Performance Analysis of Advanced Decision Forest
Algorithms in Hyperspectral Image Classification
Ismail Colkesen and Omer Habib Ertekin
Abstract
In this study, the performances of random forest (
RF
), rotation
forest (RoF), and canonical correlation forest (
CCF
) algorithms
were compared and analyzed for classification of hyperspec-
tral imagery. For this purpose, the Airborne Visible/Infrared
Imaging Spectrometer (
AVIRIS
) Indian Pine (
IP
), the Reflective
Optics System Imaging Spectrometer University of Pavia, and
the
AVIRIS
Kennedy Space Center (
KSC
) data sets were used as
main data sources. In addition to the confusion matrix–de-
rived accuracy measures (overall accuracy, kappa coefficient,
F-scores), the performances of the algorithms were analyzed
in detail considering three diversity measures (Q statistics,
correlations, and interrater agreements) and a kappa-error
diagram. Results showed that the highest classification ac-
curacies (87% for
IP
, 94% for
PU
, and 93% for
KSC
data sets)
were achieved with the use of
CCF
algorithm, and improve-
ments in classification accuracy were statistically significant
compared to RF and RoF. Based on the diversity measures
and the kappa-error diagram, individual learners in the
CCF
ensemble were found to be more diverse and accurate.
Introduction
Having accurate, reliable, and timely information about
spatial distributions of the natural and artificial objects on
the Earth’s surface plays a major role in many global and
local scale applications. Remote sensing technologies have
been used as an important tool for gathering such valuable
information, and their primary products, satellite imageries,
have been considered as a main data source in many studies.
Producing thematic maps representing the land cover and
land use (
LULC
) types of the Earth’s surfa
classification is a widely used technique
ingful information from the remotely se
reliability of thematic maps derived by image classification at
the point of representation of the natural and artificial surface
features is critical for the success of many studies requiring
LULC
information. Thematic map accuracy is strongly corre-
lated to the success of the classification process, and it varies
depending mainly on the
LULC
types in the study area, the
characteristics of the used images, the scale of the study, and
the method employed to perform the classification task (Lu
and Weng 2007). To date, many classification techniques and
methods have been developed and used in applications in
order to perform the classification task (Tso and Mather 2009;
Li
et al.
2014).
Innovations in remote sensing technologies and sensor sys-
tems have allowed satellite images to develop, especially in
terms of spatial and spectral resolution. Hyperspectral images
consisting of hundreds of narrow spectral bands are one of the
main products of the remote sensing technologies that make it
possible to gather a high level of spectral information from the
particular area of the Earth’s surface. Although the supervised
classification technique is one of the most common tools for
the analysis of hyperspectral imagery and producing of
LULC
map, the classification of hyperspectral imagery is still a chal-
lenging issue due mainly to the curse of dimensionality and
the limited number of training samples (Ghamisi
et al.
2014).
The number of training samples required to accurately deter-
mine class boundaries is generally considered to be a function
of the number of spectral bands (Foody
et al.
2006; Thenka-
bail
et al.
2014). Also, higher spectral resolution allows dis-
criminating different materials, resulting in a larger number
of classes to be classified. However, it is not always possible
to collect sufficient number of samples for each
LULC
class
to find effective class boundaries using traditional paramet-
ric classification algorithm (e.g., naive Bayes and maximum
likelihood). In order to overcome these problems and to a
conduct supervised classification task, the use of nonparamet-
ric classifiers or machine learning algorithms (e.g., support
vector machines and neural networks) have recently become
a vibrant research topic in the remote sensing area (Kavzoglu
and Colkesen 2009; Ghamisi
et al.
2016; Maxwell
et al.
2018).
In recent years, there has been renewed interest in the use
of ensemble learning algorithms for the classification of satel-
lite imagery (Dietterich 2000; Rokach 2010; Jurek
et al.
2014;
Gislason
et al.
2006; Kavzoglu and Colkesen 2013; Colkesen
and Kavzoglu 2017). The main idea behind the ensemble
learning is to construct a set of multiple classifiers and then
make a classification decision about the test samples by aggre-
gating their individual predictions (Kuncheva 2014). Within
the ensemble learning frameworks, decision tree–based
multiple classifier systems, such as random forest (
RF
) and
rotation forest (
RoF
), have received increased attention in hy-
perspectral image classification studies due to their ability to
n performance of a weak classifier (i.e.,
as handling high-dimensional data clas-
or example, Chan and Paelinckx (2008)
evaluated the classification performance of an
RF
and decision
tree–based AdaBoost algorithm using Airborne HyMap hyper-
spectral image having 126 spectral bands distributed between
0.4 and 2.5 µm. Performances of the algorithms were also
compared with an artificial neural network (
ANN
) classifier.
The results of the study indicated that tree-based ensemble
models showed similar classification performances, but both
outperformed the
ANN
classifier in terms of overall accuracy
(
OA
). Xia
et al.
(2014) conducted a comparative study on the
classification performances of decision tree–based bagging,
AdaBoost,
RF
,
RoF
ensemble models, and support vector ma-
chines (
SVM
) using three well-known hyperspectral data sets.
Results showed that the decision tree–based
RoF
algorithm
using principal component analysis produced higher clas-
sification accuracy compared with other methods in terms of
overall accuracies. More recently, a novel ensemble learning
method, called canonical correlation forest (
CCF
), based on
Ismail Colkesen and Omer Habib Ertekin are with Department
of Geomatics Engineering, Gebze Technical University, Gebze-
Kocaeli, 41400, Turkey (
)
Photogrammetric Engineering & Remote Sensing
Vol. 86, No. 9, September 2020, pp. 571–580.
0099-1112/20/571–580
© 2020 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.86.9.571
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
September 2020
571
