Image Processing Reference
In-Depth Information
6.4.2
IRS-1C Data: Anatomy of a Metropolis
Spectral mixture analysis, as applied to the V-I-S model, is beginning to emerge in
the literature (Rashed et al.
2001
; Wu and Murray
2003
). Rashed et al. (
2001
)
employed four-spectral-band Indian IRS-1C data to study the anatomy of Greater
Cairo in terms of end-member fractions through spectral mixture analysis (SMA).
Key features of this investigation are:
•
Spectral mixture analysis
•
Decision tree analysis
•
Comparative accuracy assessment
Four end-members were selected: vegetation, impervious surface, soil, and shade
from pure pixel end-member samples. The resulting fractions were used to classify
the urban scene through a decision tree classifier. The V-I-S model provided the link
to the spectral mixing procedure because the spectral contribution of each of the
three V-I-S components, and shade, can be resolved at a sub-pixel level through
SMA. Shade was added as an end-member to accommodate the shadowing of tall
buildings, a common feature in large cities, and grouped spectrally with water as dark
objects. The end-member spectra were derived from the image. Two different soils
were tested, one lighter, one darker. A four end-member SMA with the darker soil
produced the best results.
Figure
6.12a-d
displays the four fractions. The shade/water fraction highlights the
through-flowing Nile River. A binary decision tree was used to further classify the
results into eight cover and land use types. The results were tested against two tradi-
tional per pixel classifiers, maximum likelihood (ML) and minimum distance to means
(MDM). Accuracies were determined through use of the co-registered 10 m IRS pan-
chromatic band and field familiarity. Table
6.5
shows that, as hypothesized, the SMA/
decision tree fractions are more accurate than either the ML or MDM classifications.
This study shows the value of the V-I-S model in providing a direct identification
of biophysical features so important in urban area remote sensing. The study also
demonstrates the flexibility of the model in sub-pixel analysis, SMA endmember
determination, and adding shadowing as an important element of urban features of
varying heights, although spectrally virtually indistinguishable from water.
6.4.3
Thematic Mapper Data: Multivariate Calibration
and Color Display
Card (
1993
) used the six reflective bands of TM data to distinguish four V-I-S
classes: green vegetation, dry vegetation, soil, and impervious surface, in Salt Lake
City. Enlarged black and white photography (1:7,920 scale) was used for calibration
and accuracy assessment. Key features of this investigation are:
•
Transforming raw TM data to radiance values
•
Dark object subtraction for atmospheric correction