Colourful Ecology Part 1: Selecting colours for scientific figures from an image using community ecology methods

[This article was first published on rstats - MetaEvoPhyloEcoOmics, and kindly contributed to R-bloggers]. (You can report issue about the content on this page here)
Want to share your content on R-bloggers? click here if you have a blog, or here if you don't.

It used to be that I would make all my figures in black and white. It was just simpler that way, since most print journals charged extra money to print colour figures, and I did not have any money to spare. But with the move in science publishing from mainly print journals to mainly online (rich) content, colour is no longer a restriction for scientific figures. This is good news because colour can not only make your figures look great, they can help you pack more information into a scientific visualization. But the opposite side of this coin is that if colours are badly chosen, they can make your figure much much worse. This then leads to the question I used to not have to answer: what colours should I use?

I have found several R packages really useful in this regard. RColorBrewer is a great package which has a small number of palettes to choose from, but all of them are pre-selected so that their colours will look good together. Recently I have become a fan of the wesanderson package by Karthik Ram, which makes available palettes drawn from the wonderfully quirky film canon of the director Wes Anderson. Want your figure to conjure the feeling of “Fantastic Mister Fox”? Or a my personal favourite: “Rushmore” (does that age me?). Just install wesanderson and grab the palette! Jo-Fai Chow also has an excellent blog post with links to more interesting color packages, some of which I have not checked out yet.

But what if I want to get colours that are inspired by any arbitrary image I like, or a non-arbitrary image that may be related to the subject of the figure? I started thinking about this when I had an exchange on twitter with Jo-fai Chow and Karthik Ram, e.g.

Soon thereafter, Jo-fai had already made a basic function, which I will make use of below:

Since then, Jo-fai has posted a blog post, describing his function and some other work. It’s a great post, check it out!

Here is an example of using Jo-fai’s function using one of my favourite images of one of my favourite animals, which I will use throughout this post. The photo was taken by Jurgen Otto, a very amazing photographer.

A posing PeacockSpider

## install Jo-Fai Chow's package
## extract 5 colours
pal_spider1 <- extract_colours("",6)
## plot extracted palette
par(mfrow = c(1,2))
pie(rep(1, 6), col = pal_spider1, main = "Palette based on Peacock Spider", cex.main=0.8)
hist(Nile, breaks = 7, col = pal_spider1, main = "Palette based on Peacock Spider", cex.main=0.8)

A six colour palette based on a image of a Peacock Spider, generated using extract_colours()

Jo-Fai Chow's method uses k-means clustering to extract a colour palette. This works by splitting the pixels into k groups based on their similarity in RGB colour space. Once it has found these k clusters, extract_colours takes the average colour of each cluster, and returns these cluster averages as the palette. I think this is a great way to do it, and offers up a number of ways to tweak the results. For one, there are literally hundreds of methods of performing clustering (k-means is one of the most popular).

However, once we extract colours, how can we select colours useful for scientific figures?

What do we want in scientific figure colours?

There is generally two types of palettes we might want to use in a scientific figure

  • A palette of contrasting colours, to effectively distinguish between different classes
  • A palette consisting of a colour gradient, to effectively represent differing levels of a continuous variable

In this post, I am mostly going to play around with ideas about how to get a palette of the first kind from an image of something such as our little spider friend above. The second kind I will cover in a subsequent blog post.

I do this by applying ideas from community ecology to colours. The basic idea is to treat the components of colour (such as Red, Green, and Blue if we are using RGB colorspace) as traits that each pixel in an image has (think of the pixels as species in the community of the image). We then want to assembly a reduced community of pixels which maximizes the average distance in their colour traits. Ecologists often measure the traits of organisms and ask whether communities are dispersed in trait-space or clustered, and so have developed a number of metrics to measure trait dispersion. Here, I simply apply one of these metrics to colour palettes, and maximize it in a simple brute-force way.

