IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 14, NO. 11, NOVEMBER 2017 2167
Weight-Based Rotation Forest for Hyperspectral
Image C lassification
Wei Feng and Wenxing Bao
Abstract— In this letter, we propose a new weight-based
rotation forest (WRoF) induction algorithm for the classification
of hyperspectral image. The main idea of the new method is to
guide the growth of trees adaptively via exploring the potential
of important instances. The importance of a training instance is
reflected by a dynamic weight function. The higher the weight
of an instance, the more the next tree will have to focus on the
instance. Experimental results on two real hyperspectral data sets
show that the WRoF algorithm results in significant classification
improvement compared with random forests and rotation forest.
Index Terms— Classification, hyperspectral image, random
forests (RFs), rotation forest (RoF), weight.
I. INTRODUCTION
C
LASSIFICATION is one of the major tasks in remote
sensing information processing. Classification of hyper-
spectral data is usually more difficult than other remote sensing
imagery due to issues, such as the high ratio of feature to
instance and the redundant information in the feature set [1].
While most learning systems suffer from the intractability
issue known as the curse of dimensionality, studies have
demonstrated the successful application of classifier ensemble
techniques to hyperspectral classification [2]–[6].
Ensemble learning, also calle d committee-based learning,
is an effective method to develop accurate classification sys-
tems [7]. It is appealing, because it is able to boost weak
learners, which are slightly better than random guess to strong
aggregated learners, which can make very accurate predictions.
Boosting [8] and bagging (the acronym of bootstrap aggre-
gating) [7] are major ensemble learning methods. Diversity,
which is the difference among the individual learners, has
been recognized as a very important characteristic in classifier
combination [9]. It can be used effectively to reduce the
variance error without increasing the bias error by ensemble
methods [10]. In o rder to encourage diversity within bagging,
random forests (RFs) [11] are proposed. The RFs is a combi-
nation of tree predictors in which the decision trees [12] are
constructed using a resampling technique with replacement;
they randomly sample the attributes and choose the best split
Manuscript receive d April 26, 2017; revised August 20, 2017; accepted
September 11, 2017. Date of publication October 13, 2017; date of current
version October 25, 2017. This work was supported by the National Natural
Science Foundation of China under Grant 61461003. (Corresponding author:
Wei Feng.)
W. Feng is with the Geo-Resources and Environment Laboratory,
Bordeaux INP, 33600 Pessac, France (e-mail: wei.feng@ipb.fr).
W. Bao is with School of Computer Science and Engineering, North Minzu
Uni versity, Yinchuan 750021, China (e-mail: bwx71@163.com).
Color versions of one or more of the figures in this letter are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/LGRS.2017.2757043
among those variables rather than the best split among all
attributes. Important advantages, such as running efficiently on
large data bases, handling thousands of input variables without
variable deletion and low time cost make RFs widely attract
the interest of researchers [13], [14].
The rotation forest (RoF) [20] method draws upon the idea
of RFs, but aims at building more accurate and diversified b ase
classifiers. It splits randomly the feature space into several
subspaces, applies principal component analysis (PCA) [15]
separately on each subspace, and repeats the aforementioned
process to generate the diversified training data sets and
base classifiers for different feature subspaces. Studies have
demonstrated the successful application of RoF to remote
sensing imagery [2], [3], [16]–[18]. Moreover, the RoF was
found to provide more satisfactory result with respect to
bagging, AdaBoost, and RFs ensembles in the classification
of hyperspectral data [2].
Recently, several approaches have proposed to improve the
performance of RoF [6], [17], [18]. Lu et al. [6] increase the
classification accuracy of RoF by improving the effectiveness
of base classifiers. A cost-sensitive decision tree is recom-
mended to replace the standard decision tree [12] as a base
classifier in their experiment. Li et al. [17] construct the RoF
with improved performance by using an ensemble AdaBoost
instead of a single classifier as a basic classifier. Xia et al. [18]
proposed the high-performance RoF via building decision trees
for each subfeature set. However, these methods treat all
the instances equally, a nd the potentials of the infor mative
instances do not be taken into account. Furthermore, these
algorithms generate base classifiers independently of one
another, and some of these base classifiers not only increase
the computation complexity of the algorithm but also decrease
the ensemble performance.
The major contribution of this letter is to propose a novel
weight-based RoF (WRoF) algorithm. The main idea of the
new method is to guide the growth of trees adaptively via
exploring the potential of important instances. The impor-
tance of a training instance is reflected by a dynamic weight
function. The remainder of this letter is o rganized as follows.
In Section II, we introduce the dynamic weight function and
the proposed method WRoF. The experimental studies are
presented in Section III. Section IV presents the conclusions
and the fu ture work of this letter.
II. M
ETHODS
Our WRoF algorithm is inspired by boosting. However,
boosting updates the weights of training samples at each
1545-598X © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.