Training a model

The built-in analyses find things somebody wrote a rule for: heaps, slopes, water, trenches. A trained model finds whatever you show it.

Weeds. A particular species of tree. Water troughs. Rock outcrops. Damaged sections of fence. Anything you can recognise in the imagery and are prepared to outline a few dozen times.

How it works

You draw outlines around examples on a survey you have already processed. The platform cuts the imagery into tiles, learns what the outlined things look like, and produces a model. You then run that model over any survey as an analysis, and it finds the same kind of thing on ground it has never seen.

The model learns from pixels and, if you ask it to, from height as well — the canopy height model is a layer it can use, and "tall and green" is a much easier thing to learn than "green".

Making one

Open Models and create one. Name it after the thing, not the project: Parkinsonia, not Ellerslie north 2026. You will run it on other surveys.

Choose the survey to label on and the layers the model may look at. The orthophoto always. Add the canopy height model when height distinguishes the thing from what surrounds it.

Outline examples. Draw around each one on the map. Follow the edge reasonably closely — a loose box around a tree teaches it that the grass beside the tree is part of the tree.

Train. It takes a while. You get a score when it finishes.

How many examples

Three is the bare minimum the platform will accept. A couple of dozen is where it starts working, and a hundred is where it gets good.

More important than the count is the variety. Twenty examples all from one corner of one paddock in one light teach a model about that corner. Twenty spread across the survey — different backgrounds, sizes, sun angles, some partly obscured — teach it about the thing.

Include the hard ones. The small ones, the ones half in shadow, the ones next to something that looks similar. A model trained only on obvious examples finds only obvious examples.

Label everything in view

This is the mistake that costs the most. The platform cuts your survey into tiles and every tile is training data — including the ones with nothing outlined in them, which teach the model what background looks like.

So if you outline nine trees in a corner of the paddock and leave the other four hundred alone, every tile containing an unlabelled tree teaches the model that trees are background. It trains directly against the thing you are teaching it. The loss falls, the tiles look right, and the score sits near zero.

Either label every instance in the ground you are using, or restrict the labelling area to a region you have labelled completely. The second is usually easier and works just as well.

Reading the score

Training reports a score after it finishes, and part of the way through.

• Below 20% — something is wrong. Usually too few examples, or unlabelled instances teaching against you, or the thing genuinely does not look different from its background at this resolution.
• 20–50% — it will find some. Useful as a first pass over ground you were going to check anyway.
• Above 50% — good. It will find most of them and be wrong sometimes.
• Above 80% — very good, and worth being slightly suspicious of. Check that the thing is not trivially distinguishable in a way that will not hold on other ground: a species that happens to be the only tall thing in this paddock, or an object that only appears on one soil colour.

The score is measured against examples the model was not trained on, so it is not simply reporting how well it memorised.

Resolution

A model is trained at a particular ground sample distance, and it remembers it. When you run it on another survey the platform resamples that survey's imagery to match — so a model trained on 2 cm imagery works on a 4 cm survey, at the detail 2 cm implies.

Run it on imagery much coarser than it was trained on and it will find less. There is nothing to be done about that: the pixels are not there.

Versions

Every training run produces a version. They are kept, and you choose which one is live. Train again with more examples, compare the scores, and switch — or switch back if the new one is worse. Results already produced by an older version stay as they were.

Running it

A trained model appears in the analysis list alongside the built-in ones. Pick it, draw an area if you want, run it. The result behaves like any other: outlines you can click, fold together, or throw out. See Analyses.

Labelling on a tablet

Labelling is the part worth doing in the field, where you can see whether the thing in the imagery is the thing you meant. The API supports it, so a companion app can list models, fetch imagery, and post outlines. See The API.