First a quick function to extract the colours from an online or local image. This code is based on the extract_colours() function in the rPlotter package written by Jo-fai Chow. It uses the EBImage package to load the image.

colour_extracter <- function(img_file = "", rsize=100) {


  ## Read Image
  img <- readImage(img_file)

  ## Resize Image (make it smaller so the remaining tasks run faster)  
  if (max(dim(img)[1:2]) > rsize) {
    if (dim(img)[1] > dim(img)[2]) {
      img <- resize(img, w = rsize)
    } else {
      img <- resize(img, h = rsize)


First thing, it would be cool to visualize the distribution of colours
in an image to get an idea of where the major colour gradients are.

This code requires the packages EBImage, reshape2 and lattice.

## extract the colours!
## reshape
img_melt <- melt(img)
img_rgb <- reshape(img_melt, timevar = "Var3", idvar = c("Var1", "Var2"), direction = "wide")
## get just the colours
rgbs <- rgb(spider.cols)
## save abundances for later
rgb_dat <- table(rgbs)
## get rid of visually indistinguishable colours
spider.cols <-
colnames(spider.cols) <- c("Red","Green","Blue")
rownames(spider.cols) <- names(rgb_dat)
## plot a spinning 3D plot of the RGB values

for (i in seq(6,360,by=6)) print(cloud(Blue~Red*Green, spider.cols, pch=19, col=rgb(spider.cols), screen=list(z=i,x=-60), perspective=TRUE, scales=list(draw=FALSE), xlab="", ylab="", zlab="", zoom=1, par.settings = list(axis.line = list(col = "transparent"))))

An animated gif showing Peacock Spider colours in Red-Green-Blue space

This animation gives a nice 'fingerprint' of the images colours. Running down the middle—along the 1:1:1 line—are colours with equal red, green, and blue representation, which are colours with low saturation. These tend to be covered up by colour around the outside, so luckily they tend to be the less interesting colours. The brighter, higher saturation colours ring around the low saturation core, and tend to form 'wings' of dominant colours that loop out and along the light to dark axis (the 1:1:1 line). Here we can see one very large wing of red and orangish colours, and a slightly smaller blue wing. Peaking out the bottom is a smaller green wing.

The question now is how to choose colours from this spectrum that are visually distinctive for figures? Well—to start— if you are plotting on a white background, you will want to get rid of very light colours. Since I often use ggplot and its default light grey background, I don't mind light colours. But I do want to get rid of low saturation colours, both because these are 'greyish' and don't stand out against a grey background, but also because they tend to be visually less appealing when you are plotting in full colour.

One way to get rid of low saturation colours would be to remove any colours whose RGB values that were too close to the 1:1:1 line. But an easier way is to convert your RGB colours into HSV (Hue, Saturation, and Value) colours, and then remove any colours that are below a threshold saturation.

## convert colours to hsv scale
hsv.col<-t(rgb2hsv(t(spider.cols), maxColorValue = 1))
## remove colours with saturation less than 25%<-spider.cols[hsv.col[,"s"]>0.25,]
## plot the result
cloud(Blue~Red*Green,, pch=19, col=rgb(, screen=list(z=-45,x=-60), perspective=TRUE, scales = list(col = "black"), par.settings = list(axis.line = list(col = "transparent")))

Colours in Red-Green-Blue space after removing low saturation colours

That's better. Notice the one outlier pixel which is apparent in the above figure: one very bright green out on the upper-right. Otherwise most of the greens in this image are much duller.

Now, I will apply a technique used in ecology to visualize the similarity of organisms in their traits. First we calculate the distance of species in trait-space (in this case, the distance of the colours in colour-space), and then cluster the species (or colours) into a tree-like structure. Think of it as a phylogenetic tree of colours! First I will try an heirarchical clustering method known as UPGMA, a fast method of tree construction.

