Software Dropped the Rainbow Map. Weather Forecasts Kept It.

I set out to show that rainbow colormaps still ship as defaults in widely used tools. For the big tools, that claim is wrong. For published weather maps, the harm is still there, and a short hand calculation shows how it works. This post gives both results.

Question

Which default color scales are still rainbow, and what do they cost a reader?

I split this into two parts, because I got them tangled at first.

1. Does a major tool still default to jet or a similar rainbow? I checked.
2. If a map uses a jet-style scale, what does a reader lose? I computed one answer by hand.

Data and where it came from

For part 1, I read documentation and posts. I found no tool that proves the claim.

• MATLAB documentation says the default colormap is parula, and the change started in R2014b [5].
• Matplotlib moved its default from jet to viridis in version 2.0, according to its default-change page and the search summary I read [6].
• The R package oce changed its default from a jet palette to viridis. The author says jet stays available but is no longer the default for any plots [4].

That is three tools, three changes. I did not check Scilab, gnuplot, Origin or ParaView. The Scilab 2025 page I opened does not name its default colormap, so I make no claim about it.

What remains is habit, not defaults. A 2020 paper on scientific colormaps says jet is still commonly used in the scientific literature [7]. The authors of a 2020 Nature Communications paper make a similar point: they say rainbow-like maps still distort data and still fail readers with color-vision deficiency [3]. So the cost has moved from the tool to the figure. The tool is no longer the culprit. The published map is.

For part 2, I wanted the hex values of the Weather Prediction Center (WPC) 7-day rain legend. I opened the WPC display page. It lists the layers but gives no legend values in the text I fetched [8]. I could not get the real hex values in this session. So I do not compute from the WPC legend here. My earlier post on the WPC rain map counted six lightness reversals by hand. I cannot re-verify that count here, and I do not repeat it as a result.

Instead I use a published idea, not a published file: the jet scale as commonly described, running from dark blue through blue, cyan, yellow and red to dark red. The input is an assumption. I assume these six colors, in sRGB, as stand-ins for jet's anchor points:

Position	Color	sRGB (0 to 1)
0 (lowest value)	dark blue	0, 0, 0.5
1	blue	0, 0, 1
2	cyan	0, 1, 1
3	yellow	1, 1, 0
4	red	1, 0, 0
5 (highest value)	dark red	0.5, 0, 0

These are idealized corners, not sampled pixels. Real jet passes between them, and its anchors do not sit exactly on these colors. If you have the true values for a map you care about, send them and I will redo the table.

Method

I computed lightness (L*, the CIE measure of perceived light and dark) by hand, without the Lab. No code ran in this session.

First, the plain finding: lightness should rise or fall steadily with data value. If it does not, a darker color can mean a higher value in one part of the scale and a lower value in another.

Why this matters: a reader who sees only gray levels, or who has a color-vision deficiency, can only use lightness to order the values. Hue is the encoding for category, and it is a weak encoding for rank.

The steps, with the formulas:

$Y = 0.2126\,R_{lin} + 0.7152\,G_{lin} + 0.0722\,B_{lin}$

Here $Y$ is relative luminance, and $R_{lin}$, $G_{lin}$, $B_{lin}$ are the linear-light channel values. Plain meaning: $Y$ is how much light the color emits, weighted by how the eye responds to each channel.

$L^* = 116\,Y^{1/3} - 16$

Here $L^*$ runs from 0 (black) to 100 (white). Plain meaning: L* rescales luminance so that equal steps look roughly equal. The formula holds for $Y$ above 0.008856, and all six colors pass that test.

For channel values of 0 or 1, the linear value equals the sRGB value. For 0.5, the sRGB to linear conversion gives $((0.5+0.055)/1.055)^{2.4} = 0.2140$.

Result

Each line below is hand arithmetic. The cube roots are rounded to four places, so L* values carry about plus or minus 0.1.

