over specific regions using linear and nonlinear regression models, neural networks, K-nearest-neighbor, texture analysis, or
vegetation indices like
NDVI
. Semi-process-based approaches infer biomass from canopy parameters estimated by inverting veg-
etation canopy models. Plants absorb strongly in the visible, most notably the visible red, due to photosynthetic and accessory
pigments in the plant tissue. Plants reflect strongly in the near infrared (
NIR
) due to little absorbance of plant material in this
region and scatter by plant cell walls. Whether empirical or semi-process-based, biomass models are typically parameterized by
wavelengths around the transition between red absorption and
NIR
reflectance, termed the “red-edge” (Horler
et al
., 1983). The
major limitations of these biomass estimation techniques in general are red-edge bands may not account for important physical
and physiological processes related to biomass (Baret
et al
., 1989), varying soil conditions, or crop-specific traits (Haboudane
et al
., 2004). Hyperspectral remote sensing analysis, which commonly involves hundreds or thousands of
HNB
s in the optical
range (400 to 2500 nm), can reveal spectral regions that account for crop-specific physiological traits and minimize soil back-
ground effects.
Ustin
et al
. (2004) gives a brief review of hyperspectral analysis using current airborne (e.g.,
AVIRIS
) and spaceborne (e.g.,
Hyperion on board the
EO-
1 satellite) remote sensing sensors. The review shows that hyperspectral image analysis has been
used to measure several biochemical plant properties related to crop biomass: Leaf Area Index (
LAI
), fraction of photosynthet-
ically active radiation, moisture status, and stress. Carotenoids, anthocyanin, and chlorophyll increase the range over which
light energy can be absorbed for photosynthesis, as well as increase light-use efficiency and provide protection from tempera-
ture extremes and ultraviolet radiation (Ustin
et al
., 2009). These
pigments are sensitive to light from 510 to 520, 540 to 560, and
700 to 730 nm, respectively (Gitelson
et al
., 2006). Chlorophyll
concentration, for example, decreases as crops reach senescence,
which causes a blue shift in the red-edge (680 to 730 nm). Crop
stress related to nutrient (nitrogen) deficiencies is also sensitive
to the light characteristics in this region (Perry and Roberts,
2008). Leaf water absorbs strongly in select bands across the
NIR
(700 to 1000 nm), Short-Wave Infrared 1 (
SWIR
1: 1000 to 1700
nm), and Short-Wave Infrared 2 (
SWIR
2: 1700 to 2500 nm). Crop
water content, which is related to biomass, is particularly sensi-
tive to light at 970 and 1180 nm (Champagne
et al
., 2003), while
water stress, a physiological constraint, is better estimated using
metrics sensitive to pigment concentration (Perry and Roberts,
2008). Structural carbon (lignin-cellulose) contained primarily in
dry plant residues, absorbs strongly in the
SWIR
2 (Asner, 1998).
Lignin-cellulose bands are therefore used primarily to estimate
dry plant matter biomass. Given the number of wavelengths
involved in hyperspectral analysis, several data reduction and
mining techniques have been explored to develop empirical
NB
biophysical crop models (Thenkabail
et al
., 2004; Thenkabail
et
al
., 2002). These techniques are empirical and therefore may not
be transferable due to a small sample size and limited research
area, yet produce better estimates compared to crop biophysical
estimates derived from model inversion, an approach that will be
more comparable as the physical detail of process-based models
improve (Casa and Jones, 2004).
In this study, we use three common spectral transformations
and empirically-based biophysical modeling techniques to iden-
tify important
HNB
s sensitive to aboveground fresh crop biomass
(gm
-2
). The study uses a large two-year dataset of ground-based
spectroradiometric and aboveground fresh crop biomass data,
which is derided into independent subsets for model calibration
and validation. The dataset spans the Central Valley of California
and includes measurements from the four largest water users
(alfalfa, cotton, maize, and rice).
Methods
Study Site and Data Collection
A large field campaign in the Central Valley was conducted
during the warm crop season (boreal summer) in 2011 and 2012,
in which crop biophysical and spectroradiometric data were
collected to develop a remote sensing-based crop water produc-
tivity model for the entire state. The Central Valley stretches over
700 km and covers an area of over 100,000 km
2
between the Coast
range to the west and Cascade, Sierra Nevada, and Tehachapi
ranges to the east (Plate 1).
Above-ground fresh biomass spot samples and spectra were
collected for California’s four largest water users (alfalfa, maize,
cotton, and rice) during important phenological stages (sprout-
ing, flowering/silking, and grain/bud-filling). The samples were
Plate 1. The extent of major water users (alfalfa, cotton,
maize, and rice) in California defined by the National Agricul-
tural Statistics Service Cropland Data Layer (
data.gmu.edu/
) in 2012 overlaid with the five hundred and
twenty-seven field spectra and above-ground wet biomass
sample pairs in black taken during major phenological
phases over two growing seasons (2011 and 2012) in the
Central Valley of California.
N
= 44, 65, 84, and 85 for rice,
alfalfa, cotton, and maize, respectively.
758
August 2014
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
