[Defended thesis] Etienne David: The challenge of robust trait estimates with DeepLearning on high resolution RGB images

Etienne is one of the #DigitAg labelled PhDs

Etienne defended his thesis on 2nd November at 9:30 AM in AgroSciences, at the University of Avignon (Campus Jean Henri Fabre – room B020 ). 

The challenge of robust trait estimates with DeepLearning on high resolution RGB images

  • Start Date: October 2018
  • University: University of Avignon
  • PhD School: A2E – ED 536 Agrosciences & Sciences
  • Field(s): Computer Sciences, Agronomy
  • Doctoral Thesis Advisor: Frédéric Baret, Inrae, UMR Emmah Avignon
  • Co-supervisors : Frédéric Baret and 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

Abstract: High throughput plant phenotyping, especially in the context of open field acquisitions, relies on the interpretation of data from different sensors implemented on various vectors such as tractors, robots or drones. Initially, these data were interpreted using remote sensing algorithms that exploit the spatial resolution of the signal. Since 2015, however, progresses of “Deep Learning”, based on the training on examples, has already obtained promising results for measuring the rate of cover, counting plants or organs. It uses learned convolution layers, can take advantage of the spatial organization of the signal. The advantage of these methods is that they are based on Red-Green-Blue (RGB) sensors, which are much less expensive than multi- or hyperspectral imagers. However, these methods are sensitive to changes in the distribution between the data used in training and the predicted data. In practice, variable prediction errors from site to site can be observed using these methods. The objective of the thesis is to understand the causes of these variations and propose solutions for reliable phenotypic trait estimates using Deep Learning. The study focuses on detecting plants and organs from high-resolution RGB images acquired in the field. Our work first focused on the constitution of diversified image databases from different locations and stages of development for plant emergence (maize, beet, sunflower) and wheat ears, which allowed the publication of two annotated databases, grouping 27 acquisition sessions for the drone and 47 for the ear detection. The datasets demonstrate the performances difference between the published results and ours due to the change in distribution. To go beyond the limits of the usual methods, we organized two data competitions, the Global Wheat Challenges, in 2020 and 2021, which allowed us to obtain solutions trained for robustness on a different data set than the training one. The analysis of the solutions showed the importance of the training strategies for robustness beyond the architectures used. We have also shown that these solutions can be effectively deployed as a replacement for manual counting. Finally, we have demonstrated the inefficiency of training functions designed for robust training. Our work opens the prospect of a better evaluation of Deep Learning in the context of high-throughput phenotyping and thus of confidence in its use in real-life conditions.

Contact:  etienne.david [AT] inrae.fr​

Social Networks: ResearchGateTwitter – GitHub