cable to
lem. The
y of a
the need
If this
practice.
fulfil our
d to the
POT-Pan
juired on
tion with
cted less
ch more
c image.
case) the
| because
far better
was not
>st suited
nentation
a quality
a minor
/ proving
ulored to
yyed to a
6
ntical
fication
3x3
ow with
security
%
4 | 99.6
4 | 97.3
7| 95.0
21 97.1
97
)7
ween the
ation had
ult of the
pretation.
1e remote
! one can
see, that the results are very similar with the most
obvious difference in the classes „Fields‘‘ and
Settlement". It looks like the satellite image analysis
overestimates the ,Settlement" class at the cost of
Fields".
Column 3 shows the number of pixels which are identical
if compared pixel by pixel. While the percentage of
identical pixels is very high for the classes "Water",
"Forest" and "Fields", the class ,Settlement shows a
poor result. The main reason for that inaccuracy is the
uncertainty of the texture analysis for the ,Settlement"
segmentation, which tends to extend the class into
surrounding areas (usually fields) while the visual
interpreter has decided to set the class boundary much
tighter to recognizable houses. The poorer result is not
necessarily a pure drawback of the computer analysis. We
should be aware that already the class definition is not
very clear: Are backyards ,Settlements", for instance?
(see figure 5)
81% of the controlled pixels are classified with high
certainty. Column 4 lists the number of those pixels, that
are identical with the visual interpretation. In this group
the percentage of successful classification is, not
unexpectedly, greater and significantly greater for the
previously mentioned uncertain classes.
Additionally we wanted to take into consideration a slight
inaccuracy of the geometric rectification. We know, and
we must expect in any case, that our layers of the various
data sources do not geometrically fit to each other within
the accuracy of one pixel. In order to allow a small
displacement of some 10 m, we perform the identity
check again by comparing in 3x3 window instead of pixel
by pixel (i.e. a Ix] window). Column 5 lists the number
of pixels whose classes are identical. The total number of
identical pixels rises from 90% to 93%, while there is
almost no difference for the pixels classified with high
certainty (column 6). They are usually not boundary
pixels and therefore not influenced by the window check.
QU a xai :
a. :
Fig 4: Comparison of classification and visual interpretation of
the class ,,Forest*
Figure 4 and figure 5 show details of the comparison. The
left image contains the classification result, the right one
the panchromatic satellite image. Superimposed on both
are the class boundaries of the visual interpretation from
the orthophoto. The forest boundary in figure 4 gives a
rough idea about the geometric accuracy of the image
rectification. The vector image of the forest boundary has
the same shape as in the satellite image but from the
slight displacement in the order of 2 to 3 image pixels we
must conclude inaccurate geometric rectification. On the
left hand image the reference boundary is clearly outside
the classified forest by some 15 m. Figure 5 reveals one
typical problem of the settlement classification that has
been already concluded from table 2. The visual
interpreter found much smaller areas than the automatic
algorithms, that overestimates settlement areas by
including gardens and backyards in the class.
the class ,,Settlement*
7 CONCLUSION
The completion and the update of a nation-wide
information system (such as the DLM of the Austrian
BEV) with landuse data can be carried out successfully
up to a certain accuracy by classifying high resolution
data together with multispectral data and if possible with
information from an already existing GIS. All the data
sources are subjected to special classification procedures
and then linked to each other in a rule based hierarchical
classifier. The quality of the result depends on the
resolution of the original images, of course, but also on
the classes. For not well-defined classes like ,,Settlement*
one must expect a need for some additional work for
visual post-interpretation or field verification. Still, the
automatic procedure can be enormously cost-saving as
the process works automatically to a high extent and for
all uncertain program decisions the user receives hints
through reliability or certainty codes where a closer check
is recommended. The practical example proved that with
current satellite sensors, such as IRS-1C, SPOT and
Landsat TM, a geometric acurracy of about 15 m x 15 m
can be achieved. The classification compared to a visual
interpretation yielded an average success rate of up to
93%. With the new generation of satellites that have been
announced for the near future and with adapted
interpretation algorithms (texture analysis will become
increasingly important) the overall quality may even be
improved. As a summarising result and a proof of the
quality figure 6 shows the landuse layer with the situation
layer of the Austrian map 1:50000 superimposed on it.
8 REFERENCES
Ament R. (1997) Orthobilder in der ATKIS-Fortführung.
In Fritsch, Hobbie (Eds): Photogrammetric Week '97,.
Wichmann Verlag Heidelberg, pp. 127-134.
Fórstner W. (1991), Statistische Verfahren für die
automatische Bildanalyse und ihre Bewertung bei der
Objekterkennung und -vermessung, Heft Nr. 370. DGK,
München.
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 279
