90 
tions often used. Our aim however was always to develop operational classification procedures with rare use of 
auxiliary data. Digital elevation models are always necessary, performing geometric corrections. It’s derivatives 
can be used in the process. To discriminate the water we used the slope information. In our rugged testsite only a 
few non-water pixels are really horizontal. Merging multiresolution data was not suitable for our purpose to clas 
sify on raw data too. 
5.1. Classification Accuracy 
The classifications were validated in a geometric sense comparing pixel per pixel with the appropriate 
groundtruth. Classifications in raw geometry were additionaly geocoded and again compared with the ground 
truth in Swiss coordinates. Overall accuracy , the number of correct classified pixels divided by the total pixels, is 
a standard measure to verify all classes at the same time. Class specific accuracy , the number of correct classified 
pixels of a class divided by it’s total number of pixels, is used to measure the quality of the forest classification. In 
a forest / non-forest classification we need to pay attention to both accuracies at the same time. To illustrate the 
problem of overall accuracy we assume the problem of classifying a class ‘x’, which covers only a few pixels in 
an image. Classifying the whole image as not ‘x’ results in a good overall accuracy, but a catastrophic ‘x’ specific 
accuracy. According to the groundtruth, our testsite is covered by 25 % water and 29 % forest. Overall accuracy 
will therefore significantly change when missclassifying one class extremely. The results of all different combina 
tions of classification comparisons are shown below. 
I 
raw raw geo raw raw geo raw raw geo raw raw geo 
geo i 11 geo i 11 geo ¡ 11 geo i 11 
ill ill 
■ overall accuracy 
□ forest accuracy 
■ non - forest accuracy 
Figure - 3: 
Forest classification accuracy on original resolution levels. ( combination of raw & geo in name 
means classification in raw and comparison in geocoded geometry) 
6 - RESULTS ON ORIGINAL LEVEL 
First of all, the classification accuracies data set’s, which were not illumination corrected, compared on geocoded 
level must always be identical, on the one hand for classifying raw and on the other hand for classifying geocoded 
data. The reasons for this are given in the nearest neighbour based geocoding algorithm, which does not touch 
pixel values and who overlays identical groundtruth pixels on each image pixel of both geometries. XS was tested 
against the forest percentage mask, resized from the 10 m polygrid generated mask ( see chapter - 3.2 ). Pan data 
was tested against the polygrid generated 10 m mask. Best overall accuracy is obtained using Pan data ( combined 
with slope). But this is a result of enormously increasing specific accuracy of non-forest paired with the forest / 
non-forest relation of 1 to 3. Over all set’s there is a significant trend receiving better accuracies classifying on 
data in raw geometry, especially visible on XS data. Illumination correction, applied specifically on forest, does 
rise the accuracy by over 11%. On Pan data the results are worse, owing the high digital numbers value range due 
to near infrared overlap ( see chapter - 4.2 ). The best way, classifying forest with SPOT is obviously to perform 
an illumination correction on raw XS data, classifying and geocoding the result. But as seen in chapter - 3.3 the 
forest percentage mask differs up to 5 % from vector groundtruth. Instead of reaching 81. 6% forest accuracy we 
must correct the result down to 77%. Using Pan data will give the same accuracies under those circumstances. 
7 - RESULTS ON A 100 M LEVEL 
The final purpose of our studies should result in a forest / non-forest map of Switzerland in 100 m raster grid cell 
size. To test the capabilities of SPOT data for this purpose, all classification results were resized, using the per 
centage method, to 100 m raster size and compared with the different 100 m forest groundtruth data set’s. Resiz