Abstract:
Abstract: Plant phenotype has been one of the most important indicators to select and breed superior varieties in modern agriculture. The traditional phenotypic analysis of leafy vegetables cannot fully meet the requirement of large-scale production in recent years, such as the slow speed, large error, and limited dimension. Moreover, most morphological measurements have been currently confined to only several plants (such as maize, cotton and sorghum), particularly with the complex procedure and low automation. Taking the Chinese cabbage sprout as the research object, this study aims to extract the key phenotypic parameters of the plant using the high-throughput reconstruction and automatic segmentation of stem and leaf. Firstly, a multi-view geometry was utilized to reconstruct the three-dimensional model of Chinese cabbage sprout from sequence images. A series of pre-processing operations were used to establish the three-dimensional model of pure plants with the actual scale, including scale recovery, background removal, point cloud denoising, and uniform simplification. Secondly, the stem and leaf organs were automatically segmented using the convexity criterion and random sampling consistency. Thirdly, the principal component analysis and directed bounding box were combined to measure the plant height and width for the phenotypic parameters. The number of leaves was counted by the cluster segmentation of leaf clusters. The shortest path searching was selected to accurately calculate the leaf length and width. A greedy projection triangulation was used to calculate the leaf area. An HSV model and mean color histogram were utilized to measure the color characteristics for distinguishing health status of plants. The classification accuracy, recall, and F1-score were 0.922, 0.938, and 0.930, respectively. Finally, the segmentation experiments were carried out on the six varieties of Chinese cabbage sprouts at different growth stages. A comparison was then made with the ground truth. It was found that the parts belonging to stem and leaf were segmented correctly. The average precision, recall, and F1-score of stem and leaf organ segmentation were 0.961, 0.940 and 0.943 respectively, indicating better performance than the smoothing threshold-based and the color-based region growth. In addition, 50 samples were tested to verify the measurement. A regression analysis was performed between the algorithmic and manual measurements of Chinese cabbage sprout plant traits. The experimental results showed that the determination coefficients of plant height, plant width, leaf length, leaf width, leaf area, and leaf number were 0.987, 0.982, 0.984, 0.985, 0.922, and 0.924, respectively. The mean absolute percentage errors were 1.659%, 1.643%, 1.417%, 2.486%, 8.258%, and 6.000%, respectively. Among them, the Root Mean Square Error (RMSE) of plant height, plant width, leaf length, and leaf width were 0.261, 0.313, 0.174, and 0.100 cm, respectively. The RMSE of leaf area and number were 1.608 cm2, and 0.283, respectively. Consequently, the automatic measurement was realized for the seven key phenotypes of Chinese cabbage sprouts, including the plant height, leaf length, and color in lower error. Therefore, the segmentation and measurement can be expected to extract the plant phenotypic parameters with irregular leaf shape, especially in the young leaves of plant canopy, with high efficiency of phenotype extraction. The finding can provide an effective technical means for the efficient and accurate phenotypic analysis of leafy vegetables.