A Multiclassifier and Decision Fusion System for
Hyperspectral Image Classification
Zhen Ye
School of Electronics and
Information, Northwestern
Polytechnical University
Xi'an, China
yzh525@gmail.com
Mingyi He
School of Electronics and
Information, Northwestern
Polytechnical University
Xi'an, China
myhe@nwpu.edu.cn
Saurabh Prasad
Electrical and Computer
Engineering Department,
University of Houston
Houston, TX, USA
saurabh.prasad@ieee.org
James E. Fowler
Geosystems Research
Institute, Mississippi State
University
Starkville, MS, USA
fowler@ece.msstate.edu
Abstract—In this paper, a windowed three-dimensional
discrete wavelet transform (3DDWT) is employed to extract
spectral-spatial features for hyperspectral image classification;
these features quantify local orientation and scale
characteristics. Since single subband (i.e., the "LLL" subband)
is deficient in exploiting useful information, all eight subbands
(LLL, LLH, LHL, LHH, HLL, HLH, HHL, HHH) from a
single-level, dyadic 3D DWT are first fused to overcome the
small-sample-size problem. The studies reported in this paper
are conducted within the context of multi-classifiers and
decision fusion systems that are designed to handle the
high-dimensional 3D DWT feature spaces. Two decision fusion
rules—majority voting (MV) and logarithmic opinion
pool(LOGP) are employed and studied for the final
classification of hyperspectral dataset. Experimental results
show that the proposed fusion algorithms substantially
outperform traditional single-classifier methods (LDA-MLE,
LFDA-GMM, and SVM-RBF) and a single classifier algorithm
based on the windowed 3D DWT structure (3D
DWT-LFDA-GMM).
Keywords—Wavelets; hyperspectral imagery; decision fusion
I. INTRODUCTION
Airborne hyperspectral imagers such as the Airborne
Visible /Infrared Imaging Spectrometer (AVIRIS) imply the
exploration and the collection of a huge number of spectral
bands [1]. The high dimensionality of hyperspectral imagery
(HSI) is expected to provide a much detailed spectral
response for each pixel. However, compare with the large
dimension, the sample size is often limited and insufficient
for most classifiers, such as the maximum-likelihood
estimation (MLE) [2] classifier and the Gaussian mixture
model (GMM) [3] classifier. In order to solve
small-sample-size problem, dimensionality reduction, feature
extraction or feature selection is usually employed as the
pre-processing of hyperspectral image classification.
Traditional approaches that are used to reduce the
dimensionality of hyperspectral data include principal
component analysis (PCA), independent component analysis
(ICA), and Fisher's linear discriminant analysis (LDA) [2]. A
key limitation to techniques such as PCA and LDA is that
they assume that the class-conditional distributions are
Gaussian [4]. However, real-life observational data are often
not Gaussian and, in extreme cases, may be multimodal [5].
In [5], the authors combined local Fisher's discriminant
analysis (LFDA) [6] with GMM to exploit the rich statistical
structure of HSI data. The limitation of this method is that it
rely on spectral properties of the data only, whereas the
important spatial information is excluded.
In this paper, a novel classification method based upon a
windowed 3D DWT feature selection [7] [8] and decision
fusion is proposed. In this method, after a single-level 3D
DWT using Haar filters, all eight coefficients are selected as
features for hyperspectral classification. Following this, we
employed LFDA for projecting the features onto a lower
dimensional subspace optimized for classification and a bank
of GMM classifiers for obtaining local classification results.
A decision fusion strategy is then invoked to combine these
results into a single classification decision per pixel. Two
decision fusion approaches-majority voting (MV) and
logarithmic opinion pool (LOGP) [9] are studied in this
paper. The experimental results demonstrated that the
proposed fusion system outperforms a single classifier
algorithm based on the windowed 3D DWT structure
(3DDWT-LFDA-GMM) [7] under small-sample-size
situations as well as strong additive noise environments.
The remainder of this paper is organized as follows. Sec.
II describes the proposed classification system, including the
windowed 3D DWT feature extraction, LFDA dimension
reduction strategy, Gaussian mixture model, and the
proposed fusion algorithms. In Sec. III, we describe the HSI
dataset. A detailed discussion of the experimental results is
presented. Finally, Sec. IV concludes this paper.
II. P
ROPOSED SYSTEM
A. Feature Extraction by 3D DWT
In this work, a windowed 3D DWT is employed to
extract spectral-spatial features of a hyperspectral image for
classification. Specifically, we perform a sliding window
analysis, wherein a spatial window of size B
u
B pixels is
scanned across the image (see [8]). 3D DWT coefficients are
501
978-1-4673-6322-8/13/$31.00
c
2013 IEEE