Etienne is one of the #DigitAg labelled PhDs
Design of Deep Learning methods for crops characterization
- Start Date: October 2018
- University: Avignon University
- PhD School: A2E – ED 536 Agrosciences & Sciences
- Field(s): Computer Sciences, Agronomy
- Doctoral Thesis Advisor: Frédéric Baret, Inra EMMAH Avignon
- Co-supervisors : Frédéric Baret et Samuel Thomas, Arvalis institut du végétal
- Funding: Cifre Arvalis
- #DigitAg: Axes 5 & 6 – Challenges 2 & 3
Keywords: High throughput phenotyping ; Deep Learning; Traits; LAI; Remote sensing, Machine Learning, Computer vision; Close range romete sensing
The recent introduction of digital technology in agriculture provides a massive amount of data through fixed field sensors (IoT for Internet of Things), or mobile sensors carried by drones, robots or tractors. These proxi-detection techniques provide access to certain crop characteristics that can be monitored during the growth cycle. Their integration into structure-function models should make possible to accurately describe certain processes in the functioning of plants. This description, which is a priori independent of environmental conditions, can thus be linked to the genotype. It would then be possible to choose, in a given pedo-climatic context, which is the best combination of operating characteristics leading to the best productivity taking into account environmental and economic constraints.
The mass of data thus produced, most of the time in the form of images, contributes to the emergence of the “Big-Data” in agriculture. The transformation of this mass of data into useful information now appears to be the main limiting factor. Recent developments in computer science, particularly in machine learning, make it possible to effectively exploit this mass of information, particularly that contained in images. These techniques therefore represent an immense potential for crop characterization.
The thesis focuses on wheat and maize crops. Three important issues will be addressed: – Plant and organ localization: The lab’s recent work shows a superiority of deep learning algorithms in organ localization and counting ; – Estimation of the leaf index. The leaf index is a dimensionless quantity that represents the leaf area per unit of soil area. IPM is relatively easy to estimate at early stages, but access is very imprecise for high IPM levels ; – Extraction of descriptors of vegetation cover structure: The introduction of new proxy-detection techniques such as LiDaR and very high resolution RGB or multi-spectral imaging provides access to new indicators on plant structure.
Contact: etienne.david [AT] inra.fr