[PhD student] Frédérick Fabre Ferber : Small datasets and prediction in Artificial Intelligence, towards a better cohabitation

Small datasets and prediction in Artificial Intelligence, towards a better cohabitation: Application to the sustainable management of weediness in agricultural systems in La Réunion

I did my training at the University of La Réunion, with a Master's degree in Computer Science where I specialized in Machine Learning. During my M2 internship, I worked on the prediction of sugarcane weeds using supervised learning methods. This internship gave me a taste for research into a computer application for agronomy, and naturally I decided to do a thesis to go further and solve these kinds of problems, which are important to me as a native of the island.

My subject concerns the processing of agronomic datasets, most of which are unsuitable for supervised learning, using mathematical, statistical and computer science methods, with the aim of improving the performance of learning algorithms in this data context.

• Starting date : March 2022
• Research unit: UPR Recyclage et Risque, Cirad
• University : Université de La Réunion
• PhD school : Sciences Technologies et Santé - ED STS
• Scientific field : IT
• Thesis management : Jean-Christophe Soulié, UPR Recyclages et risques, Cirad et Jean Diatta, Université de la Réunion
• Thesis supervisors: Odalric-Ambrym Maillard – Dominique Gay – Thomas LeBourgeois – Sandrine Auzoux
• Funding: Bourse Régionale de recherche (La Réunion)
• #DigitAg : Labeled thesis

Keywords : Machine Learning, Core methods, Agronomy

Abstract: Several studies have been conducted on weed management. The Deci-Florsys project determines weed dynamics by simulation using agro-environmental indicators. Another project concerns the recognition of weeds by spectral image analysis. Machine learning algorithms are used to identify and discriminate the different species. However, they do not consider the tropical weed flora and do not try to directly predict weediness. The thesis will apply artificial intelligence to predict the weed flora of agricultural systems in Reunion in a tropical context. We present a non-exhaustive list of some works that will be completed during the thesis. Different scientific issues are identified that impact the performance of prediction algorithms on small observed data sets. Different concepts concern the adaptation of learning algorithms to take into account missing values, their sensitivity in a context of unbalanced data with high bias (fairness), the exploitation of relationships between the variables to be predicted and finally, the heteroscedastic aspect of the data. In the literature, a certain gap is noted between the classical works on prediction and the concepts mentioned above. However, these works have been carried out on specific tasks using complete data sets. This thesis will contribute to enrich the knowledge on these recent concepts in the literature and to apply them on small observed data sets