isoscapeisometric maps of real places

Open this in the explorer

Tutorial: map any place from open data with an AI chat, using Rapa Nui (Easter Island)

Tutorial 1. Drafted with Claude; see the artist statement.

isoscape draws isometric maps of real places from open data: buildings and roads from OpenStreetMap, and the shape of the ground from public elevation models. It is free and open source. You choose a frame on the world, and the data decides the rest. The reference site has every published map.

This tutorial teaches a way to use isoscape with an AI chat to explore any place on Earth. You ask questions in plain words. The chat finds the open data for the place, checks what it is and what it allows, draws the place, and looks at the result with you. You need no GIS software and no programming.

The worked example is Rapa Nui (Easter Island), a place with no public lidar, drawn from open data alone:

Rapa Nui seen from the south, with Rano Kau and Orongo at the lower left, Hanga Roa beside the airport, Terevaka at the top, and Rano Raraku, Ahu Tongariki and the Poike peninsula at the right

Fig. 1. The whole island, about 23 km across, drawn with isoscape from open data: OpenStreetMap for roads, buildings and land cover, and the Copernicus 30 m elevation model for the ground, raised 2× so its relief shows at this size. Open it in the viewer. Working draft.

What your chat needs

The chat has to be able to do three things:

This example was made in the Claude desktop app with Claude (Opus 5.5), with a folder on the computer connected to the chat. Nothing here needs a terminal: you ask in plain words, and the chat runs the commands and shows you the results.

Using Claude Code instead? Everything below works the same from a terminal. Clone isoscape, start Claude Code in that folder, and give it the same prompts.

If your chat cannot run code or fetch files, ask it for the exact commands, run them yourself, and paste the output back. It is slower, but the checking still works.

The method: ask, then check

A chat will answer a question about data from memory if you let it, and memory is where it goes wrong: a survey that was flown is not always one you can download, and a license read two years ago may have changed. So ask it to check. For each dataset it finds, have it answer these questions, and to open the source when it does:

  1. Does it exist and can I get it? A download link, an API, or only a paper that used it.
  2. Who made it, and how? Lidar from a plane, radar from orbit, photographs from a drone.
  3. What does it cover, and how fine is it? The area, and the size of one cell.
  4. What license? Whether you may share pictures made from it, and whether you may sell them.
  5. What can it not show? A 30 m cell cannot hold a 5 m platform.

Then have it fetch the data, measure it, draw it, and look at the picture. Numbers can look fine on a map that is visibly wrong. Every fault found in this example was found by looking.

The worked example

These are the prompts used, lightly trimmed, and what came back.

1. What could be drawn?

Here's a new challenge: Easter Island. Do an analysis of what you should be able to render. All island, the well-known landmarks, and then maximum zoom levels with LiDAR terrain preserved.

The chat searched for elevation data, read the sources, counted what OpenStreetMap holds for the island, and reported back. The first finding was that the best data is out of reach. An airborne lidar survey of the island was flown in April and May 2016, paid for by its heritage office, the Secretaría Técnica de Patrimonio Rapa Nui, and a 2026 study of sea-level rise used it [Sci. Rep. 2026], but we found no public download. The findings, checked again for this page:

DataDetailOpen?
Airborne lidar, 2016, for the Secretaría Técnica de Patrimonio Rapa Nuinot publishedno public download found
Copernicus DEM GLO-3030 m, a surface model from radar, worldwideyes, free with credit
Terrain Tiles (SRTM here)30 m, a worldwide mosaicyes, free with credit
NOAA NCEI Easter Island model3 arc-seconds (about 90 m), land and sea floor, 2017yes, US public domain [NOAA]
Drone model of the Rano Raraku quarry, 2025a 3D mesh of the quarry onlyyes, with limits (see below)
OpenStreetMaproads, about 4,100 buildings (most in Hanga Roa), land cover, named placesyes, ODbL

The answer to "how close can it go" came from the cell size: the whole island prints crisply, district views of a few kilometers soften, and close-ups of single sites need lidar that is not public.

