Computational Models Growth & Development

Deep learning techniques used to predict missing data

A new model produces 3D trees that are close to reality using predicted data.
You can listen to this page as an audio file.

Plant growth modelling can help us predict how trees will respond to climate change throughout their lifetime. Accurate predictions require detailed information about how plant growth is controlled by endogenous processes, driven by the expression of the plant’s genotype, and of exogenous processes, driven by the interaction between the plant and its environment.

Unfortunately, it is difficult if not impossible to observe and measure endogenous effects on plant processes that regulate the growth, branching and death rates of plant structures.

Imagine measuring the rate of root growth over the lifetime of a tree.

This makes estimating the parameter values for endogenous growth equations required for modelling whole-plant growth challenging. Parameters that cannot be directly measured and only estimated are termed “hidden parameters.”

CIRAD researcher Dr. Jean-François Barczi and colleagues overcame this lack of data using deep learning techniques. Their paper, published in in silico Plants describes their method to predict values for these hidden parameters using generative deep neural networks. As a result, they were able to accurately model the impact of environment on trees.

This paper describes the development of RocoCau, a new structural whole plant growth model that describes shoot and root growth and root/shoot interactions (fig. 1). RoCoCau was linked to TOY, a new functional model plugin that simulates interactions between shoot and root compartments of trees facing varying climates.

Figure 1: Workflow for predicting tree architecture under varying environmental conditions using RoCoCau+TOY.

The hidden parameters of TOY were calibrated using model inversion. That is, the authors identified the correct model input values by assessing the accuracy of the model output those parameters produced. To do this, the authors ran RoCoCau+TOY simulations using 360,000 random TOY hidden parameter and climate values. A deep neural network was trained on this simulated database to predict the correct hidden parameter values of TOY. The trained network was then validated on a separate database to check if the predicted input values were able to produce model outputs similar to the outputs produced using the original values.

Figure 2: Predicted versus true values for the impact of light and water availability on architectural development.

They found that the datasets were able to produce simulated trees that are close to reality. Using predicted hidden parameter values, RoCoCau+TOY was able to predict the impact of water and light availability on the architectural development with 98% accuracy (fig 9). The accuracy of predicted shoot death threshold, branching threshold, and apical growth factor was 95% (fig 8).

Figure 3: Predicted versus true values of the 3 hidden shoot parameters: death threshold, the branching threshold, and the apical growth factor.


Abel Louis Masson, Yves Caraglio, Eric Nicolini, Philippe Borianne, Jean-Francois Barczi, Modelling the functional dependency between root and shoot compartments to predict the impact of the environment on the architecture of the whole plant. Methodology for model fitting on simulated data using Deep Learning techniques, in silico Plants, 2021, diab036,

This manuscript is part of in silico Plant’s Functional Structural Plant Model special issue.

The source code of the RoCoCau simulator with the plant parameter files used for this study are freely available at The source code of the neural network calibration tool can be downloaded from

%d bloggers like this: