Phenotypic measurements of broadleaf tree seedlings based on improved UNet and Pix2PixHD

The nondestructive, high-precision measurement of the phenotypic parameters of broadleaf tree seedlings is critical for seedling growth monitoring. In this paper, we take broadleaf tree seedlings as the research object and design a complete set of equipment, models, and methods ranging from automatic tree seedling image acquisition to seedling image segmentation, branch and leaf separation, image restoration of occluded branches, and final plant phenotype measurement. The experimental results show that the mean intersection over union (mIoU) of the proposed segmentation model for tree seedling branches and leaves reaches 87.95 and 98.37 %, respectively, and that the mean pixel accuracy (mPA) reaches 93.16 and 99.24 %, respectively. The structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR) of branch restoration reach 98.5 % and 41.48 dB, respectively. By calculating the phenotypic parameters of tree seedling, we can keep the mean average precision error (MAPE) of the tree seedling height, ground diameter, canopy width, and canopy layer within 6 %. The results indicate that the proposed methods can more accurately extract the branch and leaf regions of a tree seedling and recover the missing parts of branches and trunks, providing a new nondestructive method of plant phenotypic measurement for broadleaf tree seedling cultivation and growth monitoring. •The process and methods for measuring the phenotypic parameters of broadleaf tree seedlings are thoroughly detaile.•The proposed segmentation network reduces artifacts like hollow regions and edge mis-segmentation for small targets.•The paper redefines metrics for plant height and canopy width in small seedlings, providing low cost and high accuracy.•Separating stems and leaves before image completion significantly improves the accuracy of phenotyping.