Detection of strawberry bloom phenology based on YOLO-RCMC and flower opening scale

Accurate estimation of the phenological distribution of strawberry flowers is crucial for effective management because it directly affects strawberry fruit growth quality. To address the challenges of insufficient detection accuracy and poor real-time performance of existing methods, we proposed a deep learning approach for estimating flower phenology using a model called flower-opening-YOLO-RCMC, which stands for You Only Look Once – Rectangular Self-Calibration Module (RCM) and RCM Variant Module (C2f-RCM and C3k2-RCM). First, the transition between the Early and Full phases of strawberry flower phenology is difficult to distinguish with the naked eye, which limits the precision of dataset segmentation. To address this issue, this study established a correlation between flower opening and flower phenology through pollen viability experiments and calculations of flower opening, thereby facilitating precise dataset division. Subsequently, a deep learning model, YOLO-RCMC, was developed to detect strawberry flower phenology. This model incorporates an RCM into its backbone to enhance its ability to extract the foreground phenological features in complex field environments. In addition, the RCM variants were integrated into the head of the model to reduce classification errors and improve detection accuracy. The experimental results identified the threshold for flower opening in both the Early and Full stages. The proposed YOLO-RCMC achieved the shortest inference time (9.1 ms) and the best performance with a mAP50 of 89.0% among the YOLO models. This method enables accurate detection of strawberry flower phenological distribution across varying distances and angles, providing valuable decision-making support for robotic flower management and pollination.