Geo-OLMs Repo

Setup

1. Clone this repo

2. Setup OpenAI API key and store in environment variable OPENAI_API_KEY (see OpenAI API Dev Quickstart).

export OPENAI_API_KEY='your-api-key-here'

3. Install requirements: slight chance the list isn't up-to-date! need to check, so better that you install inside a virtualenv

pip3 install -r requirements.txt

Data

We consider several open-source datasets (xview1, ...). For agentic experimentation purposes, we're not directly touching the underlying imagery or SQL tables. Instead the list of images with their metadata (dates, coordinates, etc.) are preprocessed into DataFrames. The instructions how ingest each dataset can be found under datasets/geeo.

** Note: We've preprocessed the files. The link to download the *.gpkg files is [send over email]. Unzip this and have all files under datasets/gpkgs/geeo25.

Getting Started [WiP]

The main.py launches the geo-agent. We are updating the flow by adding config file support, but for now:

python main.py

Evaluation [WiP]

Output is stored under ./results/. We are updating the flow on how metrics are reported, but for now:

python3 run_eval.py

Citation

@inproceedings{stamoulis2025geoolm,
  title={Geo-OLM: Enabling Sustainable Earth Observation Studies with Cost-Efficient Open Language Models \& State-Driven Workflows},
  author={Stamoulis, Dimitrios and Marculescu, Diana},
  booktitle={ACM Computing and Sustainable Societies (COMPASS)},
  year={2025},
  note = {(Accepted)}
}