AI livestock monitoring that works where there is no broadband

Why lameness goes unnoticed

Lameness is a leading cause of poor welfare and economic loss in sheep farming, and prevalence frequently exceeds 10% of a flock. It is also one of the most treatable conditions on a farm, provided it is caught early. Three things stop that happening.

It is spotted when it is already severe

Manual detection depends on someone watching animals move and noticing a change. Habituation works against that: a mild limp seen every day stops registering until it becomes obvious. By then treatment takes longer, the animal has lost condition, and if the cause is contagious it has had weeks to spread through the flock.

Checking every animal does not scale with the labour available

Greek sheep farming runs on fragmented smallholdings with an ageing workforce and rural depopulation. Hiring is often not economically possible on a family farm. Daily individual observation of a few hundred animals competes directly with every other job, and it loses.

There is no usable connection at the pen

Most precision livestock systems assume Wi-Fi, LoRaWAN or cellular. That assumption excludes exactly the farms that need help most: rugged terrain, mountain grazing, sites where the nearest reliable link is a long way from the animals. A monitoring system that needs an uplink is not available to them at any price.

What the system does

A camera watches a passage the animals already use. For each animal that walks past, the pipeline detects it, extracts skeletal keypoints, and scores the gait across a sequence of frames rather than judging a single image. The output is one record per animal per pass: an identifier, a score between 0 and 1, a severity band, a timestamp and a location.

Everything runs on the farm. A single-board computer with an accelerator handles detection, pose estimation and temporal analysis inside a 500ms budget per segment, which is fast enough to tag the animal before it leaves the frame.

The farmer sees a dashboard: total animals counted, healthy against flagged, a calendar of daily lameness counts, and a list of individual animals with severity filters and timestamps. Alerts carry a snapshot so a flagged animal can be checked without walking to the pen first.

What it measured in the field

ELDER declared its targets before the work started. These are the final results against them, including the one we missed.

In the final validation run, 32 of 37 animals were classified correctly: 11 of 12 lame animals found, 21 of 25 healthy animals cleared, at a threshold we raised from 50% to 70% during the trial because wet floors and head movement were generating false limps.

Farms without broadband

Most precision livestock systems assume Wi-Fi, LoRaWAN or cellular, which restricts them to flatland and peri-urban farms. Fragmented smallholdings on rugged terrain, which is most of Greek sheep farming, get excluded by that assumption.

Because only the conclusion leaves the farm, an event fits inside the 200-byte limit of narrowband IoT over non-terrestrial networks. We validated this against a testbed emulating low Earth orbit, with store-and-forward behaviour and delivery delays from minutes to hours, and kept full data integrity through field testing at well under 10% connectivity uptime.

This works because lameness is a chronic condition and an alert an hour late is still actionable. The full architecture is written up here.

What it does not do

It does not replace the shepherd, and at a 16% false-positive rate it cannot. In a 500-animal flock that is roughly 80 alerts a day for sound animals, which is enough to make a person stop reading them. The system is a screening tool with a human verifying alerts from the attached image, and the labour saving is real but smaller than a vendor would claim.

It does not diagnose. It scores gait. What is causing the gait is a question for a vet.

It was validated on sheep, not cattle or goats. And it will not distinguish a clinically lame animal from one that simply needs its hooves trimmed unless your own labelling makes that distinction consistently, which took us longer to work out than it should have.

Questions we get asked