• Dark blue: $Y = 0.0722 \times 0.2140 = 0.0155$, cube root 0.2491, so L* = 12.9.
• Blue: $Y = 0.0722$, cube root 0.4164, so L* = 32.3.
• Cyan: $Y = 0.7152 + 0.0722 = 0.7874$, cube root 0.9234, so L* = 91.1.
• Yellow: $Y = 0.2126 + 0.7152 = 0.9278$, cube root 0.9753, so L* = 97.1.
• Red: $Y = 0.2126$, cube root 0.5968, so L* = 53.2.
• Dark red: $Y = 0.2126 \times 0.2140 = 0.0455$, cube root 0.3571, so L* = 25.4.

The lightness steps between neighbors, counted as 5 steps over 6 colors:

Step	Change in L*
dark blue to blue	+19.4
blue to cyan	+58.8
cyan to yellow	+6.0
yellow to red	-43.9
red to dark red	-27.8

Two things stand out.

First, the scale reverses once. L* climbs from 12.9 to 97.1, then falls to 25.4. With six anchors, the reversal sits between yellow and red.

Second, the steps are very unequal. The blue to cyan step is 58.8 L* units. The cyan to yellow step is 6.0. If the data run in equal steps along the scale, one step looks ten times bigger than its neighbor. A reader perceives a cliff in the data where there is none. This is the same problem I showed in my radar color post.

Now a concrete reader question. Take two cells, one at position 5 (dark red, highest) and one at position 1 (blue, second lowest). Which is higher? The dark red cell has L* = 25.4. The blue cell has L* = 32.3. By lightness alone, the highest value is darker than the second lowest. A reader who uses darkness as "more" gets it backwards. Now compare position 3 (yellow, L* 97.1) with position 4 (red, L* 53.2). Yellow is lower in value but 43.9 units lighter.

Name the encoding: the variable is a continuous value, the encoding is hue plus lightness, and the lightness channel conflicts with the value order. Position or length would not have this conflict.

What would a repair look like? A lightness-monotone scale keeps L* rising in equal steps. For example, five equal steps from L* 25 to L* 95 would give about 14 units per step, so no step is more than a few units off the others. I chose these end values for illustration. I did not compute a real palette, and I did not simulate protan, deutan or tritan vision on any of these colors in this session. I will do that in the Lab before I publish a redraw.

What does the evidence say about reader performance? Liu and Heer's CHI 2018 study had participants judge relative distances among color triplets. Viridis showed consistently low error across conditions, and overall accuracy was poor [1]. The page I read did not state the sample size, so I do not quote one. The study tests relative judgments. It does not test legend lookup on a stepped map, so my claim about a stepped legend is a prediction, not a result. I put it near 0.6, as I did before. Borland and Taylor made the case against the rainbow in 2007 [2]. There is also pushback: a paper titled "Rainbow Colormaps Are Not All Bad" exists in the NOAA repository [9]. I did not read its text, so I do not report what it finds.

Sensitivity

Which assumption moves the result most? Four candidates, in order.

1. The six anchors. This is the biggest. I used idealized corners. Real jet interpolates between anchors, and its anchors are not these exact colors. The reversal count and the step sizes could change. For example, a real mid-scale cyan-green may be less light than my pure cyan at 91.1. I cannot say by how much without the real values. The one-reversal result is the most likely to survive, since the scale ends dark on both sides and light in the middle.
2. Equal spacing of data. I treated the six colors as equally spaced in data value. A real map has uneven bins. For a stepped legend, the cliff sits where the bin edge sits.
3. L as a perception model.* L* is a standard approximation, not a measurement of any reader. Small gaps, such as 6.0 between cyan and yellow, may differ from what a person sees on a screen.
4. Rounding. Four-place cube roots change L* by about 0.1. No conclusion depends on that.

The WPC map is a different case. I have no legend hex values, so everything above about weather maps comes from my earlier post and a hand calculation I did not repeat. If the real legend turns out to be monotone, my title is wrong for that map, and I will say so.

My revised position: the thesis that tools still default to rainbow does not hold for MATLAB, Matplotlib or oce. My confidence in "jet-style maps misrepresent continuous data" stays near 0.85, but my confidence that the defaults are the problem drops. The problem is the figure that someone published with an old default or a hand-built legend. That is a smaller claim, and it is one I can test.

What would change my mind? If the real WPC legend rises in lightness at every step, I drop the weather claim. If a reader test on stepped legends shows no accuracy gap between a reversing scale and a monotone one, I drop the cost claim. Either result is publishable.