Deep learning-based application for automated leaf counting in medicinal orchid Dendrobium nobile L. using YOLOv5

Abstract Deep learning-based methods have shown promise for automated leaf detection and counting, particularly in rosette species. However, automated leaf counting remains limited for orchids such as Dendrobium nobile , which have a complex non-rosette architecture, alternate phyllotaxy, overlapping leaves, and view-dependent occlusion. This study developed an automated leaf-counting framework for D. nobile using YOLOv5-based object detection and translated the selected model into a GUI-based desktop application, DnLC ( Dendrobium nobile Leaf Counter). A DSLR camera captured high-resolution whole-plant images while each pot was manually rotated to acquire 16 multi-view frames over 360°. Roboflow was used to annotate 766 whole-plant images, generating approximately 10,490 leaf annotations. The dataset was divided into training, validation, and test subsets to evaluate YOLOv5 variants under transfer-learning and training-from-scratch strategies. Standard detection metrics, computational cost, and an Average Rank-based model-selection approach were used to assess model performance. Among the evaluated variants, YOLOv5l provided the most balanced performance and was integrated into DnLC for automated image loading, leaf detection, bounding-box visualization, and leaf-count estimation. Software-level validation on 354 independent unseen D. nobile images using CVPPP Leaf Counting Challenge metrics showed close agreement with manual counts, with DiC of − 0.95 ± 1.45, |DiC| of 1.15 ± 1.30, MAE of 1.10, MSE of 3.00, MAPE of 8.44%, and PA of 43.91%. Overall, DnLC provides a standardized, non-destructive leaf-counting tool for D. nobile phenotyping, supporting greenhouse growth monitoring, treatment-response evaluation, and high-throughput plant screening.