[Qgis-user] Use Case: Segmentation (for photo-interpretation) without Classification?

[chris hermansen]

Laurent and list,

On Mon, Jun 30, 2025 at 3:17 AM celati Laurent via QGIS-User <
qgis-user at lists.osgeo.org> wrote:

> Hello,
> i’m taking the liberty to post a message in order to share an use case.
> Our need/purpose is to find existing methods/tools that would allow to
> perform segmentations with the aim of helping botanists for pre-map
> physiognomic units before to go on the field. In other words : to produce a
> delineation (polygonal layer) of physiognomic units. In my humble opinion,
> this is an unusual need/use case because the segmentation is NOT
> intended/aimed as a first step before classification. The goal is to
> reproduce as closely/faithfully as possible the botanist’s
> photo-interpretation. We need to find the right balance/parameters between
> the number of segments, their sizes, and their compactness (shape). The
> idea would be to perform relevant segmentation across the entire area being
> processed, neither over-segmented nor under-segmented.
> With optimal parameters/configuration and input variables/data that can
> vary depending on the environment (urban, agricultural, wooded, herbaceous).
>
> I’ve done some tests, notably via Qgis with OTB Approaches Region Merging:
>
> https://www.orfeo-toolbox.org/CookBook-7.0/Applications/app_GenericRegionMerging.html
> I haven’t really tested other traditional approaches (such as cluster mean
> shift or watershed, for example). Until now, I haven’t tested AI/DL because
> I thought there wouldn’t be any segmentation models trained on IRC ortho DB
> images that could meet our needs.
> This is why I prioritized “traditional” segmentation tools/approaches.
>
> My tests were done using satellite spot6/7 images and Infrared Colour
> aerial orthographic databases (franch mapping agency). Tests with the
> aerial orthographic database are more resource-intensive and
> time-consuming. Furthermore, this OTB tool tested (GRM) is not suitable for
> large images. The results are quite good, particularly for agricultural and
> urban areas.
> For forested areas, it’s more complicated. And in order to improve the
> results, it may be necessary to consider integrating textures or indicators
> (NDVI or other) ?
>
> if you have any comments or suggestions for guidance regarding
> methods/tools/data, I would be happy to read you.
>

 In my experience, which relates to classification / interpretation of
imagery for purposes of classifying forest ecosystems, shrub and tree
height estimation and species identification form a fundamentally important
part of image segmentation.

I have never seen this done successfully without stereo imagery.  So in my
mind at least, any image segmentation carried out on "mono" images is going
to struggle to identify complex forest structure such as multi-story
stands, at least based on pattern detection and spectrographic
information.  But I may be wrong or at least outdated in my thinking, so I
would love to be corrected!

Of course LiDAR can achieve decent vegetation profiles and I know that many
proponents of the use of LiDAR claim effectiveness at identifying crown
shape and therefore species, at least in temperate forests.  I suppose
LiDAR is also effective at seeing through overstory.

I also tend to think that "one time" vegetation inventory methods aren't
all that useful in and of themselves, but rather must be incorporated into
an ongoing inventory / update program.  This speaks to dealing with changes
in sensor ability over time so that one can claim that "changes" are real
and not just a result of improved detection capability.

Again, I would love to find that I am in error with respect to any of my
claims!

Finally, on the topic of attribute classification, there are many
classification systems that rely on "expert interpretation" of ground and
vegetation conditions - for example, wetland classification systems used in
many parts of Canada.  These systems do not typically collect physical
parameters, but rather rely on the interpreter to "integrate" the presence
of absence of those physical parameters into a class of some kind.  This
document is a good example of that kind of integrative thinking:

https://open.alberta.ca/dataset/e3c7894f-875c-4a54-8248-7f5368acd82c/resource/2b01b2b8-91b3-411a-8806-01cf5df306c1/download/srd-alberta-wetland-inventory-classification-system-version-2-0-8342.pdf


-- 
Chris Hermansen · clhermansen "at" gmail "dot" com

C'est ma façon de parler.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.osgeo.org/pipermail/qgis-user/attachments/20250630/d02f6f36/attachment.htm>