Actinidia chinensis Phenology Balanced Dataset (Portugal, 2025)

The Actinidia chinensis Phenology Balanced Dataset is a stratified and class-balanced subset derived from the Multi-Modal Actinidia chinensis Phenology Dataset, prepared explicitly for training hierarchical detection models that target female flower phenological staging.

The source material comprises the Labelled Images component from the original dataset, which contains smartphone-acquired imagery of kiwifruit reproductive structures annotated in Pascal VOC format according to a 17-class hierarchical taxonomy organised across three levels: structure (bud, flower, fruit), gender (female, male), and BBCH-adapted phenological stage.

The balanced dataset was generated through a two-stage processing pipeline:

1. Split stage: The source imagery was partitioned into test (99 images) and train+validation (1 556 images) subsets using stratified sampling to ensure proportional class representation in the test set across all hierarchical classification levels.
2. Balancing stage: Images exhibiting disproportionate annotation density for over-represented classes were systematically removed from the train+validation subset to facilitate subsequent balancing operations (1 311 images retained). Augmentation operations were applied iteratively until target class distributions were achieved (1 311 original + 649 augmented). Composite operations were constructed by combining one geometric transformation (flip, scale_rotate and downscale) with one appearance transformation (bright_contrast, grid_distortion, grid_dropout, unsharp and motion_blur).  All augmentations were implemented using the Albumentations library, which includes bounding box coordinate transformation.

Phenological stage labels for bud and male flower classes were consolidated to their parent categories, reducing the taxonomy from 17 to 9 object classes while preserving full phenological granularity for the female flower pathway.

The processed dataset is organised into two directories (test and train+validation), each containing an Images folder with JPEG files and an Annotations folder with corresponding Pascal VOC XML files.

The hierarchical detection pathway comprises three sequential classification tasks. The following tables present the class distribution for each level, demonstrating the balance achieved through the processing pipeline. Image counts indicate images containing at least one annotation of that class; individual images may contain multiple classes.

The first level performs structure detection, distinguishing between bud and flower structures:

The second level performs gender classification on detected flowers. Images containing only bud, fruit, or no annotations serve as background for this classification task:

The third level performs phenological stage classification on detected female flowers. Images containing only male flowers, generic flowers, buds, fruit, or no annotations serve as background for this classification task:

The train+validation subset should be used with k-fold cross-validation for model development and hyperparameter optimisation. The test subset is reserved for final model evaluation and should not be used during training or validation. Users requiring the complete 17-class taxonomy or male flower phenological staging should refer to the original dataset.