112 
  
  
  
  
1985 CORINE MAP 
  
   
  
   
    
   
  
  
  
241 
Interpretation 
112 difference 
211 
     
Real land 
cover change ? 
Difference 
of class 
     
  
  
  
1991 JRC - EMAP SURVEY MAP 
Figure 3. Intercomparison of Original Land Cover Map 
and JRC Survey 1991 - Loures Test Site 
recognised and the parcel (with a different extent) was 
labelled with class 142 ("sport and leisure facilties"). 
Again this is not an entirely appropriate description. In 
fact a class discrepancy of this magnitude is quite seri- 
ous -the parcel is within the group of "agricultural" 
classes in the 1985 map and in the group of "artificial" or 
man-made classes in the 1991 survey. An area of be- 
tween 1 and 2 square km. thus changes between major 
class groups without any real change on the landscape. 
The repetition of this kind of discrepancy at a European 
scale between map revisions would lead to highly erro- 
neous land cover change statistics. 
Figure 3 shows the same kind of comparison on a test 
Site near the town of Loures. Here there is a very 
marked difference between the parcel boundaries since 
there are no dominating landscape features to impose a 
clear structure to the map. In fact the area is a dormitory 
town to the north of Lisbon. The mapped area contains 
some large modern housing estates intermingled with 
older low density housing, some light industry, some ar- 
eas of natural vegetation, and some agriculture and 
546 
abandoned land. When the area is generalised to 25 
hectare parcels (as a minimum) it is extremely difficult to 
decide where the boundaries should be optimally placed. 
The result is two very different maps. Certainly there 
have been some real land cover changes in this area be- 
tween 1985 and 1991. However the statistical signifi- 
cance of these changes will be lost in the general sub- 
jective confusion. The over-riding conclusion from such 
comparisons is that a land cover map revision procedure 
relying on the repetition of a subjective human process is 
unlikely to lead to a statistically useful result. The re- 
quirement for a more automatic, repeatable, procedure 
is paramount. 
4. TOWARDS AUTOMATED LAND COVER 
MAP REVISION 
4.1 Implications of the Portugal Study 
One of the most important conclusions arising from the 
1985-1991 map intercomparisons is that an attempt to 
revise the CORINE land cover maps by undertaking a 
re-mapping exercise (following the original procedure) 
will not yield a result which has a good statistical rela- 
tionship with the original map on account of the subjec- 
tivity inherent to the photo-interpretation and spatial gen- 
eralisation processes. Hence a change detection ap- 
proach is more suitable using the original maps as the 
base-line against which changes should be observed. 
This effectively perpetuates the subjectivity of the first 
edition maps which may lead to the perpetuation of cer- 
tain kinds of errors in the generalisation. However by fol- 
lowing this change detection approach the statistical 
continuity is ensured allowing important studies to be un- 
dertaken on changes in land use over time. The use of 
the original map as a base-line for change detection also 
facilitates easier use of remote sensing data in the up- 
dating procedure since imagery can be partitioned ac- 
cording to the original parcel boundaries and decisions 
on revising class labels and/or re-drawing parcel 
boundaries are local rather than global. 
4.2 Stages of an Automatic Revision Procedure 
At the present time, the procedure for carrying out "auto- 
matic" updating is still under investigation. However, 
some of the key steps can be outlined. 
One of the most important tasks is initially to determine 
which of the original land cover map parcels have 
changed in a significant way. This can be done in princi- 
ple by detecting changes in the imagery within each of 
the original parcels. However this is a non-trivial problem 
because it is not possible to compare satellite radiances 
in old and new imagery on a pixel-by-pixel basis alone. 
The use of old and new images on slightly different days 
of year, with different illumination conditions, slightly dif- 
ferent vegetation conditions, soil wetness etc. can create 
significant pixel radiance differences even though the 
main land cover features have not changed. For this rea- 
son it is much more acceptable to work in the classifica- 
tion domain -i.e. to initially classify pixels and then to 
make comparisons between old and new to detect 
changes. However this procedure will not be useful if 
conducted on a pixel-by-pixel basis since small scale