Multi-Modal Actinidia chinensis Phenology Dataset (Portugal, 2023-2024)

The Multi-Modal Actinidia chinensis Phenology Dataset provides two complementary components organised hierarchically across plant structures, gender, and phenological stages. The labelled images component provides 1 665 annotated images (1 024 × 1 024 pixels) with bounding box annotations in Pascal VOC XML format, covering phenological stages from bud emergence to fruit set. The georeferenced videos component provides 24 MP4 videos (3 840 × 2 160 pixels) with corresponding GPX coordinate files and Excel validation files containing manual ground truth counts across four sampling zones. 

The labelled images component contains 1,665 annotated images (1024×1024 pixels) with bounding box annotations in Pascal VOC XML format, covering phenological stages from bud emergence to fruit set. Images were selected from the Actinidia chinensis Bud and Flower Dataset (Portugal, 2024) and the Actinidia chinensis cv. 'Hayward' Flower Dataset 2024 based on uniqueness calculations using the Fiftyone platform to identify and remove duplicate or near-duplicate images, avoiding accidental concept imbalance that can bias model learning. Selected images were preprocessed and standardised to 1024×1024 pixel resolution to optimise compatibility with computer vision architectures.

The georeferenced videos component contains 24 MP4 videos (3 840 × 2 160 pixels and 30 fps) with corresponding GPX coordinate files and Excel validation files containing manual ground truth counts. Data were collected from an orchard in Vila do Conde, Portugal (41°17'27.48"N, 8°38'57.57"W) using a Xiaomi Poco X6 Pro smartphone across four 5×5 meter sampling zones distributed throughout the orchard.

Applications include:

• Training and validation of computer vision models for phenological stage detection
• Benchmarking automated counting algorithms
• Spatial-temporal analysis of phenological heterogeneity
• Optimisation of assisted pollination strategies