[GRASS-user] object-based classification (vs pixel-based)

GRASS GRASS user
1

Hello Moritz,

Hello,

Some of my colleagues use eCognition for remote sensing data
classification. After having done so for teaching, we are now looking
into the possibilities of replacing proprietary with open source
solutions in research.

eCognition uses an object-based approach to classification as opposed to
pixel-based. As they put it in their brochure:

"The technology examines pixels not in isolation, but in context. It
builds up a picture iteratively, recognizing groups of pixels as objects.
Just like the human mind, it uses the color, shape, texture and size of
objects, as well as their context and relationships to draw the same
conclusions and inferences that an experienced analyst would draw."

The issue about how to emulate the eCog-approach into GRASS has been around for a while. Here are my two eurocents of knowledge:

eCog consists of several algorithms. First, a segmenting tool (patented...) is used to divide a multispectral image stack into segments. Then, a knowledge-based (fuzzy) merging process gets you to the object-representation.

So to do it with GRASS, we need to have a multispectral segmenting tool plus "the object stuff".

I am not sure if there are any FOSS-segmenting tools in the works that do their job without signature files [medical imaging tools, maybe?].

For the next step, the abstraction from spatial segments to objects, object-oriented grass-friendly environments such as R or CLIPS/CAPE could be used to create a prototype/demonstrator. However, the performance for large data-sets might be significantly below a c/c++ implementation.

Peter

Is something like this possible in GRASS ?
Or are there any other open source solutions using this approach ?

Moritz

--
Dr. Peter Löwe
<peter.loewe@gmx.de>

GMX FreeMail: 1 GB Postfach, 5 E-Mail-Adressen, 10 Free SMS.
Alle Infos und kostenlose Anmeldung: http://www.gmx.net/de/go/freemail

2

AFAIK, SPRING, a GIS program developed by the Brazilian National Space
Research Institute, has excellent classifications algorithms. I once
heard that a comparison was made between several algorithms (including
e-cog), and SPRING was considered the best (don't know the title or
author of the article, though. sorry).

The problem is that SPRING is free as in free beer, not FOSS. But
TerraLib is FOSS, and pretty much everything that is implemented in
SPRING is being (or has been) implemented in TerraLib, so maybe here
is a way.

HTH

Carlos

On Nov 23, 2007 7:49 AM, "Peter Löwe" <peter.loewe@gmx.de> wrote:

Hello Moritz,

> Hello,
>
> Some of my colleagues use eCognition for remote sensing data
> classification. After having done so for teaching, we are now looking
> into the possibilities of replacing proprietary with open source
> solutions in research.
>
> eCognition uses an object-based approach to classification as opposed to
> pixel-based. As they put it in their brochure:
>
> "The technology examines pixels not in isolation, but in context. It
> builds up a picture iteratively, recognizing groups of pixels as objects.
> Just like the human mind, it uses the color, shape, texture and size of
> objects, as well as their context and relationships to draw the same
> conclusions and inferences that an experienced analyst would draw."

The issue about how to emulate the eCog-approach into GRASS has been around for a while. Here are my two eurocents of knowledge:

eCog consists of several algorithms. First, a segmenting tool (patented...) is used to divide a multispectral image stack into segments. Then, a knowledge-based (fuzzy) merging process gets you to the object-representation.

So to do it with GRASS, we need to have a multispectral segmenting tool plus "the object stuff".

I am not sure if there are any FOSS-segmenting tools in the works that do their job without signature files [medical imaging tools, maybe?].

For the next step, the abstraction from spatial segments to objects, object-oriented grass-friendly environments such as R or CLIPS/CAPE could be used to create a prototype/demonstrator. However, the performance for large data-sets might be significantly below a c/c++ implementation.

Peter

> Is something like this possible in GRASS ?
> Or are there any other open source solutions using this approach ?
>
> Moritz

--
Dr. Peter Löwe
<peter.loewe@gmx.de>

GMX FreeMail: 1 GB Postfach, 5 E-Mail-Adressen, 10 Free SMS.
Alle Infos und kostenlose Anmeldung: http://www.gmx.net/de/go/freemail
_______________________________________________
grass-user mailing list
grass-user@lists.osgeo.org
http://lists.osgeo.org/mailman/listinfo/grass-user

--
+-----------------------------------------------------------+
              Carlos Henrique Grohmann - Guano
  Visiting Researcher at Kingston University London - UK
  Geologist M.Sc - Doctorate Student at IGc-USP - Brazil
Linux User #89721 - carlos dot grohmann at gmail dot com
+-----------------------------------------------------------+
_________________
"Good morning, doctors. I have taken the liberty of removing Windows
95 from my hard drive."
--The winning entry in a "What were HAL's first words" contest judged
by 2001: A SPACE ODYSSEY creator Arthur C. Clarke

Can't stop the signal.

3

Hello Peter,

On 23/11/07 08:49, Peter Löwe wrote:

The issue about how to emulate the eCog-approach into GRASS has been
around for a while. Here are my two eurocents of knowledge:

eCog consists of several algorithms. First, a segmenting tool
(patented...) is used to divide a multispectral image stack into
segments. Then, a knowledge-based (fuzzy) merging process gets you to
the object-representation.

So to do it with GRASS, we need to have a multispectral segmenting
tool plus "the object stuff".

I am not sure if there are any FOSS-segmenting tools in the works
that do their job without signature files [medical imaging tools,
maybe?].

I have the feeling that this is the most important part.

For the next step, the abstraction from spatial segments to objects,
object-oriented grass-friendly environments such as R or CLIPS/CAPE
could be used to create a prototype/demonstrator. However, the
performance for large data-sets might be significantly below a c/c++
implementation.

Is this automatic in eCognition or does it demand user input ? If the latter and IIUC, the main challenge here is a fuzzy classification algorithm, or ? Implementations of that shouldn't be too difficult to find.

Thanks for your explanations.

Moritz