85 
COMPARISON OF FOREST / NON-FOREST CLASSIFICATION IN RAW AND 
GEOMETRICALLY CORRECTED IMAGES OF AN ILLUMINATION 
CORRECTED SPOT DATA SET 
TOBIAS W. KELLENBERGER, STEFAN SANDMEIER, KLAUS I. ITTEN 
Remote Sensing Laboratories 
Department of Geography 
University of Zurich-Irchel, Winterthurerstr. 190, CH-8057 Zurich, Switzerland 
Phone:+41 1 257’52’46/FAX:+41 1 362’52’27 
ABSTRACT: 
The objective is to test the usefulness of SPOT data for a digital forest / non-forest classification of Switzerland in 
a 100 m raster scale. Preprocessing of binary groundtruth data, e. g. forest boundaries, is examined. Specially the 
influence of vector to raster representation and scaling problems between different raster resolutions is analysed 
towards area and shape change of forest. Different successions of SPOT data correction steps for geometry and 
illumination effects are tested with regard to the forest / non-forest classification accuracy on original satellite data 
resolution and hectare scale. 
KEY WORDS: Forest Classification, SPOT, Illumination, Resampling, Groundtruth, Scale Change 
1 - INTRODUCTION 
In 1988, the Department of Geography of the University of Zurich was entrusted by the United Nations Environ 
mental Programme (UNEP), to establish a digital forest / non-forest map of Switzerland, using optical, multispec- 
tral satellite data. While developing the procedures with different satellite data sets maximum use of GIS 
technology was made, since it was esteemed essential, that the methodology developed should be applicable also 
to other temperate zone forests. In a first step, an approach using NOAA/AVHRR data, was made to gain a forest 
/ non-forest classification over whole Switzerland in a 1km grid cell size (Itten et al., 1991a; Itten et al., 1991b). 
Therein a forest / non-forest specific classification accuracy of over 80.6 % was achieved. This work had been fin 
ished in early 1991. The purpose of the actual second step is to obtain a 100 m forest / non-forest map with addi 
tional forest stand differentiation on a 100 m grid base. Basic research on Landsat-TM data, was done fore 
geometric (Itten et al., 1993) and radiometric (Itten et al„ 1992; Meyer et al. 1993) corrections. 
In this study SPOT data is used to test at one hand the suitability of high geometric resolution satellite data 
and at the other hand the impact of different successions of geometric and radiometric correction steps on forest / 
non-forest classification accuracy. This was done on original SPOT data resolution and hectare scale using excel 
lent groundtruth data. No atmospheric correction, based on atmospheric codes, was applied to this study. Prob 
lems dealing with binary groundtruth preprocessing and scaling are intensely discussed and analysed . 
2 - BASE OF THE STUDY 
2.1. Testsite 
The selected testsite ‘Beckenried’ is situated in the mountainous pre-alps of the Canton Nidwalden in Central 
Switzerland. The site covers nearly the area of the Swiss topographic map sheet, 1:25’000, sheet 1171 ‘Becken 
ried’. For the study, some smaller section was selected, which can be described completely by an undivided 
number of pixels in the resolution of the input satellite data sets (i. e. 10 m and 20 m for SPOT and 25 m for fur 
ther involved Landsat-TM data). The testsite is dominated by a section of the Lake of Lucerne, covering 25 % of 
the area; a heterogeneous mixture of conifers and deciduous trees leaving only small patches of clearly discerni 
ble stands, covering over 29 % of the area, and mixed rural meadows with small villages at borders of the lake. 
Terrain elevation range from 434 to 2404 m a. s. 1. with pronounced steep slopes.