerial photography were
Vegetation indices complimentary. The SPOT image provided informa-
tion about areas that could not be reached by ground
Often vegetation indices can be useful for highlight- survey. Data obtained during late winter or early
ing wetlands. The vegetation indices can be used in spring would be optimal for studying wetland vegeta-
visual interpretation of wetland boundaries and ex- tion patterns in the Okavango Delta because the
tents or used in a classification algorithm to map contrast between dry and wetland environments is
wetlands and other landcover types. For example, the very pronounced.
normalized difference vegetation index (NDVI) and Johnston and Barson (1993) evaluated Landsat TM
principal components often enhance spectral sepa- data for inventorying and classifying Australian wet-
ration of coastal wetland features (Budd and Milton lands. They concluded that for reconnaissance map-
1982; Gross et al. 1987). The NDVI and the first two ping simple density slicing of bands that are related to
principal components of a September 3-band SPOT physical parameters such as vegetation indices
multi-spectral image were used in a maximum likeli- (biomass / productivity), middle infrared (soil mois-
hood classification of northeastern Florida (Lee and ture), and visible blue (water depth / turbidity) may be
Park 1992). This study area contained a variety of dry as effective as more complicated statistical classifica-
land and wetland vegetation including salt and fresh- tions. Wetlands in their study areas included herb-
water marshes and hardwood and cypress swamps. dominated shallow freshwater marshes and meadows,
Overall classification accuracy was 87.8% for 10 deeper marshes and open water, both fresh and saline,
classes. and red-gum marshes (Eucalyptus camaldulensis
The spectral indice TM4 / TM2 has been found Dehnh.). Multi-temporal imagery (winter / spring and
useful for separating water from land (Raabe and summer) provided information on seasonal variability
Stumpf 1996), although it generally combines water, in water regimes and vegetation. They used a simple
wetlands and urban (Dobson, personal communica- density slicing (threshold) of Landsat TM band 5 to
tion). Dobson (personal communication) found that provide information on the extent and location of
TM5 / TM2 and (TM5*TM7) /(TM51TM7) separate wetlands. All pixels below a threshold value of TM5
urban areas from water and wetlands; TM5 / TM2 is were mapped as wet. The threshold value was set
especially good for discriminating marshes. These empirically and depends on the types of wetlands and
indices have been used with success in coastal areas. conditions at the time the imagery is collected, which
NDVI is also useful for separating water from dry makes the choice of image date very important. Using
land. density slicing of TM5 they were able to identify 95%
393

of permanent open water or marshes, but only 50% of or narrow patches of vegetation and at the boundaries
freshwater meadows. They had little success mapping between different vegetation types (Figure 3). Al-
vegetation types and concluded that classification of though the boundaries in this figure are drawn as lines
satellite imagery is more related to vegetation density, as in most maps, in reality the boundaries are typically
vigor and moisture status than to species distribution. zones with one vegetation type grading into another.
Satellite remote sensing data should be seen as com- These fuzzy transition zones increase the number of
plimentary to conventional mapping techniques; the mixed pixels. In unsupervised classification, mixed
data can provide an overview of a large region and pixels are typically dealt with by using mixed labels,
update information on water regime and vegetation i.e. forested wetland / emergent macrophytes, because
productivity. there is no method for extracting information about
the proportion of each species using traditional per-
Mixed pixels pixel classifiers. In supervised classification, mixed
pixels usually result in misclassification because the
Fuzzy classification, mixtures estimation, spectral mixed pixels are different spectrally from pure pixels
mixture analysis and subpixel classification are all of the individual vegetation types. With fuzzy classifi-
ways of dealing with the problem of mixed pixels, cation, mixtures estimation, spectral mixture analysis
that is, when there is more than one vegetation type and subpixel classification information is obtained
within the satellite sensor’s instantaneous field of about the fraction of different classes within a mixed
view. Since satellite sensors such as SPOT and Land- pixel. Fuzzy classification is discussed in detail in
sat TM have spatial resolutions of 20 m and 30 m Jensen (1996) and the other techniques are described
respectively, often there is more than one vegetation below.
type within one pixel. This is especially true for small
Mixtures estimation

Work and Gilmer (1976) evaluated Landsat-1 MSS


data as an alternative to aerial photography to de-
termine the amount of surface water in the prairie
pothole region of east-central North Dakota. They
studied seasonal (May to July) and annual differences
(July 72 to July 73) in the amount of water surface.
They used a simple level-thresholding of MSS-7 (near
infrared) data to distinguish water from non-water
surfaces. Water was distinguished by accepting pixels
with low radiance values as water and rejecting all
pixels above a certain threshold as non-water. The
decision boundary for separating water from non-
water terrain features was located using histograms.
Because of mixed pixels at the edge of water, surface
areas of water features were underestimated. Small
ponds and those of irregular shape (a high ratio of
perimeter length to area) had more errors in their
classification. Generally a pond had to be 0.4 ha (1
acre) in size to be recognized, but then it had to be
wholly within one pixel and not spread over several
pixels. Practically it was problematic whether ponds
Figure 3. Depending on the spatial resolution of the satellite sensor 0.4–1.6 ha (1 to 4 acres) were recognized. Above 1.6
and the spatial characteristics of the different classes, mixed pixels ha ponds were almost always recognized although
may be very common. The different classes will also intermix at the their size was underestimated. To be able to identify
boundaries, or have fuzzy transitions, again resulting in mixed
pixels. Notice how small or narrow patches of different classes, smaller ponds, they also calculated the fraction of a
compared to the spatial resolution of the satellite sensor, will result pixel composed of open water using a technique
in mixed pixels. called ‘‘proportion estimation’’ or ‘‘mixtures estima-
394

tion’’ which determines the proportion of each class in different classes on the final image. The different
a pixel or if it is an unknown class. With this tech- vegetation groups included flooded forests and mac-
nique there are two constraints: 1) at least n-1 spectral rophytes. This analysis provided more information
bands are needed to estimate mixtures of n classes; 2) about the spatial heterogeneity of the Amazon River
the signatures for the classes in the mixture must not floodplain than would have been impossible with a
be similar and no one signature must approach the traditional per-pixel classifier. In addition it was rela-
weighted average of the other signatures. In this tively easy to choose pixels for the three endmembers
study, bands 5 and 7 were used for the proportion to create fraction images.
estimation to distinguish three different cover types:
water, bare soils, and green vegetation. The propor- Subpixel classification
tion of water in a pixel was calculated and cutoffs .
86% were mapped as totally water and , 30% as not Huguenin et al. (1997) used subpixel processing to
water. Proportion estimation significantly improved classify bald cypress and tupelo gum wetlands in
pond shape recognition. Ponds larger than 0.5 ha (1.3 Georgia and South Carolina. Subpixel processing is
acres) were consistently recognized and many ponds defined as the search for specific materials of interest
as small as 0.13 ha (0.33 acre) were identified. Thus from within a pixel’s mixed multi-spectral image
this technique can estimate the proportion of different digital number spectrum. This method has advantages
cover types in one pixel, but by applying threshold over spectral mixture analysis (linear mixing model)
cutoffs pixels can be mapped as one cover type. and fuzzy set classification (Jensen 1996), which also
gives subpixel information, because the overall com-
Spectral mixture analysis position of each pixel is not constrained to be some
combination of the defined image classes or end
Spectral mixture analysis was used to determine the members. The steps in subpixel processing include
vegetation cover on the Amazon River floodplain and signature derivation for a material of interest and
the relative concentration of suspended sediments in classification of each pixel identifying the fraction of
surface waters of the inundated floodplain (Mertes et material of interest present. Therefore for each materi-
al. 1995). Bands 1–5 and 7 of three Landsat TM al of interest a separate classification must be done.
images, acquired when water levels were high and the The fraction image pixel values vary from 0.0 to 1.0.
floodplain was inundated, were used. Spectral mixture In this case two classifications were done, one for bald
analysis, also known as linear mixing model, enables cypress and one for tupelo gum. A May (after spring
the extraction of information regarding the surface leaf-out) Landsat TM image was acquired for the
materials present in a pixel. Spectral mixture analysis classification. Large scale (1:7000) color IR aerial
calculates the least-squares best fit for each pixel photography was used to identify stands of cypress
along mixing lines bounded by spectra for endmem- and tupelo in 4 study areas, sometimes even indi-
bers and in this way accounts for each pixel’s vari- vidual trees, for the training (2 sites) and test data (2
ation in the mixture composition. An endmember sites). NAPP photography was used for outside the
ideally represents a pure component of the mixtures study areas. Field work was used to identify areas of
present in the pixels. Three endmembers were chosen homogeneous stands and mixed stands. Traditional
for each image: vegetation, shade, sediment-water. classification methods (ISODATA clustering, maxi-
Spectral mixture analysis produced a fraction image mum likelihood, minimum distance to means) were
for each endmember, which showed the fraction from done by an independent research team using the same
0 to 1.0 that was contributed by that endmember. The TM image, CIR aerial photography and field data. The
fraction images could be viewed individually or com- other researchers also used the same training areas
bined. Generally the spectral mixture analysis method and classified the same training and test areas. The
can detect subpixel occurrences of vegetation, shade, minimum distance to means was the best traditional
and sediment but it is not appropriate for detecting classifier. The subpixel processing method classified
subpixel occurrences of specific species of vegetation. tupelo with 91% accuracy and cypress with 89%
Therefore Mertes et al. (1995) passed the fraction accuracy while the minimum distance to means clas-
images through a parallelpiped classifier with thres- sifier was 85% accurate for tupelo and 71% accurate
holds for 5 different vegetation, 5 different sediment- for cypress. The subpixel processing method had high
water and 5 different shade groups for a total of 14 accuracy for both pure stands and mixed stands.
395

Fuzzy classification, mixtures estimation, spectral forested wetlands in Maine (Sader et al. 1995). Ancil-
mixing analysis, and subpixel classification are all lary data included hydric soil data, NWI maps, a
ways to handle the problem of mixed pixels. Depend- DEM, and hydrography. The rule-based model was
ing on the spatial resolution of the satellite sensor and compared to conventional classifiers: an unsupervised
the size of wetland vegetation patches, mixed pixels classification, a tasseled-cap transformation, and a
can be quite common. If more detailed information on hybrid classification, which combined unsupervised
the composition of mixed pixels is desired, these cluster statistics with supervised forest training statis-
techniques may be very useful. If information on a tics in a maximum likelihood classifier. July Landsat
specific vegetation species is desired, then the subpix- TM imagery was used. TM wavebands are highly
el classification method is preferred. Of course even correlated; therefore bands 3,4,5 were chosen to re-
with a subpixel classifier, similar to traditional per duce data volume yet retain most of the information of
pixel classifiers, a plant species must be spectrally the seven bands. These bands were chosen because
separable from other species in the study area to be previous research found them to be optimal for classi-
correctly identified. fication of wetlands. The rule-based model consisted
of four hierarchical layers. Each layer used the GIS
Ancillary data data and weights were assigned depending on whether
the criteria were met or not. Each pixel was assigned a
Researchers have usually found that using ancillary value depending on whether or not it was forested,
data in addition to multi-spectral satellite data im- according to the unsupervised classification. Weights
proves classification results. Ancillary data that have were assigned according to whether it was a palustrine
been used to assist in identification of wetlands in- forested wetland in the NWI layer, then whether the
cluded: soils data (Bolstad and Lillesand 1992; Ernst soil was hydric, then whether slope was , 5% and
and Hoffer 1979; Sader et al. 1995; Yi et al. 1994), finally whether it was in a three pixel buffer around
topographic or elevation data (Bolstad and Lillesand hydrographic features. NWI, hydric soils and slope
1992; Hinson et al. 1994; Park et al. 1991; Sader et al. were the most important variables in the model. The
1995), NWI maps (Sader et al. 1995), and forest rule-based GIS classification did not have a signifi-
inventory maps (Gluck et al. 1996). The use of forest cantly higher accuracy than the hybrid classification.
inventory maps (Gluck et al. 1996) was previously Both methods had overall accuracies in the 80%
described in detail in the section on Principal Com- range. Results might be improved if the hybrid classi-
ponents Analysis. Ancillary data have been incorpo- fication was used in the GIS model instead of the
rated in layered classifications (Ernst and Hoffer unsupervised classification.