## generate a euclidean distance matrix from all remaining colours (note: with rsize = 100, this creates a matrix with 38,715,600 elements which takes up about 300mb, be careful if using even greater rsize!)
## use heirarchical scaling to generate a dendrogram of the colours (using UPGMA: "average")
col_clust<-hclust(col_dists, "average")
## convert dendrogram into a phylogenetic tree 
## plot tree!
par(mfrow = c(1,1))
col_tree$tip.label <- rep("_______", length(col_tree$tip.label))
plot(col_tree, show.tip.label=TRUE, type ="fan", no.margin=TRUE, tip.color=rgb(, underscore=TRUE, cex=1.8)

Rainbow colour around a dendrogram, showing how the colour cluster in an heirarchical clustering framework

I think the heirarchical clustering causes issues because it forces the resulting distances to be ultrametric (every tree-tip is an equal distance from the tree-root), which probably does not represent colours very well. Instead, let's try neighbour-joining, which simply creates a tree by joining nearest neighbours.

## try using neighbour-joining: this takes much longer!
col_tree2 <- nj(col_dists)
col_tree2$tip.label <- rep("_________", length(col_tree2$tip.label))
plot(col_tree2, show.tip.label=TRUE, type ="fan", no.margin=TRUE, tip.color=rgb(, underscore=TRUE)

Neighbour-joing tree of colours

## Another way of visualizing that
plot(col_tree2, show.tip.label=FALSE, type ="unrooted", no.margin=TRUE)
tiplabels(pch=19, col = rgb(, cex=0.7)

Another way of showing the same neighbour-joing tree. This time it is unrooted (as it probably should be)

Looks good! But we immediately start to see a problem by visualizing the colours like this. For one, the blueish colours have large branch-lengths, yet they appear to all be very similar, at least to my eye. It turns out that human colour perception is quite biased with regards to certain hues. For example blue hues are difficult to distinguish by eye, such that some colours that are very different in RGB space can appear quite similar. On the other hand, red and orangish hues can appear quite different, despite being very close in RGB space. To account for this, other colour-spaces have been developed which attempt to get closer to how humans percieve colours. An example of this is the CIELAB colour-space, or L*A*B* for short. This space consists of three dimensions. The first, known as L* corresponds to luminance, and is a very good measure of how humans perceive light vs. dark. L* ranges between 0 and 100, with high values being light colours. The other two elements are a* and b*, which are the position of the colour on the red/magenta to green scale (-100 to 100), and the blue to yellow scale (-100 to 100), respectively. I therefore decided to transform the image's colours to CIELAB space using the colorspace package before analyzing further:

spider.cols.LAB <- as(RGB(as.matrix(, "LAB")
spider.cols.LAB <-
cloud(L~A*B, spider.cols.LAB, pch=19, col=rownames(spider.cols.LAB), screen=list(z=-45,x=-60), perspective=TRUE, scales = list(col = "black"), par.settings = list(axis.line = list(col = "transparent")))

Peacock Spider colours in L*a*b* space

col_dists_LAB <- dist(spider.cols.LAB)
col_tree_LAB <- nj(col_dists_LAB)
col_tree_LAB$tip.label <- rep("_________", length(col_tree_LAB$tip.label))
plot(col_tree_LAB, show.tip.label=TRUE, type ="fan", no.margin=TRUE, tip.color=rgb(,

A colour-tree in CIELAB colour-space

## compare with RGB
par(mfrow = c(1,2), mar=c(2,2,2,2))
plot(col_tree2, show.tip.label=FALSE, type ="unrooted", no.margin=FALSE, main = "RGB")
tiplabels(pch=19, col = rgb(, cex=0.7)
plot(col_tree_LAB, show.tip.label=FALSE, type ="unrooted", no.margin=FALSE, main = "L*A*B*")
tiplabels(pch=19, col = rgb(, cex=0.7)

Comparing the RGB and CIELAB colour-spaces

par(mfrow = c(1,1))

In the new colour-space, we can see that the blue hues now take up a much smaller volume, reflecting human's poor ability to distinguish these hues. On the other hand the red and orange hues have expanded to fill more space.

