---
name: using-olmoearth
description: Plans and runs satellite analyses of a place with Ai2's OlmoEarth models, choosing between a plain vegetation index, a released fine-tuned model, OlmoEarth Studio, embeddings or local inference, then turns the result into a where-to-survey-first list with honest limits. Use when someone wants to map habitat change after a storm, fire or clearing, map mangrove, forest or land cover for their site, asks whether OlmoEarth suits their question, or needs an OlmoEarth result checked before acting on it.
license: MIT
---

# Using OlmoEarth for a place

OlmoEarth is a family of open Earth-observation foundation models from Ai2
(the Allen Institute for AI). It reads Sentinel-2, Sentinel-1 and Landsat
imagery and produces embeddings or, after fine-tuning, maps. This skill takes
a question about one place to a map a field team can act on.

Work through the steps in order. Each ends on a check; move on when it passes.

## 1. Frame the decision

Write one sentence: *[who] will use [output] to decide [action] at [place]
during [time]*. Add the unit that gets a prediction (10 m pixel, survey cell,
mangrove stand, transect) and the smallest change that would alter the action.

Check: the sentence names a person and an action. "Understand the mangroves"
fails; "the ranger team picks which six mangrove stands to walk in March" passes.

## 2. Say what this session can do

- **Coding agent with a shell** (Claude Code, Codex, Cursor, Gemini CLI):
  plan, download imagery, run models and write files.
- **Chat-only app**: plan, explain, write the scripts and review results the
  user pastes back. It cannot download imagery or run a model. Say so once,
  then hand over runnable steps.
- **Studio user**: OlmoEarth Studio is a hosted web app with an account
  request. Plan the labels and settings; the user clicks through it.

Check: the user knows which of these they are getting.

## 3. Test whether a satellite can see it

Answer in writing before choosing a model:

- How wide is the thing? Sentinel-2 pixels are 10 m (some bands 20 m or 60 m).
  A feature needs several pixels across it. A 15 m mangrove fringe is one or
  two mixed pixels.
- When did the event happen, and which months before and after are comparable?
  Compare the same months across years so seasons and tides do not pose as change.
- How cloudy is the place in those months? Humid tropics and islands often
  lose most scenes.
- What would count as truth? Field records, drone flights, a published map, or
  nothing. Say which.

Check: every answer is written down, including "no ground truth".

## 4. Choose the smallest route

Read [choosing-a-route.md](references/choosing-a-route.md). Always compute a
plain index baseline such as NDVI change; many storm and clearing questions end
there. Use OlmoEarth when the index cannot separate the classes you care about,
when you need a map of a type (mangrove, forest) rather than greenness, or when
a released model already fits.

Check: you can name the route, the baseline, and why the route beats doing
only the baseline.

## 5. Confirm current facts

Model names, access rules and install steps change. Read
[olmoearth-facts.md](references/olmoearth-facts.md), then open its linked
changelog and repository pages before installing or quoting a model name.
Record the model ID, weight revision and date checked.

Check: every OlmoEarth name you will use appears on a live Ai2 page today.

## 6. Run small, then scale

Follow [running-it.md](references/running-it.md). Run one small window end to
end first. Look at the imagery, the cloud mask and the output together on a
map before tiling the whole area.

Check: one window produced an output that lines up with the imagery at the
coastline or a road.

## 7. Check the result

Apply [checking-results.md](references/checking-results.md). Treat a
surprisingly good score as a bug until you find why.

Check: you can state which claim the evidence supports (for example "detects
canopy loss at this site") and which it does not.

## 8. Hand over survey priorities

Follow [survey-priorities.md](references/survey-priorities.md) to turn the map
into a ranked list of places to visit, each with the reason it was chosen and
what the visit should record.

## Done when

The user has, in files they can open:

- a georeferenced map (GeoTIFF or GeoJSON) with dates, sensor, model and version;
- the baseline map made the same way, for comparison;
- a ranked survey list with coordinates and reasons;
- a short limits note covering clouds, resolution, truth data and what the map
  cannot tell them;
- the commands or notebook that made all of it.

In a chat-only app, "done" means the plan, the scripts and the limits note.