1979). Often the ancillary data are added as data Bolstad and Lillesand (1992) used topographic
layers in a GIS (Yi et al. 1994). Rule-based models position, soil texture and Landsat TM data in a rule-
have also incorporated ancillary data (Bolstad and based expert system to classify different land cover
Lillesand 1992; Sader et al. 1995). types in northern Wisconsin. Topographic position
Wetlands in northeastern Indiana were classified data were derived from USGS 7.5 minute quad maps
using a layered classification procedure that included and grouped into four categories: wetland, upland,
soils data and spectral data (Ernst and Hoffer 1979). water and road. Soil texture data came from the
The layered classifier first stratified the image into USDA Soil Conservation Service soil survey. In 5 of
upland and bottomland sites based on soils data. The 6 of the different trials of the rule-based method,
layered classifier allowed hardwood swamps to be which varied by study site, year of image, and
distinguished from upland hardwoods with 66% ac- operator, the rule-based method had better accuracy.
curacy. There was still some confusion where areas of On average the classification accuracy of the rule-
bottomland soil were too narrow to be identified in the based method was 83% compared to 69% for maxi-
soil data layer. The layered classifier improved the mum likelihood. However the rule-based expert sys-
separation of wetlands from non-wetlands (84.3% vs. tem also took much more time to develop. Therefore
71.7%). wetland scientists would need to decide on a project
basis how much time and analyst expertise was avail-
Rule-based classifiers able and the accuracy level needed.
Lunetta and Barlogh (1999) used a simple rule-
A rule-based GIS model was developed to identify based GIS model to classify wetland and upland
396

vegetation types in Maryland and Delaware. They and time they will take over conventional techniques.
used two dates of Landsat TM imagery, June 11 1988 Project managers must look at the tradeoffs of time
(leaf- on) and April 20 1986 (leaf-off), chosen to be and expertise versus improved accuracy given their
after a precipitation event during a normal precipi- wetland mapping goals and the resources they have
tation year. Gray-level thresholding techniques of available.
Landsat TM band 5 in the April image were used to
determine the extent and distribution of non-hydric Comparing satellite imagery classification to NWI
and hydric soils. Density slices of TM band 5 were
observed until one was found which separated wet- Generally small wetlands and long narrow wetlands
lands from uplands. An initial classification map was that are shown on the NWI are not detectable with
created using the June imagery. For the initial classifi- satellite imagery (Park et al. 1993; Forgette and Shuey
cation, the image was clustered to identify signatures 1997). Other differences between NWI maps and
of unique spectral classes. Then the cluster statistics wetland maps created from satellite remotely sensed
were input into a maximum likelihood classifier. imagery are described in the examples below. Similar
Stereo pairs of color infrared aerial photos, topog- differences may also be found between any detailed
raphic maps and soil maps were utilized to assign land wetland map derived from aerial photography or
cover classes to the clusters. Classes were wetland fieldwork and maps derived from satellite imagery.
(forest, shrub, emergent, agriculture) and upland (for- Ducks Unlimited was interested in wintering
est, shrub, herbaceous, agriculture). Finally using GIS habitat for ducks in California’s Central Valley, espe-
analysis, areas with hydric soils were reclassified to cially rice fields that were not on NWI maps (Kempka
wetland categories, those with non-hydric soils to et al. 1992). To identify the wetlands they performed
upland categories. Classification accuracy using the unsupervised classifications with typically more than
two dates of imagery was much higher (88% com- 230 separate spectral classes per Landsat TM scene.
pared to 69%) than from the single image date. This NWI maps of the Central Valley were out-of-date
example once again demonstrated how density slicing because they were created from photos taken in the
of TM5 can be used to separate wetlands from up- 1970’s. When they compared NWI maps to the classi-
lands. It also showed how a simple GIS procedure can fied Landsat TM image, there were differences caused
be used after classification to improve accuracy. by different sensor resolutions, different image acqui-
Fuller et al. (1998) also used simple GIS pro- sition dates, different classification procedures, and
cedures after classification to improve wetland classi- different seasons with different areas inundated. They
fication. They mapped wetlands in Sango Bay on concluded that Landsat TM and NWI provide com-
Lake Victoria in Uganda with Landsat TM data. Field plimentary data.
data were used to locate areas for training with the Matrix-overlay analysis methods were used to com-
supervised classification. However after classification pare the amount of agreement between the USFWS
some errors were easily identified where cover types NWI map and the classified map of Texas coastal
were mapped out of their normal context. Thus they wetlands (Hinson et al. 1994). NWI classes were
created a mask for cultivated and built up classes and cross-referenced to the NOAA C-CAP classification
filtered out the misclassified wetland classes for an scheme. Open water in the Gulf of Mexico was
overall classification accuracy of 86%. omitted from the comparison because it would have
Ancillary data can be very useful for improving made the agreement artificially high. Overall agree-
wetland classification. However, to improve the clas- ment was 76.3%. Differences in the NWI map and
sification, the ancillary data itself used must be accur- classified image map were caused by: 1) inconsisten-
ate and must be registered accurately to the satellite cies in C-CAP and NWI classification schemes; 2)
images. With recent technology advances, digital interpretation of boundaries between upland and wet-
ancillary data such as DEMs and soil maps are often land (NWI was more conservative while on the classi-
easy to acquire. Layered classifiers and rule-based fied image transitional pixels were usually classified
classifiers that incorporate this ancillary data typically as wetland); 3) narrow classes with a high proportion
have higher classification accuracy than conventional of edge to area showed less agreement than large
statistical techniques such as maximum likelihood. blocky classes; and 4) errors in each map were
These techniques vary in how much analyst expertise compounded by the overlay process.
