# Phenomics and high-throughput phenotyping

Why phenotyping is the binding constraint on genomic selection, what a sensor actually measures as opposed to what the trait is, and how field spatial variability is handled.

Sequencing got cheap and measuring plants did not, so the limit on breeding moved to the field.

Source: https://en.bioecon.ru/docs/agri-food/crop-biotech/phenomics-high-throughput-phenotyping/
Updated: 2026-09-06



The cost of genotyping has fallen by orders of magnitude; the cost of characterising what a plant does has not. That asymmetry moved the bottleneck. A genomic prediction model is fitted to observed phenotypes, so its accuracy is capped by the quality of those observations — measurement error enters directly, and a trait recorded sloppily behaves in the model like a trait with low heritability. High-throughput phenotyping exists to raise the number, precision and frequency of those measurements, not to discover anything by itself.

## What the instruments actually measure

Nothing on a phenotyping platform measures a trait. Each sensor measures a physical quantity from which a trait is inferred.

Multispectral and hyperspectral cameras record reflectance. Vegetation indices computed from red and near-infrared bands track leaf area and canopy greenness well at low to moderate cover, and then saturate: once the canopy is closed, additional biomass changes reflectance very little, which is exactly the range where yield differences accumulate. Thermal imaging records canopy temperature, and canopy temperature below air temperature indicates transpiration and therefore open stomata and available water — a genuinely useful proxy, and one strongly confounded by wind, humidity, soil background and time of day, so it needs reference surfaces and tight acquisition windows. LiDAR and stereo imaging give canopy height and structure. Chlorophyll fluorescence reports on photosystem II efficiency.

Every one of these requires calibration against destructive measurement — cutting, drying, weighing, counting — and the calibrations are specific to crop, cultivar group, growth stage and often site. A model trained in one season transfers to the next imperfectly, which means the destructive sampling never goes away; it becomes the smaller, more valuable part of the workload.

Roots are the standing gap. The traits that matter most under water limitation are below ground, and there is no non-destructive field method with useful throughput. Excavation scoring, rhizotrons and X-ray tomography each give part of the picture at low throughput or under artificial conditions.

## The variability problem

A field is not uniform. Soil depth, texture, drainage and historical management vary at scales of metres, and that variation frequently exceeds the genotype differences being measured. Phenotyping faster does not fix this — it produces more precise measurements of a confounded quantity. The answers are experimental and statistical: replicated and augmented designs with repeated checks, and spatial models — row–column effects, nearest-neighbour adjustment, two-dimensional splines — fitted to separate field position from genotype. A platform without that design behind it generates data and not information.

Aerial platforms add something design alone cannot: repeat measurement. Flying a trial weekly turns a single endpoint into a growth curve, and dynamic descriptors — the rate of canopy closure, the timing and slope of senescence — are sometimes more heritable and more predictive than the endpoint itself.

Controlled-environment phenotyping is precise and translates poorly. Pot volume constrains roots, light is uniform, and there is no competition; rank orders obtained in a growth chamber routinely fail to reproduce in the field. The precision is real; the question is what it is precise about.