Surf-plotting scattered data in Matlab (or: Delaunay interpolation without a grid)

Matlab has a number of methods for interpolating data, both for data that is sampled on a regular grid and for data that is “scattered”, or randomly distributed.

Surface plotting

Plotting surfaces over grid points is easy using Matlab’s surf command, and interpolation of that data to get smoother plots is straightforward. For example,

[x,y,z] = peaks(10);

will plot:


Generally I recommend avoiding 3D plots, so in 2D (view(2)):


The variables x and y are 10×10 matrices defined by (the equivalent of) [x,y]=meshgrid(linspace(-3,3,10)), and z is the value at each point in (x,y) space.

If you have sampled data with non-uniform spacing, however, it’s not as obvious how to plot that data. In this case your data would be organised with x, y, and z as column vectors with each measurement per row. The structure of the surf command simply doesn’t handle data in this form, since the data isn’t organised in a way that allows adjacent points to be connected.

Matlab provides commands for analysing this data using Delaunay Triangulation, for which it also provides an analogue to surf. Here’s an example:

N = 15;
rng(13); % repeatable random values
x = randn(1,N); % X & Y coords
y = randn(1,N);
z = randn(1,N); % "value" at each coordinate

After a little digging, you’ll find that the easiest way to plot this data as a surface is:


Seems slightly redundant, but easy enough.

Surf plots with coarse grids

The problem is that surface plots are poor at visualising data over coarse meshes, since it’s their corners which define the values. This can be seen in the following example, using peaks to plot a surface:


The first row shows plain surf. It can be seen that the colours of each “patch” don’t do a good job of representing their height (or “value”). In the second row, Matlab’s interpolating shader is used to “improve” this. It’s an improvement, but distortions are obvious where there are large changes in both directions. (Colours are smeared along diagonals, basically.)

Finally, in the third row an interpolation function is used to generate more data points. This is quite a coarse interpolation, but couldn’t go finer without the edges taking over in the 3D plot. The plot looks far more convincing. Gridded interpolation can be done in a number of ways; for this example I used

F = griddedInterpolant(x',y',z');
[x2,y2] = ndgrid(linspace(-3,3,20));

(The transpose being necessary from coordinates output from peaks. Good old Matlab.)

Surf plots with coarse scattered data

But what about scattered data? While Matlab does provide scatteredInterpolant, this form is only really convenient if you want to interpolate onto a regular grid. What if you just want to improve the look of your overly-coarse surface plot? Here’s the example using trisurf described above, with and without interpolated shading (shading interp):


Note that without the interpolated shading, the plot is basically unusable. But even with interpolated shading, there’s significant distortion along certain diagonals. I’ve left out the edge lines in the bottom-right graph to emphasise this. If we resampled this function over a regular grid using scatteredInterpolant, we’d end up with points outside the original region, which might not always be appropriate. We could then delete the outsiders, but that would require additional processing. Wouldn’t it be nice to simply resample within the original?

Delaunay interpolation

To do this requires a couple steps. First, extract the Delaunay Triangularisation of our points:

DT = delaunay(x,y);

This function returns a list of 3xM indices into x and y, where M is the number of triangles. So if we extract the coordinates for each triangle and average them, we obtain the mid-point of each, including an interpolated value:

x2 = mean(x(DT),2);
y2 = mean(y(DT),2);
z2 = mean(z(DT),2);

This figure shows the grid that results after interpolating once and re-applying the Delaunay Triangularisation: (black before, red after)


Note that the Delaunay algorithm doesn’t simply divide up the triangles into smaller components. We can then perform this iteratively to interpolate to an arbitrary level of detail:


And after removing the grid lines and increasing the order, the results look pretty nice:


Interpolation will never solve the problem of under-sampling if there’s information missing, but it definitely helps to draw more pleasing figures.

Although pretty basic, I’ve uploaded this code to the Matlab file exchange under the name delaunay_surf, and added it to my generic Matlab repository on Github: It’s pretty basic:

p = delaunay_surf(x,y,z,'N',4);

with a few additional options. Check out the example file for more, er, examples.

Happy plotting!


Git support in Matlab

I have a love/hate relationship with Matlab, the details of which I probably obliquely mention here from time to time. The biggest reasons that I stick with Matlab are (a) I have to teach my students to use it, and (b) for all its warts, its IDE is quite nice for code authoring/debugging, exploring data sets, and even drawing somewhat sophisticated 3D diagrams.

I’ve just installed Matlab 2015a on my laptop, and am pretty happy to also see that Matlab now supports Git: (this appears to have been added in 2014b, so I’m a little behind the times)

Screen Shot 2015-05-09 at 1.32.57 pm

I usually use a lightweight GUI to perform simple Git operations (GitX on Mac) and the command line for anything more sophisticated. Most common Git operations can be performed from a clunky interface that involves right-clicking on a file, selecting “Source Control” from the menu, and then selecting from a list. Rather like TortoiseGit on Windows, I guess. There are a few impedance mismatches, such as the term “Compare to Ancestor”, which is Matlab’s version of performing “git diff” to see what you’ve changed in the current file. I think trying to abstract multiple version control systems under a consistent set of Matlab commands is probably a recipe for disaster, long term, but I can understand the motivation.

I like that the Git status is indicated in the file listing, but I suspect I’ll rarely, if ever, use the Matlab interface to actually perform source control. I’m a little disappointed that there isn’t a better user interface than delving into contextual menus and receiving little more than plain info in pop-up windows. For example, what I’d like to see would be to double click a “changed” icon (blue square in the above screenshot), which would show the current differences from the last commit, and an integrated text field and nice big button to commit the changes. Not as easy, but more likely to get keep people within Matlab.

What I’m really hoping Git support in Matlab makes it possible to slowly introduce my undergraduate students to Git (engineers learning source control is a subject I hope to discuss another time), and perhaps more importantly I hope it will help facilitate collaboration with research students that work on my code. Fingers crossed.