Biomass Modeling of Four Leading World Crops
Using Hyperspectral Narrowbands in Support of
HyspIRI Mission
Michael Marshall and Prasad Thenkabail
Abstract
New satellite missions are expected to record high spectral
resolution information globally and consistently for the first
time, so it is important to identify modeling techniques that
take advantage of these new data. In this paper, we estimate
biomass for four major crops using ground-based hyper-
spectral narrowbands. The spectra and their derivatives
are evaluated using three modeling techniques: two-band
hyperspectral vegetation indices (
HVIs
), multiple band-
HVIs
(
MB-HVIs
) developed from Sequential Search Methods (
SSM
),
and
MB-HVIs
developed from Principal Component Regres-
sion. Overall, the two-band
HVIs
and
MB-HVIs
developed from
SSM
s using first derivative transformed spectra in the visible
blue and green and
NIR
explained more biomass variability
and had lower error than the other approaches or trans-
formations; however a better search criterion needs to be
developed in order to reflect the true ability of the two-band
HVI
approach. Short-Wave Infrared 1 (1000 to 1700 nm)
proved less effective, but still important in the final models.
Introduction
California leads the United States (US) in agriculture, receiv-
ing more than 40 billion dollars in revenue or approximately
12 percent of
US
agriculture revenues (CDFA, 2013). Seven of
the top ten agricultural producing counties (Fresno, Tulare,
Kern, Merced, Stanislaus, San Joaquin, and King) are in the
Central Valley of California. More than 60 percent of crops by
area are irrigated, which accounts for 75 to 80 percent of the
state’s annual water budget (USDA, 2009). Growing and com-
peting urban and domestic use, declines in non-local sources
(e.g., Colorado River), and environmental legislation in the
Sacramento-San Joaquin Delta have put considerable strain
on agricultural water resources (Faunt, 2009). Alfalfa, the top
water user by crop type, accounts for nearly a quarter of the
state’s irrigation water, while yielding only 4 percent in crop
revenues. The other large water users include maize, rice, and
cotton, and generate an additional 10 percent total in crop
revenues. Other factors notwithstanding, the disparity be-
tween crop water use and yield affords an opportunity for tre-
mendous water savings that incurs a small cost to the state’s
economy (Umbach, 1997). Increasing crop water productivity
(
WP
), defined here as the ratio of actual marketable crop yield
to actual seasonal crop water consumption or evapotranspira-
tion (Zwart and Bastiaanssen, 2004), involves integrated wa-
ter management strategies that focus on irrigation techniques
and scheduling, soil amendments, seed preparation, tilling
practices, water harvesting, crop species, and variety selec-
tion (Ali and Talukder, 2008), as well as economic incentives,
such as water markets and crop subsidies (Sunding, 2000).
Remote sensing-based models of
WP
facilitate the design, im-
plementation, and assessment of these strategies at low cost
and over large areas (Bastiaanssen
et al
., 2000).
Pinter
et al
. (2003) reviews remote sensing approaches in
agriculture, focusing on broad-band, hyperspectral narrow
bands (
HNB
s), and thermal band based vegetation indices and
their application to crop yield estimation and the manage-
ment of water, nutrients, and pests. Actual marketable crop
yield can be estimated empirically by remote sensing-based
indices that are sensitive to several crop biophysical and
biochemical properties. Biomass production is the net gain in
carbon and energy through assimilation, which is a measure
of physiological efficiency and is therefore a common deter-
minant of empirically-based yield models. The harvest index,
for example, is a crop-specific coefficient developed from
field studies that reduces crop biomass to yield (Hay, 1995).
Given the sensitivity of the index to crop variety, management
practices, and external environmental factors, physiological-
ly-based crop growth models where biomass is an integral
model component have been developed (Doraiswamy
et al
.,
2003).
Field measurements of crop biomass are time-consuming,
destructive, and difficult to extrapolate over large areas, so
remote sensing biomass techniques have been developed to
overcome these limitations. Early attempts used the Normal-
ized Difference Vegetation Index (
NDVI
) to estimate biomass,
as the index is sensitive to plant “greenness,” however, results
were less accurate and consistent compared to
NDVI
estimates
of primary productivity (Box
et al
., 1989). Lu (2006) pro-
vides a review of empirical and semi-process based biomass
modeling techniques using broad-band multi-spectral remote
sensors, radar, and lidar, with an emphasis on forest land
cover. Empirical models of biomass are typically measured
Michael Marshall is with Climate Change, Agriculture,
and Food Security, World Agroforestry Centre, P.O. Box
30677-00100, Nairobi, Kenya, and the Southwestern Geo-
graphic Center, United States Geological Survey (USGS),
2255 Gemini Dr, Flagstaff, AZ (
.
Prasad Thenkabail is with the Southwestern Geographic Cen-
ter, United States Geological Survey (USGS), 2255 Gemini Dr,
Flagstaff, AZ.
Photogrammetric Engineering & Remote Sensing
Vol. 80, No. 8, August 2014, pp. 757–772.
0099-1112/14/8008–757
© 2014 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.80.8.757
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
August 2014
757
