The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
311
Image Object Information
Evaluation of Class: VL -> BS, Bl, BV, FA. FN
and (min)
not homogeneous by area and SO
Classification value of change by norm_Zt
Classification value of decrease in vegetation >0.5
increase of StdDevs
Feature Value Weight
not applicable
not applicable
not applicable
not applicable
Value
0.457
0.457
0.457
1.000
1.000
1.000
44 Vi ►: ►» \ Features / Classification / Class Evaluation f~
Image Object Information
Term
Feature Value
Weight
Value
Evaluation of Class: VL -> BS. BI. BV. FA. FN
0.000
and (min)
0.000
no! homogeneous by area and SO
not applicable
0.018
Classification value of change by norm Zt
0 784306338
not applicable
0.784
Classification value of decrease in vegetation >0.5
1
not applicable
1.000
increase of StdDevs
not applicable
0.000
<
>
4L 4 | ► »;\ Features / Classification /Class Evaluation /
Figure 8: Evaluation of change-indication for objects to be
marked as change (top) (here from agriculture to not vegetated)
and no change (bottom).
Identity Nom
'1 - Jon Pt n
1 VU
<Top-n»Mtiâvër>" jE]f
Location'
3.373.155,303 ! -
Field
VST H
FID
244?
5twc
Polygon
Osssf**
0,457103
l
Oks/i_2
1
3
I
OKSfi.4
0,542897
Oâisfied
1
FID i
2447
mMCisjxic
Via
Pt I DeCOVE
0,176056
K2„DeCOVE
0,175056
«.3J>COVE
0,060028
PM.DeCOVE
0,070473
Pts DeCOVE
0,070423
Ft6_D«COVE
0,070123
Pt7J>«C0V£
0,052817
Pt8J>eCOVE
0,052817
0,035211
TÔC151 £*•:.-
VLg
TOaS2_OeC
Vis
TOC153 Dec
VLk
TOCLMjDeC BSg
Toass
-B5S
TOCLS6 D»C
61!
TOaS7JDeC
BSm
Toasej*:
BVs
Toasojxc
8Ss
Ptljiew
0,15594
Pt2_new
0,15594
P43_i»w
0,07797
PMJ19W
0,06
Figure 9: Change-objects with class-assignments from De-
COVER_tO (dark red labels) and indicated most probable class-
assignments for tl (black labels) with a-priori Pt values (black
lables) and Pt-values adjusted by //-values of indication classes
(purple labels). Right: object table, whereas BSg is marked as
the most probable class in tl
This way, for each object its change indication and possible
transition can be evaluated (see Figure 8 and 9).As Figure 8
shows, an object is only marked as changed if for none of the
indication classes a degree of membership of p - 0.0 is given.
Additionally, it demonstrates the reliability of an indicated
change by the overall membership degree given through the
fuzzy-and operator. In the examples given, the top result indi
cates a clear change, since p to almost each of the indicator-
classes is 1.0. A reasonable way to combine the a-priori Pt-
values with the degrees of membership in order to obtain one
value to indicate the plausibility of an indicated change is to
calculate the mean of the product of the Pt-value of an indicated
transition and the //-values to the indicator-classes. Figure 9
outlines the conjunction of a-priori Pt-values with membership-
values to describe the transition of an object from agriculture
(VL) to non-vegetated (BS, BI, BV, FA, FN). The marked
object in Figure 9 is the same as in Figure 8 at top. Although the
highest a-priori transition probabilities are given to VLg, VLs
and VLk {grassland, other permanent crop and complex agri
cultural) the indication-classes indicate a transition to non-
vegetated. Thus, not the agricultural classes are given as the
most probable (most plausible) classes but the most probable
class out of the non-vegetation category, in the example BSg
{low density urban area).
3. CONCLUSIONS AND OUTLOOK
The paper demonstrates the current stage of the development of
automated procedures to outline potential changes and to evalu
ate them automatically as far as possible within the context of
the DeCOVER project. Methods were outlined which are capa
ble to detect and outline changes in multi-temporal image data
on the basis of per-pixel and per-object change indications.
These outlined changes are analyzed by means of methods of
object based image analysis and fuzzy class assignments. It has
been demonstrated, how the fuzzy-membership of a detected
change to defined indication classes in conjunction with a-priori
knowledge about transition probabilities can be combined in
order to give evidence about a change in terms of the De
COVER nomenclature. The current stage of these developments
has to be seen as still prototypical. Expected aspects that emerge
to become objects of further research in this field are: improv
ing the reliability of a-priori transition probabilities by taking
sound analyses of historic information and spatial context into
account. Especially within an operative environment it seems to
be reasonable to adapt the current probabilities in accordance
with changes already detected (“self-learning Pt-matrix”). Spa
tial context has not been considered yet, although its influence
on the probability of a transition cannot be denied. However, in
order to focus on potential change-areas and to give indications
about what could have happened, the current status seams to be
capable to reduce manual effort. Within the context of change
segmentation on the basis of indicators, there is still demand
about optimizing algorithms and their parameterization.