# SelvaBox: 83,000 Tropical Tree Crown Bounding Boxes on 4.5 cm Drone Orthomosaics from Brazil, Ecuador and Panama (CanopyRS)

## Facts

- Provider: SelvaBox / CanopyRS authors
- License: CC-BY-4.0 (https://huggingface.co/datasets/CanopyRS/SelvaBox)
- Size: 67 files
- Format: Parquet (in sampled files)
- Stored at: Hugging Face, https://huggingface.co/datasets/CanopyRS/SelvaBox
- Access: public, readable without an account
- Category: Environmental Science
- Source: https://huggingface.co/datasets/CanopyRS/SelvaBox
- Page: https://scigantic.com/archive/7be6f6e4-cf9b-5168-9494-fdfee4919c31
- Keywords: Environmental Science, image segmentation, benchmark, Parquet, Hugging Face

## About

A high-resolution benchmark for individual tree crown detection in dense tropical rainforest canopies, introduced in the SelvaBox paper (ICLR 2026, arXiv 2507.00170) and used to train and benchmark the CanopyRS tree detection and segmentation pipeline. Fourteen drone (UAV) RGB orthomosaics from Brazil, Ecuador and Panama were resampled to 4.5 cm ground sampling distance and tiled, with more than 83,000 unique human-drawn bounding boxes around tree crowns. Training tiles are 3555 x 3555 pixels (160 m) with 50 percent overlap; validation and test tiles are 1777 x 1777 pixels (80 m) with 50 and 75 percent overlap so the largest crowns fit in a tile, giving 585 train, 387 validation and 1,477 test tiles. Labels are binary tree versus background, without species. Data are Parquet shards under data/ (about 31 GB) with TIFF images and COCO-format boxes, loadable or streamable with datasets.load_dataset; GeoPackage annotations and AOIs are on the gpkg branch. Useful for forest monitoring, canopy mapping, tree counting and object detection under heavy overlap. Licensed CC BY 4.0 by the authors; cite the SelvaBox paper (arXiv 2507.00170).

## Load the data

```python
from datasets import load_dataset
ds = load_dataset("CanopyRS/SelvaBox")
print(ds)
```

## Data structure

File types in sampled files: Parquet

Example file: `data/test-00000-of-00024.parquet`

## Open it

- In a hosted notebook with the data mounted: https://scigantic.com/archive/7be6f6e4-cf9b-5168-9494-fdfee4919c31
- Full record as JSON (schema card included): https://api.scigantic.com/api/archives/7be6f6e4-cf9b-5168-9494-fdfee4919c31
- Over MCP: `get_schema_card` with id `7be6f6e4-cf9b-5168-9494-fdfee4919c31` at https://api.scigantic.com/mcp

## Related topics

- [Parquet datasets](https://scigantic.com/topics/parquet)
- [Benchmark datasets](https://scigantic.com/topics/benchmark)
- [Image segmentation datasets](https://scigantic.com/topics/image-segmentation)
- [Environmental Science datasets](https://scigantic.com/topics/environmental-science)
