Thesis of Ambre Mychalski


Subject:
Multimodal Artificial Intelligence Methods for Varietal Identification of Wheat and Sorghum Using Images of Living Plants and Herbarium Specimen

Start date: 01/11/2026
End date (estimated): 01/11/2029

Advisor: Laure Tougne Rodet

Summary:

The PhD project is part of the PEPR Past2ECO programme, which aims to elucidate the genetic and evolutionary mechanisms underlying crop adaptation to climate change by integrating the historical and contemporary diversity of wheat and sorghum. The challenge is to link genetic, phenotypic and environmental diversity, together with agricultural practices, in order to predict and harness adaptation for sustainable agroecological systems. Within this framework, it is essential to make the most of both historical and contemporary biological resources related to wheat and sorghum.

Regarding historical resources, herbarium collections represent an exceptional source of information, providing access to crop biodiversity from past centuries (Meineke et al., 2018). However, these data remain largely underexploited because verifying (e.g. correcting approximate identifications in some herbarium collections), processing and extracting information from historical specimens cannot be carried out manually, given the sheer volume of data and the level of taxonomic expertise required. AI-based identification at the varietal level would make it possible both to curate these herbarium collections and to extract taxonomic and morphological information about crops over several centuries. This would offer a unique opportunity to characterise historical cultivated varieties and identify the impact of climatic and anthropogenic changes on them.

Regarding contemporary resources, identifying varieties in current crops makes it possible to analyse phenotypic plasticity in cultivated systems, particularly in varietal mixtures and complex agroecosystems. Here too, manual identification is extremely challenging due to the volume of data and the level of expertise required. AI-based taxonomic identification at the varietal level would therefore make it possible to extract and exploit complementary information from herbarium specimens and modern crops, in order to identify traits of agronomic interest and gain a better understanding of how wheat and sorghum respond to environmental changes.

The PhD project aims to advance computer vision for the taxonomic identification of plants at the varietal level, using images of both living plants and herbarium specimens. It seeks to significantly improve taxonomic identification performance and to gain a better understanding of the attention mechanisms employed by neural networks during the identification process. In particular, it may contribute to the curation of the herbarium collection used within the framework of the Past2ECO project. To achieve these objectives, the thesis will focus on wheat and sorghum, for which variety identification and characterisation are particularly challenging.

The PhD project is structured around four main tasks, which may evolve over time depending on the doctoral candidate's interests and the progress of the state of the art.

1. Addressing domain adaptation challenges between images of living plants and herbarium specimens through supervised generation approaches (Goodfellow et al., 2020; Jing et al., 2019; Shorten and Khoshgoftaar, 2019) and unsupervised approaches (Ajakan et al., 2014; Wang et al., 2020).

2. Leveraging Pl@ntNet data and models (Joly et al., 2016; Lefort et al., 2024) to expand the PEPR Past2ECO database and implement various transfer learning approaches (Azad et al., 2024), with the aim of developing new models and/or new Pl@ntNet branches specialised in identifying the varieties targeted by the PEPR Past2ECO project.

3. Investigating the attention mechanisms of the AI models developed during the PhD project for variety identification, in order to highlight the morphological structures that are characteristic of each variety (Covert et al., 2023).

4. Exploring the relationships between phenotype (images and geometric morphometric data) and genotype (DNA) using multimodal AI models such as CLIP (Radford et al., 2021).