This is really fantastic to see, and I hope it is read widely. Anything that can get rid of jet is a godsend for better figures.
My only minor quibble is the use of "scientifically" in phrases like "scientifically derived color maps." Since there's no general "scienctific" concept, and it's actually about the color values perceptual consistency, using a term that reflects that specific "scientific" aspect would be a lot clearer. The terms "scientific" and "unscientific" are highly subjective and do not convey much useful meaning, IMHO. And when an article is read by non-scientists, which I hope this article will be, using terms like "unscientific" reinforces the false popular perception that scientists are commonly dividing the world into "scientific" and "unscientific" like a 1950s Time magazine article. But again, very minor quibble in what is a very important message.
This is a weird article. Jet is effectively a divergent colormap and allows highlighting of extemes, but none of the other colormaps they consider are useful in this context. Jet isn't competing with batlow and viridis, it's competing with BwR.
Yeah batlow looks better than jet on a photograph, but that doesn't mean it's necessarily better for communicating data that doesn't share the underlying features of photographs. I'd like to see whether jet or batlow has better accuracy when asking readers to eyeball data values from -1 to 1 based on color.
I personally avoid jet etc, but the motivations they give here seem to miss the point of why these colormaps are used.
I wouldn't describe a rainbow colormap as divergent.
A divergent colormap transitions from one color to another, with neutral color in the middle (e.g., from blue to red, passing through white). A rainbow color map transitions between a series of spectrally-orderd distinct colors (red, orange, yellow, green, blue, indigo, violet); this gives rise to apparent discontinuities (e.g. [0]).
If you want a divergent color scheme to highlight the extreme values, then use one (see, e.g. [1]) rather than a rainbow scheme.
Jet transitions from red to blue, with green as a neutral color in the middle. It happens to transition through yellow and cyan to give more possible shades (~more resolution). It's a bad color map for various reasons, but I think it's perfectly fair to call it divergent.
Not all rainbows are divergent, but if you look at jet you can clearly see it's designed to be roughly symmetric about the mean. Rainbow and divergent are not mutually exclusive.
I have rarely seen Jet used as a divergent colourmap (i.e. in places where the data is roughly centred around 0), and it would do a particularly poor job at this. There are two perceptual peaks in lightness, in the cyan and yellow sections, and neither of these are centred.
Even worse, you can can see in the visualisation of various types of colourblindness that with protanopia only the cyan peak is visible, and with tritanopia only the yellow peak is visible.
If your data is centered around zero (i.e. lots of uninteresting data points near zero), why would you want your perceptual peak there? I can't think of any divergent colormap that has this feature. In jet the cyan and yellow regions highlight small deviations from zero, which are usually more interesting than zero itself.
I understand that jet is poor for accessiblity, and I don't use it for that reason, but in my experience jet is used frequently for divergent data and highlighting extreme values. The article's argument for perceptually uniform colormaps in this context is weak to non-existent. Surely roma wouldn't look very good on a photo of Curie.
Why is perceptual uniformity (in each direction) less important for diverging colour maps? Used as a diverging map with zero at green, jet ramps up in intensity up to yellow/cyan, and then ramps down again. As a result it draws attention to mildly deviating areas.
It is still important, and jet certainly has major flaws. I just think it's not so useful to compare jet to non-divergent colormaps. Any divergent colormap would look bad on a photograph of a human, which seems to be the main evidence provided in this paper.
In a deleted reply to my comment, a user referenced a paper by Borkin [1], that actually compares the quality of decisions made using jet vs other divergent colormaps. Jet loses. It's a shame this sort of argument wasn't included in the ncomms article, because imo it would do a much better job of convincing people who like jet.
I think the reason the example with the photo is useful is because it shows how the colourmap creates features where none exist in the data (due to the nonlinear gradient). When looking at the photo with Jet I can see discrete areas of flat colour, like the Obama 'Hope' poster, but really these boundaries are just artifacts of the colourmap.
I think the same thing happens in the arteries example in that SciPy talk. Sure it could be useful in some scenarios to highlight specific parts of the spectrum (like you mentioned with small deviations from the zero point) but at the end of the day the colourmap doesn't know anything about the data and shouldn't make any assumptions. Highlighting certain small intervals in the data could be better achieved with preprocessing, i.e. some polynomial scaling function.
If you look at the photo of Curie using any divergent colormap, you get roughly the same effect. The problem is that the colormap is divergent and replaces white with a dark color, not anything intrinsic to jet. If you map this image with BWR or roma you end up with similarly unappealing results.
IMO roma is actually the worst of the bunch, despite being the most 'scientific.' Whoever decided that deep, saturated blue and dull brown-red are equally distinct from beige needs to reevalute their model of color perception.
> I just think it's not so useful to compare jet to non-divergent colormaps.
I agree, in the "this should be so obvious that it doesn't need an article written about it" sense. But the continued prevalence of jet used as a sequential colour map suggests that the comparison is necessary.
> Jet is effectively a divergent colormap and allows highlighting of extemes
I think one of the issues is that it doesn't actually highlight the extremes, but rather some arbitrary values in the middle: The values with the brightest colours as well as the steepest increases in brightness are in the cyan and yellow sections in the middle, whereas the extremes are dark red and dark blue - so attention is actually pulled away from the extremes.
Likewise, it highlights some contours but will obscure others - and it's difficult to predict which contours will be highlighted and which will be obscured.
The problem with jet is you don't necessarily perceive the ordering. For example would you even notice on a chart colored with jet if all the yellows were missing?
The jet colormap is horrible, I hate it with a passion, and I know its deficiencies when compared with properly designed colormaps. But I use it all the time. In certain fields (eg FEM applied to mechanical engineering), it's just one of those absolutely counterproductive standards we have to deal with all the time.
