GIS-
internal
model
image
analysis
related
object
model
users
internal
purely
QlS-internal
Image analysij
IQIS interface!
image analysis^
internal
purely Image
analysis related
user's
object
model
image
analysis
model
Fig. 5. The different parts of the object related models in image
analysis and GIS-analysis: the image analysis related object
model and the user's object model are linked by the Image
Analys/s-GIS-interface, it belongs both to the interpretation
model and to the GIS-internal object model. The interpre
tation model as well as the GIS-internal object model both
contain parts which are not relevant to the other model.
The analysis itself depends on user requirements which
either are already (implicitly) contained in the object
model, have to be provided during the analysis or result
from the interpretation model, e.g. in form of a GIS-query.
In these cases only a small part of the object model, e.g.
the upper levels of a specialization hierarchy, may be
available, i.e. formalized. This may prevent or at least
hinder a smooth link with other information sources being
available.
We will come back to this model.
Interpretation models
Interpretation models5) describe the result of the analysis
step. It in principle contains all aspects of the object
derivable from the images, and possibly other information
sources, as far as the object model allows. Thus not only
the user relevant parts of the scene are contained in the
interpretation model but also all other parts which are
necessary to complete the analysis, as e.g. parts of
objects or background which the user is not interested in.
This would imply all aspects of the object to be observ
able in principle, thus the interpretation model and the
object model to coincide.
But much more likely the interpretation model will only
cover parts of the object model namely those which a
careful analysis of the design has found to be observable
and excluding those parts of the object model which are
only needed within the analysis procedures, e.g. all kinds
of preprocessing steps. This model then can be seen to
form the interface between image analysis and GIS-
application (fig. 5).
This interpretation model moreover has to be distin
guished from the (object) model used in a GIS, the whole
analysis process is designed for. This GIS-internal object
model on one hand will not contain irrelevant detail visible
in the scene, which might be necessary for the analysis
itself, and on the other hand will contain aspects of the
scene which may not be relevant to the image analysis
but only to the GIS-user. The different parts of the object
related model are shown in fig. 5.
The description of the interpretation process up to now
did not explicitly include prior knowledge not being of the
form of an object model. Specifically existing maps or
other knowledge sources are not mentioned. However
maps or even non geometry related information may also
be seen to be objects which have to be observed, from
which features or signatures have to be derived which
require a symbolic description and have to be analyzed 6).
All these information sources may be contained in a GIS.
Therefore not only play photogrammetry and remote
sensing the role of a special data source for GISs but also
vice versa: GISs play the role of a general modelling tool
for analyzing image data providing the interpretation tools
for image analysis. This requires a smooth interaction
between the different models discussed above and a
much more detailed analysis of the task of a GIS.
Fusing information during image analysis
Zooming into the analysis model reveals several infor
mation and knowledge sources to take crucial parts in the
analysis process (fig.6). Those are at least images, image
models, non-image data, strategic knowledge and the
human analyst.
Formally the fusion process to be discussed may be inter
preted as inverting the observation equations: o. f
(unknown parameters p; image models; non-image data),
i to yield p (or, image models; non-image
data; strategic knowledge; human analyst), where the
parameters include geometric, physical parameters,
classes, relations etc. Whereas the relations between the
different models have been discussed before this para
graph deals with the inherent complexity and therefore
difficulty of the analysis process. It is caused by:
the incomplete knowledge about the observation func
tion e.g. when modelling the reflectance function;
the incompatibility of the object models resulting in the
incompatibility of the image models, e.g. when linking
optical and radar images;
the heterogeneity of the non-image data, specifically
of maps, e.g. resulting from different purposes, dif
ferent age or different scale;
the hitherto unstructured strategic knowledge, e.g.
when trying to reduce the inherent high computational
complexity of matching algorithms and the nearly un-
formalizable capabilities of the human analyst.
human
Image
Analysis
image
models
analyst
strategic
knowledge
non-Image data
(maps,
images
5) Recall that we use the notion interpretation to describe the result
of the analysis, thus meaning a state rather than a process.
6) Analyzing and interpreting existing maps actually follows the
same steps and requires the same type of models.
interpretation
Fig. 6. Some of the information sources to be fused when analyzing
image data.
376
NGT GEODESIA 93 - 8