et al. is invaluable. They have implemented an innovative
spectral matching technique (SMT; also see Thenkabail et
al., 2007 for concept of SMTs) approach that automatically
determines and labels crop types by matching the class
spectral signatures with the ideal or reference spectral
libraries developed using hyperspectral Hyperion data. The
Z-score distance SMT that they use accounts for the variation
of pixel spectra by measuring the distance between the
class spectra and the reference spectra in units of standard
deviation. They used this SMT method to identify and label
11 agricultural crops classified using Landsat TM data. Their
study established that the accuracies of their automated SMT
provided: A. 6 to 11 percent greater accuracies relative to
ISODATA classification followed by manual identification
of classes, and B. 12 percent greater than the spectral angle
mapper (SAM) followed by manual identification of classes.
However, the accuracies were 2 to 12% lower than the
Maximum Likelihood Classification followed by manual class
identification. Their method can be used to automatically
identify and label classes classified using multispectral or
hyperspectral data. The automated SMT by Parshakov et al.
is novel and clearly demonstrates pathway to identify and
label crop types and other land use classes automatically
saving time and removing user bias. However, there will
be complexities of applying automated SMTs over large
and complex areas. Nevertheless, by building adequate and
accurate ideal or reference spectral libraries of agricultural
crops, vegetation categories, and other land use classes as well
as by further development of automated SMTs for specific
regions of the world, automated class labeling proposed
and demonstrated by Parshakov et al. will become accurate,
unbiased, rapid, and widely implementable.
The study by Nadiminti et al. address the important issue
of high dimensionality of hyperspectral data, ways and
approaches to overcome them, and the benefit of doing so to
overcome Hughes Phenomenon (Note: Hughes phenomenon
means that when the dimensionality of data increases, the
training sample number should also increase in order to
maintain precision of classification, alternatively we need
to increase the number of training samples which can be
often resource prohibitive. Thereby, in order to process
hyperspectral data effectively, it is necessarily to reduce the
dimensionality of hyperspectral data or increase the sample
number of training data used in classification. Including
highly correlated bands (e.g., R-square >0.9) in analysis either
makes no difference to classification accuracies or, many a
times, actually leads to decrease in classification accuracies.
This is because highly correlated bands provide same,
duplicate, information whereas the training samples remain
the same in spite of increase in number of bands. Nadimiti
et al. used hyperspectral Hyperion images of three seasons
(Monsoon, winter, and summer) over tropical forests to
classify and separate three species: Teak, Bamboo, and mixed
forests. Data dimensionality reduction was explored using
Kernel Principal Component Analysis (k-PCA), Independent
Component Analysis (ICA), and Principal Component
Analysis (PCA). Their results re-enforced the recent
findings elsewhere (Thenkabail et al., 2013) that 4 to 8 %
hyperspectral narrowbands (HNBs) provide optimal results,
leaving the rest of the bands redundant. In kPCA, for example,
10 kernel principal components or HNBs, selected based
on eigenvectors (factor loadings), explained 99% variability
in data when 179 Hyperion HNBs were used in analysis.
Thereby, the study establishes the fact that, often, HNBs that
adjoin one another are redundant for a given application.
Thereby, identifying redundant bands help overcome Hughes’
Phenomenon.
Hyperspectral data is in itself a great advancement over
broadband multispectral data. This is now an established
fact (Thenkabail et al., 2011). Nevertheless, there is
considerable scope for improvement in our understanding of
agricultural crops and vegetation communities by combining
multiple sources of remote sensing data. This aspect is well
illustrated by Zhang et al. in their paper on studying wetland
vegetation communities of Florida everglades by combining
hyperspectral data with Light Detection and Ranging (LiDAR)
data. They studied 13 common everglades vegetation
communities using 224 band hyperspectral Airborne Visible/
Infrared Imaging Spectrometer (AVIRIS) data acquired at
12 m spatial resolution and a Leica ALS-50 LiDAR system
collecting small footprint multiple returns, and intensity at
1060 nm wavelength with average point density for the study
area of 1.18 pts/m
2
. They showed by fusing the hyperspectral
and LiDAR data and using the 3 machine learning algorithms
[Random Forest (RF), Support Vector Machine (SVM), and
k
-Nearest Neighbor (
k
-NN)] it is possible to increase the
overall accuracy by as much as 10% (from 76% overall
accuracy using Hyperspectral data alone to 86% when
both Hyperspectral and LiDAR are used) in classifying 13
everglade wetland vegetation communities.
Sanchez et al. combined hyperspectral data with thermal and
radar data to retrieve soil moisture from agricultural fields.
They used data from Compact Airborne Spectrographic
Imager (CASI 550) sensor and Thermal Airborne
Spectrographic Imager (TASI 600) and combined with
Airborne L-band (ARIEL-2) to retrieve soil moisture from
irrigated-, rainfed-, and fallow- farmlands that include cereals,
sunflower, vineyards, and fallow-farmlands. They showed
that hyperspectral bands and indices (HNBs and HVIs) had
significantly better correlation with observed soil moisture
when integrated with the land surface temperature (LST) and
brightness temperature (BT) rather than when they were used
alone. They specifically recommend indices derived using
hyperspectral wavebands in the red-edge and near infrared
rather than visible. Even through microwave L-band data is
widely used for soil moisture retrieval, using that data along
with hyperspectral data has significant advantages.
Awad et al. clearly establish that classification accuracies can
be substantially increased using hyperspectral narrowband
data as opposed to multispectral broadband data. They used
CHRIS PROBA hyperspectral data to classify Stone Pine
forests and compare the results with Landsat ETM+ classified
results. Using the 63 band spaceborne CHRIS PROBA
hyperspectral data they were able to establish an increased
accuracy of as much as about 30%. The producer’s, User’s,
and overall accuracies in classifying stone pine forests using
CHRIS PROBA were 90% or higher whereas using 6 non-
thermal Landsat ETM+ these accuracies were around 60%.
722
August 2014
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
