Detección automática de estados fenológicos en Rosa spp. mediante redes neuronales convolucionales YOLOv8

This study evaluated the performance of YOLOv8 convolutional neural network models for the automatic detection of phenological stages in greenhouse-grown cut roses (Rosa spp.). Image acquisition was conducted in a commercial greenhouse in Tocancipá, Colombia, using a ground-based mobile platform equipped with RGB cameras, thereby avoiding the operational limitations of unmanned aerial vehicles (UAVs) in enclosed environments. Images were collected during five sampling periods using a Nikon camera mounted on the mobile platform across five hydroponic benches, each divided into five 6.4-m plots, for a total of 25 plots. In total, 2,000 images and 4,653 annotated objects were obtained across 9 classes (8 phenological and 1 multipurpose). Model performance was evaluated using precision, recall, F1-score, mAP50, and mAP50–95. Individual models outperformed the multipurpose model, with the C_stage model achieving an F1-score of 0.87 in validation and 0.84 in testing. The multipurpose model required extending training to 200 epochs to achieve convergence, resulting in improved performance (F1-score = 0.75 and Precision = 0.78 in validation; F1-score and Precision = 0.72 in testing), indicating its potential for simultaneous multi-stage detection under greenhouse conditions. Correlation analysis showed that object size was the main factor influencing model performance (r ≥ 0.90). At the same time, the number of labeled samples per class had only a weak relationship with the metrics. This explained the higher accuracy in phenological stages with larger and more distinctive floral structures (C_stage, S_color) and the lower performance in early stages (rice, chickpea), whose buds occupied less than 0.3% of the image area.