2. Draw it, and compare the sources

The next session started from a handoff note that held the findings above, with this prompt:

Build the first isoscape Tutorial: how to explore a place with a chat, using Rapa Nui (Easter Island) as the worked example. [...] Show me a rough draft of the whole-island map and a one-page outline of the tutorial before writing it out, and ask me your open questions as multiple choice.

The note asked for both worldwide sources to be fetched and compared by looking.

isoscape takes a region file (a center, a frame size, a view direction and an elevation source), downloads the data once, and draws from the cached copy. The two worldwide 30 m sources gave the same island with different faults:

Rano Kau, Hanga Roa and the south coast drawn from Copernicus GLO-30: smooth slopes and clean sea cliffs
The same frame drawn from Terrain Tiles: grainier slopes and vertical banding on the sea cliffs
⤢

Fig. 2. The same frame from the two 30 m sources. Drag the divider, or move it with the arrow keys. Terrain Tiles (right) shows vertical banding on Rano Kau's sea cliffs and grainier slopes; Copernicus (left) was used for every view on this page.

3. Look, and fix what is wrong

Looking at the first draft turned up a green field floating in the sea off Hanga Roa. The chat traced it to an OpenStreetMap outline tagged as a nature reserve: the Coral Nui Nui underwater park, which isoscape had drawn like a park on land. The fix was a small change to isoscape so that a reserve lying wholly in the sea is not drawn as ground, checked against every published map to make sure none of them changed. It is the kind of fault no count would have shown.

The chat put each choice like this one as a multiple-choice question, so steering the work took a click: skip reserves at sea everywhere, work around this one feature, or leave it and explain it. The answer was to fix it everywhere.

4. Zoom in until the data runs out

The handoff note named the districts to try and asked the chat to say plainly where 30 m terrain stops being honest. Another multiple-choice question settled how far to go: these three views, then the quarry model if its files and license allowed.

Rano Kau's crater with its reed-covered lake, Orongo on the rim at the left, and the sea cliffs at the front

Fig. 3. Rano Kau and Orongo, about 3.8 km across, at true height. The crater reads well; the sea cliffs on the near side come out as soft, shaded slopes because 30 m cells round them off. Open it.

The Poike peninsula with Rano Raraku's crater at the lower left and Ahu Tongariki on the bay between them

Fig. 4. Poike, Ahu Tongariki and Rano Raraku, about 6.6 km across. The peninsula's shape is clear; its cliffs and the small cones on its summit are softened. Open it.

Hanga Roa's streets and houses along the coast, with Ahu Tahai on the shore at the top

Fig. 5. Hanga Roa and Ahu Tahai, about 3 km across. Here the buildings carry the picture: OpenStreetMap gives most of them a number of floors. The ground under them is nearly flat at this scale. Open it.

At this size a 30 m cell is several millimeters on a print, and that is the limit. A view of one ahu, a platform a few meters high, would be modeled parts set on ground the data cannot describe, and it would have to say so.

A dataset with strings attached

The search also found a 3D model of the Rano Raraku quarry, made from drone photographs for a 2025 study by Lipo, Hunt, Pakarati and others [PLOS One 2025]. It is the one place on the island where public data is finer than 30 m. Before using it, the chat opened the dataset's record [Zenodo] and found two things that change the plan:

So the quarry is the next step: convert only the part of the mesh that covers the quarry into a height grid, add it to isoscape as a source that carries its noncommercial license, and draw the quarry close up. When you find data like this, ask about its size, its format and its license before any download. Here the license was fine for teaching, and the 12 GB mesh is a project of its own.

Limits and care

Make your own

The code and the region files for every view on this page are on GitHub. To start your own map, open a chat that can run code and ask it to use the GitHub repo busse/isoscape to make an isometric map of your place. How it's made explains each step isoscape takes. Prints of the published maps are sold at chrisbusse.com, which supports the project.

Sources

Open this in the explorer