Unsupervised Object-Based Differencing for
Land-Cover Change Detection
Jinxia Zhu, Yanjun Su, Qinghua Guo, and Thomas C. Harmon
Abstract
One main problem of the spectral decomposition-based
change detection method is the lack of efficient automatic
techniques for developing the difference image. Tradition-
al techniques generally assume that gray-level values in a
difference image are independent and multitemporal images
are co-registered/rectified perfectly without error. However,
such assumptions are often violated because of the inevitable
image misregistration and the interference of correlations
between spectral bands. This study proposes an automated
method based on the object-based multivariate alteration
detection/maximum autocorrelation factor approach and the
Gaussian mixture model-expectation maximization algorithm
to obtain unsupervised difference images. This procedure is
applied to bi-temporal (2005 and 2006)
SPOT-HRV
images at
Panyu District Ponds, China. Results show that the proposed
method successfully excludes the correlations of spectral
bands and the influence of misregistration, as evidenced by a
higher accuracy (up to 93.6 percent). These unique technical
characteristics make this analytical framework suitable for
detecting changes.
Introduction
The use of satellite images for change detection analysis has
increased over the past two decades because of the increased
spatial, temporal, and spectral resolution, more accurate
geo-referencing procedures, and the release of archival data
(Coppin
et al.
, 2004). Accompanying these developments,
there is a need for better solutions to locate and character-
ize land-cover changes. Both supervised and unsupervised
approaches have been used for detecting changes (Castellana
et al.
, 2007; Longbotham
et al.
, 2012). Although supervised
approaches offer advantages in recognizing land cover tran-
sitions and processing multisource images, the associated
ground truth labeling can be a difficult task. Consequently,
unsupervised change detection methods are still commonly
used, especially for historical images. High spatial resolu-
tion remote sensing images provide unique opportunities for
the detailed characterization and monitoring of landscape
dynamics. In this study, we mainly focus on the widely used
unsupervised change detection techniques for very high reso-
lution (
VHR
) imagery based on image differencing.
Traditional methods are based on two assumptions: (a)
the independence of the difference images (Chen
et al.
, 2011;
Ghosh
et al.
, 2011; He
et al.
, 2011; Gong
et al.
, 2012; Bovolo
et
al.
, 2012; Chen
et al.
, 2013; Neeti and Eastman, 2014; Subud-
hi
et al.
, 2014; Dronova
et al.
, 2015), and (b) no error in the
difference image co-registration or rectification (Singh, 1989;
Coppin and Bauer, 1996; Verbyla and Boles, 2000). However,
these two assumptions may often be violated, and therefore
lower the change detection accuracy. For example, although
current techniques can produce concentrated change informa-
tion with uncorrelated variables, the interference of correla-
tion between different bands is often neglected (Nielson
et al.
,
1998; Im and Tullis, 2007; Zhang
et al.
, 2007; Brown
et al.
,
2007; Meola and Eismann, 2008). For example, bands 4 and
5 from the same Advanced Very High Resolution Radiometer
(
AVHRR
) image typically have a correlation of 0.99 (Nielsen
et
al.
, 1998). This strong correlation could result in the detail
information of original images being neglected. Although the
commonly used principal component (PC) transformation
produces concentrated change information (Hotelling, 1936;
Besic
et al.
, 2015), which can remove the correlation between
bands from a single-date image, removing the correlation of
bands from multi-temporal images remains a challenge (Gal-
laudet and Simpson, 1994).
Moreover, it has been widely proven that the effect of
sub-pixel misregistration between two images is not negligible
(Coppin and Bauer, 1996; Dai and Khorram, 1997; Eismann
et
al.
, 2012; Falco
et al.
, 2016). Townshend
et al.
(1992) found
that over 40 percent of the error in the actual difference of
NDVI
values was caused by one-pixel misregistration. Town-
shend
et al.
(1992) and Dai and Khorram (1998) demonstrated
that a registration accuracy of less than one-fifth of a pixel
was required to generate a change detection result with less
than 10 percent error. Theiler and Wohlberg (2012) addressed
that one of the most confounding sources of incidental dif-
ferences was the inevitable imprecision in the co-registration
process. Shi and Hao (2013) showed the quantification of the
relations between registration/rectification errors and image
edges could improve the understanding of spatial distribution
of change detection errors. According to Chen
et al.
(2014),
high spatial resolution imagery typically had higher spectral
variability within neighboring pixels than low-resolution
datasets, which could further exaggerate the influence of
misregistration. Jabari and Zhang (2016) found that the relief
displacements of elevated objects in
VHR
images usually led
to significant misregistration that negatively affected the
accuracy of change detection. Although many researchers
have focused on correcting registration errors (Dai
et al.
, 1996;
Chatelain
et al.
, 2007; Zhu
et al.
, 2011; Stow
et al.
, 2016;
Barazzetti, 2016; Jabari and Zhang, 2016), few studies have
Jinxia Zhu is with the Institute of Land and Urban-Rural De-
velopment, Zhejiang University of Finance & Economics, 18
Xueyuan Road, Hangzhou, China 310019
Yanjun Su and Thomas C. Harmon are with the School of
Engineering, Sierra Nevada Research Institute, University of
California, Merced, 5200 North Lake Road, Merced, CA 95343
Qinghua Guo is with the School of Engineering, Sierra Neva-
da Research Institute, University of California, Merced, 5200
North Lake Road, Merced, CA 95343, and the State Key Lab-
oratory of Vegetation and Environmental Change, Institute of
Botany, Chinese Academy of Sciences, Beijing 100093, China
(
)
Photogrammetric Engineering & Remote Sensing
Vol. 83, No. 3, March 2017, pp. 225–236.
0099-1112/17/225–236
© 2017 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.83.3.225
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
March 2017
225
