PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
August 2015
637
An Unsupervised Urban Change Detection
Procedure by Using Luminance and Saturation
for Multispectral Remotely Sensed Images
Su Ye and Dongmei Chen
Abstract
Unsupervised change detection techniques have been widely
employed in the remote-sensing area when suitable reference
data is not available. Image (or Index) differencing is one of the
most commonly used methods due to its simplicity. However,
past applications of image differencing were often inefficient
in separating real change and noise due to the lack of steps for
feature selection and integration of contextual information.
To address these issues, we propose a novel unsupervised
procedure which uses two complementary features, namely
luminance and saturation, extracted from multispectral images,
and combines T-point thresholding, Bayes fusion, and Markov
Random Fields. Through a case study, the performance of our
proposed procedure was compared with other three unsuper-
vised change-detection methods including Principle Component
Analysis (
PCA
), Fuzzy c-means (
FCM
), and Expectation Maxi-
mum-Markov Random Field (
EM-MRF
). The change detection
results from our proposed method are more compact with less
noise than those from other methods over urban areas. The
quantitative accuracy assessment indicates that the overall ac-
curacy and Kappa statistic of our proposed procedure are 95.1
percent and 83.3 percent, respectively, which are significantly
higher than the other three unsupervised change detection
methods.
Introduction
There is a growing interest in monitoring land-use/land-cover
change as it provides up-to-date information for many appli-
cations. Employing remote-sensing (RS) technology has been
critical for keeping track of land-use/land-cover transition at
a variety of spatial scales (Rogan and Chen, 2004; Hussain
et
al.
, 2013). Compared with traditional monitoring methods
(such as field surveying), RS-based change detection can
better allow for processing large areas, producing quantita-
tive results and offering repeatable procedures (Coppin
et al.
,
2004).
Numerous state-of-the-art approaches have been developed
to analyze RS imagery for change detection. These methods
are usually categorized into supervised and unsupervised
methods, according to the availability of adequate reference
data (Bruzzone and Prieto, 2000; Bruzzone and Prieto, 2002;
Fernandez-Prieto and Marconcini, 2011). The advantage of
supervised change detection is the capability of labeling the
type of change (the detailed “from-to” information) based on
given training samples. However, the generation of suitable
multi-temporal reference data to characterize all the classes is
usually a difficult task, especially for historical images (Lu
et
al.
, 2004). Compared with supervised methods, unsupervised
ones can be much more cost-effective since no reference data
is required. In spite of being unable to offer the information
on categories of land transition, the changed/no-change
detection is often acceptable for many practical applications
(Hussain
et al.
, 2013).
Image differencing (or index differencing) is one of the
most commonly used methods for unsupervised change
detection (Bruzzone and Prieto, 2002; Rogerson, 2002; Lu
et
al.
, 2004). Compared with other unsupervised approaches,
such as Principle Component Analysis (Deng
et al.
, 2008)
or clustering algorithms (Bruzzone and Prieto, 2000), image
differencing is much cheaper computationally, and it is easier
to interpret its results (Lu
et al.
, 2004; Hussain
et al.
, 2013).
The basic idea for image differencing stems from the fact
that the physical status of land area can be characterized by
certain feature indices derived from the remotely sensed data;
when we analyze targeted features from bi-temporal images,
the larger its deviating values from means of unchanged class
appear to be, the more likely it is that change has occurred in
the corresponding area. The useful features for image differ-
encing can be defined as digital number in a single spectral
band, vegetation indexes (Singh, 1989), principle component
(Deng
et al.
, 2008), or texture index (Tomowski
et al.
, 2010).
Feature-differencing values of interested areas are usually
passed to a thresholding strategy to separate “no-change” and
“changed” class for the final result map.
However, image or index differencing often exhibits incon-
sistent performances, as it makes its decision relying only on
single feature analysis. For most urban change-detection tasks,
when single feature differencing is applied, we may have (a)
real change information corresponding to transition between
different land-cover types which are usually of interest, and
(b) noisy change identification due to other factors, such as
seasonal growth or local illumination variance. In the compli-
cated practical scenes, clusters of real and noisy changes are
sometimes mixed together in the feature space; thus, we are
unable to completely separate them by using a single thresh-
olding value. In this sense, fusion techniques merging multiple
difference images have been introduced to improve detection
accuracy (Le Hégarat-Mascle and Seltz, 2004; Du
et al.
, 2012),
since different features might offer complementary informa-
tion about the patterns to be classified (Kittler
et al.
, 1998) .
The second issue with traditional image differencing is
that global analysis of difference image fails to account for
local spatial information influencing the reliability of final
result. To address this issue, one solution is incorporating the
direct difference of certain texture indices for change detec-
tion (Li and Leung, 2002; Tomowski
et al.
, 2010). Another
method is applying Markov Random Fields (MRFs) models
Department of Geography, Queen’s University, Kingston, On-
tario, K7L 3N6 (
).
Photogrammetric Engineering & Remote Sensing
Vol. 81, No. 8, August 2015, pp. 637–645.
0099-1112/15/637–645
© 2015 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.81.8.637
