Last year, in a post titled Unweaving the rainbow, Matt Hall described our joint attempt to make a Python tool for recovering digital data from scientific images (and seismic sections in particular), without any prior knowledge of the colormap. Please check our GitHub repositoryfor the code and slides, andwatch Matt’s talk (very insightful and very entertaining) from the 2017 Calgary Geoconvention below:
One way to use the app is to get an image with unknown, possibly awful colormap, get the data, and re-plot it with a good one.
So it might come as a surprise to some, but this post is a lifesaver for those that really do like rainbow-like colormaps. I discuss a Python method to equalize colormaps so as to render them perceptual. The method is based in part on ideas from Peter Kovesi’s must-read paper – Good Colour Maps: How to Design Them – and the Matlab function equalisecolormap, and in part on ideas from some old experiments of mine, described here, and a Matlab prototype code (more details in the notebook for this post).
Let’s get started. Below is a time structure map for a horizon in the Penobscot 3D survey(offshore Nova Scotia, licensed CC-BY-SA by dGB Earth Sciences and The Government of Nova Scotia). Can you clearly identify the discontinuities in the southern portion of the map? No?
OK, let me help you. Below I am showing the map resulting from running a Sobel filter on the horizon.
This is much better, right? But the truth is that the discontinuities are right there in the original data; some, however, are very hard to see because of the colormap used (nipy spectral, one of the many Matplotlib cmaps), which introduces perceptual artifacts, most notably in the green-to-cyan portion.
In the figure below, in the first panel (from the top) I show a plot of the colormap’s Lightness value (obtained converting a 256-sample nipy spectral colormap from RGB to Lab) for each sample; the line is coloured by the original RGB colour. This erratic Lightness profile highlights the issue with this colormap: the curve gradient changes magnitude several times, indicating a nonuniform perceptual distance between samples.
In the second panel, I show a plot of the cumulative sample-to-sample Lightness contrast differences, again coloured by the original RGB colours in the colormap. This is the best plot to look at because flat spots in the cumulative curve correspond to perceptual flat spots in the map, which is where the discontinuities become hard to see. Notice how the green-to-cyan portion of this curve is virtually horizontal!
That’s it, it is simply a matter of very low, artificially induced perceptual contrast.
Solutions to this problem: the obvious one is to Other NOT use this type of colormaps (you can learn much about which are good perceptually, and which are not, in here); a possible alternative is to fix them. This can be done by re-sampling the cumulative curve so as to give it constant slope (or constant perceptual contrast). The irregularly spaced dots at the bottom (in the same second panel) show the re-sampling locations, which are much farther apart in the perceptually flat areas and much closer in the more dipping areas.
The third panel shows the resulting constant (and regularly sampled) cumulative Lightness contrast differences, and the forth and last the final Lightness profile which is now composed of segments with equal Lightness gradient (in absolute value).
Here is the structure map for the Penobscot horizon using the nipy spectum before and after equalization on top of each other, to facilitate comparison. I think this method works rather well, and it will allow continued use of their favourite rainbow and rainbow-like colormaps by hard core aficionados.
Perceptual rainbow palette – Matlab function and ASCII files
In my last post I introduced cubeYF, my custom-made perceptual lightness rainbow palette. As promised there, I am sharing the palette with today’s post. For the Matlab users, cube YF, along with the other palettes I introduced in the series, is part of the Matlab File Exchange submission Perceptually improved colormaps.
For the non-Matlab users, please download the cubeYF here (RGB, 256 samples). You may also be interested in cube1, which has a slightly superior visual hue contrast, due to the addition of a red-like color at the high lightness end but at the cost of a modest deviation from 100% perceptual. I used cube 1 in my Visualization tips for geoscientists series.
Perceptual rainbow palette – preformatted in various software formats
The palettes are also formatted for a number of platforms and software products: Geosoft, Hampson-Russell, SMT Kingdom, Landmark Decision Space Geoscience, Madagascar, OpendTect, Python/Matplotlib, Schlumberger Petrel, Seisware, Golden Software Surfer, Paradigm Voxelgeo. Please download them from my Color Palettes page and follow instructions therein.