397

Change detection oxbow lakes, lagoons and pans with hydrophytic reed
and grass vegetation. In a change detection study,
With data available from Landsat MSS since 1972, ideally all the images would be from the same sensor
AVHRR since 1979, Landsat TM since 1982, SPOT to avoid differences caused by different sensors.
since 1986 and IRS since 1991, repeat coverage However because of budget limitations of this project,
satellite remote sensing is often used to evaluate it was not possible to use all Landsat TM images,
change over time in wetland ecosystems. General which would have been preferred because of the
information on change detection techniques can be middle infrared band. Therefore Landsat MSS images
found in Singh (1989), in (Coppin and Bauer 1994) were also used. Dry season (September) images were
and in remote sensing texts such as Jensen (1996). chosen because this is when the wetland vegetation is
References such as these should be consulted to most easily distinguished from the surrounding up-
determine which change detection procedure is most land vegetation. Differences caused by changes in
appropriate for the desired project and also for im- vegetation phenology were limited by using near
portant factors affecting change detection studies. anniversary date images. A hard copy of the most
One important factor in change detection studies is to recent satellite image was taken into the field to aid in
use dates of imagery such that the wetlands are in the gathering ground reference data. Using a GPS, 87
same phenological state from year to year. Other sites were located in large areas of the same cover
important factors include using imagery from the type. Steps in the change detection procedure were: 1)
same sensor taken at the same time of day and careful atmospheric correction of latest (reference) image, 2)
geometric registration of the images. Many change geometric registration of all images and normalization
detection studies of wetlands have been done. In a of the other images with the corrected reference
study of the Matagorda Bay estuarine system on the image, 3) image classification, and 4) change de-
Texas Coast, Weismiller et al. (1977) evaluated four tection. A hybrid classification approach was used.
different change detection techniques. Jensen et al. Field sites were clustered and labels assigned to the
(1993b) used Landsat TM imagery for detecting water clusters. Areas with very similar spectral signatures
level changes in a coastal region near Charleston were again clustered and assigned labels resulting in
South Carolina. SPOT panchromatic data and image fifteen final classes. Using these fifteen classes, the
algebra change detection techniques were used to map rest of the image was classified with a supervised
seasonal changes in waterlilies and cattails in a fresh- maximum likelihood classification, resulting in a
water reservoir located on the Savannah River Site in classification accuracy of 73%. The other three im-
South Carolina (Jensen et al. 1993c). Changes in ages were classified in this way also. The assumption
vegetation types and water boundaries due to flooding was made that the 15 classes on the reference image
were analyzed in the Peace-Athabasca Delta, a 4000 also existed on the earlier images and that there were
km 2 wetland complex, in northeastern Alberta, no classes on the earlier images that did not exist on
Canada, using Landsat MSS imagery (Wickware and the latest image. To improve the accuracy of the
Howarth 1981). Haack (1996), Haack and Messina change detection, the 15 detailed classes were reduced
(1997) used Landsat MSS images from 1973, 1979, to 6 broad classes. This was done because in change
and 1989 to examine the growth of the Omo Delta at detection studies that compare classified images, it is
the northern end of Lake Turkana in Ethiopia and important to have a high classification accuracy in
Kenya, Africa. Change in habitat types as determined each of the images so that real changes are found
from a Principal Components Analysis of Landsat TM instead of misclassification errors.
imagery was used to monitor seagrass on the east A major wetland change detection effort has been
coast of Bahrain (Zainal et al. 1993). More detailed undertaken by NOAA, which started the Coastal
examples of change detection studies are given below. Change Analysis Program (C-CAP) to detect and
Munyati (2000) monitored change in a floodplain monitor changes in coastal upland, wetland and sub-
wetland system in southern Zambia with Landsat mersed vegetation landcovers (Dobson et al. 1995).
MSS and TM images. The goal of the project was to C-CAP monitors coastal regions of the United States
determine changes in the wetland caused by drought every one to five years depending on the anticipated
and less water reaching the wetland because of dams rate of change and the availability of suitable remote
and irrigation use. The floodplain zone consists of sensing and field data. Data from C-CAP will be used
398

to improve understanding of coastal landcovers and zonal statistics were not representative over very areas
their relationship to the health of marine resources and and so polygons larger than 100 ha were visually
to provide information to habitat managers on the assessed for changes. The accuracy for all changed
success or failure of habitat management policies and wetlands was 90 percent for the rule-based system,
programs. Satellite imagery is used to monitor coastal with accuracy for the five dominant NWI systems
wetlands and adjacent uplands and a combination of (PFO, PEM, PUB, PSS, PAB) ranging from 87 to 94
aerial photography and satellite imagery is used to percent. Rule-based methods are initially labor-inten-
monitor submersed vegetation. Landsat TM data was sive. However, once the key rule (in this study, the
preferred over SPOT imagery for the change analysis standard deviation threshold of mean modulus value)
because it is less expensive and covers a larger area. is found, they can be applied quickly. Houhoulis and
TM imagery also has better spectral resolution and Michener (2000) note pixel vector modulus may be
spectral bands (bands 5 and 7) that are more suitable useful for other change detection studies because it is
for wetland delineation. In addition TM data have easy to calculate, captures spectral variation, and
been routinely collected for a longer time in US reduces data volume and processing time. It may be
coastal areas. To evaluate NOAA’s C-CAP change possible to use a similar method to detect changes on
detection protocol, Jensen et al. (1993a) used multi- other aerial photography or fieldwork based maps.
temporal Landsat TM data to study changes in two
study areas in coastal South Carolina. They looked at
four post-classification change detection procedures, Conclusions
including the use of a ‘‘change / no change’’ mask. By
using the spectral information of the unclassified For more effective use of satellite remote sensing,
images to differentiate between areas of change and wetland managers should be aware of the limitations
no change, the ‘‘change / no change’’ mask can mini- and advantages of satellite data and should chose from
mize errors in change detection caused by errors in their available wetland mapping options accordingly.
either dates of classification. Satellite remote sensing is especially appropriate for
Houhoulis and Michener (2000) developed a rule- initial reconnaissance mapping and continued moni-
based model to detect wetland change and update toring of wetlands over large geographic areas. Al-
NWI data in the Georgia coastal plain. This area has though satellite remote sensing cannot provide the
forested and herbaceous wetlands interspersed in detailed information available from aerial photo-
agricultural fields, which leads to difficulties in image graphy or field studies, it can provide complimentary
classification with the confusion between wetlands information to these conventional mapping tech-
and forests and agricultural areas. First they did an niques, such as updated information on water regime
unsupervised classification of October date SPOT and vegetation productivity. It can also identify areas
multispectral imagery with twenty classes. The twen- where changes are occurring and where more detailed
ty classes were aggregated into six broad categories: information must be gathered. One promising tech-
1) water, 2) wet soil, 3) forest, 4) scrub / shrub, 5) nique for updating wetland maps is a rule-based
cultivated, and 6) bare soil. Pixel vector modulus was wetland change detection method (Houhoulis and
calculated, (sqrt(b 21 1 b 22 1 b 23 / N)), which represents Michener 2000).
