89 
Resizing raster based binary forest groundtruth, outgoing from a adequate resolution, which represents the 
forest shape well, will always be better than direct rastering from vector data. Even if a binarization level of the 
percentage coverage is not possible. Nevertheless area differences by small scale changes of ±5 % are detected. 
4 - PREPROCESSING 
4.1. Geometric Preprocessing 
The SPOT data set used in this study has been geocoded to the rectangular coordinate system of the Swiss topo 
graphic maps. The rectification also included geometric correction of relief displacement due to variation in ter 
rain elevation (Itten et al., 1993). To keep the intensity values of the scene unchanged, a ‘nearest neighbour’ 
resampling algorithm instead of often used ‘cubic convolution’ technique was applied. The geometric correction 
is a reversible process. Therefore all the groundtruth reference data in Swiss coordinates, forest vs. non-forest, the 
water boundary, digital elevation model and its derivatives illumination and casthadow have been transformed 
back into the SPOT raw geometry, to perform classification and radiometric correction on raw data too. The geoc 
oding process is based, beneath other steps, on a backward transformation, which does not touch the raw data pix 
els in geocoding. 
4.2. Radiometric Preprocessing 
The intensities measured by a sensor are influenced by sensor induced and scene related effects. Sensor effects are 
dealing with technical problems, such as calibration of detectors (see chapter - 2.2), filtering, system stability etc. 
Scene related effects include reflectance properties of the object, topographic influences, illumination properties, 
atmospheric effects etc. 
In a first step atmospheric influences and adjacency effects are neglected ( with regard to the airplane 
watervapour contrail a negative influence in the appropriate, small region must be expected). Under those 
assumptions we can state, that in the visible and infrared bands the direct sun radiation is the only illuminating 
factor. Treating the forest as a Lambertian reflector (a rather brutal assumption), a statistic-empirical Minnaert 
slope-aspect correction (Meyer et al., 1993) was applied to the data set. The Minnaert illumination correction is a 
common cosine correction of sun incidence angle, modified by the Minnaert constant k. The parameter k was 
empirically determined by a linear regression, with optimization for each band onto the forest, as represented in 
groundtruth. The correction was done on SPOT data in raw or geocoded geometry. An ideal slope-aspect correc 
tion will remove topographically induced illumination variations. As visible consequence is the loss of the relief 
effect in the scene. For later discussion it should be mentioned, that the Minnaert constant is correlated to the 
dynamic range of the band applied. SPOT XS bands 1 and 2 will be less corrected than band 3. SPOT’s panchro 
matic band, with its overlapping into the near infrared, is corrected less than XS band 3, but significantly more 
than the visible XS bands. 
5 - CLASSIFICATION 
One of the main topics of this study is to detect changes in forest / non-forest classification accuracies using SPOT 
data sets with the same preprocessing applied, but in different successions. Beside the two SPOT data set’s ( Pan 
chromatic and 3-band XS), six further set’s were generated applying geometric (chapter - 4.1) and/or illumination 
corrections (chapter - 4.2 ) in all possible combinations: 
set 1 Panchromatic, raw geometry : Pan raw 
set 2 Panchromatic, raw geometry, illumination correction applied: Pan illu 
set 3 Panchromatic, geocoded: Pan geo 
set 4 Panchromatic, geocoded, illumination correction applied: Pan geo illu 
set 5 3-band XS, raw geometry : XS raw 
set 6 3-band XS, raw geometry, illumination correction applied: XS illu 
set 7 3-band XS, geocoded: XS geo 
set 8 3-band XS, geocoded, illumination correction applied: XS geo illu 
On each set, a classification of forested and water covered areas was performed, using a standard parallel 
eppiped classifier (PPD). To obtain the input classification parameters, the forest and water groundtruth was 
applied as training areas, neglecting sites influenced by cast shadows. Classification on SPOT XS data set’s was 
done using band 1, to separate forest, and band 3 to classify water bodies. There is no way to separate water and 
forest on single band Pan data. Using the water mask or combining 20 m XS with 10 m Pan are some of the solu