osapiens Challenge Makeathon 2026

Detecting Deforestation from Space

This repository contains the materials for the osapiens Makeathon 2026 challenge on deforestation detection from multimodal satellite data. The goal is to build a system that identifies deforestation events after 2020 using noisy, heterogeneous geospatial inputs and weak supervision signals.

Start Here

If you are new to the repository, use these files in this order:

1. osapiens-challenge-full-description.md for the written challenge brief and context.
2. challenge.ipynb for the full walkthrough of the dataset structure, label encodings, visualizations, and submission example.
3. download_data.py for the dataset download entrypoint used by the project.

Repository Guide

• challenge.ipynb: Main challenge notebook with data layout, modality descriptions, label definitions, examples, and submission guidance.
• osapiens-challenge-full-description.md: Full challenge description.
• download_data.py: Downloads the challenge data from S3 into ./data.
• submission_utils.py: Utility for converting prediction rasters into submission-ready GeoJSON.
• Makefile: Convenience targets for environment setup and data download.

Setup

Create the virtual environment and install the dependencies:

make install

Download the dataset:

make download_data_from_s3

This uses download_data.py and stores the files under:

data/makeathon-challenge/

Dataset Layout

After downloading, the notebook expects the data in the following structure:

data/makeathon-challenge/
├── sentinel-1/
│   ├── train/{tile_id}__s1_rtc/{tile_id}__s1_rtc_{year}_{month}_{ascending|descending}.tif
│   └── test/...
├── sentinel-2/
│   ├── train/{tile_id}__s2_l2a/{tile_id}__s2_l2a_{year}_{month}.tif
│   └── test/...
├── aef-embeddings/
│   ├── train/{tile_id}_{year}.tiff
│   └── test/...
├── labels/train/
│   ├── gladl/
│   ├── glads2/
│   └── radd/
└── metadata/
    ├── train_tiles.geojson
    └── test_tiles.geojson

Explore the Notebook for the Full Challenge Walkthrough