the Euclidean distance of each pixel from the origin in Techniques for improving the classification of wet-
multidimensional space. NWI data was used as a lands with satellite remote sensing data include the
zonal layer and the mean modulus and majority use of multi-temporal imagery and ancillary data.
landcover class in each NWI polygon was calculated. Multi-temporal imagery allows for the highest accura-
Statistics were generated by NWI system and majority cy in wetland identification and discrimination from
landcover to provide a basis for generation of wetland other land cover types. The ideal dates will differ
change rules. Wetlands typically had pixel vector depending on the type and location of wetlands being
modulus values of less than 40, regardless of NWI studied. Ancillary data, if available, should be used to
system. Thresholds for wetland change detection were aid classification. If possible wetlands should be
developed by adding one standard deviation to the separated from other land cover types prior to their
mean of the modulus histograms plotted by NWI classification. This separation is often done with
system and majority landcover class. One exception ancillary data or a density slice of Landsat TM band 5.
to the rule was for very large wetland polygons. The Layered or rule-based methods generally provide
399

better results than conventional statistical classifica- Butera M.K. 1983. Remote sensing of wetlands. IEEE Transactions
tion methods, often because of their use of ancillary on Geoscience and Remote Sensing GE-21: 383–392.
Cairns S.H., Dickson K.L. and Atkinson S.F. 1997. An examination
data. If detailed information at a finer spatial res- of measuring selected water quality trophic indicators with SPOT
olution than the satellite sensor is required, then a satellite HRV data. Photogrammetric Engineering and Remote
subpixel classification method should be used. Sensing 63: 263–265.
The combination of radar data together with optical Carter V., Garrett M.K., Shima L. and Gammon P. 1977. The Great
data for wetland identification and mapping has not Dismal Swamp: management of a hydrologic resource with the
aid of remote sensing. Water Resources Bulletin 13: 1–12.
yet been exploited to its fullest extent. Radar is useful
Chopra R., Verma V.K., and Sharma P.K. 2001. Mapping, moni-
for studying wetlands because of its ability to dis- toring and conservation of Harike wetland ecosystem, Punjab,
tinguish between flooded and non-flooded areas, even India, through remote sensing. International Journal of Remote
in forests. In addition, radar data can be collected in Sensing 22: 89–98.
almost all weather conditions, a characteristic that is Coppin P.R. and Bauer M.E. 1994. Processing of multitemporal
Landsat TM imagery to optimize extraction of forest cover
especially important in areas with frequent cloud
change features. IEEE Transactions on Geoscience and Remote
cover. In conclusion, classification of wetlands with Sensing 32: 918–927.
satellite imagery is difficult but useful. Satisfactory Daily G.C. (ed.) 1997. Nature’s Services: Societal Dependence on
results will require specialized techniques as de- Natural Ecosystems. Island Press, Washington, 329 pp.
scribed in this paper on a case by case basis. Dobson J.E., Bright E.A., Ferguson R.L., Field D.W., Wood L.L.,
Haddad K.D. et al. et al. 1995. NOAA Coastal Change Analysis
Program (C-CAP). NOAA technical report NMFS 123. US
Department of Commerce, Seattle, Washington, USA, 92 pp.
References Doren R.F., Rutchey K. and Welch R. 1999. The Everglades: a
perspective on the requirements and applications for vegetation
Ackleson S.G. and Klemas V. 1987. Remote sensing of submerged map and database products. Photogrammetric Engineering and
aquatic vegetation in Lower Chesapeake Bay: a comparison of Remote Sensing 62: 155–161.
Landsat MSS to TM imagery. Remote Sensing of Environment Ernst C.L. and Hoffer R.M. 1979. Digital processing of remotely
22: 235–248. sensed data for mapping wetland communities. LARS Technical
Ackleson S.G., Klemas V., McKim H.L. and Merry C.J. 1985. A Report 122079. Laboratory for Applications of Remote Sensing,
comparison of SPOT simulator data with Landsat MSS imagery Purdue University, West Lafayette, 119 pp.
for delineating water masses in Delaware Bay, Broadkill River, Ernst-Dottavio C.L., Hoffer R.M. and Mroczynski R.P. 1981.
and adjacent wetlands. Photogrammetric Engineering and Re- Spectral characteristics of wetland habitats. Photogrammetric
mote Sensing 51: 1123–1129. Engineering and Remote Sensing 47: 223–227.
Anderson J.E. and Perry J.E. 1996. Characterization of wetland FGDC 1992. Application of satellite data for mapping and moni-
plant stress using leaf spectral reflectance: implications for toring wetlands. Technical Report 1, Wetlands Subcommittee.
wetland remote sensing. Wetlands 16: 477–487. Federal Geographic Data Committee (FGDC), Washington, DC,
Barbier E.B., Burgess J.C. and Folke C. 1994. Paradise Lost? The USA, 32 pp.
Ecological Economics of Biodiversity. Earthscan, London, 267 Forgette T.A. and Shuey J.A. 1997. A comparison of wetland
pp. mapping using SPOT satellite imagery and national wetland
Bartlett D. and Klemas V. 1980. Quantitative assessment of tidal inventory data for a watershed in northern Michigan. In: Trettin
wetlands using remote sensing. Environmental Management 4: C.C. (ed.), Northern Forested Wetlands; Ecology and Manage-
337–345. ment. CRC Lewis Publishers, Boca Raton, Florida, USA, pp.
Bartlett D.S. and Klemas V. 1981. In situ spectral reflectance studies 61–70.
of tidal wetland grasses. Photogrammetric Engineering and Franklin S.E., Gillespie R.T., Titus B.D. and Pike D.B. 1994. Aerial
Remote Sensing 47: 1695–1703. and satellite sensor detection of Kalmia angustifolia at forest
Best R.G. and Moore D.G. 1979. Landsat interpretation of prairie regeneration sites in central Newfoundland. International Journal
lakes and wetlands of eastern South Dakota. In: Deutsch M., of Remote Sensing 15: 2553–2557.
