A high-quality apple flower detection dataset for precision agriculture

Accurate detection of apple flowers and buds at the ground level is a prerequisite for precision orchard management, yet it remains constrained by the scarcity of densely annotated, in-canopy datasets. In this study, we present a dataset for apple flower and bud detection collected in a commercial apple orchard in Shaanxi Province, China. The dataset comprises 3,117 images and 52,181 annotated instances, capturing the structural complexity of tree canopies under variable illumination and partial occlusion. A key characteristic of this dataset is the concurrent presence of fully opened flowers and unopened buds, alongside considerable morphological variability arising from differential petal loss. High-precision bounding boxes distinguish the two classes, with annotations provided in XML, YOLO (TXT), and COCO (JSON) formats to ensure broad compatibility. Standardized train/validation/test splits (8:1:1) are supplied, and baseline performance is established using representative architectures, including YOLOv11n, RT-DETR, and DEIMv2. Evaluation results reveal a clear performance gap between buds and flowers, underscoring the inherent detection difficulty of buds due to morphological ambiguity and occlusion. By providing this high-quality dataset, we aim to support the development of robust vision algorithms for early bloom monitoring, facilitating critical management decisions such as site-specific flower thinning and yield estimation.