Using a community phylogenetics metric to choose a palette with visually distinctive colours

Putting all of this together, I put together a function to select visually distinctive colours from a palette of colours. It does the following:

  1. Gets rid of colours of low saturation and/or low or high luminosity.
  2. Converts colours to CIELAB space to align them more with human perception
  3. Maximizes the mean distance between a set of colours taken from the full palette. I call this measure Mean Perceptual Distance (MPD), which is based on Mean Phylogenetic Distance (Webb 2000) from ecology.

Mean Phylogenetic Distance is simply the mean pairwise distance between all species in a community, the distance being measured by how far apart two species are on a phylogenetic tree. Here, I simply replace the phylogenetic distances with the Euclidean distance between colours in CIELAB space, so that Mean Perceptual Distance is the mean distance between all pairs of colours in a sub-palette.

Currently the maximization is done using a brute force approach, suitable for reasonably sized input palettes, which simply generates a large number of random palettes, calculated MPD, and then returns the palette with the higher MPD. Obviously, this does not guarantee the best palette is found, but will usually find a good solution as long at the number of reps is high enough (10,000 by default).

Here is the function:

Now let's try it out! First I extract a largish palette using Jo-fai Chow's excellent function (256 colours to start).

## For now my function is in my own fork of woobe's rPlotter repository
bigpal <- extract_colours("", 256, 200)
## get 5 best 6 colour palettes for MPD
newpal <- mpd_select_colours(bigpal, sat.thresh = 0.25, dark.thresh = 0.1 , ncolours = 6, nreturn = 5)
## plot palettes
spider <- readImage("")
h <- split.screen(c(1,2))
par(mar = c(0,0,0,0)) # set zero margins on all 4 sides
plot(x = NULL, y = NULL, xlim = c(0,1500), ylim = c(0,1500), pch = '',
    xaxt = 'n', yaxt = 'n', xlab = '', ylab = '', xaxs = 'i', yaxs = 'i',
        bty = 'n', asp=1) # plot empty figure
rasterImage(spider, xleft = 0, ybottom = 0, xright = 1500, ytop = 1500) # plot jpeg
h <- split.screen(c(5,1))
for (i in 1:length(newpal)) {
  pie(rep(1, 6), col = newpal[[i]])

The five most visually distinctive palettes found for the Peacock Spider image


Note that the five palettes are in no particular order at the moment (i.e. the one at the top is not necessarily the best of the five).

Here are a few more examples, starting with a photo taken (credit: Anna Simonsen) during our field work in the Kimberley. And then a poster from one of my favourite movies :).

Five distinctive palettes for a photo of Kimberley field work

Future Directions

Some cool things to try next:

  • Take into account colour-blind friendly colours.
  • Maximize MPD using a more elegant approach. Perhaps a numerical optimization method which supports discrete parameters, such as simulated annealing. I have considered actually simulating competition amongst colours, treating them as species whose similarity in colour components determines their competition coefficients. Basically running a Lotka-Volterra competition model until the number of 'surviving' colours equals the desired number of colours. This should be pretty easy to do. Look for another post on this soon.
  • Utilize information on the abundance of different colours in the image. The MPD metric in ecology can be abundance-weighted, so that very abundant colours can contribute more to the mean distance than rarer colours.

I will also write up some of the colour distribution visualization methods into functions and make them available in the near future. In part 2, I will look at ways to try and extract a gradient of colour from an image.


Webb, CO. 2000. “Exploring the Phylogenetic Structure of Ecological
Communities: An Example for Rain Forest Trees.” The American
156 (2) (August): 145–155.

To leave a comment for the author, please follow the link and comment on their blog: rstats - MetaEvoPhyloEcoOmics. offers daily e-mail updates about R news and tutorials about learning R and many other topics. Click here if you're looking to post or find an R/data-science job.
Want to share your content on R-bloggers? click here if you have a blog, or here if you don't.

Never miss an update!
Subscribe to R-bloggers to receive
e-mails with the latest R posts.
(You will not see this message again.)

Click here to close (This popup will not appear again)