Choosing Colour Ramps for Continuous Raster Data

Classify the data first, then pick a ramp that matches it:

import matplotlib.pyplot as plt

# Sequential: a quantity with a low and a high, no special middle
ax.imshow(ndvi, cmap="YlGn", vmin=0.0, vmax=0.9)

# Diverging: a quantity with a meaningful zero
from matplotlib.colors import TwoSlopeNorm
ax.imshow(ndvi_change, cmap="RdBu", norm=TwoSlopeNorm(vmin=-0.4, vcenter=0.0, vmax=0.6))

The choice is not decorative: a ramp that is not monotonic in lightness invents structure the data does not contain. This page belongs to rendering rasters with matplotlib in Visualization, Tiling & Web Delivery.

Perceptual Uniformity, and What Breaks Without It

A perceptually uniform ramp increases steadily in lightness from one end to the other, so equal steps in data value produce equal apparent steps in colour. A rainbow ramp does not: it has bright yellow in the middle and darker colours at both ends, which makes the middle of the range look like a boundary and compresses detail at the extremes.

Lightness against position in the ramp A perceptually uniform ramp rises steadily in lightness from dark to light across its whole length. A rainbow ramp rises, peaks sharply at yellow around sixty per cent, then falls again, so two different data values map to the same apparent lightness and the peak reads as an edge. Lightness should rise once, steadily light dark uniform rainbow the yellow peak reads as a boundary A uniform ramp in swatch form: each step is the same perceptual distance from the last.

The consequence for remote sensing is concrete. An NDVI map in a rainbow ramp shows a hard line at the yellow transition that readers interpret as a real vegetation boundary, and two genuinely different values — one on each side of the peak — render at the same lightness, so a greyscale printout collapses them.

Environment & Setup

Package Version pin Used for
matplotlib >=3.8 Colour maps, norms and the rendering
numpy >=1.23 Luminance conversion for the greyscale test 
pip install "matplotlib>=3.8" "numpy>=1.23"

Complete Working Example

import matplotlib.pyplot as plt
import numpy as np
from matplotlib.colors import TwoSlopeNorm

SEQUENTIAL = {"ndvi": "YlGn", "elevation": "cividis", "reflectance": "gray",
              "temperature": "magma", "precipitation": "Blues"}
DIVERGING = {"change": "RdBu", "anomaly": "BrBG", "difference": "PuOr"}


def render(arr, extent, *, kind: str, vmin: float, vmax: float,
           diverging: bool = False, label: str = ""):
    """Render a continuous raster with a ramp appropriate to its structure."""
    if diverging:
        if not (vmin < 0 < vmax):
            raise ValueError("a diverging ramp needs values on both sides of zero")
        cmap = DIVERGING.get(kind, "RdBu")
        norm = TwoSlopeNorm(vmin=vmin, vcenter=0.0, vmax=vmax)
        kwargs = {"cmap": cmap, "norm": norm}
    else:
        kwargs = {"cmap": SEQUENTIAL.get(kind, "viridis"), "vmin": vmin, "vmax": vmax}

    fig, ax = plt.subplots(figsize=(8, 8))
    im = ax.imshow(arr, extent=extent, interpolation="nearest", **kwargs)
    ax.set_aspect("equal")
    ax.ticklabel_format(style="plain")
    fig.colorbar(im, ax=ax, shrink=0.75, pad=0.02).set_label(label or kind)
    return fig, ax


def greyscale_test(fig) -> np.ndarray:
    """Render the figure and return its luminance, for the legibility check."""
    fig.canvas.draw()
    rgba = np.asarray(fig.canvas.buffer_rgba(), dtype="float32") / 255.0
    return (0.2126 * rgba[..., 0] + 0.7152 * rgba[..., 1] + 0.0722 * rgba[..., 2])

The vmin < 0 < vmax guard is worth having because a diverging ramp applied to strictly positive data is a common and confusing mistake: half the ramp goes unused, and readers who know the convention will assume the light end means negative values.

