#3043: Change default color table
--------------------------+---------------------------------------
Reporter: wenzeslaus | Owner: grass-dev@…
Type: enhancement | Status: new
Priority: blocker | Milestone: 7.2.0
Component: Raster | Version: svn-trunk
Resolution: | Keywords: r.colors, d.rast, rainbow
CPU: Unspecified | Platform: Unspecified
--------------------------+---------------------------------------
Comment (by wenzeslaus):
Replying to [comment:5 mlennert]:
> I think that if we provide a default color table (here ''map'') of which
we say that it is good (better than rainbow), we should make our criteria
more explicit.
Article on Gnuplotting (citing Moreland, 2009) lists these features for a
default color table and I'm adding my notes in the nested lists:
* The map yields images that are aesthetically pleasing
* Although some people consider `rainbow` pretty or are simply used to
it, my claim is that it is objectively worse because things which contain
less colors or less styles in general is considered a better graphical
design.
* `grey` is OK but black and white is a bit boring and then there is the
"show color feature" thing Markus mentioned in comment:2
* In cartography black and white have usually special purpose (like text
and its halo), so this excludes color tables which start or end with one
of them.
* The black-body radiation color tables which include or are similar to
color tables like `sepia` in GRASS GIS, or `hot`, `inferno`, `plasma`, and
`magma` in Matplotlib are often recommend, but I haven't seen them used
that much. There might a actually a aesthetic reason for it. The Better
Default Colormap for Matplotlib video suggests that the reason is that
these don't have a green color in them. However, in general, I think they
are nice. Gnuplot default is actually similar to `inferno`, `plasma`, or
`magma`.
* The cubehelix color tables are often regarded as "rotten melon",
although some variations look nice and some are very much like `viridis`,
`plasma`, and others. See seaborn and palettable for examples.
* `inferno`, `plasma`, `magma`, and `viridis` were considered for
Matplotlib and `viridis` was selected because it was the one most people
liked. As mentioned above, it was likely visual because unlike others, it
had green in (the measurable/theoretical features were similar with the
other three).
* The map has a maximal perceptual resolution
* In comment:9 showed that our `rainbow` doesn't always have it. Some
revised rainbow-like color tables may have it quite high.
* Interference with the shading of 3D surfaces is minimal
* This rules out the `grey` and also the ones which go from very low
brightness (black) to very high brightness (white).
* The map is not sensitive to vision deficiencies
* Although there is many different color blindness variations and
sometimes the percentages of affected people are relatively low, this is
often a requirement and my guess is that, similarly to web design,
everybody at the end benefits.
* This rules out anything which contains red and green at the same time
like `rainbow` or `gyr`.
* This also likely rules out anything which does not look good when
converted to gray scale as the intensity is probably the last thing which
is preserved. (I would need to do more here.) However, there is the
benefit of printing in gray scale without much loss.
* The order of the colors should be intuitively the same for all people
* I guess the usual issue here is who can name colors of rainbow in the
right order? Additionally, does the color table include all colors or does
it skip some?
* This likely rules out anything with a lot of colors, so probably also
the rainbow-like color tables which fix some of the classic rainbow color
table problems.
* Simpler color tables with one or two colors like `summer` from
Matplotlib doesn't require any knowledge of order.
* As the order of colors is not clear even few colors may create
confusion. Exception might black-body radiation color tables where the
order might be intuitive (but that might be just result of the brightness
as discussed in the following point).
* I think this is also where the diverging color tables like
`differences` or `curvature` fail here as defaults because the choice of
order of individual colors is arbitrary and the often pronounced middle
would be unsuitable. However, !ParaView uses diverging blue-white-red as a
default and there are is some literature which suggests them.
* Brightness seems to be only generally accepted natural ordering. This
is in favor of `grey` and all color tables which look like grey when
printing in black and white. This is the case for cubehelix color tables
and `inferno`, `plasma`, `magma`, and `viridis`.
* The perceptual interpolation matches the underlying scalars of the map
* (I would need to do more reading here but) I think that it means that
human should be able to guess what is the color in between two colors in
the color table and the result should be the same as what is a more
detailed color table would actually say. Simple color tables with
predictable changes in brightness would fulfill this.
> Why do you find viridis good (better than other color tables) ?
As far as I can tell, the general trend in default color tables is
simplicity, i.e. one, two or three colors, excluding white and black,
clear brightness increase and black and white printing friendly, and color
blind friendliness. Matplotlib's plasma and viridis fulfill those well
together with some color tables from the cubelix and black-body radiation
families (and perhaps some others designed in some other way but
intentionally or unintentionally fulfilling the criteria - a lot of work
would be needed, however, to find and check those).
> Just aesthetics ?
I think it is the final factor after the other things are fulfilled since
many color tables would fulfill the other requirements, but wouldn't be
nice (some of the cubehelix ones, for example). There are two things which
support viridis as color table accepted by people for aesthetic reasons.
viridis has won voting over inferno, plasma, and magma preceding its
inclusion to Matplotlib and Matlab actually has similar color table to
viridis called parula (and I don't think they would select something they
wouldn't consider nice). viridis' authors actually claim (in the video)
that viridis is little better than parula according to its behavior in
some color space (default of both in the past was jet which is similar to
rainbow); some people actually mention aesthetics as well, but that's
quite subtle difference in case of viridis and parula.
> What is good for matplotlib, might not be good for GRASS...
They actually considered even terrain with shading as a use case. So, the
question should be why it wouldn't be good for GRASS when it is good for
Matplotlib?
> More fundamentally, and to repeat myself: IMHO there is '''no''' good
default color table. Color tables qualities are context and target
dependent. I personally thus prefer not to pretend that we can provide a
good color table by default.
However, people simply use the default anywhere from tutorials to articles
and the reasons range from trusting the defaults (because why would
authors put there bad defaults?) to liking the rainbow color tables (which
are considered bad for several reasons). Thus, I think we need to provide
the best default we can. This doesn't mean that we should not promote
usage of a proper color table for a given case. For example, digital
elevation models, temperatures or water depth are good examples where the
default is not appropriate.
''(I'm not including any links because Trac spam filter won't let me.)''
--
Ticket URL: <https://trac.osgeo.org/grass/ticket/3043#comment:10>
GRASS GIS <https://grass.osgeo.org>
