AerialYield-B2D dataset offers 30K annotated blueberries for ripeness AI
514 greenhouse images, 5 ripeness stages, and 30,195 labeled berries...
A new computer vision dataset aims to close the gap in agricultural AI for soft fruits. AerialYield-B2D, introduced by Iyyakutti Iyappan Ganapathi and colleagues, provides 514 real greenhouse images with 30,195 individually annotated blueberries spanning five ripeness stages: green immature, pale pink, pink-turns-purple, fully ripe, and over-ripe. Each image comes with class-specific binary masks, overall berry masks, semantic label maps, and image-level count tables, making it ready for training segmentation and counting models. The dataset was assembled from 424 smartphone images, 67 video-derived frames, and 23 DJI Fly drone captures, reflecting realistic greenhouse conditions.
Notably, the authors are careful to distinguish berry counts from yield: this release does not include harvest weight or per-area measurements, so models should treat counts as ripeness-stage proxies rather than production estimates. The dataset also includes SHA-256 hashes, source metadata, and recommended train/validation/test splits to ensure reproducibility. For researchers working on ripeness classification, class-imbalance handling, or drone-based orchard monitoring, AerialYield-B2D offers a dense, labeled benchmark that could help move precision agriculture beyond single-ripe-stage detection.
- 514 RGB images with 30,195 annotated blueberry instances across 5 ripeness stages
- Includes binary masks, semantic label maps, image-level count tables, and train/validation/test splits
- Built from smartphone, video, and DJI drone footage for controlled-environment agriculture
Why It Matters
Precision agriculture needs dense ripeness data; this dataset enables more accurate AI for harvest timing and yield forecasting.