other statistical techniques, but are less flexible than two-
band
HVI
s. Given the success of these techniques compared
to previous ground-based
HNB
studies, further investigation
is required with space-borne hyperspectral sensors to prove
their validity across scales.
The single band analysis of the untransformed spectra
shows that biomass is negatively (positively) correlated to
a large portion of the visible (
NIR
) spectrum, reflecting the
preferential absorption (scatter) of plant chlorophyll as
crops mature through the growing season. Correlations vary
across each crop type, perhaps reflecting their unique canopy
architecture and physiology. And this undoubtedly impacts
the significance of various bands when transformed and/or
analyzed using two-band
HVI
s,
SSMs
, or
PCA
. Rice biomass,
for example, correlated with
HNB
s over a much narrow range
than alfalfa and cotton. Rice has an erectophile (vertical) leaf
area distribution (
LAD
), especially during midday and un-
der well-watered (irrigated) conditions (Moran
et al
., 1989)
when spectra were collected, while cotton and alfalfa have a
planophile (horizontal)
LAD
. Light is attenuated much more
readily in erectophile canopies, while planophile crops, on
the other hand, show greater absorption (scatter) in the visible
(
NIR
) bands, because their leaf area has greater exposure to
light during midday under well-watered conditions. Alfalfa
and cotton also showed similarities in
HNB
selection during
the model-building process, highlighting the potential of
aggregating these crops and those with similar structure in
future biomass models. The positive correlation between
SWIR
bands and rice biomass is surprising, as leaf water content
and biomass increase as plants mature before senescence, typ-
ically leading to greater absorption in the
SWIR
. The presence
of water in the rice fields could offer an explanation. The rice
under study is submerged in 20 cm of water on average when
the spectra are collected. When rice is sprouting, water is
plainly visible and therefore
SWIR
absorption by the spect-
roradiometer is high. As the plant matures and its density
increases, however, the canopy obscures the signal, leading
to relatively less
SWIR
absorption. This hypothesis could be
explored further, as in Gnyp
et al
. (2014), by performing the
analysis at each major development stage and potentially re-
veal other important
SWIR
bands that predict growth for each
successive stage.
The study was purposely performed over various soil
types, irrigation methods, and at different times of the water-
ing regime, as these can influence the performance of visible,
NIR
, and
SWIR
spectra in measuring crop biomass. The higher
correlations particularly in the
NIR
and
SWIR
, after soil and wa-
ter effects were minimized using the first derivative transfor-
mation, reveal the impact background can have on estimating
biomass at longer wavelengths. This is reiterated by the higher
positive (
NIR
) correlations, particularly around strong absorp-
tion bands, after the transformation for crops with relatively
less dense canopies and more exposure to soil (alfalfa and cot-
ton) compared to rice. The second derivative transformation
further reduces solar illumination effects (Tsai and Philpot,
1998), particularly for rice in the early growth stage when
the background consists mainly of water (Gnyp
et al
., 2014).
Although the second derivative transformed spectra perform
well in the additive models, under validation, the correlations
are lower, perhaps reflecting the sensitivity of the transforma-
tion to over-fitting or poor performance of this transformation
T
able
2. T
he
T
op
T
en
T
wo
-
band
hvi
s
for
the
C
alibration
S
ubset
,
R
anked
U
sing
R
2
; R1
and
R2
correspond
to
E
quation
1
Crop
R1
R2
R
2
Alfalfa
682
458 0.81
1256
438 0.81
1205
438 0.81
1629
438 0.81
993
438 0.81
1619
438 0.80
993
428 0.80
1629
428 0.80
1619
428 0.80
Cotton 794
458 0.86
763
438 0.86
763
428 0.86
794
489 0.84
1023
438 0.84
1013
438 0.84
1034
438 0.84
1044
438 0.84
1054
448 0.84
Maize
1205
773 0.46
1205
794 0.45
1215
773 0.43
1225
773 0.42
1649
773 0.42
1629
773 0.41
1639
773 0.41
1619
773 0.40
1235
773 0.40
Rice
1225
529 0.82
1225
519 0.81
1225
509 0.81
1225
539 0.80
845
539 0.79
1498 1225 0.78
1215
529 0.78
1235
529 0.78
1235
519 0.78
T
able
3. R
esults
of
the
F
orward
A
ddition
ssm
R
esults
for
U
ntransformed
,
F
irst
D
erivative
(1),
and
S
econd
D
erivative
(2) R
eflectance
and
E
xplained
V
ariance
of
B
iomass
from
the
C
alibration
S
ubset
for
E
ach
C
rop
T
ype
Crop Type Transformations Band Centers (nm)
R
2
Alfalfa
Untransformed 672; 733; 1447
0.74
(N=65)
Derivative(1)
560; 1447; 1770
0.77
Derivative(2)
448; 763; 1750
0.82
Cotton Untransformed 672; 963; 1114
0.87
(N=84)
Derivative(1)
438; 672; 1145
0.87
Derivative(2)
478; 661; 1155
0.83
Maize
Untransformed 763; 845; 1023; 1094
0.52
(N=85)
Derivative(1)
672; 773; 783
0.58
Derivative(2)
702; 1134; 1155; 2042 0.51
Rice
Untransformed 428;1094
0.78
(N=44)
Derivative(1)
1336
0.82
Derivative(2)
1104;1155
0.85
766
August 2014
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
