and green canopy combinations were moderately skewed, but
the correlation maxima were still high around match posi-
tion 0; most important of all, the cross correlograms of yellow
and brown desiccated canopies combinations were mostly
indistinguishable. Since the differences between yellow and
brown desiccated canopies were proven to be small and hard
to distinguish, these canopy types were combined into a single
desiccated canopy type for further analysis.
Feature-selected Wavelengths
Figure 3 plots
ISI
values based on the reflectance of green, des-
iccated and dead canopy spectra as a function of wavelength.
As previously explained in the Instability Index Section, low
ISI
values were expected to correspond with low levels of
similarity between canopy types and/or low levels of variabil-
ity within each canopy type indicating high separability. In the
end, the feature-selected wavelengths with low
ISI
values were
found in four spectral regions: one red region (645-693 nm),
one
NIR
region (735-946 nm), and two
SWIR
regions (1,960-2,090
nm and 2,400-2,478 nm). In previous studies, these regions
were sensitive to changes in chlorophyll, damaged leaf layer
structures and loss of water content (Knipling, 1970; Heller,
1978; Boochs
et al
., 1990; Carter, 1993; Radeloff
et al
., 1999;
Lentile
et al
., 2006; Inoue
et al
., 2008; Piekarczyk
et al
., 2012).
Figure 3. ISI values based on the reflectance of green, desic-
cated, and dead canopy spectra, plotted against wavelength.
Atmosphere absorption regions were excluded.
Classification of Full-range and Feature-selected Spectra
RF
confusion matrices,
OOB
error, kappa, and run time were
used to compare the classification performance of full-range
and feature-selected spectra (Table 2 and Table 3). In general,
desiccated canopies were the most difficult to distinguish and
caused most of the classification errors among the three canopy
types. User’s accuracy for green canopies was lowest (70.6 per-
cent) using the full-range spectra, due to the misclassifications
of green and desiccated canopies. In addition, Table 2 showed
that using the feature-selected spectra instead of the full-range
spectra, the OBB error dropped from 17.9 percent to 16.4
percent. Correspondingly, the kappa coefficients of confusion
matrices increased from 0.678 to 0.730. The computation time
of feature-selected spectra decreased significantly compared to
full-range spectra. Using a desktop computer, the run time of
RF
classification for feature-selected spectra was 0.26 seconds,
while the time for full range spectra was 1.37 seconds.
In comparison with the full WorldView-2 spectra, the
feature-selected set of WorldView-2 spectra tended to have
higher accuracy or lower
OOB
error (Table 3). These results
indicate that a process of feature selection is likely beneficial
for improving classification accuracy and reducing the compu-
tation time. Kappa values indicate that separation of canopy
types improved when WorldView-2 bands were used, indicat-
ing that classification of green, desiccated, and dead tamarisk
canopies may not require
SWIR
bands or hyperspectral data.
Based on high accuracies for feature-selected canopy and
simulated WorldView-2 spectra, the same feature-selected
bands were applied to classification of four classes: green
canopy, yellow desiccated canopy, brown desiccated canopy,
and dead canopy. Table 4 shows the classification perfor-
mance for using feature-selected canopy spectra for separa-
tion of desiccated yellow and desiccated brown classes.
User’s accuracy for brown desiccated canopies was 0 percent,
due to the misclassification error between yellow and brown
desiccated canopy spectra. In comparison to the classification
performance based on three canopy types (Table 2),
OOB
error
increased to 32.84 percent from 16.4 percent and kappa coef-
ficient decreased to 0.526 from 0.730 (Table 4). Using the best
classification for three canopy types (four-band WorldView-2
spectra), classification for four classes was attempted again.
OOB
increased from 13.43 percent to 28.36 percent, and at
the same time, kappa coefficient decreased dramatically from
0.776 to 0.585 (Table 3, Table 5). Similar to the results from
CCSM
, Table 4 and Table 5 indicate that combining desiccated
canopies into one type was necessary.
Discussion and Conclusions
Our study contributes to continuing efforts for evaluating the
impacts of tamarisk bio-control in the western United States,
and extends our understanding of vegetation disturbance
monitoring. Compared to previous studies using high spatial
resolution or hyperspectral imagery to map vegetation canopy
disturbance, our analysis using field spectra had similar
classification accuracy. Using high spatial resolution (<4 m)
QuickBird-2 images, Wulder
et al
. (2008) mapped mountain
pine beetle “red attack” with 89 percent to 93 percent accura-
cies in British Columbia, Canada. Based on airborne hyper-
spectral imagery (PROSPECTIR-VS, 2 m spatial resolution)
and field measurements, Santos
et al
. (2010) mapped asymp-
tomatic, senesced and dead trees in a pine forest of southeast-
ern United States using a decision tree method. Stressed tree
mapping had a kappa coefficient of around 0.70, indicating
that asymptomatic trees are likely to have significantly higher
reflectance in the red-
NIR
regions, senescent trees are likely
to have significantly lower reflectance in the
NIR
regions, and
dead trees are likely to have significantly higher reflectance in
the
SWIR
regions.
Our analysis applied wavelength selection techniques on
canopy spectra collected
in situ
before classification by
RF
,
resulting in an increased classification accuracy of tamarisk
canopy types compared to using the full wavelength range
(Asner
et al
., 1998; Somers
et al
., 2010; Somers and Asner,
2013). Using selected wavelengths (bands) that display high
between-class variability and low within-class variability,
not only was higher classification accuracy achieved, but the
computational time was greatly reduced (Table 2 and Table
3). In addition, feature-selected WorldView-2 bands demon-
strated better classification performances over field spectra for
separating green, desiccated, and dead canopy types. Somers
and Asner (2013) hypothesized that redundant spectral infor-
mation in hyperspectral data caused decreased accuracy and
additional computational time in the spectral mixture analy-
sis. Redundant spectral information may have caused lower
classification accuracies for full spectrum canopy spectra and
simulated WorldView-2 spectra used in this study.
In conclusion, this study proposed a methodology to
improve remote monitoring of tamarisk bio-control by a
combination of spectral analysis techniques (
CCSM
,
ISI
, and
204
March 2015
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
