other
their
jn to
natic
and
such
tliers
s by
yulti-
The
) not
atial
hood
atial
, and
d to
nong
eling
and
non-
y as
the
and
for
yring
es 1S
yatial
tions
ip is
ım et
t the
1sual
ough
the
/exit
pond
sd in
also
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
In general, most existing methods only consider the non-
spatial attributes of a data set, or only consider the spatial
relations and ignore the semantic relations. Further, as all
geographic phenomena evolve over time, temporal aspect
should also be considered (Yao, 2003). How to detect the
temporal outliers or spatial-temporal outliers has been seldom
discussed. Moreover, spatial and temporal relationships exist
among spatial entities at various levels (scales, Yao, 2003).
Such relationships should be considered and reveled in spatio-
temporal outlier detection. Our approach will build on
existing approaches to evolve into a new methodology, which
addresses the semantic aspects and dynamic aspects of spatio-
temporal data in multi-scales.
2. ST OUTLIER DETECTION: PROBLEM
DEFINITION AND PROPOSED ALGORITHMS
Here we adopt the definition of spatial outlier to spatio-
temporal domain and define a spatial-temporal outlier (STO)
to be a spatial-temporal referenced object whose thematic
attribute values are significantly different from those of other
spatially and temporally referenced objects in its spatial
or/and temporal neighborhood.
In order to detect STOs from a data set, the existing methods
for spatial outlier detection can be modified to address the
semantic aspects and dynamic aspects of spatio-temporal data
in multi-scales. Since clustering is a basic method for outlier
detection, we start with it by including the semantic
knowledge in the process. Then, the multi-scale property of
natural phenomena is considered. If a spatial object (which is
created from clustering) disappears after aggregation, it might
be a potential STO. Since spatial objects are dynamic, the
verification of STOs will consider the temporal continuity in
addition to spatial continuity. Therefore a four-step approach
is proposed to identify the spatio-temporal outliers (see Figure
1). Here we call it a multiscale approach since the aggregation
and verification compare the change between two consecutive
scales in space and time.
2.1 classification (clustering)
This involves the classification or clustering of the input data
based upon the background knowledge of the data. The
clustering method is designed based upon the prior-
knowledge and characteristics of the data. If the data is raster-
based images, supervised classification might be applied or a
classifier might be built based upon prior-knowledge
(semantics-based approach). Other methods such as neural
network can also be applied if no prior-knowledge about the
data available. The purpose of this step is to form some
regions that have significant semantic meanings.
2.1 aggregation (filtering)
This step aggregates the clustered result in the previous step
in order to check the stability of the data. In this step, outliers
might be merged out. So it is also called filtering of noises.
2.3 comparison (identifing the merged objects)
This step compares the clustered results derived from Step 1
With the results derived from Step 2 and identifies the objects
1009
that are missed (or filtered) in Step 2, which are potential
STOs. In these step, the comparison is implemented at two
consecutive spatial scales.
2.4 verification (checking temporal neighbours)
This step checks the temporal neighbours of the potential
STOs identified in the previous step. If the semantic value of
such a STO does not have significant differences with its
temporal neighbours, this is not a STO. Otherwise, it is
confirmed as a STO.
Sampling
Data
Classification
ST Objects
Aggregation
v
ST Objects
(reduced scale)
Comparison
Y
Potential
STOs
Verification
y
Identified STOs
Figure 1. Four steps to detect the Spatio-temporal Outliers (STOs).
3. EXPERIMENTAL RESULTS
3.1 Data sets
Ameland, a barrier island in the north of the Netherlands, was
chosen as a case study area. The process of coast change
involves the erosion and accumulation of sediments along the
coast, which is scale-dependent in space and time. It can be
monitored through the observation of annual changes of
landscape units such as foreshore, beach and foredune.
The landscape units are defined based upon water lines. The
foreshore is the area above the closure depth and beneath the
low water line, beach is the area above the low water line and
beneath the dune foot, the foredune is the first row of the
dunes inland from dune foot. Based on height observation, it
is possible to derive a measure of foreshore, beach and
duneness. Height observations have been made by laser
scanning of the beach and dune area and by echo sounding on
the foreshore. These data have been interpolated to form a full
height raster of the test area. In the following analysis, the error
of the height raster, which was used as the original fine
resolution DEM, is ignored.
