Post-doctoral researcher in computer science / deep learning / remote sensing

The agroecological transition requires the development and the assessment of new multiperformant, resilient and sustainable agroecosystems. For this purpose, high-throughput observation tools based on close range imagery and LiDAR appear as essential for rapidly characterizing the state and the dynamics of such cropping systems, which are often mixed crops. In this context, the Deep4Mix project aims to evaluate the potential of deep learning algorithms and various sensors (RGB cameras and LiDAR) for field monitoring of the dynamics of the proportion and structure of species in a crop mixture.

You will be working at the joint research units CAPTE (CAPTeurs et TElédetection ) and EMMAH (Environnement Méditerranéen et Modélisation des AgroHydrosystèmes) in the INRAE center of Avignon. CAPTE develops observation systems (vectors and sensors) and data processing methods for high-throughput crop phenotyping. You will be working closely with CAPTE researchers and engineers, as well as with other partners of the Deep4Mix project : the research units LIRMM (Laboratoire d'Informatique, de Robotique et de Microélectronique de Montpellier) and AMAP (botAnique et Modélisation de l'Architecture des Plantes et des végétations), specialized in deep learning and funders of the API Pl@ntNet , and the research unit AGIR (Agoécologie – Innovations - Territoires), specialized in agroecology.

Your mission will be to develop deep learning / machine learning methods that will make it possible, first, to identify the different plant species in the RGB images, and, second, to characterize their structures.

More specifically, you will have to :

• Build a dataset of annotated RGB images to train deep learning algorithms. You will gather datasets already annotated by the different partners in previous studies, and datasets published in the literature. You will also be in charge of the annotation of additional images, together with INRAE technicians.
• Develop deep learning models to identify plant species in the images. You will compare various segmentation model architectures (semantic segmentation and instance segmentation), as well as various training strategies and data augmentation methods. The contribution of plant height information derived from LiDAR will also be investigated.
• Develop deep learning or machine learning based algorithms to estimate the proportion and the leaf area of every species, based on the segmentation results obtained at the previous step and the LiDAR data. The estimation results will be assessed using the available destructive measurements.

Special conditions of activity: short trips to Montpellier and/or Toulouse will be possible to interact with the other partners of the Deep4Mix project.