John Deere reports its camera-based See & Spray system ran across more than five million acres during the 2025 growing season, cutting nonresidual herbicide use by nearly 50% on average. That number, not any datasheet spec, is the current benchmark for optical sensing in agriculture: the question is no longer whether cameras and spectrometers can measure plants, but whether the measurement turns into a field action that pays.

A synthesis from the European Photonics Industry Consortium (EPIC), written by photonics technologies manager Jérémy Picot-Clémente, maps how imaging, spectroscopy, and controlled illumination now span the food value chain—from drones and field robots to packing lines and greenhouse fixtures.

What can hyperspectral imaging actually deliver?

Hyperspectral imaging records many contiguous spectral bands at every pixel, so each spatial point carries a full spectrum sensitive to pigments, water, tissue structure, and chemical composition. Paired with calibrated analytical models, that data can grade produce or flag defects that an RGB camera misses.

HySpex, for instance, applies hyperspectral systems to berry analysis: spatially resolved spectra show where a defect sits, which suits automated sorting rather than reporting a single average value. The trade-off is data volume and processing load.

Multispectral cameras cut that burden by selecting a handful of bands. CSEM, the Swiss innovation center, builds multispectral instruments with embedded machine learning that trade off spectral content, spatial resolution, and acquisition speed. But indices like NDVI (normalized difference vegetation index, computed from red and near-infrared reflectance) only interpret correctly when you account for canopy structure, soil background, and growing conditions.

A common engineering pattern: use hyperspectral measurements during development to find the informative wavelengths, then deploy a cheaper multispectral instrument. Ocean Optics' compact spectrometers support the same logic for point measurements—handheld or vehicle-mounted devices that answer one narrow question and skip spatial coverage entirely. More spectral channels don't automatically mean more value.

Can NIR measure sugar without cutting the fruit open?

Near-infrared (NIR) spectroscopy links optical absorption—molecular vibration features—to internal composition. Because food scatters light and is heterogeneous, quantitative predictions need calibration against reference measurements, then validation across representative samples. Sensor performance alone doesn't establish the accuracy of a composition estimate.

Norway's DigiFoods research center shows the method working at speed. Nofima and SINTEF developed noncontact NIR measurements for cherry tomatoes, with trials on moving produce linking predicted sugar content to sensory sweetness. Related work measures strawberries on the plant under daylight. That bridges lab spectroscopy, packing lines, and pre-harvest measurement.

Miniaturization helps. Hamamatsu Photonics offers miniature Fourier-transform NIR engines built on MEMS interferometry with InGaAs (indium gallium arsenide) detectors, plus InGaAs line sensors for NIR and SWIR inspection on conveyors. But geometry matters as much as size: reflection, transmission, and interactance arrangements probe different sample volumes and shift sensitivity between surface, internal tissue, and stray light. A classifier trained on one cultivar or season may simply not transfer to another.

What does active illumination add?

LiDAR (laser ranging that reconstructs terrain and canopy geometry) supplies crop height, volume, and structure, helping separate architectural changes from genuine optical-response changes. Laser-induced fluorescence offers a physiological readout: chlorophyll emission reflects the competition between photochemistry, fluorescence, and heat dissipation. Fluorescence intensity alone, though, isn't a universal health indicator—it demands a defined protocol and, ideally, photosystem-efficiency parameters.

Avantes supplies compact UV–visible–NIR spectrometers with dedicated illumination and collection optics; Photonics Bretagne in France builds multispectral LiDAR demonstrators that measure leaf reflectance. In both cases the design work lives in working distance, collection efficiency, and rejection of ambient light.

Can light itself control growth?

Horticultural LEDs act as actuators, not sensors. Blue and red wavelengths steer photosynthesis and development; far-red and UV treatments alter morphology. Responses depend on species, growth stage, dose, and environment—no universal lighting recipe exists.

Seoul Semiconductor supplies horticultural LED technology, and Crocus Labs builds systems with independently controlled spectral channels and sensor feedback, adjusting output as available daylight changes rather than holding a fixed level. Their value, the EPIC analysis stresses, should be judged on crop performance and electricity consumption together.

What stands between a good sensor and adoption?

Deployed systems prove the concept. Yara's tractor-mounted N-Sensor reads canopy reflectance to drive variable-rate nitrogen application. TOMRA reports more than 800 installed sorting machines for fresh potatoes and vegetables.

The barriers are economic and practical. OECD research identifies cost, limited relevance, complexity, skills requirements, and distrust of algorithms as the main obstacles to digital agriculture. The central challenge: demonstrating credible return under one particular farm's conditions. An accurate measurement has limited value if interpreting it needs specialist expertise or disrupts an established workflow.

EPIC's prescription for suppliers is blunt—design and test instruments with farmers, agronomists, and equipment makers from the start; validate across seasons, varieties, and operating conditions; report failures transparently; and add automatic calibration checks, simple interfaces, and leasing or cooperative-ownership models so not every farm must buy its own hardware.

The next advance in agricultural optics, the analysis concludes, will be measured by how reliably optical information becomes useful action—not by spectral channel count.