You are using an outdated browser that does not fully support the intranda viewer.
As a result, some pages may not be displayed correctly.

We recommend you use one of the following browsers:

Full text

Fusion of sensor data, knowledge sources and algorithms for extraction and classification of topographic objects
Baltsavias, Emmanuel P.

International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999
oblique looking bands. Furthermore, the urban/man-made
construction class often appeared overestimated in the
combined approach, which, in relation to the failure of the
anisotropy approach to distinguish accurately the bare soil
signature from the urban area signature, could lead to the
conclusion that this phenomenon is a result of the synergy effect
between multispectral and anisotropic information.
Among the classes that were common in all three approaches,
the forest types could be well recognised in all of them. An
information that is missing in the anisotropy approach is the
presence of grass at the forest edges, which is presented
accurately by the other two approaches.
The class "peas" was more accurately estimated by the MS and
the combined approaches. Slight differences between the MS
and the combined approaches could be noticed through a
detailed examination of the whole classified scene. These
differences concern the capability of the maximum likelihood
classifier to classify all pixels of the same plot to the same class
(cultivation). In this respect, better performance was observed
by the combined approach.
The overall accuracy of the combined and the MS approach
classification was more or less at the same level, with the
combined approach being slightly better. The accuracy of the
anisotropy approach was lower than the other two, which was
expected, since the other two approaches, apart from bands 1
and 4, use part or most of the information that the anisotropy
approach uses.
Due to time constraints, the investigations on the Diimast
dataset could not be performed with the completeness and
accuracy, which are requested for reliable conclusions. The
rectification was performed without DEM. No atmospheric and
terrain relief corrections were applied to the dataset.
Additionally, the evaluation of the potential of the anisotropy
approach was limited due to the unfavourable illumination-to-
sensor geometry. In spite of these facts, the false colour
composite of the stereo data (6/7, 6, 7) showed a surprisingly
high information content. In any case, the results allow
discussing some trends.
The "anisot" FCC displays texture and structure elements more
clearly, improving the detection of infrastructure and
settlements. The detailed analysis showed partially different
features in the two FCCs.
The example of winter wheat indicates that during specific
stages of crop development, stand structure and/or individual
plant architecture are more universal object characteristics than
the multispectral ones.
The example of mowed meadows demonstrates a situation,
well-known from the side-looking perspective of a walker,
which is also valid for remote sensing data analysis, i.e. that a
plot looks green despite a ground coverage of almost 5%. With
the anisotropy approach, it seems that is possible to detect
sparse vegetation in a very early development stage.
A point that should be noticed in the anisotropy approach is the
hot spot and the shadow effect. At object boundaries where also
significant height differences are observed, a buffer zone
appears, where pixels have a distinct radiometric behaviour, as a
consequence of the particularity of the irradiation at these
regions. The result is that these pixels are often misclassified. In
this case, they were often classified as urban/man-made
construction class.
An analysis of the results of the multispectral versus the
anisotropic evaluation, both the visual as well as the computer-
based one, indicates that the classification is determined by
different biophysical parameters. In case of vegetation, the
multispectral approach explores mainly the absorption of
incoming radiation by pigments and water, while the anisotropy
information is due to stand structure and plant architecture,
which are often an effect of phenologic and physiologic status.
The hypothesis that multispectral and anisotropic information
are complementary could be proven at least for the main
landcover classes forest, agricultural areas, settlements, water
bodies and infrastructure.
The synergistic potential of the combined use of multispectral
and anisotropic information could not be demonstrated as
clearly as in the investigations described by Schneider et al.
(1999). For the evaluated dataset, the increase in classification
accuracy for the combined multispectral and stereo band
analysis is not significant. The low difference between the
backscatter signal of the two stereo bands should be the reason.
Despite this result of the computer-based analysis, the results of
the visual interpretation let us hope that a further substantial
increase can be expected using common classification routines,
which consider also textural and context information like the
DELPHI2™ approach (deKok et al., 1999).
For more detailed investigations on this topic, field
measurements, approximating the BRDF, have to be performed,
preferably simultaneously to stereo data acquisition from space
or airborne sensors.
Last but not least, it is worth noting, that the mode D band
combination of MOMS-2P has been proven to lead to better
results than the mode 3 data of the D2 mission. The concept of
optimised spectral coverage in the visible range by substituting
band 3 (red) by band 1 (blue) led to a real improvement,
especially for visual interpretation. By combining the
multispectral bands 1 and 4 with the panchromatic stereo bands,
which provide the spectral information from blue to red, almost
true colour images can be produced.