The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
mapping but rather as a “notification/indication” of change with
a possibility to indicate geometric and attributive change occur
rences (i.e. class transition(s)). Therefore, the output of the De-
COVER change-detection procedure is a GIS-data-layer
(change-layer), which holds the geometry of detected changes
(change-objects) plus information about the assumed changes
together with indications for the plausibility of these assump
tions. Thus, it relates to objects as part of an existing database
but compares two images at different dates on a per-pixel level,
giving a pixel- and segment-based indication of LU/LC change.
Methods for the comparison of images from different dates may
be grouped into those which use univariate image differencing
alone (Singh, 1989, Fung, 1990), methods to compare vegeta
tion properties like NDVI or Tasseled Cap Tranformations
(Richards, 1993), or change vector analysis (Lambin, 1994,
Bruzzone et al, 2002). A comprehensive overview of existing
pixel based techniques, their advantages, disadvantages and
resulting accuracies is given by Lu et al, 2004. Relating pixel-
based indications to objects means developing an approach for
integrating these indications into an object based analysis.There
are several ways to implement such a procedure as has been
shown by different authors (Schopfer, 2005, Busch et al., 2005
and Gerke et al., 2004). Because users of DeCOVER prefer
some type of a “change notification”, which might also be use
ful for their specific application, like updating of user operated
databases (i.e. ATKIS) by proprietary techniques an attempt as
described in the following chapters has been made to set up and
realize a prototypical framework for CD. The framework is split
into two main modules: a focusing module and a classification
module. Both have been designed and realized in close coopera
tion between the company GeoData Solutions (GDS) and IPI.
2.1 Focusing Module
Within the CD-concept this module plays an important role,
since it outlines potential changes by generating change-
segments based upon per-pixel change indicators. Steps of pre
processing, necessary to render the two images comparable in
both the spatial and spectral domains are included (co
registration, radiometric normalization). Figure 2 shows an
outline of the focusing module.
Figure 2: The Focusing Module.
With respect to the spatial domain miss-indications by displaced
pixels in both images should be avoided. However, due to dif
ferences in illumination and view angles some “change noise”
is still unavoidable. In order to outline changes indicated by a
per-pixel comparison of the images within the focusing module,
an image segmentation based upon one or more pixel-based
change-indicators is applied. This way, not the changes them
selves, but borders between different change indication-values
are generated, which consequently leads to a spatial differentia
tion of change- (high indicator values) and no-change-areas
(low indicator values) in terms of changes in signal. These seg
ments can be subsequently handled as image objects (of change)
which consequently offers the whole palette of object based
image analysis (see Blaschke, T. et al, 2008; Schöpfer, E.,
2005). Although all of these advantages of segmenting indicated
changes, the typical drawbacks of this approach cannot be de
nied: the generated image objects should ideally represent true
change-objects, which means: each change should be repre
sented by only one image-object (no over- or under
segmentation). This in turn leads to the problem of finding suit
able segmentation algorithms and parameterizations.
change-object
jjjj Pt-Matrix
PKClass to. Class t1)
* oeCOvER to indication A indication B indication C
^{indication A) (Vindication B) (Vindication Q
îndcato* h
Indicator h
□
M№)
rTfcw
Hi.)
Mil»)
ru
□
Mil»)
□
M(i-5
I Un
:3->
Md')
‘"lud--)
□
mUO
HK)
-—H
Imp*)
Pt' = Pt A ^indication A) A ^indication 8) A p (indication C) A .
Figure 3: Illustration of determining plausible changes (Pt’) for
a segmented change-object by combining fuzzy-assignments to
indication-classes (indication A to C) using different statistical
parameters on a per-object basis for each indicator with a-priori
probabilities of change (Pt). The segmentation is based upon
one or more pixel-based change-indicators (II to In).
Last but not least, comparing for each change-object its class
assignment in the mapping of the GIS-database (DeCOVER tO),
the a-priori probabilities (Pt) of change for this class and the
statistics of the underlying (indicator-) images leads for each
segment to an estimation, whether the segment outlines a
change and if so whether the type of change, i.e. the new class
can be determined automatically or manually. Vice-versa, each
object of the DeCOVER-tO-mapping can be marked as changed
as soon as it covers or overlaps a change-object.
2.2 Segmentation and Classification Module
Within the classification module in principle the last step of the
focusing module is recursively applied for each change-object
until the most plausible change, i.e. the most plausible class at
the second point of time (tl) can be determined. In the worst
case no such evidence can be given except that a change in sig
nal has been detected, but its class-assignment in tl remains
unclear. In all other cases each change-object can be assigned to
one or more classes with a certain degree of a-priori probability