Comparison of South America maps using, from left to right: ROYGBIV (from this post) , classic rainbow, cubeYF, and grayscale
Again, there is little doubt in my mind that cubeYF does a superior job compared to the other two rainbow palettes as it is free of artefacts  and more similar to grayscale (with the additional benefit of color).
The ROYGBIV and cubeYF map have been included in Marek Kultys’ excellent tutorial Visual Alpha-Beta-Gamma: Rudiments of Visual Design for Data Explorers, recently published on Parsons Journal for information mapping, Volume V, Issue 1.
An online palette testing tool
Both cubeYF and cube1 feature in the colormap evaluation tool by the Data Analysis and Assessment Center at the Engineer Research and Development Center. If you want to quickly evaluate a number of palettes, this is the right tool. The tool has a collection of many palettes, organized by categories, which can be used on 5 different test image, and examined in terms of RGB components and human perception. Below here is an example using cube YF.
Then I wondered: what if I used this to tell me something about a color palette’s mood? The circular histogram of colors reminded me of the Harmonic templates  on the hue wheel from this paper And so I created fat colorbars using the three palettes I used in the last post, saved them as images, and run the monitor with them. Here below are the results for Matlab jet, Industry Spectrum, and cubeYF. Looking at these palettes in terms of harmony I would say that jet is not very harmonic (too large a portion of the hue circle; the T template, which is the largest, spans 180 degrees), and that the spectrum is terrible.
CubeYF is also exceeding a bit 180 degrees, but looks very close to a T template rotated by 180 degrees (rotations are allowed). So perhaps I could trim it a bit? But to me it looks a lot nicer and gives me a vibe of really good mood, and reminds me of one of those beautiful central american headdresses, like Moctezuma’s crown.
 Looking at the intensity of the colorbars may help in the assessment: the third and fourth colorbars are very similar and both look perceptually linear, whereas the first and second do not.
 Quoted from Richard’s blog post: “… in the middle is a circular histogram of the colours (spectral shades) in the image, and gives an idea of how much of each colour there is. Up the left is a histogram of image brightness (lightness of colour), and up the right is a histogram of colour saturation (vibrancy)”.
 Quoted from the paper’s abstract: “Harmonic colors are sets of colors that are aesthetically pleasing in terms of human visual perception. If you are interested in this idea there is a set of slides and a video on the author’s website
In my essay I started with the analysis of the spectrum color palette, the default in some seismic interpretation softwares, using my Lightness L* profile plot and Great Pyramid of Giza test surface (see this post for background on the tests and to download the Matlab code). The profile and the pyramid are shown in the top left image and top right image in Figure 1, from the essay.
In the plot the value of L* varies with the color of each sample in the spectrum, and the line is colored accordingly. This erratic profile highlights several issues with spectrum: firstly, the change in lightness is not monotonic. For example it increases from black (L*=0) to magenta [M] then drops from magenta to blue [B], then increases again and so on. This is troublesome if spectrum is used to map elevation because it will interfere with the correct perception of relief, particularly if shading is added. Additionally, the curve gradient changes many times, indicating a nonuniform perceptual distance between samples. There are also plateaus of nearly flat L*, creating bands of constant color (a small one at the blue, and a large one at the green [G]).
The Great Pyramid has monotonically increasing elevation (in feet – easier to code) so there should be no discontinuities in the surface if the color palette is perceptual. However, clearly using the spectrum we have introduced many artificial discontinuities that are not present in the data. For the bottom row in FIgure 1 I used my new color palette, which has a nice, monotonic, compressive Lightness profile (bottom left). Using this palette the pyramid surface (bottom right) is smoothly colored, without any perceptual artifact.
This is how I created the palette: I started with RGB triplets for magenta, blue, cyan, green, and yellow (no red), which I converted to L*a*b* triplets using Colorspace transformations, a Matlab function available on the Matlab File Exchange. I modified the new L* values by fitting them to an approximately cube law L* function (this is consistent with Stevens’ power law of perception), and adjusted a* and b* values using Lab charts like the one in Figure 2 (from CIELab Color Space by Gernot Hoffmann, Department of Mechanical Engineering, University of Emden) to get 5 colors moving up the L* axis along an imaginary spiral (I actually used tracing paper). Then I interpolated to 256 samples using the same ~cube law, and finally reconverted to RGB .
