Road discrimination

[Marcus Tooze]

> > I am looking for a method to extract roads automatically 
> > in remote sensed imagery such as Landsat TM or SPOT.
> > Does anyone have idea of a sophisticated approach to the problem ?
> > Information on previous works or publications would help me very
> > much.
> > 
> It's just like carving an elephant: start with a block of wood, and take
> away everything that doesn't look like an elephant.
> 
> Although I'm not the epitome of an imagery whiz, I seem to recall that
> even SPOT data gives you an awful lot of trouble finding classifications
> that hold still for the kind of linear features you're looking for. 

The usual approach is to use n edge detection algorithm. Of course
this is gonna detect a lot of edges you don't wanr, and of course, miss
a lot of edges you DO want due to tree cover/shadow  or whatever etc. etc.
It's pretty much a none starter to try and do this. 

An alternative would
be to use the imagery as a back drop cell display in v.digit in GRASS and
use the mouse to quickly digitize along the roads on-screen....no map
reg. hassles, and your extracting the roads off the image, albeit manually.
Note: you've obviously got to rectify/register the imagery first!!

Marcus Tooze
USACERL-TAC

Disclaimers etc etc.