Improved YOLOv8 Model for Phenotype Detection of Horticultural Seedling Growth Based on Digital Cousin

Abstract

Crop phenotype detection is a precise way to understand and predict the growth of horticultural seedlings in the smart agriculture era to increase the cost-effectiveness and energy efficiency of agricultural production. Crop phenotype detection requires the consideration of plant stature and agricultural devices, like robots and autonomous vehicles, in smart greenhouse ecosystems. However, collecting the imaging dataset is a challenge facing the deep learning detection of plant phenotype given the dynamic changes among leaves and the temporospatial limits of camara sampling. To address this issue, digital cousin is an improvement on digital twins that can be used to create virtual entities of plants through the creation of dynamic 3D structures and plant attributes using RGB image datasets in a simulation environment, using the principles of the variations and interactions of plants in the physical world. Thus, this work presents a two-phase method to obtain the phenotype of horticultural seedling growth. In the first phase, 3D Gaussian splatting is selected to reconstruct and store the 3D model of the plant with 7000 and 30,000 training rounds, enabling the capture of RGB images and the detection of the phenotypes of the seedlings, overcoming temporal and spatial limitations. In the second phase, an improved YOLOv8 model is created to segment and measure the seedlings, and it is modified by adding the LADH, SPPELAN, and Focaler-ECIoU modules. Compared with the original YOLOv8, the precision of our model is 91%, and the loss metric is lower by approximately 0.24. Moreover, a case study of watermelon seedings is examined, and the results of the 3D reconstruction of the seedlings show that our model outperforms classical segmentation algorithms on the main metrics, achieving a 91.0% mAP50 (B) and a 91.3% mAP50 (M).