Sometimes I have tried using side-by-side figures with other colormap, usually to prove some specific point, but changing the habits of a whole field is really difficult. It is even more difficult when there is not a clear alternative, but several to choose from; I have seen a few other people using alternative colormaps too and they did not use the same as me (or I did not use the same as them).
To give a more extreme example: if a doctor has to check my cancer tests and he has been diagnosing tumors using the jet colormap for decades, I would not feel comfortable if my results were given to him using the much better cividis colormap.
I'm not entirely sure how this got past the Nature editors, since a lot of this is old hat, having been discussed widely, and for many years.
The message is fine, as far as it goes.
However, I would argue that perceptual neutrality is not the only measure of a colour scale.
I hate jet as much as the next person, but I still do use the similar turbo scheme (which is less 'jagged' in perceptual space) when I teach. Why? Because it lets me show an image of a complex field (perhaps sea-surface temperature) and point to a spot of some value, and ask students to visually trace nearby spots of the same hue. It doesn't really matter if some students perceive the same hue for two different actual hues. So long as the neighbouring hues differ, the students can trace a local contour. I need that to explain certain topics.
Using a colour scheme with just a couple of hues makes it very, very, difficult to trace contours. It's even worse when an image is projected onto a screen, in a classroom or at a meeting. Speaker: "See this slightly yellowish green on the colourbar, at 10C? Where is it in the map?" Audience: "I dunno, when's the next coffee break?"
Essentially, colour maps with a lot of hues can make images into something like a contour diagram. This can help, because some fields are so ragged that contours are difficult to understand.
Sometimes, a linear mapping between data values and the perceived colour is the end goal. Other times, not.
There's a kind of colourmap which is a luminance gradient in a single hue, repeated over and over again with different hues. See the various "ramps" colourmaps in ImageJ (where colourmaps are called "look-up tables", or LUTs):
Nice to see this again, but the viridis colormap and the related talks were spread widely a few years ago. What does the Nature article mean now, is the publishing process really this slow?
The first thing I did when I opened the article was to search for viridis. It is in there, but in a sidebar. As others have pointed out, it's weird to make such a big deal about the same sort of thinking which has been done quite well previously, and then bury it in your notes.
If you haven't already watched this, here's the SciPy2015 presentation on the development of viridis. It helps to explain the importance of color maps and how human visual color perception works.
IBM did research back in the 90s on perceptually-based colormaps and how to best represent various types of data within the color dimensions of luminescence, saturation and hue [1]. For exmpale, they found that,
(1) Hue was not a good dimension for encoding magnitude information, i.e. rainbow color maps are bad.
(2) The mechanisms in human vision responsible for high spatial frequency information processing are luminance channels. If the data to be represented have high spatial frequency, use a colormap which has a strong luminance variation across the data range.
(3) For interval and ratio data, both luminance- and saturation-varying colormaps should produce the effect of having equal steps in data value correspond to equal perceptual steps, but the first will be most effective for high spatial frequency data variations and the second will be most effective for low spatial frequency variations.
I disagree with blanket statements like this. Rainbow color maps aren't ideal for showing relative difference in values. But they are great if you want to be able to tell what the value (range) in a given location is. This is almost impossible with a 10-shade grey-scale legend or a height map.
Disclaimer: I use rainbow color maps all the time because I'm often interested in critical value areas of my plots.
Color blind people likely wont be able to read your rainbow plots though. For example, for me Jet is just Blue - White - Green - Dark Green, with white being left of the middle. (Blue + Green is white since I don't see red)
So for me reading a Jet plot is basically impossible, they just doesn't make sense. Like, if a value is weak it is the brightest white, then if it gets even lower it gets blue. It peaks at green in the middle and then goes darker and darker as it gets stronger. Totally unintuitive.
Edit: Even worse is that how I see it depends strongly on what medium you use for the red and green channels. On some screens/paints red is basically black while others are almost fully green, depending on exactly where in the spectrum it is.
Anything that can help me think less about colormaps when producing figures is a clear plus. I've actually started to prefer grayscale figures for this reason.
This is what really confused me when they showed the black and white image as the ground-truth in Fig.1, so what exactly is the batlow colour map adding in this example?
They go too far in the other direction in catering to all possible colorblindnesses. They seem bland and not much different from grayscale. Give me a perception-corrected rainbow, not these two color slimy things.
With temperatures, the blue-red color-map is makes intuitive sense because, culturally, we already have an association of blue = cold and red = hot.
The issue though is where to draw the center between blue and red? Like, when do we transition from light blue to light red? The average temperature? the freezing point? And even then, what color goes in the middle? White, grey, and black are common, but not necessarily natural or perfect options.
One solution is to use a sequential, non-diverging color map that starts at white/light red, and gets more red with higher temperature (or blue, or green). But then we miss out on that intuitive blue/red relationship, and people may have a hard time interpreting "light red" as a low, cold temperature.
> [scientists] fail to condemn the proliferation among themselves if
I don't think scientists should condemn one another for their choice of colors in graphs. Make your argument for one set of colors vs another and scientists can decide for themselves if they agree.
My only minor quibble is the use of "scientifically" in phrases like "scientifically derived color maps." Since there's no general "scienctific" concept, and it's actually about the color values perceptual consistency, using a term that reflects that specific "scientific" aspect would be a lot clearer. The terms "scientific" and "unscientific" are highly subjective and do not convey much useful meaning, IMHO. And when an article is read by non-scientists, which I hope this article will be, using terms like "unscientific" reinforces the false popular perception that scientists are commonly dividing the world into "scientific" and "unscientific" like a 1950s Time magazine article. But again, very minor quibble in what is a very important message.