asures 
Mean, 
). The 
;ted for 
1dow 
0). 
th PCI- 
matrix 
or the 
OT-XS 
es the 
for the 
ou find 
texture 
ı and 
vely. 
und 30 
2 pixels 
k were 
Ision of 
ta are 
97) the 
iometric 
clearly 
ige, the 
Adaptive Image Fusion and the Local Mean Matching 
procedure is developed to maintain the radiometric 
characteristic of the input data. 
In the data fusion procedures only the SPOT-XS'95 and 
the IRS-1C '97 data were used according to their similar 
phenological characteristics. 
BROVEY-Transformation 
The Brovey transformation ( in POHL 1996 ) was 
developed by Bob Brovey combining spatially high 
resolution data sets with multispectral data. 
DN, d 
Die TN DN, DN. DN res (1) 
POHL (1996) 
The advantage of Brovey-transform is based in the very 
naturally looking image and a better fusion result in 
comparison to IHS-transform ( Stuart Nixon, 1995 ). 
Additionally to the Brovey transformation the Adaptive 
Image Fusion STEINNOCHER (1997) procedure was 
applied to the SPOT XS- and the IRS-1C data set . 
Adaptive Image Fusion (AIF) 
Using a moving window the AIF will detect spectral class 
borders in the subset based on the high resolution data of 
the PAN - image. For all pixels of the same spectral class 
inside the border a mean value will be computed and 
stored in the output image. 
The procedure of AIF is described in the IEEE 
International Geoscience and Remote Sensing 
Symposium Proceedings 1997, 
Local Mean Matching 
The Local Mean Matching procedure was developed by 
De Bethune, 1997 to maintain the spectral characteristic 
of multispectral input data. 
= 
3E * i,j(w,h) 
1, j(w,h) 
Fa represents the fusion value at the position i,j, while 
H; is the variable for the high resolution and Lu, for the 
low resolution data at position i,j. The ratio is calculating 
the mean value of a window of the multispectral image 
divided through the window mean value of the high 
resolution image. The window size in this procedure was 
fixed to 3x3 pixels. 
The three fusion images were classified using the same 
training data set like in the layerstack classification. 
Classification 
For the classification of the data sets the two common 
used classification procedures of 
Maximum Likelihood (ML) 
Minimum Distance (MD) 
were used, because these procedures are standards 
available to most end-users. 
Accuracy Assessment 
The selection of reference samples in the investigation 
area was based on random sample points. In contrast to 
the stratified random sampling, which is preferred to 
analyse the accuracy of the classes, the random 
sampling in this project was use to evaluate the different 
classification results in the term of total classification 
accuracy. 
The calculation of the random sample points was carried 
out by the use of ERDAS IMAGINE accuracy assessment 
tool. The geodetic coordinates of the random points were 
exported to a FORTRAN-Program, which calculated 
squared areas around this points. After that all reference 
areas were overlaid on the topographic map and printed 
out as detailed map in the scale of 1: 5.000. 
Field Work 
The discovery of the sample areas in the field was planed 
by the use of a GPS-System and the topographic map. 
For the GPS work two MAGELLAN-GPS-Systems from 
the department of Forest Biometrics were available. 
These systems are working without the online correction 
of the position. The position can be corrected using a 
postprocessing procedure. 
Already during the first measures under the crown closure 
the sampling of 4 GPS-satellites for the calculation of 
position was impossible or needs more than 30 Minutes 
per sample point. This was a intolerable handicap of the 
accuracy assessment. 
The consequence of this handicap was the direct use of 
detailed maps for discovery of the sample areas in the 
field. Arrived at the sample areas the predesigned 
admission maps were completed with vector graphics for 
landuse classes and a special code number. 
Reference data into GIS 
Based on the geodetic random sample points the 
integration of the reference data maps was carried out. 
The vector graphics were digitised, attributed with the 
landuse code and stored as a coverage. 
The classified images from all classification procedures 
were converted from raster to vector coverages using the 
command "GRIDPOLY" in ARC/INFO. The goal of this 
process was to overlay the classification on the reference 
cover to calculate the classification accuracy. 
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 373