International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999
In
14
Above applications are in 2D and increasingly in 3D, while
multitemporal (4D) approaches are still rare.
Quite long is the list of problems, which are encountered with
respect to data and information fusion:
• Differences between landuse (provided by GIS) and landcover
(provided in images).
• Lacking procedures for interpretation and quality control of
fused images.
• Fusion and the mixed pixels problem.
• Distortion of spectral properties with pixel-based fusion
techniques (partially avoidable by feature-based fusion).
• Different levels of data quality regarding geometric accuracy
and thematic detail.
• Large differences in fusion of multitemporal data and change
detection.
• Differences between the data in spatial and spectral resolution
(centre and width of band), as well as polarisation and angular
view.
• Data generalisation in map and GIS data.
• Different levels of data abstraction and representation
(resolution/scale).
• Models of objects used in image analysis are often simplistic,
limited in number and not general enough; on the other hand,
generic models may be too weak and broad.
• Different data representations, even for the same object or
object class (roads, road networks).
• Different data structures for the same object (e.g. raster,
vector, attribute).
• Lack of accuracy indicators for the components to be fused.
• Data are often inhomogeneous, i.e. acquired by different
methods, several analysts etc.
• Algorithms to fuse information are restrictive, mechanical, not
intelligent enough.
• Abrupt decisions (as humans take sometimes) are not
permitted by algorithms. Rules/models (e.g. roof ridges are
horizontal) always have exceptions, but still should be used,
e.g. with associated probabilities which can be updated by
accumulation of knowledge, processed data etc.
• Rules/models differ spatially (e.g. buildings in Europe differ
from those in developing countries) and in time (old buildings
differ from new ones).
• Architecture of systems is complex, processing requirements
high, commercial systems or support tools are limited or non
existent.
• Gap between research and practice (which is typical for not
matured scientific areas). 3
3. USE OF GIS DATA AND MODELS IN IMAGE
ANALYSIS
GIS data and generally object models are generally used to
provide information (geometrical, spectral, textural, functional,
temporal etc.) about the target object(s), its attributes, as well as
other objects and information related to the target ones. GIS
information can be used in image analysis for various purposes:
• Provision of initial approximations for some unknown
parameters and thus reduction of the search space and
increase of the probability of success.
• Provision of clues for target objects based on information on
other objects, related to the target ones, e.g. use of existing
information on road network to detect buildings.
• Quality control of the results, e.g. by serving as "ground
truth".
• In classification, e.g. in supervised classification to
automatically select training areas, and in unsupervised one to
automatically assign detected information classes to thematic
classes.
• Use of provided cues and information in various stages of
object recongnition and reconstruction, e.g. to: (a) find
objects, e.g. buildings by extracting the 3D blobs of a given
DSM ; (b) exclude wrong hypotheses and detect blunders, (c)
exclude regions that are impossible for a target object, e.g.
building or road in water surfaces.
• Use of data to support hypothesis generation about the model,
e.g. try to infer the roof type (or some possible ones) based on
the given building outline.
In most cases, the integration of images analysis and GIS can
not lead to full automation. Thereby, human interaction and
intervention becomes necessary. The important points are how
and when this interaction should occur. Generally, the
interaction can (a) be preventive or have a guidance character,
and (b) be corrective. Usually, the first case occurs at the
beginning of the processing, the second one at the end. Whether
the first or second approach is more appropriate depends on
how much the quality and efficiency (time aspects) of the whole
process are improved, as the result of such interaction. In this
respect, preventive approaches seem to be preferable. Human
intervention is also necessary to define the framework of the
solution to a given problem:
• Analysis of the problem, definition of the strategy.
• Selection of building blocks that should be used (data,
knowledge sources, processing methods etc.).
• Decision on interactions between the blocks and definition of
the processing flow.
4. PREREQUISITES FOR INFORMATION FUSION
Before fusing and integrating information, several prerequisites
should be fulfilled:
Co-registration. By this we mean that the different components
should be compatible/comparable with respect to various
aspects: spatial (data should refer to the same area and
coordinate system), temporal, spectral (inch appropriate
corrections due to terrain relief, atmosphere, sensor calibration),
resolution (pixel footprint, number of bits per pixel). Spatial co
registration is always a prerequisite. An overview of co
registration and geocoding methods is given in Raggam et al.,
1999 (these methods should be appended by the increasingly
used direct sensor data geocoding methods, using integrated
GPS and INS). Depending on the application and the level of
fusion, co-registration with respect to some of the above aspects
is not necessary. E.g. while image fusion of multispectral and
high resolution SAR imagery might be incompatible and not
appropriate, integration of object cues from such images might
be feasible and desirable. The same applies to temporal co
registration, e.g. while for object extraction multiple data of the
same o
propert:
change
is not a
operatic
have ti
transfer
Same s
direct c
road ce:
Clear o
not, as
user, an
an exan
part of
exampli
networl
centerli
two-diri
bicycle
intersec
general]
in some
Need o
data its
Unfortu
which i
among
transfer
Quality
combim
often n<
DTM ol
for a cl
or may
success]
needed,
better
appropr
and fusi
Regardi
• The c
will a
be me
• The d
from
often
Instea
detect
comin
locali:
• Deep
disad^
appro
