PERS March 2015 Members - page 203

and Green, 2009) were calculated to compare the RF classifica-
tion of green, desiccated, and dead tamarisk canopies using
full-range field spectra, feature-selected field spectra, 8-band
simulated WorldView-2 spectra, and bands 5 through 8 from
simulated WorldView-2 spectra.
Results
CCSM Results
Green, desiccated, and dead tamarisk canopy spectra had
important differences in their reflectance spectra, but yellow
desiccated and brown desiccated spectra were similar in terms
of amplitude and shape (Figure 1). Mean green and desiccated
canopy spectra possessed a steep increase in reflectance beyond
visible wavelengths (>700 nm, referred to as the “red edge”),
while the mean dead canopy spectrum showed a continual,
gradual increase in reflectance. The contrast between the red
and NIR spectral regions was highest for the mean green canopy
spectrum, similar for both mean desiccated canopy spectra, and
lowest for the mean dead canopy spectrum. The region with
the highest reflectance (750-1,300 nm), called the “near infrared
plateau,” contains a unique plant spectral feature. Two liquid
water absorption features were evident on the NIR plateau
centered near 980 and 1,200 nm for both green and desiccated
canopy spectra. The NIR plateau is produced by high internal
leaf scattering at cell wall interfaces (Nilson and Olsson, 1995;
Kokaly et al., 2003). On the canopy level, NIR reflectance is fur-
ther enhanced by multiple leaf layers which can amplify the al-
ready large difference in red and NIR reflectance of single leaves
(Knipling, 1970). These spectral features are also influenced
by proteins, lignin and cellulose (Kokaly et al., 2003). Previous
studies have indicated that the red and NIR regions can be used
effectively for monitoring the active biomass of plant canopies
and even vegetation vigor (Tucker, 1979, Rock et al., 1988).
The SWIR spectral region has been identified as sensitive to
vegetation moisture and senescence, and thus may be suitable
for discrimination of green, desiccated and dead vegetation
(Nagler et al., 2000; Nagler et al., 2003; Inoue et al., 2008;
Piekarczyk et al., 2012). In the SWIR region, differences in am-
plitude can be observed among all the tamarisk canopy types
(Figure 1), due to the changes in foliar water content (Gates
et al., 1965): the mean dead canopy spectrum had the high-
est reflectance value, the mean green canopy spectrum had
the lowest, and the mean desiccated canopy spectra (brown
and yellow) were in the middle but similar. The mean and
±1 standard deviation dead canopy spectra show evidence of
lignin and cellulose absorption in the SWIR between 2,000 and
2,200 nm, while this spectral feature is not readily apparent
in green or desiccated canopy spectra (Figure 1).
CCSM results were consistent with the visual assessment of
reflectance spectra (Figure 2): in comparison to other canopy
type combinations, spectra of yellow and brown desiccated
canopies were very similar with little difference, and the dead
canopy spectrum was significantly different from the spectra of
other canopy types. The cross correlogram with the reference
spectrum itself was parabolic and nearly symmetrical with a
correlation maximum up to 1 at match position 0, indicating a
perfect match. Canopy type combinations using the green, yel-
low desiccated, and brown desiccated canopy spectra as refer-
ence spectra showed that the dead canopy spectrum was dis-
tinct. The dead canopy cross correlograms were highly skewed
with much smaller correlation coefficients, and the correlation
peak was shifted towards positive match positions (Figure 2).
The cross correlograms for yellow/brown desiccated canopy
T able 2. RF Confusion Matrices for Dead, G reen, and Desiccated C anopies Using Full-Range and Feature-Selected Canopy Spectra
Full-range spectra Dead canopies Green canopies Desiccated canopies User’s accuracy (%)
Dead canopies 12 0 3 80.0
Green canopies 0 12 5 70.6
Desiccated canopies 2 3 30 85.7
Producer’s accuracy (%) 85.7 80.0 78.9
OOB error (%), Kappa and run time (seconds) 17.9 0.678 1.37
Feature-selected spectra
Dead canopies 13 0 2 86.7
Green canopies 0 13 4 76.5
Desiccated canopies 2 3 30 85.7
Producer’s accuracy (%) 86.7 81.3 83.3
OOB error (%), Kappa and run time (seconds) 16.4 0.730 0.26
Table 3. RF Confusion Matrices for Dead, Green, and Desiccated Canopies Using Simulated Full and Feature-S elected WorldView-2 Spectra
Full WV2 spectra Dead canopies Green canopies Desiccated canopies User’s accuracy (%)
Dead canopies 13 0 2 86.7
Green canopies 0 13 4 76.5
Desiccated canopies 3 1 31 88.6
Producer’s accuracy (%) 86.7 92.3 83.8
OOB error (%), Kappa and run time (seconds) 14.93 0.753 0.08
Feature-selected spectra
Dead canopies 12 0 3 80.0
Green canopies 0 14 3 82.4
Desiccated canopies 2 1 32 91.4
Producer’s accuracy (%) 85.7 93.3 84.2
OOB error (%), Kappa and run time (seconds) 13.43 0.776 0.07
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING March 2015 203