Wiesnet D.R. and Rango A. (eds), Fifth Annual William T. Fuller R.M., Groom G.B., Mugisha S., Ipulet P., Pomeroy D.,
Pecora Memorial Symposium on Remote Sensing 1979. Ameri- Katende A. et al. 1998. The integration of field survey and
can Water Resources Association, Souix Falls, South Dakota, remote sensing for biodiversity assessment: a case study in the
USA, pp. 499–506. tropical forests and wetlands of Sango Bay, Uganda. Biological
Best R.G., Wehde M.E. and Linder R.L. 1981. Spectral reflectance Conservation 86: 379–391.
of hydrophytes. Remote Sensing of Environment 11: 27–35. Gammon P.T., Rohde W.G. and Carter V. 1979. Accuracy evalua-
Bolstad P.V. and Lillesand T.M. 1992. Rule-based classification tion of Landsat digital classification of vegetation in the Great
models: flexible integration of satellite imagery and thematic Dismal Swamp. In: Deutsch M., Wiesnet D.R. and Rango A.
spatial data. Photogrammetric Engineering and Remote Sensing (eds), Fifth Annual William T. Pecora Memorial Symposium on
58: 965–971. Remote Sensing 1979. American Water Resources Association,
Budd J.T.C. and Milton E.J. 1982. Remote sensing of salt marsh Souix Falls, South Dakota, pp. 463–473.
vegetation in the first four proposed thematic mapper bands. Gilmer D.S., Work E.A. Jr, Colwell J.E. and Rebel D.L. 1980.
International Journal of Remote Sensing 44: 303–314. Enumeration of prairie wetlands with Landsat and aircraft data.
400

Photogrammetric Engineering and Remote Sensing 46: 631– Photogrammetric Engineering and Remote Sensing 66: 205–
634. 211.
Gluck M., Rempel R. and Uhlig P.W.C. 1996. An evaluation of Huguenin R.L., Karaska M.A., Blaricom D.V. and Jensen J.R. 1997.
remote sensing for regional wetland mapping applications. For- Subpixel classification of bald cypress and tupelo gum trees in
est Research Report No. 137. Ontario Forest Research Institute, Thematic Mapper imagery. Photogrammetric Engineering and
Sault Ste Marie, Ontario, Canada, 33 pp. Remote Sensing 63: 717–725.
Gomarasca M.A., Lozano-Garcia D.F., Fernandez R.N. and Johan- Hutton S.M. and Dincer T. 1979. Using Landsat imagery to study
nsen C.J. 1992. Analysis of seasonal variation in the Niger river the Okavango Swamp, Botswana. In: Deutsch M., Wiesnet D.R.
and Rango A. (eds), Fifth Annual William T. Pecora Memorial
interior delta using satellite data. Geocarto International 7: 61–
Symposium on Remote Sensing 1979. American Water Re-
73.
sources Association, Souix Falls, South Dakota, pp. 512–519.
Goodin D.G. 1995. Mapping the surface radiation budget and net
Jensen J.R. 1996. Introductory Digital Image Processing: A Remote
radiation in a Sand Hills wetland using a combined modeling /
Sensing Perspective. 2nd edn. Prentice-Hall, Upper Saddle
remote sensing method and Landsat thematic mapper imagery.
River, NJ, USA, 318 pp.
Geocarto International 10: 19–29.
Jensen J.R., Christensen E.J. and Sharitz R. 1984. Nontidal wetland
Gross M.F., Hardisky M.A. and Klemas V. 1988. Effects of solar
mapping in South Carolina using airborne multi-spectral scanner
angle on reflectance from wetland vegetation. Remote Sensing of data. Remote Sensing of Environment 16: 1–12.
Environment 26: 195–212. Jensen J.R., Cowen D., Althausen J., Narumalani S. and Weath-
Gross M.F., Hardisky M.A., Klemas V. and Wolf P.L. 1987. erbee O. 1993a. An evaluation of the Coast Watch change
Quantification of biomass of the marsh grass Spartina alternifl- detection protocol in South Carolina. Photogrammetric En-
ora Louisel using Landsat Thematic Mapper imagery. Photo- gineering and Remote Sensing 59: 1039–1046.
grammetric Engineering and Remote Sensing 53: 1577–1583. Jensen J.R., Cowen D.J., Althausen J.D., Narumalani S. and Weath-
Gross M.F., Hardisky M.A., Wolf P.L. and Klemas V. 1993. Rela- erbee O. 1993b. The detection and prediction of sea level
tionships among Typha biomass, pore water methane, and re- changes on coastal wetlands using satellite imagery and a geog-
flectance in a Delaware (USA) brackish marsh. Journal of raphic information system. Geocarto International 4: 87–98.
Coastal Research 9: 339–355. Jensen J.R., Narumalani S., Weatherbee O. and Mackey H.E. Jr
Haack B. 1996. Monitoring wetland changes with remote sensing: 1993c. Measurement of seasonal and yearly cattail and waterlily
an east African example. Environmental Management 20: 411– changes using multidate SPOT panchromatic data. Photogram-
419. metric Engineering and Remote Sensing 59: 519–525.
Haack B. and Messina J. 1997. Monitoring the Omo River delta in Jensen J.R., Hodgson M., Christensen E., Mackey H.E., Tinney L.
East Africa using remote sensing. Earth Observation Magazine and Sharitz R. 1986. Remote sensing of inland wetlands: a
6: 18–22. multi-spectral approach. Photogrammetric Engineering and Re-
Hardisky M.A., Gross M.F. and Klemas V. 1986. Remote sensing of mote Sensing 52: 87–100.
coastal wetlands. BioScience 36: 453–460. Jensen J.R., Ramsey E.W., Mackey H.E. Jr, Christensen E.J. and
Hardisky M.A., Smart R.M. and Klemas V. 1983. Seasonal spectral Sharitz R.R. 1987. Inland wetland change detection using aircraft
characteristics and aboveground biomass of the tidal marsh plant, MSS data. Photogrammetric Engineering and Remote Sensing
Spartina alterniflora. Photogrammetric Engineering and Remote 53: 521–529.
