A Simple But Effective Landslide Detection
Method Based on Image Saliency
Bo Yu, Fang Chen, Shakir Muhammad, Bin Li, Li Wang, and Mingquan Wu
Abstract
Effective large-scale landslide mapping is becoming signifi-
cantly important for analyzing natural hazards and providing
landslide locations rapidly for emergency response. Change
detection and machine learning methods are commonly used
for landslide detection. Change detection mostly relies on
several experienced parameters that users have to tune for
different images, which limits the practical application. The
training machine learning model consumes much time, and
it is limited to specific imaging conditions. In this paper, a
simple method for landslide detection using a fixed param-
eter by calculating image saliency is proposed. Landslide is
detected as a saliency object within the background of vegeta-
tion and bare rocks. It is fast and robust for the experimental
images, and outperforms the state-of-the-art, semi-automatic
method in terms of accuracy and computing time. Given the
high efficiency and robustness of the proposed method, it is
applicable to practical cases for hazard estimation.
Introduction
Numerous residents living in mountainous areas are prone to
the destructive natural hazard of landslide. It is mainly caused
by anthropogenic activities (such as deforestation, cultivation,
construction, and traffic vibrations), earthquakes, volcanic
eruptions and ground water pressure (Schuster and Krizek
1978). Based on the report from the
UN/ISDR
(United Nations
International Strategy for Disaster Reduction) and
CRED
(Center
for Research on the Epidemiology of Disasters), landslide ranks
3
rd
among the top 10 natural hazards in causing human deaths
(Martha, 2011; OFDA/
CRED
, 2006). During the recent decades,
landslide has received substantial attention for its tremendous
impact on human loss (Carrara
et al.
, 1999; Corominas
et al.
,
2014; Hervás
et al.
, 2003; Huang and Fan, 2013; Keefer and
Larsen, 2007; Lu
et al.
, 2014; Lu
et al.
; 2015; Metternicht
et
al.
, 2005; Qiu, 2014; Van Westen
et al.
, 2003). Upon landslide
occurring, immediate actions are necessary to rescue the resi-
dents and decrease human loss. Therefore, timely information
about landslide position, area, and destruction level is ex-
tremely important. Apart from this, fast and accurate landslide
detection is also useful for understanding landslides in a large
area and predicting them in the future (Li
et al.
, 2016).
Current research in landslide detection are generally based
on field survey and remote sensed data. Field survey, as a
traditional method, is time-consuming and limited to the geo-
graphical characteristics of research area (Galli
et al.
, 2008).
In contrast, remotely-sensed data is easy to access, and covers
a wide range of area. Remote sensing technology is playing a
dominant role in providing landslide information for policy
makers before and after a disaster (Metternicht
et al
., 2005;
Tralli
et al.
, 2005). With the development of remote sensing
technology, the spatial and temporal resolution of remote
sensed data are becoming higher and higher. Apart from opti-
cal data, the development of Digital Elevation Model (
DEM
)
(Kimura and Yamaguchi, 2000), Synthetic Aperture Radar
(
SAR
) (Di Martire
et al.
, 2016) and Light Detection and Rang-
ing (lidar) benefits mapping landslides to a large extent (Baldo
et al.
, 2009; Razak
et al.
, 2011; Travelletti
et al.
, 2014).
Change detection is the most commonly used method
in mapping landslides for its high efficiency and accuracy,
whether in terms of object-based or pixel-based method
(Akcay and Aksoy, 2008; Cheng
et al.
, 2013; Holt
et al.
, 2009).
From the perspective of object-based change detection method
in landslide detection, each image is segmented into objects
with different sizes and shapes, and each object is composed
of several pixels with similar spectral and textural character-
istics. Ideally, in an object-based method, landslide should
be segmented into an object or a few objects, so that it can be
extracted easily. Daniels (2006) incorporated expert knowl-
edge in building rules to generate object and classify landslide
using
DEM
and optical images; and got higher accuracy than
traditional nearest neighbor classification method. The criteria
proposed in visual interpretation (Soeters and van Westen,
1996; Van Westen
et al.
, 2008) are used to gather pixels to
form an object (Martha, 2011), and obtains 69.1 percent accu-
racy. Object-based method is suitable for high spatial resolu-
tion or very high resolution images, such as QuickBird (Lu
et
al.
, 2011) and panchromatic images (Martha, 2011). However,
high spatial resolution images (i.e., spatial resolution of higher
than 10 meters) are typically not free of charge. That makes
large-scale research much too expensive. Moreover, the per-
formance of object-based method strongly depends on various
feature selections when segmenting image into objects and
merging objects to “meaningful” objects. The features include
Bo Yu is with the Key Laboratory of Digital Earth Science,
Institute of Remote Sensing and Digital Earth, Chinese
Academy of Sciences, Beijing 100101, China.
Fang Chen is with the Key Laboratory of Digital Earth
Science, Institute of Remote Sensing and Digital Earth,
Chinese Academy of Sciences, Beijing 100101 China; the
Hainan Key Laboratory of Earth Observation, Institute of
Remote Sensing and Digital Earth, Chinese Academy of
Sciences, Sanya 572029, China; and the University of Chinese
Academy of Sciences, Beijing 100049, China
(
).
Shakir Muhammad is with the Department of Geography,
University of Peshawar KPK Pakistan, Pakistan.
Bin Li is with the Key Laboratory of Digital Earth Science,
Institute of Remote Sensing and Digital Earth, Chinese
Academy of Sciences, Beijing 100101.
Li Wang and Mingquan Wu are with the The State Key
Laboratory of Remote Sensing Science, Institute of Remote
Sensing and Digital Earth, Chinese Academy of Sciences,
Beijing 100101, China.
Photogrammetric Engineering & Remote Sensing
Vol. 83, No. 5, May 2017, pp. 351–363.
0099-1112/17/351–363
© 2017 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.83.5.351
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
May 2017
351
