Viability detection model of tomato nursery using RGB images and deep learning yolo

Abstract

The viability of tomato (Solanum lycopersicum) seedlings is a critical factor determining the success of crop production and minimizing economic losses in agricultural operations. Currently, evaluating seedling quality in nurseries heavily relies on manual visual observation, which is subjective, labor-intensive, and inefficient for large-scale implementations. To address these limitations, this study aims to develop an automatic instance segmentation model using a deep learning approach based on the You Only Look Once (YOLO) architecture to measure the viability of tomato seedlings non-destructively. The methodology involved acquiring 1875 high-resolution RGB images of tomato seedlings in a controlled greenhouse environment. The dataset was annotated into four specific growth categories (non_germinated, radicle_emergence, cotyledon, true_leaf) and utilized to train YOLOv8-Seg and YOLO11-Seg models. The performance of the models was evaluated using standard computer vision metrics. The results showed that the YOLOv8-Seg model achieved the highest detection performance with a mean Average Precision (mAP@0.5) of 92.5%, a precision of 90.8%, and a recall of 87.3%. In comparison, the YOLO11-Seg model obtained an mAP@0.5 of 91.9%. Furthermore, the YOLOv8-Seg model demonstrated a rapid inference speed of 227 Frames Per Second (FPS) or 4.4 milliseconds per image. These quantitative findings demonstrate that the YOLO-based approach provides a reliable, highly accurate, and efficient solution for standardizing real-time seedling selection processes in modern smart nurseries.