Planimetric features 
Coordinates are related 
geographic reference. 
The nature and the density of the extracted features 
fit with those usually found in topographic 1:50 000 
or 1:100 000-scale maps. 
to a cartographic or 
3. COMPUTER ASSISTED EXTRACTION 
A standard product means a constant price and 
delay, whatever the amount of extracted features is. 
This goal cannot be reached with manual process: 
The time spent by an operator to acquire features 
manually is roughly linear with the amount of data. 
Our goal is double: Cutting down and smoothing 
production costs. 
It is well known that fully automated extraction of 
relevant features in images is a hard work. 
It implies a semantic analysis of the whole image. 
Past and present research efforts have shown very 
little progress in that matter. Artificial Intelligence 
techniques together with image processing are 
involved but connection between these two domains 
is still difficult. 
Our approach is more pragmatic, it is based on the 
following considerations: 
A features extraction process is made of three 
sequential steps, identification - acquisition - quality 
control. 
- Features identification is quite an easy task for 
human brain whereas automatic systems cannot 
achieve reliable results. 
- On the other hand, acquiring accurately a lot of 
coordinates with a digitizing device is a rather slow 
an boring job. A computer assistance should 
certainly speed up the process if properly designed. 
- Two kind of errors may be done while extracting 
features, geometric errors (a coordinate does not fit 
with the true feature position) and syntactic errors 
(topological graph inconsistency). 
Geometric errors are easily detected when extracted 
data are superimposed on the raster image displayed 
on the screen. 
Syntactic errors cannot be fully detected by visual 
control. An automated checking process is needed to 
get reliable results. 
Those simple remarks have led us to design semi- 
automated tools making the most of both humans 
343 
and computers performances. 
These tools perform supervised classification of land 
cover areas, delineation of linear networks and 
edition of topological relations. 
General purpose computers have been chosen 
instead of specialized hardware because the 
performance growing rate of standard hardware is 
higher and its lower price enables a constant 
improvement of equipment. 
3.1 Land cover extraction 
Well known spectral classification techniques 
produce good performances for vegetative and 
hydrographic items extraction (ref. Haralick). 
They are quite inefficient for man-made features (like 
built-up areas) where spatial texture is much more 
significant than spectral information. 
These techniques can be extended to texture 
classification if a texture analyser is provided (ref. 
Zhang). That seems to be useful for fine and regular 
urban structures. 
Supervised or unsupervised segmentation methods 
based on region growing or split and merge 
algorithms have been tested on SPOT images (ref. 
Bretaudeau). The requirement of finely tuned 
parameters, depending on image properties, makes 
these methods impracticable in an industrial context. 
In order to unify procedures and thus reduce 
operating complexity, we have implemented a single 
criterion for land cover extraction, that is maximum 
likelihood classification with normal distributions 
assumption and supervised training. 
We think it is a rather general criterion, even if real 
spectral distributions are not often normal nor 
unimodal, because any statistical distribution can be 
approximate by a finite union of normal distributions. 
In that case, operator must define sub-classes during 
the training phase and then, merge the 
corresponding classified pixels. 
The most important point is the training phase, the 
samples designation should be as fast and accurate 
as possible. 
Our system use interactive designation with visual 
feedback: 
The operator draws, with a mouse, polygonal 
samples on the raster image displayed in a window. 
The classified pixels are displayed on request in real 
time (delay ~ 1 second). So, the operator can, at any 
time, test the fit of the samples set. 
Classification may be operated in global or local 
mode. In local mode, the operator extracts 
interactively a single area entity pointed at, using the 
mouse. 
In any case, the classification process can be 
thought as an explicit extension of a (implicit) 
description of the land cover given by human visual 
interpretation. 
3.2 Networks extraction 
Manual vectorization, one point after another, of 
linear features is extremely time consuming. 
Human operators, for bio-mechanical 
cannot achieve both precision and speed. 
reasons, 
Automatic methods have been investigated in many