cial image
:: ERDAS
1992) and
man et al,
1983). The
o delineate
hree image
obtained to
d accuracy
m various
t of the GIS
e employed
on spectral
eled cap
iches were
on scheme
tland
id
tland
|
Accuracy
of stratified
d accuracy
uggested as
The Federal
* 3x3 pixels
analysis, the
d into four
d
land
id
pland
griculture
sources and
on of these
port wetland
NWI maps,
‚odel (DEM)
and hydrography derived from digital line graph (DLG) compiled
bv the USGS. as the components of GIS data base. Through
digitizing. editing. vector to raster conversion and registering
procedures, the separate GIS layers were prepared in raster
format using ERDAS software.
35 Statistical Analysis Model
The cross tabulation error matrix method was applied to evaluate
the statistical association between the GIS variables and their
contribution to predict the location of forested wetland.
Furthermore, two analytical models support and verify the results
derived from the cross tabulation error matrix method.
Consequently, a discrete multivariate analysis technique was
applied to assess and compare classification confusion matrices.
3.6 Integrated Remote Sensing and GIS Model
The integrated model was a logical approach based on an
integration of four weighted GIS layers combined with the best
classified image that stratifies the forest and non-forest area. The
coding of the two basic groups (forest and non-forest) is 16 and
0 respectively. Each basic group was separated into two groups
according to the rule of accumulated weight factor in the analysis
matrix. The aggregated higher weight in each group will be
assigned as forested wetland and other wetland separately. The
aggregated lower weight in cach group will be assigned as
forested upland and other upland. The procedures of constructing
the integrated model follow three steps:
1) Evaluate and weight the GIS layers: Each GIS layer was
evaluated by cross-tabulation and analytical models to see what
percentage of the forested wetland types conformed to the
reference plots. The higher co-occurrence of GIS layers reflects
higher probability of detecting forested wetland.
2) Assign binary recoding system to exaggerate a maximum
range for data analysis: According to the evaluation results, a
binary weight was assigned as a confidence value corresponding
lo forested wetland probability. The weighted GIS layers
separate all combination levels by at least one unit and extend the
possible levels to the maximum. In this case, four GIS layers will
be assigned a weight of 1, 2, 4, 8 respectively. An analysis image
with 32 combination levels (16 for each forest or non-forest) can
be generated by map algebra of the additive approach with four
GIS layers.
3) Formulate a general rule to optimize aggregation process:
An analysis matrix can be generated by cross tabulating the
reference data against the additive image. The aggregation rule
d defined by setting the balanced ratio threshold in the two
basic land cover groups (forest and non-forest). The balanced
lalios divided forest area into forested wetland and forested
upland and non-forest area into other wetland and other upland.
The higher accumulated ratio represented the forested wetland
and the other wetland and the lower accumulated ratio represent
the other upland in each group. The final ratios between
Categories will control the producer agreement of the confusion
421
matrix. The balanced ratio is intended to develop a homogeneous
confusion matrix which reflects the highest agreement between
user and producer accuracy.
4.0 RESULTS AND DISCUSSION
4.1 Comparison of Conventional Image Classification
Results Between Two study Sites
Table 1 indicates that the accuracy (overall and Kappa) improved
consistently in three conventional image classification
approaches at both study sites. The overall accuracy was similar
between study sites. However, the overall Kappa of Orono were
10% higher than Acadia. Compared to Acadia study site, the
Orono study site had more comparable user and producer
agreement which led to form a homogeneous confusion matrix.
The poorer overall Kappa of the Acadia study, especially the
unsupervised approach, might be explained by the difference in
the physical characteristics of the two study sites. The larger area
has potential to affect the signature evaluation during the
aggregation process. The highly uneven sample size made
Acadia difficult to form a homogeneous confusion matrix. In this
case, the Kappa correction might more properly represent the
accuracy assessment between the two study sites
The different sample size used in the three approaches was the
result of the majority rule plot editing procedure. This might
relate to the different classification methods. For example, the
supervised classification produced greater scatter of classified
pixels representing different cover types in the image. The pixel
of the same cover type (e.g. forested wetland) did not always
cluster within a 3X3 pixel neighborhood. As a result, the sample
was reduced when many sample plots failed to meet the majority
rule criterion.
All in all the results might be indicative of the Landsat TM
sensors’ insensitivity to distinguish spectral similarities of
forested wetland versus forested upland. The accuracy of
conventional classification methods might be affected by the size
of study area. The conventional image classification approaches
indicated very little confidence in the spectral based methods to
delineate forested wetlands.
4.2 Evaluation of GIS Layers
The GIS layers were evaluated for the prospect of modeling. They
were evaluated in cross tabulation which was performed in the
same manner as accuracy assessment of the individual GIS layer.
Hydric soils, slope, and National Wetland Inventory data were the
most important variables in the integrated GIS model. However,
the relative importance of the model variables in predicting
wetland conditions differed between study sites
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996