Automated Hyperspectral Vegetation Index
Retrieval from Multiple Correlation Matrices
with HyperCor
Helge Aasen, Martin Leon Gnyp, Yuxin Miao, and Georg Bareth
Abstract
Hyperspectral vegetation indices have shown high potential
for characterizing, classifying, monitoring, and modeling of
vegetation and agricultural crops. Correlation matrices from
hyperspectral vegetation indices and plant growth parame-
ters help select important wavelength domains and identify
redundant bands.
We introduce the software HyperCor for automated pre-
processing of narrowband hyperspectral field data and
computation of correlation matrices. In addition, we propose
a multi-correlation matrix strategy which combines multiple
correlation matrices from different datasets and uses more
information from each matrix.
We apply this method to a large multi-temporal spectral
library to derive vegetation indices and related regression
models for rice biomass detection in the tillering, stem elonga-
tion, heading and across all growth stages. The models are
calibrated with data from three consecutive years and validat-
ed with two other years. The results reveal that the multi-cor-
relation matrix strategy can improve the model performance
by 10 to 62 percent, depending on the growth stage.
Introduction
Plants have a decisive role in the ecosystems of the Earth as
they cover more than two-thirds of the land surface and are
able to produce organic compounds through photosynthesis
(Jensen, 2007). Due to the increasing demand for energy and
food with limited or even diminishing agricultural areas, the
productivity of existing arable land must be enhanced (Then-
kabail
et al
., 2012). This may be accomplished by precision
agriculture by implementing more efficient management
methods supported by remote or proximal sensing (Moran
et
al
., 1997; Viacheslav
et al
., 2012; Mariotto
et al
., 2013).
Hyperspectral sensors monitor agricultural crops by col-
lecting continuous narrow band spectral reflectance. With the
resulting hyperspectral narrow bands (
HNBs
), dynamic changes
in the biophysiology and biochemical compounds of plants
can be detected (Gitelson, 2011; Thenkabail
et al
., 2011; Zhu
et al
., 2011). Additionally, hyperspectral close range sensors
such as field spectrometers provide higher spectral resolution
than airborne or spaceborne systems. Thus, they represent a
valuable data source for calibration and simulation of potential
upcoming sensors as well as for precision agriculture applica-
tions (Milton
et al
., 2009; Roberts
et al
., 2011; Mulla, 2013).
However, several challenges occur from the huge amount
of data collected by hyperspectral sensors. These include
data handling issues caused by the dataset size, redundancy
problems from the multicollinearity of bands, and the curse of
high dimensionality, describing classification and processing
issues with increasing number of spectral bands (Bajwa and
Kulkarni, 2011). Additionally, appropriate analysis techniques
for the desired purpose have to be selected and applied to
the data: techniques such as library matching and continuum
analysis provide the opportunity to use the information of the
(whole) spectrums shape (Mutanga
et al
., 2004; Bajwa
et al
.,
2011). On the other hand, a large number of bands contained
in a spectrum are redundant and a few selected
HNBs
focus-
ing on important parts of the spectrum already reach high
classification accuracies (Stroppiana
et al
., 2011; Ma
et al.
2013; Thenkabail
et al
., 2013). Careful selection of the latter is
therefore suited to overcome the curse of dimensionality.
Optimized
HNBs
can be used to characterize, differentiate,
and model vegetation (Galvao
et al
., 2011; Thenkabail et al.,
2011). For agricultural crop monitoring approaches, research-
ers have investigated appropriate hyperspectral vegetation in-
dices (
HVIs
) for biophysical parameters such as biomass (e.g.,
Hansen and Schjoerring, 2003; Koppe
et al
., 2010; Gnyp
et al
.,
2013) and
LAI
(e.g., Shibayma and Akiyama, 1989; Habou-
dane
et al
., 2004), as well as biochemical parameters such as
nitrogen (e.g., Hansen and Schjoerring, 2003; Stroppiana
et
al
., 2009; Li
et al
., 2010; Yu
et al
., 2013), chlorophyll (e.g.,
Gitelson and Merzlyak, 1996; Miao
et al
., 2009), and plant
stress (e.g., Mahlein
et al
., 2013). Nevertheless, there is still a
demand for systematic analysis of the
HVIs
. Huete
et al
. (1994)
and Running
et al
. (1994) defined the requirements for vege-
tation indices (
VIs
) which are also valid for
HVIs
. Thus, further
Helge Aasen is with the Institute of Geography, University of
Cologne, Albertus, Magnus-Platz, 50923 Cologne, Germany
(
).
Martin Leon Gnyp is with the Institute of Geography, Uni-
versity of Cologne, Albertus, Magnus-Platz, 50923 Cologne,
Germany, the International Center for Agro-Informatics and
Sustainable Development, and Research Centre Hanninghof,
Yara International, 48249 Duelmen, Germany.
Yuxin Miao is with the College of Resources and Environmen-
tal Science, China Agricultural University, 100193 Beijing,
China, and the International Center for Agro-Informatics and
Sustainable Development.
George Bareth is with the Institute of Geography, University
of Cologne, Albertus, Magnus-Platz, 50923 Cologne, Germany,
and the International Center for Agro-Informatics and Sus-
tainable Development.
Photogrammetric Engineering & Remote Sensing
Vol. 80, No. 8, August 2014, pp. 785–795.
0099-1112/14/8008–785
© 2014 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.80.8.785
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
August 2014
785