Ramps for Specific Products

Sensible defaults by product NDVI suits a yellow to green sequential ramp. Elevation suits a perceptually uniform ramp such as cividis rather than a terrain scheme. A change image needs a diverging ramp pinned to zero. A single reflectance band is clearest in greyscale, which is also how the sensor measured it. Defaults worth starting from product ramp why NDVI YlGn, sequential more vegetation reads as more green; no false midpoint elevation cividis, sequential uniform and colour-blind safe; terrain ramps imply land cover NDVI change RdBu, diverging gain and loss must be distinguishable at a glance single band gray colour adds nothing to a single radiometric quantity Deviating from these is fine — but do it because the data asks, not because the default looked dull.

The greyscale row deserves a defence, since a grey image looks unfinished next to a coloured one. For a single reflectance band there is exactly one quantity, and greyscale represents it with maximum fidelity and no implied categories. Adding a colour map to it is decoration that costs interpretability, and it is the reason so many published band images have an inexplicable orange cast.

Where colour genuinely earns its place is in distinguishing a quantity from its surroundings — an NDVI layer over a true-colour basemap, for instance — or in making a diverging quantity’s sign readable at a glance. Those are the situations the ramps above are for.

Verification

import numpy as np

lum = greyscale_test(fig)
inner = lum[80:-80, 80:-80]           # crop away the axes and colour bar
contrast = float(inner.max() - inner.min())
assert contrast > 0.35, "structure disappears in greyscale — the ramp is not uniform"
print(f"greyscale contrast: {contrast:.2f}")
The greyscale test, side by side A uniform ramp converted to luminance keeps its gradient from dark to light, so every value remains distinguishable. A rainbow ramp converted to luminance shows the same mid-grey at two different data values, and the structure collapses into repeated bands. uniform ramp rainbow ramp five distinct greys — nothing lost the ends collapse onto each other Top row: the rendered colours. Bottom row: their luminance.

Running the greyscale test in code rather than by eye makes it a gate rather than an intention. A perceptually uniform ramp applied to data with real structure gives a luminance range well above 0.5; a rainbow ramp on the same data typically gives under 0.3, because its light and dark ends both map to mid-grey.

A second check worth doing once per project is to simulate a colour vision deficiency on a rendered figure, which several small libraries do in a line. Around one in twelve men has some form of red-green deficiency, and a ramp that fails for them fails for a meaningful share of any audience.

Common Errors

The diverging ramp’s neutral colour is not at zero

A plain vmin/vmax was used instead of a norm. Use TwoSlopeNorm with vcenter=0.0, and accept the asymmetric colour bar that results.

The map looks banded

The colour map was given too few levels, or BoundaryNorm was applied to continuous data. Use the continuous ramp directly unless the data really is categorical.

Two figures of the same product look different

Different vmin/vmax were derived per scene. Fix the limits for the collection, exactly as for a composite stretch.

Frequently Asked Questions

Q: What is wrong with a rainbow colour ramp? It is not perceptually uniform, so equal steps in value look like very different steps in colour, and the sharp yellow-to-green transition creates an apparent boundary where the data is smooth. It also collapses for viewers with a colour vision deficiency, who see much of it as one band.

Q: How do I test a ramp quickly? Render the image, convert it to luminance, and look at it. If the structure is still legible, the ramp is monotonic in lightness and will survive printing, projection and colour-blind viewers. If it dissolves into bands, choose another ramp.

Q: Should elevation use a terrain colour scheme? Only when the green-to-brown convention genuinely helps a reader, and never for a quantity that is not elevation. Terrain ramps are not perceptually uniform and their colours imply land cover, so using one for rainfall or temperature actively misleads.

Q: Does the same advice apply to tile services? Yes, and it matters more there because a tile URL’s colour map parameter is often chosen once and then used by everyone. Pick the ramp with the same care as for a figure, record it next to the product, and use the identical name in both places so the map and the report agree.