Forest Classification of Multitemporal Mosaicked Satellite Images
U. Schmitt and G. S. Ruppert
Institute of Digital Image Processing
Joanneum Research, Graz, Austria
scu@pdib40.joanneum.ac.at
rup@pdib40.joanneum.ac.at
Commision VII, Working Group 5
KEY WORDS: Photogrammetry, Remote Sensing, Forestry, Classification, Mosaic, Neural Networks
ABSTRACT
Experiences with the classification of forest condition in Carinthia, one of the nine federal provinces of Austria, are described,
Three Landsat TM satellite scenes had to be mosaicked to cover all of Carinthia. Two such mosaicked scenes, one from
June and one from August, were used for classification to take into account the different phenological stages of the forest.
Additionally, the scenes were used to replace areas covered by clouds with cloud free data from the respective other scene. Very
much effort was put into this classification project to receive knowledge on choosing classifieres — we tried a statistical and à
neural network classifier — optimize the training process — true error rate approximation, generalization — and judge the final
performance — statistical or human expert. Criteria for the final classification result was not only the true error rate but also
the confusion matrix giving the desired importance for each class. An additional subjective criteria was the comprehensibility of
the training and the classification results and the visual appearance of the classification border lines of the individual Landsat
TM images of one scene of Carinthia.
KURZFASSUNG
Für das Bundesland Kärnten wurde eine flächendeckende Klassifikation des Waldzustandes aus Satellitenbilddaten durchgeführt.
Um ganz Kärnten abdecken zu können mußten drei Landsat TM Szenen zu einem Mosaik zusammengesetz werden.
Der Artikel beschreibt wesentliche Aspekte der Datenaufbereitung, der Merkmalsauswahl, der Klassifikation unter beson-
derer Berücksichtigung verschiedener Klassifikatoren, der Mosaikbildung aus mehreren Landsat TM Szenen sowie der
Qualitätsbeurteilung der Klassifikationsergebnisse.
1 INTRODUCTION
C Table 1: Satellite images.
In support of the forestry framework for the province of Sensor Scene Acquisition date
Carinthia a forest classification based on satellite images has Landsat TM | 192-27 quarter August 9, 1992
been performed. This shall record the actual state of the for- 191-27/28 floating June 15, 1992
est for the whole province. Satellite images serve particularly 191-27/28 floating August 18, 1992
well for this type of problem because they allow to derive 190-27/28 quarter, flt. | June 22, 1991
forestry parameters that are not — or at least not in a suitable 190-27/28 quarter, flt. | August 14, 1993
scale — available from maps or other sources. The parameters
which had to be recorded by means of the classification were
the actual edge of the forest, the forest type (4 classes), the — not be covered completely by a singel scene. Nearly cloudless
stand age (3 classes), and the stand density (2 classes). Landsat TM images covering all of Carinthia were available
for the peroid of 1991 to 1993 (tab. 1), in which floating scene
191-27/28 coveres nearly all of Carinthia except of relatively
small parts in the East and West. For preprocessing of the
satellite images - geocoding and topographic normalization -
a digital elevation model (DEM) was necessary. This DEM
was available in a 50 m raster and was resampled to a 25 m
raster.
2.2 Ground Truth
The classification results afterwards shall be integrated into a
geographic information system (GIS) together with other in-
formation that is required for the forestry framework, such as
digital elevation models (DEMs) and geological maps. There
they shall be jointly processed and overlayed for planning and
analysis purposes. In this connection the classification results
present themselves as the most actual GIS layer.
The work steps necessary for realising the task and the em-
ployed methods shall be presented in the following exposition. The forest condition is mainly defined by three parameters:
Forest type, stand age, and stand density. For these and
2 DATA SPECIFICATION some additional forest parameters ground truth derived by
- field work of the client was available. It was more or less
2.1 Satellite Data and DEM equally distributed all over Carinthia. Ground truth had to
The area to be classified covered approximately 15.000 km?. undertake extensive post-processing and selection by visual
Lansat TM data were suitable for giving an overview of the control before it could be used as training set. Following
forest condition over such a large area. Furthermore, neither selection criteria were applied:
SPOT nor Russian KFA data was available with blanket cov-
erage within an acceptable period. Therefore, the analysis
was based on Landsat TM data. It was necessary to form
a mosaic of several satellite scenes, because Carinthia could e location error due to image distortion
e location within the central satellite scene
e cloud coverage
602
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996
