al., 1996). Further details and presentation of the resulting forest map and its accuracy will be 
reported elsewhere. 
The unsupervised procedure involved principle component analysis of six months of regional 
NDVI and Ts data to reduce inherent data redundancies (e.g. temporal correlation) and to 
remove residual noise. The most significant principal components were then clustered using 
ISODATA (Ball and Hall, 1965) and each cluster label assigned a land cover association. 
This was performed by examination of cluster labels co-located with pre-classified MARS test 
images. The clustered AVHRR data (1.1km pixels) were resampled using the nearest 
neighbour resampling scheme to the same pixel resolution as the pre-classified MARS test 
images (20m pixels) permitting a one to one pixel mapping between the two scales of data. 
The unassigned resampled clusters were compared with the pre-classified MARS classes by 
examination of the histograms of the two data sets. Cluster labels were assigned a land cover 
association defined by the majority MARS class falling within each unique cluster label. For 
most test images it was necessary to perform this procedure iteratively by agglomerating the 
MARS test image classes together. Test images falling within the same region were treated 
independently in this manner and then the land cover associations compared to ensure within 
region consistency. The associations were then applied to the rest of the clustered AVHRR 
data lying in the region. Urban classes were labelled directly by examination of the Digital 
Chart of the World (DCW) database. 
PRELIMINARY RESULTS 
Preliminary examination of the land cover results revealed a strong spatial correspondence 
between the cluster patterns and the urban areas defined in the DCW, giving some confidence 
in the unsupervised classification procedure. A further indicator of the consistency of the 
results was that spatially continuous classes occurred across region boundaries even though 
the AVHRR data were classified independently in each region. 
Figure 2 illustrates the relationship between the percentage cropland class surface area derived 
from the land cover map and EUROSTAT cropland percentage surface area statistics for 
NUTS-2 regions in Belgium, France, Germany and Holland. Only NUTS-2 regions with less 
than 10% cloud coverage were examined as cloudy AVHRR pixels could not be classified. A 
total of 27 NUTS-2 regions were examined. The NUTS-2 region boundaries do not 
correspond to the ecosystem regions boundaries used to perform the classification. The simple