472
Photo 6
Photo 7
2.2 From higher altitudes
From 1000 m and 3000 m altitude only groups of trees
or stands can be detected. With increasing pixel size
one has more problems in finding satisfactory homo
geneous training areas for the computer classifica
tion,especially in forests in Germany.
From 1000 m altitude individual tree^shadow, slades
between trees, roads etc. can be distiguished with
some acuracy, but from 3000 m altitude and from the
higher satellite data the contrast of these features
will decrease and so the quality of class seperation
decreases. But this effect can also be an advantage.
Dead trees which are normally salvaged rapidly in
German forests do not appear in the commonly used
statistics of forest damage inventories. But as more
trees are removed from a stand ground vegetation or
soil is increasingly detected. This causes a higher
reflection and a higher standart deviation of the
pixels to represent this stand and so more pixels will
be classified in a higher damage class. This classi
fication might give more accurate information than
the ground inventories which do not count salvaged
trees.
In pict. 1 the results of photointerpretation and
computer classification for the same test sites are
shown.
(Picture 1 )
In dense old stands good correspondence between CIR-
Photointerpretation and the computer classification
was found. Worse results were found in steep slopes
of mountains and on stands of lower density.
The following photos show the results of computer
classifications from different altitudes on a testsite
in Southemwest of Germany.
3 Conclusion
After many computer classifications of daimaged fo
rests one can say that it is possible to classify
with good accuracy healthy and severely damaged fo
rest stands from 1000 and 3000 m with an airborne
scanner and with less accuracy from satellite data.
From satellite data only two classes can be sepera-
ted: healthy and severely damaged forests, and these
only when larger areas are affected. Significant
problems exist in the middle damaged class S2 (26 to
60% needleloss). The wide distribution of this class
makes it difficult to define exact class boundaries,
so quite often pixels will be classified to the
neighbor classes SO-1 or S3. For better results
additional information should be included in the
computer classification process, for example texture,
terrain models, stand density etc. to minimize the
still existing misclassifications.
In the continuing project it is planned to investi
gate in this.
ACKNOLEDGEMENT:
The authors thank Mr. H.P. Kienzle, H. Schneider and
R. Waltenspiel of the Department of Photointerpre
tation and Remote Sensing at the University of Frei
burg for developing special software.
The authors also thank Mr. V. Amann from the DFVLR
Oberpfaffenhofen for aquiring the airborne scanner
data.
LITERATURE:
A. Kadro. Investigation of spectral signatures of
differently damaged trees and forest stands using
airborne multispectral data.
Proceedings of IGARSS'84 Syrnp. , Strassburg, 27. -
30. Aug. 1984.
A. Kadro, S. Kuntz, C. Kim. Entwicklung eines Ver
fahrens zur Waldschadensinventur durch multispek
trale Fernerkundung.
1. Statuskolloquium des PEF vom 5. - 7. Maerz 1985,
Karlsruhe.
A. Kadro. Investigation of Spectral Reflectance Pro
perties of Forest Damage Using Multispectral Data.
3rd Int. Colloquium; Spectral Signatures of Objects
in Remote Sensing, Les Arcs, 16. - 20. Dec. 1985.
G. Hildebrandt, A. Kadro, S. Kuntz. Entwicklung ei
nes Verfahrens zur Waldschadensinventur durch mul
tispektrale Fernerkundung.
Zwischenbericht fuer das 2. Statuskolloquium des
PEF, 4. - 7. Maerz 1986 in Karlsruhe.
A. Kadro. Determination of Spectral Signatures of
Different Forest Damages from Varying Altitudes of
Multispectral Scanner Data.
Int. Symp. on Remote Sensing, 25. - 29. Aug. 1986,
Enschede.