described.
one from
he forest.
ene. Very
ical and a
> the final
? but also
sibility of
| Landsat
geführt.
| werden.
er beson-
owie der
on date
), 1992
1992
8, 1992
1991
4, 1993
cloudless
available
Ing scene
relatively
ig of the
lization -
his DEM
) a 25 m
ameters:
1ese and
rived by
a or less
| had to
yy visual
ollowing
e location error due to mapping
e sufficient size of stand
e homogenity of area
e clear assignement to one forest class
e representativity for all classes
e variation of grey values within one class
e overlay of spectral feature space.
Finally, 299 of 583 stands from ground truth data could be
used for training and verification. The training areas were
more or less equally distributed all over Carinthia, but not
equally distributed concerning the forest classes. This was
one reason, why classification had to be performed in three in-
dependent steps, classifying each forest parameter separately.
Furthermore, the ground truth data covering the smaller east
and west images did not allow to classify these images in-
depentently from the larger central scene (compare section
6).
3 DATA PRE-PROCESSING
For mosaicking as well as for overlay of the satellite images
high-precision geocoding with subpixel accuracy was essen-
tial. This was reached by using a parametric, sensor-specific
mapping model taking into consideration image distortions
caused by topographic relief by using a DEM. With this ap-
proach, subpixel accuracy could be reached. Furthermore,
preprocessing of satellite images included atmospheric cor-
rection based on LOWTRAN7 and meteorological data as
well as radiometric correction of topographic effects in each
separate band. For example, the influence of relief on the sig-
nature of satellite images is described by [Schardt, 1990] for
spruce and beech stands in the Black Forest, Germany. The
results are shown for band TM 4 in fig. 1. Due to different
180 120 140 160 180 200 220 248
ILLUMINAT ION
Figure 1: Signature of closed beech (+) and spruce (e) stands
in Landsat TM band 4 depending on illumination [Schardt,
1990].
Illumination the spectral signature of the same class changes
significantly with varying slope and aspect in mountainous
areas. Thus the radiometric correction is even more impor-
tant for alpine regions as is characteristic for Carinthia. The
topographic influence was reduced by radiometric correction
using the Minneart model ([Colby, 1991]).
4 SIGNATURE ANALYSIS AND FEATURE
SELECTION
In order to assess the feasibility of the task, first, results of
earlier signature analyses were consulted. In general, the best
suitable features for the classification of various forest param-
eters are well known from literature (e.g. [Coenradi, 1992,
Horler & Ahern, 1986, Schardt, 1990]). The Landsat TM
spectral bands TM 1 and TM 3 show a greater chlorophyll
absorption, thus, the reflection is decreasing with increasing
vegetation cover, whereas it is increasing in bands TM 2 and
TM 4. The bands TM 5 and TM 7 are more sensitive to the
biomass in general and to the leaf water content then to the
green vegetation. Here the reflection also is decreasing with
increasing vegetation cover. Band TM 1 is less suitable due
to its sensitivity to atmospheric effects and bands TM 5 and
TM 7 are strongly correlated. Therefore, the Landsat TM
bands TM 2, TM 3, TM 4, and TM 5 are contain the most
information for forest applications.
Besides that, the feature selection was based on further inves-
tigations such as the analysis of cluster diagramms, the calcu-
lation of correlations between features and forest parameters,
and the analysis of statistical results from classification tests
with diverse feature combinations, taking into consideration
the most suitable phenological stages.
5 CLASSIFICATION
The classification was based on ground truth, which served
as training as well as verification data sets. lt had to be
performed separately for each forest parameter due to un-
derrepresentation of some classes in the ground truth data.
The advantage of this approach was the selection of the most
suitable feature combination for each forest parameter.
The actual edge of the forest, that is the separation of forest
from non-forest areas, was classified by using a combination of
thresholds in band TM 2 of the August scene (for separation
of forest from other vegetation and non-vegetation) and TM 4
of the June scene (for separation from water). The forest
type (4 classes from coniferous over two mixed forest classes
to deciduous) were best separated in TM bands TM 3, TM 4,
and TM 5 of June. The stand age (3 classes) also influenced
mainly the TM bands TM 4 and TM 5, best classification
results were derived by using band TM 5 and the ratio of
band TM 4 and TM 3 of June. Finally, the stand density (2
classes) was classified using TM bands TM 2, TM 3, TM 4,
and TM 5 of August as features.
As age of treee stand and tree speciec composition both in-
fluence the same features or spectral signatures respectively,
it would be meaningful to classify these parameters jointly.
Due to the unfavourable distribution of training sites within
the subclasses defined by that this was not possible. For
the same reason it was also impossible to take into account
further factors such as the altitude or the ground cover.
5.1 Classifiers
In a broad variety of applications we acquired much expe-
rience with Maximum Likelihood [Bahr, 1985], a statistical
classifier widely used in remote sensing applications. It is very
important to find the best features — be that single channels
or combinations — for Maximum Likelihood to yield in good
performance. Therefore preprocessing is demanding both in
terms of knowledge and time. On the other hand there is no
necessity to care of overlearning (bad generalisation). The
603
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996