International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
200 
is a step with far-reaching consequences. To cope with this, soft 
classifiers (Schowengerdt, 1997) are a useful alternative. Fuzzy 
logic decision rules belong to these soft classifiers and offer in 
combination with feature selection a large reduction in 
complexity and a proper aid to group the spatial objects into 
meaningful classes. 
1.2. The object-oriented vision 
In this article, the important difference of an object-based 
classification method is explained, using a specific type of 
landscape feature, the forest stand. Object-based classification 
starts with the crucial initial step of grouping neighbouring 
pixels into meaningful areas, especially for the end-user. This 
means that the segmentation and object (topology) generation 
must be set according to the resolution and the scale of the 
expected objects. 
For foresters, the typical spatial objects can range from 
forest/non-forest-masks to forest-stands and even crown 
surfaces, depending upon the resolution of the images. For 
example, the commercial coniferous stands in Germany are 
often 3 to 15 ha in area, so their expected area is typically 
around 35 to 200 pixels in a standard Landsat scene. In airborne 
scanners such as the DP A, each tree crown has a diameter of 
about 10 to 20 pixels. Typical crown discrimination is possible 
with at least around 2 m resolution, using a rule of thumb that 
proper object discrimination is possible if the object area is at 
least 10 x 10 pixels. This statement is not a contradiction to the 
case, where a single pixel, representing a road crossing, is used 
as ground control point in geometric rectification. In the latter 
case, the neighbourhood characteristics are more important than 
its absolute radiometric value. 
The first segmentation layer will be used to detect (sub)-objects, 
which are locally homogeneous. Segmentation layers with 
higher tolerance levels will be defined through the relationship 
with the so-called sub-objects. At the basic level of sub-objects, 
classical maximum likelihood algorithm or nearest neighbour 
functions can still be used, as the radiometric properties are 
important and feature space is the domain, where decisions are 
being made. For higher level objects or super-objects, the 
spatial context is more important than the object mean grey 
value. Fuzzy rules play here a more important role. 
According to user preferences, objects of interest are grouped 
into a class. This capability of starting queries, including object 
topology, directly after the segmentation process, offers a large 
advantage over pixel-based methods. 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 set of 
condensed crisp end-membership functions. Integrating expert 
knowledge is also possible using GIS layers; they can function 
as boundary decision layers in the segmentation process. The 
thematic output of the object-based classification is typically a 
GIS layer, immediately usable in GIS analysis. 
2. FEATURE SPACE AND STATISTICS 
The object-oriented approach is also a philosophy of improved 
image understanding. Human vision is very capable of detecting 
objects and object classes within an image. To pursue this type 
of analysis, it is important to stay close to the ‘intuitive’ image 
understanding. Consequently, classical algorithms of pixel 
based image analysis are becoming less important. The spatial 
context plays a modest role in pixel based analysis. Filter 
operations, which are an important part of the pixel based 
spatial analysis, have the limitation of their window size. In 
object analysis, this limitation does not exist. The spatial 
context can be described in terms of topologic relations of 
neighbouring objects. Below, some typical examples are 
presented to show the advantages of the object-oriented 
approach. Three major domains are discussed: 
• The object analysis in feature space, using fuzzy logic 
• The impact of high resolution images and their statistical 
problems 
• The integration of high resolution panchromatic data into 
multispectral analysis 
2.1. Feature space and fuzzy logic 
In image analysis, the important decisions are made within the 
feature space, where the characteristics of the different bands 
are being mapped as a set of vectors representing the pixel- 
value. The mathematical assumption is the uncorrelated 
behaviour of the N different bands in N dimensional space. 
Fuzzy logic rules can be applied to a set of projections of the N 
dimensional feature space upon an N x 1 dimensional set of 
lines, which do not have to be uncorrelated. Representing an 
ellipse in 2 dimensions as a box with 4 comer points, where 
each pair of comer points belong to an uncorrelated band, 
seems contra-intuitive. This procedure can nevertheless be 
useful for a few classes only but not for the image as a whole. 
This is exactly what is done in the case of a 2-step image 
classification, where a box classifier is used to mask out the 
water bodies or the cloud/shadow cover in the first step and the 
resulting output is used in a normal maximum likelihood 
classification. The ellipse in a 2D space can be also modelled 
through a polygon in which each pair of points belongs to an 
image axis, correlated with other ones. For a 2-D space, this has 
no particular advantage. When increasing the number of bands, 
however, the simplification of a hyperellipse to a hyperpolygon 
is becoming an interesting alternative. This procedure is not 
new, as feature extraction and soft classifiers are known from 
the remote sensing literature (Schowengerdt, 1997). In this 
study, by using the fuzzy logic concepts, the feature extraction 
is defined for each class using a selected set of correlated or 
independent axes system. In the example of Fig. 2, the band 1 is 
based on the ‘first moment’ filter technique. This is described in 
the ENVI software as follows: ‘These moment operators such as 
first and second moment options are simple texture measures 
utilising the moments of the grey level histogram of the 
processing window (see Russ, 1992). The first moment is a 
measure of the contrast in an image. The second moment is a