Sensing 49: 85–92. Johnston R.M. and Barson M.M. 1993. Remote sensing of Aus-
Harper J. and Ross G.A. 1982. Digital analysis of Landsat data in tralian wetlands: an evaluation of Landsat TM data for inventory
the Athabasca Delta. In: Foreman J. (ed.), International Society and classification. Australian Journal of Marine and Freshwater
for Photogrammetry and Remote Sensing IV 1982. American Resources 44: 235–252.
Society of Photogrammetry and American Congress on Survey- Kasischke E.S. and Bourgeau-Chavez L.L. 1997. Monitoring south
ing and Mapping, Crystal City, Virginia, USA, pp. 319–327. Florida wetlands using ERS-1 SAR imagery. Photogrammetric
Hess L.L., Melack J.M. and Simonett D.S. 1990. Radar detection of Engineering and Remote Sensing 63: 281–291.
flooding beneath the forest canopy: a review. International Kauth R.J. and Thomas G.S. 1976. The tasseled cap-a graphic
Journal of Remote Sensing 11: 1313–1325. description of the spectral-temporal development of agricultural
Hewitt M.J. III 1990. Synoptic inventory of riparian ecosystems: crops as seen by Landsat, Symposium on Machine Processing of
the utility of Landsat Thematic Mapper data. Forest Ecology and Remotely Sensed Data 1976. Laboratory for Applications of
Management 33 / 34: 605–620. Remote Sensing, West Lafayette, Indiana, USA, pp 41–51.
Hines M.E., Pelletier R.E. and Crill P.M. 1993. Emission of sulfur Kempka R.G., Kollasch R.P. and Koeln G.T. 1992. Using GIS to
gases from marine and freshwater wetlands of the Florida preserve the Pacific flyway’s wetland resource. GIS World 5:
Everglades: rates and extrapolation using remote sensing. Jour- 46–52.
nal of Geophysical Research 98: 8991–8999. Kindscher K., Fraser A., Jakubauskas M.E. and Debinski D.M.
Hinson J.M., German C.D. and Pulich W. Jr 1994. Accuracy 1998. Identifying wetland meadows in Grand Teton National
assessment and validation of classified satellite imagery of Texas Park using remote sensing and average wetland values. Wetlands
coastal wetlands. Marine Technology Society Journal 28: 4–9. Ecology and Management 5: 265–273.
Hodgson M.E., Jensen J.R., Mackey H.E. Jr and Coulter M.C. Klemas V. 2001. Remote sensing of landscape-level coastal en-
1987. Remote sensing of wetland habitat: a wood stork example. vironmental indicators. Environmental Management 27: 47–57.
Photogrammetric Engineering and Remote Sensing 53: 1075– Kushwaha S.P.S., Dwivedi R.S. and Rao B.R.M. 2000. Evaluation
1080. of various digital image processing techniques for detection of
Houhoulis P.F. and Michener W.K. 2000. Detecting wetland coastal wetlands using ERS-1 SAR data. International Journal of
change: a rule-based approach using NWI and SPOT-XS data. Remote Sensing 21: 565–579.
401

Lee C.T. and Marsh S.E. 1995. The use of archival Landsat MSS Palylyk C.L., Crown P.H. and Turchenek L.W. 1987. Landsat MSS
and ancillary data in a GIS environment to map historical change data for peatland inventory in Alberta, Symposium ’87 Wet-
in an urban riparian habitat. Photogrammetric Engineering and lands / Peatlands 1987. Edmonton, Alberta, Canada, pp 365–371.
Remote Sensing 61: 999–1008. Park R.A., Lee J.K. and Canning D.J. 1993. Potential effects of
Lee J.K. and Park R.A. 1992. Application of geoprocessing and sea-level rise on Puget Sound wetlands. Geocarto International 4:
simulation modeling to estimate impacts of sea level rise on the 99–110.
northeast coast of Florida. Photogrammetric Engineering and Park R.A., Lee J.K., Mausel P.W. and Howe R.C. 1991. Use of
Remote Sensing 58: 1579–1586. remote sensing for modeling the impacts of sea level rise. World
Lee K.H. and Lunetta R.S. 1996. Wetland Detection Methods. In: Resources Review 3: 184–205.
Lyon J.G. and McCarthy J. (eds), Wetland and Environmental
Penuelas J., Gamon J.A., Griffin K.L. and Field C.B. 1993. Asses-
Applications of GIS. Lewis Publishers, New York, pp. 249–284.
sing community type, plant biomass, pigment composition, and
Llewellyn D.W., Shaffer G.P., Craig N.J., Creasman L., Pashley D.,
photosynthetic efficiency of aquatic vegetation from spectral
Swan M. et al. 1996. A decision-support system for prioritizing
reflectance. Remote Sensing of Environment 46: 110–118.
restoration sites on the Mississippi River alluvial plain. Con-
Pietroniro A., Prowse T., Hamlin L., Kouwen N. and Soulis R.
servation Biology 10: 1446–1455.
1996. Application of a grouped response unit hydrological model
Lo C.P. and Watson L.J. 1998. The influence of geographic sam-
to a northern wetland region. Hydrological Processes 10: 1245–
pling methods on vegetation map accuracy evaluation in a
swampy environment. Photogrammetric Engineering and Re- 1261.
mote Sensing 64: 1189–1200. Pietroniro A., Prowse T.D. and Lalonde V. 1995. Classifying terrain
Louiselle S., Bracchini L., Bonechi C. and Rossi C. 2001. Model- in a muskeg-wetland regime for application to GRU-type distrib-
ling energy fluxes in remote wetland ecosystems with the help of uted hydrologic models. Canadian Journal of Remote Sensing
remote sensing. Ecological Modelling 145: 243–261. 22: 45–52.
Luczkovich J., Wagner T.W., Michalek J.L. and Stoffle R.W. 1993. Pope K.O., Rejmankova E., Savage H.M., Arredondo-Jimenez J.I.,
Discrimination of coral reefs, seagrass meadows, and sandy Rodriguez M.H. and Roberts D.R. 1994. Remote sensing of
bottom types from space: a Dominican Republic case study. tropical wetlands for malaria control in Chiapas, Mexico.
Photogrammetric Engineering and Remote Sensing 59: 385– Ecological Applications 4: 81–90.
