dominant tree species and/or the hvdrologic characteristic of
wetlands are the primary features used to separate upland from
forested wetlands (Tiner. 1990). Due to the limitations of the
photo interpretation process, certain wetland types such as
evergreen forested wetlands. temporarily flooded deciduous
forested wetlands. and hvdrologically altered forested wetlands
arc among the most difficult wetlands to detect. In such cases,
subtle photo signatures, topographic position on the landscape,
collateral hydric soil information and field work must be closely
examined to aid in the interpretation process (The Federal
Geographic Data Committee. 1992).
2.2 Satellite Imagery for Wetland Mapping
The use of satellite imagery for wetland mapping provides a
number of advantages over conventional aerial photographs. For
example, wetlands are very dynamic and have tremendous
seasonal or yearly changes. Furthermore. satellite sensors cover
broader wavelengths through their optical scanner systems. Each
detector of the satellite scanner is positioned to record specific
wavelength of energy. Although aerial photography may be
appropriate for high resolution cartography, satellite imagery is
better suited and less costly for rapid, repeated observations over
broad regions assuming equipment and experienced human
analysts are available to process the data (Ferguson et al., 1993).
Besides, the major advantage of satellite data is it's digital
format, making automatic analvses possible and the integration
with GIS.
Satellite imagery may offer an efficient means of identifying
forested wetlands in the future. Landsat MSS, TM and SPOT
HRV imagery have been used successfully to detect major
categories of wetlands (Haddad and Ekberg, 1987; Jensen et al,
1993). However. they have not been used previously to map or
monitor forested wetlands for regional or national coverage.
2.3 Modelling Approaches
The primary reason for employing an integration of GIS and
expert system ( ES ) is to reduce human labor and improve the
consistency of results. A knowledge-based GIS and ES to
manage wetland was developed by the USFWS (Wei et al,
1992). The model, based on Mitchell's (1991) wetland value
assessment ~~ methodology, integrated ~~ knowledge-based
capabilities into a GIS. However. the approach was limited to
monitoring, producing maps, and analyzing changes in wetland
habitat and indicated a need to integrate knowledge of wetland
expertise (Wei et al., 1992).
Hepner et al. (1990) compared neural networks ( NN )
techniques with conventional supervised classification method.
They concluded that NN offered a potential approach to land
cover classification. The advantage of NN is it's abilitv to handle
multispectral, multitemporal, and multisource spatial data (Civco
and Wang, 1994). A NN model which incorporated multidate
Landsat TM image with ancillary spatial information had an
accuracy of approximately eight percent greater than a traditional
single date image classification approach. However, the result for
forested wetland were rather poor (Civco and Wang, 1994).
420
3.0 METHODS
3.1 Equipment and Data Acquisition
The experiments utilized three different commercial image
processing and GIS computer software packages: ERDAS
(1992), ARC/INFO (ESRI, 1990), IDRISI (Eastman, 1992) ang
three statistical programs: Cart methodology (Breiman ef 4
1984), LIMDEP (Greene, 1992), Kappa (Congalton, 1983). The
Landsat TM July, 1991 images were acquired to delineate
forested wetland and other vegetation types using three image
classification methods. Aerial photographs were obtained to
support standard plot creation for data analysis and accuracy
assessment. Land characteristics data derived from various
sources were essential components in the development of the GIS
data base.
3.2 Conventional Image Classification
Conventional techniques of image classification were employed
to understand the limitations of classification based on spectral
signature ^ characteristics. ~~ Unsupervised, ^ tasseled cap
transformation, and hybrid classification approaches were
conducted. The generalized land cover classification. scheme
includes the following classification types:
1. Urban / Open / Agriculture — 6. Softwood Wetland
2. Softwood Upland 7. Mixed Wetland
3. Mixed Forested Upland 8. Hardwood Wetland
4. Hardwood Upland 9. Shrub Upland
5. Non-forested Wetland
3.3 Sample Design for Data Analysis and Accuracy
Assessment
The purpose of the sample design was to create a set of stratified
random plots as reference for data analysis and accuracy
assessment. Three pure pixels on a side (3X3) were suggested a
a minimum area to identify an object on an image (The Federal
Geographic Data Committee, 1992). À plot size of 3x3 pixels
was defined as the sample unit. In order to facilitate analysis, the
nine land cover classification scheme was grouped into fou
super categories as follows:
1. Forested Wetland 3. Forested Upland
Softwood Wetland Softwood Upland
Mixed Wetland Mixed Upland
Hardwood Wetland Hardwood Upland
2. Other Wetlands 4. Other Upland
Non-forested Wetland ~~ Urban / Open / Agriculture
Shrub Upland
3.4 GIS Data Base Development
The GIS layers were derived from existing map Sources and
selected under the assumption that the combination of these
layers indicate a physical environment that may support wel
conditions. Consequently, this investigation selected NWI maps,
hydric soils mapped by the SCS, a digital elevation model (DEM)
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996