836 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 9, NO. 5, SEPTEMBER 2012
Multitemporal Image Change Detection
Using a Detail-Enhancing Approach With
Nonsubsampled Contourlet Transform
Shutao Li, Member, IEEE, Leyuan Fang, Student Member, IEEE, and Haitao Yin
Abstract—In this letter, we propose an unsupervised approach
for change detection in multitemporal satellite images based on
a novel detail-enhancing algorithm. The multitemporal source
images are first used to generate the difference image, which is
decomposed into low-pass approximation and high-pass direc-
tional subbands by the nonsubsampled contourlet transform. The
coefficients from the directional subbands are fused at intrascale
and interscale to extract the meaningful details of the difference
image. After that, the extracted details are injected into one
base image selected from the approximation subbands, which
results in a detail-enhanced difference image. For each pixel in the
enhanced difference image, a dimension-reduced feature vector
is created using the principal component analysis (PCA). The
final change detection map is achieved by clustering the feature
vectors using a PCA-guided k-means algorithm into “changed”
and “unchanged” classes. Experimental results demonstrate the
superior performance of the proposed approach compared with
several well-known change detection techniques.
Index Terms—Change detection, detail-enhancing strategy, dif-
ference image, nonsubsampled contourlet transform (NSCT),
principal component analysis (PCA)-guided k-means.
I. I NTRODUCTION
C
HANGE detection directly compares a pair of input im-
ages acquired on the same geographical area at different
times to identify changes that may have occurred. In recent
years, fast and accurate detection of a changed region has
played a very essential role in many applications, such as urban
studies, disaster management, and agricultural surveys.
The existing approaches for detecting changes can be cate-
gorized as either supervised or unsupervised. Compared with
the supervised techniques, the unsupervised counterparts [1]–
[5] do not require the availability of a ground truth or any other
additional information for the classifier training, thus attracting
more interests [5]. Most of the unsupervised methods are de-
veloped based on the analysis of the difference image, which
can be created by calculating the difference of the input images
[1]–[5]. In [1], Bruzzone and Prieto introduce two automatic
techniques for change detection using the difference image. In
Manuscript received October 25, 2011; revised December 3, 2011 and
December 24, 2011; accepted December 28, 2011. Date of publication
February 21, 2012; date of current version May 29, 2012. This work was sup-
ported in part by the National Natural Science Foundation of China under Grant
61172161, by the Scholarship Award for Excellent Doctoral Student granted by
the Chinese Ministry of Education, and by the Fundamental Research Funds for
the Central Universities, Hunan University.
The authors are with the College of Electrical and Information Engineering,
Hunan University, Changsha 410082, China (e-mail: shutao_li@yahoo.com.cn;
fangleyuan@gmail.com; ocean_waves@126.com).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/LGRS.2011.2182632
Fig. 1. Structure of the NSCT with two levels of decomposition and eight and
four directional subbands in the first and second scales, respectively.
[2], Celik proposes a multiscale Bayesian framework which
performs the expectation-maximization (EM)-based approach
over the subbands resulted from the dual-tree complex wavelet
transform (DT-CWT) decomposition of the difference image.
In general, these approaches provide impressive results for
detecting the changes between the two input images. However,
all these techniques model the changed and unchanged classes
in the difference image by the simple Gaussian assumption, and
therefore, their performances often deteriorate for real images.
Recently, an efficient change detection method [3] has been
proposed through the principal component analysis (PCA) and
the k-means clustering. The PCA is employed for extracting
the feature vectors, which are forwarded to the k-means clus-
tering to compute the final change detection result. In [4], the
feature vector is extracted from the subbands created by the
undecimated discrete wavelet transform (UDWT), and the final
changes are detected by applying the k-means algorithm on
the extracted feature vectors as well. Overall, both the afore-
mentioned two methods perform well when the changes in the
difference image are distinct. However, they fail to characterize
geometrical details of the difference image sufficiently, thus
resulting in improper detection around the detailed regions.
Moreover, since the k-means algorithm used in these methods
will easily get stuck in local optima, they are prone to produce
false detections without a good initial centroid.
In order to address the aforementioned limitations, we
propose a novel unsupervised method for change detection
in multitemporal images, which is based on a new detail-
enhancing strategy. In this strategy, the difference image is
first decomposed into one low-pass approximation subband and
multiple high-pass directional subbands at each scale using the
nonsubsampled contourlet transform (NSCT) [6]. Through
the intrascale and interscale fusion on the directional subbands,
the meaningful details of the difference image can be extracted.
Then, the extracted details are merged into one selected base
1545-598X/$31.00 © 2012 IEEE