International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
162 
these two techniques do not supply stable results for the 
fusion of IRS image data. 
- Above all, HPF, LMM, LMVM and WTA were very stable 
(i.e. independent of the test areas). 
- The best results regarding color reproduction were with the 
LMM and the LMVM techniques using small filter sizes 
(less than 15 x 15 pixels). 
- The best results regarding the reproduction of the spatial 
information were obtained with the HPF and the PCA 
techniques. 
3.3. Statistical Control 
To examine the preservation of the spectral information, first 
the bandwise correlation coefficients between the fused images 
and the LISS images were checked for all test areas. The 
calculated coefficients confirm the results obtained by visual 
comparison, namely the fact that the local working methods 
(LMM, LMVM, and HPF) are more independent of the chosen 
test area and the spectral characteristics of the surfaces. These 
three methods keep the spectral information of the original data 
very precisely (up to 99% with small filter sizes). An 
enlargement of the filter size reduces the correlation between 
original data and fused images. Further tests showed that the 
correlation coefficients increase only very slightly (3rd decimal 
place) with filter sizes below 7x7 pixels. Filter sizes of 25 x 25 
pixels and more largely reduce the correlation of the near 
infrared, particularly with the LMM and LMVM techniques. 
These results are confirmed by using additional test criteria as 
the bandwise standard deviations and the percentage of small 
grey value differences between the LISS and the individual 
fused images. The percentage of small differences was highest 
with the LMVM technique. This method with a window size of 
5x5 pixels was the best at preserving the spectral 
characteristics. 
It should be noted that, by applying the HPF technique, the 
change of the pixel values is similar in all three bands because 
this method is only a bandwise addition of the high frequencies 
of the panchromatic image. The comparison also showed, that 
HPF causes an influence of the original spectral values that is 
stable and similar in all test areas, but quite large. 
In summary, fusion methods using local filters provide more 
stable results regarding the preservation of the geometrical 
characteristics (HPF), as well as the spectral characteristics 
(LMM and LMVM). These results apply to the given spectral 
characteristics of the IRS-1C bands (green, red and near 
infrared). 
4. CLASSIFICATION APPROACH 
4.1. Multispectral Classification 
In this first classification phase a conventional, multispectral 
classification was applied to the fused IRS data. The produced 
intermediate result provides a set of spectrally rather 
homogeneous landcover classes and thus reliable for 
identification of landcover classes, like water or forest. On the 
other hand, it represents a rough classification and thus a basic 
structure for these classes, while classes with high spectral 
overlap are difficult or impossible to separate. To improve the 
classification quality, it is necessary to include textural and 
morphological image features of the panchromatic data. 
A multistep, hierarchical procedure was applied, developed in 
earlier projects for classification of both satellite-based and 
airborne, multispectral scanner data (Netzband, 1998). In a first 
step, an unsupervised classification (i.e. without signature 
analysis by the analyst) is executed, which supplies 15 classes. 
These classes have to be assigned to landuse types by 
interactive, visual check and postprocessing or, if necessary, 
aggregated. Furthermore, it is important to separate individual 
classes that are spectrally unique. The class separation was 
performed by a multispectral, supervised classification, in which 
each identified class was "extracted" by masking it in the 
intermediate result, in order to exclude it from the following 
classification steps. For the classification, a parallelepiped 
classifier was used. In this procedure, pixels, which do not 
belong to clusters of the spectral signatures, are not classified, 
and pixels in the overlap area of two clusters are classified 
according to the Maximum Likelihood method. The resulting 
classes can be overlaid as masks on the final result image and 
can be stored as independent layers. 
In detail the following classes could be separated: 
- forest (differentiated according to leaves and coniferous 
forest surfaces), 
- fruit-tree covered agricultural surfaces as well as grassland 
and meadow surfaces in the inner urban areas, 
- vegetationless agricultural surfaces, 
- water surfaces, 
- open soil such as sand pits, building sites, fallow land etc. 
In the course of the investigations, it appeared meaningful to 
classify single predominantly large-area classes, like forest and 
agricultural surfaces, from the original unfused LISS data (with 
23 m resolution) and overlay the results on the classification 
from the fused data. The aim was to increase the classification 
accuracy for these classes (since there was no influence due to 
the stripe noise of the panchromatic data) and to eliminate 
wrongly classified pixels of the fused data inside these large 
areas. 
The classification quality has not been examined in detail yet. 
After a rough estimation it lies between 80 % and 85 % as was 
expected. The classes "built-up areas" and "sealed open areas" 
posed particular problems for the multispectral classification, 
since they are spectrally not well separable from bare soil, as it 
occurs in agricultural but also urban areas (fallow land, dry 
meadows and grassland). It is therefore necessary, to include 
non-spectral morphologic image characteristics into the 
classification procedure.