Satellites resolve the canopy surface at ten metres and a point probe resolves a few cubic centimetres, and every agronomic decision is made in the band between them. Precision agriculture reports what it measured; the record this review searched did not include a system reporting what it could not see. That is a measurement gap before it is a hardware one, and it is closable: the unknown fraction of a volume is computable the moment sensing is posed as one field over a volume rather than three stacked data products. What binds the closing is physics, and the physics is unusually neat because in agriculture the barrier and the signal are the same substance. Soil moisture attenuates the radio that would measure soil moisture; canopy density occludes the fruit whose count is set by canopy density. The engineering changes that move it are already on the bench: magnetic-induction and acoustic carriers where radio drowns, semantic scene completion where light stops, and soil microbial fuel cells reported at sixty-eight times the power a buried node needs. What becomes possible is a farm that states its own uncertainty, which is the first thing a supervisor, a regulator and an insurer all need and none can currently get.
One of eight, and only one of them is physics. How we read a frontier →
Physical AI and integrated sensors for agriculture surveillance.
Agriculture is instrumented at two extremes. Above, satellites resolve the canopy surface at ten metres, all weather, every few days. Below, a point probe resolves a few cubic centimetres of soil. Between them lies the volume where irrigation, nitrogen, harvest timing, disease response and yield are all actually decided, and this review did not locate an agricultural sensing system that reports what fraction of that volume it cannot see. The track treats a farm the way TR-2026-08 treats a room: not as a stack of data products but as one field defined over a volume by a mesh of nodes, some carried by machines, some anchored in the ground. Smart seeds and beacons are the anchored nodes. Drones and under-canopy rovers are the carried ones. Neither is the system. The system is the field they jointly resolve, and its honest output includes the part it still cannot see.
The physics is unusually clean, and it is what makes agriculture the hard case for embodied perception rather than an easy application of it. In a field the barrier and the signal are the same substance. Soil moisture attenuates the radio that would measure soil moisture, worst in exactly the moist and clay-rich soils worth reading. Canopy density occludes the fruit whose count is set by canopy density. Grain mass insulates the hot spot that grain mass creates. The occlusion is not noise on the measurement. It is the measurement, inverted.
The track runs on four legs at once. The technology and the science are the mesh and the barriers it has to see through. The economics ask what one reduced percentage point of unknown volume costs in sensors, joules and dollars per hectare, a curve nobody has drawn. The workforce leg is the lab's own subject: a farm service technician already installs sensors, calibrates controllers and troubleshoots communication networks, which is a network-operations job wearing an agricultural title, and it needs one competency the airspace work did not, calibration literacy, the ability to judge whether a mesh is still telling the truth after a season in the weather.
The word in the title is already legal territory, which is why the track sits in this lab. Idaho prohibits surveilling a farm by drone without the owner's written consent. Nebraska has made the farmer the owner of data their drones and combines produce. Meanwhile the European Union already conducts satellite surveillance of farm parcels for policy compliance, and from 2026 Norway requires salmon operators to publish real-time water-quality telemetry. The same instrument is a diagnostic, a compliance record and a competitor's intelligence, depending only on who holds the key. Who can read which voxel is the policy question and the technical question at once, and a volumetric field can be scoped to a region of space in a way a pile of files cannot.
OmniSense, the framing this imports ↗TR-2026-32, the agricultural estate ↗
Open the full view ↗Start with the easiest volume in agriculture. Heat moves through grain by diffusion, so a temperature probe reaches √(αt). Set the bin and the time you would accept, and read the fraction of the grain no sensor touches.
The same question, asked where the money is. A soil-carbon project pays for sampling that runs a quarter to two fifths of its budget, then makes a claim about a volume. The claim holds up well for a field average and rests, for everything below the sampling spacing, on a quantity geostatistics has been publishing all along without anyone reading it as a coverage figure: the nugget, the share of variance at separations shorter than any pair of samples. It does not fall when you add cores at the same spacing, and additionality and loss events are questions about exactly that share, because they ask whether something happened somewhere rather than what the field averages.
Third barrier, third physics, same shape of answer. A camera cannot count fruit the canopy hides, so machine-vision yield estimation counts what it sees and multiplies by a correction factor calibrated on sample trees. Single-side visibility measured across real apple orchards runs 40.85% to 79.83%, so the factor those orchards imply runs from 1.25 to 2.45. The published objection to hand calibration is not that the arithmetic fails but that choosing a representative tree is problematic given how much canopy density varies between trees, and that is the same division the other two instruments found: the factor is an average, averages answer the block question, and thinning, irrigation and harvest scheduling are per-tree decisions.
Open the full view ↗Foliage extinguishes a sight path the way it extinguishes light, so visibility follows Beer–Lambert in the leaf area between camera and fruit. Set the canopy and watch the correction factor move, then watch what it costs per tree against per block.
Open the full view ↗Set the sampling design a carbon project would actually run, then read what it bought: the volume physically sampled, the areal coverage the autocorrelation range grants it, and the variance no density at that spacing can reach.
↓ White paper · PDFRead onlineTR-2026-44 · research / position preprint