There was quite a bit of trial and error involved, but I am very happy with the results. In the animations below I compare the spectrum and the new palette, which I call cubeYF, as seen in CIELab color space. I generated these animations with the method described in this post, using the 3D color inspector plugin in ImageJ:
OK, the new palette looks promising, insofar as modelling is concerned. But how would it fare using some real data? To answer this question I used a residual gravity map from my unpublished thesis in Geology at the University of Rome. I introduced this map and discussed the geological context and objectives of the geophysical study in a previous post, so please refer to that if you are curious about it. In this post I will go straight to the comparison of the color palettes; if you are unfamiliar with gravity data, try to imagine negative residuals as elevation below sea level, and positive residuals as elevation above seal level – you won’t miss out on anything.
In Figures 3 to 6 I colored the data using the above three color palettes, and grayscale as benchmark. I generated these figures using Matlab code I shared in my post Visualization tips for geoscientists: Matlab, and I presented three of them (grayscale, Spectrum, and cubeYF) at the 2012 convention of the Canadian Society of Exploration Geophysicists in Calgary (the extended abstract, which I co-authored with Steve Lynch of 3rd Science, is available here).
In Figure 3, the benchmark for the following figures, I use grayscale to represent the data, assigning increasing intensity from most negative gravity residuals in black to most positive residuals in white (as labeled next to the colorbar). Then, I used terrain slope to create shading: the higher the slope, the darker the shading that is assigned, which results in a pseudo-3D display that is very effective (please refer to Visualization tips for geoscientists: Surfer, for an explanation of the method, and Visualization tips for geoscientists: Matlab for code).
Figure 3 – Grayscale benchmark
In Figure 4 I color the pseudo-3D surface with the cubeYF rainbow. Using this color palette instead of grayscale allows viewers to appreciate smaller changes, more quickly assess differences, or conversely identify areas of similar anomaly, while at the same time preserving the peudo-3D effect. Now compare Figure 4 with Figure 5, where we use the spectrum to color the surface: this palette introduces several artefacts (sharp edges and bands of constant hue) which confuse the display and interfere with the perception of pseudo-relief, all but eliminating the effect. For Figure 6 I used Matlab’s default Jet color palette, which is better that the spectrum, and yet the relief effect is somewhat lost (due mainly to a sharp yellow edge and cyan band).
Figure 4 – cube YF rainbow
Figure 5 – Industry spectrum
Figure 6 – Matlab Jet
It looks like both spectrum and jet are poor choices when used for color representation of a surface, with the new color palette a far superior alternative. In the CSEG convention paper mentioned above (available here) Steve and I went further by showing that the spectrum not only has these perceptual artifacts and edges, but it is also very confusing for viewers with deficient color vision, a condition that occurs in about 8% of Caucasian males. We did that using computer software  to simulate how viewers with two types of deficient color vision, Deuteranopia and Tritanopia, would see the two colored surfaces, and we compare the results. In other words, we are now able to see the images as they would see them. Please refer to the paper for a full discussion on these simulation.
In here, I show in Figures 7 to 9 the Deuteranope simulations for cubeYF, spectrum, and jet, respectively. In all three simulations the hue discrimination has decreased, but while the spectrum and jet are now even more confusing, the cubeYF has preserved the relief effect.
Deuteranope Simulation of cube YF
Deuteranope Simulation of Industry spectrum
Deuteranope Simulation of Matlab Jet
That’s it for today. In my next post, to be published very shortly, you will get the palette, and a lot more.
 An alternative to the method I used would be to start directly in CIELab color space, and use a some kind of spiral *L lightness profile programmatically. For example:
– Using 3D helical curves from: http://www.mathworks.com/matlabcentral/fileexchange/25177-3d-curves
– Using Archimedes spiral
– Expanding on code by Steve Eddins at Mathworks (A path through L*a*b* color space) in this article , one could create a spiral cube lightness with something like:
%% this creates best-fit pure power law function
% Inspired by wikipedia - http://en.wikipedia.org/wiki/Lightness
%% this makes cielab real cube function spiral
radius = 50;
theta = linspace(0.6*pi, 2*pi, 256).';
a = radius * sin(theta); b = radius * cos(theta);
Lab1 = [L2, a, b]; RGB_realcube=colorspace('RGB<-Lab',(Lab1));
 The simulations are created using ImageJ, an open source image manipulation program, and the Vischeck plug-in. I later discovered Dichromacy, anther ImageJ plug-in for these simulations, which has the advantage of being an open source plugin. They can also be performed on the fly (no upload needed) using the online tool Color Oracle.
