91 
ing 10 m classification results, based on Pan data, was performed on a 50-100 % and a 51-100 % binarization 
decision level ( see chapter - 3.2), XS based 20 m data indeed could be binarized on 100 m cell resolution without 
problems. The accuracies of all set’s were calculated relative to the 100 m poly grid, the 50-100 % and the 51-100 
% resized groundtruth masks. The differences of accuracy results obtained over those three reference mask differ 
not significantly ( max. ± 0.4 % ). Therefore only the comparison with the 50-100 % resized mask, which corre 
spond most to the vector forest area, is discussed here. 
100.00 
95.00 
90.00 
85.00 
80.00 
75.00 
70.00 
65.00 
60.00 
XS )€ >S >6 Pan Pan Pan Pan Pan Pan Pan Pan 
raw raw geo geo raw raw raw raw gao geo gæ gao 
ill ill 50% 51% III ill 50% 51% III ill 
50% 51% 50% 51% 
■ overall accuracy 
□ forest accuracy 
■ non - forest accuracy 
Figure - 4: Forest classification accuracy on 100 m level relative to resized original 10 m groundtruth, bina 
rized on 50 - 100 % level 
Compared to the original level accuracies the overall accuracy of the XS data sets rise up to 3%. For forest 
accuracies an improvement of only 1 % can be detected. As seen in figure - 2, the used 100 m groundtruth differs 
1 % relative to the vector mask. The better results are therefore not significant. Pan data classification overall 
accuracy and forest specific accuracy on 100 m level is 1% less than on original 10 m level. Analog to XS data, 
this single percent can be explained by accuracy groundtruth difference of 10 m and 100 m level and is therefore 
not significant. 
8 - CONCLUSIONS AND OUTLOOK 
Groundtruth representation and preprocessing are the main topics to achieve good classification results. Conver 
sion of vector based groundtruth data to raster representation must be done with respect to the object boundary 
variations scale. Even if a low resolution grid cell size is needed, it is better to convert vector data first into an 
object resolution common high resolution grid. Then the cell size can be magnified area proportionally into the 
final resolution. Resizing binary raster data on area base can produce in some cases decision problems when the 
goal cell covers an input area occupied by exactly 50 % binary information. 
For a forest / non-forest classification SPOT XS or SPOT Pan (combined with slope) are suited equally 
good. Accuracies of classified forest of around 80 % were reached. Correction of illumination using the Minnaert 
method is not suitable for Pan data. On XS data however the illumination correction gives 10 % better results. 
Applying the Minnaert correction on raw XS geometry is recommended before geocoding the data. In general, 
classifying on raw geometry will give slightly better results. Classification comparison of original satellite data 
resolution compared with 100 m resolution shows no significant difference. 
The next step in our work will be to involve Landsat-TM data into the same classification process. The 
purpose is to prove, that the lower geometric, but better radiometric resolution of Landsat-TM, outperforms the 
‘geometric’ systems of SPOT. Research is also done on atmospheric corrections and object reflection property 
studies, to improve the classification accuracy. 
9 - ACKNOWLEDGMENT 
This work is part of the SRSFM-Project (Swiss Remote Sensing Forest Mapping Project) at the Remote Sensing 
Laboratories of the University of Zurich, which started in 1988 and is supported by the Swiss Government, 
UNEP/GRID and ESA. DEM data courtesy: Swiss Federal Office of Topography January 11, 1994.