51 
lecessity to DATA 
.and cover 
integration 
illy require 
:r labels to 
Hie cost of 
. increases, 
int of user 
)re reliable 
' European 
d using an 
d using a 
68 relatively cloud-free AVHRR mosaics of Europe were selected from the data archive of the 
SAI Monitoring Agriculture by Remote Sensing (MARS) project (Meyer-Roux and Vossen, 
1993). The mosaics were acquired from March to October 1993 and cover a geographical 
area from the Portuguese coast to central Crete and from northern Algeria to southern 
Sweden. Each mosaic was made from between 3-6 AVHRR-LAC (1.1km pixel) afternoon 
pass images which had been atmospherically corrected and then thresholded to remove 
missing, sea and cloud pixels (Vowles, 1991). The 68 AVHRR mosaics were reduced into 
eight monthly maximum value composites to lessen the amount of data to be processed and to 
reduce undesirable atmospheric effects (Roy, 1996). NDVI values were extracted from the 
AVHRR composites using the red and near infrared pixel values (Curran, 1983) and Ts 
values were extracted using the thermal infrared pixel values (Price, 1984). 
fication at 
tion of the 
1 Belward, 
The AVHRR data were stratified into 13 ecosystem regions and 82 homogeneous forest 
regions for production of the land cover and forest cover maps respectively. In both cases the 
1 and near 
difference 
al activity, 
AVHRR data were classified independently on a regional basis. The ecosystem regions were 
defined by a recent European Commission study at a scale of 1:2.5 million using topographic, 
soil and climate variables (European Commission, 1995). The homogeneous forest regions 
n biomass 
, statistical 
land cover 
oration of 
useful for 
ape, 1989; 
rising land 
id recently 
ambin and 
were defined by stratification of the ecosystem regions using six forest variables (Kennedy et 
al., 1995; European Commission, 1995). Large inland water bodies were also defmed. The 
regions were registered with the AVHRR data and stored in a GIS. 
Class labelling was performed to define land cover classes in the unsupervised classification 
procedure using MARS pre-classified test images selected from the SAI data archives. The 
test images were classified conventionally by classification of SPOT and LANDSAT TM 
images using extensive field data (Meyer-Roux and Vossen, 1993). Each test image has a 
pixel dimension of 20m and is composed of 2000 by 2000 pixels and typically defines up to 
12 agricultural classes and a variable number of masks for other land cover classes. 
A georeferenced database of forest and non-forest pixels was used in the supervised 
classification procedure for forest/non-forest class training. The database was defined by 
morphologically filtering (Serra, 1986) a pre-existing digital forest map. The forest map was 
produced with an overall classification accuracy of 82.5% by unsupervised classification of 
72 AVHRR-LAC images selected from 1989 to 1992 (Hausler et al., 1993). The filtering