ALPMON FINAL REPORT - ARC systems research

Contract ENV4-CT96-0359 ALPMON
forest. Furthermore, streams could not be classified due to their representation in mixed pixels
together with bank vegetation. On the other hand the auxiliary data supported classification itself by
establishing rules for selected classes (e.g. sealed surfaces do not occur on steep slopes).
Partly, additional features derived from the high resolution satellite data sets (see WP8) were
integrated in the classification process. Texture features were used for the delineation of sealed
surfaces in the Dachstein and Cordevole test sites. The forest / non-forest separation in the Dachstein
test site was based on a fused Landsat TM/Spot pan image.
Some main difficulties crystallised in most classifications:
� Forest age in many test sites could not be classified due to the natural appearance of the forest
with a mixture of all tree age categories. Only the Dachstein test site is dominated by managed
forest with a common age class within one stand.
� Forest type in some test sites is mainly mixed, with only a few pure coniferous or broad-leaved
stands.
� Tree species composition is an important information for two applications (national park
management and avalanche risk). However, some tree species could not be separated (spruce,
pine, and stone pine). Dwarf mountain pine could not be separated sufficiently with the satellite
data available for most test sites. This is on one hand due to spectral similarity with other
coniferous trees and, on the other hand, due to mixture with other tree species in one stand.
However, in the Dachstein and the Mangfall test site it was possible to classify dwarf mountain pine
and green alder by means of winter respectively spring images, where the low growing trees are
covered with snow whereas the high growing trees are snow free.
� Wet lands were well recognised in the Mangfall test site by using the thermal infrared band of
Landsat TM (TM6). However, this approach is not applicable e.g. to the Dachstein test site, as wet
land areas here are relatively small and beyond the resolution of TM6 with 120m.
� Sealed surfaces cannot be separated from rock, gravel, and soil based only on the spectral
characteristics of these classes, as the spectral signatures are too similar. Additional information
has to be introduced to classify this category with higher accuracy, or the information must be
taken from GIS data.
On the other hand, most categories required by the customers could be classified with sufficient
accuracy. In general no problems occurred with the classification of the forest type in those areas,
where pure coniferous or broad-leaved stands are common. Forest age could be classified at least in
some test sites, where pure age stands occur. In the Dachstein test site it was possible to separate
pure larch stands from mixed spruce/larch stands and pure spruce stands (possibly mixed with pine
and /or stone pine). In the Dachstein and Mangfall test sites even dwarf mountain pine and green alder
could be classified by means of winter satellite data. Furthermore, some classes could be divided into
subclasses (e.g. pasture and meadow, carbonate and silicate rocks) giving more detailed information.
Problems with the classification of water bodies or sealed surfaces could be overcome by introducing
auxiliary data into the classification process. In general, it can be stated that the outcome of
classification was satisfactory with respect to the local as well as the European wide applications. This
was confirmed by the customers in the first ALPMON customer work shop held from February 4-5,
1999 at Innsbruck, as well as in the final project evaluation.
2.2.7 Verification of Classification Results (WP10)
The verification phase was intended to give evidence on the quality of the classification results
(Congalton, 1991). The results of the verification are confusion matrices for each classification giving
true error rate estimates for each class and for the total classification. The accuracy of the
classification results was determined for all classified parameters based on the ground information
collected at each sample plot. The statistical analyses of accuracy included Kappa statistics, confusion
matrices, regression analyses of observed versus estimated values, etc.
All analysis of classification results was based on the use of an error matrix or contingency table. It is a
very effective way to represent accuracy, in that the accuracy of each category is plainly described
along with both the errors of inclusion (commission errors) and errors of exclusion (omission errors)
present in the classification. Error matrix can be used for a series of descriptive and analytical
statistical techniques.
JR, RSDE, ALU, LMU, Seibersdorf, WSL 23