International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999
199
OBJECT-BASED CLASSIFICATION AND APPLICATIONS IN THE ALPINE FOREST ENVIRONMENT
R. de Kok, T. Schneider, U. Ammer
Chair for Landuse Planning and Environmental Protection, Faculty of Forest Science, University of Munich, Am Hochanger 13, D-
85354 Freising, Germany, ammer@abies.lnn.forst.uni-muenchen.de, Roeland.dekok@lrz.uni-muenchen.de
KEYWORDS: Image Classification, Context, Segmentation, Fuzzy Logic, Forestry.
ABSTRACT
Context based classification is an important field of study in digital image analysis. Neighbouring pixels may possibly have almost
equal grey values. The information of local homogenous patterns in a patch based landscape organisation has lead to studies,
assuming the organisation of landscape patterns as being a complex of local spectral distributions. Image segmentation techniques
are well known and an advanced algorithm is used in this study. New sensor generations meet the strong market demands from end-
users, who are interested in image resolution that will help them observe and monitor their specific objects of interest. The mixed
pixel problem and the increasing difficulties in the spectral analysis of high resolution images make it necessary to develop additional
methods of classification. The key factor is to concentrate on the spectral properties of the objects of interest. This has important
consequences. The increasing resolution (<5 m) leads to very complex spectral analysis. Fuzzy logic decision rules offer here a large
reduction in complexity and a proper aid to group the spatial objects into meaningful classes. In this study, a new software package is
used for object-based classification, developed by DELPHI2™ Creative Technologies.
With this software, the segmentation procedure has to be set according to the image resolution and the scale of the expected objects.
For foresters, the typical spatial object can range from forest-stands to crown surfaces. According to user preferences, objects of
interest are grouped into a class. The fuzzy logic decision rules for class membership are the framework in which the expert
knowledge has been embedded. The synergy of the spectral properties, the neighbourhood object influences and the expert
knowledge lead to powerful ways of object membership decision rules. The fuzzy logic rules guarantee the transparency of the
decision rules and reduce complexity to a condensed crisp set of end-membership functions. Integrating GIS layers is equally
possible. The output of the object-based classification is typically a GIS layer.
1. INTRODUCTION
1.1. History and user demands
Digital image classification has been based upon three principal
methods:
Pixel based spectral signature, pixel-statistics from GIS objects
and context oriented pixel classification (Carl, 1996). Since
several decades classification beyond spectral signature alone
has been recognised as an important study field. In a digital
landscape image, the chance of a pixel containing the same
value as its neighbours is much more likely than another
random pixel in the image, a feature that is very useful in image
compression techniques. Of course, this depends on the
relationship between a chosen spatial resolution and the type of
landcover. In general, a typical landcover class contains a
considerable amount of pixels. The information of local
homogenous patterns in a patch based landscape organisation
has lead to several studies, assuming the organisation of
landscape patterns as being a complex of local spectral
distributions, belonging to specific landscape features. Image
segmentation techniques have been developed since a few
decades and the original work done by Kettig and Landgrebe
(1976) and the theories of Cross and Mason (1988) and Gorte
(1996) in which the image segmentation philosophy is
thoroughly explained still have high theoretical value. In this
study, the "Image Analysis" software from DELPHI2™,
Creative Technologies (a Munich software firm) is used to
explore the possibilities of advanced segmentation and object-
oriented applications.
End-users are familiar with very high resolution data from aerial
photographs, which fulfil most of the user needs. Multispectral
image information in more than three bands, however, is only
possible with digital scanners. When this information is needed,
aerial scanners become much more expensive. The cost of
visual interpretation is the bottleneck for an increase in their
practical utilisation. Considerable improvements in automatic
image analysis are essentially dealing with cost reduction of
image interpretation. Meanwhile, the amount of images
covering certain parts of the Earth is increasing so rapidly that
automatic image analysis offers the sole solution to extract
important information. The applications community is
interested in image resolutions that will help them observe and
monitor their specific objects of interest. The increasing
resolution and the physical properties of objects in the images
with a geometrical resolution of less than 5 m leads to very
complex spectral analysis (Kenneweg et al., 1991). The mixed
pixel problem and the increasing difficulties in the spectral
analysis of high resolution images make it necessary to develop
additional methods of classification. The key factor is to
concentrate on the spectral properties of the (spatial) objects of
interest, instead of the class statistics of the whole image. This