[GRASS-dev] nc_spm_08_grass8?

Maris Nartiss maris.gis at gmail.com
Thu Sep 9 12:22:59 PDT 2021


2021-09-09 19:19 GMT+03:00, Vaclav Petras <wenzeslaus at gmail.com>:
> On Thu, Sep 9, 2021 at 4:46 AM Maris Nartiss <maris.gis at gmail.com> wrote:
>
> In the documentation, we moved from (5 calls):
>
> ```
> g.region raster=lsat7_2002_10
> i.group group=lsat7_2002 subgroup=res_30m input=...
> v.to.rast input=training output=training use=cat label_column=label
> i.gensigset trainingmap=training group=lsat7_2002 subgroup=...
> i.smap group=lsat7_2002 subgroup=res_30m signaturefile=...
> ```
>
> to (16 calls):
>
> ```
> g.mapset mapset=PERMANENT
> r.support map=lsat7_2002_10 bandref=TM7_1
> r.support map=lsat7_2002_20 bandref=TM7_2
> r.support map=lsat7_2002_30 bandref=TM7_3
> r.support map=lsat7_2002_40 bandref=TM7_4
> r.support map=lsat7_2002_50 bandref=TM7_5
> r.support map=lsat7_2002_61 bandref=TM7_61
> r.support map=lsat7_2002_62 bandref=TM7_62
> r.support map=lsat7_2002_70 bandref=TM7_7
> r.support map=lsat7_2002_80 bandref=TM7_8
> g.mapset mapset=user1
> g.region raster=lsat7_2002_10
> i.group group=lsat7_2002 subgroup=res_30m input=...
> v.to.rast input=training output=training use=cat label_column=label
> i.gensigset trainingmap=training group=lsat7_2002 subgroup=...
> i.smap group=lsat7_2002 subgroup=res_30m signaturefile=...
> ```
>
> This seems to be making GRASS more difficult to use instead of making it
> easier to use or at least keeping the status quo.

And we gained ability to use signatures from one scene to classify
different scenes. Sucks if all you do is just tutorials and do not
attempt to use GRASS for any serious work. Have you been trying to use
GRASS to classify a large set of satellite imagery? In GRASS 7 you
must digitize training areas for each scene separately as you can not
use signatures from one scene (imagery group) to classify other scene
(signatures are located inside an imagery group). In GRASS 8 it is no
more – draw training areas once, generate signatures once and apply
signatures to as many scenes as you have (including in a different
mapset!). Heck, they would work out of box even in a different
location (import/export/management module is still a TODO).

> Changing the sample dataset sounds like hiding the issue rather than
> solving it. Having an ultra convenient sample dataset doesn't make the
> software easier to use.

Simplicity is not an end goal. GRASS is a tool for serious data
analysis and thus having to prepare data is nothing unusual. There are
other easier to use tools out there that will happily allow to do
shit-in, shit-out data analysis if one does not care about quality.
(Just take a look at the terrible hack i.signature.copy in addons – a
silent data corruption waiting to happen.)

> Additionally, given that the band references seem highly experimental, I
> don't think they should be required for any workflow at least from the user
> point of view.

Band references could be assigned automatically during import iff the
source has complete metadata (e.g. importing from Sentinel SAFE
format).
Portability of signature files depends on band references. I will not
see any reason why to land i.svm into main if I will not be able to
use portable signatures for automatic classification of large
datasets. Think big and not on a tutorial level. GRASS 7 imagery
classification tools are not HPC ready, GRASS 8 are (for data-parallel
approach).

Looking forward to my task of classifying 66100 imagery scenes with GRASS,
Māris.


More information about the grass-dev mailing list