the time–space variations of the factors affecting the measurements as detected
naturally on the field. In addition, even though variability is high, the level of accuracy
of the methods in collecting spectral data of different land covers on the field is at an
acceptable level (Van Aardt and Wynne, 2007). Statistical techniques and quantitative
methods applied on the field hyperspectral data prior to further digital image analysis
aim to solve the variability between different land cover targets, which, in the case of
vegetation cover, is often very low due to the similar canopy architecture, color, or
physicochemical properties (Williams, 1991; Zang et al., 2008).
One of the main applications of field spectroradiometry is in investigating the
potential means of mapping vegetation physiognomic types along different environmental or phenological gradients. For instance, Artigas and Yang (2006) analyzed
dominant saltmarsh species in New Jersey meadows in the fall using the VNIR
spectrum from 350 to 950 nm with the Mann–Whitney U-test at p < 0.01 and
suggested the orange and red segments of the VIS and few bands in the NIR as
unique for discrimination. Psomas et al. (2005) have used the same statistical test for
comparison between the spectral signatures of four dry-mesic grassland types in
Switzerland. The statistically significant seasonal variation between the spectral
reflectance of the vegetation demonstrated the potential of using spectral information
for discriminating different grassland type changes during one growing period on
satellite imagery. Manevski et al. (2012), using the Mann–Whitney U-test at p < 0.01,
have recently found the distinctive phenology of three Mediterranean trees and shrubs
and the seasonal effect of rapid flowering in spring as important factors determining
the spectral discrimination at the field scale. That vegetation cover spectral variability
is generally not parametric deserves more attention. Even if normality is assumed
when a large number of sampled spectra are available (the central limit theorem), such
analysis in addition requires the assumption of equal reflectance variances to be met
between the targets being compared. This is something which should be considered
(Figure 15.4) but is often not discussed or it is ignored.
Field spectroradiometry has also been used for estimating and mapping vegetation
quality in terms of macroelements such as nitrogen. For example, Mutanga et al.
(2003) investigated the discrimination of the tropical pasture grass Cenchurus ciliaris
under different nitrogen treatments at the canopy level. The results from the
parametric ANOVA tests at p < 0.001 with a 95% confidence level (CL) coupled
on both unaltered and continuum-removed spectra showed significant differences
between the nitrogen treatments, demonstrating the possibility to map variation in
pasture quality using hyperspectral remote sensing.
It is well known that the high-dimensional complexity of hyperspectral data
imposes some issues in terms of image processing algorithms, extensive field
campaigns and high cost. Therefore, identification of the optimal bands is required
for discriminating and mapping vegetation cover without losing important information. For that purpose, Adam and Mutanga (2009) have proposed a hierarchical
method which initially involves one-way ANOVA as an attempt to spectrally
discriminate papyrus (Cyperus papyrus L.) from three other species within the
300–2500-nm spectral region, measured in summer in swamp wetlands in South
Africa. Their results demonstrated that papyrus was significantly different then its
FIELD SPECTRORADIOMETRY
299
naturally on the field. In addition, even though variability is high, the level of accuracy
of the methods in collecting spectral data of different land covers on the field is at an
acceptable level (Van Aardt and Wynne, 2007). Statistical techniques and quantitative
methods applied on the field hyperspectral data prior to further digital image analysis
aim to solve the variability between different land cover targets, which, in the case of
vegetation cover, is often very low due to the similar canopy architecture, color, or
physicochemical properties (Williams, 1991; Zang et al., 2008).
One of the main applications of field spectroradiometry is in investigating the
potential means of mapping vegetation physiognomic types along different environmental or phenological gradients. For instance, Artigas and Yang (2006) analyzed
dominant saltmarsh species in New Jersey meadows in the fall using the VNIR
spectrum from 350 to 950 nm with the Mann–Whitney U-test at p < 0.01 and
suggested the orange and red segments of the VIS and few bands in the NIR as
unique for discrimination. Psomas et al. (2005) have used the same statistical test for
comparison between the spectral signatures of four dry-mesic grassland types in
Switzerland. The statistically significant seasonal variation between the spectral
reflectance of the vegetation demonstrated the potential of using spectral information
for discriminating different grassland type changes during one growing period on
satellite imagery. Manevski et al. (2012), using the Mann–Whitney U-test at p < 0.01,
have recently found the distinctive phenology of three Mediterranean trees and shrubs
and the seasonal effect of rapid flowering in spring as important factors determining
the spectral discrimination at the field scale. That vegetation cover spectral variability
is generally not parametric deserves more attention. Even if normality is assumed
when a large number of sampled spectra are available (the central limit theorem), such
analysis in addition requires the assumption of equal reflectance variances to be met
between the targets being compared. This is something which should be considered
(Figure 15.4) but is often not discussed or it is ignored.
Field spectroradiometry has also been used for estimating and mapping vegetation
quality in terms of macroelements such as nitrogen. For example, Mutanga et al.
(2003) investigated the discrimination of the tropical pasture grass Cenchurus ciliaris
under different nitrogen treatments at the canopy level. The results from the
parametric ANOVA tests at p < 0.001 with a 95% confidence level (CL) coupled
on both unaltered and continuum-removed spectra showed significant differences
between the nitrogen treatments, demonstrating the possibility to map variation in
pasture quality using hyperspectral remote sensing.
It is well known that the high-dimensional complexity of hyperspectral data
imposes some issues in terms of image processing algorithms, extensive field
campaigns and high cost. Therefore, identification of the optimal bands is required
for discriminating and mapping vegetation cover without losing important information. For that purpose, Adam and Mutanga (2009) have proposed a hierarchical
method which initially involves one-way ANOVA as an attempt to spectrally
discriminate papyrus (Cyperus papyrus L.) from three other species within the
300–2500-nm spectral region, measured in summer in swamp wetlands in South
Africa. Their results demonstrated that papyrus was significantly different then its
FIELD SPECTRORADIOMETRY
299