Louis RyanComputer vision · Ireland

Project · 2026 · client not named

Counting lambs from a drone

What was needed
A sheep farmer needed an accurate lamb count across four hill fields at lambing. Two people walking the ground took most of a morning. The number was still an estimate.
What I built
A fixed flight path for an off-the-shelf drone, and software that finds and counts every lamb in the photographs it brings back. The count arrives on his phone as a number and a marked-up map.
What changed
The count takes about four minutes instead of a morning, and it is repeatable, so week-to-week changes are now visible. It misses lambs sheltering under hedges, and it says so rather than quietly under-reporting.
How long it took
Five weeks, including two lambing visits to get photographs in real conditions.
The drone on the ground at the edge of a field before a counting flight.
Nothing clever here. A fixed height, a fixed speed, and a straight line. Consistency in the photographs did more for the count than anything in the software.

The story

The first thing I got wrong was the height. I flew high to cover the ground in fewer passes, and at that height a lamb was about eight dots across. The software found roughly half of them.

Dropping to a lower altitude and accepting more passes fixed most of it. That is the sort of thing that only shows up in the field, which is why I ask to visit before quoting.

The second problem was shelter. Lambs sit under hedges and trees, and no camera looking straight down will see them. Rather than pretend otherwise, the count comes in two parts. A firm number for open ground, and a flag on the areas that could not be seen.

He walks those few spots himself. That took the job from "a clever gadget" to something he actually uses at six in the morning.

A drone photograph with every lamb boxed and numbered, total shown in the corner.
The count he sees on his phone. Each box is one lamb. Tapping a box shows the photograph it came from, so he can check any number himself.
The same photograph with two lambs under a tree marked in pink as missed.
The two it misses under the hedge. I show him this picture too. A count you can trust is one that tells you where it is unsure.
For the technically minded

Small-object detection on 20-megapixel stills, tiled into overlapping crops so lambs are a workable size in each tile, with overlapping detections merged afterwards. Training data was about 3,000 hand-boxed lambs from two visits, deliberately including rain, low sun and shadow. Everything runs on his laptop from an SD card — no cloud, no subscription, and no connectivity needed in the yard. He owns the code.

The measured miss rate on open ground was about one in fifty. Under cover it is far worse, which is why cover is reported rather than counted.

Topics this used

Something similar on your land or your line?

Tell me what you are counting and how often. I will say plainly whether a camera is the right answer.