The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
307
and an indicated type of change (e.g. increase or decrease of
vegetation). Thereby, the a-priori probability is taken from a so-
called transition-probability-matrix (Pt-matrix), which will be
explained in chapter 3.1 in more detail, whereas the indicated
type of change is given by the fiizzy-combination and -analysis
of indicators (object-based statistics) and derived indicator-
classes (see chapter 3.3). Thus, the combined analysis of a-
priori probabilities of transition (Pt) together with fuzzy-
membership-degrees (p) to indicator-classes can be seen as a
plausibility check of change, whereas the most plausible change
(Pt’) can be understood as the most likely change. Referring to
the envisaged content of the change-layer, for each change-
object its most plausible change (Pt’), together with its indica
tions (membership-degrees to indication-classes) can be deter
mined (see Figure 3). Thus, after applying some generalization
rules to the outlined change-objects the output of the classifica
tion module is exactly the desired change-layer. Regarding the
three-level-hierarchy of the DeCOVER-nomenclature (see De-
COVER, 2008) it is obvious that some changes can only be
assigned in the first or second class-level automatically, while
others are even assignable in the third level. Therefore, as soon
as for a detected change no clear assignment can be given, it has
to be determined manually or left as not assignable.
3. IMPLEMENTING THE CONCEPT
As described in chapter 2, besides the delineation of changes,
each object of the change layer shall be given one or more
likely classes for tl and some tangible terms of expressing the
plausibility for each assumed tl-class. A very central point in
determining the plausibility of a detected change is the consid
eration of the a-priori probability of a class to exchange from
the current (tO) class assignment to another in tl.
3.1 Setting-up the Pt-Matrix
Assuming that the DeCOVER class-hierarchy and its nomencla
ture applies at the point of time tl the same way, as it did in tO
and assuming that a change indicated by the comparison of the
image data of tO and tl indicates a change of class assignment,
then in principle such an indicated change simultaneously indi
cates a change from the class tO to another out of the 38 possible
classes at tl at the indicated position. Since we know a-priori
that some transitions or changes of an object are very unlikely
or even impossible within a given time, e.g. such as from dense
urban area to glacier within a period of two years, while others
are relatively likely, e.g. from arable land to sparse urban area
or construction site within the same period, we can assume that
for an indicated change if there was arable land before the
probability of being sparse urban area or construction site at
the indicated position now is higher than being a glacier there.
I.e.: we can skip procedures which tend to verify unlikely tran
sitions. The probability of transitions from one class to others
can be recorded in an n x n matrix whereas n is the number of
classes. However, for some transitions reliable values are hard
to determine without expert knowledge or analyzing historic
mappings and statistics in detail. Thus, we decided to fill the Pt-
matrix in collaboration with the experts of the DeCOVER con
sortium, whereas each Pt-value is normalized to the range of
[0.0; 1.0]. By sorting each Pt-vector according to the Pt-values
one obtains the most and least probable classes. Nevertheless,
each Pt-value has to be seen as a heuristic value and does not
claim to be 100% true.
3.2. Creating change-objects by segmenting per-pixel indica
tors
In order to keep the results based on the IKONOS- and SPOT5-
data comparable, for all further investigations the IKONOS-1
(blue) and SPOT5-4 (swir) channels were skipped. As reported
in Lohmann, P. et al., 2008 several pixel-based change indica
tors have been investigated regarding their suitability for a fo
cusing module as described here. Therefore a comparison be
tween the indicators and a manual change classification regard
ing the categories Urban and Vegetation has been undertaken.
The results were relatively poor due to several reasons. One of
them is the “change-noise” which is mostly caused by different
illumination situations. However, it turned out in this investiga
tion, that the principal components (PC) generated out of the
channels of tO and tl and the (normalized) difference of corre
sponding channels (Diffnorm) show the best results. As the
authors note, one of the reasons for the relative poor results lies
in the reference used, which was a manual digitizing of recog
nizable changes. Respectively, objects were compared to pixels.
This means changes and “change-noise” was compared to a
noise-free reference, which leads to an overestimation of false-
positives and false-negatives on a per-pixel-comparison. In
order to suppress “change-noise” and to obtain contiguous areas
of change and no change respectively, an image segmentation
based upon the per-pixel change indicators appears to be rea
sonable for the following reasons:
• The border of each change-object is generated along
steep changes of change-indication depending on the
thresholds of the used algorithms.
• Thus, the generated objects can be regarded as more
or less homogeneous areas of change or no change given
by the indicators used for delineation.
• Shape, texture and pixel statistics of each object can
be used to analyze and classify it at least as change- or no
change-object.
Since the so called multi-resolution segmentation (MRS) de
scribed by Baatz & Schàpe, 2000 uses criteria of homogeneity
in color (here in change-indication) and shape it seams to be
adequate to generate reasonable change-objects based on the
per-pixel indicators. This means the criterion of color-
homogeneity of the MRS is generated by the per-pixel-indicator
values, which finally leads to objects of homogeneous change-
indication. As reported by several authors (Meinel, G., et al,
2001a; Meinel, G., et al, 2001b; Neubert & Meinel 2002a; Neu-
bert & Meinel 2002b), a critical point of the MRS is its parame
terization, i.e. to find a well balancing between over- and under
segmentation of the desired objects. To overcome this drawback,
an over-segmentation - i.e. too small neighboring objects with
similar indication values - can be merged according to their
mutual difference of indication values if the difference is below
a defined threshold. Thus, the resulting change-objects are gen
erated in a first step according to their homogeneity of indica
tion values and shape and in a second step according to their
similarity of indication values. Because of the results in Loh
mann, P. et al., 2008 as a first mind it seams to be convincing to
use the difference channels based upon the (normalized) green-,
red- and nir-channel (Diff norm) and the principal components
of the tO and tl channels (PC) as input for the segmentation.
However, regarding the information content in terms of change
or no-change both indicators are relatively redundant - espe
cially PC2 und PC4 are highly correlated to the differences of
the three spectral channels (see Table 2). Additionally, the in
terpretation of PCs - especially of temporal PCs is relatively
ambiguous so that the information content of each single PC is
quite unclear. In the data present, regarding table 2 and table 3
in conjunction, it is quite unclear, whether PC4 reflects the dif
ference in the red and green channels or just the