Lesson 04: Colour and Accessibility

Colour is the element of a figure most likely to cause problems for readers and most likely to be handled carelessly. This lesson covers how to choose palettes that are accurate, accessible, and still visually effective.

Acknowledgement

The discussion of palette choice in this lesson draws on Ciaran O'Hare's HowToMakeAPlot, which contains practical examples of colourblind simulations and palette comparisons for physics figures.

Why colour choice matters

Roughly 8 % of men and 0.5 % of women have some form of colour vision deficiency (CVD), most commonly red-green. On top of that, many readers will encounter your figure in a printed PDF, a black-and-white photocopy, or a low-contrast projection.

A figure that relies on colour as the sole distinguishing feature excludes a significant fraction of your audience. The fix is not to abandon colour. It is to use colour correctly.

Types of colour palette

Choose the palette type to match the type of data:

Type When to use Examples
Qualitative Distinct categories with no ordering Model A vs B vs C, different experiments
Sequential Ordered data from low to high Flux intensity, temperature, probability
Diverging Data diverging from a meaningful midpoint Residuals, signed differences, correlation

Using a sequential palette for categorical data is a common mistake, as the implied ordering misleads the reader.

Qualitative palettes

For line plots and markers comparing distinct categories, use a qualitative palette from ColorBrewer.

The Nestling style uses Dark2, a six-colour qualitative palette that is:

  • Distinguishable by the most common CVD types (deuteranopia, protanopia, tritanopia).
  • Distinguishable in greyscale when combined with different line styles.
  • Distinguishable on screen and in print.

Dark2 hex codes:

DARK2 = ["#1b9e77", "#d95f02", "#7570b3", "#e7298a", "#66a61e", "#e6ab02"]

These are already set as the default colour cycle in nestling.mplstyle.

Other well-tested qualitative palettes from ColorBrewer:

  • Set1: high-contrast, slightly garish; works when CVD is not a concern.
  • Set2: softer, but less distinguishable in greyscale.
  • Paired: six pairs of light/dark variants; good for showing related quantities.

Explore all options interactively at colorbrewer2.org. The site lets you filter by CVD-safe and print-safe status.

Sequential and diverging colormaps

For heatmaps, 2D histograms, and scatter plots coloured by a continuous quantity, use a perceptually uniform colormap.

Why "jet" is wrong

The default matplotlib colormap for many years was jet (rainbow). It has two fatal flaws:

  1. Not perceptually uniform. Equal steps in data produce unequal steps in perceived brightness, creating false features and hiding real ones.
  2. Not greyscale-friendly. Distant hues in jet map to the same grey value, making features invisible in greyscale.
Colormap Character Good for
viridis Blue → green → yellow General purpose, default choice
cividis Blue → yellow, CVD-optimised When CVD accessibility is critical
plasma Purple → orange → yellow High-contrast alternative to viridis
inferno Black → red → yellow Dark backgrounds, signal detection

All four are perceptually uniform and CVD-safe.

fig, ax = plt.subplots()
sc = ax.scatter(x, y, c=z, cmap="viridis")
plt.colorbar(sc, ax=ax, label=r"$z$")
Colormap Character Good for
RdBu_r Red → white → blue Signed residuals, correlations
coolwarm Blue → white → red Same, with more saturation
seismic Same family Geophysics convention

For diverging maps, always set vmin and vmax symmetrically about the midpoint:

import numpy as np
vmax = np.abs(residuals).max()
sc = ax.scatter(x, y, c=residuals, cmap="RdBu_r", vmin=-vmax, vmax=vmax)

Combining colour with other channels

Because colour alone is unreliable, always pair it with a redundant encoding:

import matplotlib.pyplot as plt
import numpy as np

E = np.logspace(0, 6, 300)

fig, ax = plt.subplots()
ax.loglog(E, 1e-18 * (E / 1e5) ** -2.7,
          color="C0", linestyle="-",  label="Model A")
ax.loglog(E, 1e-18 * (E / 1e5) ** -2.0,
          color="C1", linestyle="--", label="Model B")
ax.loglog(E, 1e-18 * (E / 1e5) ** -2.4,
          color="C2", linestyle=":",  label="Model C")
ax.legend()

Here colour AND line style distinguish the curves. The figure is still readable when printed in black and white.

For scatter plots, pair colour with marker shape:

ax.scatter(x_a, y_a, c="C0", marker="o", label="Sample A")
ax.scatter(x_b, y_b, c="C1", marker="s", label="Sample B")
ax.scatter(x_c, y_c, c="C2", marker="^", label="Sample C")

Testing your figure for accessibility

Greyscale test

Convert your saved PDF or PNG to greyscale (in Preview on macOS, or any image editor) and check whether all curves and regions remain distinguishable. If they do not, add line style or marker differentiation.

CVD simulation

Several online tools simulate how your figure appears with common colour vision deficiencies:

The ColorBrewer site also labels palettes as "colorblind safe". Filter by this when choosing.

matplotlib's built-in greyscale preview

A quick in-notebook check:

import matplotlib.pyplot as plt
import matplotlib.cm as cm

fig.savefig("check.png")
# Reload and desaturate
img = plt.imread("check.png")
grey = img @ [0.2126, 0.7152, 0.0722, 0]   # luminance weights
plt.imsave("check_grey.png", grey, cmap="grey")

Colour in context: a few more rules

  • Avoid red and green together for any information-carrying distinction. This is the most common CVD combination.
  • Don't use colour to encode the same information as position. If the x-axis already distinguishes the curves, adding colour is redundant and distracting.
  • Use transparency (alpha) with care. Overlapping transparent regions can create new colours that were not in your palette and that are hard to predict under CVD.
  • Match your colour choices to the journal. Some journals have house styles or restrictions on colour figures (especially print journals with per-colour-page charges).

Summary

Do Don't
Use Dark2 or another CVD-safe palette from ColorBrewer Use the default matplotlib colour cycle for publications
Use viridis / cividis for continuous data Use jet or rainbow
Pair colour with line style or marker shape Rely on colour as the only distinguishing feature
Test in greyscale and with a CVD simulator Assume your monitor is representative
Set symmetric vmin/vmax for diverging maps Let matplotlib choose limits automatically for residuals

Resources