PE&RS August 2014 - page 759

evenly distributed across crop type and phenological stages,
as the same fields were visited during each trip. Spot samples
of aboveground fresh biomass were measured with a spring
scale (precision = 0.1 g). A square-meter quadrant was con-
structed and used to define the area over which aboveground
fresh biomass spot samples were extrapolated, yielding units
of gm-2. Approximately 2 to 4 spot samples were collected for
each of the more than 520 frames used in the analysis. The
frames (Plate 1) were distributed across the Central Valley’s
diverse soil types and climatology. Spot samples of abo-
veground fresh biomass were extrapolated by multiplying the
average mass of the spot samples by the number of plants in
the corresponding square meter area. The number of plants
in each frame was counted manually from photos cropped to
the sample frame and taken during the sprouting phase using
a red-green-blue camera. Each photo was collected at a fixed
height (1.5 m for alfalfa, cotton, and rice, and 2.5 m for maize)
at nadir during ±2 hours of solar noon on near cloudless days
to minimize solar illumination effects caused by sun angle
and shadowing. Height was measured for alfalfa, cotton,
and rice using a meter stick and telescoping measuring rod
for maize. The biomass of alfalfa was measured differently,
because it is a perennial and grows in clusters. Spot measure-
ments of the ground length and width of clusters were taken,
averaged, and used to calculate the area of a cluster, assuming
the clusters form ellipses. From the photos, the vegetation
fraction was estimated and multiplied by a square meter and
then divided by the area of a cluster to estimate the number
of alfalfa plants, which was then multiplied by the average
mass of spot samples taken. This procedure was deemed
appropriate to avoid the unnecessary destruction of crops in
each field. The mean and standard deviation of the biomass
square-meter samples taken across crop type and phenologi-
cal stage are shown in Table 1.
Table 1. Summary Statistics of Aboveground B iomass Samples T aken per Crop
T ype and Visit (1 = sprouting, 2 = flowering/silking, and 3 = grain/bud- fill-
ing) for cotton, maize, and rice . Because Alfalfa is Harvested T hroughout
the Growing Season, the Visit does not Necessarily Match the Phenological
Stage . N is the Number of Samples and σ is the Standard Deviation.
Crop Visit N Mean (gm-2) σ (gm-2)
Alfalfa 1 54 7694.46 11117.34
2 57 3187.17 3241.97
3 60 14647.11 13715.95
Cotton 1 58 800.95 759.89
2 59 7908.44 3590.35
3 57 9713.79 5924.20
Maize 1 54 8148.00 3646.59
2 60 13824.53 2973.18
3 59 12186.12 2942.21
Rice 1 44 800.39 487.94
2 50 2703.15 1089.45
3 47 2574.64 1006.03
Field spectra were collected for each sample frame with
an Analytical Spectral Devices (ASD: www.asdi.com)portable
spectroradiometer (Field Spec Pro 3), which has an optical
range of 350 to 2500 nm resampled to 1nm resolution. A
pistol grip and 18° field of view (FOV) fore-optic were attached
to the fiber optic cable bundle, which records light from the
crop canopy. The 18° FOV fore-optic was selected, because
it shows greater spatial and spectral uniformity over the area
collected, compared to smaller fore-optics employed with the
ASD (MacArthur et al., 2012). As with the photos, each series
of spectra were collected at a fixed height (1.5 m for alfalfa,
cotton, and rice, and 2.5 m for maize) above the surface at na-
dir between ±2 hours of solar noon on near cloudless days to
minimize solar illumination effects caused by sun angle and
shadowing. The Field Spec Pro 3 records raw radiance values,
which were collected 30 times or more under sub-optimal
field conditions (e.g., high winds) and averaged internally.
Raw radiance (Wsr-1m-2) was converted internally to percent
reflectance using calibration spectra collected every 2 to 10
minutes depending on field conditions and near each sample
frame with a white standard panel composed of BaSO4. Over
the course of each visit, current generated by the detector (i.e.,
“dark current”) or passing the fore-optic from dark objects
to bright objects or vice versa typically saturated the signal,
which necessitated detector re-optimization. The internally
processed reflectance spectra were collected at five different
locations within each square-meter frame to reduce uncertain-
ties in spectroradiometric FOV and mixed effects from crop,
soil, and shadow. Five reflectance spectra collected randomly
within each frame where biomass was estimated, yielded
nearly 3,600 recorded spectra over the two-year period.
Data Processing
The spectra underwent additional preprocessing steps, before
statistical models relating biomass to spectral reflectance were
developed. The ASD Field Spec Pro 3 fiber optic cable collects
light and diverts it to three detectors: 350 to 1050, 900 to 1850
nm, and 1700 to 2500 nm. The spectroradiometer automat-
ically rectifies the overlap between each sensor, but slight
differences across detectors were observed, so inter-sensor
normalization was performed by multiplying visible/NIR and
SWIR2 detector reflectance by near-edge SWIR1 detector ratios.
This had a minimal impact on the spectra and was used
primarily for visualization perhaps. The standard deviation
of the five sample spectra was computed in the SWIR2, which
was typically the most spurious, and spectra with values
greater or less than one standard deviation were omitted,
before final averaging. Once the spectra were averaged, so that
one spectrum represented one sample, wavelengths where
strong greenhouse gas absorption occurs, were omitted. These
included 350 to 390nm (O3), 1350 to 1450 nm (H2O and CO 2),
1790 to 2000 nm (H2O and CO2), and 2300 to 2500 nm (H 2O
and CO2). In order to reduce the number of wavelengths
analyzed, the 1 nm spectra collected by the spectroradiometer
were averaged to 10 nm HNBs considering the full width at
half maximum, but matching the 400 to 2500 nm 10 nm HNBs
detected by the Hyperion sensor. In order to reduce potential
data redundancies and make computations more efficient,
previous studies have used similar 10 nm HNBs to build accu-
rate empirical hyperspectral biophysical models (Thenkabail
et al., 2004; Thenkabail et al., 2002). In addition, in a process
known as the Hughes effect (Hughes, 1968), biophysical sim-
ulation accuracy may decrease with the initial introduction
of more predictors during model-building. The aggregations
yielded 196 10 nm matching Hyperion bands, as some of the
242 Hyperion bands were erroneous and were not included in
the analysis. For the remainder of this paper, the 196 HNBs are
expressed using the 10 nm wavelength centroids.
Spectra were inherently mixed, because leaf litter, soil, and
other background features interfere with the vegetation signal.
In remote sensing, methods such as band ratioing reduce the
effect of background features, but assume that these features
vary consistently across samples and wavelengths (Hall et
al., 1990). First (Demetriades-Shah et al., 1990), second (Hall
et al., 1990), and both first and second (Elvidge and Chen,
1995) derivatives were used to transform spectra in order to
reduce the effects of soil background, as the rate of change of
the background signal tends to change more gradually than
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING August 2014 759