International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
Figure 9. Initial 3D Regions, 3D Regions after refinement,
Boundary of 3D Regions.
This defect is compensated by incorporating the relevant
information in image space. The corresponding areas in image
space for the 3D object are determined by the inverse solution
of collinearity condition equations (Figure 10).
ri
Figure 10. The Superimposed boundaries of the extracted
objects in image space.
Figure 11 presents the final extracted regions in image spaces.
As this figure shows our feature level fusion strategy has
successfully identified the presence of sub-regions within the
initial regions and hence the 2D regions are subdivided
accordingly to separate segments.
i
Figure 11. Final extracted regions in image space
4. DATA FUSION IN DECISION LEVEL
The decision level fusion is performed by using the identity
declaration provided by each source of information. The fusion
of the identity declaration is then made by using /dentity based
methods such as MAP and Dempster-Shafer methods (Shefer,
1976) or Knowledge based method such as Expert knowledge
Neural network and Fuzzy logic methods (Lin and Lee, 1996).
It is Important to note that decision level fusion use: (1) feature
extraction, transforming the raw signal provided by the sensor
into a reduced vector of features describing parsimoniously the
original information, and (2) identity declaration or object
recognition that assigns a quality class to the measured produce
based on the feature extraction process (Figure 12).
572
„BERGE NENNEN RE ae
S Feature p» _ Object %
4 Extraction Recognition |
pum es mE Object Object
Feature fp Object — g—»| Recognition §—
Extraction Recognition Fusion
è ere [ERY
Feature „| _ Object
Sn Extraction Recognition
Figure 12. Object/Decision Level Fusion
4.1. Case Study — Automatic Change Detection
Monitoring of changes in topographic digital vector maps is one
of the main requirements of urban planners, urban decision-
makers and managers (Dowman, 1998; Armenakis et. al., 2002;
Kim and Muller, 2002). However, in practice the processes for
analysing the changes are the manual methods like on-screen
change detection that are time consuming and expert dependent.
The availability of the new generation commercial high
resolution satellite imageries, due to their wealth of information
content, have opened a new era in the problem of automatic
change detection and consequently the digital vector maps
updating. Therefore, automatic change detection has been an
area of major interest in remote sensing and GIS for the last few
years (Peled, 1993; Darvishzadeh, 2000; Dowman, 1998;
Armenakis et. al., 2002; Kim and Muller, 2002; Schiewe 2002;
Shi and Shibazaki, 2000).
Nevertheless, most of the existing methods for doing change
detection process are basically optimized to use information of
one sensor imagery and in addition, by employ parametric
methods, object's fuzziness behaviour and the possibility for
introducing training potentials are basically neglected. In this
case, an attempt has been made to design a system that
integrates all above features in a total and comprehensive
automatic change detection solution. The approach presented
here takes advantage of the concept of fusion in two levels of
feature and decision. That is, information fusion to exploit the
multi-level characteristics of the objects and logic fusion for
enhancing the learning and hence recognition abilities of the
system (Figure 13).
Change
Figure 13. Decision Level Fusion
4.2. Experiments and Results
The proposed automatic change detection methodology was
tested on a 1:1000 scale digital map and a pan-sharpen
IKONOS scene of the city of Rasht, Iran (Figure 14). The maps
have been produced in 1994 from 1:4000 aerial photographs by
National Cartographic Centre (NCC) of Iran. The satellite
Interne
imager
years ti
IKONC
occurre
Figur
corresp
The ob
differen
establisl
change
still far
change
presente
based or
5. Concl
Data fus
rapid gr
the pre:
advancei
selected
that can
approacl
