Fields of The World is an open-source project (links below). We hope to grow the dataset in the future with community contributions. 📄 arxiv.org/abs/2409.16252 🌐 fieldsofthe.world/ Data beta.source.coop/kerner-lab/… Code github.com/fieldsoftheworld/… Requests github.com/fieldsoftheworld/…

Models pretrained on Fields of The World perform better in transfer learning than models pretrained on more geo-limited datasets (e.g., Europe only). FTW models also show impressive zero-shot capabilities (e.g., this example in Ethiopia, where no training data was available).

I'm so excited to finally introduce Fields of The World (FTW!) - the largest & most diverse ML dataset for field instance segmentation in satellite images. FTW has >70k samples from 24 countries (4 continents), with semantic & instance labels + Sentinel-2 RGB-NIR from 2 dates. 🧵