Full text: Fusion of sensor data, knowledge sources and algorithms for extraction and classification of topographic objects

International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
finding their physical borderlines to other landuses like forests 
or agricultural areas. 
Texture analysis of panchromatic satellite imagery is one 
promising method in heterogeneous built-up settlement areas, 
because texture is a main feature to characterise them. Several 
methods have been developed to describe, classify and segment 
texture (Rosenfeld, 1998). Simple statistical approaches using 
mean or standard deviation do not take into account the spatial 
distribution of pixels. Several authors (Haralick and Shapiro, 
1992) suggested to use co-occurrence matrices (2 nd order 
statistics) considering the spectral as well as the spatial 
distribution of the image grey values. Based on the utilization of 
these matrices, Steinnocher (1997) developed a method in order 
to demarcate settlement structures from other landuses, which 
tries to make the calculated texture feature images independent 
from directional influences within the image. 
Sindhuber and Jansa (1998) apply the Forstner operator in order 
to distinguish between point features, significant edges and 
homogenous areas, combine the results with a modified 
multispectral classification and a thresholding and achieve a 
classification accuracy of more than 90 %. Initiated by the 
Centre of Earth Observation (CEO) within the ATLAS-project, 
a multiple step procedure for classifying urban areas applied on 
IRS-1C data (test area: Berlin) has been developed (see web-site 
of VRS GmbH, Leipzig: www.kayser-threde/vrs/vrs.htm). Two 
additional texture channels were calculated using a 
morphological filter as well as a variance filter with a 3x3 
window. Within the following segmentation process, the local 
contrast for each pixel is calculated and a watershed 
segmentation is carried out. The obtained results are integrated 
in an unsupervised classification. 
Busch (1998) describes a method by means of which built-up 
areas are extracted using the high spatial density of short linear 
features in panchromatic satellite images. Built-up and non- 
built-up areas are separated by a threshold, estimated from the 
calculated feature density in training sites stored in an existing 
GIS. Dowman (1998) reports on the ARCHANGEL project, 
which aims at developing a system for automatic registration of 
satellite data to maps. Within this project, segmentation 
algorithms have been developed, tested and integrated in the 
environment XPECT. 
3.1. Used and Tested Fusion Techniques 
Pixel-by-pixel image fusion aims at producing image products 
that are better suited for a given interpretation task. Spatial 
resolution and colour rendition play a decisive part in this. 
Without sacrificing the colour information required for object 
interpretation, high spatial resolution can be achieved by 
combining high resolution panchromatic images with 
multispectral ones. Important colour composites are real-colour 
and infrared representations (the latter particularly so for 
assessments of vegetation). 
Two different requests are usually pursued with the fusion of 
satellite imagery. On the one hand, it can be the improvement of 
the visual ability to interpret the image data. Here, the focus is 
primarily on the geometrical structural image characteristics. On 
the other hand, the goal can be the maximum preservation of the 
spectral information and thus, the maximum agreement of the 
fusion product with the original multispectral images. It was 
shown that a maximum preservation of both described image 
characteristics is only hardly possible (de Bethune et al., 1998). 
Nevertheless, methods are being tested, which ensure the 
optimal preservation of the spectral and the geometri 
cal-structural information in the synthetic fused image (Raptis et 
al., 1998). A further important aspect is achieving a good 
spectral and spatial quality of the fusion product for different 
landcover types and applications, and different possible spectral 
channels (Zhang and Albertz, 1998). This target is also pursued 
in this investigation using selected methods for fusion of IRS-C 
images. The stability of the procedures is checked by the use of 
different test areas. The produced fused images are compared 
visually and statistically regarding their spatial and spectral 
Among the various methods for fusion of multispectral with 
panchromatic satellite data described in the literature, some 
classical and some latest techniques were selected and checked 
whether they were suitable for the given objective using IRS-1C 
data. The RGB to IHS color space transformation (IHS), the 
transformation procedure by means of principal component 
analysis (PCA) and the Brovey fusion method (B) belong to the 
classical fusion methods. 'Local Mean and Variance Matching’ 
(LMM and LMVM) and high-pass filter techniques (HPF) (de 
Bethune et al., 1998) belong to the latest techniques. 
Table 1. Used and tested fusion techniques. 
Image fusion 
Fusion by means of RGB to IHS color 
space transformation 
Modified RGB to IHS color space 
transformation (Carper et al., 1990) 
Fusion bv principal component analysis 
Fusion using arithmetic combination of 
Fusion by means of local mean 
Fusion by means of local mean and 
standard deviation 
Fusion by means of high-pass filtering 
3.2. Visual Control 
The following conclusions can be drawn after a visual 
comparison of the tested fusion techniques: 
- The PC technique, and to a certain extent also the IHS 
technique, strongly depend on the spectral overlap of the 
original data. This overlap is not given with the IRS sensor 
due to the spectral band positions (NIR, Red, Green). Thus,

Note to user

Dear user,

In response to current developments in the web technology used by the Goobi viewer, the software no longer supports your browser.

Please use one of the following browsers to display this page correctly.

Thank you.