389. Raabe E.A. and Stumpf R.R. 1996. Image processing methods:
Lulla K. 1983. The Landsat satellites and selected aspects of procedures in selection, registration, normalization, and enhance-
physical geography. Progress in Physical Geography 7: 1–45. ment of satellite imagery in coastal wetlands. Open File Report
Lunetta R.S. and Barlogh M.E. 1999. Application of multi-tempo- 97–287. United States Geologic Survey.
ral Landsat 5 TM imagery for wetland identification. Photo- Ramsey E.W. III, Chappell D.K. and Baldwin D.G. 1997. AVHRR
grammetric Engineering and Remote Sensing 65: 1303–1310. imagery used to identify hurricane damage in a forested wetland
MacDonald T.A. 1999. Wetland rehabilitation and remote sensing. of Louisiana. Photogrammetric Engineering and Remote Sensing
In: Streever W. (ed.), An International Perspective on Wetland 63: 293–297.
Rehabilitation. Kluwer Academic Publishers, Boston, pp. 251– Ramsey E.W. III, Chappell D.K., Jacobs D., Sapkota S.K. and
264. Baldwin D.G. 1998. Resource management of forested wetlands:
Macleod R.D. and Congalton R.G. 1998. A quantitative comparison hurricane impact and recovery mapping by combining Landsat
of change-detection algorithms for monitoring eelgrass from TM and NOAA AVHRR data. Photogrammetric Engineering and
remotely sensed data. Photogrammetric Engineering and Remote Remote Sensing 64: 733–738.
Sensing 64: 207–216. Ramsey E.W. and Laine S.C. 1997. Comparison of Landsat
McCarthy T.S., Franey N.J., Ellery W.N. and Ellery K. 1993. The Thematic Mapper and high resolution photography to identify
use of SPOT imagery in the study of environmental processes of change in complex coastal wetlands. Journal of Coastal Research
the Okavango Delta, Botswana. South African Journal of Sci- 13: 281–292.
ences 89: 432–436. Reed P.B. Jr 1988. National List of Plants that Occur in Wetlands:
Mertes L.A.K., Daniel D.L., Melack J.M., Nelson B., Martinelli National Summary. US Fish and Wildlife Service Biological
L.A. and Forsberg B.R. 1995. Spatial patterns of hydrology, Report 88(24).
geomorphology, and vegetation on the floodplain of the Amazon Reimold R.J., Gallagher J.L. and Thompson D.E. 1973. Remote
River in Brazil from a remote sensing perspective. Geomorphol- sensing of tidal marsh. Photogrammetric Engineering and Re-
ogy 13: 215–232. mote Sensing 477–488.
Mitsch W.J. and Gosselink J.G. 1993. Wetlands. 2nd edn. Van Rutchey K. and Vilcheck L. 1994. Development of an Everglades
Nostrand Reinhold, New York, 722 pp. vegetation map using a SPOT image and the Global Positioning
Munyati C. 2000. Wetland change detection on the Kafue Flats, System. Photogrammetric Engineering and Remote Sensing 60:
Zambia, by classification of a multitemporal remote sensing 767–775.
image dataset. International Journal of Remote Sensing 21: Rutchey K. and Vilchek L. 1999. Air photointerpretation and
1787–1806. satellite imagery analysis techniques for mapping cattail cover-
Narumalani S., Jensen J.R., Burkhalter S., Althausen J.D. and age in a northern Everglades impoundment. Photogrammetric
Mackey H.E. Jr 1997. Aquatic macrophyte modeling using GIS Engineering and Remote Sensing 65: 185–191.
and logistic multiple regression. Photogrammetric Engineering Sader S.A., Ahl D. and Liou W.-S. 1995. Accuracy of Landsat TM
and Remote Sensing 63: 41–49. and GIS rule-based methods for forest wetland classification in
Nayak S.R. and Sahai B. 1985. Coastal morphology: a case study of Maine. Remote Sensing of Environment 53: 133–144.
the Gulf of Khambhat (Cambay). International Journal of Re- Schowengerdt R.A. 1997. Remote Sensing. 2nd edn. Academic
mote Sensing 6: 559–567. Press, 522 pp.
402

Singh A. 1989. Digital change detection techniques using remotely development for South Florida’s National Parks and Preserves.
sensed data. International Journal of Remote Sensing 10: 989– Photogrammetric Engineering and Remote Sensing 61: 1371–
1003. 1381.
Spanglet H.J., Ustin S.L. and Rejmankova E. 1998. Spectral Wickware G.M. and Howarth P.J. 1981. Change detection in the
reflectance characteristics of California subalpine marsh plant Peace-Athabasca Delta using digital Landsat data. Remote Sens-
communities. Wetlands 18: 307–319. ing of Environment 11: 9–25.
Tiner R.W. 1990. Use of high-altitude aerial photography for Wilkinson G. and Shepherd I. 1995. Testing the water: monitoring
inventorying forested wetlands in the United States. Forest and modelling wetlands using integrated GIS and RS techniques.
Ecology and Management 33: 593–604. GIS Europe 4: 38–40.
Townsend P.A. and Walsh S.J. 1998. Modeling floodplain inunda- Work E.A. Jr and Gilmer D.S. 1976. Utilization of satellite data for
tion using an integrated GIS with radar and optical remote inventorying prairie ponds and lakes. Photogrammetric En-
sensing. Geomorphology 21: 295–312. gineering and Remote Sensing 42: 685–694.
Weismiller R.A., Kristof S.J., Scholz D.K., Anuta P.E. and Momin Yi G.-C., Risley D., Koneff M. and Davis C. 1994. Development of
S.A. 1977. Change detection in coastal zone environments. Ohio’s GIS-based wetland inventory. Journal of Soil and Water
Photogrammetric Engineering and Remote Sensing 43: 1533– Conservation 49: 23–28.
1539. Zainal A.J.M., Dalby D.H. and Robinson I.S. 1993. Monitoring
Welch R., Madden M. and Doren R.F. 1999. Mapping the Ever- marine ecological changes on the East Coast of Bahrain with
glades. Photogrammetric Engineering and Remote Sensing 65: Landsat TM. Photogrammetric Engineering and Remote Sensing
163–170. 59: 415–421.
Welch R., Remillard M. and Doran R.F. 1995. GIS database

You might also like