optimal length, width, area, textural characteristics, color, and
relationships between different objects. It is difficult to apply
an object-based method with the same set of parameters to
different images, because the characteristics of different land-
slides and their backgrounds are mostly different.
Pixel-based change detection in extracting landslide is
much easier than object-based method. Normalized Difference
Vegetation Index (
NDVI
) is commonly used in change detection
(Martha, 2011; Yang and Chen, 2010), for the reason that
NDVI
is able to discriminate vegetation from other objects in the im-
age, and landslide is mostly likely to occur in vegetation areas.
Several semi-automated methods in landslide detection based
on vegetation change have been proposed and showed good
performance (Li
et al
., 2016; Zhang
et al.
, 2010). Moreover,
change detection can be synthesized with Markov random
field (Li
et al
., 2016) and Principal and Independent Compo-
nent Analysis (PCA and
ICA
) (Mwaniki
et al.
, 2017). Pixel-
based change detection method is not limited to high spatial
resolution images, it can be used on Landsat and MODIS data
as well, which are both free of charge (Lu
et al.
, 2004).
Apart from change detection, there are also researchers
detecting landslide using single post-disaster images. The
methods are mainly focused on classification algorithms and
various source data, such as
SAR
(Di Martire, Tessitore, Branca-
to, Ciminelli, Costabile, Costantini, Graziano, Minati, Ramon-
dini and Calcaterra 2016, Mwaniki, Kuria, Boitt and Ngigi
2017, Raucoules
et al.
2013),
DEM
(Kääb 2002) and single-
lens reflex camera (Travelletti
et al.
2012). The classification
algorithms comprise maximum
likelihood (Parker
et al.
2011),
K-nearest neighbor (Cheng
et al
.,
2013), Support vector machine
(Colkesen
et al.
, 2016), logistic
regression (Colkesen
et al
., 2016).
Moreover, owing to the contribu-
tions of computer vision com-
munity, landslide detection is
starting to synthesize the devel-
oped machine learning method
and computer vision method for
better performance (Cheng
et al
.,
2013; Novellino
et al.
2017).
There is one main drawback
that the methods proposed above
are difficult to apply directly to
different images without adjust-
ing the parameters, because
most methods have experienced
parameters that need the user to
tune or retrain a model accord-
ing to different landslide cir-
cumstances. Our paper intends
to propose a simple but effective
method in landslide detection,
which is directly applicable
in different images. It consists
of two steps: (a)
NDVI
change
detection between two images
of pre- and post- landslide; and
(b) calculate saliency probability
(Yildirim and Süsstrunk, 2014)
of the change image, where
saliency probability is the prob-
ability of a pixel belonging to
landslide.
Study Area
This study is conducted in
two sites. One is located along
the valley region belonging to
Kathmandu, Nepal, shown in
Figure 1. The other is situated in
Shenzhen, Guangzhou Province,
China, shown in Figure 2. In the
first study area, an earthquake
of Mw7.8 took place on 25
April 2015, located at 28.147°N,
84.708°E (Tsuchida
et al.
, 2015).
Figure 1. Research area One.
Figure 2. Research area Two.
352
May 2017
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
