Exploitation des méthodes d’apprentissage profond pour le suivi de la dynamique des espèces dans un mélange de cultures

Durée : 2022 - 2023
Programme : DigitAG
Portée : Nationale
Instance segmentation
Semantic segmentation
Automatic vegetation characterization

The agroecological transition requires the development and the assessment of new multiperformant, resilient and sustainable agroecosystems. However, these systems are currently lacking high-throughput, non-destructive and objective observation tools. The data deficiency is even more critical because of the higher complexity of agroecosystems such as mixed crops as compared to single crop systems. High-throughput observation tools based on close range imagery therefore appear as essential for rapidly characterizing and hence better understanding these new agrosystems. However, if these tools have now reached a certain maturity for the monitoring of monospecific crops, their use in agroecology remains limited. This project aims to understand to what extent close-range imagery can be used for field monitoring of the dynamics of the proportion and structure of species in a crop mixture.
The methodological approach involves a preliminary step of species identification within the canopy using deep learning models. A new database of annotated images will be therefore first created, and based on data acquired on multispecies crop canopies in the Remix project, but also on data acquired on single crop systems including weeds or not. The contribution of ancillary information (RGB images in two viewing directions, 3D LiDAR point cloud) as input to the deep models will be also investigated. Finally, once the species are identified, the project aims to estimate new traits such as the proportion of species, the corresponding leaf area or the overlapping area between these species, as well as to use the dynamics of these traits to identify key events such as the date of species cover.


  • Inria-Zenith
  • AGIR