If you are interested in the topic of color palettes for scientific data, and the rainbow in particular, I would say you ought to read this 2007 IEEE visualization paper by Borland and Taylor: Rainbow Color Map (Still) Considered Harmful. It clearly and elegantly illustrates why the rainbow palette should be avoided when displaying scientific data. I like Figure 1 in the paper in particular. The illustration shows how it is easy to order perceptually a set of 4 paint chips of different gray intensity, but not at all easy to order 4 paint chips colored red, green, yellow, and blue. The author’s argument is that the rainbow colors are certainly ordered, from shorter to longer wavelengths, but they are not perceptually ordered. In this post I wanted to extend the chips example to all 7 colors in the rainbow and try to demonstrate the point in a quantitative way.
Here below is a 256-sample rainbow palette I created interpolating between the RGB values for the seven colors of the rainbow red, orange, yellow, green, blue, indigo, and violet (ROY G BIV):
On this palette I see a number of perceptual artifacts, the most notable ones being a sharp edge at the yellow and a flat zone at the green. The existence of these edges I tried to explain quantitatively in the first post of this series.
Now, to go back to the experiment, from the original RGB values for the non interpolated colors I created the 7 color chips below . Question: can you order them based on their perceived intensity?
I think if you have full color vision (more on the topic of rainbow and impaired color vision in the next section of this post) eventually you will be able to order them as I did.If not, try now below. In this new image I converted the color chips to gray chips using the values obtained in Matlab with this formula:
Give it a try, then hover with your mouse over the image to read the intensity values.
Not surprisingly, the values are not in any particular order. This reinforces the notion that although the rainbow colors are ordered by increasing wavelength (or decreasing in this case) , they are not perceptually ordered. (See this comment to my previous post). Below I rearranged the gray chips by increasing intensity.
And now I reconverted from gray to RGB colors and adjusted the distance between each pair of chips so that it is proportional to the intensity difference between the chips in the pair (I actually had to artificially change the value for green and orange so they would not overlap). That was an epiphany for me. And the name is funny too, BIV R GOY, or YOG R VIB…
I said that it was an epiphany because I realize the implications of trying to create a palette by interpolating through these colors with those distances. So I did it, and I am showing it below in the top color palette. We jumped out of the frying pan, into the fire! We went from perceptual artifacts that are inherent to the rainbow (reproduced in reverse order from blue to red to facilitate comparison as the bottom palette) to interpolation artifacts in the intensity ordered rainbow. Hopeless!
As if what I have shown in the previous section wasn’t scary enough, I took 7 squares and colored them using the same RGB values for Red, Orange, Yellow, Green, Blue, Indigo, and Violet. Then I used the Dichromacy plug-in in ImageJ to simulate how these colors would be seen by a viewer with Deuteranopia (the more common form of color vision deficiency). I then shuffled the squares in random order on a square canvas, and numbered them 1-7 in clockwise order.
Puzzle: can you pair the squares numbered 1 through 7 with the colors R though V? I will give away the obvious one, which is the yellow:
Cannot do it? For the solution just hover over the image with your mouse. If you like the animation and would like to use it on your blog, twitter, Facebook, get the GIF file version here. Please be kind enough to link it back to this post.
When I tried myself I could not solve the puzzle, and that finally convinced me that trying to fix the rainbow was a hopeless cause. Even if we could, it would still confuse a good number of people (about 8% of male have one form or the other of color vision deficiency). From the next post on I will show what I got when I tried to create a better, more perceptual rainbow from scratch.
I find the Color Use Guidelines by Cynthia Brewer (of Color Brewer fame) very well done and extremely useful. Below is a screen captures of the main page. To see an explanation of and example for each color scheme visit the interactive guidelines.