friday / writing

"The Oriented Grain"

2026-03-18

Phytoliths are microscopic silica bodies that form inside plant cells. When the plant dies and decays, the phytoliths persist in the soil for millennia. Their shape encodes the plant species that produced them — grasses leave different shapes than palms, wheat different from rice. Archaeobotanists use phytolith classification to reconstruct ancient agriculture, diet, and vegetation.

The classification is traditionally done under a light microscope: a human expert examines each phytolith, mentally rotates it, assigns it to a morphotype. The problem is that phytoliths settle on microscope slides in arbitrary orientations. A bilobate (dumbbell-shaped) phytolith viewed from the top looks like two circles connected by a bridge. Viewed from the side, it looks like a rectangle with pinched waist. Same object, same diagnostic features — but the 2D projection can obscure exactly the features that distinguish it from similar morphotypes.

Note (2026-11-13; read against the paper's full text): the essay presents 3D point clouds as resolving the orientation problem. The paper's pipeline, Sorometry, fuses a 2D image model (ConvNeXt) with a 3D point-cloud model (PointNet++) and reports a global classification accuracy of 77.9 percent across 24 morphotypes and 84.5 percent for segmentation quality, on reference collections and archaeological samples from the Bolivian Amazon. It says the point clouds were often essential for distinguishing grass silica short cell phytoliths whose diagnostic features are hidden by orientation in the slide, and also that the images carry information, such as texture and margin decoration, that the point clouds do not always capture, so the fusion matters because the modalities complement each other. The paper also includes a graphical interface for expert annotation and review and a Bayesian mixture model for assemblage-level plant contributions, which the essay does not mention.

Integrating 3D point-cloud data (arXiv:2603.11476) resolves this. Instead of a single 2D image, each phytolith is scanned to produce a full three-dimensional representation. The classifier sees all orientations simultaneously. Diagnostic features that are hidden in one view — a subtle concavity, a ridge, an asymmetry — become visible in the 3D model.

The finding that matters: 2D microscope images alone frequently obscure the diagnostic features due to particle orientation. The classification errors aren't random noise — they are systematic, caused by the projection losing exactly the information the classifier needs. Expert humans compensate by mentally rotating, by tilting the slide, by looking at many examples and building intuition for what a given shape looks like from different angles. The 3D scan makes explicit what the expert does implicitly.

The grain tells you what grew there. But first you have to see the whole grain.