to 25 
cult to 
laced. 
there 
ea be- 
signifi- 
i sub- 
| such 
'edure 
ess is 
he re- 
edure 
m the 
npt to 
king a 
dure) 
| rela- 
ubjec- 
i| gen- 
n ap- 
as the 
erved. 
e first 
of cer- 
by fol- 
tistical 
be un- 
use of 
n also 
je up- 
ad ac- 
jisions 
parcel 
"auto- 
wever, 
armine 
have 
princi- 
ach of 
oblem 
ances 
alone. 
t days 
tly dif- 
create 
jh the 
is rea- 
sifica- 
en to 
detect 
eful if 
scale 
changes may not be meaningful for parcels of at least 25 
ha. Also small errors in old and new image co- 
registration (of the order of 1 - 2 pixels) may lead to large 
differences being detected between pixels in very patchy 
inhomogeneous zones at the pixel scale e.g. large 
houses with gardens, small-holding farms or horticulture 
with regularly interspersed buildings and cultivated plots. 
For this reason we are investigating a possible two- 
component approach to change detection in imagery as 
follows: 
(i) within each parcel boundary from the original map 
produce class histograms from old and new imagery, 
then look for significant differences in class composition 
(ii) within each parcel boundary from the original map 
perform a cross-correlation of the old and new pixel radi- 
ance data and then to threshold the result in order to de- 
tect major spatial differences. 
A combination of (i) and (ii) then permits parcels with sig- 
nificant land cover change to be flagged. However it is 
axiomatic that once a land cover parcel is flagged as 
having changed its immediate neighbours will also be 
potential change candidates because of the possible re- 
quirement to re-draw boundaries. The posible parcel 
change detection approach is shown in figure 4. 
4.3 Development of an Image Classifier with Good 
Generalisation 
One of the most important requirements of the change 
detection procedure is a classification of the old and new 
imagery into as many classes as possible with high ac- 
curacy over large geographic areas. A potential method 
for doing this is to use an artificial neural network classi- 
fier which has been demonstrated in experiments under- 
taken at the JRC to yield results considerably better than 
conventional statistical image classifiers with large num- 
bers of classes (Kanellopoulos et al. 1992). Work on fur- 
ther developing these classifiers is continuing at the 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
OLD ORIGINAL NEW 
IMAGERY MAP IMAGERY 
y J 
CLASSIFY CLASSIFY 
À i 
FORM OLD MAP PARCEL 
CLASS STATISTICS 
1 Y 
CHANGES IN CLASS MAJOR DIFFERENCE IN 
COMPOSITION ? CROSS-CORRELATION OF 
OLD AND NEW PARCEL ? 
y 
FLAG CHANGED 
PARCELS 
INVOKE REVISION 
PROCEDURE 
  
  
  
Figure 4. Proposed Procedure for Parcel Change 
Detection 
547 
JRC. The neural network approach has a distinct advan- 
tage in that it is appropriate for parallel processing. Stud- 
ies are currently being undertaken at the JRC on impl- 
menting neural networks for remote sensing data 
analysis on parallel transputer networks. 
4.4 Development of Rule-Based Boundary and Class 
Change 
Perhaps the most difficult stage of "automatic" CORINE 
map revision is the decision making involved in (a) 
changing a class label for a parcel which has been de- 
tected as 'changed' and/or (b) re-drawing the boundaries 
for a changed parcel and its immediate neighbours for 
which there are topology-preserving and topology non- 
preserving possibilities. 
We believe a rule-based approach is indispensable in 
this context and that a detailed parcel-by-parcel evalu- 
ation has to be made where the significant image class 
changes have occurred. Essentially these changes have 
to be isolated and their impact on boundaries and 
CORINE class labels carefully evaluated. This requires 
also that a useful relationship can be established be- 
tween detectable image classes and CORINE land cover 
classes. There is a possibility of building up statistics of 
image class membership within the CORINE land cover 
nomenclature clases by examining the image classes 
detected (on old imagery) inside each parcel in the origi- 
nal land cover map. It will then be possible to determine 
the most likely new CORINE class label(s) for a changed 
parcel using its new image class composition on a 
probabilistic basis. However this procedure can not be 
separated from the spatial analysis since the image 
class composition of a parcel is dependent on its bound- 
ary. Procedures are thus suggested which involve, for 
example,: (a) identifying which pixels have changed, (b) 
deciding if these represent a growth of some existing 
landscape feature or a completely new feature (c) decid- 
ing whether simply to move an existing boundary or to 
create a new polygon and (d) deciding if new CORINE 
class labels are appropriate given the new proposed par- 
cel geometry. Such complex inter-related decision- 
making in our view requires an expert system approach 
if it is to be automated. Such a system would require 
many rules and a good probabilistic evidential reasoning 
model. 
These approaches are currently being investigated at 
the JRC and are likely to form an important component 
of our work on land cover map revision. 
5. DISCUSSION 
The revision of a land cover map with continental cover- 
age is a formidable task by any approach. We believe 
remote sensing offers a possibility to do this automati- 
cally but there are many technical problems to be 
solved. Many of these problems are common to a multi- 
tude of applications of remote sensing -i.e. how to accu- 
rately extract large numbers of classes from satellite im- 
agery, how to get good performance over wide 
geographical areas, how to spatially generalise products 
in a meaningful way. Work is underway on all these is- 
sues at the JRC and we hope to exploit techniques such 
as neural networks and expert systems in helping to