[PhD’s Corner] Benjamin Deneu: Interpretability of distribution models of plant species communities learned through deep learning – Application to crop weeds in the context of agro-ecology

Benjamin Deneu is one of the #DigitAg co-funded PhDs

Interpretability of distribution models of plant species communities learned through deep learning – Application to crop weeds in the context of agro-ecology

  • Start Date: 01/10/2019
  • University: MUSE  University of Montpellier
  • PhD School: I2S (Information Structures Systems)
  • Field(s): Computer science
  • Doctoral Thesis Advisor: Alexis Joly (Inria LIRMM), François Munoz (Université Grenoble Alpes)
  • Co-supervisors : Pierre Bonnet (Cirad AMAP), Maximilien Servajean (Université de Montpellier LIRMM)
  • Funding: #DigitAg – Inria
  • #DigitAg: Axis 5, Challenge 1

Keywords: Agro-ecology, Deep Neural Networks, Transfer learning, Interpretability, Interactions, Landscape, Agricultural Practice, Biodiversity, Crop Weeds

Abstract: The modelling of interactions between biodiversity, landscape and agricultural practice is one of the major challenges of agro-ecology. Very recently, environmental species distribution models based on deep neural networks have begun to emerge. These first experiments showed that they could have a strong predictive power, potentially much better than the models used traditionally in ecology. One of their advantages is that they can learn an environmental representation space common to a very large number of species so that the prediction performance can be stabilized from one species to another. A first objective of the thesis will be to extend such transfer learning principle to the context of agro-ecology. In particular, data characterizing the landscape and the agricultural practices will be integrated for the prediction of crop weeds and/or associated functional traits. The second objective of the thesis will be to remove the lock on the interpretability of these models in order to deduce new tangible knowledge in agro-ecology. This will include qualifying the environmental representation space learned by the neural network in its terminal layers, typically the last layer of description on which the final linear regression or classification is based. The variables (neurons) in this representation space necessarily correspond to deterministic ecological and environmental patterns, but their exact nature is totally unknown. In the case of the deep agri-environmental models targeted in the thesis, these patterns will also integrate information from the landscape and agricultural practices. Their analysis will provide a better understanding of whether or not massive integrative approaches, based on a wide variety of input data, are needed, or whether they should focus on certain key factors.

 

Contact:  benjamin.deneu [AT] inria.fr, Cell: 06 89 70 36 08

 

Social Networks:  ResearchGate LinkedIn 

Papers in international journals