87 
grid size is suitable is only valid in relation to the specific forest shape. Contrary to the forest, the water area 
change, using ‘polygrid’, is less than 0.04 percent. Water is a very compact element in our testsite with boundaries 
of low structure. 
Figure - 1: Part of the forest groundtruth in original vector representation (left), derivated 10 m( centre) and 
100 m( right) raster representation using ‘polygrid’. 
3.1.1. Compactness 
Area change detection is only significant for the total groundtruth coverage. The shape of the groundtruth is as 
well influenced by the differing character of vector and grid representations. Small groundtruth patches with fine 
tailored boundaries are easy to describe in vector, but difficult or impossible in raster representation. A morpho 
logic descriptor to detect significant changes is compactness (or complexity, circularity ) by Castleman (1979), as 
described in Meyer (1992). Compactness C gives a good relationship between boundary and area in the following 
manner 
( 1 ) 
where P 2 is the square of the perimeter ( m 2 ) and where A the area ( m 2 ). 
The minimal value of compactness is 4it in the case of a circle. A regular pixel will have a compactness of 
16. The heavy black line in figure - 2 (bottom) shows the compactness of vector forest data and the derivated grid 
forest data on different levels. The enormous compactness values are typical for the forest shape in the testsite. 
The analogue values for the water compactness are about 300 times smaller. Inverse parallel to the area change, a 
significant higher compactness on 10 m resolution can be detected, due to the raster representation of small 
patches. With increasing grid cell size, the compactness goes down. Less (but bigger) pixels have smaller bound 
aries and as conclusion smaller compactness values. 
3.2. Scale Change on Raster Data 
Scaling groundtruth in raster representation, such as forest / non-forest, to satellite data resolution cannot be done 
using standard algorithms like ‘nearest neighbour’ or ‘cubic convolution’ sampling, as they do not pay attention 
to the discrete thematic data and as they only grab data in local regions around the center pixel. Resizing or resa 
mpling of thematic groundtruth must be performed by statistical interpretation of the area , covered by a goal 
pixel, on the starting coverage. This can only be done treating the pixels as areas and combining the data, if neces 
sary, over fractions of input pixels. An algorithm was developed to resize thematic grid information in a new 
scaled grid where each pixel contains the mode, median or the fraction of a selected information class of the input 
data set covered by the output pixel. 
Resizing a binary thematic groundtruth from, for example, 10 m grid resolution to 20 m resolution, select 
ing the covered area percentage as output option maintains the information of the input data and is therefore 
invertible. To obtain the corresponding binary groundtruth from the just generated 20 m level, there is a need of