Computational Models Growth & Development Life

Autonomous construction of 3D leaves

An assessment of deep generative networks in creating realistic 3D data
You can listen to this page as an audio file.

To model how plants grow and develop according to their genotypes, phenotypes, environment and/or management the use of a realistic three-dimensional (3D) canopy helps refine the predictions of light interception and photosynthesis. 

The development of 3D imaging techniques for estimating canopy structure, shoot growth and biomass has expanded during the last couple of years. However, securing sufficient 3D-scanned data is time-consuming because of the need to manually process the data. In large-scale simulations, single-plant or few-plant reconstructions are often duplicated for convenience, and therefore lack phenotypic diversity. Now deep generative models can be used to learn and create realistic 3D data.

Dr. Jung Eek Son, professor of plant science at Seoul National University, and colleagues generated leaf models and extracted their traits by using deep generative models. The authors scanned pepper plants at various stages of development using a high-resolution portable 3D scanner. Point clouds were obtained from the scans, then these were used to train the deep generative models that could generate leaves.

This image is of the workflow from 3D leaf scanning to leaf generation. In the first box, eight leaves with petioles scanned from above are shown. An arrow from this box shows that the data is divided into a training set for model training and a validation set for model validation. From here, the model is optimized.   The model that simulated the best randomly created leaves and that created leaves with the desired traits by manipulating the latent space was then selected. To represent this, a box contains 8 randomly generated leaves. A double arrow from the box demonstrates that linear interpolation led to gradual changes in leaf shape, and arithmetic operations in the latent space added or subtracted leaf traits for the existing leaves. New traits are shown in a box: seven leaves viewed from the top and sides demonstrating changes in size, inclination and curvature of the leaves selected from randomly generated point clouds. The second box portrays editing of leaf point clouds by simple arithmetic in the latent space. Various traits, such as size, inclination and curvature, were imparted to generated leaves.
Workflow from 3D leaf scanning to leaf generation.

The authors compared leaves generated using three deep generative models: variational autoencoder (VAE), generative adversarial network (GAN) and latent space GAN.

While a VAE takes raw data from an image, encodes it with lower resolution, and then reconstructs it, a GAN generates an image from noise, then discriminates the image based on the raw data to determine if it is real or fake. A latent space GAN (L-GAN) has the same basic structure and training method as a GAN but uses latent variables: features from which the learning system could detect or classify patterns in the input, instead of raw data.

Three deep generative model architectures are shown. In the first panel, VAE, the decoder acts as a generator, which inherits the structures of the AE. In the second panel, GAN, a generator, and a discriminator generate leaves from random noise. In the third panel, L-GAN, both the generator and the discriminator of the latent space operate on the latent variables.
Architectures of the three deep generative models. 

Reliable 3D phenotypes of pepper leaves were created by the deep generative models. Among the deep generative models, L- GAN showed the highest performance in generating realistic leaves.  “We compared several generative models for leaf generation. In this way, the shape of leaves could be controlled with linear interpolation and simple arithmetic operations. That is, the generative model includes morphological traits somewhere in the model parameters. The first step toward the practical use of deep generative models was achieved for autonomous creations of 3D plant models without complicated feature extraction” says Son.

While the 3D shape of leaves viewed from the top were used to train the deep generative models, the models were also able to generate images to assess latent variables such as inclination and curvature from the data.

Interpolation results of the size, inclination and curvature of the generated leaves and the distributions of leaf phenotypes. Various traits, such as size, inclination and curvature, were imparted to generated leaves.
Images created using latent variables.

Son concludes, “deep generative models can parameterize and generate morphological traits in digitized 3D plant models and add realism and diversity to plant phenotyping studies and models.”


Taewon Moon, Hayoung Choi, Dongpil Kim, Inha Hwang, Jaewoo Kim, Jiyong Shin, Jung Eek Son, Autonomous construction of parameterizable 3D leaf models from scanned sweet pepper leaves with deep generative networks, in silico Plants, Volume 4, Issue 2, 2022, diac015,

Leave a Reply

Your email address will not be published.

This site uses Akismet to reduce spam. Learn how your comment data is processed.

%d bloggers like this: