Deep Learning for Spatio-Temporal Pesticide Exposure Modeling: Extrapolation, Interpretability, and Constraints

Jury

• Romain Hérault, Professor, GREYC Laboratory & University of Caen Normandy, Reviewer 
• Paul Honeine, Professor, LITIS Lab & University of Rouen Normandy, Reviewer 
• Carole Bedos, Research Director, ECOSYS, INRAE, Examiner
• Yann Chevaleyre, Professor, LAMSADE, Paris-Dauphine University PSL, Examiner
• Celine Hudelot, Professor, MICS, CentraleSupélec, Examiner
• Olivier Allais, Research Director, Inrae, Invited Jury
• Julia Mink, Tenure Track Assistant Professor, University of Bonn, Invited Jury

Abstract

This thesis presents a deep learning pipeline for estimating airborne pesticide exposure in France. We begin by characterizing the pesticide transfer pathway, from agricultural practices to health effects. We then analyze the CNEP monitoring campaign and introduce difficulty indices for sensors and substances, in order to better characterize sources of error and heterogeneity. Building on this analysis, we construct enriched representations that incorporate the agricultural context, pesticide purchase data, and pedoclimatic variables, together with spatial encodings designed to capture local variability.
We then propose joint deep estimators tailored to sparse and censored measurements, which we validate using dedicated protocols complemented by an assessment of extrapolation risk. We highlight and diagnose monotonicity violations induced by perturbations of the input variables. Beyond prediction, the manuscript addresses the question of trust, both through a proposed conceptual interpretability pipeline and through a unified formulation of learning under operator constraints, linking our monotonicity analyses to objectives grounded in physical equations. Finally, we study the sources of stiffness in constrained learning and propose an improvement to a natural-gradient optimizer based on adaptive regularization. Overall, this thesis argues for hybrid “AIArtificial Intelligence for Science” systems that integrate expert knowledge, both in the representations and in the learning constraints, in order to produce robust and reliable estimates.