labelling. Cluster labels are then applied to the rest of the 
AVHRR data lying in the region. In many regions this 
procedure is performed iteratively using progressively 
more agglomerated classes defined by the higher spatial 
resolution image classifications. Validation of the 
classified regions is finally performed by comparison of 
the classification results with regional surface area land 
cover summary statistics. 
GIS Regions 
  
  
Single Region 
AVHRR data 
   
     
Unsupervised 
Clustering 
Class 
SPOT. 
/TM data Labelling 
EUROSTAT NUTS 
STATISTICS 
    
      
Validation 
  
* 
à 
^ 
\ 
i 
? 
\ MAP PRODUCT ; 
Noe] sus 339, M mmn ’ 
x 2e ‘ 
à * 
Saale ; Landcover ; trees 
relate aéré J 
  
Figure 1: Flowchart describing the regional approach with 
class labelling and validation procedures. 
DATA 
68 relatively cloud-free AVHRR mosaics have been 
selected from the data archive of the SAI-MARS project 
(Roy, 1996). These images cover a geographical area from 
the Portuguese coast to central Crete and from northern 
Algeria to southern Sweden, and were acquired over the 
main growing season from March to October 1993. Each 
mosaic is made from between 3-6 AVHRR-LAC (1.1km 
pixel) afternoon pass images. Missing data, water and 
clouds are thresholded out. The 68 AVHRR mosaics are 
reduced to eight monthly maximum value composites to 
356 
lessen the amount of data to be processed and to reduce 
undesirable atmospheric effects (Holben, 1986). NDVI 
values are extracted from the AVHRR composites using the 
red and near infrared pixel values (Curran, 1983) and Ts 
values are extracted using the thermal infrared pixel values 
(Price, 1984). In total six images, each composed of 2779 
by 2343 pixels with NDVI and Ts counts defined over a 10 
bit range are derived. 
The AVHRR data are stratified into 13 ecosystem regions 
and classified independently on a regional basis. The 
ecosystem were defined by a recent European Commission 
study at a scale of 1:2.5million using topographic, soil and 
climate variables (Kennedy et. al., 1995). Large inland 
water bodies are also defined and are used to assign water 
class labels. While the ecosystem regions serve as a basis 
for stratification, a further sub-division of some 
geographically large areas was found to be necessary after 
visual inspection of the clustering results. 
Class labelling is performed using MARS pre-classified 
high spatial resolution test images selected from the SAI 
data archives. In addition, urban classes are sited across 
the entire image using the Digital Chart of the World 
database (DCW) (ESRI, 1993). The DCW database 
consists of spatial and aspatial data that can be accessed, 
queried and displayed with a GIS. The populated place 
layer of the DCW depicts the urbanised areas that can be 
represented as polygons at 1:1Million scale. 
Validation of the classified regions is performed by 
comparison of the classification results with landcover 
summary statistics defined by the Statistical Office of the 
European Union (EUROSTAT). The Nomenclature of 
Territorial Units for Statistics (NUTS) has been established 
to provide a uniform breakdown of territorial units for the 
production of regional statistics. The NUTS regions use a 
common landuse nomenclature and are hierarchically 
defined at different scales based on the institutional 
divisions in force in the Member States. 
RESULTS 
A strong spatial correspondence between the cluster 
patterns and the urban area as defined in the DCW was 
generally observed across Europe, and led to some 
confidence in the assignment of the urban class labels. 
Furthermore 10 MARS high resolution test images across 
four independent regions were examined and showed a high 
degree of surface area correlation between the labelled 
AVHRR land cover classes and the MARS test images 
(e.g. cropland 0.61< r < 0.99). However, the results must 
be regarded with some caution as area estimates made by 
pixel counting over large regions are biased when there are 
mixed pixels and when the classification accuracy is not 
high (Czaplewski, 1992). A further indicator of the 
consistency of the results is the fact that spatially 
continuous AVHRR class labels occur across ecosystem 
region boundaries even though the AVHRR data have been 
classified independently in each region. Figure 2 and 3 
illustrates a strong relationship between the AVHRR 
cropland and forest classes and the validation statistics 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996 
  
der 
val 
we 
les: 
po: 
of 
reg 
10 0 ^n An en en 
ESTIMATED AVHRR CLASSIFICATION PERCENTAGE CROP AREAS 
0 
Fig 
30 40 
20 
10 
ESTIMATED AVHRR CLASSIFICATION PERCENTAGE FOREST AREAS 
"m 
i 
tro 
This 
remot 
of a 
scale 
proce: 
requir 
reduci 
indica 
imple; 
the as 
requir 
could 
manne