New desktop tool
Train archaeological site-sensitivity models from your own LiDAR on your desktop, then import the model into the Ology iPhone/iPad app and predict fully offline — no cell signal, no cloud. Train once on ground you know; predict anywhere.
The trainer in action
Work down the input cards on the left, press Run Training, and watch each pipeline stage light up with its own timer. When it finishes, the Result tab shows the headline accuracy and buttons to open the report, reveal the model file, and open the validation map.
What it does
The trainer builds a predictive site-sensitivity model. It learns, from the places where sites are already known, what the terrain tends to look like where people lived — then it scores every patch of a new area for how much that ground resembles known-site terrain. The output is a heat surface: brighter means more likely to reward survey.
The model never tries to “see” a house pit or a scatter in the imagery. It asks: is this spot the kind of place — near reachable water, on a flood-safe bench, gently sloped, sheltered — that people historically chose? That is a question terrain can answer, which is why it works even for sites that leave no visible trace on the surface.
A prioritisation aid, not a site locator: a high score means “worth looking here first,” not “a site is here.” Ground-truth everything.
Every terrain predictor comes straight from your DEM — nothing to download.
Two model types
Transparent arithmetic. Every number is auditable — you can say exactly why a spot scored high, like “sites are 3.4× more likely on flood-safe terraces.” Fast, and the right first model.
Learns interactions FR cannot: “south-facing matters, but only on gentle slopes near water.” On tabular terrain data, gradient boosting is the current state of the art — it consistently matches or beats neural networks in independent benchmarks — and the trainer still reports each factor’s importance.
Searches model configurations and keeps the one that best predicts ground the model didn’t train on — not the one that best memorises.
Trains several models and averages them for a steadier surface. Exports to the same file format, so the app runs it with no changes.
Ensemble disagreement becomes a genuine confidence measure: “high sensitivity and the models agree” is a stronger survey target.
Proof it works
Tick “Held-out validation map” and the trainer runs an honest experiment: it splits your area into a training region and a spatially separate held-out region, trains only on the training region, then checks whether the real sites in the unseen region land in the model’s high-scoring zones.
You get a map and a headline number — of the held-out sites, how many the top-scoring 20% of that ground captured. A model with no skill captures about 20%. A good model captures several times better than chance on ground it never saw. That is the figure, and the picture, to show a reviewer or a SHPO.
How to use it
Then press Run Training. AirDrop the finished .json to your iPhone → Ology → Terrain → Archaeological Site Sensitivity → Use model → import.
Documentation
The guide assumes no machine-learning background. It explains in plain language what the trainer is doing under the hood, why the method is sound, how to run it, and how to read the results — including the held-out validation map, the ethics and honest caveats, and troubleshooting for macOS and Windows.
The model in the app
AirDrop the trained .json to your iPhone or iPad and the Site Sensitivity tools take over: pick your model, apply it to any map area, flag the top-scoring ground, and walk to it. Desktop-trained and in-app-trained models live side by side.
Get the trainer
The Ology Model Trainer runs on macOS and Windows (Python 3.12/3.13). It is currently distributed directly to interested archaeologists and cultural-resource professionals while it is refined. Tell me a little about your work and I will follow up with the training kit.
Site locations are protected under NHPA/ARPA and state law. Your training data never leaves your machine — the trainer runs entirely locally, and models derived from confidential sites should be handled with the same care.
Downloads
Public downloads will appear here as they are released. For now, the user guide is available to everyone and the trainer kit is available by request above.
Publicly released models will never include or reveal confidential site locations — model outputs are probability surfaces, not site records.
Desktop to field
The trainer and the app are two halves of one workflow: heavy LiDAR processing and honest validation on your computer, then a single portable model file that predicts offline anywhere the app can open a map. Both are aids to professional judgement, not replacements for it.