IEEE Geoscience and Remote Sensing Magazine - September 2015 - 20

often unsatisfactory [119]. This is a direct result
of the fact that PCC does not take into account
the temporal correlation. However it has been
Multiclass change
widely employed in the literature both at pixel
RefeRence
KinD of Data
Detection technique
and region level [127]-[130]. Attempts to inSupervised Post-classification
Howarth et al. (1981)
Multispectral
crease PCC accuracy have been done by using
Comparison
Singh (1989)
SAR
more than two images in the fusion step [131].
Hall et al. (1991)
Multisensor
Supervised direct multidate classification
Bruzzone et al. (1997)
Smits et al. (1999)
(DMC) [1],[113],[132], unlike PCC, takes into
Serra et al. (2003)
account the dependence existing between two
Pacifici et al. (2007)
images of the same area. The main idea of
Ling et al. (2011)
Kempeneers et al. (2012)
such a technique is to characterize pixels by
Supervised Direct-Multidate
Schowengerdt (1983)
a vector obtained by stacking the feature vecClassification
Singh (1989)
tors related to the images acquired at the two
Hall et al. (1991)
different times. Then the identification of the
Bruzzone et al. (1997)
Volpi et al. (2013)
land-cover transitions is carried out by considCompound Classification and
Bruzzone et al. (1997)
ering each transition as a single class and by
Domain adaptation
Bruzzone et al. (1999)
training a classifier to recognize such transiFernández Prieto et al. (2001)
tions. It is worth noting that a complex conCossu et al. (2005)
Serpico et al.
straint to satisfy for using this technique is to
Demir et al. (2012, 2013)
have a training set composed of training pixels
Unsupervised approaches
Byrne et al. (1980)
related to the same points on the ground at the
Häme et al. (1998)
two times. In addition, training pixels should
Bovolo et al. (2007)
represent accurately the proportions of all the
transitions in the whole area of interest. This
represents a serious drawback as, in real applications,
it is difficult to obtain training sets with such
constraints). It becomes even more complex since there
characteristics. In [121] Schowengerdt remarked that,
is a need of multitemporal information. Semi-supervised
since spectral and temporal features have equal status in
approaches represent a trade-off between the two abovethe combined data set, they cannot be easily separated in
mentioned conditions. It is worth noting that all the
the pattern recognition process. As a consequence, class
techniques based on classifiers cited in this section are inlabelling may be difficult if relatively simple classificatrinsically suitable to be used with different kind of data
tion algorithms are considered. In [132] DMC has been
and also with multisensor,
adapted to the use in VHR images.
multisource and multiresoA more realistic approach to fusion at decision level
lution information.
is compound classification (CC) [113]. Similarly to the
The post-classification
SuperviSed and partially
DMC, also in this case the objective is to perform the
comparison (PCC) (also re(Semi) SuperviSed
classification of pixels of the two images according to the
ferred to as delta classification
methodS baSed on
maximization of the posterior joint probability of classes.
[1]) is the simplest technique
Conditional probabilities of classes can be estimated
among
fusion
at
decision
claSSification are
according to different techniques and under different
level
approaches.
It
performs
intrinSically Suitable to
assumption on their statistical distribution. On the one
change
detection
by
comparproceSS multiSenSor/
hand, with respect to the PCC, the CC technique allows
ing the classification maps
multiSource data.
the temporal correlation between images to be considered
obtained by classifying inin the change-detection process. On the other hand, with
dependently the two images
respect to the DMC method, the CC technique allows the
considered. For each change
constraints related to the training sets to be relaxed [113].
the land cover transition is
In particular, training pixels should not necessarily be
obtained in an explicit way. The main advantage of delta
related to the same area on the ground [113]-[114].
classification lies in the fact that multitemporal images are
In real problems it may happen that, given a series of
classified independently, thereby minimizing the problem
multi-temporal images, a ground truth is not available
of radiometric calibration. Although PCC has been extenfor all the items of the series. In such realistic cases, the
sively used in several applications, its performance strongly
aforementioned supervised approaches cannot be emdepends on the classification accuracies of the classifier apployed. In [122], an ensemble of non-parametric mulplied to each single image. After multitemporal fusion the
titemporal partially-supervised classifiers was defined
accuracy is close to the product of the accuracies yielded by
and integrated in the context of a multiple classifier
the independent classifiers [1],[113], making the approach
table 4. suMMaRY of the Main techniques foR Multiclass change
Detection (the table is not eXhaustiVe).

20

ieee Geoscience and remote sensing magazine

september 2015
Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - September 2015
