/ReforesTree

🌴 A dataset for estimating tropical forest biomass based on drone and field data

Primary LanguageJupyter Notebook

ReforesTree 🌴

We are excited to share the ReforesTree dataset! 🎉

We introduce the ReforesTree dataset in hopes of encouraging the fellow machine learning community to take on the challenge of developing low-cost, scalable, trustworthy and accurate solutions for monitoring, verification and reporting of tropical reforestation inventory.

illustration of all sites

This is a dataset for the following 6 agroforestry sites

In alphabetical order

  1. Carlos Vera Arteaga
  2. Carlos Vera Guevara
  3. Flora Pluas
  4. Leonor Aspiazu
  5. Manuel Macias
  6. Nestor Macias

Dataset Components

For each site the data we publish consists of four components free for use:

  1. 🛸 Raw drone RGB images (see wwf_ecuador)
  2. 🌴 Hand measured tree parameters (diameter at breast height, species, biomass, and location) of every tree (see field_data.csv)
  3. 🔲 Set of bounding boxes of trees for each site cleaned by hand and labeled as banana or not banana (see annotations/cleaned)
  4. ↔️ Mappings of these bounding boxes with tree labels based on GPS location (see mappings/final)

Usage

Dataloader in PyTorch

Thanks to the torchgeo team, you can download the dataset through the ReforesTree data loader

from torchgeo.datasets import ReforesTree
ds = ReforesTree(root="data/reforestree/", download=True, checksum=True)

Raw data

You can download the raw data from dropbox and put the "data" folder in the main repo. All processed data is available directly to use, but if you want to process it yourself, feel free to only download "www_ecuador" and "field_data.csv" and follow the tutorial below.

Alternatively, we are hosting a version of the dataset in zenodo.

Tutorial

In the tutorial you'll find the steps to recreate (and hopefully improve) the dataset and how to use it.

Please read our paper here. For any questions, please reach out to gyri.reiersen@tum.de or david.dao@inf.eth.ch