Image based root system phenotyping : a new pipeline using DeepLearning techniques and plant modeling

Details

Supervisors

Lobet, Guillaume ; Draye, Xavier

Faculty

Faculté des bioingénieurs

Degree label

Master [120] : bioingénieur en sciences agronomiques

Abstract

The roots of plants play major roles in their development but their study remains very complex. In the recent years we have seen the development of more and more sophisticated modalities to study them. These new techniques have made possible the expansion of a still underexplored field of biology, the Root Phenotyping. The availability of immense quantities of photos in the field of phenotyping represents a great opportunity to apply Artificial Intelligence techniques. The objective of this master thesis was therefore to participate in this effort by using advanced Machine Learning techniques to establish links between the parameters that are easy to extract (number of tips) and the more useful but less accessible information (distance between lateral roots). The first part of this work aimed to show that it is possible to use artificial intelligence and in particular Deep Learning (DL) technique to address this issue. The second part examined to which extent these techniques are applicable to real phenotyping research. Also, a key part of using a DL algorithm is the need of a high quality training database. For this purpose we used CRootBox, a structural model for root system. CRootBox allowed us to generate as many roots systems as we wanted and to train a neural network on these virtual roots. We have succeeded in this work in demonstrating that it is possible to determine certain biological parameters that are expensive to extract (i.e. distance between lateral roots) from readily available data. The results obtained using real scans of Arabidopsis thaliana Col0 root systems further validated the predictive potential of our Deep Learning approach. This method and the different scripts have been developed to be the basis for future research. In particular, the whole code has been designed so that only a few modifications are needed to change the objective of the simulation.