86 
2.2. SPOT Data Set 
Due to the fact, that the foliage of the various forest types is only fully developed in summer time and due to the 
negative sun angle influence on cast shadows and illumination in fall, winter and spring, satellite data of midsum 
mer was selected. The base dataset consists of a SPOT - 2 scene (054/255) of 30 June 1991, in multispectral mode 
(band XS1 to XS3) and panchromatic mode, both processed at level lb. The scene is nearly cloudffee, apart from 
a airplane induced contrail, crossing the scene in the southern part from east to west. 
The SPOT data set suffers from relevant striping, induced by detector intercalibration problems. These 
effects have been corrected, using a modified morphological filtering by Banon and Barrera ( Banon et al., 1989), 
as described by Ehrler (Ehrler, 1993). 
2.3. Digital Elevation Model 
A digital elevation model from the Swiss Federal Office of Topography with a resolution of 25m in x and у and 
0.1 m in z was available for the whole testsite. The model is based on the topographic map sheet 1:25’000, with 
contour intervals of 20 m. The model was resampled according to the SPOT data resolution to 10 m and 20 m x 
and у resolution, using a ‘bilinear’ interpolation method. From this model slope, aspect and illumination as well 
as a mask for cast shadow were generated. 
The illumination, defined as the cosine of the solar incidence angle, thus representing the portion of the 
direct solar radiation hitting a pixel, is basic to any radiometric terrain correction. Illumination depends directly 
from the relative orientation of the pixel towards the sun’s actual position. 
2.4. Groundtruth Masks 
For a forest / non-forest classification comparable groundtruth, the green plate from the topographic map sheet 
1:25’000, ‘Beckenried’, was scanned using an Optronics drum scanner and vectorized on a Intergraph system. 
The vector data was transferred to an ARC/INFO system. 
Water boundaries were digitized manually directly from the topographic map sheets and added to the for 
est coverage on the ARC/INFO system. They will be needed in the classification process to separate forest and 
water patches. 
3 - GROUNDTRUTH PREPROCESSING 
To validate any satellite data classification, accurate groundtruth data sets must be available in the corresponding 
grid cell size. Unfortunately groundtruth often is in vector format or must be digitized into vector format from 
topographic map sheets, aerial photographs etc. A vector to raster conversion must then be added. In the case of 
available groundtruth in pixel format, the grid resolution sometimes differs from the satellite data. The following 
experiments were done to evaluate the effects of vector to raster conversions and grid resolution changes. The 
results will be used to assess classification comparisons to the groundtruth. 
3.1. Vector Raster Conversion 
As mentioned above, the forest / non-forest mask was generated in vector file format on an ARC/INFO system. 
Forest boundaries are represented as high accuracy polygon chains on cm accuracy level (see figure - 1). For test 
ing the forest / non-forest classification a raster representation of the forest groundtruth in satellite resolution is 
needed. But the transition between vector and raster results in decrease of information, such as different forest 
shapes and changing total area. 
For testing purposes, the forest and water coverage groundtruth in vector format was converted into 10 m, 
20 m, 25 m and 100 m grids, using the ARC/INFO task ‘polygrid’ of version 6. 1. Input and output total area cov 
ers identically the selected testsite ‘Beckenried’. Area difference is one of several possibilities to describe the 
change in raster vector conversion of thematic data. The bold black line in figure - 2 (top) shows the change of the 
forest area in relation to the original vector coverage using the polygrid conversion. It is remarkable, that the 10 m 
raster forest coverage only differs by 0.1 % relative to the vector data. But in representation of larger cell sizes the 
total forest area increases up to 106% of the vector area. For the interpretation of the forest coverage change on 
different raster resolution levels the following effects must be considered. The fine structured forest boundaries 
with variance on a 5 to 10 m level, as seen in aerial photographs, can be represented quite good by the 10 m raster. 
With increasing raster cell size indeed these fine structures are lost. The conclusion, that at most a 10 m resolution