Summary

The diversification of agricultural systems is crucial for agroecological transition and adaptation to climate change, requiring better use of genetic resources tailored to specific crops. Hybrid varieties have a strong potential for adaptation, but several obstacles limit their analysis. Available data are heterogeneous and often incomplete, preventing their efficient combined use. Additionally, only a small proportion of possible hybrid combinations have been characterized, and lineage populations are imbalanced in terms of available data. The proposed thesis aims to develop statistical learning methods to predict the performance of maize hybrids, focusing on the stability and variability of performance, as well as direct prediction of performance in different environments. These methods will utilize advanced statistical models known for their predictive capacity and interpretability, and will integrate neural networks to enhance large-scale analysis and enable transfer learning between hybrids. This work should significantly contribute to improving maize yields in Europe, which are particularly affected by climatic variations.

Contact :

• François Victor