coexisting species in 412 wavelengths at p < 0.05 (95% CL), with maximized
discrimination in the red edge (741–746 nm) and the NIR (982–1297 nm). Similarly,
Lacar et al., (2001) explored the differences between four grape vine (Vitis vinifera)
varieties, that is, cabernet sauvignon, merlot, semillon, and shiraz, and their pairs, in a
southern Australian vineyard. The authors found ANOVA coupled with the Tukey
posthoc test [National Institute of Standards and Technology (NIST), 2003] powerful
enough to identify the red edge (∼720 nm) followed by the green reflectance peak and
its wings in the VIS where the field reflectance spectra showed the greatest statistically
significant difference.
Post hoc tests are designed for situations in which significant test results have
already been obtained and additional exploration is needed on which means are
significantly different from each other. A post hoc test would control the type of error
for all possible comparisons between the groups being analyzed. Manevski et al.
(2011) have used the Bonferroni post hoc test (Abdi, 2007) at both p < 0.05 and
p < 0.01 to study the spectral discrimination of common Mediterranean tree and shrub
land cover at the field scale. The Bonferroni test is conservative and strict when the
number of comparisons exceeds the number of degrees of freedom between groups.
However, the type II error rates are high for individual tests, that is, the Bonferroni
overcorrects for type I error. Though significant spectral differences were concluded,
significant data reduction was not achieved. This might be due to the use of the
continuum-removed reflectance, which enhanced absorption pits and reflection peaks,
yielding significantly high number of wavelengths where the reflectances of the plants
are statistically different.
While originally used in spectroscopy of rocks and minerals, the continuum
removal has been recently used in field spectroradiometry of vegetation cover.
Schmidt and Skidmore (2003) attempted to discriminate 27 different Dutch saltmarsh
vegetation types measured in the 400–2500-nm spectral region in summer using the
Mann–Whitney U-test at p < 0.01. Despite the finding that for some of the vegetation
species there is no single band that is significantly different to all other vegetation
communities, still the authors found the approach robust enough to determine the NIR
plateau around 771 nm as an important spectral region for vegetation discrimination.
The red edge at 707 nm was reported as having the lowest frequency of statistically
significant different medians of reflectance of vegetation types.
Finally, the synergistic use of field spectroradiometry and hyperspectral remote
sensing imagery for land cover characterization has not been as extensively explored.
Methods such as spectral distance analysis or spectral feature selection algorithms have
been applied in some studies for selection of a single band or the “best” band
combination for vegetation species discrimination. Yet, one can argue that such
methods of dimension reduction objectively evaluate the intrinsic redundancy of the
hyperspectral data, taking into consideration the biochemical and physical processes
such data mimic in vegetation land cover. Moreover, most of those field scale studies
have not been tested on hyperspectral data acquired from airborne sensors, with only
few previous studies bearing explicit discrimination (van Aardt and Wynne, 2001,
2007). Continuation of such work has special importance in demonstrating a limited
FIELD SPECTRORADIOMETRY
301
discrimination in the red edge (741–746 nm) and the NIR (982–1297 nm). Similarly,
Lacar et al., (2001) explored the differences between four grape vine (Vitis vinifera)
varieties, that is, cabernet sauvignon, merlot, semillon, and shiraz, and their pairs, in a
southern Australian vineyard. The authors found ANOVA coupled with the Tukey
posthoc test [National Institute of Standards and Technology (NIST), 2003] powerful
enough to identify the red edge (∼720 nm) followed by the green reflectance peak and
its wings in the VIS where the field reflectance spectra showed the greatest statistically
significant difference.
Post hoc tests are designed for situations in which significant test results have
already been obtained and additional exploration is needed on which means are
significantly different from each other. A post hoc test would control the type of error
for all possible comparisons between the groups being analyzed. Manevski et al.
(2011) have used the Bonferroni post hoc test (Abdi, 2007) at both p < 0.05 and
p < 0.01 to study the spectral discrimination of common Mediterranean tree and shrub
land cover at the field scale. The Bonferroni test is conservative and strict when the
number of comparisons exceeds the number of degrees of freedom between groups.
However, the type II error rates are high for individual tests, that is, the Bonferroni
overcorrects for type I error. Though significant spectral differences were concluded,
significant data reduction was not achieved. This might be due to the use of the
continuum-removed reflectance, which enhanced absorption pits and reflection peaks,
yielding significantly high number of wavelengths where the reflectances of the plants
are statistically different.
While originally used in spectroscopy of rocks and minerals, the continuum
removal has been recently used in field spectroradiometry of vegetation cover.
Schmidt and Skidmore (2003) attempted to discriminate 27 different Dutch saltmarsh
vegetation types measured in the 400–2500-nm spectral region in summer using the
Mann–Whitney U-test at p < 0.01. Despite the finding that for some of the vegetation
species there is no single band that is significantly different to all other vegetation
communities, still the authors found the approach robust enough to determine the NIR
plateau around 771 nm as an important spectral region for vegetation discrimination.
The red edge at 707 nm was reported as having the lowest frequency of statistically
significant different medians of reflectance of vegetation types.
Finally, the synergistic use of field spectroradiometry and hyperspectral remote
sensing imagery for land cover characterization has not been as extensively explored.
Methods such as spectral distance analysis or spectral feature selection algorithms have
been applied in some studies for selection of a single band or the “best” band
combination for vegetation species discrimination. Yet, one can argue that such
methods of dimension reduction objectively evaluate the intrinsic redundancy of the
hyperspectral data, taking into consideration the biochemical and physical processes
such data mimic in vegetation land cover. Moreover, most of those field scale studies
have not been tested on hyperspectral data acquired from airborne sensors, with only
few previous studies bearing explicit discrimination (van Aardt and Wynne, 2001,
2007). Continuation of such work has special importance in demonstrating a limited
FIELD SPECTRORADIOMETRY
301