Automating leaf measurement

Undergraduate research at Kent State University, advised by Dr. Andrea Case. It sat inside the NSF programme BEE (DEB-2015606), which uses Lobelia sect. Lobelia as a model for why close relatives do or don’t live alongside each other; my piece was the phenotype side. The Lobelia silhouettes came earlier, in 2019: that one is reconstruction for depiction and keeps true size, this one is segmentation for measurement and normalises size away.

Why bother

Leaf shape varies between close relatives in the same place, and that variation is tied to how a leaf works: light interception, thermoregulation, water-supply trade-offs, plasticity (Nicotra et al., 2011; Tsukaya, 2018). Testing any of it across a clade needs shape as a number, for a lot of plants.

Lobelia sect. Lobelia is a good test case because it is awkward: 23 species as scoped in 2021, 26 in the 2025 phylogenomic revision (Godden et al., 2025), from the tall red-flowered L. cardinalis to the small aquatic L. dortmanna. Several grow as basal rosettes, where herbarium-vision work generally assumes separated, planar leaves on a stem. A rosette pressed flat is a pile of overlapping blades radiating from one point.

The corpus

Every measurement on this page comes from plants collected in the field and taken apart by hand — dismembered and photographed so the leaves lie flat, separated and unobscured, then cropped leaf by leaf.

Nine species are represented:

Sampling is uneven in both directions, and they are different problems. Twenty-one glandulosa plants contribute 48 leaves; three apalachicolensis plants contribute 41. Few leaves per plant limits what you can say about one individual; few plants per species limits what you can say about the species, and no amount of leaves off those three plants fixes it.

The pipeline

“Semi-automated” is the operative word. A human is in the loop throughout:

1. Voucher collection from multiple sites.
2. Digitization and “dismemberment.” The specimen is taken apart and photographed, so leaves lie flat, separated and unobscured.
3. Whole-plant traits in ImageJ against the 1 cm scale standard: base-to-first-leaf, base-to-first-flower, stem thickness at base and at first flower. Images are rotated to a common axis, which nearly every morphometrics tool assumes.
4. Leaf cropping (ImageJ), one sheet becoming many single-leaf images. Leaves too folded or torn to read are excluded by explicit criterion, not by eye.
5. Area and perimeter via the LeafArea R package driving ImageJ, calibrated at 85 px/cm:

library(LeafArea)
run.ij(set.directory = ".../leafcrops",
       distance.pixel = 85,
       known.distance = 1,   # cm
       trim.pixel     = 0)

1. Analysis in R / RStudio.

Thresholding each crop gives the binary mask the measurements come from, and it is where specimen condition matters most:

It is not a rare problem, and it is the reason for the attrition in the table above: of those leaf components big enough to measure at all, a third were rejected as too damaged.

Existing tools were surveyed first. Most assume material digitized herbarium sheets never supply:

The binding constraint was rarely the model. These tools assume idealized input, which pressed specimens, and flattened rosettes especially, do not provide.

What it showed

Perimeter scales with area differently in different species. That is a shape statement, not a size one: a leaf gaining perimeter quickly as it gains area is narrower, or more dissected, or more toothed.

Shape differs between species, and between populations of the same species. A clade-scale story has to hold at both levels.

Nine species, not ten. The largest component in a mask is often a stem fragment, and L. canbyi’s only mask is nothing else: 3,257 px at 0.55 solidity against 0.93–0.97 for a clean blade. Including it would have shown a stem fragment under the species name.

A second method, two years later

In 2026 I re-analysed the same 104 masks by an independent route: 486 leaves from 88 specimens across 8 species, each outline resampled to 128 pseudo-landmarks, aligned, scaled to unit centroid size and ordinated by PCA. It uses neither area nor perimeter, and shares no code with the 2024 pipeline.

PC1 carries 48% of shape variance and is almost exactly leaf breadth, correlating with measured width-to-length at r = 0.978. Species order along it monotonically, L. glandulosa at 0.109 to L. apalachicolensis at 0.439, and it holds essentially all the species signal: between-species share (η²) 0.493 on PC1 against 0.006 on PC2. PC2 carries another 26% of shape variance but splits by plant rather than species, η² 0.211 by specimen: leaf-to-leaf variation within one individual.

Classifying from outline alone, cross-validated grouped by specimen so no test plant’s leaves appear in training, gives 0.372 against a 0.166 permuted-label null and a 0.191 majority-class baseline. Roughly twice chance: a genuine species character, if a weak one. A naive split letting one plant’s leaves span train and test scores 0.430, a 6-point gap that measures the pseudo-replication in treating ten leaves off one plant as ten observations.

Against 2024 the comparison is species by species rather than leaf by leaf. L. glandulosa is the outlier here and is also the species the 2024 plot picks out with the steepest slope. They part company on L. elongata: perimeter grows with the square root of area for a fixed shape, so a straight-line slope depends on the size range a species spans, and elongata reaches ~30 cm² where the others stop near 5–10. Its gentle slope is partly a size effect; on the dimensionless index it sits mid-pack, which sharpens the original result rather than undoing it.

Damage shifts PC1 without accounting for it. Solidity, the filter the pipeline already uses to reject torn leaves, still tracks PC1 among the leaves that passed it (r = 0.360), and within species too (mean 0.308): a torn leaf reads as narrower. Residualising it out costs little, η² 0.493 → 0.434 and accuracy 0.372 → 0.360. Damage carrying no shape information at all classifies at 0.195 against a 0.191 baseline, which is what licenses reading PC1 as shape rather than preservation. One caveat: L. glandulosa is both the most damaged species and the outlier carrying the agreement above.

What this does not show. Sampling is uneven (93 leaves from puberula, 26 from inflata; 21 specimens for glandulosa, 3 for apalachicolensis). Absolute size is discarded by construction though size is a real diagnostic character, and venation, dentition and pubescence are not in an outline at all. PC1 is interpreted against a width-to-length measurement of the same outlines rather than an independent character set, so it describes this material rather than validating against an external standard.

Status

Completed undergraduate work, recorded here rather than maintained.

Code, protocol, notebooks and ledger: musharna/lobelia-leaf-morphometrics. No specimen imagery is included: the raw sheets carry all-rights-reserved notices burned into the pixels regardless of the licence field on the aggregator record. Images belong to their holding institutions; the GBIF download is CC BY-NC 4.0.