CanopyRS (Canopy Remote Sensing) is a pipeline designed for processing high-resolution geospatial orthomosaics to detect, segment, and (in the future) classify trees of various forest biomes. It supports state-of-the-art model architectures spanning both CNNs (Faster R-CNN, Mask R-CNN, RetinaNet) and transformers (DINO, Mask2Former, SAM 2, SAM 3). The pipeline includes components for tiling, detecting, aggregating, and segmenting trees in orthomosaics. These components can be chained together based on the desired application.
📖 Full documentation: hugobaudchon.github.io/CanopyRS
- [2026-02-02]: 📄 We released SelvaMask, a new preprint introducing a densely annotated tropical tree crown instance segmentation dataset and benchmarking SOTA methods like SAM3!
- [2026-01-26]: 🎉 Our paper SelvaBox, which introduces CanopyRS, was accepted at ICLR 2026! Come see us in Rio de Janeiro this April!
- [2024-11-15]: 🥇 Our team Limelight Rainforest won the $10M XPRIZE Rainforest competition, in part thanks to CanopyRS and SelvaBox!
See the Installation guide and Quick Start in the documentation.
No local GPU? You can also run CanopyRS in our Google Colab tutorial, which runs the full detect + segment pipeline (SelvaBox DINO detector + SAM2) on a sample orthomosaic on a free T4 GPU and visualizes the results, and whose cells you can reuse in your own notebook to run inference on your own data if desired. Note that setup takes ~20 minutes per session and Colab can't handle large orthomosaics, so it's best for quick tests or small datasets.
The full documentation covers:
- Installation — requirements and step-by-step setup
- Presets & Model Zoo — default configs and available pretrained models
- Pipeline & Components — how to configure and run the pipeline
- Data — downloading and using datasets (SelvaBox, SelvaMask, Detectree2, etc.)
- Training — training detectors and segmenters
- Evaluation — benchmarking and finding optimal NMS parameters
- API Reference — programmatic usage
If you use CanopyRS or SelvaBox in your research, please cite our paper (accepted at ICLR 2026):
@inproceedings{baudchon2026selvabox,
title={SelvaBox: A high-resolution dataset for tropical tree crown detection},
author={Hugo Baudchon and Arthur Ouaknine and Martin Weiss and M{\'e}lisande Teng and Thomas R. Walla and Antoine Caron-Guay and Christopher Pal and Etienne Lalibert{\'e}},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=GH7z1RURL6}
}If you use other datasets that we have preprocessed, please also cite the original authors of those datasets directly.