Agriculture is the use case edge AI was made for: sensors spread over hectares with no power infrastructure, connectivity measured in messages per day, and phenomena — pests, irrigation events, animal behavior — that must be detected where they happen.
This guide covers the applications and the architecture that survives the field.
Detecting events at the node
The pattern is consistent: a cheap sensor stream plus a small model that recognizes what matters. Acoustic sensors classify pest wingbeat frequencies (insect traps that count specific species); vibration and flow meters confirm irrigation events and flag leaks; IMU collars classify livestock grazing, rumination, estrus behavior, and distress.
Each is the same architecture: local inference turns a high-bandwidth stream into a low-bandwidth event — 'target pest detected x3', 'valve left open 40 min', 'cow 47 off-feed today' — sized for the links farms actually have.
Livestock: the collar economics of grazing animals
Livestock wearables are pet-tech physics at herd scale. IMU models classify grazing, ruminating, walking, lying — behavior budgets that flag illness and estrus days before visual observation. On-device inference is forced by the economics: a thousand-head herd can't stream raw IMU over rural links.
The models are identical in spirit to pet wearables — activity classification plus per-animal anomaly baselines — engineered for harsher environments (IP67+, extreme temperatures) and multi-year battery budgets where replacement labor dwarfs hardware cost.
- Pest traps — acoustic wingbeat classification, species counting
- Irrigation — flow/vibration event confirmation, leak detection
- Livestock — grazing/rumination/estrus from collar IMU
- Environment — microclimate events, frost, fire-signature cues
Power and connectivity: the field stack
Field nodes live on solar + battery or battery alone for seasons — the duty-cycled architecture is mandatory: sleep at microamps, wake to sample, infer in bursts, transmit rarely. LoRaWAN and NB-IoT dominate telemetry with payloads measured in bytes — another reason inference happens on the node: the event fits the link; the raw data never would.
Store-and-forward resilience matters: nodes buffer events through connectivity gaps (weather, terrain, link maintenance) and sync when the link returns. A node that goes silent during a storm is a product; a node that loses the data is a liability.
Seasonality and model drift
Farm environments change more than almost any deployment — seasons alter acoustics, vegetation, temperatures, and animal behavior baselines. Models that assume stationarity degrade; the fix is seasonal retraining (the fleet's data flywheel supplying examples) and drift detection that flags when the model's confidence collapses.
This is where OTA model updates earn their keep in agriculture: a pest model retrained on this season's acoustics pushes to the fleet without a single farm visit.
Where it shows up
Automated pest traps
Acoustic wingbeat classification counts target species in real time — replacing weekly manual trap inspections.
Irrigation monitoring
Flow and vibration nodes confirm irrigation events and flag leaks or stuck valves across hectares of field.
Livestock behavior
Collar IMUs classify grazing, rumination, and estrus — flagging off-feed animals days before visible illness.
Frequently Asked Questions
Can field sensors classify pest species on-device?
Yes — insect wingbeat frequencies are distinctive signatures (different species beat at different characteristic frequencies), and compact audio models classify them on-node in real time. Automated traps that count target species without manual inspection are a shipping product category, replacing weekly human trap checks.
How long can a field node run on battery or solar?
Seasons to years — a duty-cycled node (sleep, periodic sample-and-infer, rare transmission) averages tens of microamps, giving multi-year battery life or indefinite solar+supercap operation. The budget is dominated by the radio: keep transmissions small and infrequent and the AI itself is nearly free.
What connectivity works on farms?
LoRaWAN (kilometers of range, tiny payloads, private gateways) is the workhorse for sensor events; NB-IoT/LTE-M uses carrier infrastructure where coverage exists; satellite fills the remote gap. All share the constraint — tiny, infrequent payloads — which is precisely what on-node inference produces.
Do agricultural models need retraining each season?
Often, yes — environment and phenomenon baselines shift with seasons (background acoustics, vegetation, animal behavior). Fleets that collect labeled events feed a retraining loop and push seasonal model updates OTA. Alternatively, anomaly-detection architectures that adapt slowly to the drifting 'normal' reduce the retraining burden.
Building something that should run AI on-device?
Edgehound designs, compresses, and deploys TinyML models on microcontrollers — from feasibility audit to field-ready firmware.