Canopy RS (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 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.
The full documentation covers:
- 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 (arXiv preprint):
@misc{baudchon2025selvaboxhighresolutiondatasettropical,
title={SelvaBox: A high-resolution dataset for tropical tree crown detection},
author={Hugo Baudchon and Arthur Ouaknine and Martin Weiss and Mélisande Teng and Thomas R. Walla and Antoine Caron-Guay and Christopher Pal and Etienne Laliberté},
year={2025},
eprint={2507.00170},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2507.00170},
}If you use other datasets that we have preprocessed, please also cite the original authors of those datasets directly.