International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
data were treated as independent spectral classes within
each informational class. This means that the selected
patches were not aggregated into composite statistics
for the seven classes. The likelihood index for each
patch was computed for all individual training data set
patches by the divergence index, the maximum
likelihood using the patch mean, and the patch pdf.
Small patches with fewer than six pixels were excluded
from the region-based maximum likelihood analysis,
and treated as part of the “melt pond,” to use McDevitt
and Peddada's (1998) term. There were two reasons
for identifying melting pond pixels. Firstly, because
five variables were used in this study, patches with
fewer than six pixels had less than the minimum
number of pixels potentially required to characterize
the multivariate statistics. Secondly, the melting pond
was assumed to represent objects that are not of direct
interest, but rather extraneous objects such as cars, or
chimneys on buildings.
In the third stage, each patch was classified into seven
classes by the suggested methods. For maximum
likelihood with patch pdf, the range over which the pdf
was calculated was limited to three standard deviations.
The pdf is very low outside of this range, and is not
expected to have much significance in the calculation.
Excluding pdf values greater than three standard
deviations has the advantage of reducing the computing
cost.
Figure 4 shows a one dimensional representation of the
process. Within the pdf overlap region, the decision
range was divided into ten equal cells. The
probability of the center of each cell calculated for both
the training and the patch classes, and the lower of the
two probabilities is used for the cell height. After
multiplying cell height by the width, the cell area is
calculated. The total area of the overlap is then
estimated by summing the cell areas (Figure 4). This
procedure is modified for the multivariate case by
dividing the multidimensional overlap region into 10"
cells, where n is the number of bands. For two bands
a volume of the overlap region is calculated, and for
three or more bands a hypervolume is calculated. For
this work, five bands were used, thus, 10° cells were
calculated for each likelihood index.
Training patch Tested patch
Decision making range %
Figure 4. Maximum likelihood calculation utilizing
patch pdfs.
The patch was assigned to the class with the highest
likelihood after the unknown patch is compared with
each patch in the training data set. In the next step of
the classification, melting pond pixels are classified.
These small patches are treated as noise, and therefore
assigned to an adjacent class. If the patch is
surrounded by a single class, it is assigned to that class.
In the general case, however, the patch is adjacent to
more than one class. In this case, the patch is assigned
to the adjacent class with the most similar DN values in
the green band (Band 2). A more sophisticated,
multivariate approach was not used because of the
small sample size of these patches. In the final step,
adjacent patches of the same class were merged to form
objects.
ERDAS Imagine was used to conduct the traditional
pixel-based classifications. The unsupervised
ISODATA program (Tou and Gonzalez, 1974;
ERDAS, 1999) was executed with 24 clusters. After
classification, the 24 clusters were assigned empirically
to the most appropriate class among the seven classes
based on the ground truth and knowledge of the area.
For each of the supervised classification methods, the
same training data sets were used.
S. RESULTS AND DISCUSSION
Figure 5 shows the results from the four previously
mentioned methods. To compare the accuracy of the
four methods, error matrices for the kappa index and
errors of commission and omission were produced
using the IDIRSI program ERRMAT (Eastman, 2003)
(Table 1). Ground reference maps for the accuracy
evaluation were produced using photo-interpretation
and expert knowledge for three parts of the study area:
Downtown Morgantown, a medium density residential
area, and a forested stream valley.
Table 1. Summary accuracy statistics for 7 classes by
the 4 classification methods used in this study.
Blding Road Forest Lawn
CERR 0.363 0.489 0.135 0.440
ISODATA
OERR 0.487 0.320 0.147 0.046
MHL with | CERR | 0.391 0347 | 0063 | 0366
pixel OERR | 0.202 0.412 0.133 0.214
MLH with CERR 0.309 0.291 0.104 0.304
patch mean | OERR | 0.194 0.292 0.042 0.503
Un
MLE with | CERR | om 0.2043 | 0.101 0.294
patchpdf | OERR | 0.170 0.249 0.045 0.480
Shd Veg Water Shadw Kappa
CERR 0.567 0.024 0.062
ISODATA - 0.610
OERR 0.373 0.981 0.352
MHL with CERR 0.440 0.123 0.137
1 0.687
pixel OERR 0.297 0.308 0.359
MLH with | CERR 0.360 0.095 0.048
patch = s s ; 0.735
eat OERR 0.691 0.232 0.398
MLH with; |-.CÉRR 0.306 0.042 0.054 m
patch pdf | OERR 0.603 0.089 0.234 C
The overall kappa value of the supervised pixel-based
classifications was 0.687. The lowest accuracy, 0.610,
was obtained with the unsupervised pixel-based