Colour & accessibility for clarity

Colour is the most attention-grabbing encoding channel in data visualisation — and the most commonly misused one. The core mistake is treating colour as decoration: picking shades because they look good, or because the default matplotlib cycle happened to produce them. Colour should always encode a variable or highlight a single series. When it does neither, it adds visual noise without adding information.

There is a second problem layered on top: roughly 8% of men (and about 0.5% of women) have some form of colour vision deficiency (CVD) — the most common being red-green confusion. A chart that looks perfectly clear to you may be unreadable to one person in every twelve in the room.

The fix for both problems is the same: choose palettes that match your data type, and never rely on colour alone to convey meaning.

The three palette families

Every colour scale you will ever need falls into one of three families. Matching the family to your data type is not a style preference — it is a correctness question.

Sequential — ordered magnitude

Use a sequential palette when your data runs from low to high along a single direction: population density, temperature in one season, model confidence score, time elapsed. The palette should go from a light, low-salience colour to a dark, high-salience one (or vice-versa). The eye naturally reads darker-as-more.

Good choices: viridis, plasma, cividis, magma (all perceptually uniform and CVD-safe). Avoid the classic jet or rainbow — more on that shortly.

Diverging — deviation from a midpoint

Use a diverging palette when your data has a meaningful midpoint — a value where the interpretation switches direction. Classic examples: temperature anomaly from a baseline year (negative = cooler, positive = warmer); profit margin around zero; model error above and below a target.

The midpoint should map to a neutral, low-salience colour (white or light grey). Each arm should go toward a distinct, saturated hue.

Good choices: coolwarm, RdBu, PuOr. Avoid using red-green diverging scales (a large fraction of CVD viewers cannot distinguish them).

Qualitative — unordered categories

Use a qualitative palette when your categories have no natural order: product lines, geographic regions, species, model architectures. Here you want hues that are as distinct as possible so the eye can separate them — not a gradient, which implies an order that does not exist.

Good choices: tab10, Set2, Paired. Keep the number of categories below about eight — beyond that, human colour discrimination breaks down and you should switch to labels, patterns, or facets instead.

Colour encodes, not decorates

The key principle to internalize: colour is an encoding channel. In the same way that x-position encodes the independent variable and y-position encodes the response, colour encodes a third variable — or, in the “gray + one accent” focus pattern from the annotation lesson, it highlights the single series that carries the story.

What colour must not do is decorate. Giving each bar in a bar chart a different colour when they all represent the same variable — sales in different months — adds zero information and forces the viewer to build a legend mapping they will immediately discard. It also makes the chart nearly impossible for a CVD viewer to interpret consistently.

Rule of thumb: before you assign a colour, ask “what variable does this colour represent?” If the answer is “nothing — it just looks nice,” use a single neutral colour or the gray-plus-accent pattern instead.

Accessibility: colour vision deficiency

Approximately 8% of men have colour vision deficiency. The most common form is deuteranopia (green-blind) or protanopia (red-blind), which collapse the red-green channel. This means a red-green diverging scale, or a line chart with one red and one green line, is effectively a single-colour chart for a meaningful fraction of your audience.

Three concrete rules:

1. Use CVD-safe palettes by default. viridis and cividis were explicitly designed to be distinguishable under deuteranopia, protanopia, and tritanopia, and they print well in greyscale too.

2. Never rely on colour alone. Add direct labels to lines, use different marker shapes for scatter series, annotate the interesting bars with their values. A chart that is readable in greyscale is readable for everyone.

3. Ensure sufficient contrast. Light text on a light background, or thin lines in pale colours, fail for CVD viewers and for anyone in a bright room. Matplotlib’s default line width and font sizes are often too small for presentation contexts — increase them.

The gray + accent focus principle

The lesson on annotation and focus introduced the gray + one accent pattern: render all series in a neutral grey and use a single saturated colour only on the series you want the viewer to notice. This is the highest-contrast, most accessible version of colour-as-highlight. It requires no legend, works in greyscale, and is trivially interpretable by CVD viewers because there is no colour comparison to make — there is only “the one coloured thing” and “everything else.”

Use this pattern whenever you are telling a story about one series relative to a field — the best-performing product, the anomalous region, the model that overfit.

Runnable example: viridis heatmap with direct labels

The code below builds a small heatmap using the viridis colormap and annotates every cell with its value, so the chart is interpretable without any colour discrimination at all.

Notice what the direct labels on the right panel do: you can remove the colour entirely and the information is still complete. That is the goal. The colour adds a quick visual gradient for viewers who can use it; the labels carry the meaning for everyone else.

Quick check