International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
201 
measure of homogeneity’. The axis of ‘first moment’ is not 
independent from the standard deviation in panchromatic band 
2. Local heterogeneity in the panchromatic image has also a 
strong effect on the standard deviation value in the same 
‘window’ area of the image. Nevertheless, they can be 
simultaneously used to classify areas, where spectral values in 
multispectral bands are less important compared to local 
texture. The initial development of these texture image analysis 
methods were focussed on pictorial imagery, however, they 
have equally good potential in defining pattern indices (Musick 
and Grover, 1990). Also in this study involving an alpine forest 
environment, texture patterns show important variations. 
Decisions made with correlated axes can be used in image 
classification without any violation of statistical decision rules 
for particular objects. Well-known measures in image analysis 
such as LAI and NDVI, which are projections of pixel vectors 
on correlated axes can be very helpful in characterising 
particular object classes. Fuzzy logic rules can be regarded as a 
set of hyper-box classifiers with overlap (instead of crisp ones) 
and overlap weight factors. Although the total image 
information in feature space is reduced through the set of 
projections and therefore contains less information than the 
complete feature space, this simplification can be of major 
advantage for the areas of interest. 
2.2. High resolution imagery 
With increasing resolution, the detectable objects are rapidly 
increasing in a typical landscape scene. If with a 25 m image 
resolution the mean value of a 25 Ha area represents an actually 
homogeneous surface, the increasing resolution will not deliver 
more objects. The analysis of a lake of 5 Ha and a forest of 5 Ha 
gives a proper example of the major problems related to image 
resolution and object detection with modem sensors. The 
cartographic problem of generalisation becomes an important 
issue in mapping with high resolution sensors. In digital 
generalisation, the original raw data information has not 
disappeared; the screen display is only a part of the richer image 
information stored in the internal database. 
The traditional remote sensing output product is a thematic 
layer with the same spatial resolution as the raw data. The 
surface represented by a pixel is given a class label. With high 
resolution data and object-oriented classification, the 
classification is defined by the relationship between class-object 
dependencies. In case of a water body, the surface is represented 
by a single spatial object. This is essentially different from the 
classified raster layer that represents the lake surface as N pixels 
belonging to the class ‘water’. The spatial object contains 
topology and the whole statistics from underlying band layers. 
The spatial object representing the surface of a forest stand is a 
complex mixture of crowns, forest ground, shadow areas etc. 
The class label of this spatial object can be coniferous forest, if 
the spatial analysis of the sensor data shows enough objects that 
can be related to percentage crown cover of coniferous trees. 
This decision of what is enough is exactly the moment where 
the operator defines a set of working rules that introduce the 
expert knowledge in the semi-automatic image analysis. There 
are no sharp borderlines between different classes, in this case 
between coniferous forest and mixed forest stand. Also the 
coniferous forest stands with a lot of shadow areas and almost 
no shadow areas, still belong to the same class of coniferous 
forest. So the decision rules are well defined over a set of fuzzy 
logic rules. Spectral classes, such as shadow areas should be 
separated on basis of their origin. Shadows belonging to tree 
stems show the same radiometric properties as shadow areas 
originating from terrain discontinuities. Their spatial properties 
are however different. Without additional detailed DEM 
information, the shadow areas can still be classified according 
to their origin, as the spatial composition and relative size of 
shadows of tree crowns are different than those generated by 
terrain discontinuities. This problem of radiometric classes with 
different spatial behaviour is quite common, such as water areas 
with algae, lake borders and swamp areas or inundated 
agricultural parcels. Also haze-covered areas with enough 
information of underlying landcover classes belong to this type 
of problem. For such problems, an object-based image analysis 
can offer a solution. 
2.3. Panchromatic versus multispectral data 
Typical high resolution sensors will operate in both 
panchromatic and multispectral mode simultaneously. With the 
higher spatial resolution of the panchromatic sensors and higher 
object detection capabilities, image fusion and mixed pixel 
analysis will be a serious topic. As panchromatic data is very 
sensitive to illumination conditions and highly correlated with 
the RGB bands, the grey value of the panchromatic data is often 
not used in the spectral image analysis. Second order statistics 
of the spectral bands are very powerful characteristics of this 
type of data. The significance of the second order statistics for 
spatial object discrimination (Landgrebe, 1999) is also quite 
useful in the analysis of panchromatic data. In practice, the 
second order statistics lead to more accurate classification when 
combined with multispectral data. Already the recent interest in 
this EARSeL workshop shows the importance of proper image 
fusion techniques. The biggest disadvantage of popular image 
fusion algorithms such as IHS and Brovey, is their lack of use in 
further image classification steps, especially when infrared 
channels, poorly correlating with the panchromatic data, are 
included. The integration of panchromatic data in the 
classification procedure is acceptable, if their spatial 
information is exploited. Therefore, experiments with textural 
analysis of panchromatic data gives acceptable results 
(Steinnocher, 1997). It is very important to understand the 
reason for image fusion. If the user wants a nice picture on the 
wall, standard techniques using IHS transformations fulfil the 
demand. If high resolution image information is needed to 
improve classification and understand multispectral band 
behaviour, it is much better to use an interactive display, where 
the panchromatic data layer contains the multispectral 
information ‘in the background’. This means all bands statistics 
are available upon mouse click but do not necessarily have to be 
on the display. In the object-oriented analysis, image fusion is 
no big deal. It is the automatic side-product of object building 
in the panchromatic band. Each object can be displayed 
according to user specifications. Choosing the mean value of 
the multispectral bands together with the original object layer 
gives a proper view of the segmented panchromatic layer,