What it is
Large-scale satellite dataset for field boundary instance segmentation with 1.48M image patches and 74.0M field masks across train, test, and zero-shot splits. 1.48M multi-resolution satellite patches (0.25-10 m) with 74.0M per-field instance polygons, organized into train, test, and a 100-country zero-shot test split.
Key results
- 1,478,096 image patches
- 74,015,889 field instance masks
- MIT license; 0.25-10 m imagery + 100-country zero-shot split
Useful for training and benchmarking field boundary delineation models, including zero-shot cross-country evaluation.