r-geo-viz

About this project

This repo was created in conjunction with an AI Agentic Coding workshop at the UCSB Library

• This repo demonstrates a geospatial exploratory visualization of pre-1900 wildfire data in the US Pacific Northwest
• It uses R and Terra
• The scripts were created using various LLMs underneath OpenCode, including CIT’s Gemma 4 and the DREAM Lab’s granted GCP.
• Viewing the maps against the data allowed me to conclude that any fire marked 1900 meant that it was BEFORE 1900.
  • That is a detail that was NOT stated in the metadata.

Explore the code and outputs here

Things I learned

1. Between sessions, telling the agent to run all the scripts to learn about the repo is essential
2. When moving to a new Coder workspace, the AI will iterate over R scripts to get libraries set up. It will do it largely unsupervised
3. During one iteration, I had the agent looking for its own datafiles on the web. As a skilled librarian, I was able to web search, identify and download California and Oregon fire perimeter files faster than the AI. I was able to see it get stuck fetching California fire perimeters from 2 different Arcgis Online addresses (both Living Atlas and cnra.ca.gov). It was nice to be able to pause the agent and then point it directly at the files I fetched and uploaded to my Coder workspace.
4. The models used by OpenCode possess machine vision capabilities, allowing them, to some extent, to verify the results saved to the output/ directory.

Try it yourself

Place the agents.md file and this georeferenced map of the Olympic Peninsula in Washington State in a new project space and ask the agent to do some exploratory visualization.

Questions

1. Are there areas on the scanned map of the Olympics that represent 19th century burns?
2. Do these area correspond to known digital data sources / contemporary research data?
3. Was the agent able to find fire perimeter data, or did you have to get it yourself?

Tips

• you are starting without base data, so ask opencode for a shapefile of washington counties, points of interest, or rivers
• specify where you want your output
• you may want to provide your own fire perimeter data. (Stuck? Try this one)

Didactic

Here is a more detailed outline of how OpenCode and I got to the 9 R scripts This page was started with the AI, and heavily edited by me.

Here is an unedited version of what Gemini has to say about how this was made I prompted this file because I was curious about how well Gemini could interpret its own code in new sessions.