Change Detection based on Stacked
Generalization System with Segmentation
Constraint
Kun Tan, Yusha Zhang, Qian Du, Peijun Du, Xiao Jin, and Jiayi Li
Abstract
Change detection based on a multi-classifier ensemble system
can take advantage of multiple classifiers to extract change
information in remote sensing images. In this paper, an ef-
ficient heterogeneous ensemble algorithm, i.e., the stacked
generalization (
SG
) combined with image segmentation, is
proposed to construct a simple multi-classifier ensemble
system that can offer better detection accuracy with lower
computational cost. Due to the rich spatial information in
high-spatial-resolution remote sensing images, structure
texture (morphological) and statistical texture features are
extracted to construct the input data to the ensemble system
along with spectral features. In addition, constrained analy-
sis on segmented objects integrates the smaller heterogeneity
segmentation map and pixel-wise change map to generate
the final change map. The experiments were carried out on
two
ZY-3
and a QuickBird dataset. The results show that the
proposed algorithm can integrate the advantages of both
pixel-wise ensemble and object-oriented methods, and effec-
tively improve the accuracy and stability of change detection.
Introduction
Change detection techniques based on multi-temporal remote
sensing images have been widely applied to all aspects of
national production and life, such as urban development
monitoring, mine environmental change detection, and envi-
ronmental disaster monitoring (Huang
et al
., 2013; Jawak
et
al
., 2014; Malmir
et al
., 2015).
Because of rich information of objects in high-resolution
remote sensing images, change detection of high-resolution
remote sensing images has become a popular research topic
in remote sensing applications (Wen
et al
., 2015; Huang
et al
.,
2017). According to the degree of automation, change detection
algorithms for high-resolution remote sensing images can be
broadly divided into unsupervised (Lv
et al
., 2015) and super-
vised methods (Volpi
et al
., 2013; Hou
et al
., 2015). Unsuper-
vised change detection obtains change information without ad-
ditional information. The widely used unsupervised methods
include change vector analysis (
CVA
) (Chen
et al
., 2003; Azzouzi
et al
., 2015), Otsu’s thresholding method (Bruzzone and Prieto,
2000), and Markov random field based methods (
MRF
) (Moser
et
al
., 2011; Benedek
et al
., 2015). In contrast, supervised change
detection involves extracting change information through min-
ing knowledge from prior information and mainly includes
post-classi cation comparison and direct classi cation. The
first one is to classify each temporal image via supervised clas-
sification and then compare classification maps to determine
changes, and the other one is to directly classify the images
based on selected training samples. Because of strong learn-
ing ability of individual classifiers, they are used to extract the
data information in change detection. The individual classi-
fiers that are commonly used are support vector machine (
SVM
)
(Nemmour and Chibani, 2006; He and Laptev, 2009),
k
-nearest
neighbor (
KNN
) (Guo
et al
., 2003; Roy
et al
., 2012), multinomial
logistic regression (
MLR
) (Li, Bioucas
et al
., 2012; Khodadadza-
deh
et al
., 2014), the extreme learning machine (
ELM
) (Huang
et al
., 2006; Chang
et al
., 2010). However, no single classifier is
capable of extracting all change information, and all learning
algorithms have their limits. In order to improve generalization
ability, ensemble learning (Nemmour and Chibani, 2006; Chel-
lasamy, Ferré
et al
., 2014; Roy
et al
., 2014) has been introduced
into change detection of high-resolution remote sensing images,
including both homogeneous and heterogeneous ensemble sys-
tems. The homogeneous ensemble system means that the same
classifier is used for different training samples, and the research
focus is the construction of different training samples. The
heterogeneous ensemble system is a combination of different
classifiers, and then different fusion strategies are applied to in-
tegrate the classification results to generate the final result. The
widely utilized homogeneous ensemble algorithms are bagging
algorithms (Skurichina
et al
., 2002), random subspace method
(
RSM
) (Skurichina and Duin, 2010; Xia
et al
., 2015), AdaBoost
(Woo and Do, 2015), rotation forest (Rodriguez, Kuncheva
et
al
., 2006). The heterogeneous ensemble algorithms are mainly
concerned with the choice of classifiers and fusion strategy. The
fusion strategies include majority voting (
MV
) (Rojarath
et al
.,
2017), the Dempster-Shafer (
D-S
) evidential reasoning method
(Peng and Zhang, 2017) , the fuzzy integral (F_int) method
(Nemmour and Chibani, 2013), meta-learning (Lin
et al
., 2009).
Many scholars have shown that ensemble systems can improve
the accuracy of change detection. Du
et al
. (Du
et al
., 2012)
proposed a multiple classifier system (
MCS
) based on individual
base classifiers and obtained a satisfying classification result.
Roy
et al.
, 2012 (Roy
et al.
, 2012) developed a semi-supervised
ensemble system based on multilayer perceptron, elliptical
basis function neural network, and fuzzy
KNN
techniques, and
showed satisfying detection performance.
Kun Tan, Yusha Zhang and Xiao Jin are with the Key
Laboratory for Land Environment and Disaster Monitoring of
NASG, China University of Mining and Technology, Xuzhou
221116, China (
).
Qian Du is with the Department of Electrical and Computer
Engineering, Mississippi State University, MS 39762.
Peijun Du is with the Key Laboratory for Satellite Mapping
Technology and Applications of NASG, Nanjing University,
Nanjing 210023, China.
Jiayi Li is with the School of Remote Sensing and Information
Engineering, Wuhan University, Wuhan 430079, China.
Photogrammetric Engineering & Remote Sensing
Vol. 84, No. 11, November 2018, pp. 733–741.
0099-1112/18/733–741
© 2018 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.84.11.733
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
November 2018
733
