H
yperspectral
N
arrowbands
in
the
S
tudy
of
V
egetation
and
A
gricultural
C
rops
Over the years, extensive research has been conducted in
identifying optimal (best) HNBs for study of vegetation and
agricultural crops (Thenkabail
et al.
, 2000; Thenkabail, 2002;
Thenkabail
et al.
, 2004a; Thenkabail
et al.
, 2004b; Blackburn,
2007; Thenkabail
et al.
, 2011b; Miphokasap
et al.
, 2012;
Gitelson, 2013; Mariotto
et al.
, 2013; Thenkabail
et al.
, 2013;
Marshall and Thenkabail, 2014). A review of these and meta-
analysis leads to 28 HNBs (Table 2, Figure 4a) that are non-
redundant and optimal in studying a wide range of agricultural
crops and vegetation. An overwhelming proportion of the 28
HNBs were in short-wave infrared: 7 in far SWIR (FSWIR),
and 4 in early SWIR (ESWIR). This was followed by 4 in green,
3 each in far near infrared (FNIR), near infrared (NIR), and
red-edge, 1 in red, and 1 in ultraviolet (Table 2, Figure 4a). The
advantage of using optimal spectral analysis (OSA) involving
optimal HNBs (Figure 4a, Table 2), as opposed to continuous
spectra (i.e., every HNB in 400 to 2500 nm) or whole spectra
analysis (WSA), are several. These include:
1.
Avoiding a large number of redundant data (~88%) and
focusing on utilizing non-redundant bands (~12%), which
in turn helps in overcoming Hughes’ phenomenon;
2.
Constituting specific hyper spectral vegetation’s indices
(HVIs) from HNBs;
3.
Obtaining the same or nearly the same classification
accuracies using optimal 28 HNBs as opposed to a full
range of bands (e.g., 242 bands of Hyperion), because
accuracies asymptote after a certain number of
wavebands (e.g., ~15 to 20 HNBs attain >90% accuracy
in classifying 7 vegetation classes as shown in Figure 3);
and
4.
Increasing the computation speed and optimizing the
resources in computing and analyzing;
Nevertheless, there is considerable debate for using whole
spectra analysis (e.g., continuous and entire spectra over
700-740 nm) using such methods as partial least squares
regression, wavelet analysis, continuum removal, and spectral
angle mapper (Nielsen, 2001; Delalieux
et al.
, 2009; Thenkabail
et al.
, 2011b; Mirzaie
et al.
, 2014). The use of WSA is justified
when:
1.
Spectral signatures of objects need to be matched with
spectral signatures from existing spectral library;
2.
Integrated spectra over a continuum (e.g., first-order
derivative greenness vegetation index over 600 to 760
nm, 700 to 740 nm, or integrated over other HNBs) need
to be taken advantage of; and
3.
Computing power and other resources are not a
limitation.
It must be noted that the 28 HNBs (Table 2, Figure 4a)
discussed here are limited to the 400 to 2500 nm spectral
domain. There is substantial potential to use thermal
hyperspectral wavebands (Schlerf
et al.
, 2012) in addition
to these HNBs. Therefore, there should be considerable
effect for further advances in developing optimal HNBs in
the study of vegetation and agricultural crops if we include
thermal hyperspectral bands. Moreover, recent efforts involve
combining LiDAR, Hyperspectral, and Thermal (G-LiHT)
imagery (Cook
et al.
, 2013), which will further advance our
understanding in classifying, monitoring, modeling, and
mapping vegetation and agricultural crops (Ribeiro da Luz
and Crowley, 2010).
H
yperspectral
V
egetation
I
ndices
in
the
S
tudy
of
V
egetation
and
A
gricultural
C
rops
Hyperspectral vegetation indices (HVIs) (Haboudane
et al.
,
2004; Bian
et al.
, 2010; Galvão, 2011; Roberts, 2011; Thenkabail
et al.
, 2011b; Gitelson, 2013; Thenkabail
et al.
, 2013)
allow us
to target studies on very specific characteristics of vegetation
and agricultural crops such as biomass, leaf area index (LAI),
pigments (e.g., chlorophyll, carotenoid, anthocyanin), stress
(e.g., due to drought, disease), management properties (e.g.,
nitrogen, tillage), and other biochemical properties (e.g., lignin,
cellulose, plant residue) (Haboudane
et al.
, 2004; Blackburn,
2007; Thenkabail
et al.
, 2011b). There is a potential to have an
index for each of these characteristics (Table 3). The biggest
limitation of broad band indices derived from sensors such as
Landsat is, more or less, that one index such as NDVI is used
for studying all vegetation or crop characteristics. In contrast,
HVIs have following specific advantages (Table 3):
1.
Establishe unique indices to study specific vegetation
or crop variable (e.g., hyperspectral water\moisture
indices or HWMIs to study plant water or moisture;
hyperspectral biomass and structural indices or HBSIs
to study biomass; hyperspectral biochemical indices or
HBCIs to study plant pigments and so on; see Table 3).
2.
Provide significant improvement by explaining ~10%
to 30% greater variability over broadband indices
in modeling and mapping vegetation biophysical
and biochemical properties (Haboudane et al., 2004;
Thenkabail et al., 2011b; Bolton and Friedl, 2013;
Mariotto et al., 2013; Thenkabail et al., 2013); and
3.
Create better opportunity to develop multi-band indices.
Typically, 2 to 8 bands provide best information in terms
of R-square values, beyond which addition of bands does
not increase the R-square and the relationship becomes
asymptotic (Thenkabail et al., 2004a; Thenkabail et al.,
2004b; Mariotto et al., 2013; Marshall and Thenkabail,
2014).
704
August 2014
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
