Thanks for the input! I added some more questions below.
Also, later today I will probably post to R-Sig-Geo (or some other
list?) to ask what tools people currently use, and what areas they
would find most helpful to be integrated into GRASS. That might help
focus where I should spend the time.
On Mon, Apr 2, 2012 at 6:42 AM, Moritz Lennert
<mlennert@club.worldonline.be> wrote:
On 30/03/12 23:56, Eric Momsen wrote:
I am reading more about image classification (from
http://grass.osgeo.org/wiki/GRASS_SoC_Ideas ):
4. Implement image segmentation algorithms and tools
5. Implement region-based classification
6. Implement hierarchical classification tools (e.g. being able to
create a large class "forest", with subclasses of different types of
forests)
...
Concerning the ideas:
4. Currently GRASS does not provide any image segmentation as such. i.smap
contains image segmentation in its process, but the user cannot get
segmented outputs. Many algorithms exist and its an ongoing field of
research. FLOSS software that provide such algorithms include Orfeo Toolbox
(OTB), SAGA, R, Sextante (?) and probably a whole series of others. I think
the implementation of a series of such algorithms could be a project on its
own.
Does it make more sense to implement the algorithms again, or pick the
most useful that are implemented in some other FLOSS and provide an
easy integration to access them from the GRASS front end? (I'm
thinking of v.krige which uses existing R packages to do the
processing work.)
Has Sextante or OTB been tied into GRASS in this manner?
5. One of the main applications of image segmentation today is in
region-based classification of very high resolution imagery. As with current
resolutions individual objects are composed of many pixels, it is often more
efficient to first identify "objects" or homogeneous multi-pixel regions in
the image through segmentation and then to classify these regions. OTB
provides this I think, but I don't know if any other FLOSS software does. 5
depends on 4, so it is only possible if 4. is limited to the strict minimum
in terms of segmentation algorithms and then focus is put on 5. Maybe a bit
too ambitious.
If I can use the OTB implementation from GRASS, then I will include
this as a stretch goal if time remains at the end of the summer.
6. In the current classification algorithms in GRASS each designated class
of pixels is on the same hierarchical level as others. However, it is often
interesting to provide the option to classify an image first in a rough
manner into a series of base classes (built-up, vegetation, naked soils) and
then to refine classification within each of these classes (e.g. built-up
into high-density / low-density, vegetation into forest, grasslands, etc),
but to keep the hierarchy, i.e. to allow extracting an image (and a legend)
of the classification at each level.
This sounds interesting as well. It seems I should propose either 4/5
or 6 for the summer work.
Hope this helps and maybe motivates others to join-in as mentors.
Moritz
