Skip to content

Wheat kernels

210 wheat kernels from three varieties — Kama, Rosa and Canadian — X-rayed and measured on seven geometric properties: area, perimeter, compactness, kernel length and width, asymmetry, and the length of the groove.

Projected with DruidJS: the seven columns are standardised, then new druid.UMAP(Z, { seed: 11 }). node scripts/build-dataset.mjs openml:1499 UMAP 11 regenerates it.

210 glyphs · cell 0.8 px → glyph 0.30 · 4 overlapping pairs

Reading a kernel

area
overall size
length & width
how slender or plump
compactness
how full the outline is
asymmetry
how far the tip leans
groove length
the crease down the middle
perimeter
outline weight — a ragged edge

Each measurement drives the property of the kernel it actually describes, so the shape is legible without a key — no spoke-counting.

Varieties

  • Kama
  • Rosa
  • Canadian

Rosa kernels are simply bigger; Canadian ones are small and pale. The varieties differ more in size than in shape, which the glyphs show and the projection does not.

Each kernel is drawn as a kernel. Rather than putting the seven numbers on spokes, each one drives the property of the grain it actually describes:

measurementwhat it draws
areaoverall size
length & widthhow slender or plump
compactnesshow full the outline is between base and tip
asymmetryhow far the tip leans off the axis
groove lengththe crease down the middle
perimeteroutline weight — a long perimeter for the area means a ragged edge

The varieties differ mostly in size: Rosa kernels are simply bigger, Canadian ones small and pale. That is a case where the glyphs earn their keep in an unusual way — the projection separates the varieties perfectly well, but only the glyphs tell you the separation is largely a size effect rather than a shape one.

Compare the same data on the general-purpose star glyph:

210 glyphs · cell 0.8 px → glyph 0.40 · 6 overlapping pairs

Reading a glyph

One spoke per measurement, clockwise from the top in the order below. The outline is the item's profile. Each is scaled against that column's range across the whole dataset, so two glyphs are comparable.

  1. 1area
  2. 2perimeter
  3. 3compactness
  4. 4kernel length
  5. 5kernel width
  6. 6asymmetry
  7. 7groove length

Classes

  • Kama
  • Rosa
  • Canadian

The item closest to each class's average profile — so it looks like the class rather than like an outlier.

The star is more precise — you can read a single measurement off a spoke, which you cannot do from a kernel’s outline. But it is slower to take in, and at small sizes the spokes blur into a rough disc while the kernels stay recognisably plump or slender. When the data describes a physical thing, drawing that thing tends to survive shrinking better, because the reader already knows what they are looking at.

Małgorzata Charytanowicz, Jerzy Niewczas, Piotr Kulczycki, Piotr A. Kowalski, Szymon Łukasik and Sławomir Żak Complete Gradient Clustering Algorithm for Features Analysis of X-ray Images Information Technologies in Biomedicine 15–24, 2010. doi:10.1007/978-3-642-13105-9_2

The seeds dataset, OpenML id 1499.