.sbon 
and 
but 
1sive 
ural 
scale 
Large 
les a 
1984 
and 
rely, 
272). 
were 
»rmat 
tuto 
two 
5000, 
>lour 
1984 
| the 
Maps 
logy 
ation 
>TERS 
Llary 
thin 
3 the 
3sing 
geo- 
1sing 
Lmage 
and 
map 
lata; 
natic 
cound 
natic 
> the 
jital 
the 
would 
natic 
aver, 
some problems related to distortions on 
certain zones of the image disabled that 
registration. Therefore, these bands were 
registered to the road network digital map 
by using 60 GCPs. The 1984 Landsat TM image 
was co-registered to a previously 
registered 1995 Landsat TM (existent in our 
data base) using 38 GCPs. Some difficulties 
were found on 
registration due to the 11 years gap 
between the images. In each step a first 
performing this co- 
order polynomial and a nearest 
neighbourhood resampling was used. 
Image classification at the pixel level. A 
special effort was put into the spectral 
classification of the images to derive as 
much information as possible from the 
satellite data. Since the study area 
exhibit a large spectral variability, we 
decided to stratify the study area per 
counties. This stratification allowed the 
definition of a sub-set of CLUSTERS classes 
for each county, avoiding eventual 
misclassifications with other non-existing 
spectral 
characteristics. The training areas for 
classes with the same 
classifying the 1984 and 1995 images are 
based on an exhaustive set of training 
areas that were defined in a previous study 
to classify a 1991 image for the same study 
area (CNIG, 1996). Because of the time gap 
between images and their different spatial 
resolution, the training areas had to be 
adjusted to each image and date. 
The image classification was performed 
using spectral 
informational classes, in order to avoid 
classes instead of 
classes with a large spectral diversity and 
consequently an increase on the overlap 
different 
ellipses. The several spectral classes of 
between landuse spectral 
the same informational class were 
aggregated after the image classification. 
Besides the sub-set of CLUSTERS classes 
defined for each county, other classes with 
no direct correspondence to any CLUSTERS 
class were also defined, i.e., land cover 
road pavement and bare soil. An artificial 
code was created for those classes, and the 
assignment to a CLUSTERS class was based on 
a region algorithm. The 
classification of each county was performed 
growing 
in an iterative way using the maximum 
likelihood algorithm. In each step the 
results were validated using a set of pre- 
defined test areas for each county. The 
threshold was increased in each step, 
allowing (1) the definition of new training 
areas corresponding to unclassified pixels 
and (2) the re-definition of some training 
areas that were producing 
misclassifications. 
The pixel based classification methodology 
was applied to the 1984 Landsat TM and to 
the 1995 SPOT image to produce initial 
landuse maps for each date. 
Contextual analysis. A first level of 
improvement was applied to the output maps 
of the image classification at the pixel 
level to generate a SATMAP for each date. 
Simple techniques based on the region 
growing theory were used to modify the 
boundaries of regions of a given class such 
that misclassified pixels in the 
neighbourhood could be aggregated to that 
region. At this stage, we also applied 
contextual algorithms (CNIG, 1996), for the 
identification of specific landuse classes, 
such as airports and aerodromes (class 
A32.3), Stadiums and associated sport 
fields (subset of class A502 of the 
nomenclature) and beaches and dunes (class 
E01.2). The rational behind the development 
of these algorithms was based on knowledge 
based rules used by humans in aerial 
photointerpretation. 
Landuse map improvement by ancillary data 
integration. For the second level of 
improvement, a binary mask (urban/rural) 
was integrated with the spectral data to 
isolate misclassifications caused by 
similarities between some rural and urban 
classes. To produce a binary mask for 1984, 
and since there are no SPOT images for that 
date, the binary mask produced for 1991 
(Caetano et al., 1997a) was adjusted for 
1984 by means of screen digitising using a 
colour composite of the Landsat TM image 
for that date. Since we verify that this 
procedure was efficient, the same strategy 
was adopted to generate the binary mask for 
1995. After the post-classification 
segmentation of both maps, the methodology 
{1997a) to 
discriminate CLUSTERS classes specific to 
developed in Caetano et al. 
each strata, was applied to both maps. 
Within the urban strata, 
operators were developed to identify 
several types of residential classes (level 
IV of the nomenclature) based on the 
abundance (estimated with the NDVI) and 
contextual 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 693