88 
splitting the 0 - 100 % coverage range in two classes (empty or occupied). Induced by the incremental discrete 
integer data representation in most raster image processing systems, the splitting can only be done, if the 50 % 
coverage level of a pixel does not appear. Otherwise those pixels must be added either to the empty class or to the 
occupied class. It is impossible to resize exactly a 10 m binary forest groundtruth into 20 m or 100 m. To show the 
resulting effects the 10 m forest groundtruth, derived from the vector base, was resized to different output levels. 
For each level the 50-100 % (and 51-100 % , if a binarization was not possible) forest coverage classes were cal 
culated and compared to the vector forest area ( figure - 2 top). Additionally the compactness values are reported 
below. 
resized resolution In m 
resized resolution In m 
THEMATIC 
RESAMPLING 
RESAMPLING 
RESAMPLING 
50% UP LEVEL 
51%UP LEVEL 
EXACT 
LEVEL 
Figure - 2: Area (top) and compactness (bottom) comparison of resized forest groundtruth relative to original 
vector data set. (‘POLYGRID’: vector to specific raster size conversion. ‘THEMATIC’: 10 m 
raster to specific raster size conversion) 
In cases where a 50 % binarization level was possible, the forest groundtruth differs only by 1 % relative to 
vector data. Otherwise a significant area difference in 10 m to 20 m resizing of over 5% is detected. This effect 
must be traced back to the natural 5-10 m boundary undulation of forest in our testsite, which can be represented 
quite good by 10 m pixels. With raising pixel size all curves are nearly congruent. 
3.3. Conclusions for Thematic Scale Change 
Correct transformation of binary groundtruth data, as forest / non-forest data, from vector format into raster repre 
sentation, using ARC/INFO’s polygrid task, depends on object shapes. Only if the selected output grid resolution 
is adequate to the objects boundary scale variations (10 m for forest in testsite ‘Beckenried’) almost no area differ 
ences have to be expected. Larger grid cell sizes may lead to differences of ± 6 %, which will reduce the 
groundtruth comparison capabilities to classifications significantly.